ELEMENTS OF MATHEMATICS

Springer
Berlin
Heidelberg
New York
Hong Kong
London
Milan
Paris
Tokyo
 NICOLAS BOURBAKI

Topological Vector Spaces

Chapters 1-5

Translated by H.G. Eggleston & S. Madan

Springer
 Originally published as
ESPACES VECTORIELS TOPOLOGIQUES
Masson, Paris, 1981

Mathematics Subject Classification (2000):

46AXX, 46-02, 46EXX, 46C05, 46M05

Cataloging-in-Publication data applied for

Die Deutsche Bibliothek - CIP-Einheitsaufnahme
Bourbaki, Nicolas:
Elements of mathematics / N. Bourbaki. - Berlin ; Heidelberg; New York;
London; Paris; Tokyo ; Hong Kong : Springer
Einheitssacht.: Elements de mathematique <eng!.>
Topological vector spaces
Ch. 1/5. / trans!. by H. G. Eggleston & S. Madan. - 2003

Softcover printing of the 1St English edition of 1987

ISBN 978-3-540-42338-6 ISBN 978-3-642-61715-7 (eBook)

DOI 10.1007/978-3-642-61715-7

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned,
speeifically the rights of translation, reprinting, reuse of illustrations, reeitation, broadcasting, reproduction on
microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is
permitted only under the provisions of the German Copyright Law of September 9, '965, in its current version, and
permission for use must always be obtained from Springer-Verlag. Violations are liable for prosecution under the
German Copyright Law.
Springer-Verlag Berlin Heidelberg New York
a member of BertelsmannSpringer Seience+ Business Media GmbH
http://www.springer.de
© Springer-Verlag Berlin Heidelberg 2003

The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in
the absence of a speeific statement, that such names are exempt from the relevant pro-
tective laws and regulations and therefore free for general use.
Cover Design: Design & Production GmbH, Heidelberg
Printed on aeid-free paper SPIN: 10843963 41/3142 db 543210
 To the reader

1. The Elements of Mathematics Series takes up mathematics at the beginning,

and gives complete proofs. In principle, it requires no particula knowledge of
mathematics on the readers' part, but only a certain familiarity witn mathematical
reasoning and a certain capacity for abstract thought. Nevertheless, it is directed
especially to those who have a good knowledge of at least the content of the first
year or two of a university mathematics course.
2. The method of exposition we have chosen is axiomatic, and normally proceeds
from the general to the particular. The demands of proof impose a rigorously fixed
order on the subject matter. It follows that the utility of certain considerations will
not be immediately apparent to the reader unless he has already a fairly extended
knowledge of mathematics.
3. The series is divided into Books and each Book into chapters. The Books
already published, either in whole or in part, in the French edition, are listed below.
When an English translation is available, the corresponding English title is men-
tioned between parentheses. Throughout the volume a reference indicates the English
edition, when available, and the French edition otherwise.
Theorie des Ensembles (Theory of Sets) designated by E (S)
Algebre (Algebra(l) A (A)
Topologie Generale (General Topology) TG (GT)
F onctions d'une Variable Reelle FVR
Espaces Vectoriels Topologiques (Topological Vec-
tor Spaces) EVT (TVS)
Integration INT
Algebre Commutative (Commutative Algebra(2) AC (CA)
Varietes Diflhentielles et Analytiques VAR
Groupes et Algebres de Lie (Lie Groups and Lie
Algebras (3) LIE (LIE)
Theories Spectrales TS

e) So far, chapters I to III only have been translated.

CZ) SO far, chapters I to VII only have been translated.
e) So far, chapters I to III only have been translated.
VI TOPOLOGICAL VECTOR SPACES

In the first six books (according to the above order), every statement in the text
assumes as known only those results which have already been discussed in the same
chapter, or in the previous chapters ordered as follows: S; A, chapters I to III; GT,
chapters I to III; A, from chapters IV on; GT, from chapter IV on; FVR; TVS;
INT.
From the seventh Book on, the reader will usually find a precise indication of its
logical relationship to the other Books (the first six Books being always assumed
to be known).
4. However we have sometimes inserted examples in the text which refer to facts
the reader may already know but which have not yet been discussed in the series.
Such examples are placed between two asterisks: *... *. Most readers will undoub-
tedly find that these examples will help them to understand the text. In other cases,
the passages between *... * refer to results which are discussed elsewhere in the
Series. We hope the reader will be able to verify the absence of any vicious circle.
5. The logical framework of each chapter consists of the definitions, the axioms,
and the theorems of the chapter. These are the parts that have mainly to be borne
in mind for subsequent use. Less important results and those which can easily be
deduced from the theorems are labelled as « propositions », «lemmas », « corolla-
ries », « remarks », etc. Those which may be omitted at a first reading are printed in
small type. A commentary on a particularly important theorem appears occasionally
under the name of « scholium ».
To avoid tedious repetitions it is sometimes convenient to introduce notations or
abbreviations which are in force only within a certain chapter or a certain section
of a chapter (for example, in a chapter which is concerned only with commutative
rings, the word« ring» would always signify« commutative ring »). Such conventions
are always explicitly mentioned, generally at the beginning of the chapter in which
they occur.

errors. These passages are signposted in the margin with the sign
bend »).
Z
6. Some passages in the text are designed to forewarn the reader against serious
(<< dangerous

7. The Exercises are designed both to enable the reader to satisfy himself that he
has digested the text and to bring to his notice results which have no place in the
text but which are nonetheless of interest. The most difficult exercises bear the sign ~.
8. In general, we have adhered to the commonly accepted terminology, except
where there appeared to be good reasons for deviating from it.
9. We have made a particular effort always to use rigorously correct language,
without sacrificing simplicity. As far as possible we have drawn attention in the
text to abuses of language, without which any mathematical text runs the risk of
pedantry, not to say unreadability.
10. Since in principle the text consists of the dogmatic exposition of a theory,
it contains in general no references to the literature. Bibliographical references are
 TO THE READER VII

gathered together in Historical Notes. The bibliography which follows each historical
note contains in general only those books and original memoirs which have been
of the greatest importance in the evolution of the theory under discussion. It makes
no sort of pretence to completeness.
As to the exercises, we have not thought it worthwhile in general to indicate
their origins, since they have been taken from many different sources (original
papers, textbooks, collections of exercises).
11. In the present Book, references to theorems, axioms, definitions, ... are given
by quoting successively :
- the Book (using the abbreviation listed in Section 3), chapter and page, where
they can be found, when referring to the French edition;
- the chapter and page only when referring to the present Book;
- the chapter, paragraph and section, when referring to the English edition.
The Summaries of Results are quoted by the letter R; thus Set Theory, R signifies
« Summary of Results of the Theory of Sets ».
 CHAPTER I

Topological vector spaces

over a valued division ring
§ 1. TOPOLOGICAL VECTOR SPACES

1. Definition of a topological vector space

DEFINITION 1. - Given a topological division ring K (GT, III, § 6.7) and a set E
such that E has
10 the structure of a left vector space on K;
20 a topology compatible with the structure of the additive group of E (GT, III,
§ 1. 1) and satisfying in addition the following axiom :
(EVT) the mapping (A, x) ~ AX of K x E in E is continuous,
then E is called a left topological vector space over (or on) K.
It is equivalent to saying that E is a topological left K-module (GT, III, § 6.6).
A left vector space structure relative to K and a given topology on a set E, are
said to be compatible if the topology and the additive group structure ofE are compa-
tible and if, in addition, the axiom (EVT) is valid. This is the same as saying that
the two mappings (x, y) ~ x + y and (A, x) ~ AX of E x E and of K x E, respec-
tively, in E are continuous, for then the mapping x ~ - x = ( - 1) x, is continuous
and the topology of E is compatible with its additive group structure.
If E is a left topological vector space over K, we say that E provided only with
its vector space structure, underlies the topological vector space E.

Examples. - 1) If E is a left vector space over a discrete topological division ring K,

the discrete topology on E is compatible with the vector space structure of E (this is
not so if K is non-discrete and E is not the single point 0).
2) Let A be a topological ring (GT, III, § 6.3) and let K be a sub ring of A that is
also a division ring and such that the topology induced on K by that of A is compatible
with the division ring structure of K; then the topology of A is compatible with its
left vector space structure on K.
3) Let K be any topological division ring and I an arbitrary set. On the product
vector space K~ (A, II, § 1 .5), the product topology is compatible with the vector space
structure (GT, III, § 6.4). Or we can say that the space K! of mappings of I in K with
pointwise or simple convergence topology is a topological vector space on K (TG,
X, p. 4).
4) Let X be a topological space; on the set E = ~(X; R) of finite real-valued conti-
nuous functions defined over X, the compact convergence topology (GT, X, § 1.3) is
TVS 1.2 TOPOLOGICAL VECTOR SPACES §1

compatible with the vector space structure of E on R. For, let U o be a point of E, let H
be a compact subset of X and I': be an arbitrary strictly positive number. The real-valued
function U o is bounded in H; let a = sup luo(t)1 ; if u is any point of E then for all t E H
tEH

IAU(t) - Aouo(t)1 :0( IAI·lu(t) - uo(t) I + alA - Aol·

Hence, if IA - Aol :0( I': and lu(t) - uo(t) I :0( I': for all t E H, then for t E H,
IAU(t) - AOUO(t) I :0( 1':(1': + IAol + a), which shows that the axiom (EVT) is satisfied;
similarly it can be verified that the compact convergence topology is compatible with
the additive group structure of E.
On the other hand, if X is not compact, the uniform convergence topology (in X)
is not necessarily compatible with the vector space structure of E ; for example if X = R
and if U o is an unbounded continuous function in R, then the mapping A f-7 'Au o of R
in E is not continuous in the uniform convergence topology on E.
5) Let E be a vector space of finite dimension n over a topological division ring K;
there exists an isomorphism u: K~ ---+ E of vector K-spaces and moreover, if v is a
second isomorphism of K~ on E, then we can write v = u 0 f, where f is an automor-
phism of the vector K-space K~. Consider, on K~, the product topology that is compa-
tible with its vector space structure (Example 3); since every linear mapping of K~
in itself is continuous for this topology, every automorphism of the vector space K~
is bicontinuous. Hence, if we transfer the product topology of K~ to E, by means of
any isomorphism whatever of K~ on E, the topology obtained on E is independent
of the particular isomorphism used; we call it the canonical topology on E; we shall
characterize it differently (1, § 1.3) when K is a non-discrete complete division ring
with a valuation. Every linear mapping of E in a topological vector space over K is
continuous for the canonical topology on E.

In the same way as in def 1, a right topological vector space over K, a topological
division ring, can be defined; but every right vector space on K can be considered
as a left vector space on the division ring KO opposite to K (A, II, § 1.1) and the topo-
logy of K is compatible with the structure of the division ring KO. For this reason
we usually consider only left topological vector spaces; when we speak of « topo-
logical vector space» without qualification, it is to be understood that we refer to
a left vector space.
If K' is a sub-division ring of K, and E a topological vector space over K, then
it is clear that the topology of E is still compatible with the vector space structure
of E relative to K', obtained by restricting the field of scalars to K'; we say that
the topological vector space on K', obtained by this procedure, underlies the topo-
logical vector space E on K.
In order that a topological vector space E be Hausdorff, it is necessary and suffi-
cient that for all x#-O of E, there exists a neighbourhood of 0 not containing x
(GT, III, § l.2).
Consider a topology, on a vector space E over a topological division ring K,
that is compatible with the additive group structure of E. Because of the identity
Ax - AoXo = (A - Ao) Xo + Ao(X - x o) + (A - Ao) (x - x o)

axiom (EVT) is equivalent to the following system of three axioms.

(EVT;) For all Xo E E, the mapping A f---+ AXo is continuous at A = O.
(EVT;I) For all Ao E K, the mapping x f---+ AoX is continuous at x = o.
(EVT;n) The mapping (A, x) f---+ Ax is continuous at (0, 0).
No.2 TOPOLOGICAL VECTOR SPACES TVS 1.3

In particular :

PROPOSITION 1. - For all r:x E K and every point bEE, the mapping x ~ r:xx + b
of E in itself is continuous. Further, if r:x #- 0, this mapping is a homeomorphism of E
on itself.
The second part of the proposition is a result of the fact that if r:x #- 0, then
x ~ r:x- 1 x - r:x-1b is the inverse mapping of x ~ r:xx + b.

If A is an open (resp. closed) set in E, then r:xA is open (resp. closed)

°
COROLLARY. -
in E for every r:x #- in K.
Let E and F be two topological vector spaces on the same topological division
ring K. A bijection f of E on F is an isomorphism of the topological vector space E
on the topological vector space F if and only if f is linear and bicontinuous. In parti-
cular, if y#-O belongs to the centre of K, the homothety x ~ yx is an automorphism
of the topological vector space structure of E.

2. Normed spaces on a valued division ring

°
Recall (GT, IX. § 3.2) that an absolute value on a division ring K is a mapping
~ ~ I~I of K in R+. such that I~I = if. and only if, ~ = 0, and that I~ll I = 1~1·lll I,
and I~ + 11 I ~ I~ I + III I; an absolute value defines a distance I~ - 11 I on K, and
hence a Hausdorff topology compatible with the division ring structure of K. If
I~ I = 1 for all ~ #- 0, the absolute value is called improper, and the topology that
it defines on K is the discrete topology; if, on the other hand, there exists r:x #- °
°
in K such that Ir:xl #- 1. then there exists ~ #- in K such that I~I < 1 (it is sufficient
to take ~ = r:x or ~ = r:x - 1). and the sequence (~n)n ~ 1 converges to 0, thus the topo-
logy of K is not discrete.
We recall on the other hand (GT, IX. § 3.3) that if E is a vector space on a non-
discrete valued division ring K then a norm on E is a mapping x ~ Ilxll of E in R+.
such that Ilxll = °
if. and only if, x = 0, and such that IIAxII = IAI.llxll for every
scalar AE K. and Ilx + yll ~ Ilxll + Ilyli. A distance Ilx - yll, is defined on E by
the norm, and hence a topology that is compatible with the vector space structure
of E (loc. cit.). Unless the contrary is expressly stated, a normed space is considered
in terms of the structure of the topological vector space defined by its norm. The
normed spaces are among the most important of topological vector spaces.

It is known (GT, IX, § 3.3) that two distinct norms on E can define the same topo-
logy on E; for this it is necessary and sufficient that the two norms be equivalent (lac.
cit.). The structure of normed spaces is thus richer than the structure of topological
vector spaces; if E and F are two normed spaces one must be careful to distinguish
between the idea of isomorphism of the normed space structure of E with that of F,
and the idea of isomorphism of the topological vector space structure of E with that
of F.
TVS 1.4 TOPOLOGICAL VECTOR SPACES §1

Example. - Let I be an arbitrary set of indices; it is known (GT, X, § 3.2) that a norm
Ilxll can be defined, on the set of bounded mappings x = (~,) of I in K, '@(I; K) (also
written .@K(I) or f;(I»), by Ilxll = sup I~J When I is a topological space, the set of
'EI
bounded, continuous mappings of I in K is a closed subspace of the space .@(I; K)
(GT, X, § 3.1, cor. 2). Another subspace of .@(I; K) is the setfi(I) of absolutely summable
families x = (~) (GT, X, § 3.6); we can define on this subspace another norm
Ilxlll = I I~,I, that in general is not equivalent to the norm Ilxll = sup I~,I (I, p. 23,
lEI lEI

exerc. 6); when considering fiO) as a normed space, without specifying its norm,
it is always the norm Ilxll, that is meant. We write .@(I) and £1(I) in place of .@(I; R)
andf~(I).

3. Vector subspaces and quotient spaces of a topological vector space; products of

topological vector spaces; topological direct sums of subspaces

Everything that has been said for topological modules (GT, III, § 6.6) applies
in particular to topological vector spaces. If M is a vector subspace of a topological
vector space E, the topology induced on M by that of E is compatible with the
vector space structure of M, and the closure M of M in E is a vector subspace of E.
The quotient topology of that of E by M is compatible with the vector space structure
ofE/M.
If E is a topological vector space, the closure N of {O} in E (intersection of neigh-
bourhoods of 0) is a closed vector subspace of E; the quotient vector subspace E/N,
which is necessarily Hausdorff whether E is or not, is called the Hausdorff vector
space associated with E.
Let (EJ'Ei be a family of topological vector spaces over the same topological
division ring K, and let E be the product vector space of the E,. The product topo-
logy of the topologies of the E, is compatible with the vector space structure of E.
In the product space E, the subspace F, the direct sum of the E, is everywhere dense
(GT, III, § 2.9, prop. 25).

F or certain types of topological vector spaces on the field R or the field C we define
(in II, p. 29) a topology on the direct sum of a family (E) of topological vector spaces
that is, in general, distinct from the topology induced by the product topology of the E,.

Everything that has been said on the finite direct sums of stable subgroups of
topological groups with operators (GT, III, § 6.2) applies to topological vector spaces,
replacing « stable subgroup» throughout by «vector subspace ».

Remark. - Given a closed vector subspace M ofa Hausdorff topological vector space E,
it is not necessarily the case that there exists an (algebraic) complementary vector
subspace to M that is closed in E (even ifE is a normed space; cf IV, p. 55, exerc. 16 (c»);
a fortiori there does not necessarily exist a topological complement of M in E (cf I,
p. 26, exerc. 8). However we shall see in § 2 that when K is a non-discrete valued division
ring, then every closed subspace M ofE, with finite co dimension, does have a topological
complement in E (I, p, 14, prop. 3).
No.4 TOPOLOGICAL VECTOR SPACES TVS I.5

4. Uniform structure and completion of a topological vector space

Since the topology of the topological vector space E is compatible with the addi-
tive group structure on E, it defines a uniform structure on E (GT, III, § 3); when
we speak of the uniform structure of a topological vector space we always mean
this structure unless the contrary is expressly stated. Every continuous linear mapping
of a topological vector space E in a topological vector space F is uniformly continuous
(GT, III, § 3.1, prop. 3); every mapping of E in itself of the form x ~ r:lX + b is
uniformly continuous. An equicontinuous set of linear mappings of E in F is uniformly
equicon tinuous (G T, X, § 2.2, prop. 5).
Remarks. - 1) If B is a precompact set of K, then for every neighbourhood V of 0
in E, there is a neighbourhood U of 0 in E such that BU c V. For, if W is a neigh-
bourhood of 0 in E such that W + W c V; then from (EVT;n) there is a neighbourhood
To of 0 in K and a neighbourhood U 0 of 0 in E such that To U 0 c W. As B is precompact,
there are finitely many points Ai E B (1 ~ i ~ n) such that the Ai + To cover B; from
(EVT;I) it follows that there is a neighbourhood U c U 0 of 0 in E, such that Ai U c W
for all i; clearly U has the required properties. In a similar manner (using (EVT;)
instead of (EVT;I)) it can be shown that if H is a precompact set of E, then for every
neighbourhood V of 0 in E, there exists a neighbourhood T of 0 in K such that TH c V.
2) From 1) it follows that, if B is a precompact set of K and H is a precompact set
of E, then the mapping (A, x) ~ AX restricted to B x H is uniformly continuous. For,
if V is a neighbourhood of 0 in E then there are neighbourhoods T of 0 in K, and U
of 0 in E such that TH + BU c V. Since we can write Ax - A'x' = (A - A') x + A'(x - x'),
we s.ee that for A, A' in B, x, x' in H, A - A' E T and x - x' E U, we have AX - A'x' E V,
which proves our assertion.

A topological vector space is called complete if, considering its uniform structure,
it is a complete uniform space.

DEFINITION 2. - A complete normed space on a non-discrete valued division ring is

called a Banach space.

Examples. - If K is a non-discrete valued division ring then the space .?l(l; K) (I, p. 4,
Example) is complete (OT, X, § 3.1, cor. 1). This is also true for the space Rk(l) (I, p. 4,
Example) with the norm Ilxll! = I I~,I : for, if xn is a Cauchy sequence in this space and
'EI
xn = (~n.)'EI' then for alii E I

thus, for each I E I, the sequence (~nt)n" 1 converges to a limit ~, in K. Further, for each
finite subset J of I

and it follows immediately that there exists a constant a > 0, independent of J, m, n

such that I I~ml - ~nll ~ a. Letting m tend to + 00, we deduce I I~, - ~ml ~ E
~ ~

from which IlEI I~,l ~ a + Ilxnll \0 which shows that z = (~J'EI belongs to Rk(I); further,
for all E > 0, there exists no such that for n ;;, no and for every finite set J of I, we have
I I~, - ~rnl ~ E; passing to the limit with respect to the directed set of finite subsets of!,
'E J
TVS I.6 TOPOLOGICAL VECTOR SPACES §1

we see that liz - x.lll ~ E for n ~ no, which shows that z is the limit ofthe sequence (x.)
in the normed space£~.CI).

Let K be a Hausdorff topological division ring, E a topological vector space on K

and suppose that the completed ring K is a division ring (this is so when K is a valued
division ring, GT, IX, § 3.3) then the Hausdorff completion E of E carries the struc-
ture of a complete topological vector space on K (GT, III, § 6.5); we say that E, with
this structure, is the Hausdorff completion ofthe topological vector space E, or simply
the completion ofE when E is Hausdorff.

5. Neighbourhoods of the origin in a topological vector space over a valued division

ring

DEFINITION 3. - Let K be a valued division ring and E a left vector space over K ;
we say that a subset M of E is balanced if, for all x E M and all A E K such that IAI ::::; 1,
it is true that AX E M (or in other words if AM c M when IAI ::::; 1).

PROPOSITION 2. - I n a topological vector space E over a valued division ring K, the

closure of a balanced set M, is a balanced set.

If B is the set of ~ E K with I~I ::::; 1 ; then B is closed in K. But B x M is mapped

into M by the continuous mapping (A, x) H AX; and therefore B x M is mapped
into M (GT, I, § 2.1, tho 1) which proves that M is balanced.
When M is an arbitrary set in the vector space E over a valued division ring K,
the set M1 of the AX with x E M and A E K such that IAI ::::; 1, is clearly the smallest
balanced set containing M; Ml is called the balanced envelope ofM.

PROPOSITION 3. - Let K be a valued locally compact and non-discrete division ring

and E be a Hausdorff topological vector space (resp. a topological vector space) over K.
For every compact (resp. precompact) set H in E, the balanced envelope of H is compact
(resp. precompact).
If B denotes the ball I~I ::::; 1 in K, the balanced envelope of H is H 1, the image
of B x H under the continuous mapping m: (A, x) H AX. If E is Hausdorff, if B is
compact and ifH is compact then so is B x H and therefore H l . IfH is precompact
the restriction of m to B x H is uniformly continuous (I, p. 5, Remark 2) and as
B x H is precompact, so also is its image under m (G T, II, § 4.2, prop. 2).

Note that the balanced envelope of a closed set is not necessarily closed. For example,
in R2, the balanced envelope of the hyperbqla defined by the equation xy = 1 is not
closed.

The union of a family of balanced sets in E is balanced, which implies that for
every set M of E there is a largest balanced subset N of M called the balanced core of
M; also N is not empty if and only if 0 E M. To say that x E N means that for all
A E K such that IAI ::::; 1, we have Ax EM, or again (if 0 E M) that, for allll E K with
No.5 TOPOLOGICAL VECTOR SPACES TVS 1.7

Ij.ll ~ 1, we have x E j.lM. If 0 EM, the balanced core N of M is therefore the inter-
section n j.lM. This shows in particular that if M is closed, so also is N.
11-'1'" 1

DEFINITION 4. - Let K be a non-discrete valued division ring and E be a left vector

space on K with two subsets A and B. We say that A absorbs B if there exists r:i > 0
such that AA ::::J B for every A E K with IAI ~ r:i (or equivalently tf j.lB c A for j.l #- 0
and Ij.ll :::; r:i - 1). A set A of E is called absorbent if it absorbs every set consisting of a
single point.

Let A be a balanced set of E; for it to absorb a set B of E it is sufficient that there

exists A#-O such that AA ::::J B; in fact, for Ij.ll ~ IAI, we have AA = (Aj.l-1) j.lA, and
as j.lA is balanced and IAj.l- 11 :::; 1, it follows that AA c j.lA, and thus B c j.lA. In
particular for a balanced set A of E to be absorbent, it is necessary and sufficient
that for every x E E, there exists A#-O in K such that AX E A. Every absorbent set
of E generates the vector space E. Every finite intersection of absorbent sets is an
absorbent set.

PROPOSITION 4. - Ina topological vector space E on a non-discrete valued division

ring K there exists a fundamental system mof closed neighbourhoods of 0 such that :
(EV,) Every set V Em is balanced and absorbent.
(EVIl) For every V E m and A#-O in K, we have AV E m (in variance of m under
homotheties of non zero ratio).
(EV m) For every V Em, there exists WE m such that W + W c V.
Conversely, let E be a vector space on K, and let m be a filter base on E satisfying
the conditions (EV,), (EVIl) and (EVm)' Then there exists a topology (and it is unique)
on E, compatible with the vector space structure ofE, andfor which m
is a fundamental
system of neighbourhoods of O.

By axiom (EVT;I1) we show firstly that the balanced core, V1, of V, a neighbourhood
of 0, is itself a neighbourhood of O. For there exist r:i > 0 and a neighbourhood W
of 0 such that, if IAI :::; r:i and x E W, then AX E V. Since K is non-discrete, there exists
j.l #- 0 in K with Ij.ll :::; r:i and j.l W is a neighbourhood of 0 for which j.lW c V. Also
if v E K and Ivi :::; 1 then IVj.l1 :::; r:i and thus Vj.lW ::::J V. Hence j.lW ::::J Vi and V 1
is a neighbourhood of O. Also as V is closed so also V 1 is closed. Thus the set m of
closed balanced neighbourhoods of 0 form a fundamental system of neighbourhoods
of 0 in E. By (EVT;) every neighbourhood of 0 is absorbent; furthermore m satisfies
(EVIl) (cf I, p. 3, cor.); finally, because of the continuity of (x, y) I----> X + y at the
point (0, 0), every fundamental system of neighbourhoods of 0 in E satisfies (EVm).
The set m satisfies the conditions of the proposition.
Now let E be a vector space over K, and m be a filter base on E satisfying (EV,),
(EVIl) and (EV m). The axiom (EV,) shows firstly that for all V E m, we have - V = V
and 0 E V ; these relations and the axiom (EVm) show that m is a fundamental system
of neighbourhoods of 0, for a topology on E compatible with the additive group
TVS 1.8 TOPOLOGICAL VECTOR SPACES §1

structure of E (GT, III, § 1.2). On the other hand the axioms (EVT;), (EVT;,) and
(EVT;n) are immediate consequences of (EV,) and (EV n), thus the topology defined
above satisfies the axiom (EVT), and the proposition is proved.

Remarks. - 1) In a normed space on a non-discrete valued division ring the set of

open balls (resp. closed balls) with centre 0 is a fundamental system of neighbourhoods of
o which satisfy the conditions (EV1), (EVu) and (EV m).
2) When the division ring of scalars K is the field R or the field C, every filter base
m on E which satisfies just the two axioms (EV1) and (EV m) is a fundamental system
of neighbourhoods of 0 for a topology compatible with the vector space structure of E.
In fact, we need only prove that, in these conditions, for every A # 0 in K and every
V Em there exists WE m such that AW c V. Now from (EV m) there exists WI Em
with 2 WI C V, and we deduce, inductively, that for every positive integer n, there exists
Wn E m such that 2nW n c V. As V is balanced, if we take n so large that 2n = 12n l > IAI,
then W = Wn satisfies the condition, as required.
This result does not hold for every non-discrete valued division ring K, for in such a
division ring it is no longer necessarily true that Imtl = m for every positive integer m
(t indicates the unit element of the division ring; cf I, p. 22, exerc. 1).
3) If K is a discrete division ring, conditions (EVTD and (EVT;u) are true for any
topology on E. Arguing as in prop. 4, one easily sees that if E is a topological vector
space on K, then there exists m, a fundamental system of closed neighbourhoods of 0
in E satisfying conditions (EVu) and (EV m). Conversely, if a filter base m on a vector
space E over K is such that 0 belongs to all the sets of m and (EVu), (EVm) are true,
then m is a fundamental system of neighbourhoods of 0 in a topology compatible
with the vector space structure of E.

6. Criteria of continuity and equicontinuity

Let E and F be topological vector spaces over the same division ring K; for a
linear mapping f of E in F to be continuous, it is sufficient for it to be continuous
at the origin (GT, III, § 2.8, prop. 23). This proposition generalizes as follows :

PROPOSITION 5. - Let Ei (1 ~ i ~ n) and F be topological vector spaces on a non-

n
discrete valuedfieldK. In order that a multilinear mappingfof TI Ei in F should be conti-
i= 1
n
nuous in the product space TI Ei it is sufficient for it to be continuous at (0, 0, ... , 0).
i= 1

n
Let (al' a2 , ... , an) be an arbitrary point of TI E i ; we must show that for every

°
neighbourhood W of in F there exist neighbourhoods Vi of
such that the relations Zi E V imply
i= 1
°in Ei (1 ~ i ~ n)

Now, we can write

NO.7 TOPOLOGICAL VECTOR SPACES TVS 1.9

where H varies over the 2n - 1 subsets of the set of integers { 1, 2, ... , n }, excluding

Yi = Zi if i 1= H. There exist 2n - 1 balanced neighbourhoods W H of in F such

that L W H C W; on the other hand as f is continuous at (0,0, ... , 0) by hypothesis,
°
the set { 1,2, ... , n} itself, and where uH = f(Y1' YZ' ... , Yn)' with Yi = ai if i E Hand

°
there ~xists in each Ei a neighbourhood U i of (1 ~ i ~ n) such that the n relations
Xi E U i imply that f(x!, ... , xn) En WHo As U i is absorbent, there exists Ai =1= in K
H
°
such that Aiai E Ui' Let A be an element ofK such that IAI ;;: TI IAr 1 for each subset
iEH
H; we show that the neighbourhoods Vi = A-n Ui' fulfill the required condition.
We can write uH = J.lf(x 1, ... , xn) where Xi E U i for 1 ~ i ~ nand J.l = A-np(TI Ai- 1),
iEH
P being the number of integers of { 1,2, ... , n } not in H. From the above 1J.l1 ~ 1,
hence U H E J.lWH C W H since W H is balanced. The proposition is established.

PROPOSITION 6. ~ With the same hypotheses on Ei (1 ~ i ~ n) and on F as in prop. 5, in

n
order that a set Iff of multilinear maps of TI Ei in F be equicontinuous it is sufficient
i= 1
that the set be equicontinuous at (0,0, ... , 0).
For, in the demonstration of prop. 5 the U i (1 ~ i ~ n) can be taken such that
the relation Xi E U i (1 ~ i ~ n) imply f(x 1, ... , xn) En W H for every mapping f E Iff.
H

7. Initial topologies of vector spaces

PROPOSITION 7. ~ Let (EJ'EI be a family of topological vector spaces on a topological
division ring K. Let E be a vector space on K andfor each 1 E I, let J; be a linear map-
ping of E in E, . Then the coarsest topology on E which makes each function J; conti-
nuous, is a topology :Y compatible with the vector space structure of E. Further, if
for every X E E, <I>(x) denotes the point (J;(x)) of the product space F = TI E" then
lEI
the topology :Y is the inverse image of the topology of the subspace <I>(E) of F under
the linear mapping <1>.
The last part of the proposition is a particular case of GT, I, § 4.1, prop. 3. The
proposition then follows from the next lemma.

Lemma. ~ Let M and N be two vector spaces, and g a linear mapping of M in N.

If:Yo is a topology compatible with the vector space structure ofN, then the inverse
image of :Ya by g is compatible with the vector space structure of M.
We show, for example, that (A, x) ~ AX is continuous at each point (A a , x a) of
K x M. Put Ya = g(xo). Every neighbourhood ofO in M contains a neighbourhood
- 1
°
of the form g(U) where U is a neighbourhood of in N; by hypothesis there exists
° °
a neighbourhood V of in K and a neighbourhood W of in N such that the rela-
tions A-A o E V, and y- Yo E W imply AY- AaYo E U. Thus the relations A-A o E V,
-1 -1
x-xo E g(W) imply 'Ax-Aoxo E g(U). We can show similarly that (x, y) ~ X-Y
is continuous in M x M.
TVS 1.10 TOPOLOGICAL VECTOR SPACES §1

F or each index lEI, let ~, be a fundamental system of neighbourhoods of 0 in E, .

From the definition of the topology !T, the filter of neighbourhoods of 0 for this
-1
topology is generated by unions of sets of the families .r.(~J; in other words, the
sets of the form n
-1
k
j; k (V, k ) form a fundamental system of neighbourhoods of 0 for !T,
the (lk) 1 q~n being any finite sequence of indices of I, and, for each index k, V'k
any set of ~'k'

COROLLARY 1. - Let G be a topological vector space on K. In order that a set H of

mappings of G in E be equicontinuous, it is necessary and sufficient that, for alll E I,
the set f. 0 u where u varies in H should be equicontinuous.
This is a particular case of GT, X, § 2.2, prop. 3.

COROLLARY 2. - If the spaces E, are Hausdorff, then in order that !T be Hausdorff,

it is necessary and sufficient that, for every x =F 0 in E, there should exist an index
lEI, such that fJx) =F O.
For <I>(E) is then a Hausdorff space, and in order that !T be Hausdorff, it is evi-
dently necessary and sufficient that <I> be injective; note that we can then identify E
(with 3'"") with the subspace <I>(E) of IT E, by the mapping <1>.
'EI
COROLLARY 3. - Suppose the E, are complete and <I>(E) is closed in F = IT E,. Then
'EI
E is complete in the topology !T.

For the subspace <I>(E) of F is then complete (GT, II, § 3.4, prop. 8 and § 3.5,
prop. 10), therefore the same is true of E in the inverse image topology (GT, I, § 7.6,
prop. 10 and GT, II, § 3.1, prop. 4).
* Example. - Let ~ '(R) be the space of distributions on R; for p a number such that
I ,,; p ,,; + CfJ, let j: U(R) ---> ~ '(R) be the canonical injection, which is continuous
(when U(R) carries its normed space topology and ~'(R) the strong topology). For
every distribution f E ~ '(R), denote the derivative of f by D(f); recall that f f-+ DU)
is a continuous endomorphism of ~'(R). Then let E be the vector subspace of U(R)
formed from those f E U(R) for which DU) E U(R), and confer on E the coarsest
topology making the canonical injection i: E ---> U(R) and the mapping D : E ---> U(R)
continuous (U(R) carries its normed space topology). For this topology, the space E
is complete. For, the image of E in F = U(R) x U(R) by the mapping <P:f f-+ (f, D(f»)
is closed, since it is the trace on U(R) x U(R) of the image G of ~ '(R) in ~ '(R) x ~ '(R)
by the mapping

now G is the graph of <Po, therefore closed in ~'(R) x ~'(R) (GT, I, § 8.1, cor. 2 of
prop. 2), and as <p(E) is the inverse image of G by i x i, which is continuous, we see that
<p(E) is closed in F. *

COROLLARY 4. - Let E be a vector space over a topological division ring K, and let
(!T'\EI be a family of topologies compatible with the vector space structure of E; then
the upper bound !T of the topologies !T, is compatible with the vector space structure
ofE.
No.1 LINEAR VARIETIES TVS 1.11

For, if E, denotes the topological vector space obtained from E by the topology
9";, and ft the identity map ofE on E" then '!T is the coarsest topology making the 1.
continuous.

§ 2. LINEAR VARIETIES IN A TOPOLOGICAL VECTOR SPACE

1. The closure of a linear variety

Recall (A, II, § 9.3) that in a vector space E over a division ring K, a non-empty
affine linear variety (called « linear variety» when this can cause no confusion) is the
image under a translation of a vector subspace of E.

PROPOSITION 1. - In a topological vector space E, the closure of a linear variety is

a linear variety.
Since every translation is a homeomorphism of E, it is sufficient to demonstrate
the proposition for a vector subspace M of E, and in this case, the proposition has
been proved in I, p. 4.

COROLLARY. - In a topological vector space E, every hyperplane is either closed or

everywhere dense.
In fact, the closure of a homogeneous hyperplane H can only be H or the whole
space E, since it is a vector subspace containing H (prop. 1).
A hyperplane H is closed in E if, and only if, CH contains an interior point.
The vector subspace M generated by a set A, in a topological vector space E,
is the set of linear combinations of points of A (A, II, § 1 .7, prop. 9); the closure of
Min E is, by prop. 1, the smallest closed vector subspace containing A; we say that
this is the closed vector subspace generated by A.

DEFINITION 1. - A set A, in a topological vector space E, is total if, and only if, the
closed vector subspace generated by A coincides with E (i.e. the set of linear combi-
nations of elements of A is everywhere dense).
Examples. - 1) In the normed space 96'(1; C) (on the field C) offunctions, continuous
on 1=(0, 1), with values in C, the restrictions to I of the functions x" (n E N) form
a total set, by the Weierstrass-Stone theorem (GT, X, § 4.2, tho 3). Similarly, the res-
trictions to I of the functions e2nnix (n E Z) form a total set (GT, X, § 4.4, prop. 8), in
the subspace P of 96'(1, C) formed of functions such that 1(0) = 1(1).

°
2) Every absorbent set in a topological vector space E over a non-discrete valued
division ring (and in particular every neighbourhood of in E) is a total set since it
generates E (I, p. 7). Thus a linear variety that is not dense in E is necessarily a nowhere
dense set in E (GT, IX, § 5.1) since its closure cannot contain an interior point.

DEFINITION 2. - A family (aJ'EI of points of a topological vector space E is called

topologically independent if for any K E I, the closed vector subspace generated by the
at' with 1 =f= K, does not contain aK •
TVS I.12 TOPOLOGICAL VECTOR SPACES §2

Example. - 3) In the normed space 'I&'(I; C) of continuous functions defined over

1 = (0, 1), the restrictions to I of the functions e2nnix (n E Z) form a topologically inde-
pendent family. If f(x) is the linear combination I Ck e2knix (where all but finitely
many of the Ck are zero) then

( le2nniX - f(x)j2 dx = 1 + I Ick l2 ;, I

Jo k::;f:n

and, a fortiori, by the mean value theorem

sup le2nnix - f(x) I ;, I

XEI

which shows that e2ninx does not belong to the closed vector subspace of 'I&'(l; C) gene-
rated by e2knix, k i= n.

The set of elements of a topologically independent family is called a topologically

independent set of E. Every subset of a topologically independent subset is topolo-
gically independent; every subset consisting of a single point x =1= 0 is topologically
independent if E is a Hausdorff space.
A topologically independent family is independent (in the algebraic sense; cf A,
II, § 7.1, Remark), but the converse is incorrect.
Example. - 4) In the normed space 'I&'(I; C) of functions that are continuous over
I = (0, I), the restriction to I of the functions xn (n E N) form an algebraically inde-
pendent family. But there exists a sequence of polynomials (Pn) such that pnCX2) con-
verges uniformly to x in I (GT, X, § 4.2, lemma 2) which shows that x belongs to the
closed vector subspace of '1&'(1; C) generated by the functions x 2 " (n EN).
Remarks. - 1) The family of topologically independent sets of a topological vector
space is not necessarily inductive for the relation of inclusion (I, p. 25, exerc. 2); this
situation is thus different to that for algebraically independent sets. Moreover there
does not necessarily exist in E a maximal topologically independent subset (I, p. 25,
exerc. 4), thus there does not necessarily exist a subset that is both total and topolo-
gically independent.
2) Let M be a closed vector subspace of E and (a,ltEI a topologically independent
family in the quotient space ElM. If at is any element of the class at' then from def. 2,
and the fact that the canonical mapping of E on ElM is continuous, it follows that
the family (a,),eT is topologically independent. But note that if N is the closed vector
subspace generated by the at it can happen that M n N i= {o} (I, p. 25, exerc. 2), and
hence the sum M + N is not necessarily direct in the algebraic sense (nor a fortiori
in the topological sense).

2. Lines and closed hyperplanes

PROPOSITION 2. - Every Hausdorff topological vector space E of dimension lover a

non-discrete valued division ring K is isomorphic to Ks; in fact, for every a =1= 0 in E,
the mapping ~ ~ ~a of Ks on E is an isomorphism (in other words every linear mapping
of Ks on E is an isomorphism).
As the mapping ~ ~ ~a of Ks on E is bijective and continuous (I, p. 1, def. l), it
is sufficient to show that it is bicontinuous. Let rJ. be a real number > 0, we show
that there exists a neighbourhood V of 0 in E such that if ~a E V then I~I < rJ.. As K
NO.3 LINEAR VARIETIES TVS I.13

is not discrete, there exists an element ~o E K such that 0 < I~ol < ri; but, as E is
Hausdorff, there is a neighbourhood V of 0 such that ~oa does not belong to V.
We can suppose that V is balanced (I, p. 7, prop. 4). But then if ~a E V and I~I ~ I~ol
we have I~o~ -11 ~ 1, and ~oa = (~o~ -1 )(~a) E V; since this last statement is false
we see that ~a E V implies I~I < I~ol < ri. This completes the proof.

COROLLARY 1. - In a Hausdorff topological vector space E, over a non-discrete

valued division ring K, every vector subspace D of dimension 1 is isomorphic to Ks'

COROLLARY 2. - Let E be a topological vector space over a non-discrete valued divi-

sion ring. Every vector subspace D (of dimension 1) which is the algebraic complement of
a closed homogeneous hyperplane H is also the topological complement of H.
In D, the set {O} is closed, being the intersection of D and the closed set H; D is
therefore Hausdorff. But as E/H is also Hausdorff, the canonical mapping of D
on E/H, which is linear, is also an isomorphism by prop. 2, from which the conclusion
follows (GT, III, § 6.2).

THEOREM 1. - Let E be a topological vector space over a non-discrete valued division

ring. Let H be a hyperplane in E defined by the equation f(x) = ri where f is a linear
form not identically zero. Then H is closed in E if and only !f f is continuous.
The condition is evidently sufficient (GT, I, § 2. 1, tho 1); we show that it is necessary.
We can suppose that H is a closed homogeneous hyperplane with the equation
f(x) = O. The quotient space E/H is then a Hausdorff topological vector space of
dimension 1 on K. We can write f = go <1>, where <I> is the canonical mapping of
E on E/H and g is a linear mapping of E/H on Ks; from prop. 2, g is continuous,
thus the same is true of f.

COROLLARY. - Every continuous linear form on E that is not identically zero is a

strict morphism of E on Ks'
Remark. - There are examples of normed topological vector spaces over a complete
non-discrete valued division ring, in which every continuous linear form is identically
zero (I, p. 25, exerc. 4) ; in such a space therefore, every hyperplane is everywhere dense
(I, p. 11 , corollary). .

3. Vector subspaces of finite dimension

THEOREM 2. - Every Hausdorff topological vector space E, of finite dimension n,

over a complete non-discrete valued division ring K, is isomorphic to K~ ; in
n
fact, for every basis (e)Hi~n of E on K, the linear mapping (~) f--+ L: ~iei is an iso-
i= 1
morphism of K~ on E.
Proposition 2 of I, p. 12, implies that tho 2 is true for n = 1; we argue by induction
TVS 1.14 TOPOLOGICAL VECTOR SPACES §2

on n. Let H be the vector subspace of E generated by e 1 , e 2 , .•. , en-I; the induction

n-1
hypothesis is that the mapping (~)1 ~i~n-1 f-4 L ~iei is an isomorphism of K~-1
i= 1
on H. The subspace H, being isomorphic to a product of complete spaces, is complete
(GT, II, § 3.5, prop. 10); hence it is closed in E (GT, n, § 3.4, prop. 8). Let 0 be
the subspace Ken complementary to H in E; E is the topological direct sum ofH and
0(1, p. 13, cor. 2), therefore the mapping

II

(~)1 ~i~n f-4 L1 ~iei

i=

of K~-l x Ks on E is an isomorphism.
When n > I the hypothesis that K is complete is essential for the validity of theorem 2.
In fact, let K be a non-complete valued division ring. and let K be its completion:
for each a i= 0 of K the sct K.a is everywhere dense in K, since x f-+ xa is a homeo-
morphism of K on itself If a ¢: K, the subspace K + Ka of the topological vector space K
on K is of dimension 2 on K, but it is not isomorphic to K; since every subspace of
dimension I in K + Ka is dense in K + Ka.

COROLLARY 1. - i n a Hausdo~tJ topological vector space E over a complete non-

discrete valued division ring K, every vector subspace F of finite dimension is closed in E.
For, if F is of dimension n then it is isomorphic to K~; it is therefore complete
and hence closed in E (GT, II, § 3.4, prop. 8).

COROLLARY 2. - Let K be a complete non-discrete valued division ring, and E be a

Hausdorff topological vector space of finite dimension over K. If F is any topological
vector space over K, then every linear mapping of E in F is continuous.

COROLLARY 3. - i n a Hausdorff topological vector space E, over a complete non-

discrete valued division ring, el'ery finite independent set is topological!.v independent.

COROLLARY 4. - Let E be a topological vector space over a complete non-discrete

valued division ring. If M is a closed vector subspace of E and F is a vector subspace of E
of finite dimension, then the subspace M + F is closed in E.
Write ¢ for the canonical homomorphism of E on the quotient space ElM (neces-
-1
sarily Hausdorff). Then the subspace M + F is identical with ¢(¢(F)). Now ¢(F)
is of finite dimension in ElM, therefore (cor. 1) ¢(F) is closed in ElM, and, in conse-
-1
quence, ¢(¢(F)) is closed in E.

z We note that, if M and N are any two closed vector subspaces in a Hausdorff topo-
logical vector space E. then M + N is not necessarily closed in E. * even if E is a Hilbert
space * (cf. I V, p. 64. exerc. 13, d)).

PROPOSITION 3. - Let E be a topological vector space over a complete non-discrete

No. 4 LINEAR VARIETIES TVS U5

valued division ring K. Let M be a closed veclor subspace of finite codimension n in E.

T hen every subspace N that is an algebraic complement of M in E is also a topological
complement.
In N, the set {O} is closed, since it is the intersection of N and the set M which
is closed in E; thus N is Hausdorff. As ElM is also Hausdorff, the canonical mapping
of N on ElM, which is linear and bijective, is bicontinuous (I, p. 14, cor. 2), from
which the proposition follows.

COROLLARY. - Let E and F be two topological vector spaces over a complete 110n-
discrete valued division ring. If F is Hausdorff and of finite dimension, then every
continuous linear mapping of E on F is a strict morphism.

Remark. The results of Nos 2, 3 are no longer valid when K is discrete. For example,
let K 1 be a non-discrete valued division ring and K be the discrete division ring obtained
by endowing K1 with the improper absolute value on K 1 . Then K1 is a topological
vector space of dimension lover K, but it is not isomorphic to Ks' However, we can
show that the results of Nos 2, 3 are valid even when K is discrete, provided that we
impose on the topological vector spaces considered. the property of having a funda-
mental system of balanced neighbourhoods of 0 (i.e. neighbourhoods V such that
K. V = V) (I, p. 27, exerc. 14); this condition (which is always satisfied when K is a non-
discrete valued division ring (1 I, p. 7, prop. 4) is not valid for all topological vector
spaces over K as the preceding example shows.

4. Locally compact topological vector spaces

THEOREM 3. Let K be a complete non-discrete valued division ring. If E is a Haus-

dorff topological vector space over K, which is such that some neighbourhood V of 0
in E is precompact, then E is of finite dimension. If E =1= { 0 }, then both K and E are
locally compact.
In proving the first assertion, we need consider only the case when E is complete;
for E is an everywhere dense subspace of its completion E, and the closure V of V
in E is compact and is a neighbourhood of 0 in E (GT, Ill, § 3.4, prop. 7).
We can suppose then that there is a compact neighbourhood V of 0 in E. Let
0: E K be such that 0 < 10:1 < 1; then there are finitely many points ai E V such that
Vc U (a i + aV) .
i

Let M be the finite dimensional subspace ofE generated by the a i ; it is closed in E

(T, p. 14, cor. 1). In the Hausdorff topological vector space EjM the canonical image
of V is a compact neighbourhood W of 0, such that W c aW; hence 0:- 1 W c W,
and, by induction on n, ":1-" W c W for every positive integer n. As W is absorbent,
we conclude that W = EjM; and thus ElM is compact. To complete the proof of
the first assertion in the theorem, it is sufficient, therefore, to establish the following
lemma.

Lemma 1. - Any compact topological vector space E over a non-discrete valued divi-
sion ring, is just the set { 0 }.
TVS I.16 TOPOLOGICAL VECTOR SPACES §3

Since E is complete we can suppose K is complete (I, p. 6). If E #- {O} then E

contains a line that is closed in E (I, p. 14, cor. 1) and therefore compact. This line is
isomorphic to Ks (I, p. 12, prop. 2) and hence K must be compact. Now the mapping
~ f---> I~I of K in R is continuous and thus the image of K must be bounded, on the
other hand there exists y E K with Iyl > 1, and the set ly"1 = Iyl", n E N, is unbounded.
This contradiction shows that E = {O}.
To prove the second assertion in the theorem, if E #- {O} then from the first part
of the theorem E is isomorphic to K: with n > 0; now K is complete, hence so is E,
and thus E is locally compact. But Ks is isomorphic to a line in E (I, p. 12, prop. 2)
which is necessarily closed in E (I, p. 14, cor. 1) ; it follows that K is locally compact.

Remark. - The result ofth. 3 is no longer true if K is a discrete division ring as is shown
by the example ofR (with the usual topology) considered as a topological vector space
over the discrete field Q.

§ 3. METRISABLE TOPOLOGICAL VECTOR SPACES

1. Neighbourhoods of 0 in a metrisable topological vector space

We say that a topological vector space E is metrisable if its topology is metrisable.

Relative to the structure of its additive group and of its topology, E is, therefore,
a metrisable group (GT, IX, § 3.1).
We know that, for a topological group to be metrisable, it is necessary and suffi-
cient that there exists an enumerable fundamental system of neighbourhoods of
the neutral element e, whose intersection is the single element e (GT, IX, § 3.1,
prop. 1).
Also we know that the uniform structure of a metrisable topological vector
space E, can be defined by an invariant distance d(x, y) = Ix - yl, where x f---> Ixl
is a continuous mapping of E in R+ which satisfies the conditions: 1) 1- xl = Ixl;
2) Ix + yl ~ Ixl + Iyl; 3) the relation Ixl = 0 is equivalent to x = 0 (GT, IX,
§ 3.1, prop. 3).

We saw (GT, IX,§ 3 . 1, prop. 2) how such a distance d could be defined using a decreas-
ing sequence (W /I) of neighbourhoods of 0 in E, forming a fundamental system of neigh-
bourhoods and such that W n + 1 + W n + 1 + W n + 1 C W n' When E is a metrisable
vector space over a non-discrete valued division ring K, we can also suppose that the Wn
are balanced (I, p. 7, prop. 4); if we revert to the process of definition of d (lac. cit.) we can
see that the relation IAI ~ I implies that Ihl ~ Ixi. Further the conditions (EVTi) and
(EVTi[) of I, p. 2 imply both that IAxol tends to 0 as A tends to 0 in K for every Xo E E,
and that IAoxl tends to 0 as Ixl tends to 0 for every Ao E K. Conversely, if the function Ixl
possesses all the preceding properties and if W II is the set of x E E such that Ixl ~ 2- n ,
then the Wn form a fundamental system of balanced neighbourhoods of 0 for a metrisable
topology on E that is compatible with the vector space structure ofE.

Remark. - One of the most important classes of metrisable vector spaces are the
No.3 METRISABLE TOPOLOGICAL VECTOR SPACE TVS 1.17

normed spaces (1, p. 3). But it must be noted that there exist metrisable vector spaces
whose topology cannot be defined by a norm (I, § 3, exerc. 1); we shall study important
examples later.

2. Properties of metrisable vector spaces

Every vector subspace of a metrisable topological vector space E is metrisable;

the same is true of every quotient space ElM of E by a closed vector subspace M
(GT, IX, § 3.1, prop. 4). Every product of an enumerable family of metrisable topo-
logical vector spaces is metrisable (GT, IX, § 2.4, cor. 2). If K o is a complete valued
division ring, and K is a subdivision ring everywhere dense in Ko, the completion E of
a metrisable vector space E over K is a metrisable vector space over K o (1, p. 6 and
GT, IX, § 2, No.1, prop. 1). Finally, if E is a metrisable vector space that is complete,
then for every closed vector subspace M of E, the quotient space ElM is complete (GT,
IX, § 3. 1, prop. 4).

3. Continuous linear functions in a metrisable vector space

THEOREM 1 (Banach). - Let E and F be two metrisable vector spaces over a non-
discrete valued division ring K, and let u be a continuous linear mapping of E in F.
Suppose that E is complete. Then the following conditions are equivalent:
(i) u is a strict surjective morphism.
(ii) F is complete and u is surjective.
(iii) The image of u is not meagre in F (GT, IX, § 5.2).
(iv) For every neighbourhood V of 0 in E, the set u(V) is a neighbourhood of 0 in F.
Firstly (i) implies (ii), for let u be a strict surjective morphism and N be the kernel
of u. Then u induces an isomorphism of E/N on F. But E is metrisable and complete,
hence E/N is complete (GT, IX, § 3.1, prop. 4), therefore F is complete.
Next (ii) implies (iii). Let F be complete and u be surjective. The image of u is
precisely F and therefore not meagre in F from Baire's theorem (GT, IX, § 5.3).
The following lemma shows that (iii) implies (iv).

Lemma 1. - Let E and F be two topological vector spaces over a non-discrete valued
division ring K, and let u be a continuous linear mapping of E in F such that the image
of E is not meagre. Then, for every neighbourhood V of 0 in E, the set u(V) is a neigh-
bourhood of 0 in F.
Let W be a balanced neighbourhood of 0 in E such that W + W c V (I, § 1.5,
prop. 4). Let r:J. be an element of K such that 1r:J.1 > l; then E is the union of the sets
r:J.nw where n varies in N; in fact, for all x E E, there exists f3 E K such that x E f3W
(I, p. 7, prop. 4) and there exists an integer n ~ 0 such that 1131 < 1r:J.l n, then x E r:J.nW
since W is balanced. Hence, u(E) is the union ofthe sequence of sets u( r:J.nw) = r:J.nu(W),
and as u(E) is not meagre in F, one at least of the sets r:J.nu(W) possesses an interior
point (GT, IX, § 5 .3, def. 2) and therefore u(W) has an interior point.
TVS 1.18 TOPOLOGICAL VECTOR SPACES §3

Let Yo be an interior point of u(W); since - u(W) = u(W), and therefore

- u(W) = u(W) itfollows that 0 = Yo + (- Yo) is an interior point of u(W) + u(W).
As vector addition is a continuous mapping of F x F in F, the set u(W) + u(W)
is contained in the closure of the set

u(W) + u(W) = u(W + W) c u(V);

hence u(V) is a neighbourhood of 0 in F.

Before proving that (iv) implies (i) we prove the following lemma, where we make
the convention that, in all metric spaces, Br(x) denotes the closed ball of centre x
and radius r.

Lemma 2. - Let E and F be two metric spaces, and suppose that E is also complete.
Let u be a linear mapping ofE in F having the following property: whatever the number
r > 0, there exists a number per) > 0 such that, for all x E E, we have

Bp(r)(u(X)) c u(B.(x)).

In these conditions, for all a > r, the image u(Ba(x)) contains the ball Bp(riu(x)).

I
00

Let (rn) be an infinite sequence of numbers > 0 such that r 1 = r and a = rn'
n= 1
For each index n there exists a number Pn > 0 (with P1 = per)) such that

for all x E E; we can, and will, suppose that lim Pn = O.

n-HD

Let Xo be a point ofE, and y be a point of Bp(r)(u(xO))' We shall show that y belongs
to u(Ba(xo)).
For this, a sequence (Xn)n> 0 of points of E is defined inductively such that, for
all n ~ 1, we have xnEBrn(xn-1) and u(xn)EB pn + 1 (y). If the Xi have been defined
for 0 ~ i ~ n - 1 satisfying these relations, then we have YEB pju(xn_ 1)); since

there exists a point xn E B rn (x n- 1) whose image u(xn) belongs to the neighbourhood

BPn+ 1 (y) of y, which establishes the existence of the sequence (xn)'
Since the distance of xn from x n+ p is less than rn+ 1 + rn+ 2 + ... + rn+ p ' which
is arbitrarily small when n is large, the sequence (xn) is a Cauchy sequence in E.
As E is complete, the sequence (xn) converges to a point x of E. The distance of
I
00

Xo from x is less than rn = a, thus x E Ba(xO)' But u is continuous, thus the

n= 1
sequence u(Xn) converges to u(x); also u(xn) E B pn + 1 (y), hence y = u(X) , and the
lemma is proved.
No.3 METRISABLE TOPOLOGICAL VECTOR SPACE TVS I.19

We return to the theorem and show that (iv) implies (i). Suppose that u satisfies
condition (iv). For each of the spaces E and F, consider a distance that is invarient
under translation and defines its topology (I, p. 16). By hypothesis, the set u(Br(O))
is a neighbourhood of 0 in F for every r > 0, and thus there exists a number per) > 0
such that Bp(r)(O) c u(Br(O)). By translation we conclude that Bp(r)(u(x)) c u(B/x))
for all r > 0 and all x E E. From lemma 2, for every pair of real positive numbers
(a, r), a > r > 0, we have Bp(r)(O) c u(Ba(O)); thus u is a strict morphism of E on F.
We have shown that (iv) implies (i) and the proof of the theorem is completed.

COROLLARY 1. - If E and F are two complete metrisable vector spaces over a non-
discrete valued division ring, then every bijective continuous linear mapping of E on F
is an isomorphism.

In particular, if E and F are two complete normed spaces, there exists a number
a > 0 such that II u(x) II ~ a.llxll for all x E E .

COROLLARY
.'

2. - Let E be a vector space over a non-discrete valued division ring,

let ~1 and ~2 be two topologies on E compatible with its vector space structure and for
each of which E is metrisable and complete. Then, if ~1 and ~2 are comparable, they
are identical.

COROLLARY 3. - Let E and F be two complete metrisable vector spaces over a non-
discrete valued division ring. In order that a continuous linear mapping u of E in F
should be a strict morphism, it is necessary and sufficient that u(E) be closed in F.
The condition is necessary, because if u is a strict morphism, the image u(E),
being isomorphic to the quotient Eju- 1(0), is complete (I, p. 17) and therefore
closed in F. The condition is sufficient, since, if u(E) is closed in F, then u(E) must
be a complete metrisable vector space and thus by theorem I u is a strict morphism
of E on u(E).

COROLLARY 4. - Let E be a complete metrisable vector space over a non-discrete

valued division ring. If M and N are two closed vector subspaces, that are (algebraic)
complements in E, then E is the direct topological sum of M and N.
For M x N is a complete metrisable vector space and the mapping (y, z) 1-+ Y +z
of M x N on E is continuous and bijective, therefore an isomorphism (cor. 1).

COROLLARY 5 (The closed graph theorem). - Let E andF be two complete metrisable
vector spaces over a non-discrete valued division ring. In order that a linear mapping
u of E in F be continuous, it is necessary and sufficient that its graph, in the product space
E x F, be closed.
The condition is necessary since the graph of a continuous mapping into a Haus-
dorff space is closed (GT, I, § 8.1, cor. 2). To see that it is sufficient, note that it
implies that the graph G of u, which is a closed vector subspace of the complete
metrisable space E x F, is itself metrisable and complete. The projection z 1-+ pr 1 (z)
of G on E is a bijective, continuous linear mapping, therefore an isomorphism
TVS 1.20 TOPOLOGICAL VECTOR SPACES §3

(cor. 1); since its inverse mapping is x ~ (x, u(x)), it follows that u is continuous
in E.

°
We can express this corollary in the following form: u is continuous if the following
situation holds: if the sequence (xn) of points of E both converges to and is such that
the sequence (u(xn)) converges to y, then it is necessarily the case that y = 0.

Example. - Let E be a vector subspace of the space of real-valued functions defined

on I = (0, I) ; let II I I be a norm on E, under which E is complete, and such that its
topology is finer than the topology of simple convergence. Suppose further that E
contains the set (fj 00 (I) of functions infinitely differentiable on I; we shall show that
there exists an integer k ?o 0, such that E contains the set (fjk(I) of all functions with
a continuous k-th derivative in I.
For every pair of integers m > 0, n ?o 0, let Vmn be the set of functions IE (fj"'(I)
°°
such that I f'hl(X) I ~ 11m for ~ h ~ n and for all x E I. The Vm.n form a fundamental
system of neighbourhoods of for a metrisable topology compatible with the vector
space structure of (fj "'(I), further (fj "'(I) is complete in this topology (FVR, II, p. 2,

°
tho I). Let u be the canonical mapping of (fj 00 (I) in E; we show that u is continuous.
From cor. 5 above it is sufficient to prove that if a sequence (J,,) converges to in (fj "'(I)

I is the simple convergence limit of (D. Hence there exists an integer k ?o and a
number a > such that the relation
°
and to a limit I in E then necessarily I = 0. But this is immediate since, by hypothesis,

°
h(f) = sup
xcI
Ir)(x)1 ~ a
O~h~k

implies II I II ~ I for all IE (fj "'(I).

But h is a norm on the space (fjk(I) and (fj "'(I) is a subspace that is everywhere dense
in (fjk(l) for this norm (the set of polynomials being already everywhere dense in (fjk(I),
an immediate consequence of the Weierstrass-Stone theorem). By what has gone before,
the identity mapping of (fj 00(1) (carrying the norm Pk) in E, is continuous, and so it
can be extended continuously to the whole space (fjk(I) (since E is complete). This
proves our assertion.

PROPOSITION 1. - Let E, F be two topological vector spaces over a non-discrete valued

division ring K. We suppose that:
1) E is metrisable and complete.
2) There exists a sequence (Fn) of complete metrisable vector spaces over K and,
for each n, an injective continuous linear mapping vn ofFn in F such that F is the union
of the subspaces vn(Fn)'
Then let u be a linear mapping of E in F. If the graph of u is closed in E x F, then
there exists an integer n and a continuous linear mapping un ofE in Fn such that U= vn 0 un
(which implies that u is continuous and that u(E) C vn(Fn))'
Let G be the graph of u in E x F. For all n, we consider the continuous linear
mapping wn :(x, y) ~ (x, vn(y)) ofE x Fn in E x F; as G is closed, the set w; leG) = G n
is a closed vector subspace of E x Fn ; if Pn is the restriction to G n of the first projec-
tion prp we have Pn(Gn) = u-l(vn(F n)). As Pn is continuous and G n is complete
(since G n is closed in the complete space E x FJ, Pn(Gn) is, by theorem 1, either
meagre in E or it is the whole of E. But, by hypothesis, E is the union of the Pn(Gn),
No.3 METRISABLE TOPOLbGICAL VECTOR SPACE TVS 1.21

and as E is complete, the Pn(G n) cannot all be meagre in E by Baire's theorem (GT,
IX, § 5.3, tho 1). Therefore there exists an integer n such that piG n ) = E, or in other
words u(E) C vn(Fn)' Further, as vn is injective, G n is the graph of a linear mapping
un of E in Fn , and by the closed graph theorem (I, p. 19, cor. 5) un is continuous; it
follows then from the definitions that u = Vn 0 un'
 Exercises

§ 1

1) Let Eo = Q~ be the vector space over the p-adic field Qp (GT, III, § 6, exerc. 23) which is
the product ofan enumerable infinity of factor each identical with Qp. Let P c Eo be the
set Z~, and let E be the vector subspace of Eo generated by P. On the additive group P we
consider the product compact topology of the topologies of the factors Zp, and we denote
by 5!l the filter of neighbourhoods of 0 in P for this topology. Show that 5!l is a fundamental
system of neighbourhoods of 0 in E for a topology ,0; compatible with the additive group
structure of E, that satisfies (EVT;) and (EVT;n) but not (EVT;I) (prove that the homothety
x ~ x/p is not continuous in E).

2) Let K be a non-discrete topological division ring, Ko the division ring K with the discrete
topology. The discrete topology on Ko is compatible with its additive group structure, and
when we consider Ko as a vector space over K, it satisfies the axioms (EVT;I) and (EVT;n)
but not (EVT;).

3) For every real number a > 0, let G, be the topological group R/'1Z, and let G be the
topological product group TI
G, (a varying in the set of number> 0). For every x EO R,
let t,(x) be the canonical image of x in Go; the mapping 4>: x ---> (t,(x)) is a continuous injective
homomorphism of R in G. We consider on R the topology that is the inverse image by 4>
of that of G, and denote by E the topological group formed by R with this topology. Show
that when E is considered as a vector space over R its topology satisfies (EVT;) and (EVT;I)
but not (EVT;n)'

4) Let E be a vector space over a division ring K with a valuation; we suppose that E carries
a metrisable topology compatible with its additive group structure. Suppose further that
this topology satisfies axioms (EVT;) and (EVT;I); show that if one of the two metrisable
groups K, E is complete then the topology also satisfies (EVT;n) and is, in consequence, compa-
tible with the vector space structure of E (cf. GT, IX, § 5, exerc. 23).

5) Let K be a non-discrete valued field, and S be an arbitrary infinite set.

§1 EXERCISES TVS 1.23

a) Let D = (an> be an enumerable infinity of elements of S. For every A E K such that IAI ~ 1,
let U i. be the element of the normed space ~(S), (1, p. 4, Example) of bounded mappings of
S in K, such that uiJan ) = An for all n E Nand uA(b) = 0 for bet D. Show that the family (u i)
is algebraically independent.
b) Deduce that every basis of the vector space !i6'K(S) has the same cardinality as K S (using a)
show that the cardinal of every basis of !i6'K(S) is at least equal to Card (K); note on the other
hand that Card(!i6'K(S») = Card(K S) and use A, II, § 2, exerc. 22).
e) Show in the same way that every basis of the vector spacefi(S) has the cardinality of(K x S)N.

6) Let K be a non-discrete valued division ring. Show that, for the space £L(N) of absolutely
en
summable sequences x = (~n) of elements of K, the norms Ilxll l = I I~nl and Ilxll = sup I~nl
n= 0 n
are not equivalent (ef GT, IX, § 3.3, prop. 7); show that eL(N) with the norm Ilxll is never com-
plete even if K is complete; what is its closure in ~(N) ?

-r 7) * Let A be a ring with a discrete valuation, v the normed valuation of the division ring
of fractions K of A; take the absolute value aVon K, where 0 < a < L Let E be a normed
vector space over K, for which the norm satisfies the ultrametric inequality

Ilx + yll ~ sup(llxll, Ilyll)·

a) Denote by M the set of those x E E for which Ilxll ~ 1, and by n a uniformizer of A; M

is an A-module, and M/nM a vector space on the residual division ring k = A/nA of A.
Let (eJAEL be a family of elements of M such that the images of eAin M/nM form a basis of
this vector k-space. Show that (eA) is an independent family in E and that the vector subspace F
of E generated by (eA) is dense in E.
b) If we put Ilxll! = sup I~AI, for every x = I ~AeA in F, show that on F the norms Ilxll and
A A
Ilxll! are equivalent. A

c) Let K be complete. Deduce from a) and b) that, if L is finite, the completion E of E is iso-
morphic to KL; if L is infinite E is isomorphic to the subspace '6'~(L) of !i6'K(L) formed of the
families (~A) such that lim ~A = 0 for the filter of complements of finite subsets of L.
d) We suppose K and E complete; let G be a second normed complete space over K whose
norm satisfies the ultrametric inequality. Show that on replacing (if necessary) the norm of
.P(E; G) (GT, X, § 3.2) by an equivalent one, then .P(E; G) is isometric to the vector space
of families (h)i.E~ of elements of G such that sup II YA II < + CfJ, carrying the norm sup II h II
AEL AEL
(which is also an ultrametric norm).

8) Let E be a topological vector space over a non-discrete topological division ring K. In

order that there should exist a neighbourhood of the point (0, 0) in K x E such that the
mapping (A, x) f-+ AX should be uniformly continuous in this neighbourhood, it is necessary
and sufficient that there exist a neighbourhood V 0 of 0 in E such that the sets AV 0 form a
fundamental system of neighbourhoods of 0 in E, where A varies in the set of elements =1= 0
of K. When K is a division ring with a non-discrete valuation and E is Hausdorff, show that
the uniform structure of E is then metrisable.

9) Generalize prop. 5 of I, p. 8, to the case where the spaces Ei ( ~ i ~ n) and F are topo-
logical vector spaces over an arbitrary non-discrete topological field.

10) Let E be a complete Hausdorff topological vector space over a non-discrete valued divi-
sion ring K. Denote by F a vector subspace of E, and by 'Y the topology on F induced by the
topology 'Y' of E; let mbe a fundamental system of closed, balanced neighbourhoods of 0
for the topology 'Y. Let F 0 be the vecto.!... subspace of E, generated by the closures V in E (rela-
tive to 'Y ') of the sets V E 'B; the sets V form a fundamental system of neighbourhoods of 0
for a topology 'Yo on F 0' compatible with the vector space structure of F 0; for this topology,
F 0 is complete, and the topology induced by 'Yo on'F is identical with 'Y.
TVS 1.24 TOPOLOGICAL VECTOR SPACES §1

11) In a topological vector space E over a non-discrete topological division ring K there
exists a fundamental system 1.13 of closed neighbourhoods of 0, satisfying the conditions (EVIl)
and (EV m) as well as the two following:
(EV la ) For each V E 1.13, there exists WE 1.13 and a neighbourhood U of 0 in K such that UW c V.
(EV lb ) For every x E E and every V E 1.13, there exists A =1= 0 in K such that AX E V.
Conversely, let E be a vector space over K and let 1.13 be a filter base on E satisfying the condi-
tions (EV la ), (EV lb ), (EVIl) and (EV m ). Show that there is a topology on E (and one such only),
that is compatible with the vector space structure of E, and for which 1.13 is a fundamental
system of neighbourhoods of O.

12) Let K be a discrete field, E the division ring of fractions of the ring of formal series
A = K[[X, YlJ in two indeterminate variables on K(A, IV, p. 36). For every n ;;:: 0, let Vn c A
be the set of formal series of total degree at least equal to n. Show that in E, the sets Vn form
a fundamental system of neighbourhoods of 0, for a topology compatible with the vector
space structure of E (over K), for which E is metrisable and complete; if further K is a finite
field, then E is locally compact. Show that the K-bilinear mapping (u, v) f-> uv of E x E in
E is continuous at the point (0, 0) but that there exists Uo E E such that v f-> uov is not conti-
nuous in E (for example U o = l/X).

13) Let E be a vector space of infinite dimension over R, and let:! be the family of all absorbent
and balanced sets of E. Show that:! does not satisfy axiom (EV m) (in other words is not a
fundamental system of neighbourhoods of 0 for a topology compatible with the additive
group structure of E). For this, consider an infinite independent family (en)n" 1 in E; for every
n
integer n ;;:: 1, let An be the set of points L tie i such that Itil :(; lin for 1 :(; i :(; n; let A
i= 1
be the union of the An' and V be a subspace complementary to the subspace of E generated
by the en' and write C for the set A + V; show that there exists no set ME:! such that
M + M c C.

.,-r14) Let K be a Hausdorff topological division ring, (EJ'EI an infinite family of Hausdorff
topological vector spaces on K, none of which is the single point O. We consider on F = TI E,
'EI
the topology :T, compatible with the additive group structure of F, for which a fundamental
system of neighbourhoods of 0 is formed by the products TI V" where, for each tEl, the
'EI
set V, is a neighbourhood of 0 in E (this topology is strictly finer than the product topology;
cf GT, III, § 2, exerc. 23). We denote by :To the topology induced by :T on the subspace
E = EEl E, of F; E is closed in F for the topology :T, and if each of the E, is complete, then
'EI
F is complete for the topology :T, therefore E is complete for the topology :To (GT, III, § 3,
exerc. 10).
a) Show that if there exists in K a neighbourhood of 0 bounded on the right (GT, III, § 6,
exerc. 12) (in particular if K is a division ring with a valuation), the topology :To is compatible
with the vector space structure of E. If, further, K is not discrete, then E is not a Baire space
for any topology that is finer than :To and compatible with the vector space structure of E.
b) Moreover, if there does not exist in K any neighbourhood of 0 bounded on the right (see c))
give an example of a family (EJ such that the topology :To is not compatible with the vector
space structure of E.
c) Let A = R[X] be the ring of polynomials in one variable on R. For every sequence s = (Ent" 0
of real numbers > 0, denote by Vs the set of polynomials L akX k E A such that lakl < Ek
k
for all k. Let:! be the set of the Vs where s varies in the set of sequences of numbers> O.
Show that:! is a fundamental system of symmetric neighbourhoods of 0 for a topology compa-
tible with the ring structure of A. Let K = R(X) be the division ring offractions of A; denote
by 6 the family of subsets of K of the form UO + U) ~ 1, where U varies in the set of the Vs
not containing 1 ; show that 6 is a fundamental system of neighbourhoods of 0 for a topology
compatible with the division ring structure of K, and that there does not exist in K any neigh-
bourhood of 0 that is bounded.
§2 EXERCISES TVS 1.25

d) For every Hausdorff topological division ring K, show that there exists a set I such that
on F = KI, the topology :7, defined above, is not compatible with the vector space structure
ofF.

§ 2

1) Let S be an arbitrary infinite set.

a) Show that the smallest cardinal of any total set in the normed space ~(S) of bounded
mappings of S in R (I, p. 4, Example) is equal to 2Ca ,d(S) (consider the set of characteristic func-
tions of subsets of S and note that there exists an enumerable set everywhere dense in R).
b) Show that the smallest cardinal of any total set in the normed space PI(S) (1, p. 4, Example)
is equal to Card(S).

2) In the product topological vector space E = RN over the field R, denote bye" (n E N)
the elements of the canonical basis of the direct sum R(N). Write a o = eo, an = eo + (lln)e"
for n ;, 1. Show that, for every integer n ;, 0, the a i such that 0 ,;; i ,;; n form a topologically
independent family in E, but that the infinite family (an)n;,o is not topologically independent.
If M is the closed vector subspace Ra o , the classes an of the an in ElM form a topologically
independent family (for n ;, 1), but the closed vector subspace N generated by the an' with
index n ;, I, in E contains M.

3) Let E be a topological vector space over R, and f a homomorphism of the additive group
of E in R. Show that if there exists a neighbourhood of 0 in E in which f is bounded, then
f is a continuous linear form in E. This is so in particular when f is semi-continuous (lower
or upper).

4) Denote by K the field R with the absolute value p(~) = 1~ll!2. Let E be the vector space

f
over K of the real valued regulated functions defined over I = (0, I), continuous on the

right everywhere and zero at the point I; show that onE the mapping x f-> IIxll = Ix(t)ll/zdt

is a norm. Show that for every function x ;, 0 in E, there exists in E two functions 0,
fi
Xl ;,

x z ;' osuch that x = ~(Xl + xz)and IIxIIi = IIxzlI = IIxli. Deduce that every continuous

linear form on E is identically zero.

5) Let K be a Hausdorff topological division ring of which the topology is locally retrobounded
(GT, III, § 6, exerc. 22). Extend prop. 2 of I, p. 12 and tho I of I, p. 13 to topological vector
spaces over K; similarly extend tho 2 of I, p. 13 and prop. 3 of I, p. 14 when K is also complete.

6) Let K be the topological division ring obtained by transferring the usual topology of QZ
to the field Q(fi) by the mapping (x, y) f-> X + y fi.
a) Let E be the set Q(fi) with its vector space structure over K and with the topology induced
by that of R. Show that E is a Hausdorff topological vector space, of dimension I on K, but
that it is not isomorphic to Ks.
b) Let F be the topological vector space E x E over K; in F, the hyperplane E x {O} is
closed but there is no continuous linear form f on E x E such that this hyperplane is given
by the equation f(x) = O.

7) Let K be a valued division ring which is non-discrete and non-complete, let E be the
topological vector subspace K + Ka of K where a rt K, and let F be the product space K x E.
In F, the subspace M = K x {O} is closed and of codimension 2. Let N be the complementary
subspace to M in F generated by the vectors (0, 1) and (I, a); show that F is not the direct
topological sum of M and N.
TVS 1.26 TOPOLOGICAL VECTOR SPACES §2

~ 8) Let P be a prime number, Qp the field of p-adic numbers (GT, III, § 6, exerc. 23). Let
Eo be the topological product space Qp x R; if K denotes the field Q with the discrete topo-
logy, then Eo is a topological vector space over K. Let M be the vector subspace formed by
the elements (r, r) where r varies in Q; further let 8 be an irrational number and N the vector
subspace formed by the elements (0, r8) where r varies in Q. Let E be the subspace M + N
of Eo; show that N is a closed hyperplane in E, but that there does not exist a complementary
topological subspace to N (note that M is everywhere dense in Eo).

9) Let X be a Hausdorff topological space, and let V be a vector subspace of finite dimension
n of the space 'e(X; R).
a) Show that there exist n pair-wise disjoint, open sets Vi (1 ,s; i ,s; n) in X, such that any
function f E V which is identically zero in each of the Vi' is identically zero in X (use A, II,
§ 7.5, cor. 3).
b) Let Xi E Vi for I ,s; i ,s; n. Deduce from a) that there exists a constant c > 0 such that,
for every function fE V, we have
II

sup !f(x)! ,s; c

XEX
I
i= 1
!f(x)!.

10) Let K be a locally compact non-discrete valued division ring, and E a left vector space of
finite dimension over K. Denote by m(E) the set of norms on E, which is a subspace of the
space (€(E; R) of mappings ofE, continuous (in the canonical topology), in R.
a) When we give to 'e(E; R) the compact convergence topology * (for which it is a Frechet
space) *' the set m(E) is closed in 'e(E; R), and locally compact.
b) Let Po be an element of m(E); show that there exists a continuous mapping (A, p) I--> TC,(p)
of (0, 1) x m(E) in m(E) such that TCo(P) = P and TC 1 (P) = Po for every p E m(E).

11) With the hypotheses of!, p. 23, exerc. 7 show that if K and E are complete then every closed
subspace of E has a topological complement (proceed as in a), lac. cit.).

~ 12) Let K be a locally compact valued division ring whose absolute value is non-discrete
and ultrametric. We call a norm on the left vector space E over K an ultranorm ifit satisfies the
ultrametric inequality (II, p. 2).
a) Let E be a finite dimensional left vector space over K, let a be an ultranorm on E and H
< >
a hyperplane in E given by the equation x, a* = O. Show that there exists a point Xo E E
!< >
at which the function X I--> x, a* !/a(x) attains its upper bound in E" {O} ; show that then

<x, a*> ) !<x, a*>! )

a(x) = sup ( a ( x - <xo,a*>x o '!<xo,a*>!a(x o) .

Deduce that there exists a basis (a) of E and a family (r) of real numbers > 0 such that, for
all x = I
Siai we have a(x) = sup(riISil). We say that a is in the standard form relative to
i ,

the basis (aJ

b) Let a* be the norm on E* the dual of E canonically associated with a by

a*(x*) = sup
x*o
!< x, x* >!jCJ.(x) ;
it is an ultranorm. Show that for all Xo =I 0 in E, there exists xi; E E* such that
a(x o) = !<Xo , xi;) !/a*(xi;).
c) Let a, ~ be any two ultranorms on E. Show that there exists a basis of E such that relative
to this basis a and ~ are both of the standard form (consider a point Xo E E"" {O} at which
a/~ attains its maximum; then use b) and proceed by induction on dim E).
d) Let mo(E), the set of ultranorms on E, be considered as a subspace of m(E) (exerc. 10).
Show that mo(E) is closed in m(E). Let a o be an element of mo(E); for each a E mo(E) and
for 0 ,s; t ,s; 1, let p"(t) be the set of ~ E mo(E) such that ~(x) ,s; a o(x)I-'a(x)' for all x E E.
Show that p"(t) is not empty and that TC~ = sup p"(t) is an ultranorm. Further, the mapping
§2 EXERCISES TVS I.27

(t, ex) ~ TI~ of (0, 1) x '!l1 o(E) in '!l1 o(E) is continuous and such that TI~ = ex o and TI~ = ex
(use c».
* e) Let A be the ring of the absolute value of K, m its maximal ideal such that k = A/m
is a finite field with q elements (CA, VI, § 5, No.1, prop. 2). For every ultranorm ex on E, the
image Xo of the set of values of log ex(x) for x E E ,,{ o} under the canonical mapping in the
quotient group R/(Z.log q) is a finite set having at most n = dim E elements (use a»; the
number of these elements is denoted by r(ex) and called the rank of ex. Show that r is a lower
semi-continuous mapping of '!l1 o(E) in N and that the set '!l1~(E) of the ex for which r(ex) = n
is open and everywhere dense in '!l1 o(E) (use a) and c».
f) Suppose that r(ex) = n; let (aJ be a basis of E relative to which ex is of the standard form;
show that there exists a neighbourhood V of ex in '!l1~(E) such that every ~ E V has the standard
form relative to (aJ (use b»; deduce that there is a neighbourhood W c V of ex homeomorphic
to an open set in Rn.
g) For every basis (a) ofE, show that the set of ultra norms ex that have a standard form relative
to (a) is closed in '!l1 o(E). Deduce that if ex E '!l1~(E) has a standard form relative to (a) the
same is true of all elements of the connected component containing ex in '!l1~(E). *

~ 13) * We keep the general hypotheses and the notations of exerc. 12.
a) Let L be a free sub-A-module of E of dimension n = dim E. For all x E E ",",{O}, the set
of the a E A such that ax E L is a fractional ideal of K of the form m h (h a positive or negative
integer); putting ex(x) = qh and ex(O) = 0, show that ex is an ultranorm on E. It is said to be
associated with the free A-module L.
b) Conversely, if ex is an ultranorm on E, the set Lo of the x E E such that ex(x) « 1 is a free
A-module of dimension n. If [ex] is the norm associated with L o' we have ex « [ex] « qex, and
[ex] is the lower bound of the norms associated with free A-modules and which are;;:, ex. We
have [qex] = q.[exl, and ex(x) = inf(q-'[q'ex] (x») for all x E E, where t varies in the interval

°
(0. 1). Further, the function t ~ [q'ex] (x) is left continuous in this interval.
c) With the same notations, show that for « t « 1, there are at most n distinct ultranorms
among the [q'ex]. Conversely, let L be the set of ultranorms associated with the free A-modules
of dimension n, and let (ex,)o,,;, 0 be an increasing family of ultranorms of L such that ex l = qex o'
Show that there exists a basis of E relative to which all the ex, have the standard form (if u E A
is an element of valuation 1, and L, the free A-module of the x E E such that ex,(x) « 1, consider
the vector spaces L,/uLo on k). Deduce further, that if, for all x E E, t ~ ex,(x) is left-continuous
in (0, 1) then there exists a unique ultranorm ex such that ex, = [q'ex] for all t E (0, 1).
d) The linear group GL(E) operates continuously in '!l1 o(E); show that it operates properly.
For all ex E '!l1 o(E). the stabiliser So of ex in GL(E) is the intersection of the stabilisers of the
[q'ex] for 0 « t « 1 ; deduce that So is an open compact subgroup of GL(E), and hence that
the orbit of each ex E '!l1 o(E) is a closed, discrete subspace of '!l1 o(E).
e) For every ultranorm ex E '!l1 o(E) consider the decreasing sequence of the dimensions of the
vector k-spaces L,/uLo, where L, is the A-module of the x E E such that [q'ex] (x) « 1, and
°
t varies from to 1 ; we call this sequence, the sequence of invariants of ex. In order that ex and ~
belong to the same orbit in '!l1 o(E), it is necessary and sufficient that Xo = X~ (exerc. 12, e»
and that the sequence of the invariants of ex and of ~ should be the same (use exerc. 12, b».
f) Deduce from e) that the space of the orbits '!l1 o(E)/GL(E) is isomorphic with the space
of the orbits T"/G n , where the symmetric group operates on the right on T" by
(ZI' ... , zn) ~ (za(ll' ... , za(n)' *1
14) Generalize the results of No.2 and No.3 to topological vector spaces E over a discrete

°
division ring K, such that there exists a fundamental system of balanced neighbourhoods
of in E (i.e. of neighbourhoods V such that K. V = V).

15) Let E be a normed space of finite dimension n over R or C. Ascribe to the dual E*, the
IIX II :s; 1
>
norm defined by Ilx* I = sup 1< x, x* 1 (GT, X, § 3.2). Show that there exists a basis (eJ

I For the exercises 12 and 13, see O. GOLDMAN and N. IWAHORI, The space of p-adic
norms, Acta math., 109 (1963), pp. 137-177.
TVS I.28 TOPOLOGICAL VECTOR SPACES §3

of E such that, if (en is the dual basis, we have IleJ = I ei I = I for all i. (Let (aJ be a basis
of E formed of vectors of norm 1; consider, the determinant det(~i) for each system of n
vectors Xi = 2:~iPj of norm I, and consider such a system for which the absolute value of
J
this determinant is maximal.)

§ 3

1) a) Show that, if a Hausdorff topological vector space E over a non-discrete valued division
ring K is such that every neighbourhood of 0 contains a vector subspace that is not the single
point 0, then the topology of E cannot be defined by a norm, In particular, a product of an
infinite sequence (En> of Hausdorff topological vector spaces on K, none consisting of the
single point 0, has a topology that cannot be defined by a norm.
b) Consider the product vector space E = K~; for all X = (~,,) E E, put
00

Ixl = I
11=0
2-nl~nl/(1 + I~nl)·

Show that the topology of E is defined by the distance d(x, y) = Ix - yl, that IAxI ~ Ixl if
II.,I ~ 1, IAxI ~ IA,I.lxl if II.,I ;;:, 1 and that, for all Xo E E, IA,xol tends to 0 with 11.,1.

2) Let E and F be two complete, metrisable'Vector spaces over a non-discrete valued division
ring, and let 50 be the topology of F. Let 5 be a Hausdorff topology on F, coarser than 50'
Show that if the linear mapping u of E in F is continuous for the topology 5 on F, it is still
continuous for the topology ·'To on F (use the cor. 5 of!, p. 19).
Deduce that if 5\ and 5 z are two distinct topologies on a vector space E over a non-discrete
valued division ring, compatible with the vector space structure of E, and for each of them
E is metrisable and complete, then there does not exist a Hausdorff topology on E
coarser than 5\ and 52' Give an example of two such topologies on an infinite dimensional
vector space E (note that there exist bijections of E on itself such that both the bijection and
its inverse are not continuous for a normed space topology on E).

3) Let E and F be two Hausdorff topological vector spaces over a non-discrete valued division
ring; and suppose that E is metrisable and complete. Let u be a continuous linear injection ofE
in F, and let G be a vector subspace of u(E); suppose that there exists on G a topology 5
which is finer than the topology induced by that of F, is compatible with the vector space
structure ofG and for which Gis metrisable and complete. Show that the mapping inverse to u,
restricted to G, is continuous for 5 (use I, p. 19, cor. 5).

4) Let E, F be two complete metrisable vector spaces over a non-discrete valued division ring
and let u be a continuous linear mapping ofE in F. Show that ifthere exists in F a closed com-
plementary subspace to u(E), then u(E) is closed in F (use I, p. 19, cor. 5).

5) Let E and F be two complete metrisable vector spaces over a non-discrete valued division
ring and let u be a linear mapping of E in F. Let N be the set of cluster points of u in F with
respect to the filter of neighbourhoods of 0 in E; show that N is a closed vector subspace
of F, and that, in order that u be continuous, it is necessary and sufficient that N be the single
point 0 (use I, p. 19, cor. 5). Show that N is the smallest of the closed vector subspaces M of F
such that, if <I> denotes the canonical homomorphism of F on F 1M, then <I> 0 u is a continuous
mapping of E in F/M.

6) Let E be a complete metrisable vector space over a non-discrete valued division ring K.
a) Let p be a lower semi-continuous mapping of E in the interval (0, + CfJ) of R such that
p(Ax) = IA,I.p(x) for A, i= 0 in K and x E E, such that p(O) = 0 and p(x + y) ~ p(x) + p(y)
§3 EXERCISES TVS 1.29

for any x, y in E. Show that if p is finite in E, then p is continuous (consider the closed set
B of x E E such that p(x) :( 1, and use Baire's theorem).
b) Let (Pn) be a sequence of mappings of E in (0, + (0) satisfying the conditions of a). Show
that if none of the Pn are finite in E then there exists a point x E E such that Pn(x) = + 00
for all n (same method as above).

7) Let E be a complete metrisable vector space over a non-discrete valued division ring K.
We say that a vector subspace M of E is paracomplete if there exists on M a complete
metrisable vector space structure for which the canonical injection of M in E is continuous.
a) Let M, N be two paracomplete subspaces of E such that M + Nand M n N are closed
in E. Show that M and N are closed in E. (Taking quotients by M n N reduces the question
to the case where M n N = {O}, and we can consider then the mapping (x, y) f-+ X + Y of
M x N in E).
b) Show that ifE is the union of an increasing sequence of para complete subspaces, (M)po.
then there exists an indexj such that M j = E. (Use Baire's theorem (GT, IX, &5.3, tho 1) and I,
p. 17, tho I).

8) Let E be a Banach space over a non-discrete valued division ring K. We say that a vec-
tor subspace M of E is strongly paracomplete if there exists a norm Ilxll M on M for which
M is a Banach space and the canonical injection of M in E is continuous.
a) Show that if M and N are two strongly paracomplete subs paces of E, then M + Nand
M n N are also strongly paracomplete subspaces. (On M + N, consider the norm
IlxII M + N = inf(llull M + IlvIIN)' where the lower bound is taken over all pairs (u, v) such that
x = u + v, U E M and v EN.)
b) Let M, N be two strongly paracomplete subs paces ofE such that Nand M + N are closed.
Show that M = M + (M n N) and M n N = M n N (use exerc. 7, a».
9) a) Let a, b be two points of a normed space E on the field R. Denote by 8(A) the diameter
of a bounded set A in E (using the norm metric on E) and define inductively the sequence
(Bn)n;.t of bounded sets in E satisfying the following conditions: B t is the set of those x E E
such that Ilx - all = Ilx -:- bll = ilia - bll ; for n > 1, Bn is the set of those x E Bn - t such
that Ilx - yll :( t8(B n - t ) for all y E Bn - t . Show that the intersection of the Bn is just the
single point tea + b) (note that 8(BJ :( t8(B n - t »).
b) Deduce from a) that if u is an isometry of the real Banach space E on the real Banach space
F, then u is an affine linear mapping of E on F.
 CHAPTER II

Convex sets and locally convex spaces

In §§ 2 to 7 of this chapter, we shall be concerned only with vector spaces and affine
spaces over the field of real numbers R, and when we speak of a vector space or an affine
space without giving its division ring of scalars explicitly, then it is to be understood
that this division ring is thefieldR. For vector spaces on C, see § 8.

§ 1. SEMI-NORMS

Throughout this paragraph, K denotes a non-discrete valued division ring.

1. Definition of semi-norms

DEFINITION 1. - Let E be a left vector space over K. A mapping p of E in

R+ = (0, + w(, is called a semi-norm on E if it satisfies the following axioms:
(SN1) If x E E and AE K then pCb) = IAI p(x).
(SNn) If x E E and y E E then p(x + y) :( p(x) + p(y).

Since p(x):(p(y) +p(x- y) and p(y):(p(x) +p(y-x), from p(y-x)=p(x- y),

we deduce
(1) w(x) - p(y) I :( p(x - y).

Examples. - 1) A norm on E is a semi-norm p such that the relation p(x) = 0

implies that x = 0 (I, p. 3).
2) For every linear form f on E, the function x f--+ If(x)1 is a semi-norm on E.
3) If Pi (1 :( i:( n) is a finite set of semi-norms on E, then clearly p'(x) = sup Pi(X)
l~i~n

L
II

and pl/(x) = \J.iP;(x) (where the \J. i are;;::: 0) are both semi-norms on E.
i= 1
TVS II.2 CONVEX SETS AND LOCALLY CONVEX SPACES §1

A mapping p of E in R+ is called an ultra-semi-norm if it satisfies (SN1) and the

following axiom:
(SN;I) If x E E and y E E, then p(x + y) ~ sup(p(x), p(y)).
Clearly an ultra-semi-norm is a semi-norm.
To say that the absolute value on K is ultrametric (CA, VI, § 6.2) means that it is
an ultra-semi-norm on the left vector space K s ' which is not identically zero.

PROPOSITION l. - Let E be a left topological vector space over K and let p be a semi-
norm on E. The following conditions are equivalent:
a) p is continuous in E.
b) p is continuous at the point 0.
c) p is uniformly continuous.
d) For each real number rl > 0, the set W(p, rl), of those x E E for which p(x) < r:l,

°
is open in E.
e) There exists a real number rl > 0, such that W(p, rl) is a neighbourhood of in E.

°
j) For every real number rl > 0, the Set Yep, rl), of those x E E for which p(x) ~ r:l,
is a neighbourhood of in E.
In fact, the implications c) => a) => b) => d) => e) => j) => c) follow immediately
from (SN1) and inequality (1).

COROLLARY. - If P is a continuous semi-norm on E and q is a semi-norm such that

q ~ p, then q is continuous in E.

When p is an ultra-semi-norm on E, then the sets W(p, (1) and Yep, (1) are both open
and closed. For, we have seen that W(p, (1) is open; on the other hand if z is a cluster
point of W(p, (1), then there exists Y E W(p, (1) such that p(y - z) < (1, and from (SN;I)
we have p(z) < (1, thus W(p, (1) is closed. Also, yep, (1) is closed since p is continuous;
further if p(x) ~ rx and p(y) ~ rx, then p(x + y) ~ rx by (SN;I)' and this shows that
yep, rx) is open.

2. Topologies defined by semi-norms

Let p be a semi-norm on the vector space E over K; for every rl > let Yep, rl) °
be the sJ-Wset of those x of E for which p(x) ~ rl. Clearly, if x E Yep, rl) and A E K
is such that IAI ~ 1, then AX E Yep, rl), in other words Yep, rl) is balanced. Further,
for every- Xo E E, there exists a non-zero scalar 11 E K such that 1111 ;?: p(xo) rl- \
therefore 11- 1 Xo E Yep, rl) that is to say Yep, rl) is absorbent. Finally, from (SNn),
we have Yep, rl/2) + Yep, rl/2) c Yep, rl), and from (SN1) that for every non-zero
scalar Ain K we have AV(p, rl) = Yep, IAI rl). We conclude from these remarks, by I,
°
p. 7, prop. 4, that, when rl varies in the set of numbers> (or only in a sequence of
strictly positive numbers tending to 0) then the sets Yep, rl) constitute a fundamental
°
system of neighbourhoods of for a topology compatible with the vector space
structure of E ; we say that this topology is defined by the semi-norm p. A vector space
E with such a topology is called a semi-normed space. Note that if W(p, rl) is the
subset of x of E such that p(x) < rl, then the W(p, rl) constitute (where rl > 0, or rl
No.2 SEMI-NORMS TVS II.3

varies in a strictly positive sequence of numbers tending to zero) a fundamental

system of neighbourhoods of 0 for the topology defined by p.
If r is a set of semi-norms on E, then the upper bound of the topologies defined by
the semi-norms pEr is compatible with the vector space structure (I, p. 10, cor. 4).
A fundamental system of neighbourhoods of 0, for this topology, is given by the finite
intersections n yepi' el) where Pi E r and eli > O. This topology is said to be defined
i
by the set of semi-norms r. It is the coarsest topology on E amongst those that are
invariant under all translations and for which the semi-norms pEr are continuous.
Let E be a topological vector space over K : a system of semi-norms on E, say r,
is called a fundamental system of semi-norms if the topology on E is the same as the
topology defined by r.
Let E be a vector space over K, with the topology defined by a set of semi-norms r.
For every semi-norm p, we have p(x - z) :( p(x - y) + p(y - z), which shows that
the function (x, y) ~ p(x - y) is a pseudometric on E (GT, IX, § 1 . 1) : it follows from
the definitions that, when p varies in r, the set of these pseudometrics defines the
uniform structure of the topological vector space E.

Remarks. ~ 1) The topology defined by a finite set of semi-norms Pi (1 :( i :( n)

on E, can be defined by the single semi-norm p = sup Pi' But a topology defined by
1 ~i~n

an infinite set of semi-norms cannot, in general, be defined by a single semi-norm (III,

p. 37, exerc. 2).
2) Let (~)<EI be a family of topologies on a vector space E over K, each of which
is defined by a family of semi-norms 1,. Then the topology defined by the set of semi-
norms 1 = U 1, is the upper bound of the topologies ~.
'EI
3) If10 is a set of semi-norms directed by the increasing order relation defined between
two semi-norms p, q on E by « there exists A > 0 such that p :( Aq », then a funda-
mental system of neighbourhoods of 0, for the topology defined by 1o, is obtained by
taking the sets V(p, ex) where p E 10 and ex > O. If 1 is any set of semi-norms on E, then
a filtered set of semi-norms, defining the same topology as 1, is the set 10 of upper
envelopes of all finite families of semi-norms belonging to 1.
4) Even if K = R, the topology of a topological vector space over K cannot always
be defined by a set of semi-norms (cf II, p. 24, corollary).
Example. ~ Let «j OO(R) be the vector space over R of real valued functions that are
infinitely differentiable in R. For every function and every pair of integers n ~ 0,
m ~ 1, put

(2) Pn,mU) = sup If(n)(t)1

-m:::;;t~m

with flO) = f. Obviously the Pn m are semi-norms on «j "'(R). In order that the functions
fa tend to 0 (following a filter 'ff on the set of indices) in «j "'(R) for the topology :Y
defined by the semi-norms Pn,m' it is necessary and sufficient that for all integers n ~ 0,
the functions fa(n) tend to 0 (following ff) uniformly on every compact subset of R. We
say that :Y is the topology of compact convergence for the functions f E «j "'(R) and all
their derivatives (cf III, p. 9).

PROPOSITION 2. ~ On a vector space E, let!!T be the topology defined by a set of semi-

norms r.
TVS 11.4 CONVEX SETS AND LOCALLY CONVEX SPACES §1

(i) The closure of {O} in E, for :Y, is the subset of x E E for which p(x) = 0 for
every semi-norm pEr.
(ii) If:Y is Hausdorff and r is enumerable, then :Y is metrisable.
The proposition follows immediately from the definitions and from GT, IX, ~ 2.4,
cor. 1.
Note that if :Y is metrisable, it may be that :Y cannot be defined by a single norm;
this is the case in the example given above (ef IV, p. 18, Example 4).

Let E be a vector space over K, with the topology defined by a set of semi-norms r.
Let E be the Hausdorff completion of E (I, p. 6), and t be the set of mappings p of E
in R+ where p varies in r (GT, II, § 3.7, prop. 15). By the principle of extending
inequalities, the functions PEt are semi-norms on E, and the functions P(x - y)
form a set of pseudometrics defining the uniform structure ofE (GT, IX, § I .3, prop. I).
We see, therefore, that t is a fundamental set of semi-norms defining the topology
ofE.

3. Semi-norms in quotient spaces and in product spaces

Let E be a topological vector space over K, whose topology is defined by r, a set

of semi-norms. Clearly, the restrictions of the semi-norms of r to a vector sub-space
M of E, define the topology induced on M by that of E.
Let <p be the canonical mapping ofE on the vector quotient space ElM. We show
that, for every semi-norm p on E, the function
(3) p(z) = inf p(x)
q,(x)=z

is a semi-norm on ElM. In fact, it is clear that p satisfies the condition (SN,); on the
other hand, if z', z" are two vectors of ElM, we have:

inf p(x):( inf p(x' + x")

q,(x)=z' +z" q,(x')=z',q,(x")=z"

:( inf (p(x') + p(x"))

q,(x') = z' ,q,(x") = z"

inf p( x') + inf p( x")

q,(x') = z' q,(x") = z"

which shows that p verifies (SNn). We say that p is the quotient semi-norm of p by M.
The same reasoning proves that, if p is an ultra-semi-norm, then so also is p.

This being so, we have (in the notation of No.2)

(4) <p(W(p, a)) = W(p, a) .

for every a > O. In fact, to say that p(z) < a, means that there exists x E E such that
<p(x) = z and p(x) < a, from which the relation (4) follows.
We deduce from this, that, if the set of semi-norms r is directed (II, p. 3, Remark 3),
No.4 SEMI-NORMS TVS II.S

then the quotient topology on E/M is defined by the set of semi-norms p, when p
varies in r.
If N is the closure of {O} in E, the topology of E/N is defined by the quotient semi-
norms p, where p varies in r (even if r is not filtered) : here p(x) = p(x) for every x
belonging to the class x mod N. Note that E/N is none other than the Hausdorff
space associated with E (I, p. 4).
Let E be a vector space over K and (E)tEI be a family of vector spaces over K,
where E t has the topology :T, defined by a set of semi-norms r t . For each 1 E I, let 1,
be a linear mapping of E in E,; clearly when p, varies in the set r t , then the p, 0 1,
form a set r; of semi-norms on E. The topology :!I on E. defined as being the coarsest
of all those which make all the mappings ./; continuous (I, p. 9) is then defined by
the set of semi-norms r' = U r;, this follows from the definition of neighbourhoods
tEl
of 0 for :!I (GT, I, § 2.3, prop. 4).

If the p, are ultra-semi-norms, then so are the Pt 0 .f,.

Let E be a vector space over K, with the topology :!I defined by a family of semi-
norms (P),EI; for every 1 E I, let:!lt be the topology defined by the single semi-norm p"
and denote by E t the space obtained from E using the topology :!It' Then the topology
:!I is the inverse image by the diagonal mapping L1 : E ---> TI E, of the product topo-
'EI
logy on TI E, (I, p. 9, prop. 7). For each 1 E I, write N, for the closure of {O} in E t,
'EI
and by F, = EjNt, the normed space defined by the norm p, corresponding to Pt
(II, p. 4, formula (3)); if <Pt : E, ---> F t is the canonical mapping and <P : (xJf---+(<PJxJ)
the product mapping, we know that the product topology on TI E t is the inverse
'EI
image by <p of the product topology on TI F t (GT, II, § 3.9, prop. 18). The topology :!I
tEl
is. therefore, the inverse image under the composite mapping <p 0 L1 of the product
topology on TI F,. In particular, if :!I is Hausdorff then it follows from II, p. 3. prop. 2
tEl
that the mapping <p 0 L1 is injective, therefore:
PROPOSITION 3. - Every Hausdorff topological vector space E over K, whose topology
is defined by a set of semi-norms, is isomorphic to a sub-space of a product of Banach
spaces.
If, further, the topology of E is defined by an enumerable set of semi-norms, then E
is metrisable (I, p. 16).

4. Equicontinuity criteria of multilinear mappings for topologies defined by semi-

norms

PROPOSITION 4. - Let E; (1 ~ i ~ n) and F be topological vector spaces over K;

we suppose thatJor every i, the topology ofE; is defined by a directed set of semi~norms
r;, and that the topology of F is defined by a set of semi-norms r. Then a set H, of
TVS II.6 CONVEX SETS AND LOCALLY CONVEX SPACES §1

n
multilinear mappings of TI Ei in F is equicontinuous if, and only if, for each semi-norm
i= 1
qE r, and each index i, there exists a semi-norm Pi E r i, and a number a > 0, such
n
that for each function u E H and point (x;) E TI Ei ,
i= 1

(5)

The condition is sufficient since it implies that H is equicontinuous at (0, 0, ... , 0)

and therefore everywhere (I, p. 9, prop. 6).
We show that the condition is necessary. By hypothesis, for every semi-norm q E r
and every number p > 0, we have q(u(x l , XZ' ... , Xn)) ~ P for every function u E H

°
provided that Pi(X;) ~ r:J. i are true for each index i, 1 ~ i ~ n, and certain appro-
priately chosen numbers r:J.i > and semi-norms Pi E rio As K is non-discrete,
we can also suppose that, for every i, we have r:J.i = lAd < 1 where Ai E K. Then let
n
(Xl' XZ' ... , xn) be any point of TI E i , and for each index i, let mi E Z be an integer
i= 1
such that Pi(X) ~ IAil mi + l ; this can be written as p;(Ai-mix) ~ IAil (I ~ i ~ n),
therefore, by hypothesis, we have

(6)

Suppose firstly that one of the p;(x) is zero, then we can take mi E N arbitrarily
large, therefore q(u(x l , XZ' ... , Xn)) = 0. If, on the contrary, all the p;(x) are =1= 0,
take the integer mi such that IAri + 2 < Pi (Xi) ~ IAil mi + 1 for each i; then we have
IAri < IAr 2 Pi(x), from which, by (6), the relation (5) follows with
Q.E.D.

COROLLARY. - The set H is equicontinuous if, and only if, for every semi-norm q E r,
there exists a neighbourhood of
.
° in
n
TI
i= 1
Ei , in which the functions q 0 u, for u E H,

are uniformly bounded.

The condition is evidently necessary, and the demonstration of prop. 4 shows
that it implies an inequality of the form (5) for all u E H, and therefore the equicon-
tinuity of H.
We state explicitly the particular case of prop. 4 for linear mappings.

PROPOSITION 5. - Let E, F be two topological vector spaces over a non-discrete

valued division ring K; suppose that the topology of E (resp. F) is defined by a set
of semi-norms r (resp. r'). Let H be a set of linear mappings ofE in F. Thefollowing
conditions are equivalent :
a) H is equicontinuous.
No.1 CONVEX SETS TVS II.7

b) For every semi-norm q E r', there exists a finite family (P)U"i~n of semi-norms
belonging to r and a number a > Osuch thatJor all x E E and all u E H,

(7) q(u(x)) ~ a. sup Pi(X).

1 ~i~lI

c) For every semi-norm q E r', the mapping sup (q 0 u) is a continuous semi-norm

UEH

on E.

COROLLARY 1. - Suppose that :T, :T are two topologies on a vector space E over K
I

defined, respectively, by two sets of semi-norms rand r'. :T is finer than :T' if, and
only if, for every semi-norm q E r ', there exists a finite family (Pi)1 ~i~n of semi-norms
belonging to r and a number a > Osuch thatJor all x E E, we have q(x) ~ a. sup Pi(X).
1 ~i:$;n

In fact this shows that the identity mapping of E with topology :T, on E with
topology :T is continuous.
I,

COROLLARY 2. - Suppose that the topology:T ofa topological vector space E over K is
defined by a directed set of semi-norms r; for each semi-norm pEr, let Ep be the space
obtained from E using the topology defined by p. The set E' of linear forms on E that
are continuous for :T is the union of the sets E~, where E~ is the set of continuous linear
forms in Ep (p E n.

§ 2. CONVEX SETS

1. Definition of a convex set

For any two points x, y of an affine space E, the set of points AX + IlY where
A ~ 0, 11 ~ 0, A + 11 = 1 is called the closed segment with end points x and y; it
reduces to a point when x = y. The complement of x in this segment is called the
segment with end points x, y which is open at x and closed at y; it is empty if x = y.
Finally the complement of { x, y} in the closed segment with end points x, y is called
the open segment with end points x, y; it is empty when x = y.

DEFINITION 1. - A subset A of an affine space E is convex if, for every two points
x, y of A, the closed segment with end points x, y is contained in A.
As (1 - A) a + AX = a + A(X - a), this definition is equivalent to the following:

°
the set A is convex if, for every point a E A, the transform of A by a homothety of
centre a and ratio A where < A < 1, is contained in A (in other words, A is stable
for these homotheties).

Examples. - 1) Every linear affine variety of E (and in particular the empty set) is
convex.
2) The only non-empty convex sets in R are the intervals (OT, IV, § 2.4, prop. I).
3) Let E be a vector space and Ilxll a norm on E; the unit ball B, formed by the points x
TVS II.S CONVEX SETS AND LOCALLY CONVEX SPACES §2

such that Ilxll ~ 1, is convex since the relations Ilxll ~ 1, Ilyll ~ 1, imply for 0 ~ A ~ 1
that
IIAX + (1 - A) yll ~ A Ilxll + (1 - A) Ilyll ~ A + (1 - A) = 1.

Remark. - Let A be a convex subset of a vector space E; for any scalars a. > 0 and
~ > 0 we have a.A + ~A = (a. + ~) A. In other words, for any x EO A, y EO A, there
exists z EO A such that (a. + ~) z = a.x + ~y; in fact this relation can be written
a. ~ a. ~ a. ~
Z = - - A X + - - A y and we have - - A > 0, - - A > 0 and - - A + - - A =1,
0;+1-' 0;+1-' a. +1-' a. +1-' a. +1-' 0;+1-'
from which the assertion follows, on using def. 1.

°
PROPOSITION 1. - Let (x,) be afamily ofpoints of a convex subset A; every barycentre
L A,xt of the xtformed using positive masses At (such that L At = 1 and At = except

°
for finitely many of the indices, cf A, II,~ 9.3) belongs to A.
Clearly we need only consider the case when the indices are 1, 2, ... , p and At >
for each i; the proposition is trivial if p = 1; we prove the result by induction on p.
p-1 p-1 A. .
Put J.! = L Ai> 0, and y = L
--2 Xi; the induction hypothesis implies that YEA.
i=l i=l J.!
p

Now as Ap = 1 - J.! and L Aixi = J.!y + (1 - J.!) x p ' its follows from def. 1 that
i= 1
p
L Aixi belongs to A.
i= 1

PROPOSITION 2. - Let E and F be two affine spaces and f be an affine linear mapping
of E in F; then the image of a convex subset of E under 1, and the inverse image of a
convex subset of F under fare both convex.
The image under f of the closed segment with end points x, y is the closed segment
with end points f(x), fey), hence the first statement. We deduce that the inverse
image of a closed segment of F under f contains each closed segment whose end
points belong to it; the second statement of prop. 2 follows.
In particular the image of a convex set under a homothety or a translation is a
convex set.

PROPOSITION 3. - In the affine space E, let H be a hyperplane defined by the relation

°
g(x) = 0, where g is a non-constant affine function on E. Then the half-spaces defined
by the relations g(x) ~ 0, g(x) ~ 0, g(x) > 0, g(x) < are convex.
For these are the inverse images under g of intervals of R and thus are convex.
With the notations of prop. 3 the points of a subset M of an affine space are on the
same side (resp. strictly on the same side) of the hyperplane H ifM is contained in one
°
of the half-spaces defined by g(x) ~ 0, g(x) ~ (resp. g(x) > or g(x) < m. °
PROPOSITION 4. - The points of A, a convex subset of an affine space E are strictly
on the same side of a hyperplane H if; and only if, A does not meet H.
No.3 CONVEX SETS TVS H.9

Clearly the condition is necessary. Conversely suppose that it is satisfied and let
g(x) = 0, be an equation defining H (g is an affine linear mapping of E in R). The
set g(A) is convex in R, therefore it is an interval, and 0 ¢ g(A). Hence g(x) is offixed
sign for all x E A.

2. Intersections of convex sets. Products of convex sets

PROPOSITION 5. - The intersection ofany family of convex subsets ofan affine space E
is convex.
The proposition follows immediately from def. I of II, p. 7.

PROPOSITION 6. - Let (E)'EI be a family of vector spaces, and for each t E I, let A.
be a non-empty subset ofE,. Then the set A = n
A. is convex in E =
lEI
n E"
lEI
if, and

only if, for aliI E I, the set A. is convex in E,.

In fact, each projection pr, is a linear mapping and we have A. pr,A and
-1
A = n pr,(A.); the proposition follows from props. 2 and 5 above.
'EI
COROLLARY. - In the space R" every parallelotope (GT, VI, ~ I .3) is a convex subset.
For it is the image under an affine linear mapping of a rectangular parallelepiped,
and this last is convex by prop. 6.

PROPOSITION 7. - Let A and B be two convex subsets of the vector space E. For any
real numbers rx, ~ the set rxA + ~B (set of points of the form rxx + ~y, where x varies
in A, and y in B) is convex.
For rxA + ~B is the image of the convex subset A x B of E x E under the linear
mapping (x, y) H rxx + ~y of E x E in E.

3. Convex envelope of a set

DEFINITION 2. - Given a subset A of an affine space E, we call the intersection of all

convex sets containing A, the convex envelope of A, that is to say (II, p. 9, prop. 5)
it is the smallest convex set containing A.

PROPOSITION 8. - For any family (A.)'EI of convex subsets of an affine space E, the
convex envelope of U A. is precisely the set of linear combinations I A,x, , where
lEI tEl

x, E A., A. ~ 0 except for finitely many indices) and I A, = 1.

0 for alit E I (A, =
'EI
Denote the set of these linear combinations by C, clearly C is contained in every
convex set which contains all the A, (II, p. 8, prop. I); on the other hand A, C C
for every 1. All that remains to be proved is that C is convex. Let x = I A,X"
,
y = I Il,y, be two points of C and rx be a number such that 0 < rx < 1, write
TVS 11.10 CONVEX SETS AND LOCALLY CONVEX SPACES §2

y, = r:lA, + (1 - IX) J.!, for every I E I, and let J be the set (finite) of the indices
of I for which y, -# O. We can write IXX + (l - IX) Y = I y,z" where
'EJ

belongs to A, for all I E J; but I y, = IX I A, + (l - IX) I J.!, = 1, and we see that

tEl tEl tEl

ax + (l - IX) Y E C. The proposition is proved.

COROLLARY 1. - The convex envelope of a subset A ofE is identical with the set of
linear combinations I AiXi , where (x) is any finite family of points of A, the numbers
i
Ai > 0, for all i and I Ai = 1.
i
The dimension of the affine linear variety (A, II, § 9.3) generated by the convex
set A is called the dimension of A,
Let E be a vector space. The convex envelor.~ C, of the balanced envelope of a set
A in E is called the balanced convex envelope (or the symmetric convex envelope)
of A; clearly it is the smallest symmetric convex set that contains A; it is also the
convex envelope of A u (- A), since every point of the balanced envelope of A
belongs to a segment with extremities a and - a where a E A. The set C coincides
with the set of linear combinations I AiXi where Xi E A and I IA;! ~ 1; for it is clear
i i

that this set of points is convex and contains A and - A; it is sufficient to prove
that it is contained in C, and for this we need consider only those linear combinations
for which J.! = I 1\1 > 0; we can then write I \Xi = J.! • I IXiYi with IXi = AdJ.!
i i i
and Yi = Xi' if Ai ~ 0; and IXi = - A;/J.!; Yi = - Xi if Ai < 0; clearly I IXi = 1,
i
and our assertion is proved.

COROLLARY 2. - Let f be an affine linear mapping of the affine space E in the affine
space F ; for each subset A ofE, the convex envelope off(A) is the image under f of the
convex envelope ofA.
There is a similar statement for linear mappings and balanced convex envelopes.

4. Convex cones

DEFINITION 3. -A subset C ofan affine space E is a cone with vertex Xo if C is invariant

for all homotheties of centre Xo and ratio > o.
We shall suppose in this No. and in the one following, that we have chosen the vertex
of the cone being considered, as origin in E; i.e. we suppose that E is a vector space,
and when we speak of a cone, it is to be understood that this cone has vertex O. The
set of points of the form Aa for A > 0 (resp. A ~ 0), where a is a non-null vector, is
called an open half line (resp. closed half-line) originating at O.
A cone C of vertex 0 is said to be pointed if 0 E C, and non-pointed otherwise. A
No. 4 CONVEX SETS TVS 11.11

pointed cone is either the single point {O} or is the union of a set of closed half-lines
originating at O. A non-pointed cone is the union (possibly empty) of open half lines
originating at O. If C is a non-pointed cone, then C u {O} is a pointed cone. If C is a
pointed cone, then C - {O} is a non-pointed cone.
If C is a non-pointed convex cone, then C u {O} is a pointed convex cone. However,
if C is a pointed convex cone, C - {O} is not necessarily convex. We say that a pointed
convex cone is proper if it does not contain any line passing through O. Then
PROPOSITION 9. - A pOinted convex cone C is proper if and only if the non-pointed
cone C, which is the complement of 0 in C, is convex.
If C contains a line through 0 then clearly C is not convex. Suppose now that C
is proper and let x, y be two points of C. The closed segment with end points x, y
is contained in C; ifit contains 0 then AX + (1 - A) y = 0 for some Awith 0 < A < 1,
therefore x = IlY with 11 < O. Thus C contains the line through 0 and x, contrary
to hypothesis.
PROPOSITION 10. - A subset C of E is a convex cone if and only !f C + C c C and
AC c C for all A > O.
For the condition AC c C for all A > 0 characterises the cones. If C is convex
we have C + C = + tc tc
= C (II, p. 8, Remark). Conversely, if the cone C is
such thatC + C c C, then for 0 < A < 1, wehaveAC + (1 - A) C = C + C c C,
which shows that C is convex.
COROLLARY 1. - If C is a non-empty convex cone, the vector space generated by C
is the set C - C (the set of points x - y where x, y vary in C).
For, if V = C - C, then V is non empty, we have AV = V for all A i= 0, and
V + V = C + C - (C + C) c C - C = V, which shows that V is a vector sub-
space. Finally every vector subspace that contains C also contains V.
COROLLARY 2. -lfC is a pointed convex cone, the largest vector subspace contained
in C is the set C n ( - C).
For, if W = C n (- C), then W is non-empty and AW = W for all A i= 0, also
W +W c (C + C) n (- (C + C)) c C n (- c) = W,

which shows that W is a vector subspace. Clearly every vector subspace contained
in C is also contained in W.
Obviously, if f is a linear mapping of E in a vector space F, then fCC), the image
of a convex cone C in E, is a convex cone in F. Every intersection of convex cones
(with vertex 0) in E is a convex cone. For every subset A of E the intersection of
convex cones containing A (these exist, E itself is one such cone) is the smallest
convex cone that contains A; it is called the convex cone generated by A.

PROPOSITION 11. - Let (C.)'EI be afamily of convex cones in E; the convex cone gene-
rated by the union of the C, is identical with the set ofpoints L x,, where J is any finite
'EJ
subset of I and x, E C, for all t E J.
TVS II.12 CONVEX SETS AND LOCALLY CONVEX SPACES §2

In fact, it is obvious that C, the set of such points, is a convex cone containing
the union of the C" and that it is contained in any convex cone which contains
this union.

COROLLARY. - For any subset A of E, the convex cone generated by A, is identical

with the set of linear combinations L AiXi , where (X)iEJ is any finite non-empty family
lEJ
of points of A, and where Ai > 0 for all i E J.
It is sufficient to see that, if a convex cone contains a point x =1= 0 of A then it also
contains the half-line Cx of the points AX where A varies in the set of positive numbers
and that C x is a convex cone.

PROPOSITION 12. - If A is a convex set in E, then the convex cone generated by A

is identical with C = U AA.
1->0
The set C is clearly a cone; it is sufficient to show that C is convex. Let AX, j.ly be
two points of C (A > 0, j.l > 0, X E A, YEA). Let r:l, ~ be two numbers > 0 such that
r:l + ~ = l. Then r:lAx + ~j.ly = (r:lA + ~j.l) z, with z E A, and r:lA + ~j.l > 0;
hence r:lAx + ~j.ly E C.

Remarks. - I ) With the hypotheses of prop. 12, if 0 ~ A, then the cone C is non-pointed,
thus C u {O} is proper.
2) Let A be any convex set in E; consider the convex set Aj = A x {I} in the space
F = E x R and the convex cone C with vertex 0 that is generated by A j . Prop. 12
shows that Aj is the intersection of C and of the hyperplane E x {I} in F. Every
convex set in E can, therefore, be considered as the projection on E of the intersection
of a convex cone with vertex 0 in F and the hyperplane E x {I}.

5. Ordered vector spaces

A preorder structure, on a vector space E, denoted by X ~ Y or Y >
x, is compatible
with the vector space structure of E if it satisfies the following two axioms;
(EO!) If X ~ Y then x + z ~ y + z for all z E E.
(EOn) If x > 0 then Ax >
0 for every scalar A ): O.
The vector space E, carrying these two structures, is called a preordered vector space
(resp. an ordered vector space when the relation of preorder on E is an order).
Note that axiom (EO!) means that the preorder structure and the additive group
structure of E are compatible, that is to say, E carrying these two structures, is a
preordered group (A, VI, p. 3).
Example. - On the space E = RA of all finite real-valued functions defined over A,
the relation of order given by «for all tEA, x(t) ,:;; yet) » is compatible with the vector
space structure of E.

PROPOSITION 13. - (i) The set P, of elements> 0, of a preordered vector space E,

is a pointed convex cone.
(ii) Conversely, ifP is a pointed convex cone in E, then the relation y - x E P is a
preorder relation on E, and the preorder structure that it defines is the only one that is
No.6 CONVEX SETS TVS 11.13

compatible with the vector space structure of E andfor which P is the set ofelements ~ O.
(iii) The relation y - x E P, }vith P a pointed convex cone, is an order relation on E
ifand only if P is a proper cone.
(i) Axioms (E0 1) and (EOn) imply P + PcP and AP c P for all A > O. As
o E P, it follows that P is a pointed convex cone (II, p. II, prop. 10).
(ii) Conversely, if P is a pointed convex cone, the relation P + PcP implies
that the relation y - x E P is a preorder compatible with the additive group structure
ofE (A, VI, p. 3, prop. 3); clearly writing it x< y, the set P is identical with the set of
x ~ 0; further the relation AP c P for all A ? 0 shows that axiom (EOn) is satisfied.
(iii) To say that P is proper means that P n (- P) = { 0 } (II, p. 11, cor. 2), hence
that y - x E P is an order relation.
Example. ~ * Let H be a real Hilbert space; in the vector space 2(H) of continuous
endomorphisms of H. the positive hermitian endomorph isms form a proper pointed
convex cone; this cone, therefore, defines an order structure compatible with the
vector space structure of ~H) and for which the relation A ~ B means that B - A
is a positive hermitian endomorphism. *

For any pointed convex cone P in the vector space E, the set P n (- P) is a vector
subspace, H, of E (II, p. 11, cor. 2). The canonical image P' of P in E/H is a convex
cone and the inverse image of P' in E is P. Thus P' n (- P') = {O}, and P' defines
an order structure on E/H that is compatible with its vector space structure.
A linear form f on a preordered vector space E is said to be positive if x ~ 0 in E
implies f(x) ? O. Or, alternatively, if the convex cone P of elements ~ 0 in E is
contained in the half space of those x for which f(x) ? O. Clearly, in the dual E* to E,
the set of positive linear forms is a pointed convex cone.

6. Convex cones in topological vector spaces

PROPOSITION 14. ~ In a topological vector space E, the closure ola convex set (resp.
of a convex cone) is a convex set (resp. a convex cone with the same vertex).
For,letAbeaconvexset;themapping(x,Y)f---+AX +(1- A)y,whereO < A < 1,
iscontinuousinE x EandmapsA x AinA;thus(GT,I,~2.1,th.l)itmapsA x A
in A, ~hic~ sho~s that ~ is c~nvex. Similarly, if C is a convex cone with vertex 0
then C + C c C and AC c C for all A > O.

DEFINITION 4. ~ For any set A of a topological vector space E, the intersection of

all the closed convex sets containing A is called the convex closed envelope of A; it is
the smallest convex closed set containing A.
From prop. 14, the convex closed envelope of A is the closure 0'£ the convex enve-
lope of A; it is clearly the same as the convex closed envelope of A.
Similarly we call the smallest symmetric, convex, closed set that contains A, the
symmetric convex closed envelope (or the balanced convex closed envelope) of A;
it is the closure of the symmetric c0E.vex envelope of A (II, p. 10); it is also the sym-
metric convex closed envelope of A.
TVS 11.14 CONVEX SETS AND LOCALLY CONVEX SPACES §2

PROPOSITION 15. - Let A; (1 ;:( i;:( n) be a finite number of compact convex sets in
a Hausdorff topological vector space E. Then the convex envelope of the union of the
A; is compact (and is, therefore, the same as the convex closed envelope of this union).
Let B be the compact set in Rn defined by the points 0'1' A2 , ... , An) where
n
I TI
11

A;): 0 (1 ;:( i;:( n), and )~i = 1. Define a continuous mapping of B x A; c Rn x En

i= 1 i= 1
/J
in E by the formula (A1' A2 , ... , A/J' Xl' X 2 ' ... , X/J) H I A;x i • The convex envelope C
;= 1
n n
TI TI A; is compact
1J

of U Ai is the image of B x Ai under this mapping; as B x

i=l ;=1 ;=1
and E is Hausdorff, it follows that C is compact.

COROLLARY 1. - In a Hausdorff topological vector space the convex envelope of a

finite set is compact.

COROLLAR y 2. -- In a topological vector space E, the convex envelope of a finite set

is precompac t.
In fact, let j be the canonical mapping of E in its Hausdorff completion E; if C
is the convex envelope of A, then j(C) is the convex envelope of the finite set j(A)
in E, hence j(C) is compact (cor. 1) and therefore C is precompact (GT, II, § 4.2).

PROPOSITION 16. - Let A be a convex subset, w!!..h at least one interior point X O'
of a topological vector space E. For any point x E A, every point of the open segment
with end points x O ' x lies in the interior of A.
For any point y of this segment, let f be the homothety of centre y and ratio A < 0,
which transforms Yo into x. If V is an open neighbourhood of Xo contained in A,
then fey) is a neighbourhood of x and therefore contains a point fez) E A; now

fez) - y = A(Z - y) = A(Z - fez)) + A(f(Z) - y) ,

hence y - fez) = A ~ 1 (z - fez)), so that y is transformed into z by the homo-

thety g, of centre fez) and ratio J.1 = A/(A - 1); since 0 < J.1 < 1, g transforms V
into a neighbourhood of 0 contained in A. The proposition is proved.

COROLLARY 1. - The interior A

of a convex set A, is itself a convex set; if is not A
empty, then it coincides
c
with the interior o/"A, and A is a convex set that coincides with
the closure of A.
A
It follows· from prop. 16, that if is not empty, then it is a convex set and every
point of Ais a cluster point of A. Next we show that every interior point of Abelongs
to A. Let x be an interior point of A and suppose, for definiteness that x = O. Let V _ CJ

be a symmetric neighbourhood of 0 that is contained in A and let YEA n V; now

- YEA, and therefore, by prop. 16, we see that 0 E A, if y i= 0; this is obviously
true if y = o.
No.7 CONVEX SETS TVS IU5

COROLLARY 2. - The interior C of a convex cone C, is itself a convex cone; if C

is not empty then it coincides with (he interior of C, and C is a pointed convex cone
that coincides with the closure of C.
. Since homotheties of ratio> 0 and centre 0 transform C into itself, they do the
same for C, thus Cis a cone; the remainder of tE.e corollary follows from cor. 1 and
the obvious remark that if C is not empty then C contains the vertex of C.
Let H be a closed hyperplane in the topological vector space E over R; it has an
equation of the form f(x) = Ct., where f is a continuous linear form that is not iden-
tically zero in E (I, p. 11 tho 1). The closed half spaces defined respectively by f(x) ,::;; CI.
and f(x) ): CI. are therefore closed convex sets; their complements defined respec-
tively by f(x) > CI. and f(x) < CI., are open convex sets. We say that these half-spaces
are the closed (resp. open) half spaces determined by H.
PROPOSITION 17. - In a topological vector space E, let A be a set with at least one
interior point, and such that all its points lie on the same side ()f an hyperplane H.
Then H is closed, the interior points of A lie strictly on the same side of H, and the
cluster points of A lie on the same side of H. In particular open (resp. closed) half spaces
are determined by closed hyperplanes.
In fact suppose that H contains the origin and that f(x) = 0 is an equation of
H; suppose, for definiteness, that f(x) ): 0 for all x E A. The half space formed by
the points y such that fey) > - 1 contains at least one interior point, and, by trans-
lation, the same is true of the half space of points y such that fey) > 0; this shows
that H is closed (I, p. 11, corollary). Then we know that f is a strict morphism of E
on R (I, p. 13. corollary), therefore f(A.) is an open set in R. This set cannot contain 0
or it would contain numbers < 0 contrary to hypothesis; it is thus contained in the
open interval) 0, + x{ On the other hand, the half space of those y for which
fey) ): 0 is closed and contains A, therefore it contains A.
COROLLARY. -- Let P be a pointed convex cone, with at least one interior point, of the
topological vector space E. Then each linear form f that is not identically zero on E,
and is positive for the pre order structure defined by P (II, p. 13), is necessarily continuous.
Further, ifx is interior to P then f(x) > 0 and if x is a cluster point ofP then f(x) ): O.
Apply prop. 17 to the case A = P where H is the hyperplane with the equation
f(x) = O.
Remark. - In a topological vector space E, every convex set C is connected. In
fact, if a E C, then C is a union of segments with end point a and closed at a; these
are connected and the result follows from GT, I, § 11.1, prop. 2.

7. Topologies on ordered vector spaces

Let E be an ordered vector space. A topology on E is compatible with the ordered

vector space structure ofE ifit is both compatible with the vector space structure ofE
and subject to the following axiom :
(TO) The convex cone of the x with x ): 0, is closed in E.
TVS II.l6 CONVEX SETS AND LOCALLY CONVEX SPACES §2

An ordered vector space E with a compatible topology is called an ordered topolo-

gical vector space.
Examples. - The space R" with its usual topology and the order structure that is the
product of the order structure of its factors is an ordered topological vector space. On
the other hand, for n :? 2, when R" carries the lexicographical order (S, Ill, ~ 2.6), the
usual topology is not compatible with the ordered vector space structure of R".
Let A be a set; the vector space ~(A; R) of real valued bounded functions defined
on A. with the topology defined by the norm Ilxll = sup Ix(t) I and the order structure
tEA
induced by the product order structure of RA, is an ordered topological vector space.

In an ordered topological vector space E, the set of elements x ~ 0 is closed;

since translations are homeomorphisms, we deduce that. for all a E E, the set of
elements x ): a (resp. x ~ a) is closed. Since {O} is the intersection of the sets x ): 0
and x ~ 0, it follows that {O} is closed and that E is Hausdorff.

PROPOSITION 18. - In an ordered topological vector space E, let H be a set directed by

the relation ~. If the section filter ofH has a limit in E, then this limit is the upper bound
ofH.
For, let b = lim x; for every y E H, the set of x E H such that x ): y is a set of
the section filter of H, therefore b is a cluster point of this set; but as the set x ): y
is closed in E, we have b ): y, thus b is an upper bound of H. On the other hand,
if a is an upper bound of H, then H is contained in the closed set x ~ a; as b is a
cluster point of H, we have b ~ a, which completes the proof (II, p. 72, exerc. 42).

8. Convex functions

DEFINITION 5. - Let X be a convex subset of the affine space E. A real-valued finite

function, defined over X is convex (resp. strictly convex) ifIor any two distinct points x,
y of X and any real number A, 0 < A < 1, we have:

(1) f(h + (1 - A) y) ~ Vex) + (1 - A) fey)

(resp.
(2) f(h + ('1 - A) y) < Vex) + (1 - A) fey)) .

When E = R, this definition of convex function is the same as that in FVR, I,

p. 32. Further, f is convex (resp. strictly convex) in X if, and only if, for every affine
line D e E , the restriction of f to X I i D is convex (resp. strictly convex) in X I i D.
Examples. - If f is an affine linear function on E, then f and f2 are convex functions
on E; this is obvious for f since
f(Ax + (1 - A) y) = Vex) + (I - A) f(y);
on the other hand, if ex = f(x), ~ = fey), then;
Aex 2 + (I - A) ~2 - (b + (1 - A) ~)2 = A(i - A) (ex - W :? 0
No.8 CONVEX SETS TVS 11.17

for 0 < A < 1; further, the restriction of f 2 to an affine line DeE is strictly convex
if flD is not a constant.
A real-valued function f, defined over X, is concave (resp. strictly concave) if - f
is convex (resp. strictly convex). That is to say, for every two distinct points x, y of X
and every number A, such that 0 < A < 1, we have

f(AX + (1 - A) y) ~ Af(x) + (1 - A) f(y)

(resp.
f(AX + (1 - A) y) > Af(x) + (1 - A) f(y») .

A mapping of X in R is affine ifit is both convex and concave (cl II, p. 78, exerc. 11).

PROPOSITION 19. - Let X be a convex set of the affine space E ; and let fbe a real- valued
function defined over X. Denote the set of points (x, a) E E x R for which x E X and
f(x) ~ a (resp. x E X and f(x) < a) by F (resp. F'). Then the folloWing conditions
are equivalent :
a) The function f is convex.
b) The set F in the affine space E x R is convex.
c) The set F' in the affine space E x R is convex.
We show that a) = c). Let (x, a) and (y, b) be two points of F' and 0 < A < 1,
then f(x) < a, f(y) < b and if f is convex

f(AX + (1 - A) y) ~ Af(x) + (1 - A) f(y) < Aa + (1 - A) b

which shows that the point A(X, a) + (1 - A) (y, b) of E x R belongs to F'. Thus
F' is convex.
Next we show that c) =
b). If(x, a), (y, b) are two points ofF then for every E > 0,
(x, a + E) and (y, b + E) belong to F' and, if 0 < A < 1, the same is true of
(AX + (1 - A) y, Aa + (1 - A) b + E); by the definition of F this implies that
(AX + (1 - A) y, Aa + (1 - A) b) belongs to F.
Finally b) = a), for (with the above notation), if ('Ax + (1 - A) y, Aa + (1 - A) b)
belongs to F then

f(AX + (1 - A) y) ~ Aa + (1 - A) b

provided a ~ f(x) and b ~ f(y); hence (1) follows and f is convex.

CoROLLARY. - Iff is convex in X, then for allac E R, the set of x E X such that
f(x) ~ ac (resp. f(x) < ac) is convex.
In fact, it is the projection on E of the intersection of F (resp. F') and the hyper-
plane E x {ac} in E x R.

PROPOSITION 20. - Let f be a convex function, defined over a convex set X of the affine
TVS II.I8 CONVEX SETS AND LOCALLY CONVEX SPACES §2

space E. Then for every family (X)hi~p ofp points of X and every family (A)l~i~P
p

of p real numbers, all ~ 0, such that I Ai = 1, we have:

j= 1

p p
(3) f( I
i= 1
AiXj) ~ L AJeXj) .
i= 1

Iff is strictly convex and if )'i > 0 for all i, then

p p
(4) f( L Ajx;) < L AJ(Xj),
i= 1 i= 1

unless all the Xj are equal.

The inequality (3) follows from II. prop. 19 above and II, p. 8, prop. 1. Suppose
that the Xi are not all equal (which implies p :::;, 2) and that the Aj are all > 0; then the
p
point z = L AjXj differs from at least one Xi. Suppose for definiteness that z #- Xl'
j=1

A.
I=T
P
write z = "'lX 1 + (1 - AI»)'I where Yl = I
j= 2 1
Xi· Then Yl #- Xl and, as
o < Al < 1, we have, by hypothesis,

But by (3) f(YI) ~ J2 1 _'A. Al f(x), and the inequality (4) follows.
P

9. Operations on convex functions

Let X be a convex set of an affine space E. If J; (1 ,.,:; i ,.,:; p) are finitely many convex
functions defined over X and cj (1 ~ i ~ p) are numbers :::;, 0 then the function
L ciJ;
P
f = is convex over X.
j= 1
If(f;) is any family of convex functions defined over X and if g, the upper envelope
of the family in X, is finite then g is convex.
Finally if H is a set of convex functions defined over X, and ~ is a filter on H that
converges simply in X to the finite real valued function fo, then fo is convex over X.

10. Convex functions over an open convex set

PROPOSITION 21. - Letfbe a con vex function, defined over the non-empty open convex
set X in the topological vector space E. Then f is continuous if, and only if, it is bounded
above when restricted to some non-empty open subset U of X.
The condition is obviously necessary, we prove that it is sufficient. Let Xo EX
be a point such that f is bounded above in a neighbourhood V of x o ; we show
No. 11 CONVEX SETS TVS 11.19

firstly that f is continuous at xo' By translation, we can restrict ourselves to the case
when Xo = 0 and f(x o) = 0; moreover we can suppose that the neighbourhood
V is balanced (I, p. 7, prop. 4). Suppose that f(x) ~ a in V; for every E, 0 < E < 1,
we observe that if x E EV, then xlE E V and - xlE E V. Applying inequality (1) of
II, p. 16 to the points xlE and 0 and to the number A = E, we see that
f(x) ~ Ef(xIE) ~ w; applying it to points x and - xlE and the number A = 11(1 + E),
gives f(x) ~ - Ef( - xl E) ~ - w. Thus f(x) is arbitrarily small in EV, if E is
sufficiently small, and f is continuous at x = O.
Now let y be some point of X ; since X is open, there is a number p > 1 such that
z = py belongs to X. Let g be the homothety x f-* AX + (1 - A) z of centre z and

ratio A = - 1..,
p
which transforms 0 into y; for every point g(x) E g(V), we have
from (1)

f(g(x») ~ V(x) + (1 - A) fez) ~ Aa + (1 - A) fez) .

Thus f is bounded above in a neighbourhood of y and hence, by the first part, is

continuous at y. The proposition is proved.

COROLLARY. - Every convex function f defined over an open convex set X in Rn is

continuous in X.
We can suppose that X is not empty. Then there exist, in X, n + 1 affinely inde-
pendent points a i (0 ~ i ~ n) and the convex envelope of these points, S, contains
n
the open non-empty set formed of the points I Aia i with 0 < Ai < 1 for all i and
i=O
n
I Ai = 1. By II, p. 17, prop. 20, f is bounded above in S and therefore is continuous.
i=O

In a topological vector space of infinite dimensions there exist, in general, linear

non-continuous forms (II, p. 80, exerc. 25) and thus convex functions that are not conti-
nuous at any point.

11. Semi-norms and convex sets

Let E be a vector space over R; a mapping p of E in R is positively homogeneous

if, for every A ~ 0 and all x E E we have
(5) p(AX) = Ap(X) .

A positively homogeneous function p on E is convex if, and only if, it satisfies

axiom (SN n) of II, p. 1 for all x, y ofE;
(6) p(x + y) ~ p(x) + p(y) .

In fact, if p is convex, then for x, y in E,

p(t(x + y») ~ tp(x) + tp(y)
TVS 11.20 CONVEX SETS AND LOCALLY COJ\VEX SPACES §2

°
and, by (5), this relation is equivalent to (6). Conversely, if (6) holds, then we also
have for all A such that < A < 1,
p(AX + (1 - A) Y) :( p(AX) + p((l - A) y) = Ap(X) + (1 - A) p(y)

using (5).
A convex positively homogeneous function on E is called sub-linear.
If p is a sub-linear function defined on E; then, by II, § 2.8, corollary, for all a > 0,

°
the set V(p, a) (resp. W(p, a») of points x E E for which p(x) :( a (resp. p(x) < a)
is convex; further this set is absorbent, since for all x E E, there exists A > such that
pCb) = Ap(X) < a.
There is a partial converse of this result:

PROPOSITION 22. - Let A be a convex set, containing 0, in the vector space E. For
all x E E, put

(7) PA(X) = inf p

p> O.xEpA

(0 :( p A (x) :( x). The function p A satisfies

(8)

for all x, y in E and A > 0. If V(PA'::X) (resp. W(PA' et») denotes the set ofx E E for
which p A (x) :( et (resp. P A (x) < et), then

(9)

If A is absorbent then PA is iinite (therefore sublinear).

Since the relations x E pA and AX E ApA are equivalent when A > 0, we have
PA(AX) = APA(X) for A > 0. Let x, y be two points of E. If x (resp. y) is not absorbed
by A thenPA(x) = +x: (resp. PA(Y) = + 00) and the inequality PA(X + y) :(pA(x) +PA(Y)
is obviously true. Suppose there exist et > 0, ~ > °
such that x E ::xA, and y E ~A;
then X +y E etA +~A=(et +~) A (II, p. 8, Remark); and thusPA(x +Y):(PA(X) +PA(Y)'
The inclusion A c V(p A' 1) is clearly true. The inclusion W(p A' 1) c A follows
because A is convex and contains 0. Finally if A is absorbent then PA is obviously
finite.
The function PA defined by (7) is called the gauge of the convex set A. If A is absor-
bent and symmetric, then P A is a semi-norm.

PROPOSITION 23. - Let E be a topological vector space. If A is an open convex set

which contains 0, then PA isfillite and continuous, and A = W(P A, 1). If A is a closed
convex set containing 0, then PA is lower semi-continuous and A = V(PA' 1).
If A is open and contains 0, then it is absorbent. For x E A, there exists p < 1
such that x/p E A, and thus PA(X) < 1; this, combined with (9) gives A = W(PA' 1).
Since the convex function PA is bounded above in the open set A, it is continuous
in E (II, p. 18, prop. 21).
No. I THE HAHN-BANACH THEOREM TVS II.21

Suppose A is closed and contains 0. For every x E E with PA(X) :::; 1, we have
x E pA for all p > 1, therefore x E A since A is closed; remembering (9), this shows
that A = V(PA' 1). For all Il > 0, IlA is therefore the set of x such that PA (x) :::; Il;
°
as PA (x) ~ in E, this shows that PAis lower semi-continuous in E (GT, IV, § 6.2).
A positive sublinear function p over E is the gauge of each convex set A where
W(p, 1) c A e Yep, l).

§ 3. THE HAHN-BANACH THEOREM (ANALYTIC FORM)

1. Extension of positive linear forms

PROPOSITION 1. - Let E be a pre ordered vector space and V be a vector subspace of E

such that every element ofE is bounded above by an element ofY. Given a linear form f
on V that is positive for the preordered vector space structure of V (induced by that olE)
there exists a non-empty set Sf 01 positive linear forms on E, each heing an extension
off. If hE Sf then the values h(a) for a E E lie in the interval (et'. ri"). where

(1) \i' = sup f(z) , \i" = inf f(y).

zeV,z:=:;a YEY,y~a

1. Special case.
Suppose firstly that E = V + Ra. Since the proposition is trivial if a E V. we
confine ourselves to the case a ¢c V. The conditions on V imply that the set A" of
.r E V such that a :::; y is not empty; similarly the set A' of z E V such that - z ~ - a
(i.e. z :::; a) is not empty. For YEA" and z E A', we have z :::; a :::; y, and thus by
hypothesis fez) :::; fey)· Thus 21', 21" are finite and \i' :::; 21". Any linear form fl on E
that extends f is completely determined by fl (a) and for all t~ E R and all x E V, we
have
f 1 (x + 'A.a) = f(x) + 'M1 (a).
Thus t; is positive if and only if the relations

(2) XE V, 'A. E R, x + 'A.a ~ °

imply
(3) f(x) + Af;(a) ~ 0.

° °
As f(IlX) = Ilf(x) and the relations x ~ and IlX ~ are equivalent for Il > 0,
it is sufficient to show that (2) implies (3) in the particular cases 'A. = 0, 'A. = 1 and
). = - 1. For Ie = 0, the fact that (2) implies (3) follows from the hypothesis that f
is positive. For I. = 1, to say that (2) implies (3) means that for - x E A', we have
il (a) ~ f( - x), i.e. fl (a) ~ \i': for 'A. = - 1, (2) implies (3), means that for x E A",
we have I(x) ~ fl (a), i.e. fl (a) :::; \i". The proposition is therefore proved in this
case.
TVS n.22 CONVEX SETS AND LOCALLY CONVEX SPACES §3

n. General case.
Let 3' be the set of pairs (W, g) where W is a vector subspace of E containing V
and 9 is a positive linear form on W which is an extension of f. We order 3' putting
(W, g) :( (W', g') ifW c W' and if g' is an extension of g. Clearly 3' is inductive and
by tho 2 of S, Ill, § 2.4, there is a maximal element (W 0' go)· Suppose W 0 =1= E. Then
there exists a vector b ¢: W 0' and, if WI = W 0 + Rb, the special case above shows
that there exists a positive linear form on WI which is an extension of go; this con-
tradicts the hypothesis that (W 0' go) is maximal. Thus W 0 = E, and the first part
of the proposition is proved. When a E V, the second assertion is obviously true
with cr' = cr" = f(a); if, on the contrary, a ¢: V and one puts VI = V + Ra, the
second assertion follows from the special case I of the proof

°
COROLLARY. - In a topological vector space E with a compatible preorder structure,
let P be the set of elements ;,: in E. Let V be a vector subspace of E containing at least
one interior point Xo of P. Then every positive linear form on V can be extended to a
positive linear form on E.

°
By prop. 1 it is sufficient to show that for every x E E, there exists x' E V such that
x' - x E P. Now let U be a neighbourhood of in E such that Xo + U c P. Then
°
x + Xo + U c X + P, and, hence there exists £ such that < £ < 1 and the point
y = Xo + (1 - £) x belongs to x + P; then every point of the form x + :A(y - x)
belongs to x + P for :A > 0. If we take A = 1/£, then x + :A(y - x) = AXo E V,
from which the conclusion follows.
The conclusion of the corollary is not necessarily valid if one does not assume that V
contains an interior point of P, even if E is of finite dimension and if P n V contains
points interior in V (II, p. 91. exerc. 25, b)).

2. The Hahn-Banach theorem (analytic form)

THEOREM 1 (Hahn-Banach). - Let p be a sub-linear function on a vector space E.

Let V be a vector subspace ofE and f a linear form on V such that, for all y E V, we
have fey) :( p(y). Then there exists a linear form han E that is an extension off and
such that hex) :( p(x) for x E E.
The set of pairs (x, a) such that p(x) :( a is a convex subset P of the vector space
El = E x R (II, p. 17, prop. 19), and it is clearly a pointed cone. Let VI be the sub-
space V x R of El and g(y, a) = - f(y) + a for each point (y, a) E V l ' Then 9
is a positive linear form for the preorder structure on V 1 defined by P n VI; for if
(y, a) E P n VI' then a ;,: p(y) ;,: fey), therefore 9(y, a) ;,: 0. Next let (x, a) EEl ;
we show that (x, a) is less than a point of V 1 for the preorder defined by P. H(x', a') E VI
then (x, a) :( (x', a') if, and only if, p(x' - x) :( a' - a, taking a' ;,: p( - x) + a,
we see that (0, a') of VI satisfies the requirements. Thus we can apply prop. I of II,
p. 21 ; there is a linear form u on El extending 9 and positive for the preorder defined
by P. Therefore u(O, 1) = 9(0, 1) = 1 and u is of the form u(x, a) = - hex) + a,
where h is a linear form on E that extends f; further, for all x E E and all a ;,: p(x),
we have hex) :( a, therefore hex) :( p(x). Q.E.D.
No.1 LOCALLY CONVEX SPACES TVS 11.23

COROLLARY 1. - Let p be a semi-norm on the vector space E. Let V be a vector sub-

space of E and fa linear form on V such that If(y)1 :(: p(y) for all y E V. Then there
exists a linear form h defined on E which is an extension offand is such that Ih(x)1 :(: p(x)
for x E E.
For a semi-norm q and a linear form g on E, the relation g :(: q is the same as
[g[ :(: q. The corollary follows from tho 1.

COROLLARY 2. - Let p be a semi-norm on the vector space E. Given a point Xo E E,

there exists a linear formfdefined over E, such that f(x o) = p(x o) and that I f(x)1 :(: p(x)
for all x E E.
Apply cor. 1 to the vector subspace, V, generated by Xo and to the linear form
SXo f---+ sp(xo) defined over V.
COROLLARY 3. - Let V be a vector subspace of the normed space E and let f be a conti-
nuous linear form over V; then there exists a continuous linear form h defined over E
which extends f and is of the same norm (GT. X, ~ 3.2).
Apply cor. 1, taking p(x) = [[f[[.llxll, which gives [Ihll:(: Ilfll; but clearly
I[h I ~ I f II, and the corollary follows.
The conclusion of cor. 3 is not necessarily valid for continuous linear mappings
of a normed space into an arbitrary normed space (IV, p. 55, exerc. 16, c) and V, p. 65,
exerc. 22).

§ 4. LOCALLY CONVEX SPACES

1. Definition of a locally convex space

DEFINITION 1. - A topological vector space is locally convex (real) if there exists a

fundamental system of neighbourhoods of 0 that are convex sets.
Such a space is called a locally convex space. Its topology is called a locally convex
topology.
The topological vector spaces over R which we study in the rest of this book are
nearly all locally convex.
If V is a convex neighbourhood of 0 in the locally convex space E, then V n (- V)
is a symmetric convex neighbourhood of O. As the closure of a convex set is convex
(II, p. 13, prop. 14) it follows from I, p. 7, prop. 4 that the neighbourhoods of 0 in
E which are closed, symmetric and convex, form a fundamental system of neigh-
bourhoods invariant under homotheties of centre 0 and ratio =1= O.

PROPOSITION 1. - Let IS be a filter base on a vector space E formed from sets that
are absorbent, symmetric and convex. Then the set IB of transforms of the sets of IS
by homotheties of ratio> 0 is afundamental system of neighbourhoods of 0 for a locally
convex topology on E.
Clearly IB is a filter base satisfying (EV 1) and (EVIl) of I, p. 7, prop. 4; it also satis-
fies (EVm) since if V E IS then tv tv+ = v.
TVS II.24 CONVEX SETS AND LOCALLY CONVEX SPACES §4

Note that if :Y is the locally convex topology on E having IE for a fundamental

system of neighbourhoods of 0, then the sets (lIn) V, where n varies in the integers> 0
and V varies in 6, form a fundamental system of neighbourhoods of 0 for the topo-
logy :Y. Then :Y is Hausdorif, if and only if, for every x # 0 in E there exists an
integer n and a set V E 6, such that nx ¢: V; if, further, 6 is enumerable, then the
topology :Y is a metrisable locally convex topology. Conversely, it is clear that if
:Y is a metrisable locally convex topology, then there exists an enumerable funda-
mental system of closed symmetric convex neighbourhoods of 0 for :Y.
COROLLARY. - The topology :Y of a topological vector space E, is defined by a set
of semi-norms (II, p. 3) if, and only if; :Y is locally convex.
The condition is necessary since every semi-norm on E is a convex function, and
so, for Ci > 0, the set of x E E for which p(x) ~ Ci, is convex (II, p. 17, corollary).
Conversely if V is a symmetric, closed, convex neighbourhood of 0 in E, the gauge
p of V is a semi-norm on E such that V is the set of points x of E satisfying p(x) ~ 1
(II, p. 20. prop. 23).
This shows further that a locally convex topology :Y is defined by the set of all
semi-norms that are continuous for :Y. Further, if :Y is metrisable, then it is defined
by an enumerable set of semi-norms.
From the corollary to prop. 1, all the results of § 1 on topologies defined by sets
of semi-norms apply in particular to locally convex topologies over real vector spaces.
A locally convex Hausdorif space E has a completion E that is locally convex. A
complete, metrisable locally convex space is called a Frechet space; every Banach
space is a Frechet space.

PROPOSITION 2. - Let f be a continuous linear form defined over a vector subspace M,

of a locally convex space E; then there exists a continuous linear form h that is defined
over E and is an extension off
From the corollary above and II, p. 7, cor. 2, there exists a continuous semi-norm
p on E, such that if(y)i ~ p(y) for all y E M. By the Hahn-Banach tho (II, p.23,
cor. 1) there exists a linear form h on E that extends f and is such that ih(x)i ~ p(x)
for all x E E, and this implies that h is continuous (II, p. 6. prop. 5).
Remark. - If 9 is a continuous linear mapping of M in the product space R\ then there
exists a continuous linear mapping h of E in RI that is an extension of 9 ; for writing
9 = (g), where the gl are continuous linear forms defined over M, there is an extension
h, of gl for each 1 E I, such that h, is a continuous linear form over E. The continuous
linear mapping h = (h) has the required properties.
Note that if F is a locally convex Hausdorff space and 9 a continuous linear mapping
of M in F. then there does not necessarily exist a continuous linear mapping of E in F
c».
which is an extension of 9 (IV, p. 55, exerc. 16. However there does exist such an exten-
sion when M is finite dimensional (cf cor. 2, below).

COROLLARY 1. - Let E be a locally convex space. If Xo E E is not in the closure of

{O}, then there exists a continuous linear form f defined over E with f(x o) # O.
Apply prop. 2 to the one dimensional vector space M generated by Xo and to the
linear form ~xo f---+ ~ defined over M, which, by I, p. 12, prop. 2, is continuous.
No.2 LOCALLY CONVEX SPACES TVS 11.25

COROLLARY 2. - Let M be a finite dimensional vector subspace of E, a locally convex

Hausdorff space. Then there exists a closed vector subspace N of E, which is the topo-
logical complement of M in E.
There exists a topological complement to M in E if, and only if, the identity mapping
of M on itself can be extended to a continuous linear mapping of E on M, which
mapping is then necessarily a continuous projector (GT, III, § 6.2, corollary). Now,
this follows from the remark above since M is isomorphic to a space R" (I, p. 13,
th.2).

PROPOSITION 3. - In a locally convex space E, the balanced convex envelope of a

precompact set is itself a precompact set.
Let A be a precompact set in E. Given Y, a balanced convex neighbourhood of
o in E, there exist finitely many points ai E A (1 :::; i :::; n) such that A is contained
in S, the union of the neighbourhoods ai + Y (1 :::; i:::; n). Thus C, the balanced
convex envelope of A, is contained in T the balanced convex envelope of S; but T
is contained in B + V, where B denotes the convex envelope of the finite set of points
ai' - a i (1 :::; i :::; n). Now B is precompact (II, p. 14, cor. 2); hence there exist
finitely many points bk E B (1 :::; k :::; m) such that Bk is contained in the union of
the neighbourhoods bk + V. Then C is contained in the union of the neighbourhoods
bk + 2V, and the proposition is proved.
Note that, in an infinite dimensional locally convex Hausdorff space, the convex
envelope of a compact set is not necessarily closed (II, p. 74, exerc. 3).

COROLLARY. - If; in a locally convex Hausdorff space E, a compact set X is contained

in a complete convex set (complete in the uniform structure induced by that of E) then
the convex closed envelope of X is compact.
For this envelope is a closed subset of a complete space, therefore it is complete,
but it is also precompact and Hausdorff.
However in a non complete locally convex Hausdorff space, the convex closed
envelope of a compact set need not be compact (II, p. 87, exerc. 2).

2. Examples of locally convex spaces

1) The space R" is locally convex since the open cubes with centre 0 are convex
(II, p. 9, prop. 6). This is, therefore, also true for all real topological vector spaces
of finite dimension; in fact it follows from the above and I, § 2.3, tho 2 provided that E
is Hausdorff; if not. the Hausdorff space F associated with E is of finite dimension,
therefore locally convex, and the inverse images of convex neighbourhoods of 0 in F
under the canonical mapping E ~ F are convex and form a fundamental system of
neighbourhoods of 0 in E.
2) Let E be a vector space in R, and 1.!3 be the family of all subsets of E that are
absorbent. symmetric and convex. By prop. 1 of II, p. 23 we see that 1.!3 is a funda-
mental system of neighbourhoods of 0 for a locally convex topology :!TO) on E that
TVS II.26 CONVEX SETS AND LOCALLY CONVEX SPACES §4

is the finest of all locally convex topologies on E. This topology is Hausdorff; for
let x i= 0 be any point of E; there exists a basis (i,),EI of E with an cr E I such that
e~ = x; the set of points y = I y,e, such that ly~1 < 1 is absorbent, symmetric and

convex. It does not contain x. From II, p. 24, corollary, it follows that !Yro is also
the topology defined by the set of allsemi-norms on E, thus every semi-norm is
continuous in !Yro.

°
In particular, if u is a linear mapping of E in any locally convex space F, the inverse
image, under u, of every convex neighbourhood of in F is an absorbent convex
set in E; therefore it is a neighbourhood ofO for!Yro and thus u is continuous for !Yro.
Given a convex set C in E, we say that a point a E C is an internal point of C if,
for every line D containing a, the intersection D n C contains an open segment
which contains a; in other words - a + C is absorbent. The point a of the set A
in E is interior to A for !Yro if, and only if, there exists a convex set C with a E C c A,
and such that a is an internal point of C.

More generally, let V be an affine linear variety in E, and C be a convex set contained
in V; a point a E C is an internal point of C relative to V if, in the vector subspace
Vo = - a + V, the point 0 is an internal point of the set Co = - a + c.

When E is of finite dimension, the topology !Yro is just the canonical topology on E
(I, p. 13, tho 2) ; which shows that every internal point of a convex set C in E, is interior
to C for the canonical topology (cf II, p. 74, exerc. 5).
3) Let A be a symmetric convex set in the vector space E over R. The vector
subspace F generated by A is also the convex cone generated by A, since - A = A;

°
this set is the set of AX where x E A and A E R; the set A is absorbent in F and the
sets AA where A > 0, form a fundamental system of neighbourhoods of for a
locally convex topology on F (said to be defined by A), which is defined by the semi-
norm PA' the gauge of A (II, p. 20, prop. 22); we write EA for the locally convex
space obtained by giving F this semi-norm. The space EA is Hausdorff if, and only
if, PA is a norm or alternatively A does not contain any line. If B is a second symmetric
convex set in E and if A c B, then clearly EA c E B , and the canonical injection of
E A in EB is continuous for the topologies defined respectively by A and by B. Further,
if f is a linear mapping of E in a real vector space E', then f(A) is convex and sym-
metric in E' and f is a continuous linear mapping of EA on Ef(A).
Finally, note that if E carries a topology !Y compatible with its vector space struc-
ture, and if V is a symmetric convex neighbourhood ofO for!Y, then the vector space
generated by V is identical with E, since V is absorbent, and the identity mapping
of E in Ev is continuous.

3. Locally convex initial topologies

PROPOSITION 4. -- Let E be a vector space and let (EJ'EI be a family of locally convex
spaces. For each lEI, let f.. be a linear mapping of E in E, ; then the topology .'?I
No.4 LOCALLY CONVEX SPACES TVS 11.27

on E, which is the coarsest making each mapping J; continuous, is a locally convex

topology.
Using II, p. 24, corollary, this is a particular case of the corresponding property
for topologies defined by semi-norms (II, p. 5).
In particular, every vector subspace of a locally convex space, and every product
space of locally convex spaces, is locally convex. Every projective limit of locally
convex spaces is locally convex.
Every enumerable product of Frechet spaces (and in particular every enumerable
product of Banach spaces) is a Frechet space.
Every locally convex Hausdorff space E is isomorphic to a subspace of a product
of Banach spaces and this subspace is closed if E is complete (II, p. 5, prop. 3). Every
Frechet space is isomorphic to a closed subspace of an enumerable product of Banach
spaces (loc. cit.).

4. Locally convex final topologies

PROPOSITION 5. - Let E be a vector space, and (Fa)aEA be a family of topological

vector spaces andfor each rx E A, let ga be a linear mapping ofF a in E.
(i) Denote by m the family of absorbent, symmetric convex subsets V of E such that
g;; l(V) is a neighbourhood of 0 in Fa for every rx; the family mis a fundamental system
of neighbourhoods of 0 in E for a topology :y that is compatible with the vector space
structure.
(ii) A linear mapping f of E in a locally convex space G (resp. a semi-norm p on E)
is continuous for :Y if and only if, for every index rx, f 0 ga (resp. p 0 gJ is continuous
in Fa'
(iii) The topology :Y is the finest of the locally convex topologies on E for which
the ga are continuous.
Further, the topology :Y is the only locally convex topology on E that satisfies
condition (ii) for linear mappings (resp. for the semi-norms).
As m is a filter base invariant under homotheties of ratio > 0, the assertion (i)
follows immediately from II, p. 23, prop. 1. By the definition of m, the topology :Y
is the finest of locally convex topologies on E making the ga continuous; whence (iii).
Finally, it is clear that iff is continuous, so is f 0 ga; conversely if the f 0 ga are conti-
nuous for every rx, then for each symmetric convex neighbourhood W of 0 in G,
the set g;; 1(1 -l(W)) is a neighbourhood of 0 in Fa for each rx. Now f -l(W) is
absorbent, symmetric and convex thus f - l(W) is a neighbourhood of 0 in :Y, and
f is continuous. Similarly if p is a semi-norm on E such that p 0 ga is continuous
for every rx, and if U is the set of points x E E such that p(x) < 1, then, for every rx,
the set g;; l(U) is a convex neighbourhood of 0 in Ea that is symmetric and absorbent;
thus U is a neighbourhood of 0 in E and p is continuous (II, p. 2, prop. I).
The last statement follows from S, IV, ~ 2.5, criterion CST 18.
We say that :Y is the locally convex final topology of the family of topologies :Ya
of the Fa' for the family of linear mappings gao
TVS II.28 CONVEX SETS AND LOCALLY CONVEX SPACES §4

It may happen that .oJ is not the finest of the topologies on E compatible with its
vector space structure and making the f" continuous (II, p. 75, exerc. 15; see also II, p. 75,
exerc. 14).

In the most important case E = L g~(F ~), we get a fundamental system of neigh-
bourhoods of 0 for:T as follows; for each a E A, let V~ be a symmetric neighbourhood
of 0 for :T~, form the union of the g~(V~) for a E A and denote the convex envelope
in E of this union by q(giV~))); since every element ofE is of the form L x~, where J
~E J
is a finite subset of I and x~ E g~(F~), it is immediate that q(g~(V~))) is an absorbent
symmetric convex set in E (each of the V ~ is absorbent in F J; as q (g~(V~))) contains
all the g~(V ~), it is a neighbourhood of 0 for :T. On the other hand, it is clear that for
every symmetric convex neighbourhood V of 0 for:T, we have V ::::J q(V n giF~))),
from which our assertion follows.

COROLLARY 1. - With the notations of prop. 5, let H be a set of linear mappings of E

in the locally convex space G. Suppose that E is the sum of its subspaces g~(F J; then
H is equicontinuous for .:T, if, and only if, for every r:x, the set f 0 g~ where f varies in H,
is equicontinuous in F ~.
Remembering I, p. 9, prop. 6 the argument is similar to that of (ii) prop. 5. Let
W be a symmetric convex neighbourhood of 0 in G and note that if the set f 0 g~,
where f E H is equicontinuous, then the intersection n g;; l( f - l(W)) is a symmetric
fEH
convex neighbourhood of 0 in F~. As this intersection is the same as g;; l( n f -leW))
fEH
and the set n f - leW) is symmetric and convex, everything depends on showing
fEH
n
that it is also absorbent. Now, by hypothesis, every x E E can be written as L g~;(z~),
;= 1
where z~; E F~;. To show that there exists A > 0 such that f(Ax) E W for all f E H,
it is sufficient to consider the case x = g~(z~) with z~ E F~ (since we can pass to the
general case by replacing W by Win). But this case follows from the fact that
g;; l( n f -leW)) is a neighbourhood of 0 in F~.
fEH

COROLLARY 2. - Let (JJAEL be a partition of the index set A. Let (G~)~EA be a family
of locally convex spaces and (FA)AEL be a family of vector spaces. For each A E L, let
h'). be a linear mapping ofF'). in a vector space E; for each A ELand a E JA, let g)..~ be
a linear mapping of G~ in F').. Write J;, = h).. 0 g)..~. Suppose that each F).. carries the
finest locally convex topology that makes the g)..~ (a E J,.) continuous. Then, the finest
locally convex topology on E that makes the J;, continuous, is identical with the finest
locally convex topology making the h).. continuous.
This is a particular case of S, IV, R 2.5 criterion CST 19, and can also be proved
directly using prop. 5.
No.5 LOCALLY CONVEX SPACES TVS II.29

Examples of locally convex final topologies.

I. Quotient space.
Let M be a subspace of the locally convex space F, and <I> be the canonical mapping
of F on F 1M. As the quotient topology on F 1M is locally convex and is the finest
of all the topologies (locally convex or not) which make <I> continuous, it is also the
locally convex final topology for the family consisting of the single mapping <1>.
II. Inductive limits of locally convex spaces.
Let A be an ordered set directed to the right and let (E a , fpa) be an inductive system
lim K and let J:,ct. :E -> E
of vector spaces relative to the set A (A, II, § 6.2) ; let E = ----+ v. Cl

be the canonical linear mapping for each rt. E A. Suppose that each Ea carries a locally
convex topology 5 a, and further suppose that for rt. ~ ~, the mapping fpa: Ea -> Ep
is continuous. Then we say that the locally convex final topology 5 of the family (5a )
relative to the linear mappings fa (resp. the space E carrying the topology 5) is the
inductive limit of the family (5a) (resp. the inductive limit space of the system (E a , f pa ),
or simply of the locally convex spaces EJ Recall that E is the union of the vector
subspaces fa(Ea) and that when rt. ~ ~, we have fa(EJ c fp(Ep); if we endow fa(Ea)
with the final topology for the mapping fa (which is the same as identifying fiEa)
with the quotient space Eal fa-leO)), the topology 5 is also the final topology of the
family of the topologies of the fiEa)' relative to the canonical injections (II, cor. 2
above). Further, the continuity of fpa for rt. ~ ~ implies that the canonical injection
jPa: fa(Ea) -> fp(Ep) is continuous, so that E is also the inductive limit of fa(Ea)
carrying the preceding topologies relative to the injection jpa.
Example. ~ Let X be a locally compact space and E = ff(X; R) the vector space of
finite continuous real valued functions defined over X with compact support. For
every compact subset K of X, let EK be the vector subspace of E formed by those func-
tions fEE which are such that x rt K => f(x) = O. Denote by Y K the topology induced
on EK and by Y u the topology of uniform convergence on X. The inductive limit Y of
the topologies Y K is finer than Y u ; we can show that if X is paracompact and not
compact, then Y is strictly finer than Y u (cf INT, III, 2nd ed., § 1.8). The importance
of Y lies in the fact that the linear forms on E that are continuous in Yare precisely
the real measures on X (INT, III, 2nd., § 1.3).
Remark. ~ In the last example, the topology induced by Yon EK is identical with Y K,
since by definition it is coarser than Y K and, since Y is finer than Y u ' the topology
induced by Y on EK is finer than that induced by Y u ' that is to say Y K.
This reasoning generalises immediately to an inductive limit of locally convex topo-
logies (Yo) when there is a locally convex topology Y' on E such that Yo is the topology
induced on Eo by Y'.
More generally one can ask, when we suppose that Ep c Eo and Yp is the topology
induced by Yo, under what circumstances Y induces Yo on each of the Eo. In general
this is not so (II, p. 80, exerc. 26); but we shall see in the Nos following two important
situations where this does occur.

5. The direct topological sum of a family of locally convex spaces

DEFINITION 2. - Let E be the vector space which is the direct sum (A, II, § 1 .6) of
the family of locally convex spaces (EJ'Ei. For each tEl, let 1. be the canonical injection
TVS II.30 CONVEX SETS AND LOCALLY CONVEX SPACES §4

of E, in E. By the topological direct sum of the family (EJ we mean the space E with
the finest locally convex topology which makes each 1. continuous (this topology is
called the direct sum of the topologies of the EJ
In the remainder of this No. we keep the same notations as in def 2 (unless the
contrary is expressly stated) and we identify, canonically, each E, with a subspace
of E, by means of f..
By the general description of neighbourhoods of a locally convex final topology
given in II, p. 28, we can here obtain a fundamental system of neighbourhoods of
o in E for the direct sum topology, in the following manner; for every family (V.)'EI
where V, is a symmetric convex neighbourhood of 0 in E" consider the convex enve-
lope r((V)), of the union of the V,; the r((V)) for all the families (V) (or only taking
V, for each 1 to be in a fundamental system of neighbourhoods of 0 in E) form afunda-
mental system of neighbourhoods of 0 in E.

Example. - Let (a)'EI be a basis of the vector space E and consider the canonical
topology (I. p. 2, Example 5) on each line Ra,; the direct sum of these topologies
is the finest locally convex topology on E (II, p. 26) ; in fact, if V is an absorbent,
symmetric, convex set in E, then V, = V n Ra, is a neighbourhood of 0 in Ra, and
V clearly contains the convex envelope r((V)).

PROPOSITION 6. - A locally convex topology :Y on E is the direct sum of the topolo-

gies of the E" if and only if, the following property holds: a linear mapping of E in a
locally convex space G (resp. a semi-norm p on E) is continuous, if and only if, for
every 1 E I, the mapping g 0 f.. (resp. p 0 f.) is continuous in E,.
This is a particular case of prop. 5, II, p. 27.

Recalling the definition of the direct sum of a family of vector spaces (A, II, p. 12,
prop. 6), we can say that the topology :Y is the only one for which the canonical
mapping g f---+ (g 0 f..) is a bijection

(1) 2(E; G) ----> n

'EI
2(E,; G)

for every locally convex space G.

COROLLARY. - With the notation of prop. 5, II, p. 27, suppose that E is the sum of the
gaCFa)· Let F be the topological direct sum of the family (Fa)aEA' and let ja: Fa ----> F
be the canonical injection; suppose that g: F ----> E is the linear mapping such that
go ja = ga for all a. E A. IfN is the kernel of g, then the canonical bijection FIN ----> E
associated with g is a topological isomorphism of FIN on E with the topology :Y.
This is a particular case of II, p. 28, cor. 2 remembering II, p. 29, Example I.

PROPOSITION 7. - The canonical injection j: E ----> n E, is continuous when E carries

the direct sum topology of the E, and n
'EI
'EI
E, carries the product topology. When I is
finite, this mapping is an isomorphism of topological vector spaces.
No.6 LOCALLY CONVEX SPACES TVS 11.31

The first assertion follows from the fact that the canonical injections El( -+ TI E,
'EI
are continuous for each K E L If I is finite then} is the identity mapping, and it is suffi-
cient to show that the product topology :T' is finer than the direct sum topology :T.
Now, let V be a convex neighbourhood of 0 for ,r; each set V n E, is a convex neigh-
bourhood of 0 in E, ; if n is the number of elements of I, then the set V contains the
set .!.n L
n
(V n EJ, which is a neighbourhood of 0 for :T', and the proposition is
proved.
When I is infinite, if, for each finite subset J of!, we write E J for the space IT E 1 , with
'EJ
the product topology, then E is the inductive limit of the E J (identified as subspaces of E).

PROPOSITION 8. - Let N, be a subspace of E" for every tEl,

(i) The topology induced on N = L, N, by the direct sum topology :T on E is identical
with the direct sum of the topologies of the N,.
(ii) The canonical mapping h of the topological direct sum space of the EJN, on
E/N (A, II, § 1 .6, formula (26)) is an isomorphism between topological vector spaces.
(i) With the notations introduced above, we consider x = I A,X, belonging to
,
N n r((V.)) ((A.) is a family of numbers ;:;, 0 of which at most finitely many are
non-zero, such that I A, = 1, and x, E V" for alIt E I). Since the sum ofthe N, is direct,
,
we have A,X, E N, for all 1 E I; therefore, for all 1 such that A, > 0 we also have
x, EN, n V" and x belongs to the convex envelope r((N, n VJ), thus (i) is proved.
(ii) Denote canonical mappings as follows: 1. :E, -+ E, h, : EJN, -+ E/N, p, : E, -+ EJN,
and p: E -+ E/N. For every 1 E I, h, 0 p, = p of. and the proposition follows from
II, p. 28, cor. 2 and p. 29 Example L

COROLLARY 1. - IfN, is closed in E, for every 1 E I, then N = I N, in closed in E.

For, the canonical mapping p, : E -+ E, is continuous (II, § 4.5, prop. 6) for every
1E I, hence p,-l(NJ is closed in E, and thus the same is true of the intersection
N = np,-l(NJ
'EI

COROLLARY 2. - If each E, is Hausdorff, so also is E and each E, is closed in E.

To prove the first statement apply cor. I taking N, = {O} for alIt E I; for the second
apply cor. I with N, = E, and Nl( = {O} for every K i= l.
We shall show in III, p. 21, cor. 2 that if the E, are Hausdorff and complete then
so is their topological direct sum E.

6. Inductive limits of sequences of locally convex spaces

In this No. we shall consider an increasing sequence (En) of vector subspaces of a

vector space E, such that E is the union of the En; we suppose that each En carries a
TVS 11.32 CONVEX SETS AND LOCALLY CONVEX SPACES §4

locally convex topology fin' such that, for every n, the topology induced on En by
fin + 1 is coarser than fin' and we give to E the locally convex topology.9"' that is the
inductive limit of the sequence C:1,;) (II, p. 29, Example II); these hypotheses and nota-
tions will be used throughout the rest of this No. without restatement.
It may happen that each .'Tn is Hausdorff but that fI is not; it may also happen that
for each pair of integers n, m such that n :s:; m, the subspace En is closed in Em (using
topology flm) but that En is not closed in E using fI (II, p. 80, exerc. 26).

Lemma 1. - Let ~ be a Cauchy filter on E (for fI); then there exists an integer k,
such that for all N E ~ and every neighbourhood V of 0 in E, the subspace Ek meets
N +V.
We assume the ~ontrary and obtain a contradiction. Suppose that for every k there
exists a convex neighbourhood Vk of 0 and a set Mk E ~ such that
(E k + Vk) n Mk = 0 .
Clearly we can suppose that Vk + 1 C Vk for all k. Let V be the convex envelope of
U (E k nV k ), this is clearly a neighbourhood of 0 for fl. For all n we have V c Vn + En;
k
in fact, every x E V can be written L Aixi where Ai ~ 0, L Ai = 1 and Xi E Vi n Ei
i i
for all i; now for i < n we have Xi E En' therefore L Aixi E En; and for i ~ n we have
i<n
Xi E Vn, therefore L Aixi E Vn since Vn is convex, contains 0 and L Ai
i~n i~n
:s:; 1. Hence
V + En C Vn + En for all n. This being so, let M E ~ be a set that is V-small. For
some integer m, Em n M is not empty; and we conclude that
M c Em +V c Em +V In ;

as ~ is a filter, the set Min meets M and therefore Em + V In ; we have a contradiction

which establishes the lemma.

PROPOSITION 9. - Suppose that the topology induced on En by .9"'n + 1 is identical with

fin for every integer n. Then
(i) The topology induced by fI on En is identical with fin for each n; if the fin are
Hausdorff then fI is Hausdorff
(ii) If, for every n, En is closed in En + 1 (for fin + 1)' then En is closed in E (using fI)
for every n.
(iii) If each En is complete (using .9";J then E is complete using fl.
(i) To prove the first assertion, it is sufficient to prove that the topology fI~
induced by:!T on En is finer than fin' For this, let Vn be a convex neighbourhood of 0
in En for the topology .9"'n ; we are going to construct an increasing seq uence of convex
neighbourhoods of 0 in E n+ p for fln+p,say(Vn+p)P;'I' such that V n+p n En = Vn
for every index p ~ 1. Then the union V of the increasing seq uence (V n + p) will be a
convex set such that V n Ek is a neighbourhood of 0 in Ek (using .'Tk ), for every
index k; therefore V will be a neighbourhood of 0 in E for fI and as V n En = V n'
we have proved that fin' is finer than fin'
No.6 LOCALLY CONVEX SPACES TVS II.33

To define the V n + p we proceed by induction onp using the following lemma:

Lemma 2. - Let V be a convex neighbourhood of 0 in M. a vector subspace of a

locally convex space F. Then there exists a convex neighbourhood W of 0 in F such
that W n M = V. Further, if M is closed in F, then,for every point Xo E CM, there
exists a convex neighbourhood W 0 of 0 in F such that Won M = V and Xo ¢ Wo'
In fact, by hypothesis there exists a convex neighbourhood U of 0 in F such that
U n MeV. Clearly, the convex envelope W of U u V in F is a neighbourhood of 0
in F; we show that W n M = V. For, every point Z E W is of the form A.X +(1- A.) Y
with x E V, Y E U, and 0 :::; A. :::; 1 (II, p. 9, prop. 8); if Z E M, and A. "# I then neces-
sarily y E M, therefore y E U n MeV and hence Z E V; the result is obviously
true if A. = 1. If M is closed in F, the space F 1M is Hausdorff, thus there exists a
convex neighbourhood U o c U of 0 in F such that U o does not meet Xo -+ M;
then the convex envelope Wo of U o u V fulfils the required conditions.

Returning to the theorem, to prove the second part of (i) note that if x E E then
x EEn for some n; if x "# 0 and :Yn is Hausdorff then there is a neighbourhood Vn
of 0 for :Yn , which does not contain x. We see that there is a neighbourhood V of 0
for ,Of such that V n En = V n, hence x ¢ V, and it follows that:Y is Hausdorff.
(ii) Let x E E - En; there exists m > n such that x E Em' thus, as En is closed in
Em for :Ym (because of the hypothesis that g;, + 1 induces :Yn on En for every n) there
exists in the topology :Ym a convex neighbourhood Vm of 0 in Em such that
(x + Vm) n En is empty. Now we saw in (i) that there exists a convex neighbourhood
V of 0 for :Y such that V n Em = V m; and thus (x + V) n Em = X + V m' therefore
(x + V) n Ell = 0, which proves (ii).
(iii) From lemma L if (j is a minimal Cauchy filter for ,(OJ (GT, II, § 3.2) then there
exists a k such that the trace of ~ on Ek is a filter ~k ; from (i) this last is a Cauchy
filter for:Yk and thus (jk converges in Ek by hypothesis; but as the filter on E generated
by ~k is finer than ~, we see that ~ has a cluster point for :Y and thus converges for
:Y.
When for all n the topology induced on Ell by ,Of" + 1 is just :Yn we say that :Y is the
strict inductive limit of the sequence (:Yn ) and that the space E with the topology:Y
is the strict inductive limit of the sequence of locally convex spaces En'

Remarks. - 1) Suppose that E is the union of an increasing directed, non-enumerable

family of subspaces (Ea)aeI' each E. having a locally convex topology :T., such that,
for E. c E~, the topology induced on E. by :T~ is identical with ,Y,. It may be the case
that the topology mduced on each E. by the topology :T is equal to ,Y. and that the
E. are Hausdorfl" and complete, but that E is not complete for :T (INT, III, 2nd ed., § 1,
exerc. 2),

z
2) Let F be a locally convex space, which is the union of an increasing sequence of
vector subspaces (F,,), and for each index n, let:Tn be the topology induced on Fn by the
topology .7 of F. One should beware that in general ,Y is not the inductive limit of
the :Y".
3) Suppose that E is the strict inductive limit of the sequence (En); if F is a closed
(in:T) vector subspace ofE, it may be the case that the strict inductive limit of the topo-
TVS 11.34 CONVEX SETS AND LOCALLY CONVEX SPACES §4

logies induced by the :Yn on F n En is strictly finer than the topology induced by :Y (IV,
p. 63, exerc. 10).

PROPOSITION 10. - Let E, F be two locally convex spaces. Suppose that:

1) There exists a family of FN!chet spaces (E,.), and for each a. a linear mapping
g~ : E~ -> E, such that the topology of E is the final locally convex topology for the
family (g~).
2) There exists a sequence of FN!chet spaces (Fn) and for each n a continuous linear
injection jn: Fn -> F such that F = U1n(Fn)'

Then every linear mapping u of E in F, whose graph is closed in E x F, is necessarily

continuous.
To prove that u is continuous, it is sufficient to show that for every CI., the mapping
u 0 g~ : Ea -> F is continuous (II, p. 27, prop. 5). Now the graph of u 0 ga is the inverse
image of the graph of u under the continuous mapping g~ x IF: E~ x F -> E x F,
and therefore is, by hypothesis, closed in E~ x F. We can, therefore, restrict ourselves
to the case when E itself is a Frechet space. But then the proposition is a particular
case of I, p. 20, prop. 1.

COROLLARY. - With the same hypotheses on E and F as in prop. 10 and assuming

that E is Hausdorff, then every continuous surjective mapping v of F in E is a strict
morphism.
LetNbe the kernel ofv and write N n = jn~ l(N); then themappingj~: FnlNn -> FIN,
deduced from jn by taking quotients, is injective and continuous, also FnlNn is a
Frechet space (since N n is closed) and FIN is the union of the images under K By
hypothesis, in the canonical factorisation v: F -> F IN ~ E, the linear mapping
w is bijective and continuous and its graph in (FIN) x E is therefore closed (GT, I,
§ 8.1, cor. 2 of prop. 2). By the remarks at the beginning and by prop. 10, the inverse
mapping u of w is therefore continuous and the corollary is proved.
* Prop. 10 and its corollary apply in particular when E is a complete bornological
space (III, p. 12) and F is the inductive limit of a sequence of Frechet spaces. *

7. Remarks on Frechet spaces

Weare going to consider prop. 2 of GT, IX, § 3. 1 in the case of locally convex
spaces.
PROPOSITION 11. - Let E be a metrisable locally convex space. The topology of E
can be defined by a distance that is invariant under translations, and for which the
open balls are convex.
Let (Pn)nEN be a sequence of semi-norms that define the topology of E. Let dn
be the pseudometric defined by dn(x, y) =inQpn(x- y), lin) for x, y in E; it is invariant
under translations. For every n ~ 0, and every real number R ~ 0, let Bn,R be
the set of x E E for which dn(x, 0) < R. If R ~ lin, then Bn,R = E, and in the other
case Bn,R is formed from the x E E such that Pn(x) < R; in all cases Bn,R is convex.
No.7 LOCALLY CONVEX SPACES TVS 11.35

For x, Y in E define d(x, y) = sup dn(x, y). We see immediately that d is a distance,
nEN
invariant under translations on E and defining the topology of E. For Xo E E and
R ~ 0, the open ball with centre Xo and radius R (for the distance d) is equal to
n (xo + Bn,R)' therefore it is convex.
nEN
PROPOSITION 12. ~ Let E and F be two Frechet spaces and u a continuous linear
mapping ofE on F. Then there exists a section oju that is continuous though not neces-
sarily linear.
By prop. 11 there exists a distance d in E, invariant under translations, defining
the topology of E and for which open balls are convex. Given y and y' in F, let 8(y, y')
be the distance apart of the closed sets u- l(y) and u- 1(y') in E. As u is a strict mor-
phism (I, p. 17, tho 1) the remark of GT, IX, § 3.1 shows that 8 is a distance on F defin-
ing the topology ofF. We shall construct, inductively, a sequence of continuous map-
pings (sn)nEN of F in E satisfying the following inequalities for all y E F

(2) 8(y, u(sn(Y))) < 2 - n

(3) d(sn(y),sn-1(Y)) < 2- n+ 1 (onlyifn ~ 1).

Suppose then that either n = 0, or n ~ 1 and that sn-1 has been constructed.
Let Yo E F; as u is surjective, the set u- 1(yo) is non-empty, and for n ~ 1, we have
d(u- 1(yo), sn-1(YO)) < r n+ 1 by the induction hypothesis. Therefore there exists
a point Xo of E such that u(x o) = Yo and for n ~ 1, d(xo, sn-1(YO)) < 2- n + 1. As
the mapping sn -1 is continuous, the set of points Y of F which satisfy the inequalities
8(y, Yo) < rn and d(xo, sn-1(Y)) < 2- n+ 1 is an open neighbourhood of Yo' Hence
there exist an open covering (VJiEI of F and constant mappings sn,i of F in E which
satisfy the inequalities (2) and (3) in Vi where one replaces sn by sn,i' As the space F
is metrisable, there exists a continuous partition of unity (f)iEI' that is locally finite
and subordinate to the covering (V)iEI (GT, IX, § 4.5, tho 4 and § 4.4, cor. 1).
For every y E F, put sn(Y) = L Ny). sn,i(Y)' The mapping sn of F in E is continuous;
iEI
as the open balls are convex in E and in F, the mapping sn satisfies the inequalities
(2) and (3) for all Y E F.
From inequality (3) the mappings sn : F -+ E form a Cauchy sequence, for uniform
convergence. As E is complete, the sequence (sJnEN converges uniformly to a conti-
nuous mapping s : F -+ E (GT, X, § 1 .6); formula (2) shows that u 0 s is the identity
mapping of F, thus s is a continuous section of u.

COROLLARY. ~ If L is a compact set in F, then there exists a compact set K in E

such that u(K) = L.
It is sufficient to put K = s(L), where s is a continuous section of u.

Remarks. ~ 1) The corollary to prop. 12 can also be deduced from tho 1 of I, p. 17

and prop. 18 of GT, IX, § 2.lO.
2) We keep the notations of prop. 12. Let p be a continuous semi-norm on E;
TVS 11.36 CONVEX SETS AND LOCALLY CONVEX SPACES §5

for all y E F, put q(y) = inf p(x), so that q is a continuous semi-norm on F (II,
u(x)= y

p. 4). Let <\l be a lower semi-continuous mapping of F in the interval )0. + w(

of R. We show that there exists a continuous section s of u such that po s < q + <\l.
Let So be a continuous section of u (prop. 12) and N the kernel of u. Let Yo E F,
then there exists Zo EN such that p(so(Yo) + zo) < q(yo) + <\l(yo)' There exists an
open neighbourhood W of Yo in F such that p(so(Y) + zo) < q(y) + <\ley) for all
YEW. Hence there is an open covering (W);EI of F and constant mappings t; : F ---+ N
such that p(so(Y) + ti(y») < q(y) + <\ley) for all y E Wi' As F is metrisable, there
exists a locally finite continuous partition of unity subordinated to the covering
(W)iEI' say (gJiEI (GT, IX, § 4.5, tho 4 and § 4.4, cor. 1). The mapping s of F in E
defined by s(y) = so(Y) + I gi(y). tly) fulfils the stated conditions.
iEI

§ 5. SEPARATION OF CONVEX SETS

1. The Hahn-Banach theorem (geometric form)

THEOREM 1 (Hahn-Banach). - Let A be an open convex non empty set of the topo-
logical vector space E and let M be a non-empty linear variety which does not meet A.
Then there exists a closed hyperplane H which contains M and does not meet A.
By translation the problem can be reduced to the case 0 E A, so that A is absorbent.
Let p be the gauge of A (II, p. 20) so that A is the set of points x E E such that
p(x) < 1. On the other hand, let V be the vector subspace of E generated by M;
thus M is a hyperplane in V that does not contain 0, and hence there is a unique
linear form f on V such that M is the set of points y E V for which fey) = 1. The
hypothesis M n A = 0 implies therefore that for all y E V for which fey) = 1,
we have p(y) ~ 1; as f and p are positively homogeneous we have fey) ::s; p(y)
for all y E V such that fey) > 0; finally as p(y) ~ 0 for all y E V, we see that fey) ::s; p(y)
for all y E V. By the analytical form of the Hahn-Banach theorem (II, p. 22, tho 1)
there exists a linear form h on E which extends f and is such that, for all x E E,
hex) ::s; p(x). Let H be the hyperplane in E with the equation hex) = 1. Clearly
H n V = M and H n A = 0. On the other hand the complement ofH in E contains
the open non-empty set A, therefore H is closed in E (I, p. 11, corollary).
Q.E.D.

Remarks. - 1) When 0 E M, tho 1 can be stated as follows: there exists a continuous

linear form in E, such that g(x) = 0 in M and g(x) > 0 in A (II, p. 8, prop. 4).

2) Ifwe apply theorem 1 to the case where E carries the finest locally convex topology
(II, p. 25, Example 2), and if, for the sake of simplicity, we suppose that 0 E A, then we
get the following result (that superficially does not involve topology) : if A is an absor-
bent convex set in the real vector space E and if M is a non-empty linear variety that
does not meet A, then there exists a hyperplane H containing M and such that A lies
on one side ofH. This result is not valid for every convex set A (II, p. 65, exerc. 5).
No.2 SEPARATION OF CONVEX SETS TVS 11.37

2. Separation of convex sets in a topological vector space

DEFINITION 1. - Two non-empty sets A, B of a real topological vector space E are said
to be separated by a closed hyperplane H if A is contained in one of the closed half-
spaces determined by Hand B is contained in the other closed half-space.

DEFINITION 2. - Two non-empty sets A, B of a real topological vector space are

said to be strictly separated by the closed hyperplane H if A is contained in one of the
open half-spaces determined by H, and B is contained in the other open half-space.

PROPOSITION 1. - Let A be an open non-empty convex set and let B be a non-empty

convex set in a real topological vector space E; if A does not meet B then there exists
a closed hyperplane that separates A from B.
For the set C = A - B is open, convex (II, p. 9, prop. 7) and non-empty, also
o ¢ C. By theorem 1 of II, p. 36, there exists a continuous linear form f #- 0 on E
such that fez) > 0 in C. Then, for all x E A, and y E B, we have f(x) > fey). Write
r:J. = inf f(x); r:J. is finite and we have f(x) ? r:J. for all x E A and fey) :( r:J. for all
XEA

y E B; the closed hyperplane H with the equation fez) = r:J. separates A from B.

Remarks. - 1) The hyperplane H does not meet A (II, p. 15, prop. 1 ; if A and Bare
two convex non-empty open sets that do not meet then there exists a closed hyper-
plane that separates A strictly from B.
2) However, when B is not open, it is not necessarily the case that there exists a
closed hyperplane that separates A strictly from B, even if E is of finite dimension,
and even if A does not meet B (II, p. 78, exerc. 12).

DEFINITION 3. - For a subset A of a vector space E, a hyperplane H is called a support-

hyperplane of A, if H contains at least one point of A and all the points of A li~ on the
same side of H.
Let f be a linear form on E that is not identically zero; to say that the hyper-
plane of the equation f(x) = r:J. is a support hyperplane of A means that r:J. is either
the smallest or the largest member of the set f(A) c R. In other words, there exists
a support hyperplane of A parallel to the hyperplane of equation f(x) = 0, if, and
only if, one of the bounds of the set f(A) is finite and belongs to f(A).

PROPOSITION 2. - Let A be a non-empty compact subset of a topological vector space E.

For every closed hyperplane H in E, there exists a support hyperplane of A parallel to H.
For, if f(x) = y is an equation of H, where f is a continuous linear form in E,
the restriction of f to A is continuous, therefore bounded and attains its bounds
in A (GT, IV, § 6.1, tho 1).

This demonstrates that there exist one or two support hyperplanes of A parallel
to H; the first case can only arise when A is completely contained in a hyperplane
parallel to H.

PROPOSITION 3. - In a topological vector space E, let A be a closed convex set with

TVS II.38 CONVEX SETS AND LOCALLY CONVEX SPACES
*5
a non-empty interior. Then every support hyperplane of A is closed and every frontier
point of A belongs to at least one support hyperplane of A.
Every support hyperplane of A is closed, since all the points of A are on the same
side of the hyperplane (II, p. 15, prop. 17). Also if Xo is a frontier point of A, then
o
Xo does not belong to the open non-empty convex set A; after tho 1 of II, p. 36 there
exists a hyperplane H that contains Xo and does not meet A. As A is the closure
A
of (II, p. 14, cor. 1 to prop. 16), it follows from prop. 17 of II, p. 15 that H is a
support hyperplane of A.

3. Separation of convex sets in a locally convex space

PROPOSITION 4. ~ Let A be a closed non-empty convex set in a locally convex space E

and let K be a compact non-empty convex set in E, that does not meet A. Then there
exisis a closed hyperplane H that strictly separates A from K.
For there exists an open convex neighbourhood V of 0 in E such that A + V
and K + V do not meet (GT, II, § 4.3, prop. 4). As A + V and K + V are convex
and open in E, prop. 1 of II, p. 37 shows that there exists a closed hyperplane H
that strictly separates A + V from K + V, and a jortiori A from K.
Remark. - In a Hausdorff locally convex space E, let A and B be two closed non-empty
convex sets that are disjoint, if E is finite dimensional then there exists a closed hyper-
plane that separates A from (II, p. 78, exerc. 13); but this conclusion is not necessarily
true when E is of infinite dimension (II, p. 78, exerc. 10 and 11).

COROLLARY 1. ~ In a locally convex space, every closed convex set A is the inter-
section oj the closed half-spaces which contain it.
In fact, for every point x if A, there exists a closed hyperplane that separates x
strictly from A (using prop. 4).

COROLLARY 2. ~ In a Hausdorff locally convex space, every compact convex set A

is the intersection of the closed half~spaces which contain it and which are determined
by support hyperplanes of A.
For, let Xo if A; {x o } is closed, therefore there exists a closed hyperplane H which
separates Xo strictly from A (prop. 4); let f(x) = a be an equation of H (f a conti-
nuous linear form) and suppose that j(x) > a for all x EA. If we put y = inf f(x),
XEA

the half-space defined by f(x) ~ y contains A, is determined by the support hyper-

plane of equation f(x) = y, and does not contain xo; whence the corollary.
It is possible that a closed convex set that is not compact and has no interior point,
in a locally convex space, does not have any closed support hyperplane (II, p. 86, exerc.
18: cf also V, p. 71, exerc. 11).

COROLLARY 3. ~ In a locally convex space, the closure of each linear variety M is

the intersection of the closed hyperplanes that contain M.
For all x if M, let H be a closed hyperplane that separates x strictly from M;
No.4 SEPARATION OF CONVEX SETS TVS II.39

thus M is parallel to H; the closed hyperplane H l , containing M and parallel to H

does not contain x. The corollary follows.

COROLLARY 4. ~ Let C be a closed convex set in a locally convex space E. A subset

A of E is contained in C, if, and only if, for every real valued continuous affine function
u in E such that u(x) ~ 0 for all x in C, we have u(y) ~ 0 for all y in A.
The condition is obviously necessary. Conversely we show that it is sufficient;
if a point x E A is not contained in C, there exists a closed hyperplane of equation
fez) = Ci separating x strictly from C; if we suppose for example that f(x) < Ci,
then the continuous affine function u = f - Ci contradicts the hypotheses.

COROLLARY 5. ~ In a locally convex space E, the closure of each convex cone C of

vertex 0 is the intersection of closed half-spaces containing C determined by closed
hyperpl!!:!les that pass through O.
For C is a convex cone of vertex 0 (II, p. 13, ~op. 14). For x r/: C, there exists
a closed hyperplane H that separates x strictly from C (prop. 4). It is now just necessary
to apply the following lemma :
Lemma 1. ~ Ifa cone A, with vertex 0, is contained in an open half~space determined
by a hyperplane H, then it is contained in a closed half-space determined by a hyper-
plane Ho, that is parallel to H and passes through O.
Let fez) = Ci with Ci < 0 be an equation of H, so that fez) = 0 is the equation
of Ho. If there exists z E A such that fez) < 0, then there would exist A > 0, such that
j(AZ) = Ci, and as AZ E A, this would contradict the hypothesis.

4. Approximation to convex functions

PROPOSITION 5. ~ Let X be a closed convex set in a locally convex space E. Then

every lower semi-continuous convex function f defined in X is the upper envelope of a
family offunctions that are the restrictions to X of continuous affine linear functions
in E.
For, the set AcE x R of points (x, t) such that x E X and t ~ f(x) is convex
(II, p. 17, prop. 19) and closed, since the function (x, t) 1---+ f(x) - t is lower semi-
continuous. Then let x be any point of X and let a E R be such that a < f(x). By
cor. 1 of II, p. 38, there exists a closed hyperplane H in E x R, that contains (x, a)
and does not meet A. Every linear continuous form on E x R being of the form

(z, t) ---> u(z) + At,

where A E Rand u is a continuous linear form on E, it follows that H has an equa-

tion of the form u(z) + At = Ci, and as H contains (x, a) we have Ci = u(x) + Aa.
Now the point (x,.f(x)) E A does not belong to H and therefore A # O. Dividing
by - A, if necessary, we can write the equation of H as t - a = u(z - x). As
f(x) - a > 0, we have, therefore, fez) > u(z - x) + a for all z E X and this proves
the proposition.
TVS II.40 CONVEX SETS AND LOCALLY CONVEX SPACES §6

Remarks. -1) It follows from prop. 5 that j is the upper envelope of a directed increas-
ing family of functions that are the restrictions to X of functions which are continuous
and convex in E.
2) Suppose further that X is a closed convex cone with vertex 0 and that f is posi-
tively homogeneous. Then f is the upper envelope of a family of functions which are
the restrictions to X of continuous linear jorms in E. For, let (uJ be a family of conti-
nuous affine linear functions in E of which the restrictions to X have j as their upper
envelope. Put u" = L', + Aa, where Aa E R, and where va is a continuous linear form
in E. We have lea = ll,(O) ,,:; j(O) = o. On the other hand, if x E X, we have for every
Il > 0,
1l-IAa + va(x) = 1l- 1 ()'a + V/Il X») = 1l-1ua(llx)":; 1l-1j(IlX) =j(x)
therefore ua ~ va ~ j in X so that j is the upper envelope of the va'
3) The restriction to X of a continuous affine fonction in E is a function that is
affine in X (i.e. both concave and convex II, p. 17); but it may be the case that there
exist continuous affine functions in a compact convex set X c E, that are not the res-
trictions to X of continuous affine functions in E (II, p. 78, exerc. II, c». However:

PROPOSITION 6. - Let f be an upper semi-continuous afflne function in a compact

convex set X, of the Hausdorff locally convex space E. Let L be the set of restrictions
to X of continuous affine functions in E; the set L' of the h E L such that hex) > f(x)
for all x E X, is then decreasing directed and its lower envelope is equal to f
We may suppose that X is non-empty. Let u, v be two elements of L, such that
u(x) > f(x) and v(x) > f(x) for all x E X, and let b be a constant that is an upper
bound of u and v. Let U (resp. V) be the compact convex set of points (x, t) of X x R
such that u(x) ,,:;; t ,,:;; b (resp. vex) ,,:;; t ,,:;; b), and let F be the set of (x, t) E X x R
such that t ,,:;; f(x); F is convex and closed in X x R. The convex envelope K of
U u V does not meet F, since U u V is contained in the set of (x, t) E X x R such
that f(x) < t, a set which is convex and does not meet F. As K is compact (II, p. 14,
prop. 15), we can separate F strictly from K by a closed hyperplane H in E x R.
For every x E X, the hyperplane H separates (x, f(x») strictly from (x, b), and therefore
meets the line {x} x R in a single point w(x); thus H is the graph of a continuous
affine function whose restriction w to X is a member of L, that is a lower bound
for u and v and that satisfies the inequality w(x) > f(x) for all x E X. This proves
that the set L' is decreasing directed. Prop. 5 of II, p. 39, applied to - f shows that
f is the lower envelope of L'.
COROLLARY. - Let f be a continuous affine function in X; then there exists a sequence
(h n ) of elements of L which converges uniformly to f in X.
For, prop. 6 and Dini's theorem (GT, X, § 4.1, tho 1) show that for all n there
exists hn E L such that f,,:;; hn ,,:;; f + lin.

§ 6. WEAK TOPOLOGIES

1. Dual vector spaces

Let F and G be two real vector spaces and let (x, y) ~ B(x, y) be a bilinear form
on F x G. We say that the bilinear form B puts the vector spaces F and G in duality,
No.1 WEAK TOPOLOGIES TVS II.41

or that F and G are in duality (relative to B). Recall that we say that x E F and y E G
are orthogonal (for the duality defined by B) if B(x, y) = 0; we say that a subset M
of F and a subset N of G are orthogonal if every x E M is orthogonal to every YEN
(A, IX, § 1.2).
We say that the duality defined by B is separating in F (resp. in G) if it satisfies
the following condition :
(D() For every x#-O in F, there exists y E G such that B(x, y) #- O.
(resp.
(Du) For every y#-O in G, there exists x E F such that B(x, y) #- 0.)
The duality defined by B is said to be separating if it is both separating in F and
in G. For this to be so, it is necessary and sufficient that the bilinear form B should
be separating in the sense of A, IX, § 1 . 1. More precisely we have the following result :

PROPOSITION 1. - Let F, G be two real vector spaces and B a bilinear form on F x G.

Let
dB:y 1-+ B(., y),
SB: x 1-+ B(x, .)

be linear mappings ofG in the dual F* ofF and ofF in the dual G* ofG, associated
respectively to the right and to the left of B (A, IX, § 1. 1). Then B puts F and G in
a duality separating in G (resp. in F), if and only if dB (resp. SB) is injective.

When F and G are put in separating duality by B, we often identify F (resp. G)

with a subspace of G* (resp. F*) by means of SB (resp. dB)' When we consider F
(resp. G) as a subspace of G* (resp. F*) without specifying how this identification
is to be made, we are always using the preceding identifications; the bilinear forp1
B is then identified with the restriction to F x G of the canonical bilinear form :
<
(x*, x) 1-+ x, x* > (resp. (x, x*) 1-+ <x, x* ».
Examples. -1) Let E be a vector space and let E* be its dual. The canonical bilinear
form (x, x*) 1-+ <x, x* >on E x E* (A, II, § 2.3) puts E and E* in separating duality:
for (Du) is true because of the definition of the relation x* #- 0, and we know on
the other hand, that for all x#-O in E, there exists a linear form x* E E* such that
<x, x* >#- 0 (A, II, § 7.5, tho 6), which proves (D(); the identifications of E with
a subspace ofE** is made here by the canonical mapping CE (loc. cit.).
When E is of finite dimension, the only subspace G of E* that is in separating
duality by the restriction to E x G of the canonical bilinear form, is the space E*
itself; for, E being then canonically identified with E** (loc. cit.), if we had G #- E*,
there would exist a#-O in E such that <a, x* >= 0 for all x* E G (A, II, § 7.5, tho 7),
which contradicts the hypothesis.
2) When E is an infinite dimensional vector space, and E' is a vector subspace
ofE*, the duality between E and E' defined by the restriction to E x E' of the cano-
nical bilinear form is always separating in E' ; it can be separating in E even ifE' #- E*.
The most important case occurs where E is a topological vector space.
TVS 11.42 CONVEX SETS AND LOCALLY CONVEX SPACES §6

DEFINITION 1. - By the dual of a topological vector space E, we mean the subspace

E' of E*, the dual of the vector space E, formed by the continuous linear forms on E.
When E is a Hausdorff locally convex space, the duality between E and its dual
E' is separating: this follows from the Hahn-Banach theorem (II, p. 24, cor. 1) that
<
for every x =1= 0 in E, there exists x' E E' such that x, x') =1= O.

Remarks. -- 1) Whcn E is a topological vector spacc, the dual E* of the I'ector space E
will be called the algebraic dual of E to avoid confusion. We note also that E* is the
dual of the topological vector space obtained by giving E the finest locally convex
topology (II, p. 25, Example 2).
2) The dual E' of a topological vector space does not itself carry a topology, unless
this is expressly stated.
3) If F and G c F* are in separating duality by the canonical bilinear form, then
this is also true of F and G I' for every subspace G I of F* such that G c G 1 _

2. Weak topologies

DEFINITION 2. - Let F and G be two vector spaces put in duality by the bilinear form B.
The coarsest topolor;y on F that makes all the linearjorms B(., y): x f---+ B(x, y) conti-
nuous, where y varies in G, is called the weak topology on F defined by the duality
between F and G, and we denote it by cr(F, G).
Similarly we define the weak topology cr(G, F) on G, interchanging F and G in
definition 1; this possibility of interchanging F and G applies to all the results and
definitions that follow in this paragraph.

We use the adjective« weak» and the adverb « weakly» to denote properties relative
to a weak topology a(F, G) provided there is no possibility of confusion. We shall
speak, for example. of «weak convergence» and « weakly continuous functions» etc.

When G c F*, the notation cr(F, G) will always denote the weak topology defined
by the duality corresponding to the restriction to F x G of the canonical bilinear
form (x, x*) f---+ <x, x*).
<
Without extra hypotheses on F and G, we often write x, y> for the value B(x, y)
of the bilinear form B at (x, y), provided there is no ambiguity; we shall adopt this
convention in the rest of this paragraph,
A vector space F carrying a weak topology of cr(F, G) will be called a weak space.
A weak topology cr(F, G) is locally convex (II, p. 26, prop. 4); more precisely,
it is the inverse image of the product topology of RG by the linear mapping
cI> : x f---+ « x, y) )YEG of F in RG. It is defined by the set of semi-norms x f---+ 1< x, y)1
when y varies in G (II, p. 5). For every rx > 0, and every finite family (Yi)h i<:;n of
points of G, let W(YI' ... , Yn; rx) be the set of the x E F such that I<x, Yi) ~ rx
for 1 ~ i ~ n; these sets (for rx, nand Yi arbitrary) form a fundamental system of
neighbourhoods of 0 for cr(F, G). Note that W(Yl' ... , Yn; rx) contains that vector
subspace of F, of jinite codimension, which is defined by the equations <x, Yi) = 0
for 1 ~ i ~ n.
No.2 WEAK TOPOLOGIES TVS II.43

PROPOSITION 2. - The weak topology a(F, G) is Hausdorff if and only if the duality
between F and G is separating in F.
This is a particular case of II, p. 3, prop. 2.

PROpOSITION 3. - Let F and G be two real vector spaces in duality. Every linear
form on F, that is continuousfor a(F, G), can be written as x ~ <x, y) for some y E G.
The element y EGis unique when the duality is separating in G.
For, to say that the linear form f on F is continuous for a(F, G) means that
there exists a finite set of points Yi E G (1 ~ i ~ n) such that, for all x in F,
If(x) I ~ sup I<x, y) I (II, p. 6, prop. 5). The n relations <x, Yi) = 0 (1 ~ i ~ n)
1 ~i~n

imply therefore j (x) = 0, and hence (A, II, § 7.5, cor. 1), there exists a linear combi-
"
nation y = l: AiYi such that f(x) = <x, y) for all x E F. The uniqueness follows
i= 1
from (Du).
In other words, when the duality is separating in G, and F has the topology
a(E G), then we can identify G canonically with the dual of F for this topology
(II, p. 42, def. 1).

COROLLARY 1. - A family (a) of points of F is total for the topology a(F, G) if, and
only if, for every y i= 0 in G, there exists an index 1 such that <at' y) i= O.
For using prop. 3 and I, p. 13, tho 1, the property expresses the fact that for a(F, G)
no closed hyperplane contains all the at; the corollary follows therefore from cor. 3
of II, p.38.

COROLLARY 2. - A family (aJ of points of F is topologically independent for the

topology a(F, G), if, and only if, for every index 1, there exists an element bt E G
such that: <at' bt ) i= 0 and <a K , bt ) = 0 for all K i= 1.
This means, that for all 1, there exists a closed hyperplane in a(F, G), which con-
tains all the aK with index K i= 1 but does not contain at'

COROLLARY 3. - Let G 1 and G z be two vector subspaces of F*, in duality with F

(for the restriction of the canonical bilinear form). Then a(F, G z) isjiner than a(F, G 1 )
if and only if G 1 C G z ·
The condition is obviously sufficient; conversely, if a(F, G z) is finer than a(F, G 1)'
then every linear form that is continuous for a(F, G 1 ) is also continuous for a(F, G z),
hence G 1 C G z by prop. 3.

COROLLARY 4. - Let G be a vector subspace of the dual F*, of the vector space F.
Then F and G are in separating duality (for the canonical bilinear form) if, and only
if; G is dense in F* in the topology a(F*, F).
This follows from cor. 1.
TVS II.44 CONVEX SETS AND LOCALLY CONVEX SPACES §6

3. Polar sets and orthogonal subspaces

DEFINITION 2. - Let F andG be two (real) vector spaces in duality. For every set M ofF
<
we call the polar of M, the set oj those y E G jor which x, y) ?: - I jor all x E M.
(For complex vector spaces, cf II, p. 64.)

If G l ' G 2 are two subspaces of F* such that G 1 C G 2' then the polar of M in G 1
is the intersection of G 1 with the polar of M in G 2'
When there is no danger of confusion we use MO to denote the polar, in G, of the

°
subset M of F. Similarly we define the polar in F of a set in G.
Obviously, for every scalar Ie =I- and all M c F, we have (leMt = Ie - 1 MO. The
relation MeN c F implies N° c MO; if N absorbs M then MO absorbs N°; for
every family (MJ of sets of F, the polar set of U Ma is the intersection of the polar
a
<
°
sets M~. Since, for y E MO, the closed half-spaces defined by the relations x, y) ?: - 1
contain and M, we see that if Ml is the convex envelope ofM u {o}, then M~ = MO.
Clearly M c MOO. Hence
(MOO)O c MO c (MO)OO = (MOO)O

i.e. MO = MO (cf S, III, ~ 1.5, prop. 2).

OO

If M is a symmetric subset of F, MO is a symmetric subset of G; MO is also in this

case the set of y E G such that I<x, y) I : : ; I for all x E M.

°
PROPOSITION 4. - (i) For any set M of F, the polar set MO is a convex set that con-

°
tains and is closed in G jor the topology cr(G, F).

°
(ii) If M is a cone of vertex 0, then MO is a cone of vertex and it is also the set
<
of y E G such that x, y) ?: jor all x E M.

°
(iii) If M is a vector subspace of F, then MO is a vector subspace of G, and it is
<
also the set of y E G such that x, y) = for all x EM.
<
(i) Since the linear forms y f---* x, y) are continuous for cr( G, F) the statement
follows immediately from the definitions and the fact that a half-space determined

°
by a hyperplane is convex.
(ii) If M is a cone with vertex and if x EM, Y E MO, then as Ax EM, for all Ie > 0,
<
we have Ax, y) ?: - 1, i.e. Ie< x, y) ?: - I. Since this holds for all Ie > 0, it follows
<
that x, y) ?: 0, and (ii) is proved.
(iii) Similarly, if M is a vector subspace of F, the relations x E M, y E MO imply,
<
this time, that Ie< x, y) ?: - I for all real Ie which is possible only if x, y) = 0.
IfM is a vector subspace ofF we say that MO is the orthogonal ofM in G; ifG c F*,
then MO is the intersection ofG, and ofthe subspace orthogonal to M in the algebraic
dual F* of F (A, II, ~ 2.4, def 4).
F or a vector subspace M of F and a vector subspace N of G we say that M and N
are orthogonal if M c N° (or, equivalently, if N c MO).
THEOREM 1 (The bipolar theorem). - Let F, G be two real vector spaces in duality.
For every subset M of F the polar set MOO in F of the polar set MO of M in G is the
closed convex envelope (for cr(F, G)) of M u {O}.
No.3 WEAK TOPOLOGIES TVS lI.45

We have seen that we need only consider the case where M is convex and E M.
Denote the closure, in topology cr(F, G) of M by M, then M is a convex set in F;
°
prop. 4ofII, p.44 shows that MOO ::J M. On the other hand if a E F does not belong
to M then there exists a closed hyperplane H in F which separates a strictly from M
(II, p. 38, prop. 4); since H does not contain 0, there exists y E G such that H has
the equation <x, y) = - 1 (II, p. 43, prop. 3); thus <x, y) > - I for all x E M
and <a, y) < - 1. This implies that y E MO and a ¢ MOO, and the relation MOO = M
follows.

COROLLARY 1. ~ For any family (MJ of closed convex sets of F (in the topology
cr(F, G)), each containing 0, the polar set of the intersections M = n Ma is the c(mvex
closed envelope (for cr(G, F)) of the union of the M~. a
For, if N is this convex closed envelope, then

N° = na M:o = na Ma = M

whence N = N°~ = MO.

The conclusion of cor. 1 does not necessarily hold if the M. are not convex.

COROLLARY 2. ~ For every vector subspace M of F, the subspace MOO is the closure
of M in the topology cr(F, G).
Remark. ~ Every neighbourhood of ° in G in the topology cr(G, F) contains
a neighbourhood V defined by a finite number of inequalities of the form
I<Xi' y) I ~ 1 (1 ~ i ~ n), where the Xi are arbitrary points of F. If A is the sym-
metric convex envelope of the set of the Xi' then V is the polar set A ° of A in G. We

°
can say that the polars in G offinite symmetric sets in F (or of their convex envelopes)
form a fundamental system of neighbourhoods of in G for cr(G, F). If the duality
is separating in F, these convex envelopes are compact for cr(F, G) (II, p. 14, cor. 1
of prop. 15), and of finite dimensions. Conversely every compact, convex set offinite
dimension C in F (with the cr(F, G) topology) is contained in the convex envelope
of afinite subset ofF. For, let M be a vector subspace offinite dimension containing C.
If (e)l "'i"'" is a basis ofM, we can suppose that C is contained in the closed paralle-
°
lotope centre and constructed on the vectors of the basis ei (GT, VI, § 1 .3); now
it is immediate that this parallelotope is the convex envelope of the points I" Ciei
i= 1
with ci = ± 1.
Thus we can say that (if cr(F, G) is Hausdorff) the polars offinite dimensional,

°
convex, compact sets in F (for cr(F, G) or for any Hausdorff locally convex topology
finer than cr(F, G) on F) form a fundamental system of neighbourhoods of for
cr(G, F).

COROLLARY 3. ~ Let !T be the topology of a locally convex space E and let E' be
its dual (II, p. 42, def. 1).
TVS II.46 CONVEX SETS AND LOCALLY CONVEX SPACES §6

(i) The closed convex sets in E are the same for the topology :Y and for the weak
topology cr(E, E').
(ii) For every subset M of E, the polar set MOO in E of the polar set MO of M in
E', is the convex closed envelope of M u {o} for the topology :Y.
Clearly, (ii) follows from (i) and tho 1. From the definition of the dual E', it follows
from II, p. 43, prop. 3 that the continuous linear forms on E for the topology :Y
are the same as the continuous linear forms for cr(E, E'). The closed half-spaces in E
are therefore the same for :Y and for cr(E, E') (II, p. 15, prop. 17) and the assertion
(i) follows therefore from II, p. 38, cor. 1.

4. Transposition of a continuous linear mapping

In this No., we suppose that (F, G) and (F l' G 1) are two vector spaces in duality.

PROPOSITION 5. - Let u be a linear mapping of F in Fl. The following properties

are equivalent :
a) u is continuous for the weak topologies cr(F, G) and cr(F l' G I);
b) there exists a mapping v: G 1 ---+ G such that

(1)

for all YEF and zEG I .

If these properties hold and !f the duality between F and G is separating in G, then
the mapping v satisfying (1) is unique, and v is linear.
If u is continuous for the weak topologies, then, for all Z 1 E G l' the linear form
y f---* <u(y), ZI ) on F is continuous for cr(F, G), thus (II, p. 43, prop. 3) can be written
as y f---* <y, V(ZI) with V(ZI) E G, which shows that a) implies b). Conversely, if b)
is true, for all ZI E G 1 , the linear form

is continuous for cr(F, G) : it follows from the definition of weak topologies that
u is continuous for cr(F, G) and cr(Fl' G 1 ) (I, p. 10, cor. 1). The uniqueness of v
follows from (Dn) and this uniqueness implies that v is linear.
Remark. - Suppose that the duality between F and G is separating in G and that
the duality between F 1 and G 1 is separating in G 1. If we identify G and G 1 with
subspaces of F* and Fi respectively, the conditions a) and b) are equivalent to
tu( G 1) C G; v is the restriction of the transpose tu of u (A, II, § 2.5) to G 1.
We say, simply (when there is no chance of confusion) that v is the transpose
of u (relative to the duality on the one hand between F and G and on the other hand
between F 1 and G 1) and we again use tu to denote it.

COROLLARY. - Suppose that the duality between F and G is separating in G. If u is

a linear mapping of F in F l' that is continuous for cr(F, G) and cr(F l' G 1)' then its
No.4 WEAK TOPOLOGIES TVS II.47

transpose is a linear mapping of G 1 in G, that is continuous for a( G l ' F 1) and a( G, F).

Further if the duality between F 1 and G 1 is separating in F 1 then teu) = u.
It is sufficient to exchange F and F 1 with G and G 1 in prop. 5.

PROPOSITION 6. - Suppose that the duality between F and G (resp. F 1 and G 1) is

separating in G (resp. F 1)' Let u be a linear mapping of F in F 1 that is continuous
for a(F, G) and a(F I , G 1 ). Let A be a set in F and AI a set in Fl ; then:
(i) We have (u(A))O = tu-I(AO).
(ii) We have tu(A~) c (u- 1 (A 1 ))O ; further, if A is closed, (Jor a(F l' G 1)) convex,
and contains the origin, then we have tu(A~) = (u- 1 (A 1 ))o.
Let z 1 E G l' the relation z 1 E (u(AW is equivalent to <u(y), Z I) ~ - 1 for all
YEA, and the relation tu(ZI) E N is equivalent to <y, tu(ZI) ~ -1 for all YEA
and our assertion (i) follows using (1). Next interchanging u and tu and applying (i)
to the set A~ of G 1 we get

(2)

from which, on taking polars

We have ru(A~)) c ru(A~WO by the bipolar theorem (II, p.44, tho 1); the final
statement follows from (2) and the bipolar theorem since then A~o = Al and tu(A~)
is convex and contains the origin.

COROLLARY 1. - With the notations of prop. 6, the relation u(A) c Al implies

tu(A~) eN; iffurther Al is convex, closed (for a(Fl' G 1 )) and contains the origin,
then these two relations are equivalent.
In fact, the relation u(A) c Al equivalent to A c u- I (A 1 ), therefore implies

and conversely the relation tu(A~) c A 0 implies

COROLLARY 2. - Let u be a linear mapping of F in F 1 that is continuous for a(F, G)

and a(F l' G 1 ). We have then
(3) Ker(lu) = (Im(u))O ,

(4) Imcru) = (Ker(uW .

Suppose that the dualities between F and G and between F 1 and Glare separating;
then u(F) is dense in F 1 (Jor a(F l' G 1))' if and only if tu is injective.
TVS 11.48 CONVEX SETS AND LOCALLY CONVEX SPACES §6

Apply prop. 6 with A = F and Al = {O}, using the fact that the weak topologies
a(G, F) and a(Fl' G 1) are Hausdorff. The last assertion results from (4), inter-
changing u and tu.

5. Quotient spaces and subspaces of a weak space

Let F, G be two real vector spaces in duality. Let M be a vector subspace of F,

and consider the subspace N of the orthogonal MO in G; if y I' Y2 are two points
of G that are congruent mod. N then <x, YI) = <X, Y2) for all x E M. For each
class y mod. N, denote the common value of <x, y) when y varies in y by <x, y) ;
clearly (x, j) f---+ <x, j) is a bilinear form on M x (GjN).

PROPOSITION 7. - Let M be a vector subspace of F and N a vector subspace of G

where F and G are two vector spaces in duality. Suppose that M and N are orthogonal
(which is equivalent to saying that N c MO, or M c N°). The vector spaces M and
<
GjN are then in duality by the bilinear form (x,}) f---+ x, }).
(i) The topology a(M, GjN) for this duality is induced by a(F, G) (and in parti-
cular we have a(F, G) = a(F, GjP)).
(ii) The topology CJ(GjN, M) for this duality is coarser than the quotient topology
of a(G, F) hy N; these topologies are identical if and only if M + G' = N°.
(i) Every element of GIN is a class mod. N of an element of G; if Zi (1 :( i :( n)
are elements of G and Zi (1 :( i :( n) is the class of Zi in GjN then the set of Y E M
such that 1< y, Zi) 1 :( li for 1 :( i :( n is the trace on M of the set of those x E F such
<
that 1 x, Zi) 1 : ( li for 1 :( i :( n; the conclusion follows from the definition of
neighbourhoods of 0 for the weak topology.
(ii) Let p: G ~ GIN be the canonical surjection. We show that the quotient
topology :T of a(G, F) by N is identical with a(GjN, W). As, for Z E G, y E N°,
we have <y, <y,
p(z) = z), it follows that every neighbourhood of 0 for a(GjN, N°)
is of the form p(V), where V is a neighbourhood of 0 for a(G, F) saturated for the
relation Z - Z' E N, therefore :T is finer than a(G/N, N°). Conversely let
V = W(Yl' ... , Yn; li) be a neighbourhood of 0 in G for a(G, F), where Yi E F for
1 :( i :( 11 and li > 0; we are going to see that for I :( i :( n, there exist elements
ti E N' such that if one puts V' = W(t l , ... , t n ; li), then pCV') c p(V); this will
show that a(GjN, N°) is finer than ,ry- and therefore is actually identical with :T.
Now, let L be the vector subspace of F generated by N° and the Yi' and denote
by P the complementary subspace of N° in L; it is of finite dimension, say m. Let
(X.)l'" .'" be a basis of P; the restrictions to N of the linear forms x f---+ <x J" z)
<
} -...::.}---.m

are linearly independent, since otherwise there exists x # 0 in P such that x, z) = 0

for all zEN, that is to say x E N°, which contradicts the definition of P. Thus we
conclude that for all z' E G, there exists sEN such that <Xj' Zl) = <Xj' s> for
allj; if z' = z + S, we have <x, z) = 0 for all x E P. This being so, putYi = ti + lVi'
where tiEN' and WiEP; we have <Yi'Z) = <ti,z) = <t;.Z') for 1:( i:( 11;
therefore, for all z' E V', there exists z E V such that z' - zEN, that is to say we
have p(V') c p(V).
No.5 WEAK TOPOLOGIES TVS 11.49

Returning to the case where M is any subspace of N°, note that evidently
a(G/N, M) = a(G/N, M + GO); further, from prop. 3 of II, p.43, we see that,
if y E N° is such that the linear form Z f---* <y, z) is continuous for a(G/N, M),
then necessarily y EM + GO. We conclude that the condition M + GO = N° is
necessary and sufficient for the quotient topology .'Y to be equal to a(G/N, M).

Remark. - The duality between M and GIN (where M and N are two orthogonal
subspaces) is separating in M, if and only if M n GO = {O}; it is separating in
GIN, if and only if N = MO.

COROLLARY 1. - Suppose that the duality between F and G is separating in F. For

a vector subspace M oj F the topology a(G/Mo, M) is identical with quotient topology
of a(G, F) by M if and only if M is closed for the topology a(F, G).
O
,

This follows from prop. 7 putting N = MO, and recalling that MOO is the closure
of M for a(F, G) (II, p. 45, cor. 2).

COROLLARY 2. - Ij M is oj jinite dimension n and the duality is separating in F,

then MO is oj codimension n in G. Ij M is closed jor a(F, G) and oj jinite codimension n
and if the duality is separating in G, then MD is oj dimension n.
For, G/Mo is in separating duality with M; if M is of dimension n, the same is
therefore true of G/M o (II, p.4l, example 1). If M is closed, F/M = F/MoO is in
separating duality with MO ; if F 1M is of dimension n, it is therefore the same for MO
(II, p. 41, example 1).

COROLLARY 3. - Let (F, G), (F l' G 1) be two pairs oj spaces in separating duality
and let u be a linear mapping ojF in F l' which is continuousjor a(F, G) and a(F l' G 1).
Then u is a strict morphism oj F in F l ' if and only if, Im(lu) is a closed subspace in G
jor a(G, F).
Let N = ImCu) c G; we know that N° = Ker(u) in F (II, p. 47, formula (3)).
Let p: F ~ FIN° be the canonical mapping so that u factorises as

p w
u :F ~ F IN° ~ F1 '

where w is injective. The spaces FIN° and N are in separating duality and by for-
mula (1) of II, p. 48, we have <w(y), Z1) = <y, tU(Z1) for all y E FIN° and Z1 E G 1 .
This relation shows that w is an isomorphism of FIN°, carrying the topology
a(F IND, N), on u(F) with the topology induced by a(F l' G 1). The conclusion results
therefore from cor. 1 and the definition of a strict morphism.

COROLLARY 4. - Let (F, G), (Fl' G 1 ) be two pairs in separating duality, and let u
be a linear mapping oj F in F 1 that is continuous jor a(F, G) and a(F l' G 1). Then u
is surjective, if and only if, tu is an isomorphism of G 1 (With topology a(G l' F 1)) on
tu(G 1) with the topology induced by a(G, F).
TVS II.50 CONVEX SETS AND LOCALLY CONVEX SPACES §6

For, to say that u(F) = F 1 is equivalent to saying that u(F) is closed and every-
where dense in F 1 for cr(F l' G 1); cor. 4 follows then from cor. 3 applied to 'u and
of II, p. 47, cor. 2.

Remarks. - 1) Let (F l' G 1)' (F 2' G 2)' (F 3' G 3) be three pairs of spaces in separating
duality and consider a sequence of two linear mappings
(5)
that are continuous for the weak topologies corresponding respectively with G l ' G 2'
G 3; we consider the sequence of transposed mappings

(6)
It is clear that '(v 0 u) = 'u 0 'v, therefore the relation v 0 u = 0 is equivalent to tu 0 tv = O.
The sequence (5) is exact if, and only if, the three following conditions are satisfied
a) 'u 0 tv = 0;
b) Im('v) is dense in Ker('u);
c) tu is a strict morphism of G 2 in G 1 .
This follows in effect from cor. 3 of II, p. 49 and formulae (3) and (4) of II, p. 47.
2) It must not be thought that when u is a strict morphism of F in F l ' then tu is
necessarily a strict morphism of G 1 in G; in other words u can be a strict morphism
without u(F) being closed in Fl for cr(Fl' G 1 ). This is shown by the example where
F is a non-closed subspace of F 1 and G = G liFe, U being the canonical injection.
Similarly, the fact that the sequence (5) is exact does not necessarily imply that (6)
is exact, however, if the sequence (5) is exact and if v is a strict morphism, then the sequen-
ce (6) is exact, by the remark I and by II, p. 49, cor. 3.

6. Products of weak topologies

PROPOSITION 8. - Let (F" G')'EI be a family of pairs of spaces in duality. Let F = TI F,

'EI
be the product space oj the F, and G = EB G, be the direct sum oj the G,. Ij, jor all
'EI
x = (x,) E F and all y = (y,) E G, we write <x, y) = L: <x" y,) (a sum which has
'EI
only jinitely many non-zero terms) then the topology cr(F, G) (relative to the bilinear
<
jorm (x, y) f---* x, y») is the product oj the topologies cr(F" GJ
For, given a topology:Y on F; in order that, for ally E G, the linear form x f---* x, y) <
should be continuous for:Y, it is necessary and sufficient, by the definition of x, y), <
<
that each of the mappings x f---* pr,x, y,) should be continuous for :Y, where 1
is arbitrary in I and y, in G,; but this means that each of the mappings pr, of F in
F, is continuous for :Y and for cr(F" G,) (I, p. lO, cor. 1); this completes the demons-
tration.

Remark. - The duality between F and G is separating in F (resp. in G) if and only

if for all 1 E I, the duality between F, and G, is separating in F, (resp. in G,). If the
duality between F and G is separating in F (resp. G), then, in F (resp. G), the subspace
orthogonal to one G, (resp. F,), canonically identified with a subspace of G (resp. F)
is the subspace of the product of the F K where K i= 1 (resp. the direct sum of the G K
such that K i= 1).
No.7 WEAK TOPOLOGIES TVS 11.51

COROLLARY 1. - Let F and G be two vector spaces in separating duality. /j the

space F (with cr(F, G)) is the direct topological sum oj two subspaces M, N then the
space G (with cr(G, F)) is the direct topological sum ojthe subspaces MO, N° orthogonal
respectively to M and N.
For let p: F ---> M, q : F ---> N be the projectors corresponding to the decomposi-
tion of F into the direct sum of M and N; in these conditions the mapping
(p, q) : F ---> M x N is a topological isomorphism. If M 1 = G IMo, N 1 = GIN°,
then the topologies on M and N (induced by that of F) are identical with cr(M, M 1 ),
cr(N, N 1) respectively (II, p.48, prop. 7). The mapping t(p, q): Ml x N 1 ---> G is
a topological isomorphism when we give M 1 , Nl and G the topologies cr(Ml' M),
cr(N 1 , N) and cr(G, F), by prop. 8. Under this mapping Ml (resp. N 1 ) has as its
image in G the subspace N° (resp. MO), and the topology cr(M 1 , M) (resp. cr(N l' N))
has as its image the topology induced on N° (resp. MO) by cr(G, F), from which the
corollary follows.

COROLLARY 2. - Let (e,)'EJ be a basis oj the vector space F with dual F*, and let
u: R(I) ---> F be an (algebraic) isomorphism dejined by this basis. Then the transposed
mapping tu: F* ---> RI is a topological isomorphism when F* carries the topology
cr(F*, F) and RI the product topology.
We know (A, II, § 2.6, prop. 10) that tu is a bijection, and that if for a x* E F*,
< >
we put e" x* = ~,* for allt E I, then the image tu(x*) is the vector (~,*) of Ri,
< >
so that, for all x = L ~,e, in F, we have x, x* = L ~,~,*. The corollary then
1 tEl

follows from this formula and prop. 8.

7. Weakly complete spaces

PROPOSITION 9. - Let F, G be two vector spaces in separating duality. If F is the

completion oj the space F jor the topology cr(F, G) and if we consider the canonical
injection j: F ---> G*, where G* has the topology cr(G*, G), then the continuous exten-
sion j: F ---> G* ojj is an isomorphism oj topological vector spaces.
For, we see that G*, endowed with cr(G*, G), is Hausdorff and complete (II,
p. 51, cor. 2); if we identify F by j with a vector subspace of G* then the topology
induced on F by cr(G*, G) is cr(F, G), and F is dense in G* in the topology cr(G*, G)
(II, p. 43, cor. 4); from which the proposition follows.

Vector spaces that are complete for a weak topology are therefore the duals G*
of arbitrary vector spaces G endowed with cr(G*, G); after II, p. 51, cor. 2, they are
(topologically) isomorphic to products RI of real lines. To simplify the language, we
shall call them products oj lines (for an intrinsic characterisation of these spaces see
II, p. 85, exerc. 13 and II, p. 81, exerc. 1).
We note that on G*, the cr(G*, G) topology is minimal among the weak topologies
that are Hausdorff; for, a weak topology that is coarser than cr(G*, G) is necessarily
of the form cr(G*, H) where H c G (II, p. 43, cor. 3); but if H =1= G, then there
TYS II.52 CONVEX SETS AND LOCALLY CONVEX SPACES §6

exists a linear form x* E G* that is non null and is orthogonal to H (A, II, § 7.3,
prop. 8), therefore cr(G*, H) is not Hausdorff.
We deduce from this remark that, if F, G are two vector spaces, then a linear
bijection u: G* -+ F*, that is continuous for the topologies cr(G*, G) and cr(F*, F),
is necessarily bicontinuous.
PROPOSITION 10...... Let G he a real vector space and F = G* its dual with the topo-
logy cr(G*, G).
(i) The mapping Y H yo is a hijection oj the set of vector subspaces oj G on the
set oj closed vector subspaces oj F.
(ii) Every closed vector subspace oj F is a product of lines and has a topological
complement.
By the bipolar theorem (II, p. 45, cor. 2) Y H yo is a bijection of the set of vector
subspaces Y of G, closed for cr(G, G*) on the set of closed vector subspaces of F.
But, by definition, every linear form on G is continuous for cr(G, G*), therefore
every vector subspace in G is closed, being defined by a system of equations
< y, yD = 0 (where yi E G*); this proves (i).
Now let W be a closed subspace of F; we have then W = yo with Y = WO in G.
Let Y' be a complement of V in G. We know that F = G* can be canonically iden-
tified with y* EB Y'*, and Y'* identified with yo = W (A, II, § 2.6, cor. to prop. 10);
further (II, p. 50, prop. 8) the topology cr(G*, G) can be identified with the product
of the topologies cr(Y*, Y) and cr(Y'*, Y'); this proves assertion (ii).
Though, for the topology cr(G, G*), every vector subspace of G is closed, we note that
if G is of infinite dimension then the topology cr(G, G*) is not the finest locally convex
topology on G, every neighbourhood of 0 for cr(G, G*) containing a vector subspace
of infinite dimension: it is however the finest of the weak topologies on G (II, p. 43,
cor. 3).

8. Complete convex cones in weak spaces

Lemma 1. -- Let E be a Hausdorff weak space and C a proper cone with vertex 0
in E, that is complete for the uniform structure induced by that of E. Every continuous
linear jorm in E is then the difference between two continuous linear forms in E that
are positive in C.
Let E' be the dual ofE and F be the algebraic dual ofE', with the topology cr(F, E').
Let H = CO - Co be the vector subspace of E' formed by the differences of linear
forms that are continuous in E and positive in C (II, p. 44, prop. 4). It is sufficient
to show that the orthogonal to H in F is {O} (II, p. 41, Example 1). Then let a E F
be orthogonaf to H; as a is orthogonal to CO, it must belong to the bipolar of C
in F. But E is identifiable as a subspace of F, and since C is complete, thus closed
in F, we have a E C (U, p. 44, tho 1). Similarly a is orthogonal to - CO and therefore
a E - C. As C is proper, we have a = O.

PROPOSITION 11. - Let E be a Hausdorjf weak space, and C be a proper convex cone
with vertex 0 in E and which is complete in the un (form structure induced by that of E.
No.8 WEAK TOPOLOGIES TVS II.53

Then there exists a set I and a continuous linear mapping u ofE in the product space
RI with the following properties :
a) u is an isomorphism of C on u(C) for the uniform structures induced respectively
by those of E and of RI.
b) We have u(C) c RI+.
Further, if the uniform structure induced on C by that of E is metrisable, then we
can take I = N.
Let (f)lEI be a family of continuous linear forms in E such that the finite sums of
pseudometrics of the form (x, y) f---+ 1J:(x - y)1 on C x C define the uniform structure
of C. (If the structure is metrisable we can take I = N.) By lemma 1 we can suppose
further that each of the J: is positive in C. Let u be the linear mapping x f---+ (J:(X))lEI
of E in RI. It is clear that u is continuous and that u(C) c RI+. The restriction ulC
is a uniformly continuous mapping that is surjective from C on u(C). Further if
x, y in C are such thatJ,(x) = J,(y) for alit E I, then x = y since the uniform structure
of C is Hausdorff; therefore ulC is bijective. Finally, if W is an entourage of the
uniform structure of C, then there exists a finite set J of I and a number E > 0 such
that the relations If.(x) - J,(y)1 :( E for t E J imply (x, y) E W; therefore ulC is an
isomorphism of C on u(C) for the uniform structures being considered.

COROLLARY 1. - Let E be a Hausdorff weak space and C a proper convex cone of

vertex 0 in E that is complete for the uniform structure induced by that of E. Then
the mapping (x, y) f---+ X + Y of C x C in C is proper.
Because of prop. 11, we can suppose that E = RI and that C = RI+ (GT, I, § 10.1,
cor. 1 and 4). But then the mapping (x, y) f---+ X + Y of C x C in C is written as
((~), (T])) f---+ (I;. + T]), and we can restrict ourselves to proving that the continuous
mapping f:(~, T]) f---+ ~ + T] of R+ x R+ in R+, is proper (GT, I, § 10.1, cor.3) ,
-1
Now,for all ~ E R+, we see that f(~) is the set of pairs (~, ~ - ~) such that 0 :( ~ :( ~,
therefore the inverse image by f of the interval [0, ~l is the set of the (~, T]) E R + x R +
such that ~ + T] :( ~, which is compact. The conclusion follows applying (GT, I,
§ 10.3, prop. 7).

COROLLARY 2. - Let E be a Hausdorff weak space, and C a proper convex cone

with vertex 0 in E, that is complete for the uniform structure induced by that of E.
(i) For every point a of E, the intersection C n (a - C) is compact.
(ii) Let A, B be two closed sets in C. Then A + B is a closed set in C.
(i) The set of the (x, y) E C X C such that x + y = a is compact from cor.
and from GT, I, § 10.2, tho 1, b). Now this set is also the set of the (x, a - x) for
x E C n (a - C), which proves (i).
(ii) If A and B are closed in C, then A x B is closed in C x C, therefore A + B
is closed in C after cor. 1 and GT, I, § 10.1, prop. 1.
TVS 11.54 CONVEX SETS AND LOCALLY CONVEX SPACES §7

§ 7. EXTREMAL POINTS AND EXTREMAL GENERATORS

1. Extremal points of compact convex sets

DEFINITION 1. - Let A be a convex set in an affine space E. Then we say that a point
x E A is an extremal point of A if there does not exist an open segment that is contained
in A and contains x.
In other words, the relations x = AY +0- A)Z, YEA, Z E A, y =1= Z and 0 :( A :( I
imply A = 0 or A = I (thus x = y or x = z). This implies that x cannot be the
barycentre of a set of n points Xi of A carrying positive masses unless x is one of the
Xi; for this is just the definition when n = 2; for arbitrary n argue by induction
on n, as x is the barycentre of Xl and of the barycentre y 1 of the Xi with 2 :( i :( n,
therefore x is identical with Xl or Y1' and in the second case it is sufficient to apply
the induction hypothesis.
To say that x is an extremal point of A also means that A - {x} is convex.
Examples. - 1) In the space Rn, all the points of the sphere S" -1 are extremal points

i i
°
of the closed ball Bn" For, if I yf ,,; 1, I zf ,,; 1 and < 'A < 1, the relation

is possible only if

But this implies I (Yi - ZY = 0, thus Yi = Zi for all i, which proves our assertion.
i
2) In the nOfmed space .~(N) of bounded sequences of real numbers (1, p. 4) the
extremal points of the unit ball are the points x = (S,,) such that Is,,1 = 1 for all n. For,
suppose that we had Is,,1 ,,; 1 for all nand ISpi < 1 for one index p. We can then write

1 + Sp 1- Sp
x = --2-Y + --2-z
where Y (resp. z) is the point all of whose coordinates are equal to the coordinate of x
with the same index. except in the case of index p where the coordinate is equal to I
(resp. - 1). This shows that x is not extremal, since we have IIYII ,,; 1 and Ilzll ,,; L

°
Conversely, if Is" I = 1 for all n, then x is extremal, for the relation S" = 'All" + (1 - 'A) Sn
with 111,,1 ,,; 1, ISnl ,,; 1 and < 'A < I implies Sn = lln = S,,'
3) Let u: E --> E' be an affine mapping of an affine space E in an affine space E';
let C c E, C c E' be two convex sets such that u(C) c C. If x' is an extremal point
of C and x is an extremal point of u- 'ex') n C, then x is an extremal point of C, as it
follows from def 1.

PROPOSITION 1. - Let B be the set of extremal points of A, a non-empty compact

convex set in a Hausdorff locally convex space E, and let f be a convex function defined
in A and upper semi-continuous. Then f attains its upper bound in A at one point (at
least) ofB.
No. I EXTREMAL POINTS AND EXTREMAL GENERA TORS TVS II.55

Use IY to denote the family of subsets X of A that are non-empty, closed, and such
that every open segment that is contained in A and meets X necessarily lies in X. It
has the following properties;
(i) A belongs to IY.
(ii) A point a E A is such that {a} E IY, if, and only if, a is an extremal point of A.
(iii) Every non-empty intersection X of a family (X a ) of sets of IY also belongs
to IY.
The properties (i), (ii) and (iii) follow immediately from the definitions.
(iv) Let X E IY, and let h be a function that is convex and upper semi-cuntinuous
in A; then the set Y of the points of X where the restriction hlX attains its upper
bound in X is such that Y belongs to IY.
For, hlX being upper semi-continuous in X attains its upper bound cr over X
in at least one point of X (GT, IV, § 6.2, tho 3); thus Y is non-empty, it is also closed
(GT, IV, § 6.2, prop. 1). On the other hand let x, y be two distinct points of A and let
z = AX + (1 - A) y be a point of Y such that 0 < A < I; as Y c X and X E IY,
we have X E X and y EX; on the other hand, as h is convex, we have

h(z) ~ Ah(x) + (1 - A) hey)

but as hex) ~ cr, hey) ~ cr and h(z) = cr, of necessity hex) = hey) = cr, that is to
say x E Y and y E Y. Therefore Y E IY.
With these properties established, let M be the set of x E A where f attains its
upper bound in A; by (iv), M E IY. On the other hand, by (iii) and the fact that the
sets of IY are closed subsets ot the compact set A, it follows that IY is inductive for
the order relation :::J. By tho 2 ofS, III, § 2.4, M contains a subset N which is a minimal
element of IY. We shall show that N consists of a single point and this will complete
the proof of the proposition. Since E is a Hausdorff locally convex space, it is suffi-
cient to show that every continuous linear form u on E is constant in N (II, p. 38,
cor. 1). Now it follows from (iv) that the set N' of the x E N where ulN attains its upper
bound in N is such that N' belongs to IY; since N is minimal in IY we necessarily have
N'=N.
COROLLARY. - Let A be a compact convex set in a Hausdorff locally convex space E.
Then every closed support hyperplane H of A contains at least one extremal point ofA.
For, if f(x) = cr is an equation of Hand f(x) ~ cr in A, it is sufficient to apply
prop. I to f.

THEOREM I (Krein-Milman). - In a Hausdorff locally convex space E, every compact

convex set A is the closed convex envelope of the set of its extremal points.
For, let C be the closed convex envelope of the set of extremal points of A; clearly
C c A. To see that A c C, it is sufficient to prove that, if u is an affine linear func-
tion, continuous in E and if u(x) ): 0 in C then also u(x) ): 0 in A (II, p. 39, cor. 4);
but this follows from prop. I applied to - u.

PROPOSITION 2. - Let x be an extremal point of a compact convex set A in a Hausdorff

TVS 11.56 CONVEX SETS AND LOCALLY CONVEX SPACES §7

locally convex space E. Then for every open neighbourhood V of x in E, there exists
an open half-space F in E such that x E F n A c V n A (in other words, the traces
on A of the open half-spaces containing x, form a fundamental ,I}'stem of neighbour-
hoods of x in A).
For every open half-space D ofE containing x, the set AnD is a compact neigh-
bourhood of x in A, and the intersection of all these neighbourhoods is precisely
the point x (any two distinct points can be strictly separated by a closed hyperplane
(II, p. 38, prop. 4). By prop. 1 of GT, I, § 9.2, it is sufficient to prove that the sets
An D form afi/ter base. Now if we write Lo = An (E- D), the set Lo is convex,
compact and contained in the convex set A - {x} ; if D 1 , D2 are two open half-
spaces of E containing x, the convex envelope B of LOl U L02 is therefore contained
in A - {x} ; but B is a compact set (II, p. 14, prop. 15), therefore there exists a closed
hyperplane H that separates x strictly from B (II, p. 38, prop. 4) and if the open
half-space determined by H and containing x is D, then we have LOl U L02 C L o ,
therefore An D c (A n D 1 ) n (A n D2)'

COROLLARY.- In a HausdO/jj locally convex space let K be a compact subset of a

compact convex set A. Then the following conditions are equivalent.
a) A is the closed convex envelope of K.
b) K meets every set that is the intersection of A with one of its support hyperplanes.
c) K contains the set of extremal points of A.
a) => b). Suppose that there exists a support hyperplane H of A whose equation
is f(x) = rJ., such that (H n A) n K = 0 and suppose, for example, that f(x) ~ rJ.
in A. As f(x) - rJ. > 0 for all x E K by hypothesis and as K is compact we have

~ = inf f(x) > rJ.,

XEK

and K is, therefore, contained in the closed half-space f(x) ~ ~; therefore the
same is true of the closed convex envelope A of K and this is absurd.
b) => c). Suppose that an extremal point x of A does not belong to K; there is
a neighbourhood V of x in E such that V nAn K = 0. But by prop. 2, we can
suppose that V is an open half-space defined by a hyperplane H with the equa-
tion fez) = rJ.. If for example f(x) > rJ., then for all y E K, we have fey) :( rJ., there-
fore K does not meet the intersection of A and the support hyperplane fez) = Y > rJ.
parallel to H (II, p. 37, prop. 2); this is absurd.
c) => a). This is an obvious consequence of the Krein-Milman theorem.

Remarks. - 1) Even if the vector space E is finite dimensional thc set of extremal
points of a compact convex set is not necessarily closed (I I, p. 89, excrc. 11).
2) If K is a compact set in a non complete Hausdorfr locally convex space, and A,
the closed convex envelope of K is not compact, there can be extremal points of A
that do not belong to K (II, p. 87, exerc. 2).
3) In a Banach space E of infinite dimension, it may happen that the closed ball
of centre 0 and radius 1 does not possess any extremal point (II, p. 89, exerc. 14).
4) If A is a compact convex set in a Hausdorfflocally convex space, it may happen that
an extremal point of A does not belong to any support hyperplane of A (II, p. 78,
No.2 EXTREMAL POINTS AND EXTREMAL GENERA TORS TVS 11.57

exerc. 11). The proof of theorem 1 (II, p. 56) shows that in any case A is the convex closed
envelope of the set of extremal points of A which belong to a support hyperplane.

2. Extremal generators of convex cones

Let C be a convex cone with vertex 0 in a vector space E; clearly no other point
of C than the vertex can be an extremal point; the vertex is an extremal point of C
if and only if C is pointed and proper.

DEFINITION 2. - Let C be a convex cone of vertex 0 in a vector space E. We say that

a half-line DeC originating at 0 is an extremal generator of C, ifevery open segment
contained in C, not containing 0 and meeting D is contained in D.
It comes to the same thing to say that for all xED such that x #- 0, if y #- 0,
= y + y', then, it is necessarily the case that
y' #- 0 are two points of C such that x
y E D and y' E D.

°
Remark l. - Let C be a pointed proper convex cone in E, and consider on E the
order structure for which C is the set of elements ~ (II, p. 12, prop. 13); in order
that an element of E, say x > 0, belongs to an extremal generator of C, it is necessary

°
and sufficient that every element y ~ 0, that is bounded above by x, is of the form AX
with :s; A :s; 1 : in fact, to say that y is bounded above by x means that x = y + y'
where y' E C, whence the conclusion follows.

PROPOSITION 3. - In a vector space E, let C be a convex cone with vertex 0, and let
Xo #- 0 be a point of C, and D a half-line that is contained in C, originating jrom 0 and
containing xo' Let H be a hyperplane containing Xo and not passing through O. Then
D is an extremal generator of C if and only if Xo is an extremal point of H II C.
The condition is clearly necessary. Conversely, suppose that it is satisfied; sup-
pose that there is a line D' not containing D, passing through Xo and such that
D' II C contains an open segment to which Xo belongs. Let y#-O be a direction
vector of D'; the hypotheses imply that the point (1 + A)xo + ~y belongs to C
for 111,1 and I~I sufficiently small. But then, in the plane P determined by D and D'
and carrying the canonical topology, Xo is an interior point of P II C, and it follows
that the line P II H contains an open segment contained in H II C and to which
Xo belongs. This contradicts the hypothesis.

DEFINITION 3. - Let C be a convex set in a Hausdorff topological vector space E.

A compact convex non-empty set A oj C is called a cap oj C if the complement C - A
oj A in C is convex.
Let C be a pointed convex cone with vertex 0 in E and let A be a cap of C. Write
B =C- A. For every closed half-line L c C originating at 0, the sets L II A and
L B are convex sets that are complements in L, whose union is L, and such that
II
L II A is compact. As L II A is non-empty for at least one half-line L, we see that
o E A, thus L II A is a closed segment with an end point at O. If A exists then C is
proper.
TVS 11.58 CONVEX SETS AND LOCALLY CONVEX SPACES §7

PROPOSITION 4. - Let C be a pointed convex cone with vertex 0 in E, a Hausdorff

locally convex space.
a) Let A be a cap oj C. Let p be the restriction to C oj the gauge oj A (II, p. 20).
The set of the x E C such that p(x) ~ 1 is the set A. The junction p is lower semi-
continuous and has the jollowing properties :
(i) For any x, y in C, we have p(x + y) = p(x) + p(y).
(ii) For any x E C and A E R~, we have p(AX) = Ap(X).
(iii) Ij x E C. rhe relation p(x) = 0 is equivalent to x = O.
b) Conversely, let p be a junction dejined in C with values in (0, + 00), satisjying
the conditions (i), (ii) of a). Let A be the set oj the x E C such that p(x) ~ 1. Then
A and C - A are conv~x. A is a cap, if and only if A is compact and non-empty.
The statement b) is obvious. The properties stated in a) are consequences of the
remarks preceding prop. 4 and of the prop. 22 of II, p. 20 and prop. 23 of II, p. -20
with the exception of

p(x + y) ~ p(x) + p(y) .

It is sufficient to prove this last when x =F 0 and y =F 0; we have therefore p(x) > 0,
p(y) > O. Let /1, A be two numbers > 0 such that A < p(x), /1 < p(y), and denote
the complement of A in C by B. We have x E AB, y E !lB, therefore x + Y E AB + !lB;
by the convexity of B, we have AB + !lB C (A + Il)B, whence p(x + y) > 'k + 11,
which implies the inequality stated above.

COROLLARY I. - Let C be a painted convex cone of vertex 0 in E, a Hausdorff locally

convex space and let p be the gauge of A, a cap of C. The extremal points oj A are
then the point 0, and the points x on the extremal generators orC such that p(x) = 1.
It is clear that 0 is an extremal point of A. Let x be a point on L an extremal gene-
rator of C and such that p(x) = 1. Let y, z be two points of A such that x = i(y +z).
As L is extremal, we have y = AX and z = /1X, where A and /1 are numbers ~ 0
such that i(A + 11) = I, A = Ap(X) = p(y) ~ I and 11 = IlP(x) = p(z) ~ 1, from
which ' A = 11 = 1 and hence y = z = x; so, x is an extremal point of A. Conversely,
let x =I- 0 be an extremal point of A. Obviously p(x) = 1. Let y, y' be two points
ofC such that x = y + y', and we shall show thaty, y' are proportional to x. Without
loss of generality we can suppose that the numbers A = p(y) and A' = p(y') are
finite and> 0_ Then A-ly E A, A'-l y' E A, A + A' = 1 by prop. 4, (i) and the equa-
lity x = A(A - 1y) + A'(A' - 1y') implies, by hypothesis that

COROLLARY 2. - Every point oj C that belongs to a cap oj C, also belongs to the

convex closed envelope oj the union of the extremal generators of C.
This follows immediately from cor. 1 and the Krein-Milman theorem (II, p.55,
tho 1).
No.3 EXTREMAL POINTS AND EXTREMAL GENERA TORS TVS 11.59

* Example. - Let X be a locally compact space that is a-compact. Let e be

a closed convex cone of vertex 0 in J{ + (X) with the vague topology. We shall show
that e is the union of its caps. Let (X n ) be an increasing sequence of open, relatively
compact sets of X whose union is X. Let /.1 be an element =1= 0 ofe. There exist r:t.n > 0
such that L r:t.n/.1(Xn) = 1.

For every measure VEe, put p(V) = L r:t.nv(Xn) E (0, + 00). The function p
n
on e satisfies conditions (i) and (ii) of prop. 4. It is lower semi-continuous for the
vague topology (INT, IV, 2nd ed., § 1, No.1, prop. 4). The set A of the Y E e such
that p(y) ~ 1 is therefore closed and non-empty. On the other hand, every compact
set of X is contained in one of the X n , thus A being vaguely bounded is also vaguely
compact (INT, III, 2nd ed., § 1, No.9, prop. 15). The set A is therefore a cap of e
containing /.1. *
PROPOSITION 5. - Let e be a proper convex cone with vertex 0 in E, a Hausdorff
weak space; suppose that e is complete for the uniform structure induced by that oj E,
and that there is an enumerable fundamental system of neighbourhoods of 0 in e.
Then e is the union of its caps and is the closed convex envelope of the union of its
extremal generators.
The second statement follows from the first and from cor. 2 above. Using prop. 11
of II, p. 52 reduces the proposition to the case when E = RI and e c RI+. For all
r:t. E I, denote the projection pr~ in E by f~; then f~ is a continuous linear form. On
the other hand let (V n)nEN be an enumerable fundamental system of neighbourhoods
of 0 in e. By the definition of the topology of E, for each n EN, there exists a finite
subset I n of I and a number En > 0 such that Vn contains the set Wn of the x E e
such that f~(x) ~ En for all r:t. E I n ; put J = U I n . Let y =1= 0 be a point of e, and p
nEN
be the function L A~(f~IC) where the A~ > 0 are chosen so that p(y) = 1; this is
~EJ

possible, since if f~(y) = 0 for all r:t. E J, then y E Vn for all n, which implies that
y = 0, and this is contrary to hypothesis. Now we remark that for all r:t. E I, the
function f~le is continuous at the point 0, therefore there is an n E N, such that f~
is bounded in a W n , therefore bounded above in e by a linear combination of a
finite number of functions f~ Ie, where ~ E J. It follows that if A in the set of x E C
such that p(x) ~ 1, then f~ is bounded in A for all r:t. E I. As p is lower semi-continuous
in ~, it follows that A is closed and non-empty in e and therefore is compact. Since
it is clear that p verifies the conditions (i) and (ii) of prop. 4 of II, p. 58, we see that
A is a cap in e and contains y.

Remark 2. - There exist proper convex cones that are weakly complete and which
have no extremal generator (II, p. 92, exerc. 31).

3. Convex cones with compact sole

PROPOSITION 6. - Let E be a Hausdorff locally convex space and K a convex compact
set in E which does not contain O. Then the smallest pointed cone C of vertex 0 which
TVS II.60 CONVEX SETS AND LOCALLY CONVEX SPACES §8

contains K is a proper convex cone in E and is a locally compact and complete subspace
of E ; also, there exists a closed hyperplane H in E that does not contain 0 and is such
that H meets all the half-lines originating at 0 contained in C and such that H n C
is compact. Further, ifD is the half-space containing 0 determined by H, a closed hyper-
plane with these properties, then C n D is a cap ofC andC is the union of the A(C n D)
for. A > o.
By prop. 4 of II, p. 38, there exists a closed hyperplane H which separates 0 strictly
from K. Now, the convex envelope A of the union of {o} and of K is compact (II,
p. 14, prop. 15) and is the union of the AK with 0 ~ A ~ 1. As 0 and K are strictly
on opposite sides of H, for every x E K there exists A such that 0 < A < 1 and
AX E H. As C is the union of the AA for A )! 1, we see that H meets every half-line
originating at 0 contained in C and that H n A = H n C is compact. Further,

°
C is also the union of the A(H n C) for A )! 0; let C n be the union of the A(H n C)
for ~ A ~ n. Clearly C n is the convex envelope ofthe union of {O} and of n(H n C),
therefore it is compact. Also, for all x E E, there is a closed neighbourhood V of x
in E and an integer n such that V n C C C n ; in fact, if H is defined by the equation
fez) = r:J., where r:J. > 0, it is sufficient to take for V the closed half-space determined
by nH and containing 0, where n is so large that nr:J. > f(x). This shows that C is locally
compact (taking x E C), and that it is closed in E. We can also consider K as a subset
of the completion E, therefore C is also closed in E and therefore complete.
Given a cone C and a closed hyperplane H in a Hausdorff topological vector
space E, such that H does not contain the vertex s of C and C is the smallest cone
with vertex s containing H n C, then we call the intersection H n C a « sole» of
the con~ C. Prop. 6 shows that in a Hausdorff locally convex space E, the smallest
cone of vertex 0, containing a compact convex set K to which 0 does not belong, is
a cone of compact sole, and that every convex cone having a compact sole S, is locally
compact and complete.

Examples. ~ 1) Every proper closed convex cone in E, a vector space of finite dimen-
sion, has a compact sole. In fact, by II, p. 52, prop. II we need only consider the case
where E = Rn and C = R~. If (e)Hi';;n is the canonical basis of Rn, it is clear that the
compact convex set which is the convex envelope of the ei (1 ::;; i ::;; n) is a compact
sole for R~.
* 2) If X is a compact space, then the cone At + (X) of positive measures on X, with
the vague topology, is a cone with a compact sole (INT, III, 2nd ed., § 1, No.9, cor. 3
of prop. 15). *

§ 8. COMPLEX LOCALLY CONVEX SPACES

1. Topological vector spaces over C

Let E be a topological vector space over C the field of complex numbers; the
topology of E is also compatible with the structure of the vector space over R,
obtained by restricting the field of scalars to R. We denote by Eo the topological
No.1 COMPLEX LOCALLY CONVEX SPACES TVS 11.61

vector space on R which underlies E (I, p. 2). Note that, in Eo, the mapping x f-+ ix
(which is not a homothety) is an automorphism u of the topological vector space
structure of Eo such that u 2 (x) = - x.
Conversely, let F be a topological vector space over R, and suppose that there exists
an automorphism u of F such that u 2 ~ - IF (IF is the identity automorphism of F).
We know (A, IX, ~ 3, No.2) that it is then possible to define a vector space structure
on F relative to C writing AX = IXX + ~u(x) for all A = cr + i ~ EO C and all x EO F.
Further since the mapping (cr, ~, x) f--+ crx + ~u(x) of R2 x F in F is continuous the
topology of F is compatible with the vector space structure relative to C defined above;
if E denotes the topological vector space on C defined in this manner, then F is the
topological vector space on R which underlies E.
Remark. - Given a topological vector space F over R, it is not always the case that
there exists an automorphism u ofF whose square is - IF; for example, it is not possible
to define vector space structure relative to C on a vector space over R of finite odd
dim(lnsion.

Let E be a topological vector space on C, and Eo the topological vector space on R

which underlies E. Every linear variety M in E is also a linear variety in Eo, but the
converse is false. To avoid confusion we say that a linear variety for a vector space
structure relative to C (resp. relative to R) is a complex (resp. real) linear variety.
A complex linear variety of finite dimension n (resp. of finite codimension n) is a
real linear variety of dimension 2n (resp. of codimension 2n). In order that a real
vector subspace M of E should also be a complex vector subspace, it is necessary
and sufficient that iM c M.
Recall that, if E and F are two topological vector spaces on C, then a mapping
of E in F is called C-linear (resp. R-linear) if it is a linear mapping for the vector
space structures of E and of F relative to C (resp. R); every C-linear mapping is
evidently R-linear but the converse is false. We say that a C-linear form on E is a
complex linear form and that an R-linear form on E (i.e. a linear form on Eo) is a
real linear form. If f is a complex linear form on E, it is clear that the real part g = fYtf
and the imaginary part h = § f of f are real linear forms; further, the relation
f(ix) = ijex) implies the identity hex) = - g(ix); in other words we have

(1) f(x) = (Plf) (x) - i(fYtf) (ix) .

Conversely, if g is a real linear form on E, then f(x) = g(x) - ig(ix) is the unique
complex linear form on E such that fllf = g; and f is continuous if, and only if,
g is continuous.
Now let H be a complex hyperplane in E, with the equation f(x) = rJ.. + i~, where
f is a complex linear form on E; putting g fYtf, we see that H is the intersection
of two real hyperplanes HI' H 2 with equations respectively (l(x) = rJ.. andg(ix) = - ~;
if H is closed, so also are Hl and H2 (I, p. 13, tho 1). Conversely let Ho be a homo-
geneous real hyperplane, with equation g(x) = 0 (where g is a real linear form on E);
then H, the intersection of Ho and iHo, is a homogeneous complex hyperplane, and
if f is the complex linear form such that Pllj = g, then j(x) = 0 is the equation
of H; if Ho is closed then H also is closed.
TVS 11.62 CONVEX SETS AND LOCALLY CONVEX SPACES §s

Let G be a topological vector space over R and let G(C) be the vector space on
C obtained from G by extending the field of scalars to C (A, II, § 5.1). Identify G
as a subset ofG(C) by the mapping x I--> 1 ® x. The R-linear mapping (x, y) I--> X + i.y
is then a bijection of G x G on G(C), by means of which we transfer the product
topology of G x G to G(C). Then G(C) with this topology is a topological vector
space on C. We say that G(C) is the complexified topological vector space oj G.

2. Complex locally convex spaces

To say that a subset A of a complex vector space E is balanced means that, for
all x E A, we have px E A for 0 ~ p ~ 1 and ei8x E A for all real 11.
We say that a set A of E is convex if it is convex in the real space Eo which under-
lies E. In order that a convex set A =1= 0 ofE be balanced, it is sufficient that ei8A c A
for all real 11; for this implies firstly that - A = A; as A is convex, we see that 0
belongs to A and thus pA c A for 0 ~ p ~ I.
Let E be a complex topological vector space. The smallest balanced convex
(resp. closed balanced convex) set containing a set A of E is called the balanced
convex envelope (resp. balanced closed convex envelope) of A; the balanced closed
convex envelope of A is the closure of the balanced convex envelope of A. This last
is the convex envelope of the union of the sets ei9 A; we can therefore define it as the
set of linear sums L Aixi, when (x;) is any finite family of points of A, and (A;) a family
i
of complex numbers such that L IAil ~ I. If A is precompact so also is its balanced
i
envelope (I, p. 6, prop. 3).
We say that a complex topological vector space E is locally convex if the real
underlying topological vector space Eo is locally convex, that is to say if every neigh-
bourhood of 0 in E contains a convex neighbourhood of 0; a topology :Y on E
is locally convex if it is compatible with the vector space structure of E (relative
to C) and ifE, with topology:Y, is locally convex. As in this case every closed convex
neighbourhood V of 0 contains a balanced neighbourhood W of (I, p. 7, prop. 4),
we see that V also contains U, the balanced closed convex envelope of W; in other
words the balanced, closed, convex neighbourhoods of 0 form a fundamental system
of neighbourhoods of 0 in E, invariant under every homothety of ratio =1= O.
Conversely, let E be a complex vector space and let 6 be a filter base on E formed
by absorbent, balanced convex sets. We know then (II, p. 23, prop. 1) that the set \8,
of the transforms ofthe sets of 6 by homotheties of ratio> 0, is a fundamental system
of neighbourhoods of 0 for a locally convex topology :Y on the real vector space
Eo underlying E. Further, as the sets of \8 are balanced, they are invariant under every
homothety x I--> eillx, which shows that :Y is compatible with the vector space
structure of E (over C) (I, p. 7, prop. 4).
Every locally convex topology on a complex vector space E can be defined by
a set of semi-norms, for the gauge of an open balanced convex neighbourhood of 0
is a semi-norm on E.
No.3 COMPLEX LOCALLY CONVEX SPACES TVS II.63

The ideas and results for real locally convex spaces detailed in II, p. 25 to 36, extend
to complex locally convex spaces with no modification other than the replacement
of symmetric convex sets by balanced convex sets.
A complex locally convex space is a Frechet space if it is metrisable and complete.

3. The Hahn-Banach theorem and its applications

THEOREM 1 (Hahn-Banach). - Let V be a vector subspace of E, a complex vector

space, and let f be a (complex) linear form on V and p a semi-norm on E such that
If(y)1 ~ p(y) for all y E V. Then there exists a linear form fl on E extending f and
such that Ifl(x) I ~ p(x) for all x E E.
For g = fJ1tf is a real linear form defined in V and satisfying l·y)1 :( p(y) at
every point of V; therefore there exists a real linear form gl in E extending g and
such that Igl(X)1 :( p(x) for all x E E (II, p.23, cor. 1). Let f 1(x)=gl(x)-ig 1(ix)
be the complex linear form on E of which gl is the real part (II, p. 61). For all real if

sincep is a semi-norm on the complex space E; this implies the relation Ifl (x) I :( p(x),
and the theorem is proved.

COROLLARY 1. - Let Xo be a point of a complex topological vector space E and p

be a continuous semi-norm in E ; then there exists a continuous (complex) linear form f
defined in E, such that f(x o) = p(x o) and If(x) I :( p(x) for all x E E.

CoROLLARY 2. - Let V be a vector subspace of a complex locally convex space E

and f be a (complex) linear form defined and continuous in V; then there exists a conti-
nuous linear form fl defined in E and extending f If E is l10rmed there exists such a
form fl that also satisfies II fIll = II f II.

COROLLARY 3. - Let M be afinite dimensional vector subspace of a Hausdorff complex

locally convex space E. Then there exists a closed vector subspace N of E that is a
topological complement ofM in E.
The proofs using theorem I, p. 24 are the same as those of II, p. 23, cor. 2 and cor. 3,
p. 24, prop. 2 and p. 25, cor. 2.

PROPOSITION 1. - Let A be an open non-empty convex set in a complex topological

vector space E and M be a non-empty (complex) linear variety that does not meet A.
Then there exists a closed complex hyperplane H that contains M and does not meet A.
We can suppose that 0 E M. Then there exists a closed real hyperplane Ho con-
taining M and not meeting A (II, p. 36, tho 1). As M = iM, the closed complex
hyperplane H = Ho n (iHo) has the properties required.

COROLLARY. - In a complex locally convex space E, every closed complex linear

variety M is the intersection of the closed complex hyperplanes which contain it.
TVS 11.64 CONVEX SETS AND LOCALLY CONVEX SPACES §8

In fact, for all x rt= M, there exists a convex open neighbourhood V of x that does
not meet M, and thus there exists a closed complex hyperplane H containing M
and not meeting V; a fortiori H does not contain x.

PROPOSITION 2. - Let A be a non-empty balanced open convex set of a complex topo-

logical vector space E, and B be a non-empty convex set that does not meet A. Then
there exists a continuous complex linear form f on E and a number rx > 0 such that
If(x)1 < rx in A and If(y)1 ~ rx in B.
For, there exists a continuous real linear form g on E and a real number rx such
that g(x) < rx in A and g(y) ~ rx in B (II, p. 37, prop. I). As 0 E A, we have rx > o.
We show that the continuous complex linear form f(x) = g(x) - ig(ix) and the
number rx have tile properties required. For, since f7lf = g, we have If(y) ~ rx in B.
1

On the other hand, for all x E A and all real 3, the point eiSx belongs to A, since A
is balanced, and we have f(x) = e-i:Jf(ei!)x); then there exists a number 9 such that
IfCx)1 = f7l(e i:1f(x)) = g(ei:fx) < rx, and the proposition follows.
PROPOSITION 3. - Let A be a balanced, closed, convex set in a complex locally convex
space E and let K be a non-empty compact convex set in E that does not meet A. Then
there exists a continuous complex linear form f on E and a number rx > 0 such that
If(x)1 < rx in A and IfCy)1 > rx in K.
The proposition follows from II, p. 38, prop. 4 as prop. 2 follows from II, p. 37,
prop. 1.

4. Weak topologies on complex vector spaces

The definition and results ofII, § 6, Nos. I and 2 apply without change to complex
vector spaces. If F and G are two complex vector spaces in duality by a bilinear
form B, then the underlying spaces Fo and Go are in duality by f7lB, and it follows
from II, p.61, formula (1) that the weak topologies cr(F, G) and cr(F o, Go) are
identical.

DEFINITION 1. - Let F and G be two complex vector spaces in duality. For any sub-
set M of F, the polar of M in G, denoted by MO, is the set of y E G such that
f7l«x,y») ~ - I for all XE M.
If MO is the polar of M c F in G then (AM)O = A-1 MO for all A E C*.
If M is a (complex) vector subspace of E, then MO is a closed vector subspace
<
(for cr(G, F)), since the relation f7l(A x, y») ~ - I for every scalar A E C implies
< x, y) = 0; again we say that MO is the subspace of G orthogonal to M.
If M is a balanced set in F, then MO is a balanced set in G; in this case MO is the
set of y E G such that 1< x, y) 1 :( I for all x EM; for this relation is equivalent to
f7l«1;;x,y»):( 1 for all XEM and all1;;EC such that 11;;1 = 1.
The results of II, p. 41 to 51 are also valid without restriction for complex vector
spaces.
 Exercises

§ 2

I) A subset A of a vector space E, is starshaped relative to 0 if for all x EO A and every A such
that 0 :::; A < I, the point AX belongs to A. Let A be starshaped and such that, for each x EO A,
there exists 11 > I such that Ilx EO A. Show that if, for every pair of points x, y of A we have
t(x + y) EO A, then A is convex. Give an example of a non-convex starshaped set A such that
z(A + A) cA.

2) Let A be a convex subset of an affine space E and B a set containing A. Show that, amongst
the convex sets that both contain A and are contained in B there exists at least one maximal
set; give an example where there are several distinct maximal sets.

~ 3) Let A and B be two disjoint convex sets in a vector space E. Show that there exist two
disjoint convex sets C, D in E such that A c C, BcD and CuD = E. (Apply tho 2 ofS, III,
§ 2 .4 to the set of pairs of disjoint convex sets (M, N) such that A c M and BeN and express
the fact that M and N do not meet by the relation 0 rt M - N. To show that CuD = E,
obtain a contradiction supposing that Xo rt CuD; if C' (resp. D') is the convex envelope of
C u {xo }(resp. D u {x o }), show that it is impossible that both C' n D =F 0 and C n D' =F 0.)

4) Let C be a convex cone with vertex 0 in a vector space E; if (x;) 1 <:;i<:;n is a finite family of
n
points of C such that I AiX i = 0 for a family of numbers Ai > 0, then C contains the vector
j = 1

subspace of E generated by the Xi.

5) Suppose that the vector space E has an enumerably infinite basis (en)nEN. Let C be the set
of points X = I ~nen such that for the largest index n for which ~n =F 0, we have ~n > O. Show
n
that C is a pointed convex cone such that C n (- C) = {O} and C u (- C) = E; deduce
that C is the set of elements of E that are ;::, 0 for an order structure that is compatible with
the vector space structure of E and for which E is linearly ordered. Show that on this ordered
vector space the only linear positive form is o.
TVS II.66 CONVEX SETS AND LOCALLY CONVEX SPACES §2

6) Let E be an affine space of dimension ~ 2 and let f be a bijection of E on itself; show that
if the image under f of every convex subset of E is also a convex set, then f is an affine linear
mapping (consider the inverse mapping of f and note that a closed segment is the intersection
of the convex sets which contain its extremities cf A, II, § 9, exerc. 7).

7) Give an example of two convex sets A c R, B C Rl, such that the image of the convex set
A x B under the bilinear mapping (A, x) '--> AX ofR X Rl in R2 is not convex.

8) Let(A)l ~ i~p be a finite family of convex subsets of a vector space E; let Wi be the subspace
obtained by a translation from thc affine linear variety generated by Ai (1 ~ i ~ p). If
W = f
i= 1
Wi' show that the affine linear variety generated by the convex set f
i= 1
AiAi (where
Ai are non-zero numbers) is obtained from W by a translation.

~ 9) Let A be a subset of the space R".

a) Show that the convex envelope of A is identical with the set of points t
i=O
AiX i , where Xi E A,

Ai ~ 0 for 0 ~ i < n, and I" Ai = 1. (Establish the following lemma; if p + 1 points Xi

i=O
p
(0 ~ i ~ p) form an affinely dependent system (that is to say there exists a relation I f3iXi =0
i=O

where the f3i are not all zero and f

i=O
f3i = 0) and if x = f
i=O
(XiXi, where the (Xi are ~ 0 and
p
I (Xi = 1, then there exists an index k ~ p and p numbers Yi (0 ~ i ~ p, i 1= k) such that
i=Q

Yi ~ 0 for all i, I Yi = 1 and x = I YiXi; for this compare those of the IXJf3i that are defined.)
i::!=k i*k
b) Let a be a point of the convex envelope of A which does not belong to the convex envelope
of any subset of A with at most n points. Show then. that A contains at least n + 1 connected
components. (We can suppose that a = O;let(b)o~i~"beafamilyofn + I affinelyindependent
points of A such that 0 belongs to the convex envelope of the bi (cf a)). For each index i, let
C i be the pointed convex cone of vertex 0 generated by the bj with indices.i 1= i; show that A
does not meet the frontier of any of the cones - C i .)
c) If C is a pointed cone with vertex 0 in R", show that the convex envelope of C is the set of
points I" Xi' where Xi E C for I ~ i ~ n.
i= 1

~ 10) Let C be the convex envelope of a subset A of R", and let a be an interior point of C.
Show that there exist 2n points Xi E A (l ~ i ~ 2n) such that a is interior to Co, the convex
envelope of the Xi' (Suppose a = 0, and argue by induction on n, noting that by exerc. 9a)
there exists a ~et of k .+ 1 .points Yj of A (0 ~.i ~ k, I ~ k ~ n), affinely indel?endent and
such that, if V IS the affine linear varIety generated by the Y" then 0 E V and, relatIve to V, 0 is
interior to the convex envelope of the set of the Yj' Then pioject C on E/V and show that 0 is
interior to this projection relative to E/V). Show that in the above statement 2n cannot be
replaced by 2n - 1.

11) a) Show that, in the space R", every convex set A of dimension n contains at least one
interior point (consider an affine1y independent system of n + I points of A). Deduce that
if A is everywhere dense in R" then A = R".
b) Let E be the normed space /1 (N) of absolutely convergent series of real numbers X = (1;.)
(I, p. 4); show that the set P of x, such that 1;. ~ 0 for every index n, is a proper convex cone,
which generates E but does not contain any interior point.
c) Let E be a Hausdorff topological vector space on which there exists a non-continuous
linear form f (cf II, p. 86, exerc. 17, a)). Show that the sets A and B defined by the relations
§2 EXERCISES TVS 11.67

f(x) ;;, 0, f(x) < 0 are convex, non-empty, complementary, everywhere dense and that
each of them generates E (algebraically).

12) Show that in the space R", a necessary and sufficient condition that a convex set should be
closed, is that its intersection with every straight line should be closed (cf II, p. 74, exerc. 5).

13) Show that in the space R", every non-empty open convex set is homeomorphic to R"
(use exerc. 12 of GT, VI, ~ 2).

14) Let A be a non-empty closed convex subset of E, a Hausdorff topological vector space.
a) Show that, for every a E A, the set n A(A - a) is a closed convex cone in E, with vertex 0,
1>0
independent of a. It is called the asymptotic cone of A and written CA- For every a E A, the
set a + C A is the union of {a} and those open half lines that are contained in A and have a as
an end point.
b) If x, yare two points of A such that (x + C A) n (y + C A) is a cone whose vertex Z E A,
then this cone is necessarily z + CA'
c) IfB is a second closed convex subset ofE such that An B =1= 0, then CAe, R = C A nCR'
d) Let V A be the largest vector subspace (necessarily closed in E) which is contained in CA'
Show that if <l> is the canonical homomorphism of E on ElVA' then A = <l> -l(Ao), where Ao is
a closed convex set in ElVA which does not contain any straight line.
e) In the Banach space .'?l(N) of bounded mappings of N in R (I. p. 4) give an example of a
closed convex non-bounded set A, for which C A = {O} and which is such that for every b =1= 0
in E, there exists an integer k for which (A + kb) n A = 0.

15) a) Let A be a closed convex subset of a Hausdorff topological vector space E. Iffor some
point Xo E A there exists a neighbourhood V of Xo in E such that V n A is compact, then
show that A is locally compact. Deduce that the closure in E of a locally compact convex set
is locally compact.
b) Let A be a closed convex set that is locally compact but not compact in E; show that the
asymptotic cone (exerc. 14) is not the single point {O }.

~ 16) Let A, B be two closed convex subsets of a Hausdorff topological vector space E.
Suppose further that B is locally compact and that C A nCB = {O}. Show that A - B is
closed in E. (Let b E B, and W be a closed neighbourhood of 0 in E such that B n (b + W) is
compact. Let c E A - B; for every neighbourhood V of 0 in E consider the set Mv of those
y E B such that An (c + y + V) =1= 0. Consider two cases according as to whether there
exists a V for which Mv is relatively compact, or there does not exist such a V; in the second
case, consider the filter base formed from the sets p v " = Mv n C(b + nW) where V varies in
the set of closed neighbourhood of 0 in E and n varies in N; form the cone with vertex b
generated by p v ," and its intersection with the frontier of b + W).

~ 17) In a Hausdorff topological vector space E, a closed convex set A is said to be parabolic
if, for every z ¢c A, each half-line originating at z and contained in z + C A meets A.
a) Give an example of a parabolic convex set A in R2 such that C Ais not just a single half-line.
b) Let A be a closed convex set in E such that C A =1= {O}, but such that A is not parabolic.
Show that if z ¢c A is such that z + C A contains a half-line D with end point z which does not
meet A, then neither the convex envelope of A u {z} nor the pointed cone with vertex z
generated by A is closed in E.
Further if D' is the closed half-line originating at z and opposite to D (so that D = 2z - D')
then D' + A is not closed in E, and there exists a plane P containing D and a closed convex
set B c P, such that B n A = 0 but that the distance of B from P n A (in any norm on P)
is zero.
c) In E, let A be a closed convex set that is locally compact and parabolic; show that if BeE,
is closed and convex then A - B is closed (same method as exerc. 16).
d) Let A, A' be two closed convex subsets of E that are locally compact and parabolic; show
that the convex envelope of A u A' is closed in E (same method as in exerc. 16). Give an example
TVS II.68 CONVEX SETS AND LOCALLY CONVEX SPACES §2

in R Z where A is parabolic, A' is non-parabolic and the convex envelope of A u A' is not closed
in RZ.
e) In E, let A be a closed convex set that is locally compact and parabolic; show that for all
z rf; A, the pointed convex cone with vertex z, generated by A, is closed in E (usc dl).
1) Let E 1 , E z be two Hausdorff topological spaces, Al (resp. Az) a closed convex set in El
(resp. E z) that is also parabolic. Show that the set Al x A z is parabolic in El x E z .
* g) Show that a barrelled space of infinite dimension does not contain a parabolic closed
convex set that is both locally compact and not compact.
h) Let Eo = [2(N), and in Eo, let K be the set of points x = (U such that I~nl ~ l/(n + 1)
for all n; K is compact. The cone E of vertex 0 generated by K is a vector subspace of Eo and
K is absorbent in E. Let p be the gauge of K in E; it is a lower semi-continuous function. In
the normed product space E xR, show that the set A of points (x, ~) such that ~ ~ (p(x)j2
is closed, convex, parabolic and locally compact. Show that A + (- A) and the convex enve-
lope of A u (- A) are not locally compact. *

18) Let C be a proper closed convex cone of vertex 0 in Rn. Show that the complement of the
set C n Sn-l on the sphere Sn-1 is homeomorphic to R n- 1 (make a stereographic projection
from a point of C n Sn- 10 and use exerc. 12 of GT, VI, § 2). If C contains an interior point,
show that C n Sn-l' is homeomorphic to the closed ball B n - 1 (same method).

19) a) Let A be an unbounded closed convex set in Rn, that does not contain any line, but
does contain an interior point. Show that the frontier of A is homeomorphic to R n- 1 (use
exercs. 15, b) and 18).
b) In a Hausdorlrtopological vector space E, let A be a closed convex set that does not contain
any line and is of dimension ~ 2. Show that the frontier of A is connected (use a) and GT, VI,
~ 2. exerc. 12).

20) a) In a vector space E, let A be a convex set that generates E and meets every straight line
in a set that is closed relative to the straight line. Show that the following conditions are equi-
valent:
ct:) There exists a line D such that D meets A in a compact segment that is not empty.
~) There exists a line D such that every line parallel to D meets A in a compact segment.
y) A is distinct from E and is not a half-space determined by a hyperplane ofE.
(To show that (y) => (ct:) use exerc. 14, d) of II, p. 67, and reduce to the case E = R2.)
b) In a Hausdorff topological vector space E, let A be a closed convex set which contains an
interior point. Show that if the frontier of A is a non-empty linear variety. then A is a closed
half-space (usc exerc. 14. J) of ll, p. 67, to show that the frontier of A is necessarily a hyperplane,
then apply a)).

~ 21) a) Let Ai (1 ~ i ~ r) r > n + I, be a family of convex subsets of R n such that any

r - 1 of the Ai have a non-empty intersection; show that the r sets Ai have a non-cmpty inter-
section (Belly's theorem). (Let Xi be a point of the intersection of the Aj with indexes j oft i;
there exist r numbers Ai which are not all zero and are such that :t Ai
i= 1
= 0 and t
i= 1
AiX i = 0;
in this last equation take to one side those terms with Ai ~ 0 and to the other those with
Ai < 0.)
b) Given a family of compact convex sets in Rn. show that the intersection of all the sets of
the family is non-empty if the intersection of any selection of n + I sets of the family is non-
empty.
c) In Rn, let K be a convex set and (AJl'; i';, be a family of r > n + 1 convex sets. Suppose that
for every selection of n + 1 indices (ik ) each less than or equal to r there exists a E Rn such that
a + K contains each of the A ik . Show that then there exists bERn such that b + K contains
all the Ai' Show that similar results hold if « contains» is replaced by « is contained in» or
by « meets in a non-empty set ». (For each index i, consider the set Ci of the x E R n for which
x + K ~ Ai (or x + K c Ai' or (x + K) n Ai oft 0»). Generalize to any family of compact
convex sets of Rn.
§2 EXERCISES TVS II.69

22) In R2 consider a set of 2m points of the form (a p b;), (a p bi) where b; ~ b;' for 1 ~ i ~ m.
Let n be an integer < m - 2. In order that there should exist a polynomial P(x) of degree ~ n
such that b; ~ pea;) ~ b;' for 1 ~ i ~ m, it is sufficient that, for every family (ik)1~k~n+2 of
n + 2 indices i, there exists a polynomial Q(x) of degree ~ n such that b;k ~ Q(a i ) ~ b;~ for
every integer k such that 1 ~ k ~ n + 2. (Use exerc. 21, a).)
23) Show that in a topological vector space, the convex envelope of an open set is an open set.

24) Let M be an everywhere dense convex set in a topological vector space E (el II, p. 66, exerc.
II, e)); show that, for every closed hyperplane H in E, the set H n M is dense in H (for every
point Xo E H, and every balanced neighbourhood V of 0 in E, consider the intersections of
Xo + V and the two open half-spaces determined by H, and deduce that Xo + V + V meets
H n M).
25) a) Show that, in a topological vector space, every convex set with an interior point, is
such that its frontier is nowhere dense (use prop. 16 of II, p. 14).
b) In a Hausdorff topological vector space E, let A be a closed convex set with an interior
point, and let H be a closed hyperplane that contains an interior point of A. Show that the
intersection of H and of the frontier F of A is a set which is nowhere dense relative to F (to
show that in every neighbourhood of a point ofH n F there exist points ofF not in H, reduce to
the case when E is of dimension 2).
~ 26) In a Hausdorff topological vector space E, let A be a connected closed set with the
following property: for every x E A, there exists a closed neighbourhood V of x in E such that
V n A is convex. Show that A is convex. For this establish the following statements.
a) Show that any two points of A can be joined by a broken line in A (same method as GT, VI,
~ 1, exerc. 6).
b) Show that, if two points in A can be joined by a broken line in A with n > 1 segments, then
they can also be joined in A by a broken line with n - 1 segments. (Induction on n reduces to
the case n = 2 which is equivalent to taking R2 as E; then let T be a triangle with vertices
a, b, e such that the closed segments ae, be are contained in A, but the closed segment ab is
not; consider a point of the closure of the intersection of GA and the interior of T that is
farthest from the line ab, and show that the existence of such a point contradicts the hypo-
thesis.)
~ 27) a) Let B be a non-empty closed convex subset of E, a Hausdorff topological vector
space, and let X be a non-empty compact set in E. Show that if A is a subset of E such that
A + X c B + X, then A c B (if a E A, consider a sequence (x n ) of points of X defined induc-
tively by the relation a + Xn = b n + x n + l' where bn E B). Deduce that, if A, B are two non-
empty subsets of E, using the distance in E and the proceedure of GT, IX, ~ 2, exerc. 6. Show
A + X = B + X implies the relation A = B.
b) Let E be a normed space, cr the distance function defined on the set ~(E) of closed non-
empty subsets ofE, using the distance in E and the proceedure ofTG, IX, p. 91, exerc. 6. Show
that if A, B, C are three non-empty compact convex sets in E then cr(A + C, B + C) = cr(A, B)
(if S, is the ball defined by Ilxll ~ A, note that A + S, and B + S, are closed convex sets and
use a)).
e) Deduce from a) and b) that the setst(E) of non-empty, compact, convex subsets of a normed
space E, with the distance cr, can be identified with a cone in a normed space of which the laws
of composition induce on st (E) the laws (A, B) --> A + B and (A, A) --> AA.

28) Let f be a convex function defined over the convex subset A of a vector space E.
a) Show that if A is absorbent and f is non-constant then f cannot attain its upper bound
in A at the point O.
b) Show that the subset of points of A, at which f attains its lower bound in A, is convex.

~ 29) Let E be a Hausdorff topological vector space, and C be a non-empty open convex
non-pointed cone with vertex 0, in E. A convex neighbourhood of 0 in E is denoted by V. If f
is a convex function that is defined and bounded above in C n V, show that f(x) tends to a
finite limit as x tends to 0, where x E C n V. (Let ~ = lim. sup f(x); obtain a contradiction,
x-+ O,XECn V
TVS II.70 CONVEX SETS AND LOCALLY CONVEX SPACES §2

supposing that for some a > 0, and every neighbourhood W of 0, there exists a point

f(pa) ~ ~ - 1a for °
Y E C n V n W such that fey) < ~ - a. Show that there exists a E C n V such that
< p ,,; 1; deduce that, on a line joining a point of the form pa (p suffi-
ciently small) to a point yin C n V such that fey) < ~ - a, and that is sufficiently close to 0,
there exist points of C n V where f is arbitrarily large.)
30) a) Give an example of a convex function that is defined over a compact convex subset

°
K of R2, that is bounded and lower semi-continuous in K, but is not continuous at a point
of the frontier of K (consider the gauge of a disc of which is a frontier point).
b) Deduce from a) an example of a convex function defined in an open half-plane D of R2,
not bounded above in D and not tending to a limit at a frontier point of D.
e) Deduce from a) an example of a convex lower semi-continuous function defined over a
compact convex subset A of R2 but not bounded above in A. (Take for A the set of points
(~, 11) such that ~4 ,,; 11 ,,; 1 in R2.)

"IT 31) Let Xo be a point in the closure of A, a non-empty convex subset of a Hausdorff topolo-
gical vector space E. Let f be a convex function defined over A. Use :n to denote the set of
closed half lines D which originate at x o, for which AnD (fontains an open segment with
end point Xo' The union C of the half lines D E :n is a convex cone with vertex Xo'
a) Show that, for each fixed D E :n, as x tends to Xo such that xED n A and x =1= x o , either
f(x) tends to a finite limit or to + 00.
b) Let ~ be the subset of those D E :n, for which the limit of f(x) in a) is + 00; if Xo E A then
~ is empty. Show that ~ cannot contain two opposite half lines; if D and D' are two distinct
half lines in ~ and P is the plane determined by D and D' then, either, every half line D" of :n
in P belongs to~, or D and D' are the only two half lines of~ lying in P. Deduce that if ~ =1= :n,
then no half line D E ~ contains an internal point (II, p. 26) of the cone C relative to the vector
subspace generated by C.
e) Let 0: be the set of half lines in :n that are not in ~. Show that the union of the half lines
of 0: is a convex cone, and for each half line D E 0: the limit of f(x) defined in a) is independent
of D (use exerc. 29 above); further if Xo E A this limit is ,,; f(x o), and it is equal to f(x o) when
!y contains two opposite half lines.
d) Let f be a non-continuous linear form over E (el 11, p. 86, exerc. 17, a)) and take A = E;
show that every closed half line, originating at x o ' belongs to 0:, but that
lim.inf f(x) = - 00 and lim. sup f(x) = + 00
X-+Xo

(use prop. 21 of II, p. 18).

32) Let K be a compact convex set in a Hausdorff topological vector space E and let f be
an upper semi-continuous convex function defined over K. Show that f is bounded over K.
(Observe first that f is bounded above in K; if f is not bounded below show that
lim. inf fey) = - 00 for every point x E K, and that this contradicts Baire's theorem.) Give
y-x,y-=f=.x
an example where f is not continuous.
33) Let E be a finite dimensional Hausdorff topological vector space, and let K be a compact
convex subset of E. Show that every convex function defined over K is bounded below in K
(compare exerc. 31, d).

34) Let U, V be two open convex sets in a Hausdorff topological vector space E such that
V c U and that U does not contain any half line. Let g; be a set of convex functions defined in
D, uniformly bounded above on the frontier ofU and uniformly bounded below on the frontier
of V. Show that :F is equicontinuous.

35) Let U be a non-empty open convex set in R n and :F be a set of convex functions defined
over U. Let <I> be a filter on :F that converges pointwise in U to a finite function fo; show
that <I> converges uniformly to fo in every compact subset of U (use exerc. 34).
§2 EXERCISES TVS 11.71

36) Let A be a compact convex set in Rn and B its projection on the subspace R,,-I (identified
as the hyperplane with equation ~" = 0). Show that there exist two convex functions II' I2
defined over B, such that A is identical with the set of points (x, 1;) of R" where x E B, Y E R
and fl (x) 0( 1; 0( - fz(x).
37) Lct E be a vector space; in order that a convex set F of E x R should be formed of pairs
(x, 1;) such that f(x)0( 1; (resp. f(x) < 1;) for a convex function I defined over a convex subset
X of E, it is necessary and sufficient that the projection of F on E should be identical with X
and that, for all x E X, the set F(x) of F that projects onto x should be a closed (resp. open)
interval unlimited to the right (i.e. not bounded above).
38) Let X be a convex set of an affine space E and p an affine linear mapping of E in a second
affine linear space E I . Write XI = p(X). For every real-valued function f defined in X and
every XI E XI' let
fl (XI) = inf l(x).
P(X)=Xl

Show that if I is convex and if fl (Xl) > - w for all XI E XI' then II is a convex function.
39) Let E be a finite dimensional Hausdorfr topological vector space.
a) Let \j(E) be the family of closed non-empty sets ofE, carrying the uniform structure deduced
from the uniform structure of E by the proceedure of GT, II. § 1, exerc. 5, a). Show that, the set
(f(E) of non-empty closed convex sets of E, is closed in the space \j(E). Deduce that if K is a
compact set in E, the set of non-empty closed convex sets in E that are contained in K, is a
compact set in (f(E) (ef GT, § 4, exerc. 11).
h) Let Ra(E) be the set of compact convex subsets of E that contain 0 as an interior point.
F or every set A E .Ra(E), let p A be the gauge of A (II. p. 20). Show that A f--+ PAis an isomor-
phism of the uniform subspace .Ra(E) of (f(E) on a subspace of the space (fo,(E; R) of continuous
real valued functions in E, carrying the uniform structure of compact convergence (GT, X,
§ 1. 6).

40) In a topological vector spacc E, let U be a convex neighbourhood of Xo and let f be a

real-valued continuous convex function in U. Show that there exists a convex neighbourhood
V c U of Xo and a convex continuous function j~ in E such that fliV = fiV.
~ 41) Let H be a hyperplane in a vector space E that does not contain 0 and let S be a convex
set contained in H.
a) Suppose that the intersection of S with each line in H is a compact segment. Let a, b be
two distinct points of E such that there exist two numbers A > 0, ~l > 0 for which
b + IlS c a + AS; show that if e is the point where the line joining a and h meets the hyper-
plane H' parallel to H which contains a + AS, then C E b + IlS and b + IlS is the image of
a + AS by a homothety of centre e transforming a into b. (Reduce to the case where E is of
dimension 2.)
b) With the same hypotheses on S, let a, b be two distinct points of E and suppose that there
exists a point e E E and three numbers A> 0, ~l > 0, v~ 0 such that (a + AS) n (b + IlS) = c +vS.
Show that if A (resp. B) is the cone with vertex a (resp. b) generated by a + AS (resp. b + IlS),
and H" the hyperplane parallel to H passing through e, then H" nAn B = {c} (use a».
c) Suppose that H is the affine linear variety generated by S. Let C be the cone with vertex 0
generated by S. Show that the following two conditions are equivalent:
IX) E is a lattice for the order on E of which C is the set of elements ~ O.
~) For any points x, y of E and numbers ), > 0,11 > 0 such that the set (x + AS) n (y + IlS)
is not empty, there exists Z E E and v ~ 0 such that this set is z + vS.
(To prove that IX) implies ~), reduce to the case y = 0 and use the fact that if (s) is a finite
family of points ofS and O,J is a family of real numbers such that 2.: AjSj = 0, then 2.: Ai = O.
i
To prove that ~) implies IX) use b»).
When S satisfies the equivalent conditions IX) and ~), we say that S is a simplex in E. When
E is of finite dimension, the convex envelope of a finite set of points affinely independent in H
and generating H is a simplex * (the converse is also true: cf INT. n, 2nd ed., § 2. exerc. 7»). *
TVS 11.72 CONVEX SETS AND LOCALLY CONVEX SPACES §3

42) Generalise the prop. 18 ofII, p. 16 to the case of an ordered set E with a Hausdorff topology
for which the intervals (a, -+ (and) <-, a) are closed for all a EO E.

43) Let K be a complete valued division ring of which the absolute value is an ultrametric.
In a left vector space E on K, we say that a set A is ultraconvex if the relations x EO A, y EO A,
1"'1 ~ I, 1111 ~ 1 imply Ax + I1Y EO A.
a) Generalize the prop. I, 2, 5, 6, 7 of II, p. 8 and p. 9. Show that the smallest ultraconvex
set containing a given set M is the set of linear combinations I "',x" where x, EO M and 1\1 ~ 1
for alit.
b) Suppose that E is a topological vector space over K. Show that the closure of an ultra-
convex set is ultraconvex, and that an ultraconvex set with a non-empty interior is open.
c) Let A be an absorbent and ultraconvex set in E. Show that if, for all x EO E, we put
p(x) = inf Ipl, then p is an ultra-semi-norm on E (II, p.2). Generalize prop. 23 of II, p. 20,
xepA

to the case where the absolute value of K is obtained from a discrete valuation.

§ 3

1) Let P be a proper pointed convex cone, with vertex 0, in E a vector space over Rand p
be a semi-norm on E and V the set of points x EO E for which p(x) < 1. Let M be a vector sub-
space of E and f be a linear form on M. There exists a linear form g on E, which extends f
and is such that it is ;::, 0 in P and Ig(x)1 ~ p(x) for all x EO E, if and only if, for all
x EO M n (V + P), we have f(x) > - 1. (To see that the condition is sufficient consider a
point Xo EO M such that, f(x o) = I, the cone Q of vertex 0 generated by Xo + V, and apply the
cor. of the prop. 1 (II, p. 21) to the space E carrying the relation of preorder for which P + Q
is the cone of elements ;::, 0.)
2) For a set S let F = ~(S) be the Banach space of the real-valued bounded functions in S
(1, p. 4) and let M be a vector subspace of a normed space E. Show that, for every continuous
linear mapping f of M in F, there exists a continuous linear mapping g of E in F, that is an
extension of f and such that Ilg I = I ill.

3) Let E be a vector space over R and let p be a sub linear function on E (II, p. 20). Let A be
a convex set such that inf p(y) > - CfJ.
YEA
a) Show that the function
q(x) = inf (p(x + tz) - t.inf p(y))
ZEA,t;;:: 0 yeA

is a sublinear function on E such that - p( - x) ~ q(x) ~ p(x).

b) Show that there exists a linear form h on E such that hex) ~ p(x) III E and that
inf p(y) = inf hey) (take h such that hex) ~ q(X)).
yeA yeA

4) Let A be a non-empty set of E, a vector space over R, and p a sub-linear function on E. Let
B be the set of z EO E such that inf p(x - z) ~ 0; we have A c Band inf p(x) ~ p(z) for all
xeA xeA
Z EO B; from which inf p(x) = inf p(z).
xeA zeB
a) Show that the set of the yEO E such that inf p(z - y) ~ 0 is the set B.
ZEB

b) Deduce from a) that the intersection of B and any affine line D in E is closed in D (show
that, whatever the points a, b of E, the function t-> pea + tb) is continuous in R).
c) Suppose that for each pair of points x, y of A there exists z EO A such that p(z - t(x + y)) ~ O.
Show that, for each pair of points u, v of B we have t(u + v) EO B (write
z - t(u + v) = (z - t(x + y)) + t(x - u) + t(y - v)
for x, y, z in A). Deduce that B is then convex (use b)).
§3 EXERCISES TVS II.73

d) Under the hypotheses of c), show that there exists a linear form h on E such that hex) ~ p(x)
and that we have inf p(y) = inf hey) (use c) and exerc. 3).
YEA YEA

5) Let A be a non-empty subset of a vector space E over R and let p be a sub linear function
on E. Suppose that, for every pair of points x, y of A, there exists z E A such thatp(z - (x + y») ~ 0
and that p(x) ~ 0 for all x E A. Show that there exists a linear form h on E such that hex) ~ p(x)
in E and that hex) ~ 0 for x E A. (Apply exerc. 4, c) to the union of the.!. A for n (integer)
n
~ 1.)

6) a) Let H be a hyperplane in E a vector space over R and let p be a sublinear function on E.

Let f be a linear form on H, such that fey) ~ p(y) in H. Let a be a point of GH, and let h
be the linear form on E which extends f and is such that h(a) = inf (f(y + pea - y»). Then
YEH
hex) ~ p(x) in E. Show that for every linear form g on E extending f and such that g(x) ~ p(x)
in E, we also have g(a) ~ h(a).
b) Let V be a vector subspace of E and f a linear form on V such that fey) ~ p(y) in V. Let
S be a non-empty set of E. Show that there exists a linear form h on E that extends f, such that
hex) ~ p(x) in E and that there is no other linear form g on E extending f such that g(x) ~ p(x)
in E that is distinct from h and such that g(x) ~ hex) in S. (Consider the set 0: of pairs (V', F)
where V' is a vector subspace containing V and F a linear form on V' extending f and such that
F(z) ~ p(z) in V' and further such that there is no other linear form f" on V' with the same
properties and such that f"(z) ~ F(z) in S II V'. Order 0: and use a) and tho 2 of S, III, § 2.4.)

7) Let T be a commutative monoid CA, I, § 2.1) carrying a preorder relation x ~ y such that
if x ~ y then x + z ~ y + z for all z E T. A mapping f of T in R u { - oo} is called additive
(resp. subadditive, resp. superadditive) if we have
f(x + y) = f(x) + fey) (resp. f(x + y) ~ f(x) + f(y), resp. f(x + y) ~ f(x) + fey»)
for any x, y in T.
a) If g is subadditive and increasing in T, then the function hex) = inf g(nx)/n is sub additive
n>O
and increasing; we have h ~ g and h(O) = 0 if g(O) ~ O.
b) Under the same hypotheses suppose that there exist two elements x" X z of T and two
real numbers ~p ~z such that ~, < g(x,), ~z < g(x z ) and g(x, + xz) < ~, + ~z. Let y"
yz, z, ' Zz be four elements ofT, let n" n z be two integers ~ 0 and let 0:1' O:z be two real numbers
such that
n,~, + g(z,) < 0:1' y, ~ nix, + z,
nz~z + g(zz) < O:z, yz ~ nzx z + Zz .

Show that then g(n 2 y, + n,yz) < nzO:, + n,0:2'

c) Let co be a superadditive function on T such that coCO) = 0, and let Q be an increasing
sub additive function on T such that co(x) ~ Q(x) in T. Show that there exists an increasing
additive function f on T such that co(x) ~ f(x) ~ Q(x) in T. (Remark that the set of increasing
subadditive functions g on T such that co(x) ~ g(x) ~ Q(x) in T is non-empty and inductive
for the relation ~, and take a minimal element of this set for f; show using a) that
f(O) = O. To show that there cannot exist pairs of elements of T, (x" xz) such that
f(x, + x 2) < f(x,) + f(x z ) remark that if ~j E R, and hlx) = inf(n~j + fey») where n
varies in the set of integers ~ 0 and y in the set of elements of T such that x ~ nXj + y, then
h j is increasing and subadditive in T (j = 1, 2), h/x) ~ ~j and h/x) ~ f(x) for all x E T.
Then use the definition of f and part b) to obtain a contradiction.)

~ * 8) a) Let K be a non-discrete valued division ring of which the absolute value is an

ultrametric, non linearly compact (cf CA, III, § 2, exerc. 15); then there exists a well ordered set I
of numbers > 0 and a family (B(P»)PEI of closed balls in K such that the relation P < p'implies
B(p) c B(p'), that B(p) has radius p and that the intersection of the B(p) is empty (CA, VI, § 5,
exerc. 5). For every x E K, there exists pEl such that x rt B(p); show that the number q,(x) =
TVS II.74 CONVEX SETS AND LOCALLY CONVEX SPACES §4

Ix - yl for ayE BCp) depends neither on y E BCp), nor on pEl such that x rt B(p). If pEl is such
thatxEB(p)then<p(x) ~ p.. Thisbeingso,for(xl,xz)EKZ,p~tll(xl,xz)11 = IXllifx~ = O,and
II(x l , xz)11 = IXzl <p(X 2 I XI ) If Xz =1= O. Show that II(x l , xz)IIIS an ultranorm on K (I, p. 26,
exerc. 12) and show that there does not exist any projection of norm I of K z on K x {O}.
b) Let K be a complete non-discrete valued division ring of which the absolute value is an
ultrametric and which is linearly compact. Let E be a vector space of dimension 2 on K with
an ultranorm and let D be a line in E; show that for all points x E E, there exists y E D such
that d(x, D) = d(x, y) = I x - y II (note that the intersection of D and of a ball of centre x
is a ball in D).
c) Deduce from a) and b) that for a complete non-discrete valued division ring K, of which the
absolute value is an ultrametric, the following properties are equivalent:
a) K is linearly compact.
~) For every ultranormed vector space E on K, for every vector subspace F of E and every
continuous linear form I on F there exists a continuous linear form g on E that extends I and
is such that Ilg I = II I II. (Reduce to the case where E is of dimension 2 and use b).) *

§ 4

1) Let E be a vector space and A a convex symmetric convex subset of E. Let :T, :T' be two
locally convex topologies on E and OU, OU' be the uniform structures defined by :T, :T' on E.
In order that the uniform structure induced on A by OU' should be finer than that induced by OU,
it is necessary and sufficient that every neighbourhood of 0 for the topology induced on A by:T
should be a neighbourhood of 0 for the topology induced on A by :T '.

2) a) Give an example of a non-compact closed set in R2, whose convex envelope is not closed.
b) Show that, in Rn, the convex envelope of a compact set is compact (cf II, p. 66, exerc. 9, a»).

-If 3) Let I be the compact interval [0, I] ofR and F be the vector space '(5'(I, R) of continuous
real valued functions defined in I. Let E be the product space RF; for all a E I, let Ca be the
element of E such that ca(f) = I(a) for all I E F.

f
a) Show that, when x varies in I, the set K formed by the Cx is compact in E.

b) Let A be an element of E such that A(f) = I(t) dt for all IE F (Lebesgue measure).

Show that, in E, A belongs to the closure of the convex envelope of K but does not belong to
this convex envelope (cf FVR, II, p. 7, prop. 5).

4) With the notations ofIL p. 72, exerc. I suppose also that the space E is locally convex.
a) There exists a positive continuous linear form gin E that extends 1, if and only if I is bounded
below in M n (W + P) for at least one neighbourhood W of 0 in E.
b) Given a point x E E, there exists a positive continuous linear form g in E such that g(x) = I,
if, and only if, - x rt p.

5) a) Let E be an infinite dimensional normed space and :T be its topology. Show that there
exists on E a normed space topology :T' that is strictly finer than that of:T and a normed space
topology:T n that is strictly coarser than that of:T (define the neighbourhoods of 0 for these
topologies, using a basis of E put in the form (aa,n) where a varies in an infinite set of indices A
and n in the set of integers;;. 0 and where Ilaa,nll = 1 for the given norm on E).
b) Let p be the norm defining the topology :T'. Show that, if E is complete for the topology
:T, then p cannot be lower semi-continuous in E for the topology :T (use Baire's tho cf III,
p. 25, corollary). Deduce that the convex set A defined by the relation p(x) < 1 does not
contain any interior point for:T even though all its points are internal.
c) Deduce from b), that, if E is complete for the topology :T, then there exists in E convex
sets which are not closed for :T, of which the intersection with every linear variety of finite
dimension is closed for:T (d II, p. 67, exerc. 12).
§4 EXERCISES TVS II.75

6) Let E be a vector space with its finest locally convex topology.

a) Show that every vector subspace of E is closed, and that, if M, N are two subspaces that
are vectorial complements in E, then E is the direct topological sum of M and N. If (e)tEI is a
basis for E, then E is the direct topological sum of the subspaces Ret.
b) Let F be a locally convex space whose topology is also the finest locally convex topology.
Show that every linear mapping of E in F is a strict morphism.

7) a) Let A be a convex set with at least one interior point in a topological vector space E.
Show that the set of internal points of A is identical with the interior of A (cf exerc. 5, b».
b) Show that in the normed space E = [1 (N), the convex cone P defined in II, p. 66, exerc. II , b),
generates E but does not contain any internal point.

8) Let E be a vector space with an enumerable basis and with the finest locally convex topology.
Show that, if A is a set in E whose intersection with every vector subspace of finite dimension
is closed in E, than A is closed in E (et: exerc. 5, e».
~ 9) Let E and F be two vector spaces each with its finest locally convex topology.
a) Show that ifE and F each have an enumerable basis then every bilinear mapping ofE x F
in a locally convex space G is continuous (use Du Bois-Reymond's tho FVR, V, p. 53, exerc. 8»).
b) If one ofthe spaces E, F has a basis with cardinal equal to that of the continuum. show that
there exists a non-continuous bilinear form in E x F. (Reduce to the case where E = R(N),
F = RN, so that F can be identified with E* and the bilinear forms on E x F correspond
bijectively with the linear mappings of E* in itself; then consider the identity mapping of E*,
and note that in RN, a compact set for the product topology cannot be absorbent.)

10) Let (En) be an infinite sequence of locally convex spaces and let E be the topological direct
sum of the family (EJ Show that the topology of E is identical with the topology :70 defined
in I, p. 24, exerc. 14.

II) Let I be an infinite non-enumerable set. On the vector space E = R(I), show that the finest
locally convex topology is distinct from the topology :70 defined in I, p. 24, exerc. 14; for this
prove that the set of the x = (~) E E such that II ~t I < I, is open in :7 but not in :70,
tEl
12) Let E be a vector space with an enumerable basis (en)' Let V be the balanced convex
envelope of the set of the en and let W be the balanced convex envelope of the set of points
an = en + (n - I) e 1 (n ~ I).

Let :71 (resp. :72) be the locally convex topology on E for which a fundamental system of
neighbourhoods of 0 is formed by the AV (resp. AW) for A > O. Show that :71 and :72 are
Hausdorff, but that the lower bound of :71 and :72 in the set oflocally convex topologies on E
is not Hausdortf(cf II, p. 80, exerc. 26).

13) WithLhe hypotheses of II, § 6, show that E is complete for topology:7 which is the inductive
limit of the :7n , ifand only if, for each integer n and every Cauchy filter \Y on En for the topology
induced by :7, there exists p ~ n such that \Y is convergent in Ep for the topology :7p •

14) Let E be the strict inductive limit of an increasing sequence of locally convex spaces En
(II, p. 33). Show that the topology of E is the finest of the topologies compatible with the vector
space structure ofE, whether locally convex or not, and inducing on En a coarser topology than
the given topology :7n. (Let V o be a neighbourhood of 0 for such a topology :7 and (Vn)n ;'0' a
sequence of neighbourhoods of 0 for:7 such that Vn+ 1 + Vn+ 1 C Vn for all n ~ 0; for all
n ~ I, consider, in En' a convex neighbourhood Wn of 0 that is contained in En n Vn, and
take the convex envelope of the union of the Wn in E.)

15) Let I be an infinite non-enumerable set. Let \y(I) be the family of finite subsets of 1 and E
the direct sum space R(I). For every J E \y(I), let F J be the subspace R J of E, the product of the
factors whose indices belong to J, with the product topology; let gJ be the canonical injection
TVS 11.76 CONVEX SETS AND LOCALLY CONVEX SPACES §5

of F J in E. Show that there exists a topology :Yo on E that is compatible with the vector space
structure of E, which make the gJ continuous and which is strictly finer than the finest locally
convex topology :Y which makes the 9 J continuous. {Note that the set V of the x = (~J'EI
in E such that I 1~,11/2 ~ I is a neighbourhood of 0 for a topology compatible with the vector
lEI
space structure of E, not containing any absorbent, symmetric, convex set.)

16) a) Let E be a vector space with an enumerable basis. Show that the finest locally convex
topology on E is the finest of the topologies on E (compatible or not with the vector space
structure of E) which induces the canonical topology on every finite dimensional subspace
of E.
b) Let Eo be an infinite dimensional Banach space. Let E be the vector space that is the direct
sum of Eo ofR(N) and let Ep be the subspace ofE that is the direct sum of Eo and of RP (identified
as the product of the first p factors of R(N) ; we give to Ep the product topology of those of its
factor, so that the topology ofEp is induced by that ofEp+ l ' Show that on E the inductive limit
topology of those of the Ep is not the finest of the topologies (compatible or not with the vector-
space structure of E) which induces on each Ep a coarser topology than that of Ep. We can
proceed as follows :
Ct.) Let q be a norm on Eo which defines a topology strictly coarser than that of Eo (II. p. 74,
exerc. 5). For every I: > 0 define a mapping 1. of Eo in R+ by the relation 1.(x) = sup( q(x), I: --, II xii).
Show that 1. is continuous and > 0 in Eo and that inf 1.(x) = o.
Ilx II~'
~)Let U be the subset of E formed by the (x, (tn)) such that tn < j;/n(x) for all n. Show that
U n Ep is open in Ep for all p.
y) Show that if V c U is an absorbent convex set, then V n Eo cannot contain any ball
with centre 0 in Eo.

17) For a subset A of a commutative group G, written additively, and for each n > 0 denote
n n
the set of elements of the form I xi' where Xi E A for all i by + A. We say that the set A
i= 1
n
of G is convex if, for every integer n > 0 the relation nx E + A implies x EA.
a) Show that if a commutative topological group G (written additively) is isomorphic to
a subgroup of the additive group of a locally convex vector space (with the induced topology)
then there exists a fundamental system of symmetric convex neighbourhoods of 0 in G.
b) Conversely, let G be a Hausdorff topological commutative group (written additively) in
which there exists a fundamental system, ~, of symmetric convex neighbourhoods of O. Show
that G is without torsion, and, hence, can be considered (algebraically) as a subgroup of the
additive group of a vector space on the field Q (A, II, § 7. 10, cor. I to prop. 26). For every set
V E m, let V be the set of elements rx where x E V and r varies in the set of rational numbers
such that 0 ~ r ~ 1; show that V is symmetric and convex (in the sense defined above).
Deduce, further, that if there is no open subgroup of G distinct from G itself, then the sets V
form a fundamental system of neighbourhoods of 0 for a topology compatible with the vector
space structure of E on Q (Q being given its usual topology); conclude that in this case G is
isomorphic to a subgroup of the additive group of a Hausdorff locally convex space.
c) Let G be the group R x R ordered lexicographically (A, VI, p. 7); consider the Hausdorff
topology :Yo(G) on G that is compatible with its group structure (GT, IV, § I, exerc. I). Show
that for this topology there exists a fundamental system of symmetric convex neighbourhoods
of 0, but that G is not isomorphic to any subgroup of the additive group of a Hausdorff topo-
logical vector space over R.

§ 5
I) a) Let E be a vector space. We say that a pointed convex cone C (of vertex 0) in E is maximal
if C is a maximal element of the set of convex pointed cones of vertex 0 and oft E, ordered by
inclusion. Show that a pointed convex cone C is maximal if, and only if, it is a closed half-space
defined by a hyperplane which passes through O. To establish this result, prove successively the
following properties of a maximal pointed convex cone C;
§5 EXERCISES TVS lI.77

0:) We have C u (- C) = E (argue by obtaining a contradiction).

~) If z is a non-internal (II, p. 26) point of C then - Z E C (same method). Deduce that C con-
tains internal points.
y) The largest vector subspace H = C n (- C) contained in C is a hyperplane. (Passing to
the quotient space F = EjH, this reduces to demonstrating, using ~) that if all the points of C
other than the vertex are internal, then E is necessarily of dimension 1.)
b) Give an example of a maximal non-pointed convex cone (in the set of non-pointed convex
cones of vertex 0) which has no internal point (ef. II, p. 65, exerc. 5).

2) Let N be a hyperplane in a vector subspace M of a vector space E and let A be a convex

set in E, such that all the points of A n M are on the same side of N and which also possesses
the following property; for any y =F 0 in E, there exists x E A n M such that x + Ay E A
for all Asuch that IAI is sufficiently small. Show that there exists then, a hyperplane H ofE such
that all the points of A are on the same side of H and such that H n M = N. (Reduce to the
case N = {O}; if a =F 0 belongs to A n M, consider the set U of pointed convex cones with
vertex 0 containing A and not containing - a; show that there exists a maximal element C
of U and that C is a maximal pointed convex cone (exerc. 1»). Deduce a new proof of the Hahn-
Banach theorem.

3) Let A be a convex set in a topological vector space E and Xo be a point of E. Then, there
exists a closed hyperplane H, containing x o , and such that all the points of A lie on the same
side of H if and only if there exists a non-pointed convex cone C with vertex x o, which contains
at least one interior point and does not meet A. (F or an example of a convex set A =F E which
is not contained in any half-space defined by a hyperplane, see II, p. 65, exerc. 5.)

.,-r 4) Let E be a normed space and A a complete convex set for the uniform structure induced by
that of E.
a) Let x' be a continuous linear form on E that is bounded in A. Consider a number k > 0
and the closed convex cone P in E, with vertex 0 and formed by the x E E such that
<
Ilxll <:;; k x, x' ); it is pointed and proper. Show that for the order on E for which P is the set
of elements ~ 0, the set A is inductive (use the fact that the restriction of x' to A is increasing
and bounded).
b) Deduce from a) that the set of points of the frontier F of A which belong to a support hyper-
plane of A is dense in F (Bishop-Phelps th.). (For each point Z E F, consider a point y E CA
arbitrarily close to z, and separate y strictly from A by a closed hyperplane of equation
<x, x') = 0:, with Ilx'll = 1 and use a) with k > 1, also exerc. 3 above.)

5) Let A be a closed convex set in R" and Xo be a point of CA; denote the euclidean distance
in R" by d.
a) Show, without using tho I of II, p. 36, that there exists one and only one point x E A such
that d(x o , x) = d(x o , A), and that the hyperplane orthogonal to the line joining Xo and x,
and passing through x is a support hyperplane of A.
b) Deduce from a) a new proof of tho I of II, p. 36 when the space E is finite dimensional.
(Reduce to the case when M is a frontier point Xo of A ; note that the lower bound of the distance
of Xo from support hyperplanes of A, is zero, and use the compactness of S" - 1')

~ 6) Let A be a closed set in R" with the following properties; for every x E R", there exists
one and only one point YEA such that d(x, y) = d(x, A), where d is the euclidean distance.
Show that A is convex. (Argue by reductio ad absurdum, considering a closed segment with
end points a, b in A containing a point c E CA ; there is a closed ball B of centre c contained
in CA; consider the set mof closed balls S which contain B and whose interiors do not meet
A; show that the radii of these balls is bounded above, and deduce that there exists one of
these balls So whose radius p is the largest possible. Then get a contradiction by proving that
So can only meet A in a single point, and that this implies the existence in m of a ball
of radius> p.)
7) In a Hausdorff locally convex space E, let A be a complete, convex set and let B be a pre-
compact closed convex set such that A n B = 0. Show that there exists a closed hyperplane
TVS II.78 CONVEX SETS AND LOCALLY CONVEX SPACES §5

separating A from B (argue in the completion E). Consider the case when A is finite dimen-
sional.

points, such that C A n C B = { °}

8) In a Hausdorff locally convex space, let A and B be two closed convex sets, without common
(II, p. 67, exerc. 14), and such that B is locally compact. Show
that there exists a closed hyperplane separating A from B (er II, p. 67, exerc. 16). Similarly, if

°
A and B are two closed convex cones of vertex 0, such that A n B = {o} and B is locally
compact, then there exists a closed hyperplane passing through and separating A from B
(use lemma 1 of II, p. 39).

°
9) Deduce from exerc. 8 that if V is a finite dimensional vector subspace of E and C is a closed
convex cone of vertex in E such that C n V = {O}, then there exists a support hyperplane
of C that contains V (use lemma 1 of II, p. 39).

° °
10) In the normed space E = 11(N) of summable sequences of real numbers x = (~")nEN'
let D be the line defined by the relations ~n = for n :;;, 1. Show that there exist two increasing
sequences (cxn ), (~,,) of real numbers > such that the convex set A defined by the inequalities
~o :;;, Icx"~,, - ~"I for n :;;, 1 is closed, non-bounded does not meet D and that there is no closed
hyperplane separating A from D (choose cxn and ~" so that A - D is everywhere dense).

* 11) a) Let E be a Hilbert space and F an everywhere dense subspace of the dual E' of E
that is distinct from E'; the unit ball B of E is compact for the weak topology cr(E, F), and
there exists a point a of the unit sphere through which passes no closed (in cr(E, F)) support
hyperplane of B.
b) Give E the topology cr(E, F) and consider, in the product space G = E x R the set A of
pairs (x, 1:;) such that Ilxll < 1, I:; :;;, Ilxll/(1 - Ilxll). Show that A is closed and locally compact,
but that if D is the line with equation x = a in G then DnA = 0 and there does not exist
any closed hyperplane in G separating A from D.
e) Show that, when we give E the topology cr(E, F), there exists a continuous affine real valued
function in the subspace B of E, that is not the restriction to B of a continuous affine function
in E. *
12) Consider in R 3 , the closed convex cone C defined by the relations ~1 :;;, 0, ~2 :;;, 0,
~~
is no plane through the origin °
:(; ~1 ~2' Show that the line D of equations ~1 = 0, ~3 = 1 does not meet C, but that there
containing D and not meeting C-{O}.

13) Let A and B be two closed convex sets in the space R", such that if V and W are affine
linear varieties generated by A and B respectively, then no point of A n B is both interior
to A relative to V and interior to B relative to W. Show that there exists a hyperplane separating
A from B. (By taking quotients, reduce it to the case where either one of the varieties V, W is
contained in the other or V and Ware complementary vector subspaces in E.)
14) Let A be a parabolic closed convex set (II, p. 67, exerc. 17) not containing a line. Show that
if B is a closed convex set not meeting A then there exists a hyperplane in Rn that separates
A strictly from B (if d is the Euclidean distance prove that d(A, B) > 0) (ef exerc. 12.)

15) Let S, T two finite sets in R" with no common points, and such that Card(S u T) :;;, n + 2.
In order that there exists a hyperplane separating S strictly from T, it is necessary and suffi-
cient that for every finite set F c S u T of n + 2 points, there exists a hyperplane separating
F n S strictly from F n T (use Helly's tho (II, p. 68, exerc. 21)). Show that in this statement we
cannot replace the number n + 2 by n + 1, and that the statement does not extend to the
case where Sand T are infinite.

16) Let A be a compact set with interior points in R". Show that if each frontier point of A
lies on at least one support hyperplane of A, then A is convex. (Obtain a contradiction, showing
that if x and y are two points of A such that the segment with end points x, y is not contained
in A, and if z is an interior point of A not situated on this segment, then there exists a frontier
point of A, distinct from x and y in the triangle with vertices x, y, z.)
§5 EXERCISES TVS II.79

17) In R", let A be a symmetric convex set of which 0 is an interior point and of which the
frontier does not contain any genuine segment. Let H be a homogeneous hyperplane and D
a line complementary to H. Show that there exists a point a E H II A such that at a there is
a hyperplane of support to A that is parallel to D.

18) In a topological vector space E, let Ai (1 :S; i :S; n) be n open non-empty convex sets.
a) Show that if the union of the A; is distinct from E, then every point x E E not belonging
to any of the Ai' belongs to a closed linear variety of codimension n, that contains x and does
not meet any of the A; (argue by induction on n).
b) If the intersection of the Ai is empty, show that there exists, in E, a closed linear variety of
codimension n - 1 that does not meet any of the Ai (same method).

19) Let C, C be two closed convex sets in a Hausdorff topological vector space E that are
strictly separated by a closed hyperplane H. Let H' be a closed hyperplane of support to
both C and C such that C and C lie on the same side of H'. Show that H' is the only hyper-
plane with these properties which contains H II H' and that H II H' is a support hyperplane
of the trace P on H of the convex envelope of cue. Conversely, if C and C are compact,
then for every support hyperplane D of P in H, there exists a hyperplane H' that supports both
C and C', which contains D and such that C and C lie on the same side of H'.

~ 20) Let A, B be two disjoint closed convex sets in a Hausdorff locally convex space E and
let H be a closed hyperplane separating A from B; suppose that A II H =1= 0 and that the
intersection of A II H and of every line is compact. Show that, if A or B is locally compact,
then there exists a neighbourhood V of 0 in E such that (A + V) II B is empty. (Consider
two cases according to whether A or B is locally compact; in the first case, note that there
exists a hyperplane H' parallel to H such that, if S is the set of points between Hand H', then
A II S is compact. In the second case, suppose for example that 0 E B II H; for every neigh-
bourhood V of 0 in E, consider the set (A + V) II B and consider successively the case where
this set is relatively compact for at least one V or the case when this is not so, as in exerc. 16
of II, p. 67.)

~ 21) a) In R" let a i (1 :S; i :S; n + 1) be n + 1 points that are afTinely independent. Denote
the convex envelope of the ai by S and the convex envelope of the ai with i =1= k by Fk
for 1 :S; k :S; n + 1. For each k let C k be a compact convex set containing F k' and suppose that
"+1
S is contained in the union of the C k ; show then that n Ck =1= 0. (Argue by reductio ad
k~1

absurdum and induction on n, considering the intersection C~+ 1 of the C i with indices i :S; n
and supposing that C" + 1 II C~ + 1 = 0, which would allow the strict separation of the two
convex sets by a hyperplane.)
b) Let X be a compact convex set in a Hausdorff topological vector space E, and (C)J'EL a
family of compact convex sets contained in X, such that for every set H c L having n (resp. m)
elements, the intersection (resp. the union) of the C, with indices A E H is not empty (resp. is
equal to X). Show that if m :S; n + 1 the intersection n C, is not empty. (This is effectively

proving that for any finite set H of p :;, m indices of L, we have n C, =1= 0. Argue by induc-
tion on p assuming that the result has been proved for p - 1 indices. Argue then by reductio
ad absurdum, considering for each index i E H a point a i E n
C,' and showing by the

aid of Helly's tho (II, p. 68, exerc. 21) that the a; generate a linear variety of dimension p - I.
then finally apply a).)

~ 22) In Rn let (CJl <S:i<S:", be a finite family of closed convex cones with vertex 0, such that
the sum of any n of them is distinct from R". Show that there exists a hyperplane H, passing
through 0, such that, for any index i no pair of points of C i are strictly separated by H. (Dis-
tinguish two cases :
TVS 11.80 CONVEX SETS AND LOCALLY CONVEX SPACES §5

a) Either there exists a number r < nand r indices, say, 1, 2, ... , r such that Ci for i ~ r
generates a cone which contains a vector subspace V of dimension ): r. Argue by induction
on n, projecting on the orthogonal to V.
~) Or, for all r < n, any r of the C i generate a cone C such that the maximal vector subspace
C n (- C) contained in C is of dimension < r. Consider then a set of the cones C i , maximal
with respect to being contained in a half-space; there are at least 11 cones in a maximal set. Let
r be the cone generated by the union of the cones belonging to this maximal set. If C j is a cone
which does not belong to the maximal set considered, show that C i C - r. For this, argue
by reductio ad absurdum, showing that in the contrary case there exists a frontier point of
- r (relative to the vector subspace generated by r) that is interior to C. (relative to the vector
subspace generated by C). Write such a point as the sum of the least dumber s of vectors, of
which each belongs to a cone - C i , among those C i used in defining r ; then s ~ n - 1. Prove
finally that these cones and C. generate a convex cone containing an s + 1 dimensional vector
subspace, contradicting the Iiypothesis; for this use exerc. 4 of II, p. 65.)
23) Let E be a topological vector space, and let /7' be the locally convex topology on E that
is the finest of all those that are coarser than the given topology flo on E. IfF is a locally convex
space, then the continuous linear mappings of E in F are the same for flo as for fl. There
exists a continuous linear form on E that is distinct from the null form if, and only if, there
exists a neighbourhood of 0 for .'7n whose convex envelope is not everywhere dense (for flo)
(cf I, p. 25, exerc. 4).

24) Let E be an infinite dimensional, metrisable, locally convex space.

a) Show that there exists a sequence (an) of points ofE tending to 0 and a decreasing sequence
(L n), of closed vector subspaces of E, such that Ln is of codimension n in E and that, for all n,
the point an belongs to Ln - Ln + 1 •
b) Suppose further that E is complete. Show that we can then find sequences (an) and (Ln)
verifying the conditions a), and such that in addition, for every bounded sequence of real
number (An), the series, whose general term is Anan, is commutatively convergent in E, and that
the linear mapping (~n) H L ~nan of the Banach space ~ (N) in E is injective and continuous.
c) Deduce from b) that when E is an infinite dimensional Frechet space then every basis of
E on R has cardinal at least equal to 2Card (N) (el I, p. 22, exerc. 5).
If there exists an enumerable set that is dense in E, then every basis of E has the cardinal
of the continuum.

25) Let E be an infinite dimensional Frechet space of enumerable type (therefore having an
enumerable everywhere dense subset) (cr I, p. 25, exerc. 1). Show that there exists an everywhere
dense hyperplane H of E which meets every closed, infinite dimensional linear variety of E.
(Use the existence of a basis, having the cardinal of the continuum, in each of the direction
subspaces of these varieties (exerc. 24, c» and the fact that the set of closed, infinite dimensional,
linear varieties of E also has the cardinal of the continuum (GT, IX, § 5, exerc. 17); then apply
a method of construction of a linear form on E following from S, III, § 6, exerc. 24.) The hyper-
nlane H does not contain any infinite dimensional, closed, vector subspace.

'IT 26) Let E be an infinite dimensional Frechet space of enumerable type.

a) Show that there exists a sequence (a.) of linearly independent elements of E such that
each sequence (a 2 .) and (a 2n + 1) is total (use exerc. 24, c».
b) Let F be the vector subspace of E generated by the a 2n + 1 (n EN). For every n > 0 let Mn
be the subspace generated by the a 2k with k ~ n. For each n, let <Pn be the restriction to F of
the canonical homomorphism of E on ElM., and let fl. be the topology on F which is the
inverse image under <Pn of the quotient topology, on ElM •. Show that each of the topolo-
gies fI" on F is a Hausdorff locally convex topology, but that the lower bound of the fI" in the
set of locally convex topologies on F is the coarsest topology on F.
* c) Take E to be a Hilbert space; show that we can choose the sequence (a,) so that if G
is the closed vector subspace generated by the a 4 .+ 1 , then G has infinite codimension and
so that the images of the a 2 • and the a 4n + 3 in E/G are still linearly independent. Write G. for
the subspace of E that is the sum of G and of the subspace generated by the a 4 k+ 3 with k ~ n,
§6 EXERCISES TVS 11.81

and give to G n the topology which is the inverse image under the canonical mapping res-
tricted to G n, of the quotient topology on E/Mn' Show that the sequence (G n) is an inductive
system of topological vector spaces such that G n is closed in G n+ 1 for the topology of G n+ 10
but that G n is not closed in the inductive limit space of this sequence. *

~ 27) Let E, F be two Hausdorff topological vector spaces, and X (resp. Y) a compact convex
set in E (resp. F). Let f be a real valued function defined in X x Y with the following pro-
perties :
(i) For all x E X, the mapping y H f(x, y) is lower semi-continuous in Y, and for all c E R,
the set of the y E Y such that f(x, y) ~ c is convex.
(ii) For all y E Y, the mapping x H f(x, y) is upper semi-continuous in X, and for all
c E R, the set of the x E X such that f(x, y) ~ c is convex.
Show that, in these conditions, we have
sup(inf f(x, y)) = inf(sup f(x, y)) .
XEX YEY yeY xeX

(Argue by reductio ad absurdum, supposing that there exists a number c such that
sup(inf f(x, y)) < c < inf(sup f(x, y)) .
xeX yeY yeY xeX

For all x E X (resp. all y E Y) let Ax be the set of y E Y such that f(x, y) > C (resp. B the set of
the x E X such that f(x, y) < c), which is open in Y (resp. in X); the Ax (resp. the B) form a
covering ofY (resp. X) when x varies in X (resp. y varies in Y). Show that there exist two finite
sets Xo c X, Yo c Y such that: 10 for all y belonging to the convex envelope Bo of Yo, there
exists x E Xo such that f(x, y) > c, and Xo is minimal for this property; 2o for all x belonging
to the convex envelope Ao of Xo, there exists y E Yo such that f(x, y) < c, and Yo is minimal
for this property. Then for all y E Yo, let C y be the set of x E Ao such that f(x, y) ~ c; using
exerc. 21, a) of II, p. 79, show that the intersection of the C y for y E Yo is not empty. Proceed
in the same way in Bo and obtain a contradiction.)

~ 28) Let X be a compact convex subset of E a Hausdorff locally convex space, and let f
be an upper semi-continuous convex function in X. Show that the set L of continuous convex
functions g in X such that g(x) > f(x) for all x E X is decreasing directed and that its lower
envelope is equal to f. (Let u, v be elements of L. To construct an element of L which is less
than u and v, use reasoning analogous to that of prop. 6, II, p. 40. Interpret the set K1 analogous
to the set K in this argument as the set of points situated above the graph of a lower semi conti-
nuous function that is less than u and v and strictly larger than f at every point; apply prop. 5
of II, p. 39 and Dini's tho to this function. To show that the lower envelope of Lis f, note that
f is bounded above by a constant b; (x, t) being a point of E x R situated above the graph
of f, let K' be the convex envelope of {(x, t)} u (X' x {b}), where X' is a convenient compact
neighbourhood of x in X; argue with K' as above for Kl')

29) Let X be a compact convex set in a Hausdorff locally convex space E. Let u be a lower
semi-continuous convex function in X and v an upper semi-continuous concave function in
X such that u(x) > vex) for all x E X. Then there exists an affine linear function f that is
continuous in E and such that vex) < f(x) < u(x) for all x E X.

30) Let X be a compact convex set of a Hausdorff locally convex space E. Show that the set
of lower semi-continuous convex functions in X is a lattice.

§ 6

I) Let F, G be two vector spaces in duality, such that cr(F, G) is Hausdorff. Show that if:Y is a
Hausdorff topology compatible with the vector space structure of F and coarser than cr(F, G)
(but not necessarily locally convex a priori), then :Y = cr(F, G 1)' where G 1 is a vector sub-
TVS II.82 CONVEX SETS AND LOCALLY CONVEX SPACES §6

space of G, dense in the topology cr(G, F). (Consider on F the locally convex topology :Y[
in which a fundamental system of neighbourhoods of 0 is formed by the closed, convex, balanced
sets in :Y which are neighbourhoods of 0 for cr(F, G).) Deduce that if ffo is a Hausdorff locally
convex topology on a vector space E, that is minimal in the set of Hausdorff locally convex
topologies on E (II, p. 85, exerc. 13) it is also minimal in the set of topologies (locally convex
or not) that are Hausdorff and compatible with the vector space structure of E.

2) In R", let (C)! "i"m be a family of m ~ n + 1 convex cones with vertex 0; show that if,
for any n + I of these cones there exists a hyperplane H through 0 and such that the cones
lie on the same side ofH, then there exists a hyperplane Ho such that all the cones C i (1 ~ i ~ m)
lie on the same side of Ho (ef II, p. 68, exerc. 21, a».

3) In R" let (D;)! "i"m be a family of m ~ 2n closed half-spaces determined by hyperplanes

passing through O. Show that if, for any 2n of these half spaces, there exists a point i= 0 in their
intersection, then there exists a point i= 0 in the intersection of all the Di (1 ~ i ~ m) (ef II,
p. 66, exerc. 10).

4) Let S, T be two finite sets in R", without common points, such that their union contains
at least 2n + 2 points. Then there exists a hyperplane separating S from T if, and only if, for
every finite set F c S u T of 2n + 2 points, there exists a hyperplane separating FilS from
F 11 T (use exerc. 3 and the method of II,p. 78, exerc. IS).

5) Let E be the vector space of quadratic forms on R", which is identified with the vector
subspace of symmetric square matrices in the space M"(R) of square matrices of order n on R.
We endow Mn(R) with the scalar product Tr('X. Y), which enables us to identify it with its
dual and similarly for E.
a) Let PeE be the set of quadratic forms for which the matrix has all elements ~ 0, and let
SeE be the set of positive quadratic forms in R". Show that we have P = po and S = So.
m
b) Let B be the set of quadratic forms on R" that can be written in the form I x? for same m,
j~ !
where xi is a linear form that takes values ~ 0 for all x = (x)! "i"n of coordinates Xi all ~ 0;
let C be the set of quadratic forms that are ~ 0 for all the vectors x = (x,) with coordinates
Xi all ~ O. Show that B = Co and C = BO (prove that B is closed, showing that every element
m
of B can be written in the form I Xi2, with xi positive for all x with coordinates ~ 0, and
j~ !
m ~ 2").

6) Let F, G be two vector spaces in separating duality, and A a weakly compact convex set
in F. Let C be a convex cone with vertex 0, that is weakly closed in G. Suppose that, for all
<
y E C, there exists x E A such that x, y) ~ O. Show that there exists Xo E A such that
<x o, y) ~ 0 for all y E C (apply prop. 4 of II, p. 38, to A and CO).

4[ 7) a) Let F, G be two vector spaces in separating duality and C a weakly closed convex
cone in F. Let M be a finite dimensional vector subspace of G. Show that, either there exists
Yo E C such that Yo E MO and Yo i= 0, or there exists Zo E M such that Zo E Co and Zo i= 0
(argue by induction on the dimension of M). If C does not contain any line and if the two pre-
ceding properties are simultaneously satisfied, show that Zo cannot be an internal point of Co.
b) Let the two matrices (a i), (bij) with real entries in n rows and m columns, be such that
aij > 0 for every pair (i, j). show that there is a unique value of A E R such that there are two
vectors x = (x) E Rm, y = (y,) ERn satisfying the relations x i= 0, y i= 0, Xj ~ 0, Yi ~ 0
for all i, j and finally such that
m m
(1) A I!
j~
QijXj ~ I ! bijxj
j~
for ~i~n

n
(2) A I 1 aijYi
i=
~ I"
i= 1
bijYi for ~j ~ m.
§6 EXERCISES TVS II.83

(Putting c iJ = Aa'J - b,j for I ~ i ~ n, I ~ j ~ m, and cn+,.j = o'J (Kronecker's index)

for I ~ i ~ m show that the problem reduces to finding a vector x E Rm and a vector
z = (z) E Rn+m which are non null and satisfy the relations
m
(3) I CijXj ;;:, 0 for I~i~n+m
j~ 1
n+m
(4) I
i= 1
CijZ i = 0 for I ~j ~ m

and Zi ;;:, 0 for I ~ i ~ n + m. Remark that, if (3) has a solution for one value AD of A, then
it also has a solution for A ;;:, AD, and that if(4) has a solution for AD, then it also has a solution
for A ~ AD. Finally use a).)
.~ 8) Let T be a compact space and L a vector subspace of ~(T ; R), that is of finite dimen-
sion r; give to L the norm induced by that of ~(T; R) and to its dual L * the norm
Ilx'll = sup <x,x'),sothatifBistheballllxll ~ linL,thenBOistheballllx'll ~ linL*.
Ilxllo
a) For all t E T, write e; for the linear form x f-+ x(t) on L. Show that BOis the convex envelope
of the set of the ± e;, where t varies in T (obtain a contradiction using prop. 4 of II, p. 38).
r
Deduce that every linear form x' E L * such that Ilx'll = I we can write x' = I Aie;i' where
i= 1

the ti are r points ofT and the Ie; are real numbers such that I IAil = I (ef II, p. 66, exerc. 9, a)).
i=1
b) For every y E ~(T; R), there exists a unique x E L such that Ily - xii = dey, L), if and
only iffor every non null Z E L, there exist at most r - I distinct points ti E T such that z(tJ = 0
(Haar's th.). (To show that this condition is sufficient, observe first that it is equivalent to
saying that for r distinct points ti E T (1 ~ i ~ r) the e;i are linearly independent in L*. Now
argue by assuming the conclusion is false and obtaining a contradiction. If there exist two dis-
tinct points x', x" of L such that Ily - x'il = Ily - x" I = d (y, L) then there exists Xo ELand
z E L such that, for all sufficiently small real A, we have IIY - (xo + Az)11 = dey, L). Apply the
last part of a) to the subspace L EB Ry of C(j (T ; R) and to a suitable linear form on this space
which vanishes in L. To see that the condition is necessary, note that if it is not true, then there
exist r distinct points ti E T (I ~ i ~ r) such that the e;i are linearly dependent and there exists
a function z E L that is not null and vanishes at the points t i • If a i (1 ~ i ~ r) are numbers
not all zero such that f
i= 1
aie;i = 0, consider a function w E ~(T; R) such that Ilwll = I,
wet,) = sgn(lX,) for I ~ i ~ r, and the function y = w(1 - IPzl) with IPI sufficiently small and
# 0.)
c) Suppose that T is a compact interval in R and that L satisfies the condition of Haar's tho ; let
(t')1
1
"'.'"
-...::l-....;;r
+ 1 be a strictly increasingr +sequence
1
of r + I points of T; then there exist r + I real
non zero numbers Ai such that I Aie;i = O. Show that then sgn(AJ sgn(Ai+ I) = - 1 for
i= 1
1 ~ i ~ r. (Consider separately the case r = 1 and the case r > 1. In the second case, suppose
on the contrary, that for some index i ~ r - I, the number Ai is of the same sign as \-1
or as Ai+ 1 and that Ai-I and Ai + 1 are of opposite signs. If, for example, Ai > 0, take lXi -1 > 0,
lX i + 1> 0 such that lX i - 1 Ai - 1 +lX i + 1 Ai+ 1 = 0, then z E L such that Z(t i _ 1 ) = lXi-I' z(ti+ 1) = a i + 1
and z(t.) = 0 for j distinct from i - I , i and i + 1. Deduce that z(t,) < 0 and show that this
contradicts the given hypotheses.)
d) With the same hypotheses and notations as those of c), suppose that there exists y E ~(T; R)
and z E L such that yet,) - z(t,) = (- l)ilXi with a i > 0 for I ~ i ~ r + 1. Show that we then
have dey, L) ;;:, inf a i • (Use a) applied to L EB Ry, and c).)
e) With the hyp~theses of c) let y E ~(T; R), and z be the unique point of L such that
Ily - zll = dey, L). Show that there exists a strictly increasing sequence (t,)I';i';r+l in T
such that
yet,) - z(t,) = (- l)icll y - zll
TVS II.84 CONVEX SETS AND LOCALLY CONVEX SPACES §6

with e = + 1. Conversely if z has this property, then z is the unique point of L such that
Ily - zll = dey, L). (Use c) and d». Consider the case when T is an interval ofR and when L is
the set of the rcstrictions to T of polynomials of degree < r (T chebycheffs th.).

9) Let F 'and G be two vector spaces in separating duality and A be a convex subset of F which
contains O. For every y E G, writc

HA(y) = sup(- <x,y»),

.xEA

so that 0 ~ HA(y) ~ + 00; we call HA the support function of A.

a) Show that HA is the gauge of A 0 (II, p. 20).
b) If A is weakly compact, then, for all y E G, the hyperplane with the equation <
x, y) = HA(y)
is a support hyperplane of A.
c) HA is finite and continuous for the topology cr(G, F) if, and only if, A is finite-dimensional
and bounded (in the finite dimensional vector subspace that it generates).
d) Let Ai (1 ~ i ~ p) be convex sets in F which contain 0 and Ai be real numbers;:. 0 (1 ~ i ~p);
show that the support function of the convex set A = L
AjAj is HA = LAjBA ,. If Y EGis
i i
such that the intersection C i , of Ai with the hyperplane <x, y) = HA(y), is non-empty for
<
1 ~ i ~ p, show that the intersection of A and the hyperplane with equation x, y) = HA(y)
is the set L '!ciCj'
e) Suppos~ that A is locally compact and does not contain any line. Then the set which is the
union of {O} and the set of y oF 0 such that HA(y) = + 00, H A( - Y) oF +x is the polar cone
of the asymptotic cone C A (consider the case when A is the convex envelope of {O} and of one
half-line).
f) Suppose that F is finite dimensional. Show that, A is parabolic (II. p. 67. exerc. 17) ifand only
if HA is a continuous mapping of G in R (if there exists a line paralle to a half-line of C A, which
does not meet A, note that there exists a hyperplane separating this half-line from A).

10) To each compact convex set A in E = R" containing 0, we make correspond its support
function HA by the duality between E and E* : HA belonging to the space '6'(E*; R) of conti-
nuous real valued functions in E*. We ascribe to the space '6'(E*; R) the uniform structure of
compact convergence and, to the set R~(E) of the compact convex sets in E containing 0, the
uniform structure defined in the exerc. 39 ofII, p. 71. Show that A f--+ HA is an isomorphism of
R~(E) on a uniform subspace of (&,(E* ; R).
Deduce that the mapping A f--+ A of the set Ro(E) of compact convex sets in E which contain
0

o as an interior point, on the set S~o(E*), is an isomorphism for the uniform structures of these
two spaces (ef II. p. 71, exerc. 39).

11) Let F, G be two vector spaces in separating duality. An ultrafilter U on F converges weakly
to a point Xo if, and only if, Xo belongs to the intersection of all the weakly closed convex sets
which belong to U (note that if Xo is a point of this intersection that is not a cluster point of U,
then there exists a closed half-space belonging to U and not containing x o).
Deduce from this result that, for a sequence of points (xn ) of F to be weakly convergent to a
point a, it is necessary and sufficient that a belongs to all the weakly closed convex envelopes of
the sets formed by an infinity of the terms of the sequence (use prop. 7 of GT, I, § 6.4).

12) a) Let E be a vector space and (E')'EA be an increasing directed family of subspaces of E,
whose union is E; cach Ea is supposed to carry a locally convex topology;!, such that for
a. ~ ~ the canonical injection Ea -> E~ is continuous. Let:;T be the topology on E, which is the
inductive limit of the .Y, (II, p. 29, Example II); show that the dual E' of E (for :;T) with the
topology cr(E', E) can be canonically identified with the projective limit of the duals E~, with
topology cr(E~, E,).
b) Let (X" <PaP) be a projective system of non-empty sets corresponding to a directed set of
indices A, such that <P,~ are surjective and that lim X, = 0 (S, III, § 7, exerc. 4). Put F, = R(X,)
and denote by fop : F p -> F, for a. ~ ~ the li;)ear mapping deduced canonically from <PaP
§6 EXERCISES TVS 11.85

(A, II, ~ 1 .1 L cor. 1). If, we give to each Fa the topology which is the direct sum topology of its
factors, the dual Ea = F~ of Fa' with the weak topology cr(Ea' Fa)' can be identified with the
product space R Xa , and tf.~ is an isomorphism ofE a on a closed subspace ofE~, having a topo-
logical complement in E~. Show that on E = liw Ea (for the tfa~) the topology which is the
inductive limit of those of Ea is the coarsest topology (therefore non-Hausdorff) (using a) and
noting that lim Fa = {O }).
+-

13) Let E be a vector space. We say that a Hausdorff locally convex topology fT on E is
minimal (and that E, with fT, is a space of minimal type) if there exists no Hausdorff locally
convex topology on E, that is strictly coarser than fT (el II, p. 81, exerc. 1).
a) Let fT be a minimal topology on E, and let E' be the dual ofE (when E has the topology fT) ;
show that fT = cr(E, E') and E = E'* (note that there cannot be an everywhere dense hyper-
plane in E' for the topology cr(E', E) using the cor. 3 of II, p. 43). Deduce that spaces of mini-
mal type are products of lines.
b) Show that in a Hausdorff locally convex space F, every subspace E of minimal type has a
topological complement, and in particular is closed (use a) and the Hahn-Banach tho for
extending the identity mapping ofE in itself to a mapping ofF in E).
e) Let u be a continuous linear mapping of a space E of minimal type in a Hausdorff locally
convex space F. Show that u(E) is closed in F and that u is a strict morphism of E in F (use b)
and the definition of a space of minimal type).
d) Let F be a Hausdorff locally convex space and M be a closed vector subspace of F. Show
that, if there exists a complement N of M in F that is a subspace of minimal type, then N is
a topological complement of M in F (use e)).
e) Let M be a subspace of minimal type in a Hausdorff locally convex space F; show that,
for every closed vector subspace N of F the sum M + N is closed in F (consider the quotient
space FIN and use e)). If further N is of minimal type, then M + N is of minimal type.

14) Let E be a Hausdorff locally convex space and F a locally convex space of minimal type
(exerc. 13).
a) Show that if M is a closed vector subspace of the product space E x F, its projection on E
is closed in E (use exerc. 13, e)).
b) Let u be a linear mapping of E in F. Show that if the graph of u is closed in E x F, then u
is continuous (use a)).
e) Suppose that, in E, every closed vector subspace has a topological complement (ef. V, p. 13).
Show that, in E x F, every closed vector subspace M has a topological complement. (If N 1 is
the projection of M on E and N 2 a topological complement of N 1 in E, if PI = M n F, and
P 2 is a topological complement of PI in F, show that N 2 + P 2 is a topological complement of
M in E x F, using b).)

* 15) Let E, F be two Hausdorfflocally convex space. We say that a continuous linear mapping
u: E -+ F is linearly proper if, for every Hausdorff locally convex space G and every closed
vector subspace V of E x G the image of V by u x IG: E x G -+ F x G is closed. Show that
this condition is equivalent to the following: u- 1 (O) is a subspace of minimal type of E and
for every closed vector subspace W of E, the set u(W) is closed in F. (To show that the first
condition implies the second, consider the mapping v : E -+ {O} and, giving E the topology
cr(E, E'), so that E is immersed in E'* with cr(E'*, E'), take the image under the projection
v x I E" : E x E'* -+ E'* of the closure in E x E'* of the diagonal ~ of E x E. To show that
the second condition implies the first, show that it implies that, for the topologies cr(E, E') and
cr(F, F'), the mapping u is a strict morphism and use exerc. 13, e).) *

16) Let F be a product oflines and C a closed convex set in F.

a) Show that there exists Xo E F, two sets I and J and a topological isomorphism u of F on
RI x R J such that u(xo + C) is of the form RI x A, where A is a closed convex set of R~.
(Note that we have F = G* where F has the topology cr(G*, G); consider the polar Co of
C in G, the vector subspace ofG generated by Co and a complement of this subspace.)
b) IfC does not contain any affine line, the mapping (x, y) H X + Y ofC x C in F is proper.
e) Suppose that C is a cone with vertex 0 and that the uniform structure induced on C by that
TVS 11.86 CONVEX SETS AND LOCALLY CONVEX SPACES §6

of F is metrisable. Then if the sets I and J, the point Xo and the mapping u satisfy the conditions
of a) then I is enumerable and there exists an enumerable subset H of J such that the restriction
of the canonical projection p : RI x R J ..... RI X RH to u(xo + C) is an isomorphism of the uni-
form subspace u(xo + C) of RI x W on the uniform subspace p(u(x o + C» of RI x RH.

17) Let E be an infinite dimensional vector space.

a) Show that there exist hyperplanes in E* that are everywhere dense for the topology cr(E*, E).
b) If H' is such a hyperplane. show that, in E, the only linear subvarieties oF E that are every-
where dense for the topology cr(E, H') are the hyperplanes.

'IT 18) a) In a normed space E, let A be a closed convex set oF E; show that the function
x f-> d(x, CA) is concave in A (use the fact that A is the intersection of closed half-spaces).
b) Define inductively a sequence of closed convex sets An C Rn in the following manner;
Al = R+; if Rn+l is identified with R" x R, then A n+ l is the set of pairs (x, s)
such that
x E An and that

where Ilxll is the Euclidean norm. Show that An+ I does not have any support hyperplane of
the form H x R, where H is a hyperplane of Rn and that its asymptotic cone is {O} x R+.
c) If Pnm is the canonical projection R m ..... Rn (Rm being identified with Rn x Rm-n) for m ;;, n,
show that when RN is identified with the projective limit of the projective system (Rn, Pnm )
the An form a projective system of sets and that A = I~ An is a closed convex set not relatively
compact in R N , having no closed hyperplane of support and such that C A = {O}.

19) a) Let A be a closed convex set in E, a product of lines, that is non-compact and such
that C A = {O} (exerc. 18). Show that if B = A - A and if M is the convex closed envelope
of Au (- A), then Band M contain lines (use exerc. 16, b) of II, p. 85).
b) Let At, A2 be two closed convex sets in E such that Ai + A2 is closed and none of At,
A2 , At + A2 contain affine lines. Show that C A, +Az = CAl + CAz (use exerc. 16, b) oflI, p. 85).
e) Let A be a closed convex set in E that docs not contain any affine line and M 1 , ... , Mn
closed convex sets contained in A. IfB is the convex envelope ofU M i , show that B = B + I C M ;
i
and C B = I C M , (same method).

'IT 20) Let F= RfA), G = R\ where A is any infinite set: suppose that F and G are put in
<
separating duality by the bilinear form x, y) = I x(a) yea) .
• EA
a) Let N be an additive subgroup of G; we denote by N* the subgroup of the x E F such that
< x, y) is an integer for all YEN and by N** the subgroup of the Z E G such that x, z) is <
an integer for all x E N*. If N is the closure of N for the topology cr(G, F), show that N* is
closed in F for cr(F, G) and that N** = N (to establish this last point, use GT, VII, ~ 1.3,
prop. 6, projecting N on the finite dimensional coordinate varieties of G).
b) Suppose that A = N. Let M be a closed subgroup of F for cr(F, G); show that if V is the
largest vector subspace contained in M, then M is the topological direct sum of V and of a
closed subgroup P that is a free Z-module having an enumerable base. (Consider F as the
union of an increasing sequence (Fn) of finite dimensional vector subspace and apply GT,
VII, ~ 1.2, tho 2 and ~ I, exerc. 7.) P is discrete (for the topology induced by Ci(F, G», if, and
only if, P is of finite rank.
e) Deduce from a) and b) that when A = N, every closed subgroup of G (when G carries the
product topology Ci(G. F»
can be transformed, by an automorphism of the topological
group G, in a product RI x ZJ, where I and J are two sets of N without common elements.
d) In the space E = RN, carrying the topology cr(E, E*), show that the subgroup ZN is closed
and does not contain any line, even though it is not a free Z-module (A, VII, p. 59, exerc. 8);
the results of b) do not therefore extend when A is not enumerable.
§7 EXERCISES TVS 11.87

§ 7

1) Let A be a convex set. Then, a point x EO A is extremal in A if, and only if, for any subset
B of A, the statement x belongs to the convex envelope of B, implies that x EO B.

2) With the notation ofII, p. 74, exerc. 3, let G be the vector subspace ofE generated by K U {A}.
Show that, in G, the point A is an extremal point of the closed convex envelope of K, but
that A does not belong to K (el II, p. 25, corollary).

~ 3) Let A be a convex set in a vector space E, and let x be a point of A. We call the set formed
by x, and the y =1= x in A such that the line passing through x and y contains an open segment,
that is contained in A and contains x, the jaeet of x in A. The internal points relative to the
linear variety generated by A (II, p. 26) (resp. the extremal points) of A are the points whose
facet in A is equal to A (resp. is a single point).
a) Show that the facet F x of a point x EO A is the largest convex set B c A such that x is an
internal point of B (relative to the linear variety generated by B).
b) For every pointy EO F x' the facet F y of y in A is identical with the facet of y in F x' In order
that F = F x' it is necessary and sufficient that y is an internal point of F x (relative to the
linear ~ariety generated by FJ. Deduce that, if Fx is finite dimensional, and if y is a non-
internal point of Fx (relative to the linear variety generated by FJ, then the dimension of Fy
is strictly less than that of F x'
e) A linear variety V in E which meets A and is such that for every x EO A II V, every open seg-
ment contained in A and containing x, is necessarily contained in V, is called a support variety
of A. Show that, for all x EO A, the linear variety M generated by the facet Fx of x in A is the
smallest support variety of A which contains x, and that M II A = Fx' For every support
variety V of A, the intersection V II A is the facet in A of each of its internal points (relative
to the linear variety generated by V II A).
d) Let A and B be two convex sets in E. For every point x EO A II B, the facet of x in A II B
is the intersection of the facets of x in A and in B.
e) Let B be a closed convex set in a Hausdorlf topological vector space E, and let B contain
a closed linear variety M of finite codimension n; then every facet in B of a point ofB contains
a closed linear variety of codimension n (II, p. 67, exerc. 14, d)). If A is a convex set then the
facet in A II B of a point x in A II B is of finite dimension if, and only if, the facet of x in A
is of finite dimension: further if p and q are the dimension of the facet of x in A and of the facet
of x in A II B, then p :( q + n. In particular if x EO A II B is an extremal point of A II B,
then its facet in A is of dimension :( n.
j) Deduce from e) that if A is compact, and V is a closed linear variety in E, of finite codimen-
sion n, then every extremal point of V II A is a linear combination of at most n + 1 extremal
points of A.

4) In the plane R 2, consider the convex set A formed by the points (~, 11) satisfying - 1 :( ~ :( 1,
- 1 - )1 - ~2 :( 11 :( 1 + )1 - ~2. Show that there exist frontier points of A for which
the facet is distinct from the intersection of A and of the lines of support of A passing through
this point.

5) In the Banach space I'X\N) of bounded sequences x = (~n) of real numbers, let A be the
closed convex set defined by the inequalities - 1In :( ~n : ( 1 for n ?o 1 and - 1 :( ~o :( 1.
Show that A has a non-empty interior, that the origin is a frontier point of A and that the
facet of 0 in A is not closed. If we give to A the topology induced by that of the product space
R N , show that A is compact but that the facet of 0 in A is not closed in A.

6) Let E, E' be two vector spaces in separating duality, and A be a convex set in E containing 0
and closed for cr(E, E'). For all a EO A, the set F~ of points x' EO AO such that <a, x'> = - 1
is a closed (for cr(E', E)), convex set of Ao. Show that F~ is the facet in A of each of the internal
0

points of F~ relative to the linear variety generated by F~. We say that F~ is the dual jaeet
of a in Ao. If Fa is the facet of a in A, show that F~ is also the dual facet in A of each of the
0

internal points of Fa relative to the linear variety generated by Fa; further, if A is identified
TVS II.88 CONVEX SETS AND LOCALLY CONVEX SPACES §7

with ADO, the dual facet in A of each internal point of F~ relative to the linear variety generated
by F~, contains Fa' When F~ is not empty (which is always the case when E isjinite-dimen-
sional and a oj:. 0, cf II, p. 78, exerc. 13), we say that each of Fa. F~ is the dual facet of the other.
We say that a point a E A is smooth point of A if F~ is a single point (in other words if there
exists one closed hyperplane of support of A passing through a); we say that a is a point of
strict convexity (or is an exposed point) if there exists a closed hyperplane H supporting A
so that H n A = {a}; this is the same as saying that there exists an internal point of F~
(relative to a linear variety generated by F~) which is a smooth point of AO.

which °
~ 7) Let E be a vector space of finite dimension n and let A be a closed convex set in E of
is an interior point.
a) Let F and F' be two dual facets of A and A ° (exerc. 6); if F is of dimension p and F' of
dimension q, then show that p + q :( n - 1. For every frontier point x of A, the dimension
of the facet of x in A is called the order of x, and the dimension of its dual facet in A° is called
the class of x. The order (resp. the class) of a facet F of A is by definition the order (resp. the
class) of one of the internal points ofF relative to the linear variety generated by F. An extremal
point of A is a point of order 0; a smooth point of A (exerc. 6) is a point of class 0.
b) A frontier point of A of class n - 1 (and hence of order 0) is called a vertex of A. Show
that the set of vertices of A is enumerable (consider the set of dual facets of the vertices of A,
GT, VI, ~ 2, exerc. 12).
c) Let F be a p-dimensional facet of A, and M a linear variety of dimension n - p, which
meets F in the single point a, such that a is an internal point ofF and which contains an interior
point of A Show that, if V is a support hyperplane of M n A in M, that passes through a,
then the hyperplane H generated (in E) by F u V is a support hyperplane of A
d) We say that a facet F of A of order p and of class q is an ultrajacet if p + q = n - 1; the
dual facet is then also an ultrafacet of A If a linear variety M of dimension n - p meets an
0.

ultrafacet F in a single point that is an internal point of F (relative to the linear variety gene-
rated by F), show that this point is a vertex of the convex set M n A, and conversely (use c».
Deduce that the set of ultrafacets of order p of A is enumerable. (Identify E with R n, consider
the projection of A on each of the coordinate varieties of Rn of dimension p; if the set of ultra-
facets of order p of which the projection on V is p-dimensional, is not enumerable, show that
there exists a point of V which is an interior point to a non-enumerable infinity of these pro-
jections considering the points of V with rational coordinates; then use b).) Give an example
of a convex set with a non-enumerable infinity of facets, each of which is not a single point
nor an ultrafacet.
e) If all the frontier points of A are smooth, show that the mapping, which puts each point x
of the frontier G of A in correspondence with the unique point of the dual facet of x, is a con-
tinuous mapping of G on the frontier of A (cf TG, I, ~ 9.1, corollary). In what case is this
0

mapping bijective ?

8) Let E be a vector space of finite dimension n and A be a compact convex set in E.

a) Let H be a hyperplane in E. Show that in an open half-space determined by H and contain-
ing at least one point of A, there exists a point of strict convexity of A (II, p. 87, exerc. 6). (Con-
sider, in H, a closed euclidean ball C of dimension n - 1 and of sufficiently large radius that
contains H n A, then the euclidean balls B of dimension n and of larger radius containing A
and such that B n H = C.)
b) Show that A is the closed convex envelope of the set of points of strict convexity (use a».
c) Show that every extremal point of A is a cluster point of the set of points of A of strict
convexity. (Using b) and the exerc. 9, a) of II, p. 66, note that an extremal point is the limit of
n n
a sequence of points of the form L Aimxim' where Aim ~ 0, L Aim = 1 and the xim are points
i=O i=O
of strict convexity of A; next use the compactness of A)

9) Show that in the product space E = R N, the cube IN, where I = (0, 1), is a compact convex
set with no point of strict convexity.

10) In the space R2, show that the set of extremal points of a closed convex set A is closed
§7 EXERCISES TVS II.89

(show that the set of points of A whose facet in A is of dimension 1 form an open set relative
to the frontier of A).

the union of the circle S = 0, ~ 2 + 112 - 2~ = °

11) a) In the space R 3 , consider the compact convex set A which is the convex envelope of
and the two points (0, 0, 1) and (0, 0, - 1).
Show that the set of extremal points of A is not closed in A.
b) Let A be a metrisable compact convex set in a Hausdorff topological vector space E.
Show that the set of extremal points of A is the intersection of a sequence of open sets in A.
(If d is a distance defining the topology of A, then for each integer n consider the set of points
x = ~{y + z), where y, Z are in A and dey, z) ;:0, lin.)

°
12) In the Banach space ["'(N), let en be the sequence all of whose terms are zero except the
n-th which is 1. Let A be the convex closed envelope of the set formed from and the points
e,,/(n + 1) (n ;:0, 0). Show that A is compact but that it is not identical with the convex enve-
lope of the set of its extremal points.

* 13) In the Hilbert space [2(N), let A be the set of points x = (~n) such that we have
I 22n~; ~ 1. Show that A is convex, compact and that it is the closure of the set of its extremal

points. *
14) Let E be a closed vector subspace of the Banach space [WeN), formed of the sequences
x = (~n) such that lim ~n = 0.
n~ro

a) Show that, in the Banach space E, the closed unit ball B does not have any extremal points.
b) Let u be the continuous linear form (~n) I-> I rn~n on E. Show that there does not exist
any support hyperplane of B that is parallel to the closed hyperplane with the equation u(x) = 0.
15) Let A be a compact set in the normed space E.
a) Show that the distance apart of two parallel support hyperplanes of A is at most equal
to the diameter 8 of A.
b) Show that there exist pairs of points (a, b) of A such that Iia - bll = 8; for such a pair of
points, there exist two parallel support hyperplanes of A passing respectively through a
and b and whose distance apart is 8 (consider the closed ball of centre a and radius 8).

16) a) Let A be an n-dimensional compact convex set in the space Rn, normed with the Eucli-
dean norm; for every Z E Sn _ l' denote by p(z) the upper bound of lengths of segments parallel
to the vector z and contained in A. Show that there exist two points u, v of A such that the
segment with end points u, v is parallel to z and is of length p(z); deduce that there exist two
support hyperplanes of A that are parallel and pass respectively through u and v (consider
the set A' = A + p(z) z, and separate the sets A and A' by a hyperplane).
b) Let d be the lower bound of the distances between two parallel support hyperplanes of A;
show that there exist two points a, b of A such that Iia - bll = d, and that the hyperplanes
passing respectively through a and b and orthogonal to a - b, are support hyperplanes of A
(use a).

17) In the space [W(N), let A be the compact convex set defined in the exerc. 12 of II, p. 89 and
let E be the closed vector subspace of [W(N) generated by A. Show that the lower bound of
the distance between two parallel closed support hyperplanes of A in the space E is equal
to 0, even though A is not contained in a closed hyperplane of E.

18) In a Hausdorff locally convex space E, let (K')'Ei be a decreasing directed family of convex
sets that are compact and non-empty. For all cr E I, denote the set of extremal points of K.
by A., and by F. the closure of the union of the Ap for ~ ;:0, cr, so that (F.) is a decreasing directed
family of compact sets. Let A be the intersection (non-empty) of the F., and K the intersection
(non-empty) of the K •. Show that K is the closed convex envelope of A. (If f is a continuous
linear form on E, and x. a point ofF. where f attains its maximum in F., show that fey) ~ f(x.)
for all y E K; then take a cluster point of the family (x.) following the filter of sections of 1.)
TVS 11.90 CONVEX SETS AND LOCALLY CONVEX SPACES §7

~ 19) In the space Rn, let (K.) be a family of compact sets, in number;;:, n + 1, and such
that none of them is contained in an affine hyperplane. Suppose that for every family (u.)
of affine automorphisms of Rn, if any n + 1 of the sets u.(K.) have a common point, then all
the u.(K.) have a common point. Show that under these conditions the K. are convex. (Sup-
pose on the contrary that there exist n + 1 points x1"'" xn + 1 in the same set K. and a point Xo
which belongs to the convex envelope of the set of the Xi (i ;;:, 1) but does not belong to K •.
Note that for every index i ;;:, 1, there is an affine automorphism U i of Rn and an index Ct.i
such that Xo and the x· of index j =1= i are extremal points of u;CK.), and show that the n + 2
sets K. and u;CK.) ha~e no points in common.) .

20) Give an example of a compact convex set K in R2, containing 0 and such that the cone
of vertex 0 generated by K is not closed in R 2 •

~ 21) a) In a Hausdorff locally convex space E, let A be a locally compact closed convex
cone, that does not contain any line. Show that A is a cone of compact sole (apply prop. 2
of II, p. 55, to the vertex of A, which is an extremal point of A). Deduce that there exists a
closed support hyperplane H of A, which contains the vertex s of A and is such that HilA = {s}.
b) Let A, B be two closed convex cones with vertex 0 in E, that are locally compact and do
not contain a line. Show that if A 11 B = {O}, then A - B is a closed, locally compact, cone
not containing any line (method as in II, p. 67, exerc. 16). Deduce that there exists a closed
hyperplane that supports both A and B, that separates A from B and such that
HilA = H 11 B = {O}.
e) Give an example of a locally compact closed convex cone A such that A - A is not locally
compact (ef II, p. 78, exerc. 11).
a) Let A be a locally compact closed convex set in E, which does not contain a line. Let Xo
be a point of A, C A the asymptotic cone of A (ef II, p. 67, exerc. 14) and H a support hyper-
plane of Xo + C A passing through Xo and such that (xo + C A ) 11 H = {x o }. If f(x) = a is
the equation of H and if f(x) ;;:, a in Xo + C A , then show that for every real number b, the
set of the YEA such that fey) ~ b is compact.

~ 22) By an extremal ray of a convex subset A of a vector space E we mean a closed half
line D contained in A, such that, for all XED and every open segment with end points a, b
in A, and which contains x, it is necessarily the case that a E D and bED; the end point of
D is an extremal point of A.
a) In a Hausdorff locally convex space E, show that every locally compact closed convex
set, not containing a line, is the closed convex envelope of the union of its extremal rays and
its extremal points. (Suppose the contrary, and writing B for this closed convex envelope,
note first that by exerc. 21, d), there exists a closed hyperplane H so that HilA is compact
and non-empty and H 11 B = 0. Show then that if a E HilA is an extremal point of HilA
(therefore not an extremal point of A by hypothesis) and if the open segment S with end
points b, e contained in A and not contained in H, contains a, then the line D containing S
necessarily contains a segment containing a and whose end points are extremal points of A,
or contains an extremal ray of A containing a.)
b) Prove that if E is finite dimensional, then every closed convex set in E, that does not contain
any line, is the convex envelope of the union of its extremal points and its extremal rays (argue
by induction on the dimension of E).

23) In R 3 , consider a closed convex set A with an interior point, whose fonction F contains
two open segments S, T lying in two non-parallel lines D, D' (the points of S, T are thus non-
extremal in A), and all of whose other frontier points are extremal (one shows how to define
such convex sets). For every x E R 3 , put f(x) = (d(x, D)f Let B the convex closed set in
R4 = R3 X R formed by the pairs (x, t;) such that x E A and S ;;:, f(x). Show that in B the set
of extremal points, the union of the extremal rays, the set of end points of extremal rays, and
all the unions of two or three of these sets are not closed and non-empty.

~ 24) In Rn, every intersection of finitely many closed half spaces (resp. of closed half-spaces
determined by hyperplanes passing through the same point) is called a polyhedron (resp. a
§7 EXERCISES TVS II.91

polyhedral cone). A convex set C c:: R" is locally polyhedral in a point x E C if there is a neigh-
bourhood V of x in C which is a polyhedron.
a) Show that. if a closed convex set C c:: R n is locally polyhedral at a point x E C, then the
cone with vertex x generated by C is polyhedral.
b) Show that a compact convex set in R" that is locally polyhedral at each of its points is a
polyhedron (use a».
c) Let P c:: Rn be a closed convex set with an interior point. Show that the following condi-
tions are equivalent:
ex) P is a polyhedron.
~) P has only a finite number of facets (II, p. 87, exerc. 3).
y) P is the convex envelope of a set which is the union of finitely many points and finitely
many closed half lines.
(To show that (1) implies ~) take P as the intersection of the smallest possible number of closed
half spaces, and show that the hyperplanes defining these half-spaces are generated by facets
of dimension n - I of P. To show that ~) implies y), argue by induction on n. Finally, to see
that y) implies ex), consider the polar po of P.)
d) Show that every convex polyhedron P can be written in the form Q + C p , where Q is a
compact polyhedron and C p the asymptotic cone of P. A non compact polyhedron cannot
be parabolic.
e) Show that every facet of a convex polyhedron is an ultrafacet (II, p. 88, exerc. 7. d» (argue
by induction on n).

~ 25) a) Let C c:: R" be a closed convex cone with vertex O. Show that the projections of C
on every 2-dimensional subspace ofRn are closed if and only if C is a polyhedral cone (exerc. 24).
(Reduce to the case when C contains no line : argue by induction on n using the existence
of a compact sole S of C (II, p. 90. exerc. 21. a», and project onto a hyperplane parallel to a
line joining 0 to an extremal point of S, and deduce that S is locally polyhedral (exerc. 24).
b) Deduce from a) that, if we give Rn the order for which C is the set of elements?: 0, then,
every positive linear form on any vector subspace F of Rn can be extended to a positive linear
form on R" if, and only if, C is a polyhedral cone (apply a) to the polar cone CC). If C is the
cone in R3 generated by the (~1' ~2' ~3) such that ~1 = I, ~3 ?: (~D-, and F is the subspace
~3 = 0 give an example of a positive linear form on F that cannot be extended to a positive
linear form on R3.
c) Let A be a polyhedron in R". Then, the convex envelope of A u B is closed for every poly-
hedron B, if and only if. A is compact (use exerc. 24, d».

26) a) Let E be a Hausdorff, locally convex space and let A be a cap of a convex set C in E.
If SEC is a point not belonging to A and if B is the cone with vertex s generated by A show
that the closure of B n (C n CA) is a cap in B.
b) Suppose that E is finite dimensional. Show that every cap A of a closed convex set C in E
can be obtained in the following manner: consider a facet F of C (II, p. 87, exerc. 3) and a
hyperplane H in the affine linear variety V generated by F, such that F is entirely on one side
of H, take as A the set of points of F contained between H and a hyperplane H' of F parallel
to H (use a) and prop. 4 of II, p. 38). Every extremal point of a facet of C is an extremal point
of C.
c) Give an example of a compact convex set C in a Hausdorff locally convex space E and
of a cap A of C such that A and C n CA each generate E and that A and C n CA cannot be
separated by a closed hyperplane of E (cf II, p. 78, exerc. 11).

27) Let C be a closed convex set in a product of lines, E, and let a be an extremal point of C.
Show that for every neighbourhood V of a in C. there exists an open half-space F in E such
that a E F n C c:: V. (Reduce to the case when C is compact.)

28) Let I be a non enumerable infinite set. Show that every cap of the cone Rl+ in RI is con-
tained in the sum of subs paces of the form R J , where J is an enumerable subset ofl (use prop. 4
of II, p. 38). Deduce that there are points of R~ which do not belong to any cap of R 1+, even
though RI+ is the convex closed envelope of the union of its extremal generators.
TVS 11.92 CONVEX SETS AND LOCALLY CONVEX SPACES §7

29) a) Let (En) be a sequence of Hausdorff locally convex spaces and E = TI En their product.
" there exists a cap
In each En. let C n be a convex cone with vertex 0 and An a cap ofCn" Show that
of C = TI C n which contains TI
An (argue as in prop. 5 of II. p. 59).
n n
b) Let (En. <Pmn) be an enumerable directed projective system of Hausdorff locally convex
spaces and let E = lim E" be its projective limit. For all n, let C" be a convex cone of vertex 0
such that (C n ) is a projective system of sets. Show that if, for each n, the set e" is the union
of its caps, then this is also true of C = ~ C" (use a».
In particular, if the C n are cones with
compact soles then C in the closed convex envelope of the union of its extremal generators.

30) Let E be a Hausdorff locally convex space and A a cap of C, a closed convex set in E.
Show that if a E A is an extremal point of A then the facet F of a in C (II, p. 87, exerc. 3) is of
dimension:( I (use exerc. 26 ofIL p. 91). Deduce that F n A is a cap ofC.

~ * 31) Let X be the compact interval (0, I) ofR and <I> the set formed from the continuous
real valued functions defined in X and the functions t f-+ 1t - a 1-', where a E X and 0 < C( < I
(we put 0-' = + oc for r:l > 0). In the space jlt(X) of measures on X, let uttej>+ be the set of
the measures J..L ;? 0 such that all the functions of <I> are J..L-integrable.
a) Give JI'I'; the uniform structure induced by the product structure of Rej>. Show that ./1'1';
is a proper convex complete cone for this uniform structure. (Note that for every function
f E <I> there exists 9 E <I> such that, for all E > 0, there exists u E 'b (X ; R) such that 0:( f - u:( Eg.)
b) Show that the cone .ttej>+ has no extremal generator. (Observe that if J..L E .4iej>+, then all
the measures" such that 0 :( " :( J..L belong to ullej>+')
c) Show that the set S of the J..L E ./I'Iej>+ such that J..L(l) = 1 is a sole of the cone jllej>+ and a sim-
plex in Rej> (II, p. 71. exerc. 41). *

32) Let E and F be two Hausdorff locally convex spaces, let A be a convex subset of E, and
u a linear mapping of E in F.
a) The inverse image under u of a support variety of u(A) (II, p. 87, exerc. 3. c) is a support
variety of A.
b) If A is compact and u is continuous, then every extremal point of utA) is the image under
u of an extremal point of A.
c) If A is a locally compact cone with vertex 0 and if u is continuous then every extremal
generator of u(A) is the image under u of an extremal generator of A.

~ 33) Let E be a Hausdorff locally convex space and A a subset of E.

a) Denote by r o(A) the set of points x E E such that, for every continuous linear mapping
u of E in a jinite dimensional vector space, the image u(x) belongs to the convex envelope
of u(A). This comes to the same as saying that for every closed linear variety V of E con-
taining x and of finite codimension n > O. there exists a subset of A having at most n + I ele-
ments, and of which the convex envelope meets V. Show that r o(A) is a convex set containing
A, that r o(r o(A») = r o(A) and that r o(A) is contained in the closed convex envelope of A
(use prop. 4 of II, p. 38).
b) Let (X')'EI be a family of elements of A and ("JOEl a family of positive numbers such that
I ", = 1 and that the family C",x,) is summable in E. Show that the sum s = I "ox o belongs
~ ~
to r o(A). (With the aid of a) reduce to the case where E is of finite dimension and identical
with the linear variety generated by the )'ox, ; then argue by reductio ad absurdum. considering,
for every finite subset J of I, a closed hyperplane H J that passes through s and does not meet
the convex envelope of the set of the Xo such that C( E J, then using the compactness of the
unit sphere in a finite dimensional space.)
c) Show that if A is compact. then r o(A) is identical with the convex closed envelope of A.
d) If K is a compact convex set in E, and A the set of its extremal points show that K = r o(A)
(use exerc. 22. b) of II, p. 90, and exerc. 32).
e) With the notations of II, p. 74, exerc. 3,Iet A be the set formed of the Ex' where x varies in
the set of rational numbers such that 0 :( x :( I. Show that r oCA) is distinct from the convex
envelope of A and from the convex closed envelope of A.
§7 EXERCISES TVS n.93

liT 34) Let S be a closed convex set of E, a Hausdorff locally convex space, and A a subsct
of S such that S = [' o(A) (exerc. 33), and let So be the convex envelope of A (so that S = So).
Let N be a closed convex subset of E containing a closed linear variety offinite codimension,
M =S n Nand Mo = So n N.
a) Show that M = Mo. (N ote, using exerc. 33, a), that every closed linear variety of finite
codimension in E. containing a point x E M, meet Mo, and use the prop. 4 of 11, p. 38.)
b) Suppose that for every finite subset F of A, the intersection of N and of the facet (in S)
of each point of the convex envelope of F, is compact or of finite dimension and does not
contain a line. Show then that M is the convex closed envelope of the set of its extremal points.
(By the aid of a), this reduces to proving that every point of Mo is contained in the convex
closed envelope ofthe set of extremal points of M. Use exerc. 3. e) ofII, p. 87, the Krein-Milman
tho (II, p. 55) and exerc. 22, b) of II. p. 90.) Deduce that every closed support hyperplane of
M contains an extremal point of M.

35) Let 1 be a non enumerable set, write E = R(I) and E' = E x R. Denote the canonical
basis of E by (e.)'Ei and let s be the element (0, 1) of E'. Define a separating duality between
E and E', by (e" e B ) = B'B' (e" s) = 1 for all (1 E 1. Let C be the pointed cone Rt+ in E.
a) Show that the topologies induced on C by cr(E, E') and by the norm p(x) = I Ix.1 on E
coincide. 'Et
b) Show that the uniform structure induced on C by cr(E, E') is not metrisable.

36) Consider the space E = R(N), with the weak topology cr(R(Nl, RN); let C be the closed
convex cone in E formed by the points x = (x n) such that x" ~ 0 for all n.
a) Let x = (x n) be a point of C and let J be the finite set of integers n for which xn > 0; if
m is the number of elements of J, then let A be the set of the points Y = (Yn) of C such that
y" = 0 for n E J and L Ykx'; 1 :;;;; m. Show that A is a cap of C containing X.
kEJ
b) Show that there does not exist a cap B in C such that C is the union of the sets nB for n > O.
(Let p be the restriction of the gauge of B to C; p will be finite in C and, if (eJ is the cano-
nical basis of E, we have peen) > 0 for all n (II, p. 58, prop. 4), and the points zln) = en/p(e.)
belong to B ; but show that there exists z' E RN such that the sequence ( z(n), z'») is not bounded.)

37) Let F be the Banach space /1 (N) of summable sequences x = (x,,) of real numbers and
let E be the space of sequences y = (Yn) which tend to 0; we give F the weak topology cr(F, E)
where E and F are in separating duality using the form B(x, y) = L xnJ'n'
n
a) Let C be the convex cone in F formed of the points x = (x,,) such that "" ~ 0 for all 11.
Show that C is closed in F.
b) Let A be the set of the x = (x,,) E C such that L xn :;;;; I. Show that A is a cap of C, that
is metrisable for the topology induced by that of F~ and that C is the union of the sets nA for
n > O.
c) Show that a sole of C is not compact (such a set S would be the set of the x = (x,,) E C
such that L znxn = 1. where (zn) E E and zn > 0 for all n. If em = (Bmn)n~O' the points zn-1en
belong to S," but do not form a relatively compact set in F).
d) Show that C is not metrisable for the topology induced by that of F (use Baire's th., noting
that there is no point in A that is an interior point relative to the subspace C).

38) Let E be a Hausdodf locally convex space and X a compact convex set in E. Denote by
d(X) the set of continuous affine functions in X (not necessarily restrictions to X of conti-
n uous affine functions in E, cf. II, p. 78, exerc. 11). For every real valued functionj'that is bounded
above in X. put (x) = inf(/i(x)) for all x E X where h varies in the set of fL;nction of d(X)
such that h ~ I h

a) Show that I
is an upper semi-continuous concave function. If f itself is upper semiconti-
nuous and concave, then 1= f (ef II, p. 39, prop. 5).
TVS 11.94 CONVEX SETS AND LOCALLY CONVEX SPACES §7

b) Suppose that f is upper semi-continuous. Show that 1

is the lower envelope of the func-
tion g, when g varies in a set of functions that are continuous in X of which f is the lower
envelope.
c) For all x E X, write uf/x for the set of finite families f..l = ((f..lj' x)) where the Xj are points
of X, the f..lj are numbers ~ 0 satisfying I f..lj = I, such that x = I f..ljX j . For each real valued
j j
function f bounded above in X, put f'(x) = sup I f..lJ(x) for all x E X. Show that f' is
a concave function in X and that f' ~ J.
d) Suppose that f is continuous in X. Given E > 0, let (Uk) \ ";k";N be a covering of X by convex
open sets such that the relations x E Uk' Y E Uk imply the inequality If(x) - f(y)1 ~ E. Put
A\ =U\ and, for k > I, Ak = Ukn C(U I u U 1 U U 3 u ... u Uk-I)' Show that, for all XEX,
there exists a family f..l = ((f..lk' x k») of N terms belonging to .$/x with Xk E Uk for I ~ k ~ N
such that I f..lJ(x k ) ~f'(x) - 2E (if we have the inequality I),/c}) ~ flx) - E, group
k
the Yj belonging to the same Ak together).
e) Deduce from d) that when f is continuous then f' is upper semi-continuous and f' = /'
(If U is an ultrafilter on X finer than the filter of the neighbourhoods of a point x E X and
if f'(y) ~ r for all the points Y of a set belonging to U, show that f'(x) ~ r - 2E, by making
a family f..l y E ./1/)" correspond to each y, which satisfies condition a) and proceeding to the
limit through n.)

39) Let H be a closed hyperplane in a Hausdorlf locally convex space E, that does not con-
tain 0 and let S be a compact simplex contained in H (II, p. 71, exerc. 41).
a) Let C be the cone with vertex 0 generated by S. Show that if (X)iEI and (})jEJ are two finite
families of points of C such that I Xi = I Yj' then there exists a finite family (zij)(i.jl EI x J
I I
iEI JEI
of points of C such that Xi = Zij for all i E I and Yj = Zij for all j E J (argue by induction
jEJ iEI
to reduce to the case I = J = {I, 2}).
b) Let f be a convex, upper semi-continuous function in S. Show that the function 1 (defined
in ex ere. 38) is an affine function. (First reduce to the case where f is contLnuous by using
exerc. 38, b) and II. p. 81. exerc. 28. Next use the fact iff is continuous, then f =f' (exerc. 38,
e», and show that f' is convex using a) to bound f'CCi I Xl + Ci1X 1) above when:;(l ~ 0, Ci 1 ~ 0,
Ci\ + Ci 1 = 1.)

40) Let X be a compact convex set in E, a Hausdorlf locally convex space, and let f be an
upper semi-continuous function that is bounded below. Let g be a lower semi-continuous
concave function, such that g ~ f Show (with that notations of ex ere. 38) that g ~ (Reduce 1
to the case where inf(g(x) - f(x») > o. If (fo.) is a decreasing directed family of continuous
XEX

functions such that f = inf(fo.), show that then also inf(g(x) - fo.(x») ~ 0 for Ci ~ Cia, and
XEX

so reduce the problem to the case where f is continuous. Then use exerc. 38, e).)

41) Let S be a compact simplex contained in a Hausdorff locally convex space E (II. p. 71
exerc. 41), and f an upper semi-continuous convex function that is bounded below. Let g
be a lower semi-continuous function that is concave and such that g ~ f
a) If we write u = .1 v = - (- g) .1 then u and u are affine functions such that u ~ v (use
ex ere. 39, b) and exerc. 40).
b) Show that there exists an affine function h, continuous in X and such that f ~ h ~ g
(D. Edwards' th.). (We can replace f by u and g by v. Construct three sequences (u m ), (u m ),
(hrn) of affine functions such that in X, the function Urn is upper semi-continuous, the function
Urn is lower semi-continuous and the function hrn is continuous and

Use exerc. 29 of II, p. 81, for this.)

§8 EXERCISES TVS II.95

§8
1) Extend the results of exerc. 8 of II, p. 83 to the spaces <g(T; C) and to their finite dimensional
subspaces.
2) Show that, when z varies in the unit disc Izl :( 1 in C, then the convex cone generated by
the points (z, Z2, ... , zn) in the spaces cn, is the whole of the space C". (Note that there cannot
n
exist complex numbers Ck not all zero (1 :( k :( n) such that i?4( L ck e ki8 ) ~ 0 for 0 :( 0 :( 2 IT,
k~l

using the faet that f 2"

eki6 de = 0 for every integer k oj 0.)
o
3) For the topological vector spaces on H, the division ring of quaternions, give the defi-
nitions and properties corresponding to those of this paragraph.
 CHAPTER III

Spaces of continuous linear mappings

In this chapter, all the vector spaces under consideration are vector spaces over a
field K, which may be R or C.
We recall (II, p. 2) that a semi-normed space is a vector space E endowed with a
semi-norm p and with the topology defined by p. Let r be a real number > O. The
set of all x E E such that p(x) :s; r is called the ball (closed) of radius r of E (or of p).
When r = 1, this ball is also called the unit ball.

§ 1. BORNOLOGY IN A TOPOLOGICAL VECTOR SPACE

1. Bornologies

DEFINITION 1. - A bomology on a set E is a subset \E of the set of all subsets of E

satisfying the following conditions (cf GT, X, § 1 .2, Remark 2).
(Bl) Ever,V subset of a set of \E belongs to \E.
(B2) Every finite union of a set of\E belongs to \E.
We say that \E is covering if every element of E is contained in a set which belongs
to \E, or. which is the same, if \E is a cover of E.

Example. - Let E be a metric space; the set of all bounded subsets of E (GT, IX,
§ 2, No.2) is a covering bomology on E. Let G be the group of isometries of E ; the set
of all subsets M of G such that for every x E E, the set M. x is a bounded subset
of E, is a covering bomology on G.

If \E is a bomology on a set E, a subset \E 1 of \E is said to be a base of \E if every

set of \E is contained in a set of \E 1 .
The intersection of a family of bomologies on E is a bomology; consequently
for every subset 6 of I.]3(E), there exists a smallest bomology containing 6; this
bomology is said to be generated by 6 and admits as a base the set of finite unions
of sets of 6. If E and E' are two sets, and \E (resp. \E') a bomology on E (resp. E'),
TVS 111.2 SPACES OF CONTINUOUS LINEAR MAPPINGS §I

the product bomology is the bomology on E x E' which admits the sets M x M'
as a base, where M E ~ and M' E ~'.

DEFINITION 2. ~ Let E be a vector space. A bornology ~ on E is said to be convex,

if for everyX E ~ and t E K, the homothetic set tX and the convex balanced envelope
reX) (II, p. 10) of X belong to ~.

If X and Yare two subsets of E, we have

X +Y c 2reX u Y)

AX c treX) for IAI:( t.

Consequently, if ~ is a convex bomology on E, if A is a bounded subset of K
and if X, Y belong to ~, then X + Y E ~ and A.X E~.

2. Bounded subsets of a topological vector space

DEFINITION 3. ~ Let E be a topological vector space. A subset A of E is said to be

bounded if it is absorbed by every neighbourhood 0[0 in E (I, p. 7, def 4).
In order that A be bounded, it is sufficient that A be absorbed by every neigh-
bourhood of a fundamental system of neighbourhoods of O. Since there exists a
fundamental system of balanced neighbourhoods of 0 (I, p. 7, prop. 4), this is the
same as saying that, for every neighbourhood V of 0 in E, there exists A E K such
that A c AV.
Suppose the topology of E is defined by a fundamental system r of semi-norms
(II, p. 3); then a subset A of E is bounded if and only if every semi-norm pEr is
bounded on A.
In particular, if E is a semi-normed space, a subset A of E is bounded if and only
if it is contained in a ball. In other words, if E is normed this means that A is bounded
for the metric space structure of E (GT, IX, § 2, No.2).
Remarks. - 1) If E is a semi-nnfmed space, the balls form a fundamental system
of bounded neighbourhoods of 0 in E. Conversely, if E is a locally convex topological
vector space, and if there exists a bounded neighbourhood of 0 in E, this neighbourhood
contains a convex balanced neighbourhood W, and the gauge of W is then a semi-
norm defining the topology of E.
Thus, if E is locally convex and metrizable, and if its topology cannot be defined

z
by a single norm, then there exists no distance on E defining its topology and such
that the bounded subsets for d (GT, IX, § 2, No.2) are the bounded subsets of E. More
precisely, for every distance d on E, which is translation invariant and which defines
the topology of E, the bounded subsets of E are bounded for d (III, p. 38, exerc. 3),
but the converse is false.
2) Let M be a vector subspace of E endowed with the induced topology. In order
that a subset of M be bounded in M, it is necessary and sufficient that it be bounded
in E.
3) Let N be the intersection of all neighbourhoods of 0 in E, so that E = E/N is the
Hausdorff vector space associated with E. Then N is bounded; if n: E --> E is the canoni-
cal homomorphism then a subset B of E is bounded if and only if nCB) is bounded.
4) Let E be a Hausdorff locally convex space; then for every x # 0 in E, there exists
No.2 BORNOLOGY IN A TOPOLOGICAL VECTOR SPACE TVS III.3

a continuous semi-norm p such that p(x) of 0; this semi-norm is not bounded on the
real half-line R+. x generated by x. Hence no non-null subspace of E is bounded. In
particular, a bounded subset does not contain any line.

DEFINITION 4. - Let E be a locally convex space. A bornology 5B on E is said to be

adapted to E, if it is convex, is composed of bounded subsets of E and if the closure of
every set of 5B belongs to 5B.

PROPOSITION 1. -
Let E be a locally convex space. The set of bounded subsets ofE
is an adapted bomology.
We need to establish the following properties :
a) If B is a bounded subset of E, every subset of B is bounded.
b) The union of two bounded subsets is bounded.
c) Every set that is homothetic to a bounded set is bounded.
d) The closed convex balanced envelope (II, p. 13) of a bounded subset is bounded.
If p is a continuous semi-norm on E, the balls of p are convex, balanced, closed
and the set homothetic to a ball is a ball. Hence, if p is bounded on two subsets X
and Y of E, it is also bounded on the closed convex balanced envelope of Xu Y,
and on the sets homothetic to these. This establishes properties b), c) and d), and a)
is obvious.

DEFINITION 5. - Let E be a locally convex space. The set of all bounded subsets
of E is called the canonical bomology of E.

If 5B is a set of bounded subsets of E, then there exists a smallest bomology m

adapted to E and containing 5B. The sets of m
are those that are contained in a
set homothetic to the closed convex balanced envelope of a finite union of sets of 5B.
Every adapted bomology is contained in the canonical bomology.

PROPOSITION 2. - In a locally convex space E, every precompact set is bounded.

Let A be a precompact subset of E, and V be a convex balanced neighbourhood
of O. There exists a finite sequence (a)t ";i";n of points of A such that

A c U (a i + V).
1 ~i~n

Since B = {a p ... , an} is bounded, there exists a scalar Ie such that 0 < Ie < 1
and leB c V; we have leA c leB + Ie V c V + V, from which the proposition
follows.

COROLLARY. - In a locally convex space, the set of points of a Cauchy sequence is

bounded.
In fact, this set is precompact (GT, II, § 4, No.2).

Remark 5. - In general the bounded subsets of a locally convex space E are not all
precompact (for example, if E is an infinite dimensional normed space, its unit ball
is not compact (I, p. 15, tho 3)). However, this is so if E is a weak space (II. p. 42) : for
TVS I1I.4 SPACES OF CONTINUOUS LINEAR MAPPINGS §1

the Hausdorff topological vector space associated with E is then isomorphic to a sub-
space of a product KI whose bounded subsets are precompact (cf III, p. 4, cor. 2).
For other examples, see IV, p. 18.

PROPOSITION 3. - Let A be a subset of a locally convex space E. Suppose that A

is bounded; then for every sequence (x,,) of points of A and for every sequence (A,,)
of scalars tending to 0, the sequence (A"x,,) tends to 0. Conversely, if there exists a
sequence (A,,) of non-zero scalars such that for every sequence (x,,) of points of A, the
sequence (Anx,,) is bounded, then A is bounded.
Suppose that A is bounded. If (An) is a sequence of scalars tending to 0, and V is
a neighbourhood of 0, we have AnA c V whenever n is large enough, and the first
assertion follows.
Conversely, if A is not bounded and if (A,,) is a sequence of scalars "# 0, then there
exists a continuous semi-norm p and a sequence (xn ) of points of A, such that
p(xn) ~ I~n I . We have then that P(Anxn) ~ n, and the sequence (A"x n) is not bounded.

COROLLARY. - A subset A ofE is bounded if and only if every countable subset of A

is bounded.

3. Image under a continuous mapping

° °
PROPOSITION 4. - Let E and F be two locally convex spaces and f:E ~ F a conti-
nuous mapping. Assume that f(O) = and that there exists a real number m ~
such that f(b) = Amf(x) for every A > 0. Then, if A is a bounded subset ofE, f(A)
is bounded in F.
°°
In fact, if V is a neighbourhood of in F, then f - 1(V) is a neighbourhood of
in E. If A is bounded in E, there exists A > such that A c Aj-1(V) and this implies
°
that f(A) c Amv.

COROLLARY 1. - Let E and F be two locally convex spaces, and u : E ~ F be a conti-

nous linear mapping. If A is a bounded subset ofE, then u(A) is bounded in F.

COROLLARY 2. - Let E = n Ei be the product of a family of locally convex spaces.

ieI
Then a subset of E is bounded if and only if all its projections are bounded.
M ore generally :

CoROLLARY 3. - Let E be a vector space, (F)iEi a family of locally convex spaces

and J; a linear mapping from E into Fi (for i E I). Suppose E is assigned the coarsest
topology (locally convex) for which all the J; are continuous (II, p. 26). Then, for a
subset A of E to be bounded, it is necessary and sufficient that J;(A) is bounded in Fi
for all i E I.
In fact, if A is bounded, so are the J;(A)(cor. 1). Conversely, if the J;(A) are bounded
and if p is a continuous semi-norm on E, then there exists a finite subset J of I and
No.4 BORNOLOGY IN A TOPOLOGICAL VECTOR SPACE TVS 111.5

a family (q.) 'EJ' where qj is a continuous semi-norm on F j , such that p :S; sup (q i 0 J)
J J jEJ .

consequently p is bounded on A.

COROLLARY 4. - Let E j (1 :S; i :S; n) and F he locally convex spaces, and f he a

n
continuous multilinear map from TI Ei into F. If Bi is a bounded sunset of Ei' for
;~ 1
II

I :S; i :S; n, then I( TI B;} is bounded in F.

i~ 1

COROLLARY 5. - Let E and F he two locally convex spaces and u: E --> F he a conti-
nuouS polynomial mapping. If A is a hounded suhset of E, then u(A) is bounded.

4. Bounded subsets in certain inductive limits

PROPOSITION 5. - Let (E)iEI he a family of Hausdorff locally comex spaces, and
let E he the rop%gical direct sum of this family (II. p. 29). In order that a subset B
of E he hounded, it is necessary and sufficient that there exists a finite suhset 1 of I
such that prj(B) is bounded in E; for i Eland pri(B) c {O} for aff i i J.
Let 1 be a finite subset of 1. Since the topology of E induces the product topology
onTI Ej(II, p. 30, prop. 7 and p. 31, prop. 8), it follows from III, p. 4, cor. 2 that the con-
jE]
dition is sufficient.
Conversely, let B be a bounded subset of E. Then pri(B) is bounded for all i (TTl,
p. 4, cor. 1). Therefore it is enough to prove that there exists a finite subset J of I
such that pri(B) c {O} for all i i 1. If not, then there exists an infinite sequence (iJ
of distinct elements of I and an infinite sequence (XII) of clements of B such that
pro (XII) i= O. Since E j is Hausdorff, there exists a continuous semi-norm p on E.
su~h that Pn(pri.,(xn )) '~ n. Hence p = L PII 0 prin is a continuous semi-n'~rm o~
fJ ~ 1
E and p is not bounded on B, which is a contradiction.

PROPOSITION 6. - Let E he a locafly convex space which is the strict inductive limit
of an increasing sequence (Ell) of closed vector subspaces of E (II, p. 33). A subset B
ofE is bounded!f and only if it is contained in one of the subspaces En' and is bounded
in this subspace.
The condition is sufficient, since the topology induced on Ell by that of E is preci-
sely the given topology of Ell (II, p. 32, prop. 9). To see that the condition is necessary,
it is enough (III, p. 4. prop. 3) to prove that if a sequence (xm) of points of E is not
contained in any of the subspaces Ell' then it cannot tend to O. By extracting a sub-
sequence of the sequence (x m ), we can assume that there exists a strictly increasing
sequence (n k ) of integers such that, for every index k. we have x k i Ellk and X k E Ellk + 1 •
Then there exists (II, p. 33, lemma 2) an increasing sequence (V k ) of convex sets
such thatV k is a neighbourhood 01'0 in E llk , Vk+l n Ellk = V k and such that x k i Vk+l
for every index k. The union V of the V k is then a neighbourhood of 0 in E, and we
have X k i V for all k. This proves that the sequence (x k ) does not tend to O.
TVS III.6 SPACES OF CONTINUOUS LINEAR MAPPINGS §1

The conclusion of prop. 6 is not necessarily true for a space E which is the inductive
limit of a non-denumerable directed set of closed subspaces ofE (rr III, p. 38, exerc. 7).

PROPOSITION 7. - Let (En)n;"O be a sequence of Hausdorff locally convex spaces,

and for every 11, let Un: En --+ En + 1 be an injective linear mapping which is compact
(i.e. such that There exists a neighbourhood of 0 in En whose image under Un is relatively
compact; this implies That un is continuous). Let E be the inductive limit of the system
(En' un) (11, p.29), and let vn be the canonical mapping from En into E. Then the locally
convex .Ipace E is Hausdorff. Moreover, for every subset A ofE, the following condi-
tions are equivalent :
(i) A is bounded;
(ii) there exists an integer n such that A is the image under Vn of a bounded subset
of En ;
(iii) A is relatively compact.
We identify En with a vector subspace of E (endowed with a topology finer than
the induced topology).

Lemma 1. - Under the hypothesis of prop. 7, the topology olE is thefinest topology
for which all the mappings L'n : En --+ E are continuous.
We need to prove that, if U is a subset of E such that U n En is open in En for
every n, then U is open in E; in other words, we must prove that, for every x E U,
there exists a convex balanced set V such that x + V c U and that V n En is a
neighbourhood of 0 in En for every large enough n (II, p. 27, prop. 5). For every n,
let Wn be a convex balanced neighbourhood of 0 in En such that the closure Hn
of W n in En + 1 is compact. Let x E U and let no be an integer such that x E E"o. We
shall construct, by induction, a sequence (En)n;"O of scalars > 0 such that
x + L GjH j is contained in U for n ~ no. Suppose that the Ei for i < n have
no:::;;' i:::S n
been constructed. If Il = no, set Vn- 1 = {O}; if not, set

Vn - 1 = L
no~i~I1-1
E;Hi·

Then V" _ 1 is compact in En' and a fortiori in En + 1 . Since U n En + 1 is open in En + 1 ,

there exists a scalar En > 0 such that x + Vn = X + Vn - 1 + EnHn is contained
in U (GT, II, § 4, No.3, cor.). Let V = U Vn. Then V is convex and baJanced; for
n~no

II we have V n En ::J EnHn n En ::J E"W n, hence V n En is a neighbourhood

~ /lo,
of 0 in En. This completes the proof of the lemma.
The set U = E - {O} is such that the set U n En = En - {O} is open in En
for every n, hence U is open in E, which proves that E is Hausdorff(GT, Ill, § 1, No.3,
prop. 2). It is clear that property (ii) implies (iii) and that (iii) implies (i). Finally
we show that (i) implies (ii). For this it is enough to show that if a subset A of E is
not absorbed by any of the sets L Hi' then A is not bounded. But then there
O~i::::;'n

exists a sequence (Xn)n;" 1 of points of A such that, for every n, we have xn rt n 2 L Hi.
O~i~n
No.5 BORNOLOGY IN A TOPOLOGICAL VECTOR SPACE TVS I1I.7

Then the set of the x,,/n is closed by virtue of lemma 1, since its intersection with Em
is discrete for every integer 111. The complement of the set of the .\)n is then an open
neighbourhood of 0 which docs not absorb the sequence (x). hence A is not bounded.

Remarks. - I) With the above notations, let F" be the vector space generated by
H", with a norm equal to the gauge of H II . We shall see (III, p. 8. cor.) that F" is a
Banach space. The injection from FII into Ell + 1 is compact, hence afortiori also the
injection w" from F" into F,,+ l' Further, E is the inductive limit of the inductive system
(F", wII ) of Banach spaces. For, a convex balanced neighbourhood V of 0 in E is
such that V (\ Ell absorbs H,,-l for all n ~ 1, and conversely, if a convex balanced
set W in E is such that W (\ E" absorbs H,,_ l' then W (\ E,,_ 1 contains a set homo-
the tic to W n - 1 for all n ~ 1, and hence W is a neighbourhood of 0 in E.
2) Let F be a locally convex space, k an integer ~ 0 and f: Ek -+ F a
multilinear mapping. For I to be continuous, it is necessary and sufficient
that the restriction of f to E~ is continuous for every n. We verify imme-
diately that Ek has the final locally convex topology for the family of linear
mappings 1"" x ... x v":E,, x ... x E" -+ Ex'" x E (II, p. 28, cor. 2 and p. 30,
prop. 7) and that u" x ... X u" is a compact injective linear mapping from (E,/
into (E,,+ l)k. It is now enough to apply lemma 1.

5. The spaces EA (A bounded)

Let E be a locally convex space and A be a convex balanced subset of E. We

recall that EA denotes the vector space generated by A, with PA the gauge of A,
as the semi-norm (II, p. 26, Example 3). We verify immediately that the canonical
injection of EA into E is continuous if and only if A is bounded. If, in addition, E is
Hausdorff, a bounded set A does not contain a line (Ill, p. 2, Remark 4) and so PA
is a norm (II, loc. cit.).
We shall say that a uniform space X is semi-complete if every Cauchy sequence
in X is convergent. A complete uniform space is semi-complete; but the converse
is not always true (GT, II, § 4, exerc. 4) ; however, a metrizable semi-complete space
is complete (GT, IX, § 2, No.6, prop. 9).

PROPOSITION 8. - Let E be a HausdOlfJ locally convex space and let A be a closed,

balanced, bounded and convex subset ()fE. Let (x,,) he a Cauchy sequence in EA' Then
this sequence converges in EA if and only if it converges in E.
The canonical injection from EA into E is continuous. Hence, if (x) converges
in E A, it converges in E. Conversely, suppose (XII) converges to y in E. There exists
an increasing sequence of integers (n k ) such that PA(X m - x,,) ~ 2- k - 1 if m ~ nk
and n ~ nk . Therefore the sequence (Xllk + rkA) is decreasing. Since A is closed
in E, we have ),En(x"k + 2- k A), which proves that (XIIJ, hence (:x), converges
k
to y in EA'
TVS I1L8 SPACES OF CONTINUOUS LINEAR MAPPINGS §1

CoROLLARY. - If A is semi-complete (in particular, complete) then EA is a Banach

space.
In fact, a Cauchy sequence in EA is also a Cauchy sequence for the topology of
E and is contained in a set homothetic to A, hence converges in E.

6. Complete bounded sets and quasi-complete spaces

DEFINITION 6. - A locally convex space E is said to be quasi-complete if every closed

and bounded subset ofE is complete (for the uniform structure induced by that of E).
A complete locally convex space is quasi-complete, but the converse is not always
true. * For example, if E is an infinite dimensional Hilbert space, or more generally,
an infinite dimensional reflexive Banach space, then E with its weakened topology
is quasi-complete but not complete (II, p. 51, prop. 9). *
A quasi-complete space is semi-complete, since every Cauchy sequence is con-
tained in a closed and bounded subset (III, p. 3, corollary and prop. 1). In particular,
a locally convex metrizable and quasi-complete space is complete.
In a Hausdorff quasi-complete space, the closure and the closed convex balanced
envelope of a precompact subset are compact; in fact, they are precompact (II,
p. 25, prop. 3), and complete being closed and bounded (III, p. 3, prop. 2).

PROPOSITION 9. - (i) A closed vector subspace of a quaSi-complete locally convex

space is quasi-complete.
(ii) The product of quasi-complete locally convex spaces is quasi-complete.
(iii) The topological direct sum of quasi-complete locally convex spaces is quasi-
complete.
(iv) A locally convex space which is the strict inductive limit of a sequence of closed
quasi-complete subspaces is quasi-complete.
Assertion (i) follows from Remark 2 (III, p. 2), (ii) from III, p. 4, cor. 2, (iii) from
prop. 5 (III, p. 5) and (iv) from prop. 6 (III, p. 5).

We shall see later that the quotient space of a quasi-complete locally convex space
by a closed vector space is not necessarily quasi-complete (IV, p. 63, exerc. 10).

PROPOSITION 10. - Let E be a locally convex space, M a vector subspace of E such

that every point ofE is in the closure of a bounded subset ofM. Then every continuous
linear mapping f from M into a Hausdorff quasi-complete locally convex space F
uniquely extends to a continuous linear mapping from E into F.
The hypothesis implies that M is dense in E, hence f extends uniquely to a conti-
nuous linear mapping j from E into the completion F of F (GT, III, § 3, No.4, corol-
lary). But every x E F lies in the closure ofa bounded subset B ofM; hence f(x) is in the
closure of f(B) in F. But f(B) is bounded in F, hence its closure in F is complete,
and coincides with its closure in F. This proves that j(x) E F.
No. 7 BORNOLOGY IN A TOPOLOGICAL VECTOR SPACE TVS III.9

7. Examples

a) Let X be a topological space. Let ~(X), the vector space of numerical (finite)
functions on X be assigned the topology of compact convergence (OT, X, S 1, No.3) :
this is the coarsest topology for which the restriction mappings .J1'(X) -+ ~(H)
are continuous (where H runs through the family of compact subsets of X and
where ~(H) is assigned the topology of uniform convergence). Cor. 3 of Ill, p. 4 shows
that a subset A of ,~(X) is bounded if and only if, for every compact subset H of X,
the set of restrictions to H of functions belonging to A is uniformly bounded.
* b) (Spaces of infinitely differentiable functions.) Let n ~ 1 be an integer. For
every open set U in R/I, let (6'''XJ(U) denote the vector space of infinitely differentiable
functions on U(VAR, R, 2. 3). Letfbein ~Xl(U). For every multi-index ex=(ex 1 .... , ex/l)
in Nil, 8"fdenotes the partial derivative 81"1f!8x~' ... ?X~'; this is a continuous function
on U (V AR, R, 2.3 and 2.4). For every integer In ~ 0, and every compact subset
H of U, set

(1) Pm.if) = sup 18~f(x)l·

1"I:;;m
XEH

Then Pm,H is a semi-norm on ~ Y(U).

Let !{;' X(U) be assigned the topology defined by the semi-norms Pm,H' This is
the coarsest of the topologies for which the mappings f --+ 8~l from ~ oo(U) into
.~(U) are continuous, where ~(U) is assigned the topology of compact convergence.
There exists an increasing sequence of compact subsets (H)/I '" 0 of U whose interiors
cover U; the family of semi-norms Pm.H n defines the topology of ~ oo(U) , which then
becomes a locally convex metrizable space. The space Cf3 feU) is complete; in other
words, it is a Frechet space (II, p. 24) : in fact, let (1;J be a Cauchy sequence in ~ oo(U) ;
for every ex E Nil. the sequence (8"t;J converges in the complete space 31>(U) (TO, X,
§ 1, No.5, tho 1) to a continuous function gao By induction on lexl, we deduce from tho 1
of FVR. 11, p. 2 that go = 8"go for every ex E Nil. In other words, the sequence (1;)
converges to go in ~ O0(U).

°
Let A be a subset of C(fX'(U). In order that A be bounded, it is necessary and suffi-
cient that the number sup Pm,HU) is finite for every integer m ~ and for every
fEA
compact subset H of U ; this condition means that for every ex E Nil, the sct of func-
tions a~llH for f E A is uniformly bounded for every compact H c U.
Let H cUbe compact. We denote by C(f~(U) the subspace of C(f OO(U) consisting
of those functions whose support lies in H. The space '6"(U) of infinitely differen-
tiable functions with compact support in U is the directed increasing union of
subspaces Cf3;;"(U) where H runs through the family of compact subsets of U. Each
space C(f ~(U) will be assigned the topology induced by that of (~W(U), and 'lJ oXc(U)
with the corresponding inductive limit topology. If the sets HII are such that their
interiors form a cover of U, then the space (g ooo(U) is the strict inductive limit of
TVS IILIO SPACES OF CONTINUOUS LINEAR MAPPINGS §1

the Frechet spaces ~.L.(U); it is therefore complete (II, p. 32, prop. 9) and every
bounded subset of XCV) is contained in one of the subspaces ~.L.(V) (I II, p. 5,
prop. 6). *
c) (Gevrey's spaces.) Let I be a compact interval in R. For every integer 11 ~ 0,
D,,! denotes the nth derivative of a numerical function f defined on I (whenever
this derivative exists). Let s ~ 1 and M ~ 0 be two real numbers. Let ~s.M(I) denote
the vector space of those infinitely differentiable functions f on I (FVR, I, p. 28)
for which the sequence (ID'11/M"(n !)S)1l '" 0 is bounded in the space ~(I) of all conti-
nuous functions on I (with the topology of uniform convergence). The space ~S.M(I)
is a Banach space with the norm

Ilflls.M = sup ID"f(x)I/M"(n!)s.

n ~ O,xEI

For M ~ M', we have ~l}s.\I(I) C ~s.M,(I), and

Ilflis,M' ~ IIIlls,M

for every IE ~s.M(I). Let ~lT) denote the union of the spaces ~,.:,iI) and endow
it with the inductive limit topology of the topologies of ~s.M(I).
Let M < M' and let B be the unit ball (closed) in ~s.\1(I). We will prove that B
is a compact subset of the Banach space ~s.rvl'(I). It is clear that B is closed in ~s.M,(I)
and so it is enough to prove that B is precompact in ~s.\I,(l), Let E > 0 and let N
be a positive integer such that (M/M,)N ~ E/2. Let k be a positive integer; the set
of all functions D k + if, as f ranges over B, is bounded in ~(I), hence the set of all
functions Dk{, as I ranges over B, is relatively compact in ~(I) : this follows from
the theorem of finite increments (FVR, I, p. 23, cor. 1) and from Ascoli's theorem
(GT, X, § 2, No.5). We define a norm on ~s.M(I) by

qeD = sup ID"I(x)I/M"'(n !)S .

O~n~N
xEl

The above argument shows that B is precompact for the topology associated with
the norm q; in other words, there exists a finite subset C of B such that for every
fEB, there exists gEe such that q(f - g) ~ E. Finally, for every n > N, we have

from which we get II f - gil ~ E. This proves that B is precompact in ~s.M(T),

The space ~s(I) is the inductive limit of the spaces ~sin as k ranges over N : by
prop. 7 (III, p. 6) every bounded subset of q/I) is contained in one of the spaces
~s.k(I) and is relatively compact in this space.
* d) (Spaces of holomorphic functions.) Let n ~ I be an integer. For every open
subset U of en, yt'(V) denotes the space of functions holomorphic in V, and is
assigned the topology of compact convergence in U. For every compact subset L
of C", ,iIf'(L) denotes the space of germs of hoi om orphic functions in a neighbourhood
No.7 BORNOLOGICAL SPACES TVS III.11

of L; we endow this space with the finest locally convex topology for which the
canonical mappings Te v :X(U) -+ X(L) are continuous, where U ranges over the
set of open neighbourhoods of L.
For every integer m ~ 1, let Urn be the set of points of C" which are at a distance
< 11m from L. It can be shown that the canonical mapping Te um from X(U m ) into
X(L) is injective, and that the restriction mapping from x(U rn ) into X(U p ) is
compact for p ~ m. We can then apply prop. 7 (III, p. 6). Let A be a bounded subset
of X(L); then there exists an integer m ~ 1 such that A consists of germs of func-
tions in a neighbourhood of L, belonging to a bounded set B in X(U m ). Moreover,
a mapping <!> from X(L) into a topological space T is continuous if and only if
the mapping <!> 0 Te u from X(U) into T is continuous for every open neighbourhood
U of L. *

§ 2. BORNOLOGICAL SPACES

In this paragraph, E denotes a locally convex space, and I.E its canonical bor-
nology (III, p. 3, def. 5).
Lemma 1. - Let G be a semi-normed space, p its semi-norm, and u a linear mapping
fi"om G into E. The following conditions are equivalent :
(i) u is continuous;
(ii) the image of the unit ball of Gunder u is bounded in E;
(iii) for every sequence (x") ofpoints ofG tending to 0, the sequence (u(x")) is bounded
in E.
It is immediate that (i) implies (ii) (III, p. 4, cor. 1) and that (ii) implies (iii). Let V
be a neighbourhood of 0 in E; if u- leV) is not a neighbourhood of 0 in G, then there
~. Hence the
exists a sequence (y") of points of G - u-l(V) such that p(y") :::;
n
sequence x" = ny" tends to 0 in G and u(x") ~ nY, which implies that the sequence
(u(x"») is not bounded. Therefore (iii) implies (i).

PROPOSITION 1. - The following conditions are equivalent :

(i) Every semi-norm on E which is bounded on bounded subsets ofE is continuous.
(i') Every convex balanced subset of E which absorbs the bounded subsets of E
(I, p. 7, def. 4) is a neighbourhood of 0 in E.
(ii) E is the inductive limit of the semi-normed spaces E A , where A ranges over the
directed increasing set of closed, convex, balanced and bounded subsets of E.
(ii') There exists a family (E)iEl of semi-normed spaces, and for every i E I, a linear
mapping u i : Ei -+ E such that the topology ofE is the Jinest locally convex topology for
which the ui are continuous.
(iii) For an arbitrary locally convex space F, a linear mapping u : E -+ F is continuous
if and only iffor every sequence (x,,) ofpoints in E tending to 0, the sequence (u(x,,») is
bounded in F.
TVS IIU2 SPACES OF CONTINUOUS LINEAR MAPPINGS §2

(iii') For an arbitrary semi-normed space F, a linear mapping u : E ---+ F is conti-

nuous if and only if u(X) is bounded in F for every bounded se t X in E.
It is immediate that (i) and (i') are equivalent in view of the correspondence
between semi-norms and convex, balanced, absorbent subsets (II, p. 20). If p is a
semi-norm on E, which is continuous on each E A , then p is bounded on bounded
subsets of E; hence (i) implies (ii) (II, p. 27, prop. 5). It is clear that (ii) implies (ii').
Now let (Ei' U)iEI be as in (ii') and let u be a linear mapping from E into a locally
convex space F, such that (u(x n)) is bounded in F for every sequence (XII) of points of
E tending to 0. It follows from lemma I ofIII, p. 11 that the linear mapping u 0 ui : Ei---+ F
is continuous for all i E 1. Hence, if the topology of E is the finest locally convex
topology for which the ui are continuous, then u is continuous (II, p. 27, prop. 5).
This proves that (ii') implies (iii).
It is immediate that (iii) implies (iii') (III, p. 3, cor.) Finally, ifp is a semi-norm on E,
which is bounded on bounded subsets ofE, the condition (iii') asserts that the identity
map is continuous from E into the semi-normed space (E, p); in other words, p is
continuous. This proves that (iii') implies (i).
DEFINITION 1. ~ A locally convex space is said to be bornological if it satisfies the
equivalent conditions of prop. 1.

Examples. ~ 1) Every semi-normed space is bomological.

2) In particular, every finite dimensional locally convex space is bomological.
3) On account of the transitivity of final locally convex topologies (II, p. 28, cor. 2),
we deduce at once from condition (ii') that if(EJiE' is a family oflocally convex bomo-
logical spaces and if E is assigned the finest locally convex topology for which the
linear mappings ui : Ei ---+ E (for i E 1) are continuous, then E is bomological. In
particular, all inductil'e limit, a direct sum, a quotient space olbornological spaces are
bornological spaces.
On the other hand, a closed subspace of a bomological space is not necessarily
bomological (IV, p. 63, exerc. 8).

COROLLARY . ~ Every Hausdorff and semi-complete bornological space is an inductive

limit of Banach spaces.
In fact, the spaces E A , where A is closed and bounded are Banach spaces (III,
p. 8, corollary).

PROPOSITION 2. ~ A locally convex metrizable space is bornological.

Suppose E is metrizable, and p a semi-norm on E which is bounded on bounded

subsets of E, but which is not continuous. Let A be the set of all x E E such that

°
p(x) < 1. Let (VII)n '" 1 be a decreasing sequence forming a fundamental system of
neighbourhoods of in E. Since p is not continuous, A is not a neighbourhood of 0;
hence for every n > 0, we have A:::p n-1V n and there exists a point xn in Vn' such
that n -1 Xn ¢ A, that is, p(xn) ~ n. The sequence (x,.) tends to 0, hence is bounded
(III, p. 3, corollary); this contradicts the hypothesis on p.
No. I SPACES OF CONTINUOUS LINEAR MAPPINGS TVS III.13

COROLLARY. - Every Frechet c\pace (II, p. 24) is the inductive limit of Banach spaces.

§ 3. SPACES OF CONTINUOUS LINEAR MAPPINGS

1. The spaces 23 (E ; F)

Let F be a topological vector space, E an arbitrary set, and 6 a family of subsets

of E. Consider the vector space FE with the uniform structure of 6-convergence
(GT, X, § 1, No.2). We know that this structure is compatible with the commutative
group structure of FE (GT, X, § 1, No.4, cor. 2). The topology so deduced is called the
6-topology. If X is a subset of FE, or more generally, a set with a mappingj: X --+ FE,
then the unverse image underj of the 6-topology on FE is called the 6-topology on X.

Remarks. - 1) The 6-topology is identical with the 6'-topology, where 6' denotes
the bomology generated by 6 (III, p. 1).
2) Let M E 6 and let V be a neighbourhood of 0 in F; let T(M, V) denote the set
of all f E FE such that f(x) E V for every x E M. If 6 is stable under finite unions,
the sets T(M, V) form a fundamental system of neighbourhoods of 0 for the 6-
topology of FE.

PROPOSITION 1. - Let E be a set, 6 a family of'subsets of' E, F a topological rector

space and H a rector subspace of FE. In order that the 6-topolog}' be compatible with
the vector space structure of H. it is necessary and sufficient that u(M) is bounded in F
for every u E H and every M E 6. 1/; moreover, F is locally convex, then the 6-topology
on H is locally convex.
On account of Remarks 1) and 2) above, we see that a necessary and sufficient
condition for the 6-topology to be compatible with the vector space structure of H
is that the sets H n T(M, V) are absorbent in H (I, p. 7, prop. 4); but this implies
that for every u E H, every subset M E 6, and every balanced neighbourhood V of 0
in F, there exists Ie i= 0 such that u(M) c leV; that is to say (III, p. 2) that u(M) is
bounded in F. Finally, the last assertion of the proposition follows from the fact
that if V is convex, so is T(M, V).

COROLLARY, - Let E and F be two locally convex spaces, 6 a family of bounded sub-
sets ofE, and 2(E; F) the vector space of continuous linear mappings from E into F.
Then the 6-topology is compatible with the vector space structure of 2 (E; F) and
is locally convex.
I t is enough to remark that if u is a continuous linear mapping from E into F and M
is a bounded subset of E, then u(M) is bounded in F (III, p. 4, cor. 1).

Given two locally convex vector spaces E and F, and a family 6 of bounded
subsets of E, let 23 (E; F) denote the locally convex space obtained by assigning
the 6-topology to 2(E; F).
TVS III.14 SPACES OF CONTINUOUS LINEAR MAPPINGS §3

Examples. - 1) If is is the set of all finite subsets of E, then the is-topology is

the topology of simple convergence and the space !i'G (E; F) is also denoted by
!i'/E; F). A bounded subset of !i's(E; F) is called a simply bounded subset of !i' (E ; F).
2) If is is the set of compact (resp. precompact, compact convex) subsets, then the
is-topology is called the topology of compact (resp. precompact, compact convex)
convergence and the space !i'6(E; F) is also denoted by !i'c(E; F) (resp. !i'pJE; F),
!i'c/E; F)). (Cf IV, p. 48, exerc.7.)
3) If is is the set of all bounded subsets of E, we say that the is-topology is the
topology of bounded convergence and the space !i'G (E; F) is denoted by !i'b(E; F).
4) When F = K, the space !i'(E; F) is the dual E' of E. We denote by E'G, E~ etc.
the space !i' 6 (E; K), !i'sCE; K) etc. The space E~ (resp. E~) is called the weak dual
(resp. strong dual) of E. A bounded subset of E~ (resp. E~) is said to be weakly (resp.
strongly) bounded. We observe that the weak topology on E' is none other than
0(E', E) (II, p. 42).
When E = F, we denote by !i'(E), !i'G (E) etc. the space !i'(E; F), !i'G (E; F) etc.
Letp be a continuous semi-norm on F and M a bounded subset ofE. Let

(1) PM(U) = sup p(u(x)) .

XEM

It is immediate that PM is a semi-norm on !i'(E; F) and that if r is a fundamental

system of semi-norms on F, the family of semi-norms PM' where p ranges over rand
M ranges over a base for the bomology generated by 6, is a fundamental system of
semi-norms of !i'G (E; F).
In particular, if E and Fare semi-normed spaces, and if p (resp. q) denotes the
semi-norm of E (resp. F), then the topology of bounded convergence on !i'(E; F)
is defined by the semi-norm
(2) r(u) = sup q(u(x»)
p(x)~ 1

(ct: GT, X, § 3, No.2). When we consider !i'b(E; F) as a semi-normed space, we

shall always, unless the contrary is expressly stated, mean the semi-norm (2). If F
is a normed space, the semi-norm (2) is a norm.

Remarks. -- 3) Let A be a dense subset of the unit ball of E. In view of the continuity
of u, we also have
(3) r(u) = sup q(u(x» .
XEA

For example
(4) r(u) = sup q(u(x».
pix) < 1

Since we have u(tx) = tu(x) for t E R, we also have,

q(u(x»
(5) r(u) = sup q(uCx» = sup -(-)-
pix) ~ 1 pix) oF 0 Px
whenever p i= O.
4) The formula (2) shows that the map u f-+ r(u) is lower semi-continuous on !l'sCE; F).
No.3 SPACES OF CONTINUOUS LINEAR MAPPINGS TVS IIU5

PROPOSITION 2. - Let E and F be /11"0 locally COl11'ex spaces and let G be a sel of
bounded subsets of E.
1) The G-topologl' on 2(E; F) is identical with the '~-topology, where 6 denotes
the smallest adapted, bomoloxy (III, p. 3) on E which cOl1tains2:.
2) Suppose that {O} is I/O! dense in F and let 6' be another set of hounded subsets
of E. Then the G'-/opologr is coarser than the 6-topology if alld only if 6' c 6.
Let u E 2J(E: F). M E 2: and let p be a continuous semi-norm on F. Since p 0 u
is a continuous semi-norm on E, this is the same as saying that p 0 u is bounded
above by 1 on M or on the closed, convex balanced envelope M ofM; in other words,
we have PM = PM' Moreover, it is clear that we have PAM = APM for all A > 0 and
PMuM' = SUP(PM' PM')' from which the first assertion follows, since 6 has the set
of homothetics of the closed convex balanced envelopes of finite unions of sets of 6
as a base.
We now prove the second assertion: first, if F is the base field. it follows from the
definition that the 6-topology on E' = 2(E; F) has as a fundamental system of
neighbourhoods of 0, the set of polars of the sets of ~. Let A be a bounded subset
of E, whose polar A is a neighbourhood of 0 for the ~-topology; then there exists
C

a closed convex balanced set B E 6 such that A =:J B, and so A c BOo; but by
0

cor. 3 of II, p, 45, we have Bee = B, and hence A c B and A E ~. Therefore if 6'
is a set of bounded subsets of E, the G'-topology is coarser than the G-topology
on E' if and only if 6' c 6. The general case follows immediately, since if Y E F
is not in the closure of 0, we can verify that the mapping which makes fEE' cor-
respond to the mapping x ~ f(x) y is an isomorphism of the locally convex spaces
E'~ onto its image in 2 6 ,(E; F).

2. Condition for 22: (E ; F) to be Hausdorff

PROPOSITION 3.-- Lf!t E and F be two locally COIIl'ex spaces, F hf!ing assumed
Hausdorff, and let \3 be a f{zl11ily of bounded subsets of E. If I he union A ol the sets of 6
is total in E, then the space :t'e;CE; F) is Hausdorff.
Let Uo be a non-zero element of 2(E; F); since U o is continuous and A is total
in E, there exists an Xo in A such that uo(x o) i= 0, Since F is Hausdorff, there exists
a neighbourhood V of 0 in F such that uo(x o) if v. Let M E 6 be such that Xo E M,
Then the set U of all u E 2 (E; F) such that u(M) c V is a neighbourhood of 0 in
2(E; F), and we have U o if U, hence 2(E; F) is Hausdorff.
In particular, the following topologies on 2(E; F) arc Hausdorff whenever F is
Hausdorff : simple convergence, compact convergence, precompact or compact
convex, and bounded convergence.

3. Relations between :t' (E ; F) and 2 (E ; F)

Let E and F the two Hausdorff locally convex spaces, and suppose F is com-
plete; let E be the completion of E. Since every continuous linear mapping u from E
TVS III.16 SPACES OF CONTINUOUS LINEAR MAPPINGS §3

into F extends uniquely to a continuous linear mapping Ii from E into F, we can

identify the spaces 2'(E; F) and 2'(E; F) by the mapping u f--+ u. In addition, let
6 be a family of bounded subsets of E; the 6-topology on 2'(E; F) coincides with
the 6-topology on 2'(E; F) and also with the eo-topology, where 2; denotes the
family of closures in E of sets of 6.
For example, if E is normed, the topology of bounded convergence on 2'(E; F)
is identical with the topology of bounded convergence on 2'(E; F) : for, every
bounded subset of E is contained in the closure of a bounded subset of E. Since the
unit ball ofE is the closure of the unit ball ofE, it follows from formula (3) (III, p. 14)
that ifF is a Banach space, the map u f--+ Uis an isometry from 2'(E; F) onto 2'(E; F).
We observe that if E is not a normed space, then there may exist bounded subsets
ofE which are not contained in the closure of any bounded subset ofE (for example,
if E is the weak dual of an infinite dimensional Banach space); however, this is so
if E is metrizable and satisfies the first axiom of countability (III, p. 39, exerc. 16).

4. Equicontinuous subsets of 2' (E ; F)

Let E and F be two locally convex spaces. For a subset H of Y(E: F) to be equi-
continuous it is necessary and sufficient that it is equicontinuous at the point 0 in E
(I, p. 9, prop. 6); this implies that for every neighbourhood V of 0 in F, the set
n u- leV) is a neighbourhood of 0 in E; or that for every continuous semi-norm
UEH

p on F, the function sup (p 0 u) is a continuous semi-norm on E. Moreover (I, p. 5),

UEH

H is uniformly equicontinuous. We note that the convex balanced envelope of an

equicontinuous subset is equicontinuous, since if p is a continuous semi-norm on
F and H the convex balanced envelope of H, we have, for the U j in H, the inequality
po (I
lejuJ ~
j
I
IleJ(p 0 uJ hence sup(p 0 u) = sup(p 0 u).
j ~H ~H

Consequently, the family of equicontinuous subsets is a convex bomology on

2'(E; F)(JII, p. 2, def 2).

PROPOSITION 4. - Let E, F be two locally cOl1l'ex spaces, and F be Hausdorf.f: Let

the space FE of all mappings from E into F be assigned the topology of simple conver-
gence. Then
(i) The set of linear mappingsfrom E into F is closed in FE.
(ii) IfH is an equicontinuous subset of 2'(E; F), the closure H of H in FE is con-
tained in 2'(E; F) and is equicontinuous.
We know that His equicontinuous (GT, X, § 2, No.3, prop. 6). It remains to prove
the assertion (i). Let x, y be in E and Ie, f.! in K, and let A(x, y, Ie, f.!) be the set of all
u E FE such that

u(b + flY) - leu(x) - f.!u(y) = O.

This set is closed in FE since the mapping u f--+ u(x) from FE into F is continuous
No.4 SPACES OF CONTINUOUS LINEAR MAPPINGS TVS III.17

for every x E E and since F is Hausdorff. But the set of linear mappings from E into F
is equal to
n A(x, y, Ie, 11) .
x,y,J....,J.l

Thus this set is closed in FE.

COROLLARY 1. - For an equicontinuous subset H of 2(E; F) to be relative~v compact

in 2s(E; F), it is necessary and sufficient that for all x E E, the set H(x) of all u(x)
as u ranges over H, is relatively compact in F.
In fact, this condition is necessary and sufficient for H to be compact in FE (GT, I,
§ 9, No.5, cor.).

COROLLARY 2. - Every equicontinuous subset of the dual E' of E is relatively compact

for the weak topology O'(E', E) on E' (III, p. 14, Example 4).
For, ifH is an equicontinuous subset ofE', sup lui is a continuous semi-norm on E;
UEH

in particular, for every x E E, the set H(x) is bounded, hence relatively compact in
the field of scalars.

COROLLARY 3. - In the strong dual E~ of a semi-normed space E, every closed ball

is compact for the weak topology O'(E', E).
This ball is also closed for O'(E', E).

PROPOSITION 5. - Let E and F be two locally convex spaces and let T be a total subset
of E. The following uniform structures coincide on every equicontinuous subset H of
2(E; F) :
1) the uniform structure of simple convergence in T;
2) the uniform structure of simple convergence in E;
3) the uniform structure of convergence in the precompact subsets of E.
We recall (III, p. 15, prop. 2) that the S-topology on 2(E; F) coincides with the
6-topology, where 6 is the smallest bomology adapted to E and containing S.
In the statement of prop. 5, we can therefore replace the word « total» by « every-
where dense ». The proposition then follows from the general properties of equi-
continuous sets (GT, X, § 2, No.4, tho 1).
Examples. - * 1) Let J.l be the Lebesgue measure on R, and let E be the semi-normed
space 2 P(J.l) (1 :(; p < co) (INT, TV). For every numerical function f and every real
number h, let .f;. be the function x f--> f(x - h). Clearly the mapping f f--> J;. defines a
linear isometry from E onto itself. If f is continuous and has compact support, then fh
converges to f uniformly, hence also in the mean of order p, as h tends to O. Since the
set ff(R) of all continuous functions with compact support is dense in E, and the set
of linear isometries of E is equicontinuous, it follows from prop. 5 that for every fEE,
J;. converges in the mean of order p to f as h tends to O.
For p = 1, consider the Fourier transform, which associates to each f E 2 1(J.l) the
function J on R defined by

J(y) = f e- 2inxYf(x) dJ.l(x) .

TVS IIU8 SPACES OF CONTINUOUS LINEAR MAPPINGS §3

The set of linear forms .f f-> ley) is an equicontinuous subset of the dual of 5£' 1(J.l).
On the other hand, we know that the set T of all characteristic functions of closed
bounded intervals is a total subset of 5£' 1(J.l); and we verify easily that for all .f E T,
the Fourier transform I is a continuous function tending to zero at infinity. We deduce
that this is true for all .f E 5£' 1(J.l) (<< Riemann-Lebesgue theorem »).
The relation sup 1/(y)1 ~ Ilfii 1 shows that the map f f-> I is a continuous map
.v ER
from 5£'1 (J.l) into the space .~(R) of all bounded functions on R with the structure of
uniform convergence. Since lis continuous for all f E T, it follows that lis continuous
for all .f E U(J.l). The fact that ltends to zero at infinity follows from the fact that the
subspace '?fa(R) of all continuous functions tending to zero at infinity is closed in &i9(R).
2) Let E be the space of all continuous numerical functions on R endowed with the
topology of compact convergence. Let K be a compact subset of R and let (J.ln) be a
sequence of measures on R with support in K. Suppose IIJ.lnll ~ 1 for all n. The set of
the J.ln is then an equicontinuous subset of E'. Therefore, if for every function .f E E,
we have lim J.l,,(.f) = 0, the sequence of functions x f-> fei,xdJ.ln(t) converges to 0,
lI-tifj

uniformly on every compact subset of R (since the set of functions t f-> ei'x, as x ranges
over a compact subset of R, is compact in E). *

COROLLARY. - Suppose F is Hausdorff Let H be an equicontinuous subset of 2(E; F).

If a filter <D on H converges simply to a mapping U o from E into F, then U o is a continuous
linear mapping Fom E into F, and <D converges uniformly to U o on every precompact
subset of E.
The first assertion follows from prop. 4 (III, p. 16) and the second from prop. 5
(III, p. 17).
PROPOSITION 6. - Let H be an equicontinuous subset of 2(E; F). If F is metrizable
and (f there exists a countable total set in E, then the uniform structure on H of simple
convergence in E is metrizable. If in addition, there exists a countable total set in F,
then there exists a countable everywhere dense set in H (for the topology of uniform
convergence on compact subsets of E).
Let (an) be a total sequence in E. Then the mapping u f-+ (u(a n )) is an isomorphism
from 2(E; F), where 2(E; F) has the uniform structure of simple convergence
on the set of the an' onto a uniform subspace ofFN. IfF is metrizable (resp. metrizable
and satisfies the first axiom of countability) then this is also true for FN (GT, IX, § 2,
No.4, cor. 2 and § 2, No.8, corollary), and the proposition follows from prop. 5
(III, p. 17).

COROLLARY 1. - Let E be a locally convex metrizable space, and F a normed space.

Suppose that E and F both satisfy the first axiom of countability. Then 2(E; F) is the
union of a countable family of equicontinuous subsets and there exists a countable
set in 2(E; F) which is dense for the topology of uniform convergence on precompact
subsets of E.
Let B be the unit ball of F and (V n) a countable fundamental system of neighbour-
hoods of 0 in E. For every integer n, the set Hn of all u E 2(E; F) such that u(V n) c B
is equicontinuous and 2(E; F) is the union of the Hn. The corollary then follows
from prop. 6.
No.5 SPACES OF CONTINUOUS LINEAR MAPPI!"\GS TVS III.19

COROLLARY 2. - Every closed ball in the dual E' of a normed space satisfying the jirst
axiom of countability, is a compact metrizable space for the weak topology cr(E', E).
andfor this topology there exists a counfaMe dense subset in E'.
This follows from prop. 6 and from HI, p. 17, cor. 3.

5. Equicontinuous subsets of E'

In this section, E denotes a locally convex space and E' its dual. Whenever we
talk of the polar M of a set M in E (resp. E'), we shall always mean, unless otherwise
O

stated, the polar of M relative to the duality between E and E'. Recall that if Y is a
closed convex balanced neighbourhood of 0 in E, we have Y = Y (II, p. 45, cor. 3).
OO

PROPOSITION 7. - Let M be a subset of E'. The j()llowing conditions are equivalent:

(i) M is equicontinuous;
(ii) M is contained in the polar of a neighbourhood of 0 in E;
(iii) the polar of M is a neighbourhood of 0 in E.
If M is equicontinuous, there exists a convex balanced neighbourhood Y of 0
such that iu(x)i ~ 1 for all x E Y and all u EM; then we have that M c yo and (i)
implies (ii). With the same notations, if M c yo then Y c Y c Me and (ii) implies
OO

(iii). Finally, if M contains a convex balanced neighbourhood Y of 0, then

O

M c Me, c yo and the relations x E £Y, u EM imply iu(x)i ~ £ for all £ > 0,
which proves that (iii) implies (i).
We remark that every x E E defines a mapping j(x): u H u(x) from E' into K.
Hence we can talk of the 6-topology on E, whereS is a family of subsets of E' : this
is the inverse image under j of the 6-topology on KE'. We verify immediately that
if S is a convex bomology on E', then the po lars of sets ofS form a fundamental
system of neighbourhoods of 0 for the S-topology on E. This is so, in particular,
when 6 is the family of equicontinuous subsets of E' and prop. 7 implies:

COROLLAR y 1. - The topology of E is identical )vith the topology of uniform con ver-
gence on equicontinuous subsets of E'.
More generally, let F be a locally convex space; every u E 2'(E; F) defines a map
j(u): (x, f) H f(u(x)) from E x F' into K (i.e. into R or C). This enables us to define,
on the space 2'(E; F), the topology of uniform convergence on a set of subsets of
E x F'. Tn particular:

COROLLARY 2. - Let S be a family of bounded subsets of E. TheS-topology

on 2'(E; F) is the topology of uniform convergence on sets of the form A x BeE x F',
where A is in 6, and B belongs to the family of equiconrinuous subsets ofF'.
For every u E 2'(E; F), every A E 6 and every closed convex balanced neighbour-
hood Y of 0 in F, the relation u(A) c Y is equivalent to «j(u) (A x YO) is contained
in the unit ball of K ».

PROPOSITION 8. - Let H be a family of linear mappings from E into a locally convex

space F. For H to be equicontinuous, it is necessary and sujficient that for every equi-
TVS III.20 SPACES OF CONTINUOUS LINEAR MAPPINGS §3

continuous subset X in the dual F' of F, the set of linearforms f 0 u, for f E X and u E H,
is equicontinuous.
It is obvious that the condition is necessary. Suppose it is verified, and let V be
°
a closed convex balanced neighbourhood of in F. Since VO is equicontinuous,
there cxists a neighbourhood W oro in E such that If(u(x»1 :(: 1 for all x E W, U E H
and f E VO ; in other words, u(W) c v eo = V for all u E H, hence His equicontinuous.

6. The completion of a locally convex space

THEOREM 1 (Grothendieck).- Let E be a locally convcx 5pace, and let 6 be an adapted

and covering homology on E. Let F c E* be the space of those linear forms on E
\I'hose restriction 10 each set belonging to 6 is continuous. If F is assigned Ihe 2-
IOjiO!ogy, then tlze canonical injection Fom E'e: into F extends to an isomOlphismji'o/11
the completion E'e: of Fe: onto F.
Since every simple limit of linear forms on E is a linear form (Ill, p. 16, prop. 4)
and since the bornology 6 on E is covering, it follows from GT, X, ~ 1, No.6, cor. 2
that the space F with the 6-topology is Hausdorff and complete. It is clear that Es is a
topological vector subspace of F; hence it is enough to prove that E'e: is everywhere
dense in F. This follows from the following lemma:

Lemma 1.- Let A be a closed convex balanced subset of E and let u be a linear form
on E whose restriction to A is continuous. Then for every £ > 0, there exists an x' E E'
such that
lu(x) - <x, x'>1 :(: £ for every x EA.
°
Let £ > O. There exists a closed convex balanced neighbourhood U of in E such
that lu(x)1 :(: £ for all x E UnA. We know that the polar U ofU in E* is contained
in E' and is compact for the topology O'(E*, E) (III, p. 17, cor. 2). Since the polar
A of A in E* is closed for cr(E*, E), it follows that AC + U is a closed convex subset
C
C

of E* (GT, III, § 4, No. 1, cor. 1).

Let C be a closed convex balanced subset of E. Then C is closed for cr(E, E')
(II, p. 45, cor. 3), hence also for cr(E, E*), and as a consequence, we have C = Coo
(for the duality between E and E*). As a result, we have

from which, we get

(A n ur c (A C + uey" = A' +U .

Since the linear form E - 1 U belongs to (A n U). there exist z; E A 0 and W E U C such
that u = lO(u + w). Hence x' = lOW belongs to E' and u - x' = lOU is bounded above
in absolute value by £ on A; hence the lemma.
Now let E be a locally convex Hausdorff space and E its completion. Every conti-
nuous linear form f on E extends to E by continuity: hence we have (E)' = E'
NO.7 SPACES OF CONTINUOUS LINEAR MAPPINGS TVS III.21

(III, p. 16) and every element of E defines a linear form on E'; that is, an element of
the algebraic dual E'* of E'. In addition, the duality between E (resp. E) and E' is
separating (II, p. 24, cor. 1). Consequently E and E can be identified with vector
subspaces of E'*.

THEOREM 2. - Let E be a locally convex HausdOl:ff space and E its completion; we

identify E and E with vector subspaces of E'*. Then for an element f E E'* to belong to
E, it is necessary and sufficient that the restriction off to every equicontinuous subset
of E' is continuous for the topology a(E', E).
The space E can be identified with the topological dual of E' when E' is assigned
the topology a(E', E) (II, p. 43, prop. 3); on the other hand, if 6 is the set of equi-
continuous subsets ofE', the given topology on E is the6-topology (III, p. 19, cor. 1).
Then it follows from III, p. 13, prop. 1, that the sets of 6 are bounded for a(E', E)
(cl later on, III, p. 22, prop. 9); in other words, 6 is an adapted and covering bor-
nology for the topology a(E', E). Theorem 2 is then a consequence of tho 1 if we
replace E by E' and E'e; by E.

COROLLARY 1 (Banach). - Let E be a Hausdorff and complete locally convex space.

In order that a linear form on E' be continuous for the weak topology a(E', E) (i.e.
arises from an element of E) it is sufficient that its restriction to every equicontinuous
subset of E' is continuous for a(E', E).

Remark. - Suppose in addition, that there exists a countable total set in E; then
every equicontinuous subset of E' is metrizable for the topology a(E', E) (III, p. 18,
prop. 6); therefore to verify that a linear form u on E' is weakly continuous, it is
enough to verify that for every equicontinuous sequence (x;,) in E' which converges
to 0 for a(E', E), we have lim u(x;,) = O.
n~oc

COROLLARY 2. - Let (E);EI be a family of Hausdorff local(y convex spaces and let E
be their topological direct sum. Then the canonical mappingji'om the direct sum of the
E; into E is an isomorphism. In particular, E is complete if and only if all the E; are
complete.
We know that the dual of E can be identified with the product of the duals of the
Ei (II, p. 30. formula (1)). Let u E E, and let ui E E;* be the restriction of u (considered
as an element ofE'*) to E; c E'. It is immediate that it is enough to prove that ui = 0
except for a finite number of indices i E I. Suppose on the contrary that there exists
a sequence (in)n~N of distinct indices such that ui" #- O. Then there exists Xi" E E;"
such that ui,,(x;) = n. The set H of all Xi" is equicontinuous in E' and the restriction
of U to H is not bounded, which is impossible.

7. 6-bomologies on f/! (E ; F)

Let E and F be two locally convex spaces and 6 a family of bounded subsets of E.
To say that a subset H of f/!(E; F) is bounded for the 6-topology means that for
TVS m.22 SPACES OF CONTINUOUS LINEAR MAPPINGS §3

every M E 6, every neighbourhood V of 0 in F absorbs the set H(M) = U u(M);

this is the same as saying that for every M E 6, the set H(M) is bounded in F. Equi-
valently, this means that for every neighbourhood V of 0 in F, the set n u- 1 (V)
UEH

absorbs every subset M of 6.

PROPOSITION 9. - Let E andF be two locally convex spaces and 6 afamily of bounded
subsets of E. Then every equicontinuous subset of 2(E; F) is bounded for the 6-
topology.
For, if H is an equicontinuous subset of 2(E; F) and V a neighbourhood of 0
in F, the set n u- 1 (V) is a neighbourhood of 0 in E, hence absorbs every bounded
UEH

subset of E.
A subset of 2(E; F) which is bounded for a 6-topology is not necessarily equi-
continuous, even if 6 is covering and 6 is the canonical bomology on E (IV, p. 50,
exerc. 17). In the following paragraph we shall study, under the name barrelled
spaces, the spaces E such that every simply bounded subset of 2(E; F) is equiconti-
nuous. For the present note the following result :

PROPOSITION 10. - Let E be a bornological space (in particular, a metrizable locally

convex space) and F a locally convex space. Every subset H of 2(E; F) which is
bounded for the topology of bounded convergence is equicontinuous.
For every convex balanced neighbourhood V of 0 in F, the set n u- 1 (V) absorbs
UEH

every bounded subset of E, hence is a neighbourhood of 0 in E; this proves that H

is equicontinuous.

8. Complete subsets of 26 (E ; F)

PROPOSITION 11. -- Let E and F be two locally convex spaces, (S a cover of E con-
sist ing of bounded subsets. If F is Hausdorff and quasi-complete (III, p. 8), then every
equicontinuous subset H of 2(E; F) which is closedfor the 6-topology is a complete
un(form subspace of 26 (E; F).
Since H is bounded in 2 6 (E; F) (III, p. 22, prop. 9) and closed in FE for the 6-
topology (III, p. 16, prop. 4), this follows from cor. 3 ofGT, X, § 1, No.5.

Remark 1. - Let M be a complete uniform subspace of 26 (E; F). For every set
of bounded subsets 6 ' ::::> 5 of E, the 5 '-topology is finer than the 5-topology on
2(E; F); on the other hand, there exists a fundamental system of neighbourhoods
of 0 for the (S'-topology which are closed for the topology of simple convergence
(III, p. 13, Remark 2), and a fortiori for the 6-topology. We conclude (GT, III, § 3,
No.5, cor. 1) that M is complete for the (S'-topology.

COROLLARY. - Let E and F be two locally convex spaces, H an equicontinuous subset

of 2(E; F). If F is Hausdorff and quasi-complete and if a filter <II on H converges
No.8 BANACH-STEINHAUS THEOREM TVS III.23

simpfy at all points of a total subset T ofE, then there exists a eontinuous linear mapping
u from E into F such that <l> converges uniformly to u on evelT precompact subset of E.
For, by virtue of prop. 5 (Ill, p. 17) <D is a Cauchy filter for the uniform structure
of precompact convergence in E; by prop. 11, the closure H of H in £'pc(E; F) is
complete and so <l> converges uniformly on every precompact subset ofE to a mapping
uEH.

Remark 2. ~ Let (un) be a sequence of continuous linear mappings from a Banach

space E into a Banach space F; it may happen that (un(x») has a limit at every point
of an everywhere dense vector subspace T of E, without the sequence (un) being
bounded in the normed space 2(E; F). For example, take E to be the space of all
continuous numerical functions on R, tending to zero at infinity, with the norm
Ilfll = sup If(x)1 and let T be the subspace of continuous numerical functions
XER

with compact support. The sequence of continuous linear mappings f I-> nf(n)
from E into R converges to 0 for all f E T, but is not bounded in St;,(E; R). The
same example shows that in the space 2(T; R), a sequence (c,) may be simply
convergent and non-bounded for the topology of bounded convergence.
n
On the other hand. the seq uence of continuous linear mappings f I-> I f(k)
k= 1
is a Cauchy sequence in 2(T; R) for the topology of simple convergence, but does
not tend to a limit in 2(T; R) for this topology.

PROPOSITION 12. ~ Let E be a bomologicallocally convex space, F a complete locally

convex Hausdorff space and 6 a family of bounded subsets of E containing the image
of every sequence cOl1l'erging to O. Then the space 26 (E: F) is complete.
Let <D be a Cauchy filter in ~6 (E; F). Then <D is a Cauchy filter for the topology
of simple convergence, hence converges in FE; moreover, its limit u is a linear mapping
from E into F and <D convcrges to u uniformly on every set of 3 (GT, X, § 1, No.5,
prop. 5). It follows that the image under u of a sequence converging to zero is a sequen-
ce converging to zero. hence. that u is continuous, since E is bomological (III, p. 11,
prop. 1, (iii»).

COROLLARY 1. ~ The strong dual of a bomological space is complete.

COROLLARY 2. ~ Let E be a semi-normed space, and F a Banach (resp. Frechet)

space. The space 2 b (E; F) is a Banach (resp. Frechet) space. In particular, the dual
of a semi-normed space is a Banach space.

§ 4. THE BANACH-STEINHAUS THEOREM

In this paragraph E denotes a locally convex space and E' its dual. Whenever we
talk of the weak topology on E', we shall mean cr(E', E).
TVS III.24 SPACES OF CONTINUOUS LINEAR MAPPINGS §4

1. Barrels and barrelled spaces

PROPOSITION 1. - Let T be a subset of E. The following conditions are equivalent:

(i) T is convex, balanced, closed and absorbent.
(ii) T is the polar of a convex, balanced and weakly bounded set M in E'.
(iii) There exists a lower semi-continuous semi-norm p on E, such that T is the set
of all x E E satis/ving p(x) ~ 1.
(i) => (ii) : under the hypothesis of (i), let M = TO; then M is convex and balanced
in E'. For every x E E, there exists a real number I' > 0 such that rx E T, hence
lu(x)1 ~ ~ for all u EM; in other words M is weakly bounded. From cor. 3 of II,
p. 45, we have T = MO, hence T satisfies (ii).
(ii) => (iii) : under the hypothesis of (ii), let p(x) = sup lu(x)1 for all x E E. It is
UEM

immediate that T = MO consists of all x E E such that p(x) ~ 1. The semi-norm

p on E' is lower semi-continuous, being the superior envelope of a family of conti-
nuous functions (GT, IV, § 6, No.2, corollary).
(iii) => (i) : this is clear.

COROLLARY. - The following conditions are equivalent

(i) every weakly bounded subset of E' is equicontinuous;
(ii) every convex, balanced. closed and absorbent set in E is a neighbourhood of 0;
(iii) evel]! lower semi-continuous semi-norm on E is continuous.

DEFINITION 1. - A set T satisfying the equivalent conditions of prop. I is said to

be a barrel in E.

DEFINITION 2. - A space E is said to be barrelled if it satisfies the equivalent condi-

tions of the corollary of prop. 1.

We know (III, p. 22, prop. 9) that every equicontinuous subset of the dual E' of E
is strongly and weakly bounded. We can therefore restate the definition of barrelled
spaces as follows:
Scholium. - In the dual of a barrelled space, the class of equicontinuous sets, the class
of strongly bounded sets, the class of weakly bounded sets and the class of relative~v
compact sets for the weak topology are all identical. IfE is Hausdorff and barrelled,
and if E~ is its strong dual, the polars of the neighbourhoods of 0 in one of the spaces
form a base of the canonical bomology of the other, and the polars of bounded subsets
of one of the spaces form a base for the filter of neighbourhoods of 0 of the other space.

PROPOSITION 2. - Every locally convex space E which is a Baire space (GT, IX, § 5,
No.3) is barrelled.
Let T be a barrel in E; since T is absorbent, E is the union of closed sets nT (n
integer> 0); since E is a Baire space, at least one of these sets contains an interior
No.2 BANACH-STEINHAUS THEOREM TVS III.2S

point, hence T itself has an interior point x. If x =I- 0, since - x E T, and 0 is a point
of the open segment with extremities x and - x, 0 is an interior point of the convex
set T (II, p. 14, prop. 16). Therefore T is a neighbourhood of O.

COROLLARY. - Every Frechet space (and in particular, every Banach space) is barrelled.
This follows from Baire's theorem (GT, IX, § 5, No.3, tho 1).

PROPOSITION 3. - Let (F)iEI be a family of barrelled spaces, and for every i E I, let J;
be a linear mapping from Fi into a vector space E. The space E with the finest locally
convex topology for It'hich the J; are continuous (II, p. 27, prop. 5), is a barrelled space.
Let Tbe a barrel in E. Since J; is continuous, J;-l(T) is a convex, balanced, closed
and absorbent set in Fi ; in other words, a barrel in Fi ; since Fi is barrelled, J;-l(T)
is a neighbourhood of 0 in Fp for all i E I. This implies that T is a neighbourhood
of 0 in E (II, p. 27, prop. 5).

COROLLARY 1. - Every quotient space of a barrelled space is barrelled.

COROLLARY 2. - Let (E)iEI be a family of locally convex spaces and let E be the topo-
logical direct sum of thisfamily. For E to be barrelled, it is necessary and sufficient that
each Ei is barrelled.
The condition is evidently sufficient by virtue of prop. 3; it is also necessary, by
cor. 1, since each Ei is isomorphic to a quotient space ofE (II, p. 31, prop. 8).

COROLLARY 3. - Every inductive limit of barrelled spaces is a barrelled space.

We shall prove later (IV, p. 14, corollary) that every product of barrelled spaces
is barrelled.

2. The Banach-Steinhaus theorem

THEOREM 1. - Let E be a barrelled space, F a locally convex space. Every simply

bounded subset H of £l(E; F) is equicontinuous.
For, let p be a continuous semi-norm on F; let q = sup (p 0 u). Since H is simply
UEH

bounded, we have q(x) < + Cf) for all x E E and q is a lower semi-continuous semi-
norm, being the finite superior envelope of continuous semi-norms. Since E is
barrelled, q is a continuous semi-norm and therefore H is equicontinuous.

COROLLARY 1. - Let E and F be two Banach spaces, H afamily of continuous linear

mapping ji-01n E into F; ij~ for all x E E, we have sup II u(x) II < + 00, we also have
UEH

sup !Iull < + 00.

UEH

In fact, the hypothesis says that H is simply bounded and the conclusion that it is
equicontinuous. In addition, every Banach space is barrelled (III, p. 25).

COROLLARY 2. - (Banach-Steinhaus theorem). - Let E be a barrelled space. F a

locally convex Hausdorff space, and let (uJ be a sequence of continuous linear mappings
TVS III.26 SPACES OF CONTINUOUS LINEAR MAPPINGS §4

from E into F, which converges simply to a mapping ufrom E into F. Then u E 2(E; F),
and (un) converges to u uniformly on every precompact subset of E.
The sequence (un) is, in fact, simply bounded, hence equicontinuous, and the
corollary follows from the cor. of prop. 5 of III, p. 18.
Remarks. - 1) The property expressed by cor. 2 does not imply that E is barrelled :
we shall see later that the strong dual of a Frechet space possesses this property,
while not necessarily being barrelled (IV, p. 23, cor. to prop. 2 and p. 58, exer~
2) Let E and F be two Banach spaces, and (u,,) a sequence of continuous' linear
mappings from E into F such that sup Ilu,,11 = + CIJ. Then the set X of all x E E such
that sup I u,,(x) I = + CIJ is dense in E and is the intersection of a sequence of open sets
in E. For, let X k denote the set of all x E E such that sup I un(x) I > k (for k integer> 0).
Each X k is open and X is the intersection of the X k . Since E is a Baire space, it is enough
to show that each X k is dense in E. But, if the complement of X k contains a non-empty
open set U, we would have Ilun(x)11 ~ 2k for x E U - U and, since U - U is a neigh-
bourhood of 0, we would have sup I un I < + CIJ.

COROLLARY 3. - Let E be a barrelled space, F a locally convex Hausdorff space

and 1> a filter on 2 (E; F) which converges simply in E to a mapping u from E into F.
If 1> contains a Simply bounded subset of 2(E; F), or if 1> has a countable base, then
u is a continuous linear mapping from E into F and 1> converges uniformly to u on
every precompact subset of E.
Suppose first that 1> contains a simply bounded set H ; since H is equicontinuous
(th. 1), the corollary follows from the corollary of prop. 5 (III, p. 18). If 1> has a coun-
table base, every elementary filter lJ' associated with a sequence Un (GT, I, § 6, No.8)
which is finer than 1> is then simply convergent to u in E and it follows from cor. 2
that u is a continuous linear mapping from E into F, and that lJ' converges to u for
the topology of uniform convergence on precompact subsets of E. Consequently,
the same holds for 1>, since the latter is the intersection of elementary filters, each
finer than 1> (GT, I, § 6, No.8).
We observe that a filter on 2(E; F) which converges simply and has a countable
base does not necessarily contain a simply bounded set : to see this consider the
example of the filter of neighbourhoods of 0 in 2(K; F) when the topology of F
is metrizable, but cannot be defined by a single norm.
Example. - Let E be the Banach space (over C) consisting of continuous complex
functions with period 1 in R, with the norm II f I = sup I f(x)l.

f
x

For every integer n E Z and every function f EO E, let cn(f) = f(x) e-2irrnXdx

(n-th Fourier coefficient of f) ; each of the mappings If-> c,,(f) is a continuous linear
form on E. Let (cx,,) be a sequence of complex numbers such that, for every function
fEE, the serie with the general term cxncn(.f) + cx-ncn(.f) is convergent. Under these
conditions, the mapping u:f f-> cxoco(.f) + I [cx"c,,(.f) + cx_"cn(.f)] is a continuous
n~l

linear form on E; * in other words, there exists a measure 1.1 on [0, I] such that

u(f) = ff(X) dl.l(x) for every function lEE, and CX n is the n-th Fourier coefficient
m
of 1.1. * In fact, for every integer m > 0, the mapping f f-> I CXkC k ( f) is a continuous
k= -m
No.3 BANACH-STEINHAUS THEOREM TVS III.27

linear form urn on E, and for all fEE, the sequence (urn(f)) converges to u(f), by hypo-
thesis. The assertion then follows from Banach-Steinhaus theorem, since E is barrelled.

COROLLARY 4. - Let E and F be two locally convex spaces, 6 a cover of E consisting

of bounded subsets. If E is barrelled and F Hausdorff and quasi-complete, the space
£'6 (E; F) is Hausdorff and quasi-complete.
In fact, every bounded and closed subset of £'6 (E; F) is simply bounded (because
6 is a cover of E), hence equicontinuous (III, p. 25, tho 1) and consequently is a
complete subspace of £'6 (E; F) because of prop. 11 (III, p. 22).

COROLLARY 5. - The strong dual and the weak dual of a barrelled space are quasi-
complete.

3. Bounded subsets of £'(E; F) (quasi-complete case)

THEOREM 2. - Let E be a locally convex Hausdorff space, F a locally convex spaa

and 6 a family of closed, convex, balanced, bounded and semi-complete subsets of E
(III, p. 7). Every simply bounded subset H of £,(E; F) is bounded for the 6-topology.
Let A E 6. The space EA is then a Banach space (III, p. 8, corollary), hence bar-
relled. On the other hand, the canonical image of H in £'(E A; F) is simply bounded,
hence equicontinuous (III, p. 25, tho 1). Consequently, the set of all u(x) for u E H
and x E A, is bounded in F, which proves that H is bounded for the 6-topology.

COROLLARY 1. - Let E be a locally convex Hausdorff space, F a locally convex space,

and 6 afamity ofbounded subsets ofE. IfE is semi-complete, then every simply bounded
subset of £'(E; F) is bounded for the 6-topology.
It is enough to apply tho 2, after replacing the sets of 6 by their closed, convex,
balanced envelopes, since this does not change the 6-topology.
When E is semi-complete (for example quasi-complete), we can talk of the bounded
subsets of £'(E; F) without specifying the 6-topology, since these are the same for
all the 6-topologies whenever 6 is a cover of E.

COROLLARY 2. - Every semi-complete bornological space is barrelled

Every simply bounded subset of the dual of such a space is strongly bounded
(cor. 1), hence equicontinuous (III, p. 22, prop. 10).

COROLLARY 3. - Let E be a locally convex space. Every subset of E which is bounded

for a(E, E') is bounded.
Let A be a subset of E. Saying that A is bounded for a(E, E') means that every
continuous linear form on E is bounded on A; Saying that A is bounded means that
every continuous semi-norm on E is bounded on A. Let N be the closure of 0 in
E and n the canonical mapping from E onto EjN. The continuous linear forms on
E are the mappings of the form f 0 n with f E (EjN)' and we have an analogous
characterization of continuous semi-norms on E. Replacing E by EjN and A by
neAl we can thus limit ourselves to the case where E is Hausdorff.
TVS 111.28 SPACES OF CONTINUOUS LINEAR MAPPINGS §5

Let 6 be the set of equicontinuous subsets of E'; when E' is assigned the topo-
logy cr(E', E), E can be identified with (E')'6 (III, p. 19, cor. 1). Every closed equiconti-
nuous subset of E' is compact for cr(E', E) (III, p. 17, cor. 2), hence complete for
cr(E', E). It is now enough to apply tho 2.

§ 5. HYPOCONTINUOUS BILINEAR MAPPINGS

1. Separately continuous bilinear mappings

Let E, F, G be three locally convex spaces. For every bilinear mapping u from
E x F into G, and for every x E E (resp. y E F), we denote by u(x, .) (resp. u(., y))
the mapping y f---+ u(x, y) (resp. x f---+ u(x, y)) from F into G (resp. from E into G).

DEFINITION 1. - A bilinear mapping u from E x F into G is said to be separately

continuous if, for all x E E, the linear mapping u(x, .) from F into G is continuous,
and for all y E F, the linear mapping u(., y) from E into G is continuous.
The following proposition follows immediately from the definition.

PROPOSITION 1. - For a bilinear mapping u from E x F into G to be separately

continuous, it is necessary and sufficient that for all y E F, the linear mapping u(., y)
from E into G is continuous and that the linear mapping y f---+ u(., y) from F into 2's(E; G)
is continuous.
We can also say that, to every linear mapping v E 2' (F; 2'sCE; G)) is associated
the bilinear mapping (x, y) f---+ V (y) (x), then we define a linear bijection from
2' (F ; 2's(E; G)) onto the vector space of separately continuous bilinear mappings
from E x F into G.
A separately continuous bilinear mapping from E x F into G need not necessarily
be continuous on E x F (III, p. 47, exerc. 2; cf however III, p. 30, and IV, p. 26,
th.2).
The notion of a separately continuous bilinear form on a product El x E2 of
two locally convex spaces is directly related to that of a continuous linear mapping
when El and E2 are assigned the weak topologies (II, p. 42), Suppose that (E 1 , F 1)
and (E 2, F z) are two pairs of real (resp. complex) vector spaces in separating duality
(loc. cit.); we assign to Ei (resp. F) the weak topology cr(E i, F) (resp. cr(Fi' E))
for i = 1, 2, and denote by B(E 1 , E 2 ) the vector space of separately continuous
bilinear forms on El x E 2. Applying prop. 1 to the case when G = K, we see that,
for every bilinear form <I> E B(E 1 , E 2) and every X z E E 2 , the mapping Xl f---+ <I>(x 1 , x 2 )
is a continuous linear form on E 1 , hence (II, p. 43, prop. 3) there exists one element,
and only one d<l>(X 2) E F 1 such that
(1)

for every Xl E El and X z E E z ; moreover, the mapping d<l>: Ez --+ F 1 is linear and
continuous for the (weak) topologies of E2 and of Fl'
No.2 HYPOCONTINUOUS BILINEAR MAPPINGS TVS III.29

Conversely, for every continuous linear mapping u: E Z -> F 1 the mapping

(x 1, X 2 ) I-+<D(X l' x 2) = <Xl' U(X 2» is a separately continuous bilinear form on
El x E z , and we have u = d<D. Thus we have defined an isomorphism d:<D 1-+ d<D
from B(E 1 , E z) onto 2'(E z ; F 1)' said to be canonical. Similarly the formula
(2)
defines a canonical isomorphism s: <D -> S<D from B(Ep E z) onto 2'(E 1, F z); we
have evidently the commutative diagram

(3)

where t is the isomorphism of transposition (II, p. 46, prop. 5 and corollary). In view
of the definition of weak topologies on F 1 and F z' it is immediate that when B(E 1 , E z ),
2' (E 1 ; E z) and 2' (E z ; F 1) are assigned the topology of simple convergence, the
isomorphisms of diagram (3) are topological vector space isomorphisms.

2. Separately continuous bilinear mappings on a product of Frechet spaces

PROPOSITION 2. - Let E, F and G be three locally convex spaces. Suppose that E

and Fare metrizable andE is barrelled. Let H be a set of separately continuous bilinear
mappingsfrom E x F into G. Suppose thatfor every x E E, the set of mappings u(x, .)
from F into G, where u runs through H, is equicontinuous. Then H is equicontinuous.
Let Un (resp. Vn) be a fundamental sequence of neighbourhoods of 0 in E (resp. F).
If H is not equicontinuous, there exists a closed, convex, balanced neighbourhood
W of 0 in G such that for all n, H(Un x Vn) is not contained in W. There exists then
a sequence of pairs (xn' Yn) E Un X Vn, and a sequence (un) of elements of H, such
that un(xn, Yn) 1= W. Let p be the gauge of W. For every Y E F and every u E H, the
mapping u(., y) from E into G is continuous, hence po u(., y) is a continuous semi-
norm on E. On the other hand, for every x E E, the set of mappings u(x, .)
for u E H is equicontinuous; since the sequence (Y n ) tends to 0, it is bounded, and
the set of all u(x, Yn), for n ~ 0 and u E H, is bounded (III, p. 22, prop. 9). It follows
from this that the function p'(x) = sup p(u(x, Y n)) is a lower semi-continuous semi-
UEH
11;3:0

norm (finite) on E. Since E is barrelled, p' is continuous (III, p. 24, corollary). Since
(xn) tends to 0 in E, p'(x,.) tends to 0, so that we have p'(xn) :(; 1 if n is large enough;
but then p(un(xn, Y n)) :(; p'(xn) :(; 1, hence un(xn, Y n) E W, which contradicts the
hypothesis on un' X n' Yn·
TVS 111.30 SPACES OF CONTINUOUS LINEAR MAPPINGS §5

COROLLARY 1. - Let E and F be two Prechet spaces, and G a locally convex space.
Every separately continuous bilinear mapping from E x F into G is continuous.
In fact, every Frechet space is barrelled (III, p. 25, corollary).
Let E and F be two locally convex spaces. We use .?l(E, F) to denote the space
of continuous bilinear forms on E x F, with the topology of uniform convergence
on sets of the form A x B, where A (resp. B) is bounded in E (resp. F). The formula
u(x, y) = <y, <P (u)(x) >
(for x E E, y E F and u E .?l(E, F)) defines a continuous linear injective mapping <p
from .?l(E, F) into 2"b(E; F~).
COROLLARY 2. - Suppose that E andF are metrizable and that E is barrelled Then <p
is a topological vector space isomorphism from .?l(E, F) onto 2"b(E; F~).
Let f E 2"b(E; F~). Put u(x, y) = <y, f(x) >for x E E and y E F. The bilinear form
u on E x F is separately continuous; by prop. 2, it belongs to .?l(E, F), and we have
f = <p (u). Hence <p is a linear bijection from .?l(E, F) onto 2"b(E; F~). It is immediate
that <p is bicontinuous, hence cor. 2 follows.

3. Hypocontinuous bilinear mappings

In what follows, we shall define a notion which is intermediate between that of
a continuous bilinear mapping and that of a separately continuous bilinear mapping.

PROPOSITION 3. - Let E, F, G be three locally convex spaces, 6 a family of bounded

subsets of E. Let u be a separately continuous bilinear mapping from E x F into G.
The following properties are equivalent:
a) Por every neighbourhoodW of 0 in G and every set M E 6, there exists a neigh-
bourhood V of 0 in F such that u(M x V) c W.
b) Por every set M E 6, the image of M under the mapping x ~ u(x, .) is an equi-
continuous subset of 2"(F; G).
c) The mapping y ~ u(., y) from F into 2"6 (E; G) is continuous.
a) expresses that y ~ u(., y) is continuous at the point 0, on account of the defi-
nition of neighbourhoods of 0 in 2"6 (E; G) (III, p. 13); likewise a) expresses that
the image of M under the mapping x ~ u(x, .) is equicontinuous at the point 0
(III, p. 16).

DEFINITION 2. - Let u be a bilinear mapping from E x F into G. We say that u is

6-hypocontinuous if u is separately continuous and if it verifies one of the equivalent
conditions a), b), c) of prop. 3.
The condition c) of prop. 3 shows that the notion of 6-hypocontinuous bilinear
mapping depends on 6 only through the 6-topology on 2"(E, G).
For every set:l: of bounded subsets ofF, we define similarly the notion of:l:-hypo-
continuous mapping, by interchanging the roles of E and F in prop. 3. A separately
continuous bilinear mapping u is said to be (6, :l:)-hypocontinuous if it is both 6-
hypocontinuous and :l:-hypocontinuous.
NO.3 HYPOCONTINUOUS BILINEAR MAPPINGS TVS III.31

Every continuous bilinear mapping from E x F into G is (6, ~)-hypocontinuous

for every pair (6, ~) of sets of bounded subsets: for every neighbourhood W of 0
in G, there exists a neighbourhood U of 0 in E and a neighbourhood V of 0 in F
such that u(U x V) c W; since every set M E 6 is bounded, there exists A > 0
such that AM c V, and so

u(M x AV) = u(AM x V) c u(U x V) c W .

The converse is in general false (III, p. 47, exerc. 3).

PROPOSITION 4. -- Let u be a 6-hypocontinuous bilinear mapping from E x F into G.

For every set ME 6, the restriction of u to M x F is continuou.-" and u(M x Q)
is bounded in G for every bounded subset Q of F.
The first assertion follows from cor. 3 of GT, X, § 2, No. I. Let W be a neighbour-

°
hood of 0 in G; there exists, by hypothesis, a neighbourhood V of 0 in F such that
u(M x V) c W. Since there exists A =f. such that AQ c V, we have Au(M x Q) =
u(M x AQ) c W, and this proves the second part of the proposition.

PROPOSITION 5. - Let u be a (6, ~)-hypocontinuous hilinear mapping from E x F

into G. For every pair of sets ME 6, N E~, U is uniformly continuous on M x N.
The proposition follows immediately from prop. 2 of GT, X, § 2, No.1 and prop. 5
of GT, X, § 2, No.2.

PROPOSITION 6. -- If F is a barrelled space, every separately continuous bilinear

mapping u from E x F into a locally convex space G is 6-hypocontinuous for every
set 6 of bounded suhsets of E.
In other words, the linear mapping y 1---+ u(., y) from F into 2 b(E; G) is continuous.
It is enough (III, p. 30, prop. 3) to prove that the image of every bounded subset M
of E under x 1---+ u(x, .) is equicontinuous in 2'(F; G). But, by virtue of prop. 1
(III, p. 28) this image is a simply bounded subset of 2'(F : G), and since F is barrelled,
every simply bounded subset of 2' (F ; G) is equicontinuous (III, p. 25, tho I).

Remark. - Suppose the topology of F is the finest locally convex topology on F

for which the linear mappings h~ : F cr -1- F are continuous (II, p. 27). Then condition c)
of prop. 3 (III, p. 30) shows that if E and G are locally convex, then the
bilinear mapping u: E x F -1- G is 6-hypocontinuous if and only if each of the
bilinear mappings

from E x F~ into G is 6-hypocontinuous.

Now suppose that E is a locally convex space which is the strict inductive limit
of an increasing sequence (En) of closed vector subspaces of E (II, p. 33) ; then every
set M E 6 is contained in one of the En and is bounded in this subspace (III, p. 5,
prop. 6). We denote by 6 n the family of all subsets belonging to 6 contained in En'
TVS III.32 SPACES OF CONTINUOUS LINEAR MAPPINGS §5

Condition a) of prop. 3 (III, p. 30) shows that for a bilinear mapping u: E x F -4 G

to be 6-hypocontinuous, it is necessary and sufficient that each of the restrictions
un:En x F-4G of u is 6 n -hypocontinuous.

4. Extension of a hypocontinuous bilinear mapping

PROPOSITION 7. - Let E, F, G be three locally convex spaces, G being assumed
Hausdorff; let Eo (resp. Fo) be a dense vector subspace ofE (resp. F). Let u be a sepa-
rately continuous bilinear mapping from E x F into G.
1) Ifu(E o x Fo) = {O}, then u = O.
2) Let 6 0 be a family of bounded subsets of Eo; if the restriction of u to Eo x F 0
is 6 0 -hypocontinuous then so is u.
1) By hypothesis, for all x E Eo, the continuous linear mapping u(x, .) is null on
F o, hence on F : therefore for all y E F, the continuous linear mapping u(., y) is
null on Eo, hence on E. This proves that u = O.
2) For every closed neighbourhood W of 0 in G and for every set ME 6 0 , there
exist~ by hypothesis, a neighbourhood V of 0 in F 0 such that u(M x V) c W.
But V is a neighbourhood of 0 in F; for every x E M, the relation u( {x} x V) c W
implies Q:at u( { x} x V) c W, since u(x, .) is continuous and W is closed; therefore
u(M x V) c W, which proves that u is 6 0 -hypocontinuous.
PROPOSITION 8. - Let E, F, G be three locally convex spaces; assume that G is
Hausdorff and quasi-complete. Let Eo (resp. Fo) be a dense vector subspace of E
(resp. F), 6 0 (resp. ::to) a family of bounded subsets of Eo (resp. F 0) such that every
point ofE (resp. F) is in the closure of an element of 6 0 (resp. ::to). Then every (6 0 , ::to)-
hypocontinuous bilinear mapping u from Eo x F 0 into G extends uniquely to a sepa-
rately continuous bilinear mapping u from E x F into G and u is (6 0 , ::to)-hypo-
continuous.
The uniqueness and hypo continuity of u follows from prop. 7; it remains to
prove the existence of U. For every y' E F 0' the continuous linear mapping
x' f---+ u(x', y') from Eo into G extends uniquely to a continuous linear mapping
x f---+ u 1 (x, y') from E into G (III, p. 8, prop. 10). It follows immediately that for every
x E E, the mapping y' f---+ u 1 (x, y') from F 0 into G is linear; and we shall show that
it is continuous. By hypothesis, there exists ME 6 0 , such that x E M. For every
closed neighbourhood W of 0 in G, there exists, by hypothesis, a neighbourhood
V of 0 in F 0 such that u(M x V) c W; since x f---+ u 1 (x, y') is continuous, we deduce
that u 1 (M x V) c W, and in particular u1 (x, y') E W for all y' E V. This establishes
our assertion. By virtue of prop. 7, the bilinear map u 1 from E x F 0 into G is (6 0 , ::to)-
hypocontinuous. We end the proof by interchanging the roles of E and F in the first
part of the proof, applied to u 1 •

5. Hypocontinuity of the mapping (u, v) f---+ V 0 u

PROPOSITION 9. - Let R, S, T be three locally convex Hausdorff spaces. Suppose that

the spaces ,P(R; S), 'p(S; T), ,P(R; T) are each assigned the topology of simple
No.5 HYPOCONTINUOUS BILINEAR MAPPINGS TVS III.33

(resp. compact, bounded) convergence. Then the bilinear mapping (u, v) 1--+ V 0 U from
2(R; S) x 2(S; T) into 2(R; T) is (6, '2)-hypocontinuous, where '2 is the family
ofequicontinuous subsets of 2(S; T), and 6 thefamily offinite (resp. compact, bounded)
subsets of 2(R; S).
We first prove that (u, v) 1--+ V 0 u is '2-hypocontinuous. Let H be an equicontinuous
set in 2 (S ; T), let W be a neighbourhood of 0 in T and let M be a finite (resp. compact,
bounded) subset of R. We must show that there exists a neighbourhood V of 0 in
S such that if u(M) c V and v E H, then v(u(M)) c W. But for this, it is enough
to have v(V) c W for all v E H, and the existence of such a neighbourhood follows
from the equicontinuity of H.
To see that (u, v) 1--+ V 0 u is 6-hypocontinuous, we shall prove that, for every
neighbourhood W of 0 in T, every finite (resp. compact, bounded) subset M of R
and every finite (resp. compact, bounded) subset L of 2(R; S) there exists a finite
(resp. compact, bounded) subset N of S such that the relations v(N) c Wand u E L
imply that v(u(M)) c W. Evidently it is enough to show that we can take N = U u(M),
UEL

i.e. that the set N is finite (resp. compact, bounded) whenever Land M are. This is
immediate ifL and M are finite, or ifM is bounded in Rand L is bounded in 2(R; S)
(for the topology of bounded convergence, cf III, p. 22). Finally, we show that if M
is compact in Rand L is compact in 2(R; S) for the topology of compact con-
vergence, then N is compact in S. But if uM is the restriction to M of u E L, the mapping
u 1--+ U M from L into the space C(5 (M ; S) of all continuous mappings from Minto S,
with the topology of uniform convergence, is continuous; hence the image of L
under this mapping is compact, and our assertion then follows from the continuity
of the map (w, x) 1--+ w(x) from C(5(M; S) x Minto S (GT, X, § 1, No.6, prop. 9).
In the two corollaries that follow, we assume as in prop. 9, that the spaces 2 (R ; S),
2(S; T), 2(R; T) are all three assigned the topology of simple convergence, or
all three the topology of compact convergence, or all three that of bounded conver-
gence.

COROLLARY 1. - For every equicontinuous subset H of2(S; T) the map (u, v) 1--+ V 0 U
from 2(R; S) x H into 2(R; T) is continuous.
This follows immediately from prop. 9 (III, p. 32) and 4 (III, p. 31).

COROLLARY 2. - Suppose S is barrelled. If the sequence (un) tends to u in 2(R; S)

and the sequence (v n) to v in 2(S, T), then the sequence(v n 0 un) tends to v 0 u in 2(R; T).
In fact, the sequence (v n ), being simply bounded in 2(S; T) is equicontinuous,
since S is barrelled (III, p. 25, tho 1); the corollary is then a consequence of cor. 1.
TVS 111.34 SPACES OF CONTINUOUS LINEAR MAPPINGS §6

§ 6. BOREL'S GRAPH THEOREM

1. Borel's graph theorem

THEOREM 1. - Let E be a locally convex space which is the inductive limit of Banach
spaces, Fa Souslin locally convex space, for example a Lusin space (GT, IX, § 6, No.2
and No.4), and u a linear mapping from E into F. If the graph of u is a Borel subset of
E x F, then u is continuous.
Let E j be a family of Banach spaces, and (u) a family of continuous linear mappings
u j : E j --+ E such that the topology ofE is the finest locally convex topology for which
the U j are continuous. It is enough to prove that the composed mappings u 0 u j are
continuous, or in fact (GT, IX, § 2, No.6, prop. 10) that the restriction of u 0 U j to every
closed subspace G of E j satisfying the first axiom of countability is continuous.
The graph of this restriction is the inverse image of the graph of u under the conti-
nuous mapping U j x IdF : G x F --+ E x F, hence is a Borel set in G x F. In
addition, G x F is a Souslin space and every Borel subset of a Souslin space is a
Souslin space (GT, IX, § 6, No.3, prop. 10). Th. 1 then follows from tho 4, GT, IX,
§ 6, No.8.
Remark. - Recall (III, p. 12) that every homological Hausdorff and semi-complete
space, for example every Frechet space, is the inductive limit of Banach spaces. * This is
also true for the strong dual of a reflexive Frechet space (IV, p. 23, prop. 4). *

2. Locally convex Lusin spaces

PROPOSITION 1. - Let E be a Hausdorff locally convex space. Suppose that there
exists a sequence (En)neN of Frechet spaces satisfying the first axiom of countability,
and continuous linear mappings Un: En --+ E such that E = U un (En)' Then E is a
Lusin space. neN
Let P n be the kernel of un; then un defines a bijective continuous mapping from
the quotient space En/Pn onto un(EJ Since En/Pn is a Frechet space satisfying the
first axiom ofcountability (GT, IX,§ 3, No.1), hence a polish space (GT, IX,§6, No.1,
def. 1), uiEJ is a Lusin subspace of E (GT, IX, § 6, No.4, prop. 11). Therefore by
GT, IX, § 6, No.7, cor. ofth. 3, the space E, which is regular (GT, III, § 3, No.1) is a
Lusin space.
Example 1. - Every Frechet space satisfying the first axiom of countability is a polish
space, hence a Lusin space. Consequently, so are the spaces 't&'(X), where X is locally
compact and has a countable base (the topology of't&'(X) being that of compact con-
vergence, cf GT, X, § 3, No.3, corollary and § 1, No.6, cor. 3) ; * the spaces 't&' "'(V), where
V is an open subset ofRn (III, p. 9) and Jt"(U), where V is an open subset ofcn (III, p. 10).
Prop. 1 shows that the spaces 't&'o"'(V), where V is an open set in R n, ~s(I), where I
is a compact interval in Rand s ~ 1, and Jt"(K), where K is a compact subset of cn
are all Lusin spaces (III, p. 10). *

THEOREM 2. - Let E be a locally convex space, which is the inductive limit of an

increasing sequence (En)neN of subspaces of E, endowed with the topologies of Frechet
No.2 BOREL'S GRAPH THEOREM TVS III.35

spaces satisfying the first axiom of countability. Suppose that every compact subset
of E is contained in one of the En and is compact in this space. Let F be a Frechet space
satisfying the first axiom of countability. Then the space .Pc(E; F) is a Lusin space.
The space E is bomological (III, p. 12), hence the space .PcCE; F) is complete
(III, p. 23, prop. 12). The linear mapping}: f f---+ UIEn)nEN is an injection from .PcCE; F)
into the product space n
nEN
.PcCEn; F); by virtue of the hypothesis on the compact
subsets ofE and the definition of the 6-topologies,} is an isomorphism from .PcCE; F)
onto its image (endowed with the topology induced by the product topology);
moreover, since .!.f,,(E; F) is complete, this image is a closed subspace of n
nEN
.PcCEn; F)
(GT, II, § 3, No.4, prop. 8). By GT, IX, § 6, No.4, it is therefore enough to prove that
each ofthe spaces .PiE; F) is a Lusin space. For the rest of the proof, we shall assume
that E is a Frechet space satisfying the first axiom of countability.
Since F is a Frechet space satisfying the first axiom of countability, it is isomorphic
to a closed subspace of a countable product of Banach spaces F n , each of which is
a quotient of F (II, p. 5), hence satisfies the first axiom of countability. The linear
mapping j' : f f---+ (prn 0 f)nEN is an injection from .!.f,,(E; F) into the product space
n
nEN
.PcCE ; F n ), and by using the definition of the 6-topologies and of the open sets

in a product, j' is an isomorphism from .Pc(E ; F) onto its image; moreover, since
.PcCE; F) is complete, this image is a closed subspace of
nEN
n
.!.f,,(E; F). Therefore
it is enough to prove that each of the spaces .Pc(E; Fn) is a Lusin space (GT, IX,
§ 6, No.4), and consequently, we can assume that F is a Banach space satisfying the
first axiom of countability.
The space .PiE; F) is the union of a countable family of equicontinuous and
closed subsets (III, p. 19, cor. 1 and GT, X, § 2, No.3, prop. 6). But every equiconti-
nuous subset H of .PcCE; F) is metrizable and satisfies the first axiom of countability
(III, p. 18, prop. 6 and GT, X, § 2, No.4, tho 1); if H is closed, then it is a complete
space for the uniform structure induced by that of .PcCE ; F), since the latter is complete
In other words, H is a polish space, and a fortiori a Lusin space; consequently the
regular space .Pc(E; F) is a Lusin space (GT, IX, § 6, No.7, cor. of tho 3).

COROLLARY. - The hypotheses on E being as in tho 2, assume, in addition that every

bounded subset of E is relatively compact. Then the strong dual of E is a Lusin space.
* In particular, the strong dual of a Frechet space satisfying the first axiom of counta-
bility, which is also a M ontel space, is a Lusin space. *

* Example 2. - Let U be an open subset of Rn. The corollary applies in particular to

the Frechet space E = CC OO(U); its dual CCo - OO(U) (the space of distributions with compact
support on U) is then a Lusin space.
The space CCoOO(U) is a strict inductive limit of a sequence of Frechet spaces CCKOO(U)
satisfying the first axiom of countability (III, p. 9). We can show that each of the spaces
CCK~(U) is a Montel space; in addition, every bounded subset of CCo OO(U) is contained
in one of the spaces CCK~(U) (III, p. 5, prop. 6). We can then apply the corollary of tho 2.
Then the dual CC - OO(U) of CCo OO(U) (the space of distributions on U) is a Lusin space for
the strong topology.
TVS III.36 SPACES OF CONTINUOUS LINEAR MAPPINGS §6

Similarly we prove that for every open subset U of en, and for every compact subset K
of en, the strong dual of Yf(U) and the strong dual of Yf(K) are Lusin spaces. *
Remark. ~ Let E be as in tho 2; let F be a Hausdorlf locally convex space which is the
union of the images of a sequence of continuous linear mappings Un: Fn -> F, where
each Fn is a Frechet space satisfying the first axiom of countability; then 2'c(E; F)
is a Lusin space. As in prop. 1, we first reduce to the case where each un is injective;
then, as in the proof of tho 2, we can assume that E is a Frechet space satisfying the first
axiom of countability. Then, by I, p. 20, prop. 1, 2'(E; F) is the union of the 2'(E; Fn);
moreover, the canonical injection 2'cCE; Fn) -> 2'cCE; F) is continuous (GT, X, § 1, No.4,
prop. 3). Since each of the spaces 2'cCE; Fn) is a Lusin space by tho 2, 2'(E; Fn) is also
a Lusin space for the topology induced by that of 2'cCE; F) (GT, IX, § 6, No.4, prop. 11) ;
consequently 2'c(E; F) is a Lusin space by virtue ofGT. IX, § 6, No.7, corollary ofth. 3.

* 3. Measurable linear mappings on a Banach space 1

PROPOSITION 2. - Let E be a Banach space, F a locally convex space and u a linear

mapping from E into F. Assume that for every closed subset B of F, every compact
subset X of E and every measure J.l on X, the intersection X n u- l(B) is J.l-measurable.
Then u is continuous.
First assume that F is the base field. For every compact subset X of E and every
measure J.l on X, the restriction of u to X is J.l-measurable (INT, IV). Suppose that
u is not continuous. Then we can find a sequence of points (xn) in E such that
I Ilxnll < CfJ and lu(Xn)I ~ n for every integer n. Consider the mapping
n
g: (tn) f--> I tnxn from the cube C = (0, 1) N into E; it is clear that g is continuous.
Hence f = u 0 g is measurable for every measure on C (INT, V); in particular for
the measure J.l which is the product of Lebesgue measures on the factors of C. Hence
there exists a compact subset D of C such that J.l(D) > t and such that the restric-
tion of f to D is continuous, hence also bounded. Let M be the upper bound of
If I on D and let pEN be such that p ~ 4M. Let S = (sn) and t = (t n) be two points
of D such that sn = tn for all n #- p. Then

f(s) - f(t) = u(I snxn - I tnxJ = (Sp - tp) U(Xp) .

n n

Since If(s) - f(t)1 ~ 2M and lu(Xp) I ~ p ~ 4M, we get

Isp - tpl ~ t·
The Lebesgue-Fubini theorem (INT, V, 2nd ed., § 8, No.3, cor. 2 of prop. 7) implies
that J.l(D) ~ t; this gives a contradiction. Hence u is continuous.
In the general case, for every v E F', the linear form v 0 u is continuous, by the pre-
ceding argument. Let (xn)nEN be a sequence of points of E tending to 0; then the
sequence (u(xn»)nEN tends to 0 in F, if F is assigned the topology cr(F, F'); hence this
sequence is bounded for cr(F, F') and so it is bounded in F (III, p. 27, cor. 3). Since E
is bomological (III, p. 12, prop. 2); the linear mapping u: E --> F is continuous. *

1 The results of this section depend on the book of Integration.

 Exercises

§I

°
1) Let E be a left topological vector space over a non-discrete topological field K. A subset
B of E is said to be bounded if for every neighbourhood V ofO in E there exists A oF in K such
that AB c V.

° °
a) Show that if B is bounded in E, then for every neighbourhood V of in E, there exists a
neighbourhood U of in K such that U. B c V.
b) Show that the closure of a bounded set in E is bounded. The union of two bounded sets
is bounded. Every precompact set in E is bounded. Extend the corollaries of III, p. 4, prop. 4
to topological vector spaces over K.
c) Prove that if A is a bounded set in K (considered as a vector space on the left over itself)
and B is a bounded set in E, then A. B is bounded in E.
d) Extend prop. 3 of III, p. 4 to the case where K is a metrizable topological division ring.
e) Extend the notion of a quasi-complete space and its properties to topological vector
spaces.

°
2) a) Let E be a left topological vector space over a non-discrete topological field K. Prove
that if there exists a neighbourhood V of in E which is bounded (exerc. 1), than the sets AV,
for A E K and A oF 0, form a fundamental system of neighbourhoods ofO in E. IfK is metrizable,
the Hausdorff topology associated with the topology of E (OT, III, ~ 2, No.6) is metrizable. If
K = R or K = C, the locally convex topology on E which is the finest of the topologies
coarser than the given topology on E (II, p. 80, exerc. 23) can be defined by a single semi-
norm.
b) Prove that the topology of an infinite product of locally convex Hausdorff spaces (of
which none is just 0) cannot be defined by a single semi-norm.
c) Let E be a locally convex space whose topology is defined by an increasing sequence (P n )
of semi-norms. In order that the topology of E be defined by a single semi-norm, it is necessary

kll ~°
and sufficient that there exists an integer no such that for every n ~ no there exists a number
such that Pn(x) ~ k,Pno(x) for all x E E.
d) Let E be the vector space over R consisting of infinitely differentiable numerical functions
on the internal I = (0, I). For every integer n ~ 0, let PnU) = sup (sup Ipk)(x)1) (with
O~k:'=:;n xeI
TVS III.38 SPACES OF CONTINUOUS LINEAR MAPPINGS §1

PO) = f); show that the Pn are norms on E and that the topology defined by the sequence
of norms p" cannot be defined by a single norm.

3) Let E be a metrizable vector space over R, d a translation invariant distance, compatible

with the topology of E. Let Ixl = d(x, 0) (1, p. 16). Prove that for every integer n > 0, we have
~n Ixl :( 1::.1.
n
Deduce from this that if B is a bounded subset of E, then sup
=B
Ixl < + 00 (in
other words, B is bounded for the distance d (GT, IX, § 2, No.3»). Give an example of a metri-
zable vector space E and an unbounded set in E which is bounded for the distance d (exerc. 2).

4) Let E be a topological vector space over a non-discrete metrizable topological field K.

Show that ifE is a Baire space and if in E there exists a countable base for the bomology formed
by bounded subsets ofE (III, p. 37, exerc. 1), then there exists a neighbourhood ofO in E which
is bounded, and consequently the Hausdorff topology associated with the topology of E is
metrizable (III, p. 37, ex ere. 2) (compare with exerc. 6).

5) Let E be a metrizable vector space over a non-discrete valuated field K. Prove that if (Bn)
is an arbitrary sequence of bounded subsets ofE (III, p. 37, exerc. 1), then there exists a sequence
(A,,) of scalars of 0, such that the union of the sets AnB" is bounded.

6) Let E be a locally convex space, which is the strict inductive limit of a strictly increasing
sequence (E,,) of locally convex Hausdorff spaces, each En being closed in En + 1 (II, p. 32,
prop. 9).
a) Prove that E is not metrizable (use III, p. 5, prop. 6 and the preceding exerc. 5).
b) In order that there exist a countable base for the canonical bomology of E, it is necessary
and sufficient that the canonical bomology of each En has a countable base in En'

~ 7) a) Let E be an infinite dimensional Banach space and let 6 be the family of compact,
convex and balanced subsets of E, which is a directed set for the relation c. Show that E is

°
the inductive limit of the inductive system of the Banach spaces E A, where A runs through 6.
(Prove by contradiction that a neighbourhood V of for the inductive limit topology of the
topologies of EA contains a ball with center 0; for this, note that otherwise, there will exist
a sequence (x,) of points of E such that I x" I :( 1/n2, and which will not belong to V.) Deduce
from this that there exist bounded subsets in E which are not contained in any EA for A c 6.
TI
eD

b) Let E be an infinite dimensional Banach space. On the vector space Em' where Em = E
m=l
for each m, let Y n denote the topology obtained by taking the product of the Banach space
topology on each Em for m :( n, and for m > n, the finest locally convex topology on Em;
oc
F" denotes the locally convex space TI Em with the topology Y n. Every identity map Fn --> Fn+ 1
m=l
is continuous; show that the inductive limit space of the inductive system Fn is the space F
ro
obtained by assigning to TI Em the topology which is the product of the Banach space topo-
m=l
logies on each of the factors. Deduce from this that there are bounded subsets in F which
are not bounded in any Fn'

8) Prove that in a space which is an infinite product of topological vector spaces (over R
or C) none just the point 0, there does not exist a countable base for the canonical bomology
(first show that it is enough to prove this for the space RN , and then use III, p. 38, exerc. 4).

9) Let E be the vector space over R consisting of all regulated functions on the interval I = (0, 1)

f
(FVR, I, p. 4). For every integer n > 0, let Vn be the set of all functions fEE such that

JT7(t)l dt :( lin. Show that the sets Vn form a fundamental system of neighbourhoods

of °for a metrizable topology which is compatible with the vector space structure of E and
§1 EXERCISES TVS 111.39

that for this topology the sets VII are bounded; but the convex envelope of each V n is the entire
space E. (Observe that every function fEE can be written as f = ~(g + h), where 9 and h
belong to E, and

11jjg(t)1
o
dt = 11 v1fi(t)1
0 )2
1
dt = ---= 11 v1f(t)l
0
dt.)

Deduce from this that the only locally convex topology coarser than the topology of E is the
coarsest topology on E.
10) Let (E,),EI be an infinite family of Hausdorff topological vector spaces, none just the point
0, over a non-discrete topological field K. Let E be the direct sum vector space of the E" and
:Yo the topology on E defined in I, p. 24, exerc. 14. Then a subset B of E is bounded for :Yo
(III, p. 37, exerc. 1) if and only if B is contained in a product subspace TI
E" where H is a finite
'EH
subset of I and the projections of B on each E, for 1 E H are bounded. Deduce (for K = R
or C) that, if each E, is a quasi-complete space, then E with :Yo is quasi-complete.

11) Let E be a topological vector space over a field K with a non discrete valuation.
a) For a balanced subset A of E to absorb every bounded subset (III, p. 37, exerc. 1) of E, it is
sufficient that A absorbs the set of points of every sequence (XII) tending to 0 in E. Then A is
said to be bornivorous.
b) Let u be a linear mapping from E into a topological vector space F over K. The image
of every bounded subset of E under u is bounded in F if and only if for every sequence (x,,)
of points of E tending to 0, the sequence (u(x,)) is bounded in F
C) Suppose E is metrizable. Show that every bomivorous subset of E is a neighbourhood
of 0 in E. Deduce that if u is a linear mapping from E into a topological vector space F over
K which transforms every sequence converging to 0 in E into a bounded sequence in F, then
u is continuous on E.

12) Let E be a Hausdorff topological vector space over a field K with a non-discrete va-
luation, and F a metrizable vector space over K. If u is a continuous linear mapping from E
into F such that, for every bounded subset B of F, u- 1 (B) is bounded in E, show that u is an
isomorphism from E onto a subspace of F

13) Let I be an infinite set, and (E)'EI a family of locally convex spaces, none of which is O.
Let f be a linear mapping from E = TI
E, into a Banach space F Show that if the image of
'EI
every bounded subset of E under f is a bounded subset of F, then there exists a finite subset
H of I such that for every 1 1= H, the restriction of f to E, (considered as a subspace of E) is
null. (Argue by contradiction that if not, we can construct a bounded sequence (xn) in E whose
image under f is unbounded in F)

14) Show that if the topology of a metrizable locally convex space E cannot be defined by
a single norm, then there does not exist a countable base for the canonical bomology of E
(using III, p. 38, exerc. 5, show that otherwise there will exist a bounded bomivorous set (III,
p. 39, exerc. 11) in E, and complete the argument using III, p. 39, exerc. 11, c)).

15) In a Hausdorff topological vector space E over R, let A be a compact convex set and B
a closed, convex and bounded set. Show that the convex envelope C of the union A u B is
a closed set. (Consider a point z in the closure ofC, but not in A, and reduce to the case where
z = o. Observe that there exists a neighbourhood V of 0 and a number ri < 1 such that the
relations 0 :( A :( 1, x E A, Y E B, AX + (1 - A) Y E V imply that A :( ri. Next, for every
neighbourhood W of 0 consider the set of triplets (A, x, y) such that AX + (1 - A) YEW,
o :( A :( 1, X E A, Y E B.)
16) Let E be a locally c<?nvex metrizable space satisfying the first axiom of countability,
such that its completion E is a Frechet space which satisfies the first axiom of countability.
TVS 111.40 SPACES OF CONTINUOUS LINEAR MAPPINGS §2

Show that every bounded subset B of E is contained in the closure of a bounded subset of E.
(Reduce to the case where B is countable, arranged as a sequence (X");A on the other hand,
let (p,,) be an increasing sequence of semi-norms defining the topology of E; for each integer n
consider a sequence (Y"k)D I of points ofE, which converges to x" and is such thatp,,(x,,- Y"k) ~ I
for all k ~ 1.)

-r 17) Let A be the set of increasing mappings ~ 1 from N into N; for every CI. E A, let Bo
denote the set of all points Z = (ZII) E RN such that IZIII ~ CI.(n) for all n E N.
a) Show that the sets B. form a base for the bomology of all bounded subsets of the space RN.
b) For every a. E A, the set RBo is a vector subspace ofR N, distinct from RN and dense in RN;
hence there exists a linear form j~ =ft 0 (not continuous) on RN such that fo(z) = 0 for all
Z E Bo '
c) Let E be the vector space consisting of all mappings g: a. f--> (gll(a.)) E RN from A into RN
such that for all n EN, the sum PII(g) = I Ig,,(a.)I is finite. Show that the P" are semi-norms

which define the topology of a Frechet space on E.

d) Let H be the set of all h E E such that h( a.) E RB. for all a. E A; show that H is an everywhere
dense vector subspace of E (observe that every hE E such that h(a.) = 0 except for a finite
number of values of a. E A belongs to H).
e) Let Eo c E be the vector subspace ofE consisting of all g E E such that I 1f.(g(a.))1 < + CfJ;

the mapping u: g f--> (fo( g( a.)))oEA is then a linear mapping from Eo into the Banach space
F = fl(A) (I, p. 4). Prove that u(Eo) is everywhere dense in F (observe that for every finite
subset J of A, there exists g E Eo such that g(a.) = 0 for all a. E A - J and that the f.(g(a.))
for a. E J take arbitrary values in R). Show that u-I(O) is everywhere dense in Eo (use d)).
Finally, show that for every bounded subset C ofE, there exists a. o E A such that f.o(g(a. o)) = 0
for all g E C n Eo, and deduce that the closure of u(C n Eo) in F is not a neighbourhood
of 0 in F.
f) Let G be the graph of u in Eo x F, a vector subspace of the Frechet space E x F. Show
that G is everywhere dense in E x F (observe that for every x E Eo, x + u-I(O) is dense
in E). However, show that for every bounded subset M of G, the closure M of M in E x F
does not contain the bounded set {O} x U ofE x F, where U is the unit ball in F (ifN = pr I (M),
observe that because of e), u(N) cannot contain U).

18) In the Banach space CI(N) (I, p. 4) let em be the sequence (8 m ,,),,;, 0 such that 8mn = 0 for
and 8"" = 1. Define a continuous mapping from fl(N) in R which transforms the
In =ft 11
b ,)unded seq uence of the ell into a non bounded subset ofR (use U rysohn's tho (G T, IX, § 4, No.2,
th 2)).

§2

1) Let E be a locally convex space, and :Y its topology. Amongst the locally convex topologies
on E for which the bounded sets are the same as those for /Y, there is one :Y' finer than all
the others, and this is the only one amongst these topologies which is bomological. The space
obtained by assigning E with :Y' is called the bomological space associated with E. A linear
map u from E into a locally convex space F transforms every bounded subset of E into a
bounded subset of F if and only if it is continuous for the topology :Y '.
Show that the topology :Y' is the finest of the locally convex topologies on E for which the
canonical injections EA --> E, where A runs through the family of convex, bounded and balanced
subsets of E, are continuous.

2) Let I be an infinite set and (EJ'EI a family of locally convex spaces, none of which is O.

n
a) Suppose that each space E, is bomological. Show that if, in addition, the product space
RI is bomological, then the product E = E, is bomological (using III, p. 11, prop. 1 (iii)
'EI
§3 EXERCISES TVS 111.41

and p. 39, exerc. 15, reduce to proving the following: a linear mapping I from E into a Banach
space, which transforms every bounded set into a bounded set, and whose restriction to
each E, is null, is necessarily null on E. For this consider, for every x = (x,) E E, the restriction
of I to the product of lines RxJ
b) Deducefrom a) that every product ofa sequence (En) ofbomological spaces is bomological.

I[ 3) Let E be a locally convex space, L a vector subspace of E of linite co dimension, and S

a convex balanced bomivorous set (III, p. 39, exerc. 11) in L. We shall prove that there exists
a convex, balanced and bomivorous set S' in E such that S = S' n L.
a) We can reduce to the case where L is a hyperplane such that E = L EB Ra for a point
a ~ L, and such that there exists a bounded sequence (x n) in E such that ifwe put xn = An(Yn +a)
with An E Rand Yll E L, then IAnl tends to + CIJ ; if Bo is the convex balanced envelope of the
set consisting of a and the xll ' then we have Yn + a E A,~ 1Bo for all n.

thesis, for every B E 'B, there exists PB > °

b) Let 'B be the set of all bounded, convex, balanced subsets of E which contain Bo ; by hypo-
such that 2P BB n L c S. Show that if R is the
convex balanced envelope of the union of the sets PBB for B E 'B, then we have R n L c S.

4) Deduce from exerc. 3 that if E is a locally convex bomological space, then every subspace
of E with finite codimension is bomological (ef IV, p. 64, exerc. 11).

§3

1) Let X be a Hausdorff topological space, and F a topological vector space (over R or C).
Show that on the space '&'(X; F) of all continuous maps from X into F, the topology of compact
convergence is compatible with the vector space structure.

2) Let E and F be two Hausdorff locally convex spaces, and 6 a family of bounded subsets
of E.
a) Show that if F is not just 0, then a necessary (and sufficient) condition for 2- (E; F) to
be Hausdorff is that the union of the sets of 6 is total in E (use Hahn-Banach th.). C

b) Suppose that 6 is a cover for E. Show that there exists an isomorphism from F onto a
closed subspace of 22 (E; F). Deduce that if 22 (E; F) is quasi-complete, then F is necessarily
quasi-complete.
e) Suppose that 6 is a bomology adapted to E (III, p. 3, def. 4). In order that 22 (E; F) be
metrizable, it is necessary and sufficient that F is metrizable and that there exists a countable
base (III, p. 1) for the bomology 6. In order that the 6-topology on 2(E; F) be defined by
a single norm it is necessary and sufficient that the topology of F can be defined by a single
norm and that there exists a set M E 6 which absorbs every set of 6.

3) Let E be a topological vector space over R (resp. C). Show that for every family 6 of bounded

canonically isomorphic to E. Deduce that for every integer n > °

subsets of R (resp. C) none of which is the point 0, the space 22 (R; E) (resp. 22 (C; E)) is
and every covering 6 of
Rn (resp. C") by bounded subsets, 22 (Rn; E) (resp. 22 (C"; E)) is isomorphic to En.

4) a) Let El' E z , F be three topological vector spaces (over R or C). Let I be a continuous
linear mapping from E1 into E z , and 6 1 (resp. 6 z ) a family of bounded subsets ofE1 (resp. E z),
such that 1(6 1 ) c 6 z . Show that U f-> U 0 I is a continuous linear mapping from 2 22 (E z ; F)
into 2 2 ,(E 1 ; F).
b) Let E, F be two topological vector spaces, and M be a vector subspace of E. Let I be the
canonical map from E onto ElM, and 6 be a family of bounded subsets of E. Show that the
mapping U f-> U 0 I is an isomorphism from 2 J 2 )(E/M ; F) onto the subspace of 22 (E; F)
consisting of those continuous linear mappings from E into F which are null on M.

5) Let (E.)'EA be a family of locally convex spaces, E a vector space (over the same field of
scalars as the E,), and for each rJ. E A, let h, be a linear mapping from E, into E. The space E
is assigned the finest locally convex topology for which the h, are continuous (II, p. 27). For
TVS III.42 SPACES OF CONTINUOUS LINEAR MAPPINGS §3

every r:x E A, let 6, be a family of bounded subsets of Eo, and let 6 be the union of the families
h,(6,) of bounded subsets of E. Under these conditions, show that, for every locally convex
space F, the 6-topology on 2'(E; F) is the coarsest topology for which the linear mappings
u f--> u 0 h. from 2'(E; F) into 2'c" (E.; F) are continuous. In particular, if E is the topological

n
direct sum (II, p. 29, def. 2) of the family (E')OEA (each E. being identified with a subspace of E),
then the product space 2'c" (Eo; F) is canonically isomorphic to the space 2'c (E; F),
OEA
where 6 is the union of the 6 in Il3(E).
0

6) Let (E,)tEI be a family of Hausdorff locally convex spaces, none of which is 0, let E be the
product space n
tEl
Et and F be a normed space. Show that there exists a canonical isomorphism
from the space 2'(E; F) with the topology of bounded convergence (resp. of simple conver-
gence, resp. of precompact convergence) onto the topological direct sum space of the spaces
2' (Et ; F), where each of these spaces is assigned the topology of bounded convergence (resp.
of simple convergence, resp. of precompact convergence). (Observe that if u is a continuous
linear mapping from E into F, then there exists a finite subset H of I such that u- 1 (0) contains
the product of the Et for all indices 1 ~ H.)

7) Let E, F p F 2 be three topological vector spaces, let f be a continuous linear mapping

from F 1 into F 2' and 6 be a set of bounded subsets of E; show that u f--> f 0 u is a contin uous
linear mapping from 2'c(E; F 1 ) into 2'c(E; F z)·

8) Let E be a topological vector space, with a set of bounded subsets of 6. Let (GJtEI be a
family of topological vector spaces, and F be a vector space (over the same field of scalars
as E and the G,) ; for every 1 E I, let gt be a linear mapping from F into G t. Suppose F is assigned
the coarsest topology for which the gt are continuous. Show that the 6-topology on 2'(E; F)
is the coarsest topology for which the linear mappings u f-4 gt 0 u from 2' (E; F) into 2'c (E; GJ
are continuous. In particular, if F = n tEl
n
G t , the product space
tEl
2'c (E ; GJ is canonically

identified with 2'c (E; F).

9) Let E and F be two Hausdorff topological vector spaces, and H an equicontinuous subset
of 2'(E; F). Show that if there exists a countable total set in E, and if everv bounded subset
of F is metrizable, then H is metrizable for the topology of simple convergence-in E. If moreover,
every bounded subset of F satisfies the first axiom of countability, then so does H.

10) Let E be a topological vector space, which is a Baire space, and F be a topological vector
space.

°
a) Show that, if a subset H of 2'(E; F) is bounded for the topology of simple convergence,
then H is equicontinuous (for every closed neighbourhood V of in F, consider the sets
Mn = n u-1(nV)).
UEH

b) Show that, if a subset H of 2'(E, F) is not equicontinuous, the set of all x E E such that
H(x) is not bounded in F is the complement of a first category set. Deduce from this that,
if (Hn) is a sequence of subsets of 2'(E; F) which are not equicontinuous, then there exists
an x E E such that none of the sets Hn(x) is bounded in F (<< principle of condensation of singu-
larities »).

~ 11) Let T be a metrizable topological space, E a topological vector space which is a Baire
space, and M a family of mappings from ExT into a topological vector space F, satisfying
the following conditions :
10 for all to E T, the set of all mappings x f--> f(x, to) where f runs through M, is an equicon-
tinuous set of linear mappings from E into F;

20 for all Xo E E, the set of all mappings t f--> f(x o , t) from T into F, where f runs through M,
is equicontinuous.
Show that M is equicontinuous. (Given to E T and a closed balanced neighbourhood V of °
§4 EXERCISES TVS 1II.43

in F, for every x EO E, let dx be the upper bound of the radii of all open balls with center to in T
such that, for an arbitrary point t in one of these balls, we have f(x, t) - f(x, to) EO V for all
f EO M. Show that x f-> dx is upper semi-continuous at every point Xo EO E; for this, show that
if we had dx > rJ. > dx for points arbitrarily close to x o , then for every neighbourhood W
of 0 in F, f(x o , t) - !(x o , to) would belong to V + W for d(t, to) :;:; rJ. and f EO M. Finally
use GT, IX, § 5, No.4, tho 2.)

12) Let E be a bomologicallocally convex space, and 6 be a family of bounded subsets of E

containing the image of every sequence converging to O.
a) Show that for every locally convex space F, every bounded subset of !.f~ (E; F) is equi-
continuous.
b) Show that if F is a Hausdorff and quasi-complete locally convex space, then the space
!.f~ (E; F) is quasi-complete.

13) Show that if E is a Hausdorff and semi-complete locally convex space, then for every
locally convex space F, every subset of !.f(E; F) bounded for the topology of simple conver-
gence is bounded for every 6-topology.

§4

I) Show that the completion of a Hausdortr barrelled space is barrelled.

2) Let E be a vector space over R or C. Show that E, with the finest locally convex topology
on E (II, p. 25) is barrelled. Deduce from this examples of barrelled spaces which are not
metrizable and are not Baire spaces.

3) Let E be a Hausdorff locally convex space with a countably infinite base (all)'
a) Show that E admits a countable, topologically independent base (ell) (using the fact that
every line in E has a topological complement, define the ell by induction).
b) Show that, for E to be barrelled, it is necessary and sufficient that the topology .r of E is
identical with the finest locally convex topology on E (observe that the convex balancea
envelope of every sequence (Alle ll ) is closed in E). In particular, if :Y is metrizable, E is not
barrelled (el exerc. 2).

4) Let E be a Banach space in which there exists an infinite algebraically independent sequence
(an) which is total in E (for example the space e1 (N) (I, p. 4»). Let B be a base of E containing
the an; we know (II, p. 80, exerc. 24) that B is not countable. Let (en) be a sequence of distinct
elements of B, and distinct from the all' and let C be the complement of the set of the ell in B.
Let Fn be the vector subspace of E generated by C and the ek for indices k :;:; n; E is the union
of the Fn' Let S be the unit ball in E; show that there exists an index n such that S n Fn, is
not a first category set. Deduce that for this value of n, FlI is a metrizable, non-complete Baire
space.

5) Give an example of a locally convex space which is a complete, Hausdorff Baire space,
but is not metrizable (el GT, IX, § 5, exerc. 16).

6) A locally convex space E is said to be relatively bounded if there exists a bounded barrel
in E.
a) In order that E be relatively bounded, it is necessary and sufficient that the topology of E
is coarser than a topology defined by a semi-norm. Then there exists a base for the canonical
bomology of E consisting of barrells.
b) For E to be bomological and relatively bounded, it is necessary and sufficient that the
topology of E is the lower bound of a family of normed space topologies on E (el III, p. 40,
exerc. I). Further, in order that there exist also a countable base for the canonical bomology
of E, it is necessary and sufficient that the topology of E is the lower bound of a sequence of
normed space topologies.
TVS III.44 SPACES OF CONTINUOUS LINEAR MAPPINGS §4

7) A locally convex space E is said to be infra-barrelled if every barrel of E which is bomi-

verous (III, p. 39, exerc. 11) is a neighbourhood of 0 in E. Every bomological space is infra-
barrelled; every barrelled space is infrabarrelled. Show that the completion of a Hausdorff
infra barrelled space is barrelled (use the fact that in a Hausdorff locally convex space E, each
barrel absorbs every convex, balanced, bounded and semi-complete subset of E).

8) Let (EJ'EI be a family of infrabarrelled spaces, and for every lEI, let J; be a linear mapping
from E, into a vector space E. Show that the space E, with the finest locally convex topology
for which the J; are continuous, is infra barrelled. In particular, every quotient space of an
infra barrelled space is infra barrelled ; every topological direct sum of infra barrelled spaces is
infrabarrelled.

9) Let I be an uncountably infinite set; on the direct sum vector space E = R(I), consider,
on the one hand, the finest locally convex topology Y, and on the other hand, the topology Yo
defined in I, p. 24, exerc. 14, which is locally convex; we know that Y and Yo are distinct
(II, p. 75, exerc. 11) and that E with Yo is complete (GT, Ill, § 3, exerc. 10). Show that the
bounded sets in E are the same for;:7- and ,Ufo (III, p. 39, exerc. 10) and that E with Yo is not
barrelled (observe that the set T of all x = (~,) E E such that I I~,l ~ 1 is a barrel and use
exerc. 11 of II, p. 75). lEI

10) Show that an infra barrelled space in which every closed convex balanced and bounded
subset is semi-complete is a barrelled space.

11) Let E be an infra barrelled space, F a locally convex space. Show that every subset of
!f'(E; F) which is bounded for the topology of bounded convergence is equicontinuous.

-r 12) a) Let E be a locally convex space, (An) an increasing sequence of convex balanced
sets in E such that A = U All is absorbent. Let (W n) be a decreasing sequence of convex balanced
n
neighbourhoods of 0; then the convex balanced envelope V of the Wn n All is absorbent;
if E is barrelled, V is a neighbourhood of O.
b) Let ~ be a filter on E; suppose that for every n, there exists a set Mil E ~ such that
(Mil + W n) n A21l = 0· Let VII be the convex balanced envelope of the W k n Ak for k ~ n - 1
and ofWn in such a way that Vn is a neighbourhood of 0 and that we have tvC VII for all n.
Show that (Mil + VII) n An = 0 for all n.
c) Deduce from a) and b) that ifE is barrelled and if~ is a Cauchy filter on E, then there exists
an integer N such that, for all M E ~ and every neighbourhood W of 0 in E, M + W meets AN'
(Argue by contradiction; with the notations of b), consider a set M E ~ with small order tV.)

~ 13) a) Let E be a barrelled space, (C,) an increasing sequence of convex, balanced sets
such that E = U C II . Let U be a convex, balanced and absorbent set such that for every n,

U n C n is closed in C n . Show that U is a neighbourhood of 0 in E. (Show that D c 2U, by

considering a filter ~ on U converging to a point x E E and applying exerc. 12, c)).
b) Let E be a barrelled space, (En) an increasing sequence of subspaces ofE such that E = U En'

Show that ifU is a subset ofE such that Un En is a barrel in En for every n, then U is a neigh-
bourhood of 0 in E. In particular, E is the strict inductive limit (II, p. 33) of the sequence (EJ.

~ 14) a) Let E be a Hausdorfflocally convex space, L a subspace ofE with finite codimension,
and T a barrel in L. Show that there exists a barrel T' in E such that T' n L = T (show that
we can take for T' the sum of the closure T ofT in E and of a finite dimensional compact convex
set).
b) Let E be a barrelled space, L a subspace of E, which has a complement having a countable
basis. Show that L is barrelled (use a) and exerc. 13, b)).
* c) Let E be a Hausdorff locally convex space; its completion E can be identified with a closed
§4 EXERCISES TVS III.45

subspace of a barrelled space F, which is the product of a family of Frechet spaces (II, p. 5,
prop. 3 and IV, p. 14, corollary). Let (e")"EA be the basis of a complement in F of the subspace E,
and let H" be the hyperplane in F generated by E and the e ~ for indices ~ =1= rJ.; by b), H. is a
barrelled space. For every x E E, let u(x) be the point of the barrelled space G = fl H" (sub-
"EA
space of FA) all whose coordinates are equal to x; u is an isomorphism from E onto the subspace
~ n G, where ~ is the diagonal in FA. Show that u(E) = ~ n G is closed in G, and consequently
that every Hausdorff locally convex space is isomorphic to a closed subspace of a Hausdorff
barrelled space. *

15) Let E be a Hausdorff barrelled (resp. infra barrelled) space, and E be its completion. Show
that every subspace F of E which contains E is barrelled (resp. infrabarrelled) (ef III, p. 24,
cor. and IV, p. 52, exerc. 1).

-r * 16) Let (E)'EI be an uncountable family of Hausdorff barrelled spaces, none of which are
the point 0, and let E = fl E, ; then E is barrelled (IV, p. 14, corollary). Let G be the subspace
lEI

of E consisting of all points (x) such that x, = °

except for a countable number of indices.
Every sequence of points of G which converges in E has a limit belonging to G, but G is dense
in E.
a) Show that every subset M of G' = E', which is bounded for cr(E', G) is contained in a
finite product fl E: (IV, p. 12, prop. 13), where H is a finite subset of I; consequently M is
lEH

bounded for cr(E', E). Deduce from this that G is barrelled.

b) Let F be a subspace of E such that G c FeE and such that G is a hyperplane (everywhere
dense) in F; F is barrelled (exerc. 15). Show that F is not bomological. (Argue by reductio ad
absurdum; if there were a convex, balanced and bounded set A in F such that G is an everywhere
dense hyperplane in the normed space FA (III, p. 7), then there would exist a sequence of
points of G converging to a point of F not belonging to G) (ef IV, p. 52, exerc. 2). *

17) a) Let E be a Hausdorff locally convex space, L a vector subspace of E of finite codimen-
sion, and T a bomiverous barrel in L. Show that there exists a bomiverous barrel T' in E
such that T' n L = T. (Reduce to the case where L is a hyperplane in E. Let Eo be the bomo-
logical space associated with E (III, p. 40, exerc. 1), Lo the hyperplane L with the topology
induced by that of Eo; observe that T is a neighbourhood ofO in Lo and consider the following
two cases: that Lo is dense in Eo, or is closed in Eo; show that for T' we can take the closure
T of T in E or the sum of T and a compact convex set of dimension 1).
b) Let E be an infra barrelled space, L a vector subspace of E of finite codimension. Deduce
from a) that L is infra barrelled (el IV, p. 64, exerc. 11).

-r 18) Let E be a strict inductive limit space of an increasing sequence (En) of locally convex
metrizable subs paces (II, p. 33), and let F be a vector subspace of E such that every point
of E is a limit point of a sequence of points of F.
a) If En is the closure of En in E, then E is the strict inductive limit of the sequence (EJ Let
F" be the closure of F n E" in E. Show that E is the union of the increasing sequence of sub-
spaces F".
b) Suppose E is barrelled. Show that F is bomological. (Let u be a linear mapping from F
into a Banach space G which transforms every bounded subset of F into a bounded subset
of G. Show that there exists a linear mapping from E into G, whose restriction to F is equal
to u, and whose restriction to each F" is continuous. Finally use exerc. 13, b) of Ill, p. 44.)

19) A Hausdorff locally convex space E is said to be ultrabornologieal if every convex subset

neighbourhood of °
of E which absorbs all the convex, balanced, bounded and semi-complete subsets of E, is a
in E.
a) Show that every ultrabomological space is both bomological and barrelled.

°
b) Let E be a Hausdorff locally convex space such that the closed, convex, balanced envelope
of the set of points of every sequence tending to is semi-complete. Show that if E is bomo-
TVS 111.46 SPACES OF CONTINUOUS LINEAR MAPPINGS §5

logical then it is uitrabornological. In particular every bornological and quasi-complete

space is ultrabornological; every Frechet space is ultrabornological.
c) Let (E,) be a directed increasing family of vector subspaces of a vector space E such that,
E is the union of the E". Let Y" be a locally convex topology on E, for every Ct., and let Y be
the finest locally convex topology for which the canonical injections from E. into E are conti-
nuous. Suppose that Y is Hausdorff and that, for every Ct., the topology on E. induced by Y
is Y". Show that if each ofthe spaces E" is uitrabornological, then E with Y is ultrabornological.
d) Show that every finite product of ultrabornological spaces is uitrabornological; deduce
that every topological direct sum of ultrabornological spaces is ultrabornological.
I
rye

e) Show that the product space E = Ell of an infinite sequence of ultrabornological

11=0
spaces is ultrabornological. (Let A be a convex subset of E which absorbs every convex,
balanced, bounded and semi-complete subset of E. Show that if A were not a neighbourhood
of 0 in E, then there would have existed a sequence (XII) in CA such that x" has its first n - 1
coordinates zero, but is =1= 0. Next observe that the closed convex balanced envelope of the set of
7 oc
points of such a sequence is identical to the set of points I A"X", where I 1\,1 :s; I, and
//= 0 11=0
that this envelope is a semi-complete set.)

20) Show that, for a Hausdorff locally convex space E to be ultrabomological it is necessary
and sufficient that it is the inductive limit ofa family of Banach spaces. (To see that the condition
is necessary, consider the convex, balanced, bounded and semi-complete sets B in E, and the
spaces E B . To see that it is sufficient, observe that ifE is the inductive limit of a family of Banach
spaces E., we can assume that the E, are (algebraically) subspaces of E; if V is a convex set
in E which absorbs the convex, balanced, bounded and semi-complete subsets of E, show
that V absorbs each ball B" of E" (argue by reductio ad absurdum); if V does not absorb B",
then it does not absorb a sequence (x,,) of points of B., tending to 0 in E; then use the fact
that in a Banach space, the closed, convex envelope of a compact set is compact.)

21) Show that if E is a Hausdorff locally convex semi-complete space, then the bomological
space associated with E (III, p. 40,exerc. I) is uitrabomological.

~ 22) Let E be an infinite dimensional Banach space satisfying the first axiom of counta-
bility.
a) Show that the set oX of all compact, convex and balanced subsets A of E such that EA
is infinite dimensional, is infinite and has a cardinality :s; 2Card (N) (GT, IX, § 5, exerc. 17).
For every Xo E E and every A E oX, the set Xo + A contains a free subset of cardinality 2Card (N)
(II, p. 80, exerc. 24, c».
b) Let Xo =1= 0 be in E. Show that there exists a family (YA)Ad' such that Xo and the YA form a
free family and that we have YA E Xo + A for all A E x· (well order oX and argue by transfinite
induction, using a».
c) Let f E E* be a linear form such that f(x o) = 1 and f(YA) = 0 for all A E oX, and let
H = f - 1 (0). Show that a subset M of H which is convex, balanced and semi-complete is neces-
sarily finite dimensional (observe that if not, M will contain an infinite dimensional compact
convex and balanced set A; hence YA will belong to H n (xo + M»).
d) Show that H with the topology induced by that of E is not ultrabornological, in spite of
being bomological and barrelled (III, p. 44, exerc. 14). (By using c), show that if H were ultra-
bomological, its topology would have been the finest locally convex topology, and deduce
a contradiction.)

§5

I) Let E, F and G be three locally convex spaces, 6 a cover of E consisting of bounded sets.
Show that if u is a separately continuous bilinear mapping from E x F into G such that for
every set M E 6 the restriction of u to M x F is continuous, then u is 6-hypocontinuous.
§5 EXERCISES TVS III.47

2) Let E be the direct sum space R(N), with the topology induced by the product topology
w
on RN. Show that the bilinear form ((x,,), (y,)) I-> L x"Yn on E x E, is separately continuous,
n= 0
but that for every set 6 of bounded subsets of E containing at least one infinite dimensional
bounded set, this bilinear form is not 6-hypocontinuous.
3) Let E be the space R(N) with the finest locally convex topology (II, p. 26); let F be the space
RN; the space E is ultrabornological (Ill, p. 45, exerc. 19) and complete, whilst F is metrizable
and complete. Let 6 (resp. ::t) be the set of all bounded subsets of E (resp. F). Show that the
Yc

bilinear form ((x n ), (y,,)) I-> L x"y" on E x F is (6,::t)-hypocontinuous, but is not continuous
11= 0

(ef IV, p. 48, exerc. 11).

4) Let E be a locally convex space, Fan infra barrelled space (III, p. 44, exerc. 7) and ::t the set
of all bounded subsets of F. Show that, if a bilinear mapping from E x F into a locally convex
space G is ::t-hypocontinuous, it is (6, ::t)-hypocontinuous for every set 6 of bounded subsets
of E (c1 III, p. 44, exerc. II).

5) a) Let E, F and G be three Hausdorff locally convex spaces, and u a bilinear mapping from
E x F into G. In order that there exist a balanced neighbourhood U of 0 in E such that the
set of all mappings u(x, .), where x runs through U, is equicontinuous in Sf(F; G), it is neces-
sary and sufficient that u is continuous when we replace the topology of E by the coarsest
topology for which the sets AU (A oF 0) form a fundamental system of neighbourhoods of O.
Show that if G is normed, this condition is satisfied by every continuous bilinear mapping
from E x F into G.
b) Take for E, F and G the product space R N , and for u the continuous bilinear mapping
((x,'), (y,,)) I-> (xnyJ Show that there does not exist any neighbourhood U of 0 in E such that
the set of maps u(x, .), where x runs through U, is equicontinuous in Sf(F; G).

6) Let E, F and G be three topological vector spaces. A set H of bilinear mappings from E x F
into G is said to be separately equieontinuous iffor alf x EO E, the set of linear mappings u(x, .),
where u runs through H, is equicontinuous in Sf(F; G) and if for all yEO F, the set of linear
mappings u(., y), where u runs through H, is equicontinuous in Sf(E: G).
Suppose that F is metrizable, and that E is a Baire space (el III, p. 43, exerc. 5 and V, p. 79,
exerc. IS). Show that every separately equicontinuous set of bilinear mappings from E x F
into G is equicontinuous (ef III, p. 42, exerc. II).

7) Let E, F and G be three topological vector spaces, 6 a set of bounded subsets of E, and
H a set of separately continuous bilinear mappings from E x F into G. The following pro-
perties are equivalent:
ex) For every neighbourhood W of 0 in G and every set M EO 6, there exists a neighbourhood
V of 0 in F such that u(M x V) c W for all u EO H.
~) For every set M EO 6, the image of H x M under the mapping (u, x) I-> u(x, .) is an equi-
continuous subset of Sf(F; G).
y) As u runs through H, the set of mappings y I-> u(., y) from F into Sf3 (E; G) is equicon-
tinuous.
We then say that H is a 6-equihypoeontinuous set of bilinear mappings (separately continuous)
from E x F into G. Similarly for a set 'J: of bounded subsets of F, we define the notions of a
'J:-equihypoeontinuous set and a (6, 'J:)-equihypoeontinuous set.
8) Let H be a 6-equihypocontinuous set of bilinear mappings from E x F into G (exerc. 7).
For every subset M EO 6, show that H is equicontinuous in M x F; moreover, for every
bounded subset Q of F, the union of the sets u(M x Q), where 11 runs through H, is bounded
in G.

9) Let H be a (6, 'J:)-equihypocontinuous set of bilinear mappings from E x F into G

(exerc. 7); show that for every pair of sets M EO 6, N EO ::t, H is uniformly equicontinuous in
M x N.
TVS III.48 SPACES OF CONTINUOUS LINEAR MAPPINGS §5

10) Let El' E z and F be three topological vector spaces, G 1 (resp. G 2 ) an everywhere densc
subspace of El (resp. E z), and 6 1 (resp. 6 z) a family of bounded subsets of G 1 (resp. G z ).
Let H be a set of separately continuous bilinear mappings from El x E z into F: if the set of
restrictions to G 1 x G z of the mappings u E H is (61' 6 z)-equihypocontinuous. then so
is H.

11) If F is a barrelled space, every separately equicontinuous set of bilinear mappings from
E x F into a locally convex space G is 6-equihypocontinuous for every set 6 of bounded
subsets of E.

12) Let E, F be two topological vector spaces, and let f be the bilinear mapping (x, u) I---> u(x)
from E x ~(E; F) into F; let .r be a topology compatible with the vector space structure
of ~(E; F) and finer than the topology of simple convergence. Let 2: be a family of bounded
subsets of E, U a family of bounded subsets of ~(E; F) (for the topology .r). Show that f is
2:-hypocontinuous if and only if :Y is finer than the G-topology; f is U-hypocontinuous if
and only if the sets of U are equicontinuous subsets of ~(E; F).
13) Let E, F, G be three topological vector spaces, and 2: (resp. 2) a family of bounded subsets
of E (resp. F). Let H be the vector space of 2-hypocontinuous bilinear mappings from E x F
into G.
a) Show that on H the topology of uniform convergence on sets of the form M x N, where
M E 6 and N E 2 is compatible with the vector space structure; this topology is called the
(6, 2)-topology on H. For every mapping U E H, let it be the continuous mapping x I---> u(x, .)
from E into P,! (F: G). Show that U I---> it is an isomorphism from the space H. endowed with
the (6, 'l:)-topology, onto the space ~~ (E; ~,!(F; G»).
b) Let L be a subset of H such that, for every pair (x, Y) E E x F, the set of all u(x, y), where U
runs through L, is bounded in G (simply bounded subset of H). Show that, ifE, F, G are locally
convex, and ifE and F are Hausdorff and quasi-complete, then L is bounded in H for the(6, 'l:)-
topology.
c) Let E, F, G be three Hausdorff locally convex spaces. IfE is barrelled and F quasi-complete
or barrelled, then every simply bounded subset L of H is 'l:-equihypocontinuous (III. p. 47.
exerc. 7).
d) If E and F are barrelled, and G quasi-complete, and if 6 and 'l: are coverings of E and F
respectively, then H is Hausdorff and quasi-complete for the (6, 'l:)-topology.

14) Extend the definitions and rcsults of § 5 to arbitrary multilinear mappings. Let E, F, G
be three topological vector spaces, 6 (resp. 'l:) a family of bounded subsets of E (resp. F),
and U a family of bounded subsets of the space ~~ 'l (E. F; G) of bilinear (6. 'l:)-hypoconti-
nuous mappings from E x F into G, endowed with the (6, 'I)-topology (exerc. 13). Show
that the trilinear map (x, y, u) I---> u(x, .1') from E x F X 2'~.'l (E, F; G) into G is (6, 'l:)-
hypocontinuous; in order that it is (6, U)-hypocontinuous, it is necessary and sufficient
that every set LEU is 6-equihypocontinuous (III. p. 47, exerc. 7).

~ 15) Let E be the space of all sequences x = (~"),,;,,o of real numbers such that the series
with the general term ~n is convergent. Put I x I = sup I I" ~k I·
II k=O
a) Show that I x II is a norm on E, and that E is complete for this norm.
b) Show that the vector space r1 eN) (I, p. 4), considered as a subspace of E, is everywhere
dense (for the topology of E); the topology on e1(N) defined by the norm II xiiI = I
11=0
I ~n I
is strictly finer than the topology induced by that of E.
c) Let (P,,) be an increasing sequence of finite subsets of N x N forming a cover of N x N.
For every x = (~,,) E E and every .1' = (11,,) E e1 (N), let j,,(x. y) = I ~illj' Then the sequence
(i,jEP n
U~(x, ,F») tends to a limit for every pair (x, y) E E x e1 (N) if and only if for each of these pairs
en 00

(x, y), the sequence (f.,(x, y») is bounded; the limit of f,,(x, y) is then equal to (I ~,,) ( Ill,,).
n=O 11=0
§6 EXERCISES TVS 111.49

(Using exerc. 13, c) of III, p. 48, show that the sequence of bilinear forms (f,,) is equicontinuous,
and observe that it converges in the subspace fl(N) x fl(N); conclude the argument using b).)
d) For every j E l\;, let Pi" be the smallest number of closed intervals of N whose union is
the projection ofP n n (N x {j}) onto N; let p" = sup Pj". Show that the condition obtained
jeN
in c) is equivalent to sup p" < + CfJ. (If q,,, is the characteristic function of P n' show that
"
(I
00

the norm of thc bilinear form 1;, is sup I p"Ci, j) - q,,,(i + I, j) I).)
jeN i= 0

§6

~ 1) An exhaustion of a Hausdorff locally convex space is given by a sieve C = (C", P,,),,~o

(GT. IX. § 6, No.5, def. 8) and, for every n ? 0, a mappmg P" from C" into the set of all convex
and balanced subsets of E having the following properties:
EI) E is the union of the cPo(c) where c ranges over Co;
E2) for every n and every C E c,,' cP,,(c) is the union of the cP,,+ I (c') where c' ranges over Pn-I(c);

°
E3) for every sequence (Ck)k ~ 0 such that ck E C k and ck = Pk( Ck + 1) for all k ? 0, there exists
a sequence (cPk) of numbers> such that, for every sequence (xkJ of points of E such that
°
x k E cPk(Ck ) and every sequence (/'k) of real numbers satisfying ,s; j'k ,s; Pk for all k, the series
I
7J

AkXk is convergent in E.
k~O

a) Under the above hypotheses, show that if in addition the q,..(c) are closed for CE Cn' then
I
00

we can assume that the Pk have been so chosen that we have AkXk E q,m(cm) for all m ?

f
k=m
(take the Pk such that pk,s; 1).
k~O

b) Suppose that we are given a sieve C and sequence (q,,,) of mappings into the set of convex
and balanced subsets of E satisfying El), E2) and the following condition:

(~k) of numbers> °
E3') for every sequence (ckJ UO such that ck = Pk(C k+ l ) for al1 k ? 0, there exists a sequence
such that, for every sequence (xk) of points of E with Xk E <h(ck) for all k,
the sequence of points (flkXk) is contained in a convex, bounded balanced and semi-complete
set in E.
Show that then the condition E3) is also verified (take Pk = 2- k flk)'
A locally convex Hausdorff space is said to be exhaustible if there exists an exhaustion
of E.

~ 2) Let E be a locally convex space which is a Haire space, F a locally convex exhaustible
space (exerc. 1), and (C", P", q,,,) an exhaustion of F.
a) Let u be a linear mapping from E into F and let W be a convex, balanced and absorbent
set in F. Show that there exists a sequence (c k ) such that ck E C k , ck = p/Ck+ 1) for all k ? 0,
and a sequence (mk ) of integers > 0, such that each of the sets U- 1 (q,k(Ck ) n mk W) (which
is denoted by M k ) is not a thin set in E. Show that for every E > 0, there exists a sequence
(v k) of numbers > 0 such that if the sequence (xk)u I of points of E is such that x k E vkM k
I
00

for all k ? 1. the serie U(X k ) converges in F and that its sum belongs to EW.
k ~ 1
b) Suppose, in addition, that E is metrizable and that the graph of u in E x F is closed. Show
that for every <: > 0, we have U- leW) c (1 + E) U-I(W). (Observe that if (Uk) is a countable
fundamental system of neighbourhoods of 0 in E, then for every k there exists a convex balanced
neighbourhood Vk of 0 in E such that Vk C Uk n vkM k . For every point aE II-I(W), find
k
a sequence (Xk)k~O such that Xo E u-I(W), x k E vkM k for k;;, 1 and a - I Xj E Vk for all
j~O

k ? 1, then apply a).)

TVS III.50 SPACES OF CONTINUOUS LINEAR MAPPINGS §6

c) Deduce from b) that if E is a metrizable Baire space, then every linear mapping from E
into F whose graph is closed is continuous.

3) Show that a Frechet space E is exhaustible (if (Uk) is a decreasing sequence forming a
fundamental system of closed, convex and balanced neighbourhoods of 0 in E, consider
the finite intersections of the sets (m + 1) Uk' where m, and Ie run through N).

4) a) Every closed subspace of an exhaustible locally convex space is exhaustible.

b) Let E be an exhaustible locally convex space, and u:E --> F a continuous linear surjective
mapping from E into F. Show that F is exhaustible. In particular, every quotient space of E
by a closed subspace of E is exhaustible. Every space obtained by assigning to E a Haus-
dorff locally convex topology coarser than that of E is exhaustible.

5) Let (c(m))m;'O be a sequence of sieves c(m) = (c~m), p~m))n;'O' For every n ;;:. 0, put
00

Dn = qO) X C~l~l x ... x cg') x Il {am},

m--:-:/I +1

where am = 0 for all 111 ;;:. 0: the mapping Pn : Dn+ 1 --> Dn is taken to be equal to

fi~,O) x p;/~ 1 x ... x pg') x q(n+l) x Il id",

m=1l +2

where q(n+l) is the unique mapping from C~+l) onto {O}, and idm is the identity map of {am}'
Then (D ,fin) is a sieve.
Let (E(m))m;,o be a sequence of Hausdorff locally convex spaces; we assume that for each
In there exists an exhaustion (c~m), p~n), cp~m))n? ° of E(m) Consider the Hausdorff locally convex

space E = Il E(m), and for every n, put

00

cp" = cp~O) X cp~l~ 1 X ... X cpg') X Il o/m

m=n+ 1

where o/m is the mapping from {am} into the set of convex and balanced subsets of E(m) such
that o/"(a m) = E(m) Show that (D n, Pn' CPn) is an exhaustion on the product space E.

6) Show that an inductive limit (II, p. 31) of an increasing sequence of subspaces En of a

vector space E, with topologies I". such that E" endowed with .'7" is exhaustible, is an exhaus-
tible locally convex space, if it is Hausdorff.
 CHAPTER IV

Duality in
topological vector spaces

Throughout this chapter, all the vector spaces under consideration are vector spaces
over a .field K which is either R or C.

§ 1. DUALITY

1. Topologies compatible with a duality

In this section, E and F denote two vector spaces put into duality by a bilinear
form B (II, p. 40). We recall (II, p. 41) that we defined two linear mappings

characterized by the relation

(1) B(x, y) = (x, dB(y) = (y, SB(X)

for x E E, y E F.

DEFINITION 1. - A locally convex topology :!I on E is said to be compatible with

the duality between E and F !f dB is a bijection from F onto the dual of the locally convex
space obtained by assigning the topology :!I to E.

If there exists one such topology :!I, the mapping dB is injective, that is to say,
the duality between E and F is separating in F (II, p. 41).

PROPOSITION 1. - (i) The closed convex subsets in E are the same for all the locally
convex topologies on E which are compatible with the duality between E and F.
(ii) The bounded subsets of E are the same for all the locally convex topologies
on E which are compatible with the duality between E and F.
TVS IV.2 DUALITY IN TOPOLOGICAL VECTOR SPACES §1

Let :Y be a topology on E compatible with the duality between E and F, hence

finer than cr(E, F). If a convex subset of E is closed for :Y, it is the intersection of
closed, real half-spaces (II, p. 38, cor. 1), hence it is closed for cr(E, F). This proves (i).
Assertion (ii) was proved in cor. 3 of III, p.27.
Let F cr denote the vector space F endowed with the weak topology cr(F, E). Then
the linear mapping SB maps E onto the dual (F cr)' of F cr (II, p. 43, prop. 3). Let 6
be a family of bounded subsets of F cr' By abuse of language, the inverse image under
SB of the 6-topology on (F cr)' is called the 6-topology on E. It is defined by the family
of semi-norms

(2) PA(X) = sup IB(x, y)l,

YEA

where A runs through 6. In particular, when 6 is the family of finite subsets of F,

the 6-topology is precisely the weak topology cr(E, F).

DEFINITION 2. - Let E and F be two spaces in duality. The Mackey topology on E,

denoted by T(E, F) is defined as the 6-topology on E, where 6 is the family of all sub-
sets of F whose image in E* (under dB) is convex, balanced and compactfor cr(E*, E).
When the duality between E and F is separating in F, dB is injective and the topo-
logy cr(F, E) on F is the inverse image under dB of the topology cr(E*, E) on E*.
In this case, 6 consists of all those subsets of F which are convex, balanced and
compact for cr(F, E).
In general, if F 1 = dB(F) c E*, and if we denote by (x, Y1) f---+ B1 (x, Y1) the res-
<
triction of the canonical bilinear form (x, x*) f---+ x, x* > to E x F l ' then E and
F 1 are put in duality by B 1, and this duality is separating in F l ' since by definition
we have B(x, y) = B1 (x, dB(y)), def 2 shows that T(E, F) = T(E, F 1)'

Remark 1. - Let A be a compact convex subset of a Hausdorff locally convex

space G, and let A be the closed convex balanced envelope of A. When the field K
is R, the set A is the closed convex envelope of A u (- A); when K is C, the set A
is contained in the closed convex envelope of 2A u ( - 2A) u (2iA) u ( - 2iA).
Consequently (II, p. 14, prop. 15), A is compact.
We deduce, in particular, that when the duality between E and F is separating
in F, the Mackey topology T(E, F) is also the 6'-topology, where 6' is the set of all
convex subsets of F which are compact for cr(F, E). In an analogous way we define
the Mackey topology T(F, E) on F.

THEOREM 1 (Mackey). - Let E and F be two spaces in duality; suppose that the dua-
lity is separating in F. In order that a locally convex topology :Y on E be compatible
with the duality between E and F, it is necessary and sufficient that :Y be .finer than
the topology cr(E, F) and coarser than the Mackey topology T(E, F).
Identify F with its image in E* under dB' Let 6 0 denote the set of all subsets of F
which are convex, balanced and compact for cr(F, E). By definition, T(E, F) is the
6 0 -topology on E, hence is finer than cr(E, F).
No.1 DUALITY TVS IV.3

Lemma 1. - The subspace F ofE* consists of all linear Iorms on E which are conti-
nuous lor ,(E, F).
Every element of F is a continuous mapping for cr(E, F), hence for ,(E, F).
Conversely, let IE E* be continuous for ,(E, F). There exists a neighbourhood U
of 0 in E (for ,(E, F)), such that iIi :( 1 on U; we can assume that there exists a
set A E 6 0 such that U = A In other words, I belongs to the bipolar A of A
0. 00

for the duality between E* and E. But the topology cr(F, E) on F is induced by
cr(E*, E); consequently A is convex, balanced and compact for cr(E*, E), and the
theorem of bipolars (II, p. 44, tho 1) implies the equality A = AOo. Therefore we
have that IE F, from which the lemma follows.

Lemma 2. - Let :Y be a locally convex topology on E such that every linear Iorm
on E which is continuous lor :Y belongs to F. Then:Y is coarser than ,(E, F).
Let U be the set of convex, balanced neighbourhoods of 0 for :Y. Let 6 be the
set of polars in F of elements of U. By cor. 2 of III, p. 17, we have 6 c 6 0 , and by
cor. 1 of prop. 7 of III, p. 19, :Y is identical with the 6'-topology, where 6' is the
set of po lars of sets of U in the dual E' of E. But E' c F, by hypothesis, hence every
set of 6' is contained in a set of 6; and the lemma follows.
Let :Y be a topology on E compatible with the duality between E and F. Then :Y
is coarser than ,(E, F) by lemma 2, and evidently :Y is finer than cr(E, F). Conver-
sely, F is the dual ofE for the topology ,(E, F) (lemma 1) and for the topology cr(E, F)
(II, p. 43, prop. 3), hence also for every topology intermediate between ,(E, F) and
cr(E, F).

COROLLARY. - Let p be a semi-norm on E. Thefollowing conditions are equivalent:

(i) p is continuous lor the topology ,(E, F);
(ii) every linearform I on E, such that iIi :( p, comes lrom an element of F.
=
(i) (ii) : if p is continuous for ,(E, F), every linear form I on E such that iIi :( p
is continuous for ,(E, F), hence comes from an element of F by lemma 1.
=
(ii) (i) : let:Y be the topology on E defined by the semi-norm p. If condition (ii)
is satisfied, the linear forms on E which are continuous for :Y belong to F. By lemma 2
:Y is coarser than ,(E, F), hence p is continuous for ,(E, F).

Remark 2. - * Let K be a convex subset of F which is compact for the weak topo-
logy cr(F, E) and ~ a positive measure on K. Put

p(x) = L IB(x, y)1 d~(y)

for all x E E. It is immediate that p is a semi-norm. Moreover, for every x E E, the
relation « IB(x, y)1 :( 1 for all y E K» implies that p(x) :( ~(K). This proves that
the semi-norm p on E is continuous for the Mackey topology ,(E, F). *

Example. - Let G be a locally convex space and G' its dual. On G', the weak topo-
logy cr(G', G) and the topology of convex compact convergence (III, p. 14) are
TVS IV.4 DUALITY IN TOPOLOGICAL VECTOR SPACES § 1

compatible with the duality between G' and G. In general, the strong topology
and the topology of compact convergence on G' are not compatible with the duality
between G' and G. Recall however that when G is Hausdorff and quasi-complete,
the topology of compact convergence on G' coincides with that of convex compact
convergence (III, p. 8), hence is compatible with the duality between G' and G.

DEFINITION 3. - Let E and F be two vector spaces in duality, and:Y the family of
subsets ofF which are boundedfor cr(F, E). Then the :Y-topology on F is denoted by
~(E, F).

Similarly, we define the topology ~(F, E) on F. It can be seen easily that the topo-
logy ~(E, F) is identical with ~(E, F lEO), and we can reduce to the case when the
duality between E and F is separating in F.
Remarks. - 3) Let E" denote the space E endowed with the topology cr(E, F). The
barrels (III, p. 24) in E" are the subsets of E which are convex, balanced closed and
absorbent for cr(E, F). These are none other than the polars of the subsets of F
which are convex, balanced and bounded for cr(F, E). Consequently, the family of all
barrels in E" is a fundamental system of neighbourhoods of 0 for the topology peE, F)
in E. In other words, a semi-norm on E is continuous for peE, F) if and only if it is
lower semi-continuous for cr(E, F) (ef III, p. 24, prop. l).
4) Let Y be a topology on E compatible with the duality between E and F. By
prop. I, (ii) ofIV, p. I, the topology P(F, E) on F is none other than the strong topology
on F, when F is identified with the dual of E (with the topology Y).
5) The topology peE, F) on E is finer than ,(E, F). It is not, in general compatible
with the duality between E and F (ef however § 2). In particular, a subset of E which
is bounded for cr(E, F) is not necessarily bounded for peE, F).

2. Mackey topology and weakened topology on a locally convex space

Let E be a locally convex space and E' its dual. We put E and E' in duality by
means of the canonical bilinear form (x, x') f---+ <x, x') on E x E'. This duality is
separating in E'. We shall consider three topologies on E compatible with the duality
between E and E' :
a) the given topology on E, which we shall call the initial topology, whenever
any confusion is likely to arise;
b) the topology cr(E, E'), called the weakened topology on E;
c) the topology T(E, E'), called the Mackey topology on E.
The initial topology is finer than the weakened topology and coarser than the
Mackey topology; moreover, these three topologies can be distinct (IV, p. 49,
exerc. 8).
By prop. 1 of IV, p. 1, these three topologies have the same closed convex sets,
the same barrels, the same bounded sets and the same adapted bomologies. In
particular :

PROPOSITION 2. - Let E be a locally convex space, and let A be a convex subset ofE
(for example, a vector subspace of E). The closure of A is the same for the initial topo-
logy and for the weakened topology of E.
No.2 DUALITY TVS IV.S

Remarks. ~ 1) For a family (X)iEI of elements of E to be total (resp. topologically

independent) for the initial topology, it is necessary and sufficient that it is so for
the weakened topology; this follows from prop. 2. Hence we can apply the criteria
of II, p. 43.
2) Let 51 and 52 be two locally convex topologies on E, compatible with the
duality between E and E', 51 being finer than 52' Then every neighbourhood of 0
for 51' which is convex and closed for .Y'1 is closed for 52 by prop. 1 of IV, p. l.
Consequently (GT, II, § 3, No.3, corollary) every subset ofE which is complete for 52
is so for 51 also.
In particular, every subset of E which is complete for the weakened topology is
complete for the initial topology, and every subset of E complete for the initial
topology is so for the Mackey topology. If E is quasi-complete for the weakened
topology, it is so for every topology compatible with the duality between E and E'.
If it is quasi-complete for the initial topology, it is so for the Mackey topology.
3) Suppose E is Hausdorff (for the initial topology). Let A be a subset of E which
is closed and bounded for o{E, E'), hence also for every topology compatible with
the duality between E and E'. Since A is precompact for cr(E, E') (III, p. 3, Remark 5),
assuming that A is compact for cr(E, E') is equivalent to A being complete for cr(E, E').
Therefore, on account of remark 2, we see that:

PROPOSITION 3. ~ Suppose E is Hausdorff; and E' its dual. Every subset of E which
is precompact for the initial topology and compact for cr(E, E'), is compact for the
initial topology.

4) The topology ~(E, E') (IV, p. 4, def 3) is finer than the Mackey topology. If
~(E, E') is distinct from ,(E, E'), it is not compatible with the duality between E
and E'. The space E is barrelled if and only if the initial topology is equal to ~(E, E')
(III, p. 24).

PROPOSITION 4. ~ Let E be a locally convex ~pace. The Mackey topology on E is

identical with the initial topology in each of the following cases :
a) E is barrelled;
b) E is bomological;
c) E is metrizable.
We note first that the Mackey topology on E is identical with the initial topology
if and only if every convex subset of E' which is compact for cr(E', E), is equi-conti-
nuous. This is certainly the case if E is barrelled (III, p. 24, corollary).
Suppose E is bomological; let V be a convex and balanced neighbourhood of 0
in E for the topology ,(E, E'). Let B be a subset ofE, boundedfor the initial topology.
Since B is bounded for the Mackey topology, V absorbs B, and since E is bomolo-
gical, V is a neighbourhood of 0 for the initial topology.
In case c), the space E is bomological (III, p. 12, prop. 2).
TVS IV.6 DUALITY IN TOPOLOGICAL VECTOR SPACES §1

3. Transpose of a continuous linear mapping

In this section, E j and E z denote two locally convex spaces, with respective
duals E~ and E~.
Let u be a linear mapping from El into E z . For u to be continuous when E j and
E z are assigned the weakened topologies, it is necessary and sufficient that f 0 u
belongs to E~ for all f E E~ ; this is the case if u is continuous. Then the linear mapping
f f---+ f 0 u from E~ into E~ is called the transpose of u and is denoted by tu.
PROPOSITION 5. - Let u be a continuous linear mapping from E j into E z .
(i) If E j and E z are Hausdorff then u is injective if and only if the image of tu is
dense in E~ for the weak topology cr(E~ : Ej)'
(ii) For tu to be injective it is necessary and sufficient that the image of u is dense
in E z .
A vector subspace of E2 is dense for the initial topology if and only if it is dense
for the weakened topology (IV, p. 4, prop. 2). Prop. 5 then follows from II, p. 47,
cor. 2.

PROPOSITION 6. - Let u be a linear mapping from E j into E2 which is continuous

for the weakened topologies. For i = 1, 2, let 6; be afamily of bounded subsets ofE;.
In order that tu is a continuous mapping from (E~) 62 into (E~)6' it is necessary and
sufficient thatJor every set A E 6 1 , there exist sets A j , ... , An in 6 2 and a real number
A > 0 such that A. u(A) is contained in the closed convex balanced envelope of
Al U ... U An 1.
This is an immediate consequence of prop. 2 of III, p. 15.

COROLLARY. - Let u be a continuous linear mapping from El into E 2. Then tu is

continuous when the duals E; are assigned the following topologies :
a) the weak topologies cr(E;, E;);
b) the strong topologies ~(E;, E);
c) the Mackey topologies ,(E;, E);
d) the topologies of precompact convergence.
Moreover, ifE2 is Hausdorff, tu is continuous when the duals E; are assigned:
e) the topologies of compact convergence (resp. compact convex).
The only point which requires a proof is the case c), when the topologies of E j
and E2 are not necessarily Hausdorff. Then for every linear form f E E~*, f 0 tu is
a linear form on E~ ; hence there is a linear mapping v: E~* ---+ E~* which is continuous
for the topologies o{E~*, E~) and cr(E~*, E~) and is such that dB2 0 u = v 0 dB"
where dB; is the canonical mapping from E; into E;* (i = 1, 2). Consequently, if A
is a subset of E j such that dB,(A) is convex, balanced and compact for cr(E~*, E~)
then dB2(u(A)) = v(dB,(A)) is convex, balanced and compact for cr(E~*, E~) since
the topologies cr(E~*, E~) and cr(E~*, E~) are Hausdorff.

1 In other words, u(6 j ) is contained in the smallest adapted bomology containing 6 2

(III, p. 3).
No.3 DUALITY TVS IV.7

PROPOSITION 7. - Let u: El -+ E2 be a linear mapping. We assume that u is conti-

nuous for the weakened topologies of El and E 2 .
(i) The mapping u is continuous ifEl and E2 are assigned their Mackey topologies.
(ii) IfEl is bornological or barrelled, then u is continuousfor the initial topologies
of El and E 2 ·
(iii) In order that u be continuousfor the initial topologies ofE l andE 2 , it is necessary
and sufficient that the image under tu of every equicontinuous subset of E~ be equi-
continuous in E~.
The hypothesis implies that tu is continuous for the weak topologies o{E~, E 2 )
and o{E~, E l ) (II, p. 46, corollary) hence the image under tu of a convex, balanced
and compact subset for (J(E~, E 2) is convex, balanced and compact for a(E~, E l ),
the topologies a(E~, E z) and (J(E~, E l ) being Hausdorlf. Therefore, assertion (i)
follows from GT, X, § 1, No.4, prop. 3, b). Assertion (ii) is a consequence of (i) : for, if El
is bomological or barrelled, its initial topology is the Mackey topology, and the
Mackey topology of E2 is finer than the initial topology of E 2 . Finally, the initial
topology of Ei is that of uniform convergence on equicontinuous subsets of E;
(III, p. 19, cor. 1 of prop. 7). This proves (iii).

COROLLARY. - Suppose El is a normed space. Let u be a linear mapping from El

into E 2 • The following properties are equivalent :
a) u is continuous;
b) u is continuous for the weakened topologies;
c) the image of the unit ball in El under u is bounded in E2 ;
d) for every sequence (x n) of points of El tending to 0 for the initial topology, the
sequence (u(x n )) is bounded for the weakened topology of E z .
Since El is bomological the equivalence of a) and b) follows from prop. 7; that
of a) and c) is immediate. The equivalence of a) and d) follows from prop. 1 of IV,
p. 1, and from prop. 1 of III, p. 11.

PROPOSITION 8. - (i) Let E be a normed space, with dual E'. For every x E E, we
have

(3) Ilxll sup I<x, x')I·

x'EE', II x' II '" 1

(ii) Let El and E2 be two normed spaces and u a continuous linear mapping from
El into E2 . We have

(4)

Let x E E. For every x' E E' such that Ilx'll .;:; 1, we have
I<x, x')1 .;:; Ilxll·llx'll .;:; Ilxll·
By Hahn-Banach theorem (II, p. 23, cor. 2), there exists an element x' in E' such
that Ilx'll .;:; 1 and <x, x') = Ilxll. This proves (i).
TVS IV.S DUALITY IN TOPOLOGICAL VECTOR SPACES §1

Let us now prove (ii). By formula (3) and the definition of the transpose, we have

I lull sup Illu(y')11 = sup I<X,IU(y')1

Ily'll U Ily'll u,llxll U

sup l<u(x),y')1 = sup Ilu(x)11 = Ilull·

Ilxll~l,IIY'IIU IlxilU

Remarks, - 1) Formula (3) is a particular case of (4), corresponding to the linear

mapping A f-+ AX from K into E.
> >
2) Put B(x, y') = <u(x), y' = <X, 'U(y') for x EEl' y' E E~. The above proof
shows that B is a continuous bilinear form on El x E~, with norm (GT, X, ~ 3, No.2)
equal to II u II.

COROLLARY. - Let E be a normed space satisfying the first axiom of countability.

There exists a countable subset 0 of E' - {O} such that we have

(5) Ilxll = sup I<x, ol/II~II

~ED

for all x E E.
Let B' be the unit ball of the dual E' of E with the weak topology a(E', E) assigned
to it. Then B' is a compact metrizable space (III, p. 19, cor. 2); hence there exists
a countable dense subset 0' in B'. Put 0 = 0' n (E' - {O}). Let x E E; the mapping
<
x' f--+ x, x') from B' into K is continuous, therefore

sup I<x, x')1 = sup I<x, 01 ~ sup I<x, ol/II~II ~ Ilxll·

x'EB' ~ED' ~ED

Formula (5) now follows from (3).

4. Dual of a quotient space and of a subspace

Throughout this section, E denotes a locally convex space, M a vector subspace

of E, and MO the orthogonal of M in the dual E' of E. Let p be the canonical mapping
from E onto E/M; then lp is injective, with image MO, hence defines a vector space
isomorphism (not topological)

n : (E/M)' -> M O •

Similarly, let i be the canonical injection from Minto E. Then Ii is surjective (II,
p. 24, prop. 2); its kernel is equal to M and we get a vector space isomorphism (not
O
,

topological)
l:E'/MO -> M' .

PROPOSITION 9. - (i) For a subset A of'(E/M)' to be equicontinuous, it is necessary

and sufficient that n(A) is an equicontinuous subset of E'.
NO.4 DUALITY TVS IV.9

(ii) Let 6 be a set of bounded subsets ofE, and 6 1 the set of the images of subsets
A E 6 in ElM. Then TC is an isomorphism from (E/M)~, onto M where MO is assigned
O
,

the topology induced by that ofE~.

(iii) Suppose E is a normed space then TC is an isometry from the normed space (ElM),
onto the normed subspace MO ofE'.
Let A be a subset of (ElM), and B = tp(A) c E'. Put

q(~) = sup I<~, ~') I

~'EA

for all ~ E ElM. In order that A be ~uicontinuous, it is necessary and sufficient

that the mapping q from ElM into R+ is a continuous semi-norm. This implies
that q 0 p is a continuous semi-norm on E (II, p. 27, prop. 5, (ii)). Since we have

(qop)(x) = sup l<x,x')1

x'EB

for all x E E, this in turn implies that B is equicontinuous in E', and (i) follows.
Let A E 6 and let f be a continuous linear form on ElM. For every A E R+, we
have If I :( A on peA) if and only if Itpcnl :( A on A; hence (ii).
Finally we prove (iii). Let y' be in (ElM)'. An element in ElM has norm < 1 if
and only if it is the image under p of an element of norm < 1 in E. Hence

Ily'll sup I<y, y')1 = sup I<p(x), y')1

YEEfM,IIYII<1 XEE,llxll<1

sup I<x, tp(y') I = Iltp(y')11 '

xEE,llxll <1

and tp induces an isometry from (ElM), onto MO.

PROPOSITION 10. - (i) For a subset A of M' to be equicontinuous, It IS necessary

and sufficient that it is the image under ti of an equicontinuous subset ofE'.
(ii) Suppose M is closed in E. Let 6 be a covering ofE consisting of bounded subsets
and let 6 1 be the set of subsets ofM of theform M n Afor A in 6. The bijective linear
mapping I from E~ IMo onto M~ I is continuous. It is a homeomorphism if6 is a directed
set for the relation c and consists of closed convex and compact sets for cr(E, E').
(iii) IfE is assumed to be normed, then t is an isometry from E'/Mo onto M'.
The image under ti of an equicontinuous subset of E' is an equicontinuous subset
of M' (IV, p. 7, prop. 7). Conversely, let A be an equicontinuous subset of M'. The
topology of M is defined by the set of restrictions to M of the continuous semi-
norms on E. Hence there exists a continuous semi-normp on E such that If(x) I :( p(x)
for all IE A and for all x EM. Let B be the set of all linear forms g on E such that
Igl :( p and whose restriction to M belongs to A. The set B is equicontinuous in E';
by Hahn-Banach theorem (II, p. 23, cor. 1), we have ti(B) = A, hence (i) follows.
We now prove (ii). By prop. 6 of IV, p. 6, the linear mapping ti from E~ into M~l
is continuous, and defines, by passing to the quotient, a continuous linear mapping
TVS IV.lO DUALITY IN TOPOLOGICAL VECTOR SPACES §1

t from E~/Mo onto M'ZI' Let :Y be the topology on M' obtained by transferring
that of E~/Mo by t; this is finer than the 6 1 -topology.
Suppose now that 6 is a directed set for c and consists of closed, convex, balanced
and compact sets for G(E, E'). To show that t is a homeomorphism, i.e. that :Y is
coarser than the 6 1 -topology on M', it is enough to prove that :Y is compatible
with the duality between M' and M and that every equicontinuous set in M (consi-
dered as the dual of M with .'1") is contained in the homothetic of a set belonging
to 6 1 , Since :Y is finer than the 6 1 -topology and 6 1 is a covering of M, the linear
form y' f---* (y, y') on M' is continuous for :Y for every y EM. Let f be a linear
form on M' which is continuous for:Y; then f 0 ti is a continuous linear form on Es'
The 6-topology on E' is coarser than the Mackey topology ,(E', E); for, the mapping
dB: E -* E'* is continuous for the topologies G(E, E') and G(E'*, E'), and since the
latter is Hausdorff, the image under dB of a set which is compact for G(E, E') is
compact for G(E'*, E'). By lemma 1 of IV, p. 3, there exists X o E E such that
fCi(x')) = (xo,x') for all x'EE'. In particular, (xo,x') = 0 for all x'EMo, and
since M is closed in E, we have X o E M (II, p. 45, cor. 2) ; and finally, fey') = (x o, y')
for all y' EM'. This proves that :Y is compatible with the duality between M and M'.
Now let A be a subset of M equicontinuous for the topology :Y on M'. By the
definition of .'1", and in view of the hypothesis that 6 is directed, this means that
there exists a set B E 6 containing 0 and such that the upper bound 'A of the numbers
I(y, x') I for YEA and x' E W, is finite (III, p. 19, prop. 7). Since B is closed in E,
the theorem of bipolars (II, p. 44, tho 1) shows that we have A c 'A(B n M); this
completes the proof of (ii).
We shall now prove (iii). Let y' EM'. We shall prove the formula

(6) I y' I = inf I x' II .

t;(x') = y'

By prop. 8, (ii) oflV, p. 7, we have I ti II = II i II, and so I ti I ,;:; 1, and

(7) I y' II,;:; inf I x' I .
'i(x')= y'

By Hahn-Banach theorem (II, p. 23, cor. 3), there exists a linear form x~ on E which
extends y' and is of the same norm; hence we get the inequality opposite to (7),
since ti(x~) = y'.

Remark. - We know (II, p. 48, prop. 7, (ii)) that 1 is a topological vector space isomor-
phism from E~ I MO onto M~ (weak duals). For the topology of compact convex conver-
gence, prop. 10 shows that 1 is an isomorphism from E;c/Mo onto M;c when E is Haus-
dorff and M closed in E. For the strong topologies, 1 is a continuous mapping from
E~/Mo onto M~; it is an isomorphism if E is a Banach space * or if E is semi-reflexive
and M is closed in E (IV, p. 15) *' but this is not always so if E is a Frechet space (rv,
p. 58, exerc. 5, c)).

PROPOSITION 11. - (i) The weakened topology on ElM is the quotient of that on E;
the weakened topology on M is induced by that of E.
NO.5 DUALITY TVS IV.ll

(ii) The Mackey topology on ElM is the quotient of that on E; the Mackey topo-
logy on M is finer than the topology induced by ,(E, E').
Assertion (i) follows from prop. 7 of II, p.48.
The canonical injection i: M -+ E is continuous for the weakened topologies,
hence for the Mackey topologies ,(M, M') and ,(E, E') (IV, p. 7, prop. 7). Similarly,
the canonical projection p: E -+ ElM is continuous for the Mackey topologies. We
see immediately that the quotient topology on ElM obtained from ,(E, E') is com-
patible with the duality between ElM and (ElM)" hence is coarser than the Mackey
topology on ElM, by Mackey's theorem (IV, p. 2, tho 1). This proves (ii).

5. Dual of a direct sum and of a product

For every i E I, let (E i , F) be a pair of vector spaces, set in duality by a bilinear

form Bi . We put E = TI Ei and F = EB F i , and we identify each Fi with a subspace
iE] iEI
of F. We put E and F in duality by means of the bilinear form

(8) B(x, y) = L B;CXi' y) for x = (x) and y = (y)

iEI

(the family (Bi(X i , y))iEI has finite support).

We recall (II, p. 50, prop. 8) that the weak topology cr(E, F) is the product of the
weak topologies cr(E i , FJ

Lemma 3. - (i) For every i E I, let 6 i be afamily of subsets of F i , which is bounded

for cr(F i, E); put 6 = U 6 i . Then the 6-topology on E is the product of the 6 i-
iEI
topologies on the E i .
(ii) For every i E I, let ~i be an adapted bornology on the space Ei endowed with
the weak topology cr(Ei' F), none equal to {0}. Let :3 be the family of subsets A
of E = TI Ei such that pri(A) E:3 i for all i E I. Then the :3-topology on F is the direct
iEI
sum of the :3i-topologies on the F i .
Let .'!I be the product of the 6 i-topologies. The sets of the form

A = TI A~
iEl
x TI
iEI - J
Ei

where J c I is finite and Ai E 6 i for all i E J, form a fundamental system of neigh-

bourhoods of 0 for :Y. We have A = (U A;)O, hence :Y is identical with the 6-topo-
iEJ
logy. This proves (i).
We assign the :3-topology to F and the :3 i-topology to each F i . For every subset
A ofE, we have Fi n AO = pri(A)O, hence the injection from Fi into F is continuous.
Let q be a semi-norm on F; we assume that the restriction qi of q to Fi is continuous
for all i E I. Then we can find non-empty subsets Ai E:3 i such that we have

(9) qi(y) ~ sup IBi(xp y)1 (Yi E F).

XiEAi
TVS IV.l2 DUALITY IN TOPOLOGICAL VECTOR SPACES §1

Put A = n
i.eI
Ai; then A E 3· For y = (Y;)iEI in F, we have

q(y) :'( I CJly) :'( I sup IBi(x i, y)1 = sup IB(x, y)I,
iEI iEI XiEAi XEA

where the last equality follows from (8) since the family (Y)iEI has finite support
and the Ai are non-empty and can be assumed balanced (GT, IV, § 5, No.7, cor. 2
to prop. 12). This inequality proves that q is continuous Oil F, and hence (ii).

PROPOSITION 12. - The topology ~(F, E) is the direct sum of the topologies ~(Fi' EJ
The topology ~(E, F) is the product of the topologies ~(Ei' FJ
We shall apply lemma 3 taking for 6 i the family of all subsets of Fi which are
bounded for a{F i , E) and for :3 i the family of all subsets of Ei which are bounded
for a(E i , FJ
By cor. 2 of III, p. 4, :3 is the family of all subsets of E; which are bounded for the
product topology of the aCE;, F), which is identical with aCE, F). Hence our asser-
tion on ~(F, E) follows.
We endow F = ttl Fi with the topology :Y which is the direct sum of the topo-
iEI
logies a(F i , EJ Then the dual of F consists of the linear forms r f-> B(x, y) where
x runs through E (II, p. 30, prop. 6). By prop. 1 of IV, p. 1, the topologies :Y and
a(F, E) have the same bounded sets. Assume first that that the topologies a(F;, E)
are Hausdorff. By prop. 5 of III, p. 5, these sets are contained in a subset of the form
I Bi with J c I finite and B; bounded in F j (for a(F;, E)) for all i E J. Since I Bi
~ ~

is contained in the convex envelope of U nB;, where n = Card(J), we can apply

iEJ
lemma 3, to prove the assertion on ~(E, F) in this case.
For the general case, let Ni be the intersection of all neighbourhoods of 0 for
a(F i , E), and let N = I N i ; then FIN is the topological direct sum of the F;/Ni
iEI
(II, p. 31, prop. 8); we deduce from this that every subset of F which is bounded
for :Y is contained in a set of the form N + I Bi with J c I finite and Bi bounded
ieJ
in Fi for all i E J (III, p. 2, Remark 3); since the polar of this set in E is the same as
that of I Bi' the result follows as above.
iEJ

PROPOSITION 13. - The Mackey topology -reF, E) is the direct sum of the Mackey
topologies -r(F;, E) . The topology -r(E, F) is the product of the topologies -r(Ei' FJ
The assertion on -reF, E) follows from lemma 3 (ii) and the following property:
for a closed, convex and balanced subset of F* = n
F; to be compact for a(F*, F),
ieI
it is necessary and sufficient that its projection on each F; is compact for a(F;, FJ
To prove the assertion on -r(E, F), assume first that the topologies a(Fi' E) are
Hausdorff, it is enough (lemma 3 (i») to prove that every subset A of F which is convex,
balanced and compact for a(F, E) is contained in a set of the form I Ai where
iEJ
J c I is finite and where Ai is convex, balanced and compact for a(Fi' EJ But
No.5 DUALITY TVS IV.l3

such a subset is bounded for a(F, E). By the proof of prop. 12, there exists a finite
subset J of! such that A c I F i, and it is enough to take for Ai the projection of A
iEJ
on F i .
In the general case, with the same notations as in the proof of prop. 12, we have
T(E i, F) = T(E i, F;/N) and T(E, F) = T(E, FIN) (IV, p. 2) and since FIN is the
topological direct sum of the F;/Ni' we have reduced to the preceding case.
Q.E.D.
For the remainder of this paragraph, we assume that (E)iEI is a family of locally
convex spaces. Let S denote the topological direct sum of the Ei and P, their product.
We define a linear mapping S: S' --+ n
E;, said to be canonical, by
iEI

(10) S(x') = (x'IEJiEI (x' E S')

(where S' denotes the dual of S, and E; that of EJ

PROPOSITION 14. - (i) The mapping S is an isomorphism from the strong (resp. weak)
dual of S = EB Ei onto the product of the strong (resp. weak) duals of the Ei :
iEI
(ii) For a subset A of s' to be equicontinuous, it is necessary and sufficient that
the projection of SeA) onto E; be equicontinuous for all i E I.
(iii) The Mackey topology T(S, S') is the direct sum of the Mackey topologies
T(E i, E;).
(iv) The topology ~(S, S') is the direct sum of the topologies ~(Ei' E;).
That S is bijective follows immediately from the definition of a topological direct
sum (II, p. 30, prop. 6). Assertion (i) then follows from prop. 12 of IV, p. 12, for the
strong topologies, and from prop. 8 of II, p. 50, for the weak topologies. Similarly
(iii) follows from prop. 13 (IV, p. 12) and (iv) from prop. 12 (IV, p.12).
To prove (ii), let A be a subset of S'. Put

(11) q(x) = sup I<x, x') I for XES;

x'EA

let qi denote the restriction of q to E i, whence

(12) qlx) = sup l<xi,x;)1 for XiEE i ,

x;EAi

where Ai denotes the projection of SeA) on E;. For A to be equicontinuous, it is

necessary and sufficient that q is finite (that is, that each qi is finite) and continuous.
In view of the characterization of continuous semi-norms on a topological direct
sum (II, p. 27, prop. 5), this is the same as assuming that each qi is continuous, or
in fact, that each set Ai is equicontinuous. Q.E.D.
Let <I> be the linear mapping, said to be canonical, from EB E; into the dual P'
iEI
TVS IV.14 DUALITY IN TOPOLOGICAL VECTOR SPACES §2

of P = TI E i, defined by the formula

iEi

(13) <x, <!lex') >= L <Xi' x;>

iEI

for X = (x) in P and x' = ex:;) in EEl E;.

iEI

PROPOSITION 15. - (i) The map <!l is an isomorphism ji-om the topological direct
sum of the strong duals of the Ei onto the strong dual of P = Ei · TI
iEi
(ii) For a subset A of P' to be equicontinuous, it is necessary and sufficient that
it is contained in a finite sum L
<!leA), where J c I isfinite and where Ai is equicon-
iEJ
tinuous in E; for all i E J.
(iii) The Mackey topology rep, P') is the product of the topologies ,(E i , EJ
(iv) The topology ~(P. P') is the product of the topologies ~(Ei' EJ
It is immediate that <!l is injective. A fundamental system of neighbourhoods of 0
in P consists of sets of the fonn V = Vi X TI TI
Ep where J c I is finite and Vi
iEJ iEi-J
is a neighbourhood of 0 in Ei for i in 1. The polar of V in P' is equal to L <!leV?).
This proves the surjectivity of <!l and also assertion (ii).
Assertions (i) and (iv) follow from prop. 12 (IV, p. 12) and (iii) from prop. 13
(IV, p. 12).

COROLLARY. - Every product of barrelled spaces is barrelled.

A locally convex space E is barrelled if and only if the initial topology is identical
with ~(E, E') (TV, p. 4. Remark 3). Hence it is enough to apply prop. 15 (iv).

~ 2. BIDUAL. REFLEXIVE SPACES

1. Bidual

DEFINITION 1. - Let E be a locally convex space and E~ its strong dual. The dual
of the locally convex space E~ is called the bidual of E and is denoted by E".
< >
F or every x E E, let x be the linear form x' f---+ x, x' on E'; it is contin liOUS
for the weak topology cr(E', E), hence a fortiori, for the strong topology on E' ;
therefore x E E" for all x E E. The map cE : x f---+ X from E into E" is a linear mapping,
said to be canonical.

PROPOSITION 1. - The kernel of cE:E --+ E" is the closure of 0 in E. If E is Haus-

dorff, CE is injective.
By construction, the kernel of cE is the intersection orthe kernels orthe continuous
linear fonns on E, i.e. the closure of {O} in E (II, p. 24, cor. 1).
When E is Hausdorff, we identify E with a subspace of E", by means of the mapping
CEo
No.2 BIDUAL. REFLEXIVE SPACES TVS IV.15

The strong topology on E" is the G-topology, where G is the family of all strongly
bounded subsets of E'. Since every equicontinuous subset of E' is strongly bounded
(III, p. 22, prop. 9), the initial topology on E is coarser than the topology obtained
by taking the inverse image under cE of the strong topology on E"; it can be strictly
coarser (IV, p. 52, exerc. 1). However:

PROPOSITION 2. - Suppose that the space E is bornological or barrelled. The initial

topological on E is the inverse image under CE of the strong topology on E".
For, every subset of E' which is strongly bounded is equicontinuous (III, p. 22,
prop. 10 and III, p. 24).

PROPOSITION 3. - Let E be a locally convex Hausdorff space. In order that the strong
dual E~ of E be barrelled, it is necessary and sufficient that every subset of E" which
is bounded for cr(E", E'), is contained in the closure, for cr(E", E'), of a bounded subset
of E.
The equicontinuous subsets of E" are the subsets contained in the bipolar (for the
duality between E" and E') of a bounded subset of the subspace E of E". It is now
enough to apply the theorem of bipolars (II, p. 45, cor. 3) and the definition of a
barrelled space (III, p. 24).

Remark. - Let E be a locally convex Hausdorff space, E' its dual and E" its bidual.
We have E c E" c E'*, where E'* is the algebraic dual of E'. If B is a bounded
subset of E, its closure B in E'* endowed with cr(E'*, E') is containe~ in E" : for,
the polar U = BO of B in E' is a neighbourhood of 0 in E~, and we have B c U c E". O

2. Semi-reflexive spaces

DEFINITION 2. - Let E be a locally convex space. We say that E is semi-reflexive

if the canonical mapping CE from E into E" is bijective.
This implies that E is Hausdorff, and that every linear form on E', which is conti-
>
nuous for the strong topology ~(E', E), is of the form x' f-+ <x, x' with x E E, i.e.
continuous for the weak topology cr(E', E).

THEOREM 1. - A locally convex Hausdorff space is semi-reflexive if and only if every

bounded subset of E is relatively compact for the weakened topology cr(E, E'). If E
is semi-reflexive, the strong dual E~ of E is barrelled.
The second assertion follows from prop. 3 (IV, p. 15), and the identity between
bounded subsets for the initial topology and for the weakened topology of E (III,
p. 27, cor. 3).
To say that E is semi-reflexive means that the topology on E; is compatible with
the duality between E and E', in other words, by Mackey's theorem (IV, p. 2, tho 1)
that the topology on E~ is coarser than T(E', E) (and in fact is identical with it);
by definition (IV, p. 2), this means that every closed, convex and bounded subset of
E is compact for cr(E, E'), and this is equivalent to saying that every bounded subset
TVS IV.16 DUALITY IN TOPOLOGICAL VECTOR SPACES §2

of E is relatively compact for aCE, E'), because the closed convex envelope of a
bounded subset of E is bounded (III, p. 3, prop. 1).

COROLLARY. ~ Let E be a locally convex semi-reflexive space. Every closed vector

subspace M of E is semi-reflexive; moreover, the strong topology on E'/Mo (considered
as the dual of M) is the quotient of the strong topology on E'.
Let B be a bounded subset of M. Since B is bounded in E, and the weakened
topology a(M, M') on M is induced by aCE, E') (IV, p. 10, prop. 11), the closure
of Bin M endowed with a(M, M') is compact. Hence, by tho 1, M is semi-reflexive.
The last assertion of the corollary follows from prop. 10 ofIV, p. 9, applied to
the set G of all closed, convex and bounded subsets of E.
Remarks. ~ 1) Suppose E is semi-reflexive. Every subset of E which is convex,
closed and bounded for the initial topology is compact for the topology aCE, E')
(IV, p. 1, prop. 1). * On the other hand, the unit sphere (with the equation Ilxll = 1)
of an infinite dimensional hilbertian space E is closed and bounded for the initial
topology, but is not closed for the weakened topology, even if E is semi-reflexive. *
2) By remark 3 of IV, p. 5, we can reformulate tho 1 as follows : the Hausdorff
space E is semi-reflexive if and only if it is quasi-complete for its weakened topology.
If it is semi-reflexive, then it is quasi-completefor its initial topology (IV, p. 5, Remark 2).
3) Under the hypotheses of the above corollary, the space ElM is not necessarily
semi-reflexive (IV, p. 63, exerc. 10).

3. Reflexive spaces

DEFINITION 3. ~ A local(v convex space E is said to be reflexive !f the canonical

mapping cE Fom E into E" is a topological vector space isomorphism from E onto
the strong dual of E~.
In particular, a reflexive space is semi-reflexive, hence Hausdorff.

PROPOSITION 4. - The strong dual of a reflexive space is reflexive.

This follows immediately from def 3.

THEOREM 2. - In order that a locally convex Hausdorff space E be reflexive, it is

necessary and sufficient that it is barrelled and that every bounded subset ofE is rela-
tively compact for the weakened topology aCE, E').
By tho 1 (IV, p. 15), this is the same as saying that E is reflexive if and only if it
is semi-reflexive and barrelled
If E is reflexive, E~ is reflexive (prop. 4) and consequently E is barrelled (IV, p. 15,
tho 1). Conversely, if E is semi-reflexive and barrelled, cE is a bijection and is biconti-
nuous by IV, p. 15. prop. 2, hence E is reflexive.
Remarks. - * 1) Let E be an infinite dimensional real hilbertian space. Let F denote
the space E endowed with the weakened topology. The spaces E and F have the same
dual E', and E is a reflexive Banach space (V, p. 17). Consequently, F is semi-reflexive.
No.4 BIDUAL. REFLEXIVE SPACES TVS IV.17

However, the strong topology and the weakened topology on E are distinct, hence
F is not reflexive. *
2) Let E be a reflexive space and M a closed vector subspace of E. It may happen
that neither M nor ElM are reflexive spaces (IV, p. 63, exerc. 10). * For the case of
normed spaces, see prop. 7 of IV, p. 17. *

4. The case of normed spaces

Let E be a nonned space. The strong topology on the dual E' of E is defined by
the norm

(1) Ilx'll = sup I<x, x')I,

xEE.llxll ";1

and the strong dual of E is a Banach space (III, p. 24, cor. 2). Then the bidual E"
of E is also a Banach space, for the norm defined by

(2) Ilx"11 = sup I<x', x") I.

x'EE' .11 x' II ,,; 1

By prop. 8, (i) of IV, p. 7, the canonical linear mapping cE : E -+ E" is an isometry.

Henceforth, we shall identify E with a normed subspace of its bidual E".

PROPOSITION 5. - Let E be a normed space, E' its dual and E" its bidual. The (closed)
unit ball in E" is the closure of the unit ball B in E for the weak topology cr(E", E').
By formulas (1) and (2), the unit ball in E" is the bipolar BOO of B. Prop. 5 then
follows from the theorem of bipolars (II, p. 45, cor. 3).

Remark. - A Banach space E is closed in its bidual E" for the strong topology, but
is dense for the weak topology (prop. 5).

In order that a nonned space be reflexive, it is necessary and sufficient that it is

semi-reflexive; for, the initial topology of E is always induced by the strong topology
of E". Th. 1 (IV, p. 15) then implies the following result:

PROPOSITION 6. - In order that a normed space E be reflexive, it is necessary and

sufficient that the unit ball in E be compact for the weakened topology cr(E, E').
We observe that a reflexive nonned space is complete hence a Banach space,
and that its dual is a reflexive Banach space by prop. 4 of IV, p. 16.

PROPOSITION 7. - Let E be a reflexive Banach space and M a closed vector subspace

of E. Then M and ElM are reflexive Banach spaces.
Let E' be the dual of E and MO the orthogonal of M in E'. As a normed space,
we can identify the space E'/Mo with the dual M' of M (IV, p. 9, prop. 10). Since
M is semi-reflexive (IV, p. 16, corollary), it is reflexive, hence so is E'/Mo; similarly
MO is reflexive, as also its dual E/M = ElM. oO

Examples. - 1) Lete<Xl(N) denote the Banach space of bounded sequences x = (xn)nEN

TVS IV.18 DUALITY IN TOPOLOGICAL VECTOR SPACES §2

of scalars, with the norm

(3) Ilxll = sup IXnl (I, p. 4).

nEN

Let co(N) be the closed vector subspace of 1'X!(N) consisting of sequences tending
to o. Finally, let ('I (N) be the vector space of summable sequences, endowed with
the norm

(4) Ilxlll = L
l1EN
Ixnl·

We can show (IV, p. 47, exerc. 1) that the dual ofco(N) can be identified with el(N)
in such a way that we have

(5) (x, x') = L xnx~

llEN

for all x E co(N) and x' E ('1(N). Similarly the dual of ('I(N) can be identified with
e'D(N) in such a way that we have the relation (5) for all x E ('1(N) and all x' E ('XJ(N).
Hence f'X;(N) is the bidual of co(N), and this latter space is not reflexive.
* 2) Every hilbertian space is a reflexive Banach space (V, p. 17). *
* 3) Let X be a Hausdorff topological space and fl a complex measure on X.
For every real number p > 1, the Banach space LP(X, fl) is reflexive, and its dual
can be identified with U(X, fl) with p-I + q-I = I (INT, V, 2nd edition, § 5, No.8
and IX, § 1, No. 10). *

5. Montel spaces

DEFINITION 4. - A locally convex Hausdorff and barrelled space in which every

bounded subset is relatively compact is called a Montel space.
Examples. - 1) Every finite dimensional Hausdorff space is a Montel space. A normed
space which is a Montel space is locally compact, hence is finite dimensional (1, p. 15,
tho 3).
2) With the notations and hypothesis of prop. 7 of III, p. 6, the space E, being the
inductive limit of Banach spaces, is barrelled (III, p. 25); moreover, every bounded
subset of E is relatively compact (III, p. 6, prop. 7). In other words, E is a Montel space.
In particular, Gevrey spaces (Ill, p. 10) are Montel spaces. * This is true for the space
ff(K) consisting of germs of functions analytic in a neighbourhood of a compact
subset K of C" (III, p. 10).*
3) Every strict inductive limit E of a sequence (En) of Montel spaces (II, p. 33) such
that En is closed in En + 1 for all n, is a Montel space; in fact, E is Hausdorff (II, p. 32,
prop. 9 (i»), barrelled (III, p. 25, cor. 3) and every bounded subset of E is contained
in one of the En (III, p. 5, prop. 6) hence is relatively compact in En' and consequently
also in E.
* 4) Let U be an open set in Rn and let 'tI OJ(U) be the Frt::chet space of infinitely"
differentiable functions on U (III, p. 9). We shall prove that this is a Montel space.
Since 'tI W(U) is a Frechet space, it is barrelled (III, p. 25, corollary). Let B be a bounded
subset of 'tI ctJ(U) and let K be a compact subset of U. For every a E Nn let H. K be the
set of restrictions to K of the functions a'f, as f runs through B. Let a E Nn ; for every
~ENn such that I~I = lal + 1, the set H'.K is bounded in 'tI(K) since B is bounded
in 'tI OJ(U); by VAR, R., No.2. 2.3, the set H'.K is equicontinuous, hence (GT, X,§ 2, No.5)
NO.5 BlDUAL. REFLEXIVE SPACES TVS IV.19

relatively compact in C6'(K). But the topology of C6' ""(V) is the coarsest among the
topologies for which all the maps fl-> a~flK from (6' ""(V) into 'is'(K) arc continuous.
therefore B is relatively compact in C6' OO(V) (GT, 1. § 4. No.1. prop. 3 and § 9, No.5,
corolJ arv).
Simil,;r1y. the space 'too' (U) of all infinitely dij/crentiahle ./imctiol1s \\ith compact
support in U (III, p. 9) is a Montel space. For, C6'o" (V) is the strict inductive limit of
a sequence ((;~I~,(U) of Frechet spaces (III, p. 9), and it is enough to see that each of
the spaces ({;,(JV) is a Montel space (Example 3). But a bounded and closed subset
ofC6'H~JV) is closed and bounded in C6' OO(V), hence compact in <ti. '(V). and consequently
in C6'1~(V), *

PROPOSITION 8. - Let E be a Montel space and ~ a filter on E, which converges to

a point Xo in Efor the weakened topology. If~ is a countable base, or contains a bounded
set, then ~, converges to Xo for the initial topology also.
Assume first that there exists a bounded sety in ~. The closure B of B for the
initial topology of E is b~unded; in addition, B is compact because E is a Montel
space. The topology on B induced by cr(E, E') is Hausdorff and coarser than the
topology induced by the initial topology; they therefore coincide (GT. I, § 9, No.4).
This prove the proposition for this case.
Next assume that ~ has a countable base. It is enough (GT. L § 6. No. 8, prop. 11) to
consider the case of a sequence (Xn)n;d tending to Xo for cr(E, E'). Let B be the set
of all Xn for n ;;:: O. This set is bounded for cr(E, E'), hence also for the initial
topology (III, p. 27, cor. 3). Thus we have reduced to the first case of the proof.
Every Montel space is reflexive : this follows from def. 4 and from tho 2 of IV,
p. 16. Further :

PROPOSITION 9. - The strong dual of a Montel space is {/ Montel 5pace.

Let E be a Montel space and E~ its strong dual. Since E is reflexive. E~ is barrelled
(IV, p. 15, tho 1). Since every bounded subset of E is relatively compact the strong
topology on E' coincides with the topology of compact convergence. Let B be a
bounded subset of E~; it is bounded for the weak topology cr(E', E), hence is equi-
continuous because E is barrelled. Then Ascoli's theorem (GT, X, § 2, No.4, cor.
and § 2, No.5, cor. 1) implies that the closure ofB for cr(E', E) is compact for the topo-
logy of compact convergence; therefore B is relatively compact in E~.

PROPOSITION 10. - Every metrizable Montel space satisfies the first axiom of coun-
tability.
Let E be a metrizable Montel space. We know (II, p. 5) that E can be identified
with a subspace of the product F = TI
Fn of a sequence of normed spaces, and
nEN
we can even assume that we have pr/E) = Fn for all n EN. If each of the metrizable
spaces Fn satisfies the first axiom of countability, then so does F (GT, IX, § 2, No.8),
hence also E.
We argue by reductio ad absurdum. Assume for example that Fo does not satisfy
the first axiom of countability. Let Bo be the unit ball (closed) in F 0; this is a metric
TVS IV.20 DUALITY IN TOPOLOGICAL VECTOR SPACES §2

space which does not satisfy the first axiom of countability. We shall use the following
lemma:

°
Lemma 1. - Suppose the metric space X does not satisfy the first axiom of counta-
bility. Then there exists a real number E > and an uncountable subset A in X such
that d(x, y) ~ E for all distinct x, y in A.
For every integer n ~ 1, let ~n be the set (ordered by inclusion) of subsets D
of X such that d(x, .r) ~ ln for distinct x, y in D. The set ~" is of finite character,
hence possesses a maximal element Dn (S, III, § 4, No.5). Then for all y E X, there exists
a point x in Dn such that d(x, y) < l,
n
by virtue of the maximal character of Dn'
Put 0 = U D,,; the set D is then dense in X, and since X does not satisfy the first
"
axiom of countability, D is not countable, and so one of the Dn is not countable.
Q.E.D.

°
By lemma 1, applied to Bo' there exists an uncountable subset Ao of F 0 and a
number E > such that Ilxll :S 1 and Ilx - yll ~ E for distinct x, y in Ao. We have
pro(E) = F 0 and hence there exists a subset A in E such that pro induces a bijection
from A onto Ao.

Lemma 2. - There exists a sequence (xm)m;.O consisting of distinct elements of A,

which is bounded in E.
We shall construct a sequence (xm )",;. 0 of points of A by induction; and a decreasing
sequence (Cm)m;' 0 of subsets of A satisfying the following conditions:
a) None of the sets Cm is countable.
b) For every n ~ 0, the set prk(C m ) is bounded in Fk for 0 :S k :S m.

°
c) For every I1J ~ 0, we have xm E C m - C m+ l '

°
We put Co = A. Suppose the sets C m for :S m :S n have been defined, so as
to satisfy a) and b) for 0 :S In :S n, and also the points xm in C m - C m + 1 for :S m < n.
For every integer r ~ 1, let Cn,r be the set of all x E Cn such that

Since Cn is not countable, there exists an integer r ~ I such that Cn,r is not countable.
We choose a point Xn in C",r and put Cn+ 1 = Cn,r - {x n }· Evidently Cn + 1 C Cn
and xn E C n - C,,+1' the set Cn + 1 is not countable and prk(Cn +,) is bounded in Fn
for 0 :S k :S n + 1.
We have xm E Cm' and so xm E C n where m ~ n. The projection of the sequence
(xm)m;' a on Fn is therefore bounded for all n ~ 0; in other words, the sequence
(xm)m", 0 is bounded in E, and this establishes lemma 2.
Q.E.D.
With the notations of lemma 2, the bounded sequence (xm)m", a has a limit point y
in E. Therefore the sequence (pra(xm))m;.ohas a limit point pro(y) in Fa, but this
contradicts the construction of Aa.
No.1 DUAL OF A FRECHET SPACE TVS IV.21

COROLLARY. - Let E be a metrizable Montel space. Then there exists a countable

dense set in the strong dual of E.
On the dual E' of E, the strong topology is identical with that of compact conver-
gence, since E is a Montel space. It is now enough to apply cor. 1 of prop. 6 of III,
p. 18.

Z We can show that the strong dual of a metrizable Mantel space E is not metrizable
if E is infinite dimensional (IV, p. 57, exerc. I).

§ 3 DUAL OF A FRECHET SPACE

1. Semi-barrelled spaces

PROPOSITION 1. - Let E be a locally convex space. The following conditions are

equivalent :
(i) Let U be a subset of E which absorbs every bounded subset of E, and which is
the intersection ofa sequence of convex, balanced and closed neighbourhoods of 0 in E.
Then U is a neighbourhood of 0 in E.
(ii) For every locally convex space F, every bounded subset of 'pb(E ; F) which is the
union of a countable family of equicontinuous subsets, is equicontinuous.
(iii) In the strong dual E~ of E, every bounded subse t which is the union of a coun-
table family of equicontinuous subsets, is equicontinuous.
It is clear that (iii) is a particular case of (ii).
(i) => (ii) : let H be a bounded subset of ,Pb(E; F), and let (Hn) be a sequence of
equicontinuous subsets of ,P b(E; F) such that H = U Hn' Let V be a convex, balanced
and closed neighbourhood of 0 in F. For every n, the set Wn = n u-l(V) is a
uEHn
convex, balanced and closed neighbourhood of 0 in E since Hn is equicontinuous.
The set W = n
u- leV) absorbs every bounded subset of E, since H is bounded
UEH

in ,P b(E; F) (III, p. 22), and we have W = n Wn.

n
If E satisfies (i), then the set W
is a neighbourhood of 0 in E, hence H is equicontinuous.
(iii) => (i) : let (Un) be a sequence of convex, balanced and closed neighbourhoods
of 0 in E. We assume that the set U = n Un absorbs every bounded subset of E,
n
hence that its polar UO is bounded in E~. Then the set B = U U~ is contained in UO,
n
hence is bounded in E~. If E satisfies (iii), the set B is equicontinuous in E'; conse-
quently, the polar BO = n (U~)O = n Un = U of B in E is a neighbourhood of 0
n n
in E.
DEFINITION 1. - A locally convex space E is said to be semi-barrelled if it satisfies
the equivalent conditions of prop. 1.
Every barrelled space is semi-barrelled. This is also true for every bomological
space (III, p. 22, prop. 10).
TVS IV.22 DUALITY IN TOPOLOGICAL VECTOR SPACES §3

2. Dual of a locally convex metrizable space

PROPOSITION 2. - Let E be a locally convex metrizable space and F its strong dual.
The ~pace F is complete, semi-barrelled and satisfies the following condition :
(DB) There exists a sequence (An)nEN ofbounded subsets ofF such that every bounded
subset of F is contained in one of the An'
The space E is bomological (III, p. 12, prop. 2), hence its strong dual is complete
(III, p. 23, cor. 1).
Let (Vn)"EN be a decreasing sequence of neighbourhoods of 0 in E, such that every
neighbourhood of 0 in E contains one of the V n. Let An be the polar of Vn in F.
Since E is bomological, every bounded subset of F is equicontinuous (III, p. 22,
prop. 10), therefore contained in one of the An' In other words, the space F satisfies
the condition (DB).
We now show that F is semi-barrelled. Let (U,,)nEN be a sequence of convex, balanced
and closed neighbourhoods of 0 in F. We assume that the set U = n Un absorbs
n
every bounded subset of F. We shall prove that U is a neighbourhood of 0 in F.
For this, we shall construct, by induction on the integer n ;::: 0, real numbers An > 0
and convex balanced neighbourhoods W" of 0 in F, whose which are closed for a(F, E),
and satisfy the following relations

(1) AnA" c tu n (n W;)

O:::;i <n

(2) U AAi C Wn C Un .
O~i~n

Suppose that the numbers Ai and the sets Wi have been constructed for 0 :( i < n.
By the hypothesis, the set U absorbs the bounded subsets of F; moreover, for
o :( i < n, Wi is a neighbourhood of 0 in F, hence absorbs the bounded subsets
of F. We can therefore find a number An > 0 satisfying (1). Let C denote the closed
convex balanced envelope, for a(F, E), of U AiAi; the set C is equicontinuous,
O~i~n

hence compact for a(F, E) (Ill, p. 17, cor. 2). Since Un is a neighbourhood of 0 in F,
there exists a bounded subset B ofE such that Be C tUn' Put Wn = C + BO. Since
BO is a neighbourhood of 0 in F, we see that Wn is a convex and balanced neighbour-
hood of 0 in F. In addition, C is compact and B closed for a(F, E); by cor. 1 of
C

GT, III, § 4, No.1, Wn is closed for a(F, E). Finally, we have C c tu c tUn and
BO c tUn' hence Wn C Un since Un is convex. We have thus established (2).
Put W = n W n , then W c U. By (1) and (2), we have AAi c Wj for all i and j
n

in N, and so Ai Ai c W for all i E N. In particular, W is absorbent, hence is a barrel

for a(F, E). By remark 3 ofIV, p. 4, W is a neighbourhood of 0 in F. Afortiori, U is
a neighbourhood of 0 in F, and F is semi-barrelled.
The following corollary extends the Banach-Steinhaus theorem to the dual of
a Frechet space (cl III, p. 25, cor. 2).
No. 4 DUAL OF A FRECHET SPACE TVS IV.23

COROLLARY. - Let G be a Hausdorff locally convex space, and let (un) be a sequence
oflinear mappingsfi'om F into G, converging simply to a mapping ufrom F into G.
Then u is continuous, and the sequence (u,.) converges to u uniformly on every pre-
compact subset of F.
Since F is complete, the set of all un' which is bounded for the topology of simple
convergence, is bounded in £\(F; G) (III, p. 27, cor. 1). Since the space F is semi-
barrelled (prop. 2), every countable and bounded subset of Sf b(F; G) is equiconti-
nuous by prop. I of IV, p. 21. Therefore the set of the un is equicontinuous, and the
corollary follows from III, p. 18, corollary.

3. Bidual of a locally convex metrizable space

PROPOSITION 3. - Let E be a locally convex metrizable space, E~ its strong dual

and G a Frechet space. The space Sfb(E~; G) is a Frechet space.
By prop. 2 (IV, p. 22), there exists a sequence (A,) of bounded subsets of E~ such
that every bounded subset ofE~ is contained in one of the An' Let (Vn) be a countable
fundamental system of neighbourhoods of 0 in G. Let Hmn be the set of linear map-
pings u from E~ into G such that u(Am) c Vn' Then (HmJ is a fundamental system
of neighbourhoods of 0 in Sf b(E~ ; G), and the latter space is then metrizable.
To show that Sfb(E~; G) is complete, it is enough to prove that every Cauchy
sequence (un) in this space is convergent; since G is complete, there exists a linear
mapping u: E~ -> G such that (un) converges simply to u. By IV, p. 23, corollary,
we have u E SfJE~; G). It then follows from prop. 5 ofGT, X, § 1, No.5, that (un) con-
verges to u in Sf b(E~ ; G).

COROLLARY. - The bidual of a locally convex metrizable space is a Frechet space.

4. Dual of a reflexive Frechet space

PROPOSITION 4. - Let E be a reflexive Frechet space. The strong dual E~ ofE is the
inductive limit of a sequence of Banach spaces.
Let (Vn)nEN be a decreasing sequence of convex, balanced and closed neighbour-
hoods of 0 in E, such that every neighbourhood of 0 in E contains one of the Vn •
Let An be the polar ofVn in E'. Then An is convex, balanced and compact for cr(E', E);
by III, p. 8, corollary the space EAn is a Banach space. We shall prove that E~ is the
inductive limit of the spaces E A"; in other words, that every convex and balanced
subset U ofE' which absorbs each of the An is a neighbourhood of 0 in E~. For every
n EN, choose a real number An > 0 such that AnAn c U. Let Bn be the convex
envelope of the set U A;A;; put V = U Bn , then V c U. For every n E N, the
O::s'i~n n
set Bn is convex, balanced and compact for cr(E', E) (II, p. 14, prop. 15).
Now we shall show that tv oo c V. Let x E E~ - V; for every n E N, we have

x r/= B n, and since Bn is closed for cr(E', E) there exists an element Yn in B~ such that
TYS IY.24 DUALITY IN TOPOLOGICAL VECTOR SPACES §3

<Yn' x) = 1 (II, p. 38, prop. 4). Since E is reflexive, every bounded subset of E is
relatively compact for aCE, E') (IY, p. 16, tho 2). By the definition of Bn , we have

(3)

hence the sequence (Yn) is bounded. Let y be a limit point of (Yn) for the topology
aCE, E'). We have Y E yo = n B~ and
n
<y,
x) = 1. Hence x rt tYoo, and so we
have the inclusion tY oo c Y and a fortiori, ty oo c U.
Since every bounded subset ofE~ is contained in one of the sets All' the set Y =U Bn
n

absorbs every bounded subset of E~. Consequently, yo is bounded in E, hence

ty oo is a neighbourhood of 0 in E~. Afortiori, U is a neighbourhood of in E~.

COROLLARY. - The strong dual of a reflexive Frechet space is bornological and

barrelled.
An inductive limit of Banach spaces is bomological by definition. Further, a
Banach space is barrelled (III, p. 25, corollary) and every inductive limit of barrelled
spaces is a barrelled space (III, p. 25, cor. 3).

5. The topology of compact convergence on the dual of a Frechet space

THEOREM 1 (Banach-Dieudonne). - Let E be a locally convex metrizable space.

The following topologies coincide on the dual E' of E :
a) the topology 3W of lJl-convergence, where IJl is the family of subsets of E each
consisting of points of a sequence converging to 0;
b) the topology :?Ie of uniform convergence on compact subsets of E;
c) the topology :?Ipe of uniform convergence on precompact subsets of E ;
d) the topology :?If which is the finest topology inducing the same topology as
aCE', E) on every equicontinuous subset of E'.
First observe that a subset A ofE' is closedfor :?If if and only if An H is closed
for aCE', E) for every subset H of E' which is equicontinuous and closedfor aCE', E).
The weak topology aCE', E) and :?Ipe induce the same topology on every equiconti-
nuous subset of E' (III, p. 17, prop. 5). Consequently each of the topologies :?I91 ,
:?Ie' :?Ipe ' :?If is coarser than the one following it. It is therefore enough to prove that
:?I91 is finer than .Ojf. Moreover, every translation in E' is a homeomorphism for :?If.
Hence it is enough to prove that, if F is a subset of E' which is closed for :?If' and
does not contain 0, then there exists a set S E IJl such that So n F = 0.
Let (Un)n?O be a decreasing sequence of neighbourhoods of 0 in E forming a
fundamental system of neighbourhoods of o. We shall construct, by induction on
n ~ 0, finite sets Xn such that we have

(4) Xn C Un
(5) ( U Xp)O n U~+ 1 n F = 0
O~p~n
No. S DUAL OF A FRECHET SPACE TVS IV.2S

for every integer n ;::: O. Let m ;::: 0 be an integer such that Xn has been constructed
for 0 ~ n < m and satisfies (4) and (S) for 0 ~ n < m. For every x E Urn' put

Fx={ U Xp)On{x}OnU~+lnF.
O~p<m

Formula (S) with n = m - 1 implies that n Fx = 0. Further, the set U~+ 1

XEU m

is equicontinuous, and compact for a(E', E). In view of the definition of flf' each
of the sets F x is compact for a(E', E); therefore there exists a finite subset Xm of
Urn such that n Fx = 0, i.e. relation (S) is satisfied for n = m.
XEX m

Put S = U X n. We have Xn C Up for n ;::: p, therefore S is the set of points of

n~ 0
a sequence which converges toOinE. From(S) we deduce that So n U~+l n F = 0,
and since E' is the union of the sequence of sets U~ + l ' we get So n F = 0.

COROLLARY 1. - Let E be a locally convex metrizable space. Every pre compact subset
ofE is contained in the closed convex balanced envelope of the set ofpoints of a sequence
conve.'ging to O.
This follows from the fact that the topologies flpe and ~ are identical, on account
of prop. 2 of III, p. IS.

COROLLARY 2. - Let E be a Fn?chet space. In order that a convex subset A of the

dual E' ofE be closedfor a(E', E), it is necessary and sufficient that A n U is closed O

for cr(E', E) for every neighbourhood U of 0 in E.

Since E is complete, the topology fie on E' is compatible with the duality between
E' and E (IV, p. 3, Example); consequently the closed convex subsets in E' are the
same for fie and cr(E', E) (IV, p. 1, prop. 1). The corollary then follows from the
identity of the topologies Yc and flf'

Recall (I, p. 13) that the hyperplanes of E' which are closed for cr(E', E) are the kernels
of linear forms on E' associated with elements of E. Cor. 2 therefore gives another proof
(for Frechet spaces) of cor. I of III, p. 21.

COROLLARY 3. - Let E be a Banach space and M a vector subspace of the dual E'
of E. In order that M be closed for the weak topology cr(E', E), it is necessary and
sufficient that its intersection with the unit ball (closed) in E' be closed for cr(E', E).

Example. - * Let H be a hilbertian space satisfying the first axiom of countability;

let Hcr denote the space H with the weakened topology assigned to it. Let if l(H) be
the Banach space of nuclear endomorphisms of H (V, p. 51, and TS, V); the norm
in if !(H) is defined by Ilull! = Tr((u*u)!/2). We can identify if (H) with the dual of
the Banach space if !(H) by associating the linear form <Pu : v 1-+ Tr(uv) on if !(H)
with every u E if (H). Let A be a sub-algebra of if (H), containing 1 and stable under
u 1-+ u* ; this is a von Neumann algebra if and only if it is closed in if (H) for the weak
topology cr(if(H), if l(H». From cor. 3, we deduce the following criterion : for A
to be a von Neumann algebra, it is necessary and sufficient that if (u ll ) is any sequence of
elements of A with norm ~ I having a limit u in the space ifsCH ; Hcr), then u belongs
to A. *
TVS IV.26 DUALITY IN TOPOLOGICAL VECTOR SPACES §4

6. Separately continuous bilinear mappings

Lemma 1. - Let E and F be two locally convex metrizable .ljJaces, and u be a conti-
nuous linear mapping from E~ il/"o F. Then there exiS1S a neighbourhood U orO in E~
whose image under u is hounded in F.
Let (U')lIEN (resp. (V,)lIEN) be a fundamental system of neighbourhoods of 0 in E
(resp. F). We assume that the sets Un are balanced and form a decreasing sequence.
Since u is continuous, for every 11 E N. there exists a bounded set Bn in E such that
u(B~) c V". Since Bn is bounded, there exists a real number An > 0 such that
A"B" c U". Put B = U AnBn'
nEN
We shall prove that the set B is bounded in E, in other words that for every integer
m ~ 0, there exists a real number fl > 0 such that flB c Urn' Since the sets B"
are bounded, there exists a real number ~l such that 0 < 11 :::;; 1 and such that
1l.(An B,,) c Urn for 0 :::;; 11 :::;; In; we have also A"B" c U" c U m if n > In; hence
IlB c Urn since Urn is balanced.
Let U be the polar of B in E~. This is a neighbourhood of 0 in E;, and we have
AnBo c B~. hence I'nl/(U) c V" for all n E N. Consequently u(U) is bounded in F.

THEOREM 2. - Lei E1 and E2 be two reflexive Frechel spaces, and G a locally convex
Hausdorff space. For i = 1,2, let Fi be the strong dual or E i . Then every separately
continuous bilinear mapping u: F] x F 2 -+ G is continuous.
The space G is isomorphic to a subspace of a product of Banach spaces (II, p. 5.
prop. 3). Therefore it is enough to prove the theorem under the additional hypo.thesis
that G is a Banach space. But F 1 is barrelled and F2 bomological (IV, p. 24, corollary),
and Sf' b(F 2 ; G) is a Fn!chet space (IV, p. 23, prop. 3). Let v denote the linear mapping
from F 1 into 2 J b(F 2' G) associated with u by the relation

u(x 1 • x 2 ) = v(xJ(x z) (Xl E F1, x2 E F z ).

Since F 1 is barrelled and u separately continuous, v is continuous (III, p. 31, prop. 6).
Since v is continuous, lemma I implies the existence of a neighbourhood U 1
of 0 in F 1 whose image under v is bounded in Sf' bCF 2; G). In other words, for every
bounded subset B2 in F 2' the set u(U 1 x B 2) is bounded in the Banach space G.
Let U 2 be the set of all x 2 E F 2 such that Ilu(x 1 , x2 )11 : :; I for all Xl E U l' The set
U 2 then absorbs every bounded subset; since F 2 is bomological, U 2 is a neigh-
bourhood of 0 in F 2 , and this proves that u is continuous.

§ 4. STRICT MORPHISMS OF FRECHET SPACES

For every locally convex space E, let SeE) denote the set of all continuous semi-
norms on E. For every p E SeE), let Hp denote the set of all linear forms f on E
such that If I :::;; p. The family (Hp)PES(E) is a base for the bomology consisting of
equicontinuous subsets of E'.
No.1 STRICT MORPHISMS OF FRECHET SPACES TVS IV.27

1. Characterizations of strict morphisms

PROPOSITION 1. - Let E and F be two locally convex spaces and u a continuous

linear mapping from E into F. In order that u be a strict morphism, it is necessary
and sufficient that the following condition be satisfied:
(MS) For every semi-norm p E SeE), which is null on the kernel of u, there exists q
in S(F) such that p :::; q 0 u.
Let N be the kernel and M the image of u; we introduce the canonical decomposi-
tion of u, let

E ~ E/N .fl.. M -4 F .

The continuous semi-norms on E which are null on N, are the semi-norms Pl 0 TC

where Pl ranges over S(E/N); similarly SCM) consists of the semi-norms ql for
which there exists q E S(F) with ql :::; q/F. Finally, u is a strict morphism if and only
if the bijective continuous linear mapping fl has a continuous inverse; this also
means that every semi-norm in S(E/N) is of the form ql 0 fl with ql in SCM). Prop. 1
follows immediately from these remarks.

PROPOSITION 2. - Let E and F be two Hausdorff locally convex spaces and u a conti-
nuous linear mapping/rom E into F. In order that u be a strict morphism, it is necessary
and sufficient that its transpose tu: F -+ EI satisfy the following conditions :
a) The image of tu is closed in E' for a(EI, E).
b) Every equicontinuous subset ofE contained in the image oftu is the image under
I
,

tu of an equicontinuous subset of F.
If this is so, we have Ker tu = (1m u)O and 1m tu = (Ker u)O and there exist cano-
nical isomorphisms from Coker 'u onto the dual ()f Ker u and from Ker tu onto the dual
of Coker u.
Let N be the kernel and I the image of u. By cor. 2 of II, p. 47, the kernel of tu
is the orthogonal ofl, and the closure of the image oftu for a(EI, E) is the orthogonal
N° of N. The conjunction of a) and b) is then equivalent to the following condition:
b Every equicontinuous subset of EI contained in N is the image under tu of an
l
)
C

equicontinuous subset of F.
Since N° can be identified with the dual of E/N, prop. 9, (i) of IV, p. 8, shows that
the equicontinuous subsets of EI contained in N° are the sets which are contained
in a set of the form H p ' where p is a continuous semi-norm on E, vanishing on N.
The condition b') then says that, for every semi-norm p E SeE) which is null on N,
there exists q E S(F) such that Hp c 'u(H q). By Hahn-Banach theorem (II, p. 23,
cor. 1 and 2, p. 63, tho 1 and cor. 1), we havetu(Hq) = Hqou' and the relations Hp c H p'
and p :::; pi are equivalent for all semi-norms p and pi in SeE). Consequently, the
relation Hp c 'u(H q) is equivalent to the relation p :::; q 0 u. By prop. 1, property
b implies that u is a strict morphism.
l
)

Suppose that u is a strict morphism. We have already seen that the kernel of
'u is the orthogonal of I and the image of tu is the orthogonal of N. The cokernel
TVS IV.28 DUALITY IN TOPOLOGICAL VECTOR SPACES §4

of u is the space FjI and its dual can be identified with 1° = Ker tu. Similarly, the
dual of N = Kcr u can be identified with E'jN° (IV, p. 8), i.e. with the cokemel
of tu since N is the image of I II.
Remark.- With the notations of prop. 2, properly b') also implies that u is a strict
morphism for the weakened topologies (II, p. 49, cor. 3).

PROPOSITION 3. - Let E and F be two locally convex .lj)(7("es. and II a continuous linear
mappingf;'om E into F. We assume that E is Hausdorff and that F is metrizable. For
u to be a strict morphism, it is necessary and sufficient that the image of 'u be closed
in E' for the weak topology cr(E', E).
The necessity follows from prop. 2.
Conversely, suppose that the image of 'u is closed for cr(E', E) and introduce the
canonical decomposition of u as in the proof of prop. 1. By the above remarks, the
inverse mapping 17 - 1 of fi is continuous for the weakened topologies. But the subspace
M = u(E) of F is metrizable. hence bomological (lIf, p. 12. prop. 2); consequently
(IV, p. 7, prop. 7. (ii)), 17- 1 is continuous, hence u is a strict morphism.

2. Strict morphisms of Frechct spaces

THEOREM 1. - Let E and F be two Frechet ,Ipaces and u a continuous linear mapping
from E into F. The following conditions are equivalent:
a) u is a strict morphism.
b) u is a strict morphism for the weakened topologies.
c) The image of u is closed in F.
d) tu is a strict morphism from F' into E' for the l\"eak topologies.
e) The image of ' II is closed in E' for the weak topology cr(E', E).
f) The image of' u is closed in E' for the strong topology ~(E', E).
g) 'u is ([ strict morphism ji-om F~ into E~ (the duals endowed with the topology
of compact convergence).
The equivalence of a), h) and e) follows from prop. 3 of IV, p. 28 and the remark
preceding it. That of a) and c) is precisely cor. 3 of I, p. 19. The remark of IV, p. 28,
also shows that d) is equivalent to the fact that, the image of u is closed for the weak-
ened topology cr(F, F') of F; the equivalence of c) and d) then follows from prop. 2
of IV, p. 4.
We now prove the equivalence of e) and f). It is enough to prove that f) implies
e). Suppose that the image oPu is closed for ~(E', E) in E'. On account ofthe Banach-
Dieudonne theorem (IV, p. 25, cor. 2), it is enough to prove that for every convex
balanced neigh bourhood U of 0 in E, the intersection B = 'u(F') n U is compact
for cr(E', E). The strong dual E~ of the Frechet space E is complete (IV, p. 22, prop. 2),
hence the closed subset B of E~ is complete, and so the normed space E~ is complete
(III, p. 8, corollary). Let (V n ) be a decreasing sequence forming a fundamental sys-
tem of neighbourhoods of 0 in F. Then F' is the union of sets C = V~ which are
compact for cr(F', F), hence E~ = U B n , with BI1 = E~ n tu(Cn ). Since E~ is a Baire
n
No.2 STRICT MORPHISMS OF FRECHET SPACES TVS IV.29

space, and each of the sets Bn is convex balanced and closed, there exists a real
number r > 0 and an integer n such that B cr. Bn' Then we have B = UO n tu(r. C n) ;
since the sets UO and r,C n are compact and tu is continuous for the weak topolo-
gies, B is compact for cr(E', E). This completes the proof of the equivalence of e)
and j).
Finally the equivalence of g) and the preceding conditions follows from prop. 18
of GT, X, § 2, No. 10 and the following lemma.

Lemma 1. - Let E and F be two Hausdorff locally convex and quasi-complete spaces
and u a continuous linear mapping from E into F. For tu to be a strict morphism Fom
F~ into E~, it is necessary and sufficient that the image u(E) of u be closed, and that
every compact subset of u(E) be the image under u of a compact subset of E.
By Mackey's tho (IV, p. 2, tho 1) and the fact that on E' (resp. F) the topology of
compact convergence coincides with that of convex compact convergence (IV,
p. 4), we can identity E (resp. F) with the dual E~ (resp. F~). Then u is the transpose
of t u, and the equicontinuous subsets of E (resp. F) are the relatively compact sets.
Lemma I then follows from prop. 2 (IV, p. 27), since u(E) is closed in F if and only
if it is closed for the weakened topology cr(F, F') (IV, p. 4, prop. 2).

COROLLARY 1. - Under the hypothesis of th. 1, the following conditions are equiva-
lent:
(i) u is a strict injective morphism;
(ii) tu is a strict surjective morphism for the weak topologies.
(iii) tu is sU/jective.
The implication (i) => (ii) follows immediately from the equivalence of conditions
a), d) and e) of tho 1 and from IV, p. 6, prop. 5. It is clear that (ii) implies (iii). Finally,
we prove that (iii) implies (i) : if tu is surjective u is a strict morphism by the equi-
valence of a) and e) in tho 1 ; that u is injective follows from prop. 5 of IV, p. 6.

COROLLARY 2. - Under the hypothesis of th. I, the following conditions ute equiva-
lent:
(i) u is surjective;
(ii) u is a strict surjective morphism;
(iii) tu is a strict injective morphism for the weak topologies.
The equivalence of (i) and (ii) follows from Banach's tho (I, p. 17, tho 1).
In view of the equivalence of a) and c) in tho 1, condition (ii) says that u is a strict
morphism and that its image is dense in F for cr(F, F'). The equivalence of (ii) and
(iii) then follows from the equivalence of a) and d) in tho 1 and from prop. 5 of IV,
p.6.

If u : E ---> F is a strict morphism of Frechet spaces, the transpose 'u is not necessarily
a strict morphism from F~ into E~ (IV, p. 62, exerc. 3). However, we have the following
partial result:
TVS IV.30 DUALITY IN TOPOLOGICAL VECTOR SPACES §4

COROLLARY 3. - Under the hypotheses of tho 1, the following property implies the
properties 0) /0 g) :
h) tu is a strict morphismfrom F~ into E~.
When E and F are both Banach spaces, or both Montel spaces. property h) is
equivalent to the properties a) to g) of tho l.
Suppose that'll is a strict morphism from F~ into E;. We shall prove that the
image H of tu is closed in E~. from which the first assertion of cor. 3 will follow.
Let G be the closure of the image of u in F; the space G, with the topology induced
by that of F assigned to it, is a Frechet space. The mapping u: E --* F factorizes as
u = j 0 v where j is the canonical injection from G into F and where v E 2'(E; G).
Then we have tu = tv 0 tj, where tj is surjective, by Hahn-Banach tho (II, p. 24, prop. 2);
also, tv is injective since veE) is dense in G (IV, p. 6, prop. 5). By hypothesis, the
mapping tu from F~ onto H is open; since:j is surjective and continuous, the mapping
tv induces a homeomorphism from G~ onto H. But the dual G~ of the Frechet space
G is complete (IV, p. 22, prop. 2); consequently H is complete, hence closed in E~.
If E and Fare Montel spaces, the strong topology on E' (resp. F') coincides with
the topology of compact convergence, and h) is just a reformulation of g).
If E and F are Banach spaces, so are E~ and F~, and condition h) is equivalent
to f) by the equivalence of 0) and c) applied to tu: F~ --* E~.

COROLLARY 4. - Suppose E and F are Banach spaces. For tu to be sUijective, it

r.l,
is necessary and sufficient that there exist a real number r > 0 such that II x II :( I
u(x)
for at! x E E.
This is simply a reformulation of the equivalence of the conditions (i) and (iii)
of cor. l.

COROLLARY S. - Let E and F he two Frechet !>paces and u a continuous linear mapping
from E into F. The jiJllowing conditions are equivalent:
a) u is an isomorphism /rom E onto F.
b) u is an isomorphism ji-om E onto F for the weakened topologies.
c) 'u is an isomorphism from F' onto E' for the weak topologies.
d) tu is an isomorphism from F' onto E' for the strong topologies.
e) 'u is an isomorphism from F; onto E~.
Since an isomorphism is none other than a strict bijective morphism, the equi-
valence of a) and b) follows from the equivalence of conditions a) and b) of tho l.
It is clear that a) implies each of the conditions c), d) and e).
Conversely, suppose that one of the conditions c), d) or c) is satisfied. It follows
from tho I and its cor. 3 that II is a strict morphism from E into F. and tu is evidently
bijective. Let N (resp. l) be the kernel (resp. the image) of u. Since tu is bijective,
we have Imtu = E' and Kertu = {O}, and so N° = E' and r = {O} by prop. 2
ofIV, p. 27. But N (resp. l) is a closed vector subspace ofE (resp. F), and the theorem
of bipolars (11, p. 44) implies that N = {O} and I = F, hence u is bijective. We have
therefore proved that u is an isomorphism.
No.3 STRICT MORPHISMS OF FRECHET SPACES TVS IV.31

3. Criteria for surjectivity

PROPOSITION 4. - Let E and F be two Frechet spaces, and u a continuous linear

mapping ji-om E into F. The following conditions are equivalent :
(i) u is surjective.
(ii) For every semi-norm p E SeE), there exists q E S(F) such that we have If I :( q
for every linear form f E F' satisfying If 0 ul :( p.
(iii) For every semi-norm p E SeE), there exists q E S(F) satisfying the following
property : if a linear form f E F' satisfies If 0 ul :( p, then f is null on points where q
is null andfor all y E F, r E S(F), there exists x E E with r(u(x) - y) = O.
(iv) For every semi-norm p E SeE), we have

(1) sup If(y) I < + (JJ for all y E F.

fEF
1f u1
o ~p

We shall prove the proposition according to the following logical scheme

(i)

(.. )~~~( ...)

~ II
(iV)~
11 111

If u is surjective, it is a strict morphism (IV, p. 28, tho 1) then for every semi-norm
p SeE), there exists a semi-norm q E S(F) such that, for all y E F satisfying q(y) :( 1,
E
there exists x E E satisfying p(x) :( 1 and u(x) = y. We deduce immediately that (i)
implies (ii) and (iii). It is clear that (ii) implies (iv).
To prove that (iii) implies (iv), let p and q be as in (iii). Let y E F, by (iii), there
exists x in E such that q(u(x) - y) = O. If f E F' satisfies If 0 ul :( p, then we have
f(u(x) - y) = 0, hence

If(y)1 = If(u(x» I :( p(x)

and the relation (1) is satisfied.

Finally we prove that (iv) implies (i). Let p E SeE) and let q be the superior envelope
of the functions If I for f E F' satisfying If 0 ul :( p. By (iv), q is finite on F, and is
evidently a lower semi-continuous semi-norm on F; since F is barrelled (III, p. 25,
corollary), we have q E S(F). Let Bp (resp. Bq) denote the set of all x E E (resp. y E F)
such that p(x) :( 1 (resp. q(y) :( 1). We have q 0 u :( p, and so u(Bp) c Bq . The polar
of u(Bp) in F' consists of linear forms f E F' such that If 0 ul :( p, hence IfI :( q;
in other words, we get u(Bp)O c B;, and finally that u(Bp) = Bq follows from the
theorem of bipolars (II, p. 45, cor. 3). If U is a neighbourhood of 0 in E, there exists
p E SeE) such that Bp c U, hence u(U) contains the neighbourhood Bq of 0 in F.
This implies that u is surjective (I, p. 17, tho 1).
TVS IV.32 DUALITY IN TOPOLOGICAL VECTOR SPACES §5

COROLLARY. - Suppose E and F are Banach spaces. The following conditions are
equivalent :
(i) u is surjectil'e.
(ii) There exists a real number r> 0, such that II.n,,;; r.II'u(f)I! j()r all fEFf.
(iii) For aU I' E F, \1'1' have sup If(y) I < + 00.
JEF'
irouii '" 1
The conditions (ii) and (iii) are in fact reformulations of conditions (ii) and (iv)
of prop. 4 for Banach spaces.

§ 5. COMPACTNESS CRITERIA

1. General remarks

Let A be a subset of a topological space E, For a sequence (xlI)nEN of points of A

to have a point .\ of E as a limit point, it is necessary and sufficient that the following
condition is satisfied (GT. I. ~ 7, No.3) :
(A) For every integer m ~ 0 and every neighbourhood U oj' x, there exists an
integer n ~ m such that XII E U.
A sequence of the form (Yk)kEN with Yk = x llk for a strictly increasing sequence
(nk)kcN of positive integers is called an extracted sequence of the sequence (xn)nEN'
If there exists an extracted sequence of the sequence (xlI)ncN which converges to x,
then x is a limit point of (x n ); conversely, if x has a countable fundamental system
of neighbourhoods, and x is the limit point of the sequence (XII)' then there exists
an extracted sequence of (XII) converging to x.
On account ofGT, IX. ~ 2, No.9, corollary, we conclude that when E is metrizable,
the following conditions are equivalent :
a) the set A is relatively compact in E;
b) every infinite sequence of points of A has a limit point in E;
c) from erery infinite sequence of points of A, we can extract a sequence which
converges to a point of E,
In this section, we shall extend this criterion to certain non metrizable topological
vector spaces. The following proposition enables us to reduce the study of compact
sets to that of weakly compact sets in a number of cases.

PROPOSITION 1. - Let E be a HausdO/jJ locally convex space, and A a subset of E.

Let E" denote the space E with the weakened topologv.
a) If every infinite sequence of points of A has a limit point in E, then A is pre-
compact in E.
b) In order Ihat A he relatively compact in E, it is necessary and s~dJicient that
it is preeompact in E and relatively compact in E".
We shall prove a) by reductio ad absurdum. If A is not precompact, then by
tho 3 ofGT, II, § 3, No, 7, it follows that there exists a symmetric convex neighbourhood
V of 0 in E such that A cannot be covered by a finite number of translates of V.
No.2 COMPACTNESS CRITERIA TVS IV.33

In other words, if x o, Xl' ... , X,,-l are points of A, then A cj: U (Xi + V) and so
O~i<1I

there exists a point Xli of A such that Xli - Xi ¢: V for 0 :( i < n. Then, by induction
on the integer n, we can construct an infinite sequence (X,)IiEN of points of A such
that Xli - xm ¢: V whenever n > m; since V is symmetric, we also have xm - Xli ¢: V
for m # n and the sets Xn + t V are disjoint. For every point X in E, there exists
at most one integer n ~ 0 such that XII EO X + t V, hence the sequence (Xn)IIEN does
not have any limit point. This proves a).
Now suppose that A is pre compact in E and is contained in a compact subset B
of E(J' Then B is complete in E(J' hence also in E (IV, p. 5, Remark 2). We have A c B,
hence A is relatively compact in E. The converse is evident and b) follows.

2. Simple compactness of sets of continuous functions

In this section, X denotes a compact space and ~/X) the space of continuous
functions on X, with values in the field K (equal to R or C). The space ~/X) is assigned
the topology of simple convergence on X.

PROPOSITION 2. - Let D be a dense subset of X and A a subset of the space ~,(X).

The following conditions are equivalent:
(i) A is relatively compact in ~s(X).
(ii) From every infinite sequence of elements of A, we can extract a sequence con-
verging in ~s(X).
(iii) Every infinite sequence of elements of A has a limit point in ~,(X).
(iv) Let U;,)nEN be a sequence of functions belonging to A and (Xm)mEN a sequence
of points of D. If the iterated limits

y = lim lim f.lx m ) , 8 = lim lim f.,(x m )

m--+oo II--+OC' n--+c{) m--+CXJ

exist, then they are equal. In addition, we have sup If(x) I < + CD for all X EO X.
_ fEA
(i) => (ii) : let A be the closure of A in ~s(X). Assume that A is compact, and con-
sider a sequence of functions f., EO A (for nEON). Let cp be the continuous mapping
X f---+ (f.lx))nEN from X into the metrizable space KN. The image X' of X under cp
is a compact metrizable space, since X is compact. Let E be the closed subspace
of~/X) consisting of continuous functions f on X such that the relation cp(x) = cp(y)
implies f(x) = fey) for every pair of points x, yin X. By cor. 2 ofGT, I, § 9, No.4 and
prop. 3 of GT, I, § 5, No.2, the mapping f' f---+ f' 0 cp is a homeomorphism cp* from
~,(X') onto E. Hence the set A' = (cp*)-l(A) is compact in ~s(X'), and it is clear
that there exist elements f: in A' such that CP*(f.;) = .r,; 0 cp is equal to f.,.
Since X' is a compact metrizable space, there exists a countable dense subset
D' in X' (GT, IX, § 2, No.8, prop. 12 and § 2, No.9, prop. 16). Let ,cTl (resp. 3 2 ) be the
topology on A' induced by the topology of simple convergence on D' (resp. X'). Then
3 1 is metrizable, 3 2 is compact and finer than 3 1 , hence 3 1 and .'12 coincide;
in other words, A' is a compact metrizable subspace ofeg/X'). Therefore, there exists
TVS IV.34 DUALITY IN TOPOLOGICAL VECTOR SPACES §5

a sequence U:J extracted from U:) and converging to an element f' of ~JX').
Therefore, the sequence (1"J converges to f = l' 0 <p in ~JX).
(ii) =:> (iii) : this is clear.
(iii) =:> (iv) : suppose that every infinite sequence of elements of A has a limit
point in ~s(X). Let x E X. The mapping <Px:f~ f(x) from A into K is continuous.
Consequently, every infinite sequence in <Px(A) has a limit point; since the field K
(equal to R or C) is metrizable, the set <p)A) is relatively compact in K, hence bounded.
In other words, we have sup If(x) I < co.
fEA
Let 1", x m ' y and 8 be as in (iv). Let f be a limit point of the sequence Cr.,) in ~,(X),
and let x be a limit point of the sequence (x m ) in the compact space X. For every
m, the mapping h ~ h(xm ) from ~s(X) into K is continuous. In view of the hypo-
theses, we have f(x m) = lim J,.(x m), and hence y = lim f(x m); since f: X ---+ K is
continuous and x is a limit point of the sequence (x m), we get y = f(x). In an ana-
logous way, we prove the equality 8 = f(x), whence y = 8.
(iv) =:> (i) : suppose that the set of numbers f(x), as f ranges over A, is bounded
in K for all x E X. This is equivalent to assuming that the closure A of A in the pro-
duct space K X is compact (GT, I, § 9, No.5). Suppose Q1at A is not relatively compact
in ~s(X). This means that there exists a function u E A and a point a E X such that
u is not continuous at a. Hence there exists a real number E > 0 such that in every
neighbourhood U of a, there exists a point x with Iu(x) - u(a) I ~ E.
We shall construct by induction a sequence (X")"EN of points in D and a sequence
U',)"EN of elements of A, satisfying the following relations

for m ~ I;

Iu(x) - fm(x) I ::::; - - I

1 for 0::::; i ::::; m - I ;
m+
Ifm(x) 1 for 0::::; m ::::; i.
- fm(a) I ::::; i +I
We take Xo = a with fa arbitrary in A (the set A is not empty, otherwise it will
be relatively compact in ~s(X)). Let n ~ 1 and x o' xl' ... , X,,-l' fa, j~, ... ,1,'-1
sati~y relations (1)m' (2)m for I ::::; m < nand (3)m,i for 0 ::::; m ::::; i < n. Since u belongs
to A, there exists J,. E A satisfying (2)". Let V" be the set of all x E X such that
Ifrn(x) - fm(a) I : : ; n ! I for 0 ::::; m ::::; n. Since 1" is continuous, V" is a neighbourhood
of a; choose a point x" in D n Vn such that lu(x,,) - u(a) I ~ E, hence (1)" and (3)m,n
are satisfied. Therefore, the construction can be continued.
Since u(X) is a compact subset of K, there exists a sequence (h) extracted from
ex-m ) and such that the limit y = lim U(Yk) exists. By (2)m' we have u(x) = lim j~ (Xi)
k-+oc n--+rx;

for all i EN, hence

(4) y = lim lim .f,,(Yk) .
k --+ 'lJ Tl--+ (.(;
No.3 COMPACTNESS CRITERIA TVS IV.35

On the other hand we have !;,(a) = lim .f,,(x) by (3)m.i hencef"Ca) = lim !;.(h).
i--+ oc-, k--+ .::r;.

Since Xo = a, we deduce from (2)m that lim .f,,(a) = u(a). Consequently,

n--+'X)

(5) uta) = lim lim !;,(Yk) .

II-X k-x

Finally, from (l)m' we get IY ~ uta) I ~ E, and so y =f. u(a). This contradicts asser-
tion (iv); we have thus proved that (iv) implies (i).

3. The Eberlein and Smulian theorems

THEOREM 1 (Eberlein). - Let E be a Hausdorff and quasi-complete locally convex

space, :Y a topology on E which is compatible with the duality between E and E' and
A a subset of E. For A to be relativezv compact for :Y, it is necessary and sufficient
that every infinite sequence of points of A has a limit point in E for 21.
The condition stated is obviously necessary.
Suppose now that every infinite sequence of points of A has a limit point for :Y,
hence also for the coarser topology cr(E, E'). Then A is precompact for .'!/ (IV, p. 32,
prop. 1); in order that A be relatively compact for :Y, it is necessary and sufficient
that it be so for cr(E, E') (lac. cit.). Therefore it is enough to prove the theorem when:Y
is the weakened topology cr(E, E').
Let E denote the completion ofE, which we shall identify as usual with a subspace
of the algebraic dual E ' * of E' (III, p. 21, tho 2). Let E", E" and E~* denote the spaces
E, E and E'* endowed with the topologies cr(E, E'), cr(E, E') and cr(E'*, E') respec-
tively.
Let (X;)iEI be a basis of the vector space E' over the field K. The mapping
fl-+ (f(X;»)iEI is a homeomorphism <P from E~* onto KI; for every i E I, the image
of A under the mapping x; from E into K is relatively compact: for, K is metrizable
and every infinite sequence of elements of x;(A) has a limit point. If follows that
<p(A) is relatively compact in K 1, hence that the closure A of A in E~* is compact.
Next we shall prove that A is contained in E. Let H be an equicontinuous subset
of E'; let X be its closure for cr(E', E); X is compact (III, p. 17, cor. 2). For every
x E E'*, let <Px be the restriction of x' 1-+ <x, x' >to X; let A c ~s(X) be the set of
functions <Px as x ranges over A.In view of the hypothesis on A, every infinite sequence
of elements of A has a limit point in ~,(X) ; by prop. 2 (IV, p]3), the set A is therefore
relatively compact in ~s(X). It follows that for every a E A, the function <Po on X
is continuous. The inclusion AcE then follows from tho 2 of III, p. 2l.
N ow we shall show that A is contained in E. Since A is pre compact in E" (IV, p. 32,
prop. 1), it is bounded in E" (III, p. 3, prop. 2), hence also in E (IV, p. 1, prop. 1).
Let C be the closed convex balanced envelope of A in E. Then C is bounded since
A is bounded, hence complete since E is quasi-complete. In other words, C is a
convex and closed subset of E, so also of E" (IV, p. 1, prop. 1). Since A c C and
the topology of E" is induced by that of E~*, we have A c C, and hence AcE.
TVS IV.36 DUALITY IN TOPOLOGICAL VECTOR SPACES §5

Since the topology of Eo is induced by that of E~*, the subset A of Eo is compact,

and tho 1 follows.
THEOREM 2 (Smulian). - Let E be a FYI!chet space and A a subset of E. Let Eo denote
the space E endowed with the weakened topology. The following conditions are equi-
valent:
(i) A is relatively compact in Eo;
(ii) every infinite sequence of points of A has a limit point in Eo;
(iii) from every infinite sequence of points of A, we can extract a sequence which
converges in Eo'
The equivalence of (i) and (ii) follows from Eberlein's theorem and (iii) obviously
implies (ii).
We shall prove that (i) implies (iii). Suppose that the closure B of A in Eo is com-
pact and that (X")"EN is a sequence of points of A. Let F denote the smallest closed
vector subspace of E containing the x n ' this is a Frechet space satisfying the first
axiom of countability. Since F is closed in Eo and the topology a(F, F') on F is
induced by aCE, E'), the set B n F is compact for a(F, F'). On account of the remarks
in IV, p. 32, the existence of a sequence extracted from (X")"EN converging for aCE, E')
(or, which is the same, for a(F, F')) is a consequence of the following lemma:
Lemma 1. - Let F be a Frechet space satisfying the first axiom of countability. Every
subset C of F which is compact for the topology :Y induced by a(F, F') is metrizable
for this topology.
Since the topology of precompact convergence on F' is finer than the topology
a(F', F), there exists an everywhere dense countable subset in F~ (III, p. 18, cor. 1).
Hence the set C can be identified with a subset of K D , and the topology induced
on C by that of K D , which is metrizable (GT, IX, § 2, No.8) is coarser than the topo-
logy induced by a(F, F'), for which C is compact. Hence these two topologies are
identical (GT, I, § 9, No.4, cor. 3).

Smulian's theorem can be extended to the case where E is the strict inductive limit
of a sequence of Frechet spaces (IV, p. 67, exerc. 2).

*4. The case of spaces of bounded continuous functions

For every topological space X, let yjb(X) denote the Banach space of all conti-
nuous and bounded mappings from X into K, with the norm defined by

(6) Ilfll = sup If(x) I

XEX

(GT, X, § 3, No.2). When X is compact, every continuous function on X is bounded

(GT, IV, § 6, No.1), and we write yj(X) for yjb(X).
In this and the following section, we shall use the following lemma, which is a
particular case of Lebesgue's theorem (INT, IV, 2nd ed. § 4, No.3, tho 2) on account
of the interpretation of the elements of yj(X)' as measures on X.
No.5 COMPACTNESS CRITERIA TVS IV.37

Lemma 2. - Let X be a compact space. If a sequence U;,)1JEN is bounded in C(j (X)

and converges simply on X to a continuous function 1, then fl(f) = lim fl(f.,) for every
1/- ~fJ

fl in C(j (X)'.

PROPOSITION 3. - Let X be a compact space, and let A be a bounded subset of C(j (X).
For A to be relatively compact for the topology of simple convergence, it is necessary
and sufficient that it is relatively compact for cr(C(j(X), (&'(XY).
The topology of simple convergence is Hausdorff and coarser than cr((&,(X), C(j(X)'),
hence the condition stated is sufficient (GT, I, § 9, No.4, cor. 3).
Now suppose that A is relatively compact for the topology of simple convergence.
Let U;,)nEN be a sequence of elements of A. By prop. 2 (IV, p. 33), there exists a sequence
(f.,) extracted from UJ and converging simply to a continuous function f By
lemma 2, the bounded sequence U;,) tends to f for cr(C(j(X), C(j(XY). Then Smulian's
theorem (IV, p. 36, tho 2) shows that A is relatively compact for cr(cg(X), C(j(XY).

COROLLARY. - Let S be a topological space and A a bounded subset ofC(jb(S). The

following conditions are equivalent :
(i) A is relatively compact for cr(C(jb(S), C(jb(S'));
(ii) if (J;,)nEN is a sequence of elements of A and (Xm),1JEN is a sequence of points
of S such that the iterated limits

exist, then y = 8.
Let X be the Stone-Cech compactification ofS (GT, IX, § 1, No.6) and:x the canoni-
cal mapping from S into X. Put D = :xeS). The mapping <p :f f---+ f 0 :x is an isomor-
phism from the normed space C(j(X) onto the normed space C(jb(S) ; put A = <p -leA).
Since X is compact and D is dense in X, the prop. 2 (IV, p. 33) shows that condition (ii)
is equivalent to the compactness of A for the topology of simple convergence. The
equivalence of (i) and (ii) then follows from prop. 3. *

*5. Convex envelope of a weakly compact set

THEOREM 3 (Krein). - Let E be a Hausdorff and quasi-complete locally convex

space, and let :Y be a topology on E compatible with the duality between E and E'. Let
A be a subset of E which is compact for /Y. Then the closed convex balanced envelope
C of A is compact for :Y.
We shall first make several reductions.
A) The set C is precompact for :Y (II, p. 25, prop. 3), and A is compact for cr(E, E').
On account of prop. 1 (IV, p. 32), it is enough to prove that C is compact for cr(E, E').
and so we have reduced to the case where :Y = cr(E, E').
B) Since C is pre compact and closed for cr(E, E'), it is bounded and closed for
the initial topology of E (III, p. 3, prop. 2 and IV, p. 1, prop. 1); hence it is complete
since E is quasi-complete. In other words, C is the closed convex balanced envelope
TVS IV.38 DUALITY IN TOPOLOGICAL VECTOR SPACES §5

of A in the completion E of E. Since the topology crCE, E') induces cr(E E') on E,
we have reduced to the case when E is complete.
C) Let r be the convex balanced envelope of A. Then C is the closure of r for
cr(E E'). By Eberlein's theorem (IV, p. 35, tho 1), it is enough to prove that evcry
scquence (X"\EN of points of r has a limit point for cr(E, E') in E. But xn belongs
to the convex balanced envelope of a finite subset Bn of A. Let F be the closed vector
subspace of E generated by the countable set B = U B". Then F is complete, the
"
topology cr(F, F) on F is induced by cr(E, E') and we have x" E F for all n E N. Hence
it is enough to prove that (xn)nEN has a limit point for cr(F, F), which gives the reduc-
tion to the case when there exists a countable dense set in E.
Let A be assigned the topology induced by cr(E, E'), which makes it a compact
space. We define a linear mapping u: E' -> ,(;(A) by

(7) u(x') (a) = (a, x') (a E A, x' E E').

Let (x;.)nEN be an equicontinuous sequence in E', converging to 0 for cr(E', E). Then
the sequence of functions u(x;,l is bounded in '6'(A) and converges simply to O. For
every fl E (6'(A)" we have lim ~L(U(X~») = 0 by lemma 2 (IV, p. 37). By the criterion
given in the remark in ilL p. 21, the linear form fl 0 1I on E' is then continuous for
cr(E', E) for every fl E ~(A)'. Hence there exists a linear mapping v :~(A)' -> E
satisfying the relation

(8) (u(x'), fl) = (r(~l), x')

\
(x' E E', fl E (fiCA),) .

It is clear that v is continuous if ~(A)' is assigned the topology cr(~(A)'. "&(A)) and E
the topology cr(E, E').
The unit ball (closed) B of the Banach space <'6(A) is compact for the topology
cr((g(A)', ~(A») (III. p. 17, cor. 3). Consequently, t{B) is a convex balanced and compact
subset of E for cr(E. E'). For every a E A, the continuous linear form Ea:fl---> f(a)
on (fi(A) belongs to B, and we have V(Ea) = a by formulas (7) and (8). Hence, A c v(B),
and so C c vCB). This proves that C is compact for cr(E, E').
Q.E.D.
 APPENDIX

Fixed points of groups

of affine transformations

1. The case of solvable groups

Let E be a real vector space, and K a convex subset of E. A mapping u: K -> K

such that for x, y in K and for every real number tin (0, I), we have

(1) u(tx + (I - t) y) = tu(x) + (1 - t) u(y)

is said to be an affine transformation. From relation (1) we deduce that

(2) u(I tixJ = I tiu(x)

iEI iEI

for every finite set I, points Xi in K and positive real numbers ti such that I ti = l.
iEI
Let u and v be two affine transformations on K, then the mapping u 0 v is an
affine transformation on K. If v : E -> E is a linear mapping such that v(K) c K,
the mapping u: K -> K which coincides with v on K is an affine transformation.

THEOREM 1 (Markoff-Kakutani). - Let E be a Hausdorff locally convex vector space

over thefield R, and K a non-empty compact convex subset ofE. Let be a set of conti-r
nuous affine transformations on K, pairwise permutable. Then there exists a point a
in K such that u(a) = a for all u E r.
For every u E r, let Ku be the set of all x E K such that u(x) = x. We shall show
that Ku is non-empty. Let x be a point of K; for every integer n ;;:: 1, let xn be the
1 11-1 .
element - I u'(x) of E. Since K is convex and stable under u, the points Xn belong
n ;=0
to K and since K is compact, there exists a limit point a of the sequence (Xn),,;o, l '
The mapping y f---+ u(y) - y from K into E is continuous, hence u(a) - a is a limit
point of the sequence (u(x,) - x,,)n;o, l ' But we have u(xn) - xn = ~ (u"(x) - x).
TVS IV.40 FIXED POINTS OF GROUPS OF AFFINE TRANSFORMATIONS APP.

Since K is compact, hence also bounded (III, p. 3, prop. 2), the sequence (un(x) - X)n> 1

is bounded; consequently, the sequence

n
(!
(u"(x) - X))
">1
tends to 0 (III, p. 4,

prop. 3), and since E is Hausdorff, we have u(a) - a = 0. Therefore a E Ku'

Each of the sets Ku is a closed and convex subset of the compact space K, and
we shall prove that the intersection n Ku is non-empty. For this it is enough to
UEf

prove that, for n ;?: 1, and Up ... , u" in r, the set KUl n ... n KUn is non-empty.
The case n = 1 having been considered, we argue by induction on n. Suppose
n ;?: 2 and put L = Ku I n ... n KUn _ l ' By the hypothesis of induction, L is a non-
empty compact convex subset of E. Since Un commutes with u 1 ' ... , un -1' we have
u,,(L) c L. Applying the first part of the proof to the affine transformation induced
by Un on L, we conclude that there exists a point a in L such that un(a) = a; then
a belongs to KUl n ... n K un ' which is then non-empty.

COROLLARY. - Let G be a solvable group of continuous affine transformations on K.

Then there exists a point in K which is invariant under G.
By the definition of a solvable group (A, I, § 6, No.4) there exists a finite decreasing
sequence (G)o",j"," of distinct subgroups ofG, such that Go = G, G n = {e} and
such that the group G j_ dGj is commutative for 1 ~ i ~ n. Let K j denote the set
of fixed points of G j in K. Then Kn = K. Moreover, for 1 ~ i ~ n, every element
of G j induces the identity transformation on Kj; we thus deduce an action of the
abelian group G j_ dG j on K if K j is non-empty; it follows from tho 1 that the set
K j- 1 of fixed points of G j_ dGj in K j is non-empty. By descending induction on i,
we conclude that K o is non-empty, hence the corollary.

2. Invariant means

Let X be a topological space. Let 2B(X; R) denote the real vector space consisting
of continuous bounded mappings from X into R. Endowed with the norm
II f II = sup If(x) I, this is a Banach space (GT, X, § 3, No.1); it is also an ordered
XEX

vector space, where the relation f ;?: 9 means «f(x) ;?: g(x) for all x EX».

DEFINITION l. - A positive linear form J.! on the space 2B(X; R), where X is a topo-
logical space, for which IIJ.!II = 1, is called a mean on X.

* When X is compact, a mean on X is a positive measure on X such that ~(X) = 1. *

Lemma l. - The set K of means on X is the subset of the unit ball of the dual of the
Banach !>pace E = 2B(X; R) whose elements are the linear forms J.! such that J.!(l) = 1.
It is a subset ofE' which is convex and compact for cr(E', E).
Let J.! be a linear form on E, such that J.!(l) = 1. For every function fEE, we define
the function f' E E by f'(x) = II f II - f(x) (x E X). First assume that J.! is a mean;
foreveryfEE,wehavef';?: O,henceJ.!(f');?: O,i.e.J.!CD ~ Ilfll;thereforellJ.!11 ~ l.
No.3 FIXED POINTS OF GROUPS OF AFFINE TRANSFORMATIONS TVS IV.4I

Conversely, suppose J.l belongs to E', and that 11J.l11 ::;; I; for every positive function
fEE, we have J.l(f') ::;; II f' II, hence

II fll - J.l(f) = J.l(f') ::;; II f'11 ::;; II fll ,

and finally J.l(f) ~ 0; consequently, J.l is a mean.

It is clear that K is convex; that it is compact for cr(E', E) follows from cor. 3
of III, p. 17.
Q.E.D.
Let r be a set of continuous mappings from X into X which commute pairwise.
Let Y E r. For every function fEE, we have f 0 Y E E; hence we can define an affine
transformation uy on the set K of means on X, by

UyJ.l(f) = J.l(fo y) (J.l E K, fE E).

If K is assigned the topology induced by cr(E', E), the mapping uy is continuous.

If y is a homeomorphism, UyJ.l can be deduced from J.l by transport of structure.
Finally, we have UyU y' = Uy,U y for all y, y' in r. By the Markoff-Kakutani tho (IV,
p. 39, tho 1), there exists a mean J.l on X, such that uyu = J.lfor all y E r; in other words,
J.l satisfies the relation J.l(f) = J.l(f 0 y) for fEE and y E r.
In an analogous way, the corollary ofth. 1 (IV, p. 40) implies the following result :
PROPOSITION 1. - Let X be a topological space and G a solvable group. We assume
that G operates on X on the left, in such a way that for all g E G, the mapping x ~ g.x
from X into X is continuous. Then there exists a mean on X which is invariant under G.
COROLLARY. - Let G be a solvable topological group. Then there exists a mean on
G which is invariant under the left and the right translations.
It is enough to apply prop. 1 to the solvable group G x G acting on G by
(g.g').x = gxg,-l.

3. Ryll-Nardzewski theorem

In this section, E denotes a normed space over the field Rand :Y a Hausdorff
locally convex topology on E for which the norm of E is lower semi-continuous.
These hypotheses are in particular satisfied in the following cases :
a) :Y is the topology induced by the norm of the normed space E.
b) :Y is the weakened topology cr(E, E') of the normed space E.
c) E is the dual of a normed space F and :Y = cr(F', F).
d) There exist two normed spaces F 1 and F 2 such that E = £,(F 1 ; F 2) and :Y
is the topology of simple convergence.
Unless otherwise expressely stated, the topological notions refer to the topology :Y.
Let K be a convex subset of E. Suppose that K is compact (for the topology :Y),
and that K satisfies the first axiom of countability for the distance defined by the
norm ofE.
TVS IV.42 FIXED POINTS OF GROUPS OF AFFINE TRANSFORMATIONS APP.

Lemma 2. - Suppose K contains at least two points. For every E > 0, there exists
a partition of K into two non-empty subsets KI and K 2 , having the following proper-
ties:
a) KI is convex and compact;
b) we have Ilxl - x 2 11 < E for every Xl and X2 in K 2 •
Let L be the closure of the set of all extremal points of K. By the Krein-Milman
tho (II, p. 55, tho 1), K is the closed convex envelope of L. Since K contains at least
two points, so does L. For every x E L, let Ax be the set of all Y E L such that
Ilx - yll < E/4. By the hypothesis, on K, there exists a countable subset D of L
such that L = U Ax' Since the norm is lower semi-continuous, each of the sets
XED

Ax is closed. Applying Baire's tho (GT, IX, § 5, No.3, tho 1) to the compact space L, we
see that there exists a point a in D and an open subset U in E such that L I i U is
non-empty and is contained in Aa' Since L contains at least two points, and since E
is Hausdorff, we can choose U in such a way that L ¢ U.
Let M be the closed convex envelope of L I i CU. For every real number t such
that 0 < t < 1, let M t be the set of all vectors of the form tX I + (1 - t) x 2 with
Xl EM and x 2 E K; this is a non-empty, compact convex subset of K. We shall
prove that M t i= K by reductio ad absurdum. Suppose that M t = K; then every
extremal point X in K belongs to Mp hence can be written in the form
x = tX I + (1 - t) x 2 with Xl EM and x 2 E K. This implies that X = Xl = x 2 '
and so x EM. By Krein-Milman tho (II, p. 55, tho 1), we have K = M, and K is the
closed convex envelope of L I i CU. By II, p. 56, corollary, this implies that
L c L I i CU, which contradicts the relation L I i U i= 0.
Put d = sup Ilx - yll and choose a real number t such that 0 < t < I and
XEK,YEK
t < E/4d. Put KI = M t and K2 = K - M t . By the preceding argument, the sets
KI and K2 are non-empty, and KI is convex and compact. Let M' be the closed
convex envelope of L I i U. Since K is the closed convex envelope of the set
L = (L I i CU) u (L I i U), it is also the closed convex envelope of MuM'; Let
Xl and x 2 be two points in K 2 ; for i = 1,2, there exist Yi E M, Zi E M' and a real
number cri such that 0 < cri < 1 and Xi = criYi + (1 - cr) Zi' If cri ): t, then

Xi = tYi + (l - t) { ~i ~ : Yi + ~ =- ~i Zi } ; this contradicts the assumption that

Xi $ Mi' Hence cri < t, for i = 1, 2, and so

For every point Z in M', we have liz - all < E/4, since L Ii U c A a , and so, in
particular Ilzi - all < E/4. Thus
2
IIXI - x 2 11 <L (1lx i - z;II + IIZi - all) < E.
i~ 1

This completes the proof.

No.3 FIXED POINTS OF GROUPS OF AFFINE TRANSFORMATIONS TVS IV.43

Lemma 3. - Let G be a group of continuous (for Y) affine transformations on K.

Suppose that K is non-empty and that Ilgx - gyll = Ilx - yll for all x, y in K and
all 9 in G. Then there exists a point in K which is invariant under G.
Let J be the family of non-empty subsets of K which are closed convex and stable
for G. If (LJael is a family of elements of J which is totally ordered by inclusion,
then the set L = n La belongs to J. Consequently (S, III, § 3, No.4, tho 2), there exists

an element L in J which is minimal for the relation of inclusion. We shall prove

that L reduces to a point.
We argue by reductio ad ahsurdum, assuming that L contains at least two distinct
points Xl and x 2 ' put x = (Xl + x 2 )/2 and £ = Ilxl - xzl!/2. The convex and
compact set L satisfies the first axiom of countability for the distance defined by the
norm (GT, IX, § 2, No.8). Hence we can apply lemma 2 and find a compact and
convex subset Ll of L, distinct from 0 and from L, having the following property:
perty :
(A) For every J'I and yz in L - L 1 , we have IIY1 - hil < £.
We shall prove by reductio ad absurdum that gx E L1 for all 9 E G. Let g E G
be such that gx E L - L1 then we have

Ilgxi - gxll = Ilxi - xii = Ilxl - x211/2 = £ ,

for i = 1,2. By property (A), we have gXi ELl. Since Ll is convex, we conclude
that gx = (gx1 + gX z )/2 belongs to L 1 , which contradicts the assumption.
Let L' be the closed convex envelope of the orbit Gx of x. The set L' belongs to 3.
By the preceding argument, we have L' c L 1, hence L' c L, L' 0;6 L. This contra-
dicts the minimal character of L and the proof is complete.

THEOREM 2 (Ryll-Nardzewski). - Let E be a normed space and K a non-empty

convex subset of E, which is compact for the weakened topology a(E, E'). Let G be
a group of isometric affine transformations ofK. Then there exiSTS a point in K which
is invariant under G.
For every 9 E G, let Kg denote the set of all points x in K such that gx = x; let K
be assigned the weakened topology: each set Ky is convex and closed in the compact
space K. We shall prove that the intersection n Kg is non-empty; for this, it is
geG
enough to prove that the set K91 n ... n Kg» is non-empty for every 91' ... , gn in G.
Fix 91' ... , gn and let H be the subgroup of G generated by {gl' ... , gn }. Choose
a point a in K and let L denote the closed convex envelope of the orbit Ha of a.
Let D be the countable set of elements of the form lel hi a + ... + Amh",a, where
A1 , ... , Am are positive rational numbers such that A1 + ... + Am = 1, and h1' ... , hm
are elements in H. The closure D of D for the strong topology, is convex, hence
it is closed for crCE, E') (IV, p. 4, prop. 2); therefore D = L and this proves that L
is a metric space satisfying the first axiom of countability for the distance
(x, y) I--> II x - y II. We can now apply lemma 2. There exists a point b in L which
is invariant under H, hence bE KYI n ... n Kg».
TVS IV.44 FIXED POINTS OF GROUPS OF AFFINE TRANSFORMATIONS APP.

COROLLARY. - Let E be a reflexive Banach space, G a group of automorphisms of

the normed space E, andK a subset ofE. Suppose that K is non-empty, convex, closed,
bounded and stable under G. Then there exists a point in K which is invariant under G.
Since E is reflexive, K is compact for cr(E, E') (IV, p. 15, tho 1). Moreover, every'
element of G belongs to 2(E).

4. Applications.

* A) Unitary representations of groups

Let E be a complex hilbertian space, G a group and n a unitary representation
of G on E, i.e. a homomorphism from G into the group of automorphisms of E.
Let EG be the hilbertian subspace of E consisting of all vectors invariant under
neG). For every x E E, let Kx be the closed convex envelope of the orbit of x. Fix
a point x in E.
We shall show that there exists a unique point in Kx which is invariant under neG),
namely the projection of x on EG. By IV, p. 44, corollary (applied to the underlying
real vector space to E), there exists a point in Kx which is invariant under n(G);
let a be such a point; then a E EG. Let P be the set of all y E E such that y - x is
orthogonal to EG; we see immediately that P is closed, convex and invariant under
n(G); therefore x E P, hence Kx c P and finally a E P. In other words, a - x is
orthogonal to EG; consequently a is the projection of x onto EG. *

* B) Trace o/' an operator in a hilbertian space:

Suppose that the representation n is irreducible, that is, that there exists no hil-
bertian subspace of E, distinct from {O} and from E, which is invariant under neG).
Let F = 22(E) be the hilbertian space of all Hilbert-Schmidt endomorphisms
>
of E, with the scalar product <ulv = Tr(u*v). We define a unitary representation
A from G into F by the formula

(3) A(g). u = neg) un(g)-l (u E F, g E G) .

The space FG of all elements of E invariant under A(G) consists of the Hilbert-
Schmidt endomorphisms u of E which commute with neg) for all g E G. By Schur's
lemma, such a u is a homothety. Hence we must consider two cases
1) if E is infinite dimensional, then FG = {O};
2) if E is finite dimensional, then F = 2(E) and FG = C.l E .
Applying the result of A) to the unitary representation A, we obtain the following
theorem:
Let u E 2 2(E), and let Au be the closed convex envelope in 22(E) of the set of
endomorph isms neg) un(g)-l ofE, where g runs through G. IfE is infinite dimensional,
we have 0 E Au. IfE is finite dimensional with dimension d, there exists a unique homo-

thety in Au' namely the projection ~ Tr(u).IE of u onto the subspace C.l E of 2 2(E). *
NO.4 FIXED POINTS OF GROUPS OF AFFINE TRANSFORMATIONS TVS IV.45

C) Haar measure of a compact group :

Let G be a compact group and let E = rc(G, R) be the Banach space of all real
valued continuous functions on G, endowed with the norm

(4) II f II = sup If(x)l·

XEG

For all x E G, we define the automorphisms Yx and Ox of E by the formulas

(5) yJ(y) = f(x- 1y) , oJ(y) = f(yx)

(for y EG, fE E).

Let fE E, let r J (resp. LlJ ) denote the closed convex envelope t. E, of the set of
all functions y xf (resp. oxf) as x ranges over G. We shall prove Ii, at there exists a
unique constantfunction Il(f) belonging to r J , a unique constantfunction Il'(f) belong-
ing to LlJ' and that these constants are equal.
It is clear that a continuous function on G is invariant under the automorphisms
y x (resp. ox) ofE if and only if it is constant. Now the set of all functions y J(resp. oxf)
for x in G, is compact in E, since the mapping x f---+ yxf (resp. x f---+ oxf) from G into E
is continuous (GT, X, § 3, No.4, tho 3). It follows (II, p. 25, prop. 3) that r J (resp. LlJ )
is a compact set in E for the topology defined by the norm, hence for cr(E, E'). By
the Ryll-Nardzewski tho (IV, p. 43, tho 2), there exist constant functions in r J and LlJ .
lt remains to prove that if C 1 E r J and C z E LlJ are constants, then c 1 = c z .
Let £ > O. By the hypothesis there exist points Xl' ... , x n' Y1' ... , Yn in G and
positive real numbers A1 , ..., An' Ill' ... , 11m such that

(6) Al + ... + An = III + ... + 11m = 1 .

n
(7) sup I L AJ(XiX) - C1 1 ~ £,
XEG i= 1

m
(8) sup I L IlJ(xy) - Czl ~ £.
XEG j= 1

m n

Put r = L AiIlJ(XiY). Then r - c 1 L Ilpj with aj

AJ(XiY) - c 1 ; by = L
i.j j=l ;=1
(7), we have lajl ~ £ for 1 ~ j ~ m, hence Ir - c 1 ~ £. Similarly, we prove the
1

inequality Ir - czl ~ £, hence IC 1 - czl ~ 2£. Since £ is arbitrary, we get c1 = cZ '

as asserted.
By the definition of Il(f), for every £ > 0 we can find positive numbers A1 , ... , An
n
with sum 1 and elements Xl' ... , xn in G such that I L AJ(XiX) - 1l(f)1 ~ £ for all
;= 1
XEG.
It is immediate that for 1, 9 in E and for every scalar A, we have r J+g c r J + rg
and ryJ=ArJ , hence we have the relations 1l(f+g)=Il(f)+Il(g) and Il(Af)= AIl(f)·
Therefore, the mapping 11 :f f---+ Il(f) from E into R is a mean on the compact space G
(IV, p. 40); * in other words 11 is a positive measure on G such that Il(G) = I *.
TVS IV.46 FIXED POINTS OF GROUPS OF AFFINE TRANSFORMATIONS APP.

It is immediate that ~ is invariant under the left translations of G, and the equality
~(j) = ~'(j) implies that ~ is also invariant under right translations. * In other
words, ~ is a left and a right measure on G (INT, VII, § 1, No.2, def 2). *

* D) Existence of invariant measures:

Let X be a Hausdotif topological space, ~ a positive bounded measure on X,
and G a group of homeomorphisms of X. Suppose that for all 9 E G, the measure
g.~, the image of ~L under the mapping g: X ---+ X is of base ~. Let ug be a positive
~-integrable function on X such that g.p. = Ug'~' Suppose also that there exist
two positive ~L integrable functions <p and \j! on X, which are not ~l-null and are
such that <p ~ ug ~ \j! ~-almost everywhere for all 9 E G. We shall prove that there
exists a positive lnmded measure v oF 0 on X. with base ~, and invariant under G.
Let P be the subset of the Banach space E = L leX, p) consisting of classes of
functions f such that <p ~ f ~ \j! ~-almost everywhere. Then P is compact for the
weakened topology aCE, E'). The mapping h f---> h. ~ from P into the Banach space
F = .'ib(X) of bounded real measures on X, is a bijection from P onto a subset
PI of E which is convex and compact for the topology a(F, F'). By hypothesis,
g. ~ E PI for all 9 E G. Let K be the closed convex envelope of the set of all measures
g.~. For all 9 E G, the mapping v f---> g. v is an isometric affine transformation of K.
By the Ryll-Nardzewski tho (TV, p. 43, tho 2), there exists a measure v E K which is
invariant under G. We have <p.~ ~ v, hence v oF O. *
 Exercises

§ I

1) Let A be an infinite set.

a) Let E be the Banach space Jf"(A) over R, consisting of all families x = (X')'EA of real num-
bers such that rt f---> x. tends to 0 with respect to the filter of complements of finite subsets
of A, and endowed with the norm Ilxll = sup Ix,l (when A = N, this space is denoted by
,EA
Co or co(N). Show that every continuous linear form on E can be written in a unique way as
x f---> I u.x., where (U.).EA is a family such that I IUal < + CIJ; then the dual E' of E can
~EA ~A

be identified (as a non topological vector space) with the space £1(A) (I, p. 4, Example).
h) Let F be the Banach space fl(A) (1, p. 4, Example) (when A = N, also denoted by £1). Show
that every continuous linear form on F can be written in a unique way as x f---> I UaXa' where
,EA
(U.)aEAis a bounded family of real numbers; then the dual F' of F can be identified (as a non
topological vector space) with the space .%'(A) = f'x (A).
c) Let B be an arbitrary set, (ca~)( •. ~)EA x B an arbitrary family of numbers> O. Let 0 be the
vector space of all families x = (Xa)'EA of real numbers such that, for every ~ E B, we have
p~(x) = I c,~lxal < + CIJ; the p~ are semi-norms on O. In order that 0, with the topology
aEA
defined by this family of semi-norms, be Hausdorff, it is necessary and sufficient that, for
every rt E A, there exists at least one ~ E B such that c,~ > O. Show that then 0 is complete,
and that every continuous linear form on 0 can be written uniquely as x f---> I u,xa ' where
'EA
(U,)'EA is a family of real numbers satisfying the following condition; there exists a finite
number of indices ~i E B (I :( i :( n) and a number a > 0 such that lu,l :( a. ca~; for all rt E A
and I :( i :( n. Prove the converse, and extend these results to the case where the field of
scalars is C.

2) (i) Let F and 0 be two vector spaces in separating duality. Show that, if F is relatively
bounded for O'(F, 0) (III, p. 43, exerc. 6), then 0 is relatively bounded for 0'(0, F).
TVS IV.48 DUALITY IN TOPOLOGICAL VECTOR SPACES §1

b) Let F be a vector space, and let G l ' G 2 be two vector subspaces of F* such that F is in
separating duality with G 1 and with G 2 • Show that, if F is relatively bounded for er(f', G 1 )
and er(F, G 2)' it is so also for er(F, G 1 + G 2)'
c) Suppose that F has a countable basis. Show that, for every vector subspace G of F* which
is in separating duality with F and has a countable basis, F is relatively bounded for er(F, G)
(by induction define two bases (an), (b n) of F and G respectively, such that am' bn ) = '8 mn ). <
3) Let F be a vector space. Show that, for the topology er(F, F*), every bounded subset of F
is finite dimensional. Deduce that, if F is infinite dimensional, there exist, in the completion F
of F (for er(F, F*)), compact subsets which are not contained in the closure of any bounded
subset of F (el II, p. 52, prop. 10).

-0 4) Let F and G be two vector spaces in separating duality, G (resp. F) being identified
with the dual of F (resp. G) when the latter is assigned the topology er(F, G) (resp. er(G, F)).
Let 6 (resp. '1:) be a covering of F (resp. G) consisting of convex, balanced and bounded subsets
for er(F, G) (resp. er(G, F)). Show that the following propositions are equivalent:
a) Every set M E 6 is precompact for the '1:-topology.
~) Every set N E '1: is precompact for the 6-topology.
y) On every set ME 6, the topology induced by the '1:-topology is identical with the topo-
logy induced by er(F, G).
8) On every set N E '1:, the topology induced by the 6-topology is identical with the topo-
logy induced by er(G, F).
(Use prop. 5 of III, p. 17 to show that a) implies 8) and exerc. 1 of II, p. 74 to show that 8)
implies ~).)

5) Let E and F be two locally convex Hausdorff spaces, 6 a family of subsets of E. In order
that the 6-topology on the space .P(E; F) be compatible with the vector space structure,
it is necessary (and sufficient, cf III, p. 13, corollary) that every set of 6 is bounded in E.

6) a) Let E be a locally convex Hausdortr space. For every ultrafilter U on E, let U' be the
filter for which the convex envelopes of sets of U form a base. Show that, in order that a point
of E be a limit of U for the weakened topology, it is necessary and sufficient that it is a limit
point of U' for the initial topology (use exerc. 11 of II, p. 84).
b) Let A be a convex subset of a vector space E and let :Y1, :Y2 be two locally convex Haus-
dorff topologies on E, and :Y;, :Y; the corresponding weakened topologies. Show that if the
topology induced on A by :Yl is finer than the topology induced by :Y2, then the topology
induced on A by :Y; is finer than the topology induced on A by :Y;.

~ 7) Let F be the direct sum space R(N), G the space e1(N) (I, p. 4, Example); F and G are
put in duality by the bilinear form
(x, Y) f-> I ~n 1']n

for x = (~n) E F and Y = (1']n) E G.

a) Show that, in F, every set K which is convex and compact for er(F, G) is finite dimen-
sional. (Assume the contrary; let (n k ) be a strictly increasing sequence of integers > 0, and
(a k ) a sequence of points of K such that the components of ak for indices > nk are zero, but
that of index n k is #- O. Show that there exists a sequence (lk) of numbers > 0 such that
I tk < + 00 and that in the Banach space .'?I(N) of all bounded sequences of real numbers,
I
k
the point lkak has non-zero components for indices ni , for all i; deduce that the sequence
k
p

of partial sums sJi> = I tka k cannot have a limit point in F for er(F, G).)
k~l

b) Show that in F there exist compact sets for er(F, G) which are infinite dimensional (observe
that G is the dual of F for the topology induced by the topology of the normed space .'?B(N)).
Show that there exist precompact sets in F for er(F, G) which are not relatively compact.
§1 EXERCISES TVS IV.49

e) Deduce from a) and b) that T(G, F) = cr(G, F), but that T(G, F) is distinct from the topo-
logy of uniform convergence on subsets of F which are compact for cr(F, G).

8) Let E be an infinite dimensional locally convex metrizable space, and E' its dual. In order
that the topology T(E, E') be identical with the weakened topology cr(E, E') it is necessary
and sufficient that E is isomorphic to an everywhere dense subspace of the product space
RN (resp. eN). (Observe that in E', for the bomology consisting of equicontinuous sets, there
exists a countable base (Ill, p. 1) consisting of finite dimensional sets; conclude that the vector
space E' has a countable basis).
Give an example of a locally convex Hausdorff space E for which the initial topology,
and the topologies cr(E, E') and T(E, E') are all distinct.

9) a) Let E be a locally convex, Hausdorff, bomological and quasi-complete space. In order

that the topologies T(E', E) and cr(E', E) on the dual E' of E be identical, it is necessary and
sufficient that the topology of E is the finest locally convex topology (show that every bounded
subset of E is finite dimensional).
b) Let E be a locally convex Hausdorff space, E' its dual. In order that the strong topology
~(E', E) on E' be identical with the weak topology cr(E', E), it is necessary and sufficient that
the topology of the bomological space associated with E (Ill, p. 40, exerc. 1) is the finest locally
convex topology on E.

10) Let E be a locally convex Hausdorff space, E' its dual. Show that the following proposi-
tions are equivalent :
IX) E is barrelled;
~) every weakly bounded subset of E' is equicontinuous;
y) every weakly bounded subset of E' is relatively weakly compact, and the topology of E
is T(E, E').

11) Let E be a locally convex Hausdorff space, E' its dual, IS a covering of E consisting
of bounded subsets; we assign the IS-topology to E'. Show that, for the bilinear form
(x, x') ~ <x, x' >to be continuous on E x E', it is necessary and sufficient that the topology
of E can be defined by a single norm, and that the IS-topology is the strong topology on E'
(ef III, p. 37, exerc. 2).

12) Let E be a locally convex Hausdorff space, E' its dual.

a) In order that there exists a weakly bounded and absorbent set in E', it is necessary and
sufficient that the topology of E is coarser than a normed space topology (ef IV, p. 47, exerc. 2).
b) In order that there exists an equicontinuous and weakly total set in E', it is necessary
and sufficient that the topology of E is finer than a normed space topology.
e) In order that there exists an equicontinuous absorbent set in E', it is necessary and suffi-
cient that the topology of E can be defined by a single norm.

13) Let F and G be two vector spaces over R in separating duality.

°
a) Let A be a convex set in F, not containing the origin, and compact for the topology cr(F, G);
let C be the convex cone with vertex generated by A. Show that the polar cone CD in G
has an interior point for the topology T(G, F):
b) Conversely, let C be a proper convex cone with vertex 0, closed for cr(F, G) and having
an interior point for the topology T(F, G). Show that in G there exists a hyperplane H, closed
for cr(G, F), not containing the origin, such that H n Co is compact for cr(G, F) and
(H nCO) u {O} generates Co.

14) Let E be a locally convex Hausdorff and quasi-complete space, E' its dual. Show that
on E~ the topology of compact convergence is the finest of the topologies compatible with
the duality between E and E' and which induces the same topology as cr(E', E) on every equi-
continuous subset of E' (ef IV, p. 48, exerc. 4).
TVS IV.50 DUALITY IN TOPOLOGICAL VECTOR SPACES §1

"IT 15) a) Let E be a locally convex Hausdorff space. A a convex and balanced set in E, and
u a linear form on E. Show that if A" u- 1 (0) is closed with respect to A. then the restriction
of u to A is continuous. (Tf not. show that 0 will be in the closure of the intersection of A and
a hyperplane u- 1 (ct) withct i= 0; deduce that if b E A is such that u(b) = - ct. the point b ±
will be in the closure of A "U 1(0).)
b) Let E be a real locally convex Hausdorff and complete space. E' its dual. Show that if a
hyperplane H' of E' is such that its intersection with every equicontinuous and weakly closed
subset M' c E' is weakly closed, then H' is weakly closed (use a) and Tn, p. 21. cor. 1).
c) Let E be a locally convex Hausdorff space, E' its dual and C a convex. balanced and closed
set in E. Let u be a linear form on E; show that if the restriction of u to C is continuous for
the initial topology, it is also continuous for cr(E. E') (use a)). Show by an example that the
restriction of u to the vector subspace M generated by C is not necessarily continuous (take
E = R(N) with the norm Ilxll = sup I~"I, and for C take a suitable convex set generating E).
n

16) Let F and G be two vector spaces in separating duality, the set of all linear forms x' on F
which are bounded an every subset of F bounded for cr(F, G) is called the enclosure of G
in the algebraic dual F* of F ; this is a vector subspace G of F*. which is the dual of F when
F is assigned the topology of the bomological space associated (IIT, p. 40. exerc. 1) with one
of the topologies compatible with the duality between F and G. Thcn G is said to be enclosed
in F* if G = G.
a) Let M be a closed vector subspace of F for the topology cr(F, G). Show that if G is enclosed
in F*, and if F/M is assigned the topology cr(F/M, MO), then its dual is enclosed in (F/M)*.
b) Let E be a Hausdorlflocally convex space, E' its dual. For E to be bomological, it is neces-
sary and sufficient that its topology is identical with ,(E, E') and that E' is enclosed in E*
c) Let (E)icI be a family of Hausdorff locally convex spaces, and F the direct sum space of
the E i , endowed with the topology defined in 1, p. 24, exerc. 14. Show that the dual of F is
canonically isomorphic to the subspace of the product IT E; of the duals of the E i, consisting
iEI
of all families (x;) such that x; = 0 except for a countable number of indices. (Let Vi be an
arbitrary neighbourhood of 0 in E i . Show that if x; i= 0 for an uncountable set of indices,
then therc exists a number ct > 0 and an uncountable set H c I such that there exists Xi E Vi
< >
for which Xi' X; ;?: ct for all i E H; conclude that (x;) cannot belong to F'.)
d) Show that if I is uncountable then F is not enclosed in F* ; if we take Ei = R for all i E I,
then F' endowed with the strong topology, is not complete. and there exist strongly bounded
subsets in F' which are not weakly relatively compact.

17) Let E be a Hausdorff locally convex space, E' its dual and. in E', let '!ll be the family
of convex eq uicontinuous sets. '!l2 the family of convex, relatively weakly compact subsets,
1J3 3 the family of convex strongly bounded subsets, and 1J3 4 the family of convex weakly bounded
subsets. Then 1J3 1 c l!lz c l!l3 c IJ3 4' Give an example of a space E for which the four families
of sets are distinct (take for E a product of three spaces for which we have, respectively 1J3 1 i= 1J3 2
(ef IV, p. 49, exerc. 8), IJ3 z i= 1J3 3 (exerc. 16, d)), 1J3 3 i= l!l4 (III, p. 23, Remark 2).

18) Let E, F be two vector spaces and E *, F* their respective algebraic duals. Show that
if u is a linear mapping from F* into E, which is continuous for the topologies cr(F*, F) and
cr(E, E*), then u(F*) is finite dimensional. (Use prop. 2 of IV, p. 27 to show that u is a strict
morphism, and deduce that u(F*) is a subspace of E of minimal type (II, p. 85, exerc. 13);
conclude by considering the bounded sets of this subspace.)

19) Let E and F be two Hausdorff locally convex spaces, E' and F their duals. For every
subset H of the space 2'(E; F) of continuous linear mappings from E into F, let 'H denote
the set of the transposes of the mapping u E H. For every subset M (resp. N') of E (resp. F),
let H(M) (resp. 'H(N'») denote the union of the sets u(M) (resp. 'u(N')) as u ranges over H.
a) For H to be equicontinuous. it is necessary and sufficient that, for every equicontinuous
subset N' of F, 'H(N') is an equicontinuous subset of E'.
b) Let is be a set of bounded subsets of E. Show that H is bounded in £'(E; F)for the is-topo-
logy, if and only if for every y' E F, 'H(y') is bounded in E' for the is-topology.
§1 EXERCISES TVS IV.51

c) Let S be a set of bounded subsets of E, 3 a set of bounded subsets of F forming an adapted

bomology of F (III, p. 3, def. 4). Suppose E' is assigned the S-topology and F' the 3-topology.
Then 'H is equicontinuous if and only if for every set BE S, H(B) belongs to 3. In particular,
'H is equicontinuous for the strong topologies on F' and E' if and only if H is bounded in
2(E; F) for the topology of bounded convergence.
d) Deduce from b) and c) that, if 'H is bounded for the topology of simple convergence in
2(F'; E') when F' and E' are assigned the strong topologies, then 'H is equicontinuous for
these topologies.
e) Show that if E is barrelled, the following properties are equivalent:
ex) H is simply bounded in 2(E; F);
~) H is equicontinuous;
y) 'H is simply bounded in 2(F'; E') when E' is assigned the weak topology cr(E', E);
8) 'H is equicontinuous, when E' and F' are assigned the strong topologies.
f) Show that if E is infra barrelled (III, p. 44, exerc. 7), then the properties ~) and 8) of e)
are equivalent, and are also equivalent to the following :
£) H is bounded in 2(E; F) for the topology of bounded convergence;
<p) 'H is simply bounded in 2(F'; E') when E' is assigned the strong topology.
g) Show that if E is quasi-complete, the properties ex), y) and 8) of e) are equivalent.

20) Let E be a Hausdorff locally convex space, and E' its dual.
a) Let M be a closed vector subspace of E. The topology ,(M, E' fMC) is identical with the
topology induced on M by ,(E, E') if and only if every convex balanced set in E'fMo which
is compact for the weak topology cr(E' fMc, M), is the canonical image of a convex balanced,
compact (for cr(E', E») subset of E' (cf V, p. 73, exerc. 15).
b) Let N be a dense vector subspace of E. If the topology induced on N by that of E is identical
with ,(N, E'), show that the topology of E is identical with ,(E, E').

21) a) Let E be a vector space, E* its algebraic dual. Show that the topology ,(E, E*) is
the finest locally convex topology and that the topology ,(E*, E) is identical with cr(E*, E).
b) Let E** be the algebraic dual ofE*. Show that ifE is infinite dimensional, then E is dense
in E** for all the topologies compatible with the duality between E** and E*, but that the
topology induced on E by ,(E**, E*) is distinct from ,(E, E*).

22) Let E be a Hausdorff locally convex space, E' its dual and M a closed subspace of E.
a) Show that if the closed convex envelope of a compact set in E is compact, then the topology
of compact convergence on E' /Mo (identified with the dual of M) is the quotient topology
of the topology of compact convergence on E' by MO.
b) Show that if M is infrabarrelled and if E'jMO, endowed with the topology ~(E'jMC, M)
is bomological, then the topology ~(E'/Mo, M) is the quotient of ~(E', E) by MO.

23) Give an example of a family (EJiEI of Hausdorff locally convex spaces such that the cano-
nical mapping from EEl E; onto the dual
iEI
of P = n Ei is not an isomorphism from the topo-
iEI
logical direct sum of the E; endowed with the weak topologies cr(E;, E) onto the dual P'
endowed with cr(P', Pl.

24) Let (E')'EA be a family of locally convex Hausdorff spaces, E a vector space and for every
ex E A, let fo. be a linear mapping from E, into E. On E consider the finest locally convex topo-
logy :Y for which the functions fo are continuous (II, p. 27); suppose ,,2/ is Hausdorff and
let E~ denote the dual of Eo, and E' that of E endowed with Y. Show that if, for every ex E A,
the topology of E. is identical with ,(E" E~), then the topology /2/ is identical with ,(E, E').

25) a) Show that every real (resp. complex) Banach space is isometric to a closed vector
subspace of a Banach space of the form '6'(S; R) (resp. '6'(S; e)l consisting of all real (resp.
complex) continuous functions defined on a compact space S (GT, X, § 4, No. l)(use formula (3)
of IV, p. 7).
TVS IV.52 DUALITY IN TOPOLOGICAL VECTOR SPACES §2

b) Deduce from a) that every Hausdorff locally convex space E is isomorphic to a subspace
ofa locally convex space of the form %,(L; R) (resp. 'fic(L; C)) (GT, X, § 1, No.6). In particular,
every Frechet space is isomorphic to a closed subspace of a space 'fiJL; R) (resp. 'fic(L; C)),
where L is locally compact and separable.

§ 2
1) a) Let E be a locally convex Hausdorff space, and E' its dual. Show that the following
properties are equivalent :
cz) E is infra-barrelled (Ill, p. 44, exerc. 7).
~) Every strongly bounded subset in E' is equicontinuous.
y) Every strongly bounded subset in E' is relatively weakly compact, and the initial topo-
logy of E is ,(E, E').
8) The topology induced on E by the strong topology of the bidual E" is identical with the
initial topology on E.
Then a fundamental system of neighbourhoods of 0 for the strong topology of E" consists
of the closures, for the topology cr(E", E'), of a fundamental system of neighbourhoods of 0
for the initial topology of E.
b) Prove that if E is infra-barrelled and if its dual E' is identical with its algebraic dual E*,
then the initial topology of E is the finest locally convex topology.

-If 2) a) Show that every product of infra-barrelled spaces is infra-barrelled. (Reduce to the
case of Hausdorff infra-barrelled spaces; then use exerc. I and prop. IS of IV, p. 14.)
b) Give an example of a Hausdorff and infra-barrelled locally convex space which is neither
bomological non barrelled. (Proceed as in Ill, p. 45, exerc. 16, replacing the barrelled spaces
by infra-barrelled spaces and use a).)

3) Let E be a complex Hausdorff locally convex space, Eo the underlying real locally convex
space to E, and let E' and E~ be the duals of E and Eo respectively. Show that the canonical
R-linear mapping fl--> f7lf from E' onto E~ is a homeomorphism for the 6-topologies on
E' and E~, where 6 is an arbitrary set of bounded subsets of E. From this deduce the defi-
nition of the canonical R-linear mapping from the bidual E" onto the bidual E~, which is a
homeomorphism for the weak topologies cr(E", E') and cr(E~, E~), as also for the strong
topologies ~(E", E') and ~(E~, E~); by this mapping, E (considered embedded in E") trans-
forms into Eo (considered embedded in E~).

4) Let E be a Hausdorff and infra-barrelled locally convex space.

a) Show that if the strong dual E~ of E is bomological, then the completion E of E, identified
with a vector subspace of E'* (III, p. 21, tho 2), is contained in the bidual E" of E.
b) We say that E is distinguished if every subset of E" which is bounded for the topology
cr(E", E) is contained in the closure (for this topology) of a bounded subset of E. Show that
for E to be distinguished it is necessary and sufficient that its strong dual E~ is barrelled (el IV,
p. 15, prop. 3).

5) Let E be a Hausdorff locally convex space, and E' its dual.

a) In order that the strong topology on E' be identical with ,(E', E), it is necessary and suf-
ficient that E is semi-reflexive.
b) Suppose that E is infra-barrelled. In order that the strong topology on E' be identical
with the topology of compact convergence, it is necessary and sufficient that E be a Montel
space.

-If 6) Let F and G be two vector spaces in separating duality.

a) Show that the following properties are equivalent:
cz) the space F with a topology compatible with the duality between F and G, is semi-
reflexive;
~) the space G, with ,(G, F), is barrelled.
b) Show that the following properties are equivalent:
cz) the space F, with ,(F, G), is reflexive;
§2 EXERCISES TVS IV.S3

~) the space G, with t(G, F), is reflexive;

y) F and G are barrelled for t(F, G) and t(G, F) respectively.
c) Show that every locally convex Hausdorff space E, with the topology teE, E') is isomorphic
to the quotient of a semi-reflexive space F by a closed subspace. (By III, p. 44, exerc. 14, c),
E' endowed with teE', E) is isomorphic to a closed vector subspace M of a Hausdorff barrelled
space G; take F = G' endowed with t(G', G) and use prop. 11 of IV, p. 10.}
d) Let A be an infinite set with Card(A) > ~ 1 (S, III, § 6, exerc. 10). In the product space

°
P = RA, let Eo denote the everywhere dense subspace consisting of all x = (XJ'EA such that
x, = except for a countable set of indices; the space Eo is barrelled and so is the subspace E
of P generated by Eo and the point lA in P all whose coordinates are equal to 1 (III, p. 45,
exerc. 16). Let B be a bounded subset in Eo, whose closure in P contains lA, and let 1 be a
subset of A with cardinality ~1 ; let prj denote the projection from Ponto R J; show that
there exists a subset BJ in B, with cardinality ,s; ~ l ' such that the closure of pr /B) in R J
contains I J; then the set l' :::0 1 of all indices C1 E A such that for at least one point of B J the
coordinate with index C1 is "# 0, has a cardinality equal to ~ l ' We define 10 = 1 and by induc-
tion 111+ 1 = 1~, and put H = U l n , whose cardinality is ~1' and let BH = U BJn; show that
n n
the closure of BH (and hence also that of B) contains the point (lH' 0) E RH X R A- H = P,
which does not belong to Eo. Conclude that for every bounded and compact set C in E, C n Eo
is again closed in E, and that E is not semi-reflexive.
e) Deduce from d) that the dual E' of E (which can be identified with the dual P' = R(A) of P)
is not complete for the topology teE', E), in spite of being semi-reflexive (hence quasi-complete)
for this topology (consider the linear form on E which is equal to 0 on Eo and equal to 1 at
the point lA, and use III, p. 21, tho 2).

7) Let (E)iEI be a family of Hausdorff locally convex spaces, P the product space of the E i,
and S their topological direct sum. Show that for P or S to be semi-reflexive (resp. reflexive),
it is necessary and sufficient that each Ei be semi-reflexive (resp. reflexive).

8) Let E be a Hausdorff locally convex space which is the strict inductive limit of an increasing
sequence (En) of closed vector subspaces (II, p. 32, prop. 9).
a) Show that, if the strong dual of each of the En is complete, then the strong dual of E is
complete (III, p. 20, tho I).
b) In order that E be semi-reflexive (resp. reflexive), it is necessary and sufficient that each
of the En be semi-reflexive (resp. reflexive).

9) Let (Ea)aEA be a family of Hausdorff locally convex spaces contained in the same vector
space, which is directed for the relation :::0, such that, if EJl c Eo, the topology of E~ is finer
than the topology on E~ induced by that of Ea' Let E be the intersection of the Eo, endowed
with a topology which IS the supremum of the topologies on E induced by those on the Ea'
Show that if each Eo is semi-reflexive, then E is semi-reflexive (consider an ultrafilter on a
bounded subset of E).

10) Show that every product, and every topological direct sum of Montel spaces is a Montel
space 1.

11) Let E be a Hausdorff locally convex space such that the strong dual E~ ofE is semi-reflexive.
a) Show that, on every strongly bounded subset of E', the topologies induced by cr(E', EO)
and cr(E', E) are identical.
b) Deduce from a) that E is infra-barrelled for the topology teE, E') (cf IV, p. 52, exerc. I)
and that, if E is its completion for this topology, and is identified with a subset of E'* (III,
p. 21, tho 2), then En C E. In particular, if E is quasi-complete for teE, E'), then E is reflexive
for this topology.

1 On the other hand, a closed subspace of a Montel space need not be infrabarrelled,
and the quotient of a Montel space by a closed subspace need not be semi-reflexive (IV, p. 63,
exerc. 8).
TVS IV.54 DUALITY IN TOPOLOGICAL VECTOR SPACES §2

12) Let E be a Banach spacc and E' its dual.

a) Show that the distance x ~ d(x, A) from a point x E E to a closed convex set A is a lower
semi-continuous function on E for the topology aCE, E').
b) Show that if E is reflexive, then for every closed convex subset A of E, there exists a point
Xo E A such that 'I Xo II is equal to the distance of 0 to A (use a). This point is unique if every
boundary point of the unit ball of E is extremal (II, p. 54, def. 1).
e) Suppose E is reflexive and let B be a closed convex and bounded subset in E: deduce
from a) and b) that there exist two points x E A, Y E B such that II x - y II = dCA, B) (el V,
p. 71, exerc. 8).

13) Let E be a Banach space, and M a closed vector subspace of E. Show that if M and ElM
are reflexive, then E is reflexive.

14) Let A be an infinite set.

a) Show that the strong dual of the Banach space E = ff(A) (IV, p. 47, exerc. I) can be iden-
tified with the Banach space 1'1 (A), and that the strong dual of the Banach space f'(A) can be
identified with the Banach space .~(A) = ( ' (A): deduce that E is not reflexive and that E"/E
is infinite dimensional (el TV, p. 71, exerc. 18). If A = N, then E and E' are Banach spaces
which satisfy the first axiom of countability, but E" does not (1, p. 25, exerc. I).
b) Let B" be the unit ball in E" = rCA), and let B~ be the convex set B" + (B" n E). Show
that B~ is a bounded closed convex set in E" for the strong topology, with a non-empty interior,
but does not have any extremal point. If p is the gauge of B~, show that E" endowed with the
norm p is not isometric to any dual of a Banach space and that B~ is not closed for the topo-
logy a(E", E') (although B n is compact for a(E", E') and B" n E is strongly closed).

-r 15) The notations are those of exerc. 14.

a) Let (x~) be a sequence in E' which converges to 0 for the topology a(E', E"); show that,
for every s > 0, there exists a finite subset H of A such that I Ix;,(a)I ,s; E for every integer 11.
"fH
(Argue by redueto ad absurdum : if thc property were not true, show that then there exists
a number 0 > 0, an increasing sequence (Ilk) of integers, an increasing sequence (H k ) of finite
subsets of A. such that I Ix;,(a)I ,s;"8o for n ? Ilk' I IX;,(a) I ,s;"80 for 11 ,s; Ilk and
IXEHk_ 1 a"Hk

Ix;,(alj ? ~; show that this implies a contradiction (the « gliding bump» method).)
Ildlk-Hk-l
Deduce that the sequence (x~) converges to 0 for the strong topology, although the latter is
strictly finer than the topology a(E', E").
b) Show that" if (x~) is a Cauchy sequence in E' for the topology a(E', E"), it converges to
a point in E' for this topology; in other words, E' is semi-complete (Ill, p. 7) for a(E', En).
(Show that, for every E > 0, there exists a finite subset H in A such that I IX~(Cl)1 ,s; E for
every integer n; argue by redueto ad absurdum, as in iI), and use a).)

'IT 16) With the notations of exerc. 14, let E" be the dual of E" = ex (A).
a) Let eo (for Cl E A) be the element of E" for which e.Wl = o,~ (Kronecker's symbol). Let
(Kn) be a sequence of finite subsets of A, two by two disjoint, and let (x;") be a sequence of
points in Em. Show that there exists (n k ), a strictly increasing infinite sequence of integers> 0
such that, if we put B = Y
K nk , then the elements y;, = (x~'(Cl))'EB belong to [1 (B). (Let 0 be
an arbitrary number> 0, and let (J".lmEN be a partition ofN in finite sets. Arguing by reduclo
ad absurdum, show that if x'" E E"', then there exists an integer m such that I<x", x"'>1 ,s; 0
for all x" E E" for which ilx"ll ,s; 1 and x"(a) = 0 except for indices Cl belonging to the set
UK". Apply this result successively to .Y';', x~', ... in a suitable way.)
nEJ m
b) Deduce from a) and from exerc. IS, u) that, if «) is a sequence converging to 0 in E'" for
the topology a(E"', E") and if x~' is the restriction of x;;' to the strongly closed subspace E
of E", then lim II x;;' II = 0 in E'.
1l-~ X',
§2 EXERCISES TVS IV.55

c) Deduce from b) that, the strongly closed subspace E of E" does not have a topological
complement for the strong topology. (Restricting to the case A = N, let (e;.) be the sequence
of continuous linear forms on E such that <x, e;,) = x(n) for all x E E; show that the sequence
(e;,) tends to 0 for cr(E', E), but that e;, cannot be extended to a continuous linear form x~'
on E" in such a way that the sequence (x;;') tends to 0 for cr(E"', E").)

17) Let E be a non-reflexive Banach space, E' its strong dual, E" the strong dual of E', E'"
the strong dual of E" and ElY the strong dual of E"'.
a) Show that, in E"', E' and the subspace EO orthogonal to E (when E is considered as a sub-
space of E") are topological complements, and that the projection from E'" onto E' for this
decomposition is a continuous linear mapping of norm 1. Comparing with exerc. 16, c),
deduce that the Banach space .x(A) is not isomorphic (as a topological vector space) to
the strong dual of any Banach space.
b) Show that ElY is the topological direct sum of E" and E", and also of E'C and E ; we have
E" n E = E. Let v be the linear mapping from ElY onto itself which is identity on E'o, and
OO

on E", is the projection from E" onto EO parallel to E"'; show that v is an isometry, but is
O

not continuous for the topology cr(EIV, E"').

18) Show that a Banach space E whose strong dual E' satisfies the first axiom of countability,
and which is semi-complete (III, p. 7) for the weakened topology cr(E, E'), is reflexive (compare
with IV, p. 54, exerc. 15, b».
19) Let E be a Banach space, E' its strong dual, G' a strongly closed subspace of E' satisfying
the first axiom of countability for the strong topology. Show that there exists a countable
subset of E such that, if F is the closed vector subspace of E generated by this subset, then
G' is isometric to a strongly closed subspace of the strong dual F' of F. (Suppose that the
sequence (x~) is strongly dense in G'; for every n, let x" E E such that Ilx,,11 ~ 1 and
<x"' x;,) = (1 - ~) I x;, I ; show that the strongly closed subspace Fof Egenerated by the
x" is the required space.)

lIT 20) Let E be a Banach space, E' its dual, and B the unit ball in E. In order that every point
of B have a countable fundamental system of neighbourhoods for the topology induced by
the weakened topology cr(E, E') on B, it is necessary and sufficient that E' satisfies the first
axiom of countability for the strong topology. (To show that the condition is necessary,
observe that if every point in B has a countable fundamental system of neighbourhoods for
the weakened topology, then this is also true for the closure BOO of B in E" for the topology
cr(E", E'). Therefore there exists a sequence (a;,) in E' such that every neighbourhood of 0
in sao for cr(E", E') contains the intersection of BOO and of a finite number of po lars {a;,} 0 ;

consider the strongly closed subspace W' of E' generated by the a;" and the orthogonal wm
of W' in E".)
lIT 21) Let E be a Banach space, E' its dual, B the unit ball in E and B; the closed ball with
centre 0 and radius r in E'.
a) Let M~, M~ be two vector subspaces of E', which are everywhere dense for the weak topo-
logy cr(E', E). In order that the topologies induced by cr(E, M~) and cr(E, M~) on B coincide,
it is necessary and sufficient that the strong closures of M~ and M~ in E' are identical.
b) Let M' be a vector subspace of E' which is everywhere dense for cr(E', E); let M'(l) denote
the vector subspac;e generated by the closure of M' n B~ in E' for the topology cr(E', E). In
order that M'(l) = E', it is necessary and sufficient that the weak closure of M' n B~ contains
a ball B; with r > 0 (use the fact that E' is barrelled for the strong topology).
c) Let r be the supremum of the numbers t such that the weak closure of M' n B~ contains
a ball B;; the number r is said to be the characteristic of M'. Show that r is the infimum of
the numbers sup I<x, x')l/llxll where x ranges over the set of all points oft 0 in E (use
x'EM'nB'l
the Hahn-Banach theorem).
d) Show that Ilr is the supremum of Ilxll as x ranges over the closure ofB in E for the topology
cr(E, M') (use c) and the Hahn-Banach theorem).
TVS IV.56 DUALITY IN TOPOLOGICAL VECTOR SPACES §2

e) Let M'G be the orthogonal of M' in E"; show that r = inf{llx + z"II/llxll) where z" ranges
over M'O and x ranges over the set of non-zero points in E (usc c) and the Hahn-Banach th).
Deduce that in order that M'(1) = E', it is necessary and sufficient that E + M'O be strongly
closed in E" (use Banach's th., L p. 17).
f) Let A = N x Nand E = Jf(A) (cf IV, p. 54, exerc. 14). In the space E" = f"'(A), let P
be the vector subspace consisting of all points x = (Xi) such that xij = xo)(j + 1) for all
i ;;:, O. Show that P = M" where M' is a vector subspace of E' that is everywhere dense (for
cr(E', E»), but that E + M' = E + P is not strongly closed in EN; deduce that the charac-
teristic of M' is O.

22) Let E be a Banach space, E' its dual and M' a strongly closed subspace in E' which is
everywhere dense for the weak topology on E'. We say that M is irreducible if there exists no
vector subspace N' oF M' of M' which is strongly closed and weakly everywhere dense in E'.
a) Show that M' is irreducible if and only if the orthogonal M of M' in E" is the topological
W

complement of E (for the strong topology of E"). Deduce that then M'(1) = E' (exerc. 21)
and that E is isomorphic to the strong dual of the space M' endowed with the topology induced
by the strong topology of E'.
b) Show that M' is irreducible if and only if the unit ball in E is relatively compact for the
topology cr(E, M'l (use exerc. 21, all.
c) For E to be isomorphic to a strong dual of a Banach space (for thc topological vector
space structure), it is necessary and sufficient that there exist an irreducible subspace in E'.
Deduce a new proof of the fact that the Banach space )feN) is not isomorphic to a strong
dual of a Banach space (c{ IV, p. 55, exerc. 16, c».

23) With the same notations as in exerc. 22, assume that M' is irreducible.
a) In order that the canonical mapping from E into E" /M w , which is the restriction of the
canonical mapping E" -> E"/M'o be a Banach space isometry, it is necessary and sufficient
that the characteristic (IV, p. 55, exerc. 21) of M' is equal to 1. Then, M' endowed with the
norm induced by that of E' is said to be the predual of E and E can be canonically identified
(with its norm) with the dual of the Banach space M'.
b) For every vector subspace F of E which is closed for cr(E, M'), show that the canonical
image of M' in the dual F' of F, identified with E'/Fo, is a predual of F.

24) a) Let (ak)1 ,;:k",,, be a finite sequence of points in a normed space E, and let (/"kll ';:k';:n
n k-l
be a finite sequence of numbers> 0 such that I Ak < 1; put ~lk = I - I I'i for every k;
k=l j~l

then

b) Let (anl be an infinite sequence of points in the unit ball of E, and let (An) be an infinite
11-1
sequence of numbers> 0 such that I An = 1. For every n > 0, put I!n = 1 - I Ak and
n-I
bn = I Aka k + I!n a,,; show that, for all n ;;:, 1, we have
k~1

(Apply a) by induction.)
c) Let (C n ) be a decreasing sequence of convex sets in E, contained in the unit ball, and suppose
that d(O, C I ) ;;:, e > 0 (hence, afortiori d(O, C n ) ;;:, e for all n). Let (An) be a sequence of num-
bers > 0 such that I An = 1. Show that there exists a number (l such that e ~ (l ~ 1 and a
§3 EXERCISES TVS IV.S7

sequence (xn) of points of E such that xn E en for all n, III Anxnll = rx and, for all n

I f
k=l
Akxkll :::; rx(1 - 8 I
j=n+l
Aj).

(Take Xl such that IlxI11 is arbitrarily close to d(O, e l ), then, by induction, take xn
11-1 CIj

such that I I AkXk + (I A) xnll is arbitrarily close to the infimum of the numbers
k=l j=n
11- 1 00

II I AkXk + (I A) yll where y ranges over en" Then use b).)

k= 1 j=n

-If 25) Let E be a Banach space satisfying the first axiom of countability. Show that the follow-
ing properties are equivalent :
rx) E is not reflexive.
~) For every number 6 such that 0 < 6 < I, there exists a sequence (x~) in E' such that,
Ilx~11 :::; I for all n, that the sequence (x~) converges to 0 for cr(E', E) and that the distance
of 0 from the convex set generated by the x;, is ;;: 6.
y) For every number 8 such that 0 < 8 < I and every sequence An of numbers> 0 such
that I An = I, there exists a number rx such that 6 :::; rx :::; I and a sequence (y~) of points of

E' such that IIY~ I :::; I for all n, II I

n
Any~ I = rx and I f
k=l
AkY~ I :::; rx(1 - 8 I
j=n+l
Aj) for all n.
8) There exists z' E E' such that for no X E E do we have I<x, z'>1 = Ilxll.llz'll (Theorem
of J ames-Klee).
(To see that rx) implies ~), observe that there exists z" E E" such that Ilz"ll < I and d(z", E) > 6.
If(xJ is an everywhere dense sequence in E, find the sequence (x;,) in E' such that Ilx~11 < I
and such that < x k ' x~ > = 0 for k :::; nand < x;" z" > = 8. To see that ~) implies y), use exerc. 24.
For y) implies 8), show that for all x E E we have II An < x, y~ > I < rx, with the notations
of y).)

26) A locally convex space E is said to have the property (GDF) if every linear mapping u
from E into a Banach space F which satisfies the following property, is continuous : in the
product space E x F, every limit of a convergent sequence of points in the graph r of u again
belongs to r. Every Frechet space has the property (GDF) (I, p. 19, cor. 5); this is also true of
every inductive limit of a family ofFrechet spaces (II, p. 34, prop. 10). Show that every Hausdorlf
locally convex space with the (GDF) property is barrelled. (Let V be a barrel in E, q its gauge
and H the Hausdorff space associated with E endowed with this semi-norm; show that the
canonical mapping TC from E into the completion fI is continuous by using the property
(GDF) and the fact that every linear form x' E VO can be extended uniquely to a continuous
linear form on fI, the set of these forms being the unit ball in the dual of fr.)

§ 3
-If 1) Let E be a locally convex metrizable space, and E~ its strong dual. If E~ is metrizable,
prove that the topology ofE can be defined by a single norm (use III, p. 37, exerc. 2 and p. 38,
exerc. 5 and also the fact that E is bomological).

-If 2) An infra-barrelled space is semi-barrelled. A locally convex space is said to be a (DF)

space if it is semi-barrelled and if the canonical bomology (III, p. 3, def 5) has a countable
base. Every normed space and every strict inductive limit of a sequence of normed spaces
(II, p. 33) is a (DF) space. Every strong dual of a Frechet space is a (DF) space.
a) The strong dual of a (DF) space is a Frechet space.
b) Let E be a (D F) space and let (An) be an increasing sequence of bounded, convex, balanced
and closed subsets of E such that every bounded subset of E is absorbed by one of the An'
TVS IV.58 DUALITY IN TOPOLOGICAL VECTOR SPACES §3

°
Let U be the union of the An; show that the closure U of U in E is precisely the set of all x E E
such that AX E U for :S; I, < 1. (If x ¢ AU for some A > I, then for every n, there exists a
linear form x;, E E' such that x;, E A~ and <x, x;, > = A, and the sequence (x;) is equicontinuous,
hence has a weak limit point.)
e) Show that if a (OF) space is barrelled, it is also bomological (cr Ill, p. 44, ex ere. 13, b».
Give examples of (OF) spaces which are not ultrabomological, but are bomological and
barrelled (III, p. 46, exerc. 22) and also of (OF) spaces which are not barrelled but are bomo-
logical.

'IT 3) Let E be a locally convex metrizable space, and E~ its strong dual.
a) Show that every convex balanced subset V' of E~ which absorbs the strongly bounded
subsets of E~ contains a barrel (for the strong topology) which absorbs the strongly bounded
subsets of E~. (Let (K~) be a countable base of the canonical bomology of E~ and let An be such
that A"K;, c iV'; apply exerc. 2, b) to the sequence A;" where A;, is the convex envelope of
the union of the AjKj for j :S; n.)
b) Deduce from a) that the following properties are equivalent
~) E is distinguished (IV, p. 52, exerc. 4).
~) E~ is infra barrelled (III, p. 44, exerc. 7).
y) E~ is bomological.
8) E~ is barrelled.
£) E~ is ultrabornological (Ill, p: 45, exerc. 19).
e) Show that ifE~ is reflexive, then E = E" (which is obviously reflexive) (ei IV, p. 52, ex ere. 4
and p. 53, exerc. II).

4) Let E be a locally convex Hausdorff space, E' its dual. If M is a closed vector subspace of E
which is metrizable and distinguished (IV, p. 52, exerc. 4), then the strong topology ~(E' IM c, M)
is the quotient topology by Me of the strong topology ~(E', E) (use ex ere. 3, b) and IV, p. 51,
ex ere. 22, b).

'IT 5) For every integer n > 0, let d n ) be the double sequence (a~~) (p E N, q E N) such that
a~~ = q if p :S; n and a~~ = I if p > n. Let E be the vector space of all double sequences
x = (xpq)(P.q)EN x N of real numbers such that, for every integer n > 0, the number rn(x) = I a~~ IXpql
p.q
is finite. If E is assigned the topology defined by the semi-norms rn' then E is a Frechet space
satisfying the first axiom of countability (IV, p. 47, ex ere. 1, e»); the dual E' ofE can be identified

°
with the space of all double sequences x' = (x~q) of real numbers such that for at least one
>
index n, there exists k n > such that IX~ql :S; k"a~~ for every pair (p, q); and <x, x' = I x pqX;q
P.q
(IV, p. 47, exerc. 1, e»).
° °
For every integer Po > and every sequence (mp) of integers> 0, let J(po; (m ») be the set
of pairs of integers p > 0, q > such that p ? Po and q ? mp; let "D be the filter base on
N x N consisting of the sets l(Po; (mp») and let 3' be an ultrafilter which is finer than the filter
with base "D.

°
a) Show that for all x' = (x~q) E E', the double sequence (x~q) has a limit u(x') with respect
to the ultrafilter 3'; if Vn is a neighbourhood of in E defineCl by r,,(x) :S; 1, then lu(x')1 :S; 1
for all x' E V~.
° °
b) Let U' be a neighbourhood of in E', for the strong topology, which is convex, balanced
and weakly closed, and for every n, let ~n > be such that ~n V~ c U'. For every integer

°
p > 0, let mp be an integer such that 2P+ 1 :S; r:1pmp' and let x' = (x~q) be the double sequence
with X~q = for q < m p' X~q = 2 for q ? mp' Show that x' E U' but that u(x') = 2; deduce
that u is not strongly continuous in E', while being bounded on every bounded subset of E'.
Conclude (IV, p. 58, exerc. 3) that E is not distinguished, and consequently that the strong
dual E~ is a non infra-barrelled (OF) space.
e) Using b) construct an example of a closed subspace M of a Frechet space F such that the
strong topology ~(F' IMo, M) is distinct from the quotient topology by Me of the strong topo-
logy ~(F', F) (embed E in a countable product of Banach spaces).
§3 EXERCISES TVS IV.59

~ 6) a) Let E be a (DF) space (IV, p. 57, exerc. 2), and let U be a convex set such that for
every bounded, convex and balanced subset A of E, UnA is a neighbourhood of 0 for the
topology induced on A by that of E. Show that U is a neighbourhood of 0 in A. (Let (An) be
a countable base for the canonical bomology of E (III, p. 3, def. 5). Show that, by induction
we can define a sequence (An) of numbers > 0 and a sequence (V n ) of closed, convex and balanced
X)

neighbourhoods of 0 in E such that AnAn c tu, A"A" c n V j , V" n An C U for every 11.
j~ 1
First show that if Aj and Vj have been constructed for j :( 11, then we can find A,,+ 1 such that
A,,+ 1 An+ 1 c t U and A,,+ 1 An+ 1 c Vj for j :( 11. Next prove that we can find V" + 1 such that
AjAj c V,,+ 1 for j :( 11 + I and Vn + 1 n A,,+ 1 c U; for this, letting A denote the convex
envelopeoftheAAJorj:( 11 + l,showthatwecantakeV,,+l = A + Vfor a suitable convex
balanced neighbourhood V of 0; we remark that for this it is enough to show that, if
B = An + 1 n CU, then 0 is not in the closure of the set B + 2A.)
b) Deduce from a) that if u is a linear mapping from E into a locally convex space F, such that
the restriction of u to every bounded subset of E is continuous, then u is continuous (cf IV,
p. 50, exerc. 15).

~ 7) a) Let E be a (DF) space, U a convex, balanced and closed set in E, which absorbs the
bounded subsets of E, and let (x n ) be a sequence of points of CU. Show that there exists a
neighbourhood V of 0 in E which does not contain any of the x"' (Let (An) be a countable
base for the canonical bomology of E. Show that, by induction we can define a sequence
(A,,) of numbers> 0 and a sequence (V,,) of convex, balanced and closed neighbourhoods of
o such that A"A" c n
j~ 1
Vj' A"A" c U and x" E CV" for all 11. For this, if the Aj and Vj have

been constructed for j :( 11, take An+l such that A,,+lAn+l c U and A,,+lA"+1 c Vj for all
j :( 11, then take V" + 1 containing the closure of the convex envelope of the union of the AjA j
for j :( 11 + 1.)
b) Deduce from a) that if M is a subset of E containing an everywhere dense countable set,
then the topology induced on M by the strong topology of the bidual E" of E is identical with
the topology induced by that of E. In particular, the convergent sequences in E are the same
for the topology of E and for the topology induced by the strong topology of E"; for every
metrizable subset M of E, the topology induced on M by the topology of E is identical with
the topology induced by the strong topology of E".
c) Deduce from a) that if there exists a countable everywhere dense set in E, then E is infra-
barrelled.
d) Deduce from b) and from exerc. 6 that if every bounded subset of E is metrizable for the
topology induced by that of E, then E is infra barrelled.

~ 8) Let E be a Frechet space, E~ its strong dual. Suppose that there exists an everywhere
dense sequence (x;,) in E~. Show that E satisfies the first axiom of countability. (Let (K~) be a
countable base for the canonical bomology of E~ consisting of closed convex balanced sets.
For every system (1 consisting of a point x~, an arbitrary finite number of rational numbers
m
Ak > 0 (1 :( k :( m) and m indices 11k such that x;, ric 2 I AkK~k = 2H~, let x, E E be such that
k~l

<
the hyperplane with equation x" y') = I strictly separates the two weakly compact sets
H~ and x~ + H~. Show that for every x' =I- 0 in E', there exists a system (1 such that <x" x') =I- O.
For this, consider a neighbourhood V' of 0 in E~ such that V' n (x' + V') = 0, then for each
integer m, take a rational number Am > 0 such that AmK~ c V'; use the fact that the union
U' c V' of the AmK~ is a neighbourhood of 0 (exerc. 7, and IV, p. 58, exerc. 3, b)) and drat
there exists 11 such that x;, EX' + U'.)
~ 9) a) Let E be a HausdortT semi-barrelled space, M a closed vector subspace of E and E'
the dual of E. Show that ElM is semi-barrelled and that the strong topology ~(MO, ElM)
is identical with the topology induced on MO by the strong topology ~(E', E). (Note that
it is enough to prove that a sequence (x~) in MO which converges to 0 for ~(E', E) is bounded
for ~(MO, ElM).) Deduce that if E is a (DF) space, then so is ElM (cf IV, p. 63, exerc. 8).
TVS IV.60 DUALITY IN TOPOLOGICAL VECTOR SPACES §3

b) Let E be a locally convex Hausdorff space, M a (not-necessarily closed) vector subspace

ofE. Show that if M is a semi-barrelled space, then the strong topology ~(E' IMo, M) is identical
with the quotient topology by MO of the strong topology ~(E', E). (Argue as in a).)
c) Show that a Hausdorff and quasi-complete semi-barrelled space E is complete (use b)
applied to E and E). In particular, a semi-barrelled, semi-reflexive space is complete (ef IV,
p. 52, exerc. 6).
d) Show that the completion of a semi-barrelled (resp. (DF») Hausdorff space is semi-bar-
relled (resp. a (D F) space).
e) Let (E,,) be a sequence of semi-barrelled (resp. (DF») spaces, E a vector space, and for each n,
let J;, be a linear mapping from E" into E. Suppose that E is the union of the j~(E,,) ; show that,
E is semi-barrelled (resp. a (DF) space) for the finest locally convex topology for which all
the In are continuous (first examine the case where E is the topological direct sum of the E,,).
If the E" are semi-reflexive (resp. reflexive) and if E is Hausdorff, then E is semi-reflexive (resp.
reflexive).

10) Let E be a Frechet space satisfying the first axiom of countability. Show that if in the dual
E' of E, every sequence which converges for the weak topology cr(E', E) also converges for the
strong topology ~(E', E), then E is a Montel space. (Show that every bounded subset of E'
is relatively compact for the strong topology; use GT, II, § 4, exerc. 6; then use IV, p. 53,
ex ere. 11, b).)

~ II) Let (e mn ) be a double sequence of numbers> 0 such that Cm n :( Cm+ 1" and let E be
the space of all sequences x = (x.,) of real numbers such that Pm(x) ~ I Cm" Ix,,1 < + CIJ for

every integer m. We endow E with the topology defined by the semi-norms Pm and for this
topology E is a Frechet space satisfying the first axiom of countability; the dual E' of E can
be identified with the space of all sequences x' = (x;) such that sup c,~/ Ix;,1 < + CIJ for at
n

least one m, the canonical bilinear form <x, x') being identified with I XliX;, (IV, p. 47,
ex ere. I, c». Suppose that there does not exist any subsequence (n k ) for which there is a sequence
(am) of numbers? 0 and an index mo such that Cm'''k :( amCmO.llk for all m ? mo and for all k.
Under these conditions, every weakly convergent sequence in E' is strongly convergent and
consequently (IV, p. 60, exerc. 10) E is a Montel space. (Argue by reductio ad absurdum; if
necessary make a transformation of the form (xn) H (anxn) to reduce to the case where Cma " = I
for all n and some mo, and where there exists a sequence (x,(p)p" 0 which converges weakly
to 0 in E', and is such that Ix;,(p)1 :( I for every pair (p, n), and also that there exists a bounded
set B in E, defined by Pm(x) :( bm for all m ? 0 and such that sup 1< X, x'(p) 1 ? 20 > 0 for
XEB
every integer p. Under these hypothesis, prove that there exists a strictly increasing sequence
(rq) of integers, and a sequence (x(q) of points of B such that

for each index q. Then show, by reductio ad absurdum, that for every q, there exists at least one
index Sq such that l'q < Sq :( rq + 1 and that for every integer m, we have em,s :( bm2m+ 21o,
which contradicts the hypothesis.) q

12) a) Let F be a H!;lusdorff (DF) space, and F~ its strong dual. Show that if F~ is reflexive,
then the completion F of F is reflexive and is equal to the bidual F" of F (ef IV, p, 52, exerc. 4
and p. 53, exerc, II),
b) Let E be a Frechet space. Show that if the bidual E" of E is reflexive, then E is reflexive,

13) a) Let E, F be two Frechet spaces, G a locally convex Hausdorff space and E', F', G' the
duals of E, F, G respectively. Let u be a bilinear mapping from E' x F' into G', which is sepa-
rately continuous (III, p, 28) when E', F', G' are assigned the weak topologies cr(E', E), cr(F', F)
and cr(G', G), Show that under these conditions, u is a continuous mapping from E' x F'
into G' when E', F' and G' are assigned the strong topologies. (For Z E G, put
§3 EXERCISES TVS IV.61

<z, u(x', y') > = <vz(x'), y' >where vz(x')

EO F. First show that if E' is assigned the strong topo-
logy and F the initial topology, then the set of all vZ ' as z ranges over a bounded set C in G,
is equicontinuous; for this use IV, p. 51, exerc. 19, d). Next show that there exists a neigh-
bourhood V' of 0 for the strong topology of E' such that the union of the sets vz(V') as z ranges
over C is bounded in F; for this use III, p. 47, ex ere. 5.)
b) Give an example to show that the conclusion of a) does not hold if we assume that E is
a Frechet space and F a strict inductive limit of Frechet spaces (III, p. 47, exerc. 3).
14) a) Let E be a Frechet space, E' its dual. Show that E', endowed with the topology of
compact convergence or with a finer 6-topology, is complete (cf III, p. 22, Remark 1). If E
is not reflexive, show that E' is not infra barrelled for any 6-topology which is finer than the
topology of compact convergence and coarser than 't(E', E).
b) Let (E')'EA be a family of Frechet spaces, E a vector space and for every r:x EO A, let h, be a
linear mapping from E, into E. Suppose that E, endowed with the finest locally convex topology
for which the h, are continuous (II, p. 27) is Hausdorff. Prove that the dual E' of E, endowed
with the topology of compact convergence or with any finer 6-topology, is complete (cf Ill,
p. 20, tho 1).
15) Let E be an infinite dimensional Banach space, (a,,),,,,! a countable total free family of
points of E, and let FII be the n-dimensional subspace of E generated by the all/ for m :( n.
Let Sn be the sphere with equation Ilxll = n in E; in Sn n FII let All be a finite set such that every
n
ex)

point of Sn n F" is at a distance :( lin from All' Prove that A = An is such that its inter-
n=l
section with every closed bounded set is closed, but that 0 is a limit point of A for the weakened
topology.

16) Let E be an inductive limit space of a sequence (E ) of locally convex metrizable spaces,
the canonical mappings E --> E being injective and let Ii be the union of the images of the E .
Show that the strong dua(E~ of E is exhaustible (III, p. 49, exerc. 1). (Let (Uf) be a decreasin"g
fundamental system of closed convex and balanced neighbourhoods of 0 in E ; consider the
finite intersections of the polar sets (Uf)O in E'; use the fact that for every incre~sing sequence
n
~

Cm)},,! the intersection (U!.,Y is the polar of a neighbourhood of 0 in E, and that E~ is

) ~ !
complete.)
-If 17) a) Let E be a locally convex Hausdorff space, such that the bomology consisting
of the relatively compact sets of E has a countable base (All) (Ill, p. 1). Show that (An) is also
a base for the canonical bomology (III, p. 3, def. 5). (Let C n be the relatively compact set which
is the sum of n sets each equal to An; then (C II ) is also a base for the bomology of relatively
compact sets. Argue by reductio ad absurdum, considering a bounded set B which is not
contained in any of the CII , and conclude that there exists a sequence (x n ) of points of B such
that xnln ~ An; deduce a contradiction.) Then, the space E is semi-reflexive and the closed
convex balanced envelope of every compact set in E is compact.
b) Suppose that E is infrabarrelled 'lnd that for a topology iT compatible with the duality
between E and E', there exists a countable base for the bomology of relatively compact subsets
of E for :Y. Show that then E is a reflexive (DF) space (IV, p. 57, exerc. 2). (Use a), observing
that the closed bounded sets in E are compact for ,cT, and consequently, complete for the
initial topology of E; this implies that E is barrelled). If:Y is the initial topology, then E is a
Montel space.
18) a) Let E be a locally convex Hausdorff space, and let (All) be an increasing sequence of
convex, balanced, compact sets for the weakened topology crCE, E'), such that the union of the
sets All is E, and that for every integer n and every Ie > 0, there exists m such that leAn C AII/'
Show that (An) is a base for the bomology consisting of convex and relatively compact sets
for cr(E, E'). (Observe that on E' the topology of uniform convergence on the An is 't(E', E).)
b) If E is barrelled and if there exists a sequence (An) having the properties mentioned in a),
then E is a reflexive (DF) space. (Observe that E', endowed with 't(E', E) is metrizable and that
the topology of E is ~(E, E').)
TVS IV.62 DUALITY IN TOPOLOGICAL VECTOR SPACES §4

§ 4

1) Let E and F be two locally convex Hausdorlf spaces, E' and F' their respective duals. Show
that if, for every vector subspace N of F, the topology induced on N by T(F, F') is identical
with ,(N, F'/N°) (IV, p. 51, exerc. 20), then every strict morphism from E into F for the topo-
logies cr(E, E') and cr(F, F') is also a strict morphism for the topologies ,(E, E') and T(F, F')
(ef IV, p. 10, prop. 11). Consider the case when F is metrizable.

2) Let E be a locally convex Hausdorlf space such that in the dual E' there exists an infinite
dimensional convex set B' which is compact for cr(E', E) (this condition is realized for example,
when E is an infinite dimensional vector space, endowed with cr(E, E*)). Show that there exists
a linear form u E E'* which is unbounded on B'; deduce that B' is not compact for the topology
cr(E', F), where F is the subspace E + Ru of E'*. Conclude from this that the canonical injec-
tion from E into F is a strict morphism for the topologies cr(E, E') and cr(F, E'), but not for
the topologies ,(E, E') and T(F, E').

3) Give an example of a strict injective morphism u from a Frechet space E into a Frechet
space F such that 'u is not a strict morphism from F~ into E~ (ef IV, p. 58, exerc. 5, e)).

4) Let E and F be two locally convex Hausdorlf spaces, u a continuous linear mapping from
E into F, and M an everywhere dense vector subspace of E. Show that if the restriction of u
to M is a strict morphism from Minto F, then u is a strict morphism from E into F (use prop. 2
of IV, p. 27). If in addition u(M) = F, show that for every open, convex and balanced neigh-
bourhood V of 0 in E, u(V) is in the interior of u(V n M).

5) Let E and F be two normed spaces, u a continuous linear mapping from E into F.
a) Show that if u is a strict morphism from E into F, then 'u is a strict morphism from the
strong dual F~ into the strong dual E~.
b) Suppose E is complete; show that if'u is a strict morphism from F~ into E~, then u is a strict
morphism from E into F, and 'u is a strict morphism from F' into E' for the weak topologies
cr(F', F) and cr(E', E) (consider F as a subspace of its completion).
c) Give an example where E is not complete, F is complete, 'u is a strict injective morphism
from F' into E' for the strong topologies and for the weak topologies, but u is not a strict
morphism from E into F (c:f II, p. 74, exerc. 5).
d) Give an example where E is not complete, F is complete, u is a strict injective morphism
from E into F, 'u is a strict morphism from F' into E' for the strong topologies, but not for
the weak topologies (take E to be everywhere dense in F).
e) If F is complete and if 'u is a strict morphism from F' into E' for the weak topologies, then
'u is a strict morphism from F' into E' for the strong topologies (extend u to E).
f) Give an example where E is complete, F is not complete, 'u is a strict morphism from F'
into E' for the weak topologies but not for the strong topologies (ef II, p. 74, exerc. 5).

6) Let E be the locally convex metrizable space fl (N) (I, p. 4) endowed with the topology
induced by that of the product space RN ; its dual E' can be identified with R(N) and the topology
T(E', E) is the topology induced on E' by the norm topology of eo(N) (IV, p. 47, exerc. 1).
Show that if u is a surjective continuous linear mapping from E onto a Hausdorlf barrelled
space F, then F is necessarily finite dimensional. (Observe that 'u is an isomorphism from F',
endowed with cr(F', F), onto a subspace of E', endowed with cr(E', E) ; from the fact that F is
barrelled, conclude that 'u(F') endowed with the topology induced by T(E', E), is a Banach
space, and use exerc. 24 of II, p. 80.)

7) a) Let E be a Banach space, and (XaLA an everywhere dense set in the unit sphere of E.
Let u be the linear mapping from the space fl(A)(I, p. 4) into E defined by u(t) =I t(CX)Xa
"EA
for all t = (t(CX))"EA belonging to fleA). Show that u is a strict morphism from fleA) onto E
and consequently that E is isomorphic to a quotient space of fleA).
§4 EXERCISES TVS IV.63

b) From a) deduce an example of a closed subspace of Cl(N) which has no topological com-
plement in [l(N) (for E take co(N) and use IV, p. 54, exerc. 15, b) and p. 55, exerc. 18).

~ 8) For every integer n > 0, let d n ) be the double sequence d n ) = (a~~»i" l.j" 1 defined by
a~'J) j" for every pair (i, j) such that i < n and a~:;> = in for every pair (i, j) such that i ;" n.
=
Let E be the vector space of double sequences x = (Xi) of real numbers such that, for every
integer n > 0, we have Pn(x) = I a~'J) Ix i) < + CfJ; the semi-norms Pn define a topology of
i,j
a Frechet space and of a Montel space on E (IV, p. 60, exerc. 11); the dual E' of E, which is
a (DF) space and a Montel space (hence ultrabornological and reflexive) is identical with the
space of all sequences x' = (x;) such that, for at least one index n, we have the relation
sup la~'J)I- 1 IX;jl < + CfJ (IV, p. 47, exerc. 1, c».
I.)

a) For every x = (Xi) E E, let Yj = I Xij (for all j ;" 1). Show that I IYjl < + CfJ; let u(x)
i j
denote the sequence (y) E Cl(N); show that u is a continuous linear mapping from E into
F = [l(N), and that for every sequence y' = (y) E F = fOC(N), 'u(y') is the sequence (z;) E E'
for which z;. = y'. for every index i. Deduce that 'u is an injective linear mapping from pe''(N)
onto a subspace 6fE' which is closed for cr(E', E), and consequently that u is a strict morphism
from E onto fl(N) for the initial topologies, and 'u is an isomorphism from F = fuc(N) onto
'u(F) for the weak topologies cr(F, F) and cr(E', E).
b) Let M = u-l(O); then MO = 'u(F). Show that the inverse image under 'u of the topology
induced on Me by the strong topology of E', is the topology of uniform convergence on com-
pact subsets ofF (IV, p. 28, tho 1 and p. 54, exerc. 15). Deduce that on MO, the topology induced
by the strong topology ~(E', E) is not the strong topology ~(MO, ElM), and that for the topo-
logy induced by ~(E', E), MC is not an infra-barrelled space, in spite of being a closed subspace
of an ultrabornological Montel space; on the other hand, ElM, which is the quotient of a
Frechet and a Montel space by a closed subspace, is not reflexive. Show that in ElM there
exist bounded sets which are not the canonical images of bounded subsets of E.

-r 9) Let A be a countable set. Consider three pairs of vector spaces (P, P'), (Q, Q'), (E, E'),
each of the six spaces being vector subspaces of RA and containing the direct sum subspace
R(A); in addition, suppose that for every point x E P (resp. x E Q, X E E) and every point
<
x' E P' (resp. x' E Q', x' E E'), the family (x"X~)oEA is summable and put x, x') = I xox~;
OEA

this bilinear form puts P and P' (resp. Q and Q', E and E') in separating duality.
a) Suppose that E = P n Q, E' =0 P' + Q' and E' =1= P' + Q'. Consider the linear mapping
u : x f-> (x, x) from E into F = P x Q, which is put into separating duality with F' = P' x Q'.
Show that u is continuous for the weak topologies cr(E, E') and cr(F, F) and that its image
M = u(E) is a closed subspace for cr(F, F); deduce that 'u is a strict morphism from F into
E' for the topologies cr(F, F) and cr(E', E), and that N = 'u(F) is not closed in E' for cr(E', E).
If E' is metrizable for the topology ,(E', E), deduce that 'u is also a strict morphism from F
into E' for the topologies ,(F, F) and ,(E', E).
b) In addition, suppose that E' is a Frechet space for ,(E', E). Then F IMc endowed with the
quotient topology of ,(F, F) by MO, is not semi-complete, and there exist bounded sets in
F IM c which are not relatively compact for ,(F IMo, M).
c) Under the same hypotheses as in b), let x' be an element of E' not belonging to N = 'u(F');
<
for every Y EM, let v(y) = x, x'), where x E E is the unique element such that u(x) = y.
Show that the linear form v on M is not continuous for the topology cr(F, F), but that its
restriction to every bounded subset of M is continuous for cr(M, F IMO). Deduce that L = v - 1(0)
is a vector subspace of F, whose intersection with every bounded and closed subset of F (for
cr(F, F») is closed for cr(F, F), but which is not closed in F for cr(F, F).

~ 10) a) Let G, H be two reflexive Banach spaces such that R(N) c G c H c RN (* for
example G = e'(N) and H = fP(N), with 1 < r < P < + CfJ *). Taking A = N x N, P = H(N)
(topological direct sum), P' = H'N, Q = G N, Q' = G,(Nl, E = G(Nl, E' = G'N, show that the
conditions of exerc. 9, a) are satisfied; that E' is a reflexive Frechet space, that E, F, Fare
the strict inductive limits of reflexive Frechet spaces (hence complete and reflexive).
TVS IV.64 DUALITY IN TOPOLOGICAL VECTOR SPACES §4

b) From a) and exerc. 9, construct examples with the following properties:

ex) A quotient space of a strict inductive limit of reflexive Frechet spaces which is neither
quasi-complete nor semi-reflexive.
~) A closed subspace of a strict inductive limit of reflexive Frechet spaces which is not
reflexive, and whose dual is not complete.
y) A vector subspace of a strict inductive limit of reflexive Frechet spaces En' which is
not closed (hence not barrelled) but whose intersection with each of the subspaces En is closed.
8) A non-closed subspace of the dual of a strict inductive limit of reflexive Frechet spaces,
whose intersection with every weakly compact subset is weakly compact (el IV, p. 25, cor. 2).

~ 11) a) Let E be a strict inductive limit of Frechet spaces En satisfying the first axiom of
countability, and F an everywhere dense subspace of E. For every n, put Fn = F n En; show
that E is the strict inductive limit of the Fn (el III, p. 44, exerc. 13, b)). Deduce that F is bomo-
logical. (Let u be a linear mapping from F into a Banach space L which transforms every
bounded set into a bounded set. Observe that the restriction Un of u to F n En is continuous
and can be extended to a continuous linear mapping un from F n En into L, the Un being restric-
tions of the same linear mapping v from E into L, and conclude that v is continuous.)
b) Suppose F is not closed in E but that F n En is closed in En for all n (exerc. 10, b)). Let An
be a countable dense set in En' and let G be the vector subspace of E generated by the union
of F and the An' Show that G is a bomological space in which F is a subspace with a comple-
ment which has a countable basis, but that F is not infra-barrelled (cl III, p. 41, § 2, exerc. 4
and p. 45, exerc. 17).

12) Let E, F be two Frechet spaces, u a strict morphism from E into F. Show that for every
finite rank continuous linear mapping v from E into F, u + U is a strict morphism from E
into F.

IIf 13) a) Let E be a Hausdorlf and complete locally convex space. Suppose that there exists
(Fn) a decreasing sequence of closed vector subspaces in E such that for every neighbourhood
V of 0 in E, there exists n such that Fn c: V. Show that the space E is of minimal type (II, p. 85,
exerc. 13).
b) Let E be a Frechet space satisfying the first axiom of countability, and which is not of
minimal type. Show that there exist two closed vector subspaces M, N in E such that
M n N = {O} and such that M + N is not closed. (Let (x~) denote a sequence of linearly
independent continuous linear forms on E, forming a total set for cr(E', E) (III, p. 19, cor. 2);
let Ln be the subspace of E orthogonal to the x; for indices i < 2n; let xn' Yn be two linearly
independent vectors in the complement of L" + 1 with respect to L". Let d be a translation
invariant distance defining the topology of E. Using a), show that there exists a number ex > 0
such that we can take d(O, xn) ~ ex, d(O, Y,,) ~ ex and d(x", Y,,) < 1jn. Show that ifM (resp. N)
is the closed vector subspace generated by the x" (resp. Yn)' then M and N have the properties
required; use I, p. 19, cor. 4.)
c) Let E be a closed subspace of a product Il Fo of normed spaces; show that if every closed
OEA

subspace of E generated by a countable family of points is of minimal type, then E itself is

of minimal type. (Argue by reductio ad absurdum, assuming that the projection of E on each
Fo is equal to Fa' and that for some ex, Fo is infinite dimensional.)
d) Let E be a Hausdorlf and complete locally convex space, in which every subspace gene-
rated by a countable family of points is metrizable. In order that E be of minimal type, it is
necessary and sufficient that for every pair of closed vector subspaces M, N of E such that
M n N = {O}, M + N is closed in E (use b) and c)).
14) a) Let E be a Frechet space on which there exists no continuous norm. Show that
there exists a closed vector subspace of minimal type in E (II, p. 85, exerc. 13), which is infinite
dimensional, and consequently, has a topological complement in E. (There exists a strictly
decreasing fundamental sequence (V,,) of convex, balanced neighbourhoods of 0 in E, and
a sequence (x,,) of points of E such that xn ¢ V" + 1 and such that the line passing through xn
is contained in V"' the xn being linearly independent; show that the required space is the closed
vector subspace generated by the x n .)
§4 EXERCISES TVS IV.65

b) Let E be a Frechet space whose topology cannot be defined by a single norm, but by an
increasing sequence (Pn) of norms. Let Vn be the neighbourhood defined by p,,(x) :0( 1, and
let A~ = V~ be its polar in E' ; we can assume that A~ + 1 is not contained in the vector subspace
of E' generated by A~ (IV, p. 49, exerc. 12, c». Let (x;,) be a sequence of points of E' such that
x~ E A~ and that x~ does not belong to the vector subspace generated by A~ _ 1 ; show that the
vector subspace M' generated by the sequence (x~) is weakly closed in E' (cf IV, p. 25, cor. 2)
and does not have a topological complement in E' for the topology cr(E', E). (Observe that
if there existed a weakly continuous projector u' from E' onto M', then u'(A'I) would be con-
tained in one of the sets M' n A~ by Baire's theorem, and derive a contradiction, since A~
is weakly total in E'.) Deduce that the subspace M of E does not have a topological comple-
W

ment in E.
c) Let E be a Frechet space whose topology cannot be defined by a single norm. Show that
if E is not isomorphic to a product of Banach spaces, then there exists a closed vector subspace
in E which has no topological complement. (Argue by reductio ad absurdum; let (Pn)n~ I be
an increasing sequence of semi-norms on E, defining the topology of E; let Fn = p,~ 1(0)
and let Edl be a topological complement ofF n+ 1 with respect to Fn (we put E = Fo); using
b), show that En+ 1 is a Banach space and that E is isomorphic to the product of the En (n ~ 1);
for this use I, p. 17, tho 1.)

IS) a) Lct E be an infinite dimensional normed space (real or complex). Show that there
exists a sequence (xn) in E such that, for every bounded sequence (An) of scalars, there exists
a continuous linear form x' on E such that <Xn , x') = An for all n. (Construct a sequence
<
(x~) of points in the dual E' of E and a sequence (xn) of points in E such that Xi' xj) = 8ij
and Ilx~11 :0( 2-n.)
b) For a locally convex metrizable space E to have the property stated in a), it is necessary
and sufficient that the completion of E is not a spacc of minimal type (II, p. 85, exerc. 13).

16) a) Let E and F be two infinite dimensional complex normed spaces, and u a bijective
semilinear mapping from E onto F, with respect to an automorphism cr of C, which trans-
forms every closed hyperplane of E into a closed hyperplane of F. Show that the automor-
phism cr of e is_necessarily continuous (and consequently, is either the identity or the auto-
morphism ~ f-4 ~). (Argue by reductio ad absurdum: let (x n ) be a sequence of points in E satis-
fying the condition of exerc. IS, a) and let (An) be a bounded sequence of complex numbers
such that IA~I ~ n.llu(xn)11 for all n; if x' E E' is such that <xn, x') = An for all n, consider
<
the image under u of the closed hyperplane of all x E E such that x, x') = I and derive
a contradiction.)
h) Deduce from a) that u is a continuous semi-linear mapping from E into F (IV, p. 7, corollary).
c) Let cr be a discontinuous automorphism of the field C. Show that the bijection (~n) f-4 (~~)
from eN onto itself transforms every closed hyperplane into a closed hyperplane (cf IV, p. 14,
prop. 15).

17) Let E, F be two Banach spaces, u a strict morphism from E onto F. Then there exists
a number m > 0 such that for every E E) 0, m ( and for every y E F, there exists z E E such
that u(z) = yand Ilzll :0( (m - E)-1 Ilyli.
a) Let B be a closed ball Ilx - all :0( r in E, and let co be a mapping from B into F satisfying
the following conditions : 10 sup II w(x) I = M < + CfJ; 20 there exists a number k > 0
XEB

such that for all x, x' in B, we have Ilw(x) - w(x') II :0( k Ilx - x'il (<< Lipschitz condition »).
Show that if k < m and M < rem - k), then the image of B under the mapping xf-4u(x) + w(x)
contains a ball with center b = u(a). (Show that for every y E F close enough to h, we can define
a sequence (x',)n~O of points of B such that u(x o) = y and u(x,.) = y - w(xn_ l ) for n ~ 1,
and such that (x n ) converges to a point of B.)
b) Suppose that w is a mapping from all E into F, such that II w(x) - w(x') II :0( k I x - x'il
for every x, x' in E. Show that if k < m, u + w is a surjective and open mapping from E onto F.
c) In particular, if w is a continuous linear mapping from E into F such that I wll < m, then
v = u + w is a strict morphism from E onto F. If E E) 0, m(, then for every x E v - I (0) with
Ilxll = I, there exists Xo E u- 1(0) such that Ilx- xoll :0( I wll/(m - E); if in addition II wll < m- E,
TVS IV.66 DUALITY IN TOPOLOGICAL VECTOR SPACES §4

then for every Xo EO u- 1(0) with Ilxo II = 1, there exists x EO v - 1(0) such that

Ilx - xoll :( Ilwll/(m - € - Ilwll)·

Deduce that if there exists a closed vector subspace G ofE which is the complement of u- 1(0),
then G is also the complement of v - 1(0) whenever II w II is small enough (argue by reductio ad
absurdum: show that the projection of v- 1 (0) onto u- 1 (0) cannot be contained in a closed
hyperplane of u- 1 (0); on the other hand, note that there exists a > 0 such that every point
x EO G such that Ilxll = 1 is at a distance:? a from u- 1 (0»).

18) a) Let E, F be two normed spaces, u a strict injective morphism from E into F; then there
is a number m > 0 such that Ilu(x)11 :? m Ilxll for all x EO E. Show that if w is a continuous
linear mapping from E into F such that II wll < m, then v = u + w is a strict injective mor-
phism from E into F. Moreover, for all Yo EO u(E) such that IIYol1 = 1, there exists yEO veE)
such that IIY - Yoll :( Ilwll/m; for every Y EO veE) with Ilyll = 1, there exists Yo EO u(E) such
that Ily - Yoll :( Ilwll/(m - Ilwll)·
b) Deduce from a) that, if in addition, E and F are Banach spaces, and if there exists a closed
vector subspace G of F which is the complement of the closed subspace u(E), then G is also
the complement of veE) whenever II wll is small enough (argue as in exerc. 17, c».

19) Let E be the subspace of the Banach space '6'(( - 1, I); R) of all continuous mappings
from ( - 1, 1) into R, consisting of the polynomials. Similarly, let F be the subspace of
(6'(( 0, I); R) consisting of all polynomials. Let u be the mapping which associates to each
polynomial f EO E, the polynomial t f-4 ~f(Jt) + f( - Jt)) in F. Also, let w be the mapping
from E into F which associates with every polynomial f EO E its restriction to (0, 1). Show
that u is a strict morphism from E onto F, but that for every E > 0, u + €w is not a strict
morphism from E into F.

20) a) Let E, F be two Banach spaces, u a strict morphism from" E. into F, such that u- 1 (0)
is finite dimensional. Show that for every continuous linear mapping w from E into F with
small enough norm, v = u + w is a strict morphism from E into F and dim v- 1 (0) :( dim u- 1 (0).
(Write E as the topological direct sum of u- 1 (0) and a closed subspace and use exerc. 18 of
IV, p. 66.)
b) Let E, F be two Banach spaces, u a continuous linear mapping from E into F such that
u(E) has finite codimension in F. Then u(E) is closed and u is a strict morphism (I, p. 28, exerc. 4).
Show that for every continuous linear mapping w from E into F, of small enough norm, v = u + w
is a strict morphism from E into F and
codim(v(E») :( codim(u(E»)
(consider tv = tu + tw, and use a) and IV, p. 30, cor. 3).

21) Let E and F be two Banach spaces. A continuous linear mapping from E into F is said to
be a Fredholm operator (or a quasi-isomorphism) if u- 1 (0) is finite dimensional and u(E)
has finite co dimension (this implies that u(E) is closed in F and u is a strict morphism); the
number Ind(u) = codim(u(E») - dim(u- 1 (0») is called the index of u.
a) Show that tu: F' --+ E' is also a Fredholm operator and that Ind(tu) = - Ind(u).
b) If u: E --+ F and v: F --+ G are two Fredholm operators, then so is v 0 u: E --+ G and
Ind(v 0 u) = Ind(u) + Ind(v).
c) If w: E --+ F is a continuous linear mapping with finite rank or with a small enough norm,
then u + w is a Fredholm operator and lnd(u + w) = Ind(u) (use exerc. 17, c) oflV, p. 65 and
18, b».
22) Let X be a real Banach space, E a finite dimensional subspace of X.
a) Let S be the unit sphere in X. Show that for every E > 0, there exists a finite number of
linearly independent points Zi EO S (l :( i :( r) such that for every x EO S I l E, there exists
an index i such that II x - Zi II :( €.
<
b) Let z; (1 :( i :( r) be points in the dual E' of E with II z;II :? 1 for every i, and Zi' z'.) = 8 i ·
(Kronecker's index), and let F be the closed subspace of codimension r in E which i~ ortho:
§5 EXERCISES TVS IV.67

gonal to the subspace of E' generated by the z;.

Show that for every XES n E and for every
y E F. Ilx + yll ? I-E. and in particular, E n F = {o}. so that the sum E + F is the topo-
logical direct sum.
c) Deduce from b) that the continuous projector P from E + F onto E corresponding to
the topological direct sum decomposition E EB F has a norm IIPII < I/O - E).

-If 23) Let X be a real infinite dimensional Banach space.

a) Suppose that for every A > 0, there exists a finite dimensional subspace E, of X such that
there exists no continuous projector P on X with image EA and such that I PII ~ A. Show that
every closed subspace Y of X with finite codimension has the same property as X.
b) By induction on n, show that there exists a decreasing sequence (X n ) of closed subspaces
of X, of finite codimension, and for every n, a finite dimensional subspace E" of X such that:
10 the sum El + E z + + E" is direct, and there exists a continuous projector P" on X n ,
00'

with norm ~ 2 whose image is El + E z + + E,,; 20 the space E" is contained in

00.

(I - P,,-l) (X,,-l); 3u there exists no continuous projector on X with image E" and with
norm ~ n + 2.
c) Let Z be the closed subspace of X generated by the union of the E". Show that there does
not exist a topological complement of Z in X (observe that if Q were a continuous projector
on X, with image Z, then (1 - P,,-l) P"Q would be a continuous projector on X with image E,,).

§5

I) Let I be an uncountable set, and let E be the space R(l) endowed with the topology defined
c».
in I, p. 24, exerc. 14; let E' be its dual (IV, p. 50, exerc. 16, Show that in E' there exist non
relatively compact (for cr(E', E») subsets H such that from every sequence of points of H we
can extract a sequence which converges to a point of H for cr(E', E).

2) a) Let X be a regular space and A a subset of X. Suppose that every sequence of points
of A has a limit point in X and that there exists a metrizable topology :Y on X which is coarser
than the given topology /70 , Show that the closure A of A in X is a compact metrizable space
(arguing by reductio ad absurdum, show that the topologies induced by .F and fa on A are
identical).
b) Let E be a Hausdorff locally convex space, which is the union of a sequence (En) of me-
trizable vector subspaces for the topology induced by that of E. Show that, for a subset A
of E to be relatively compact in E, it is necessary and sufficient that every sequence of points

° °
of A has a limit point in E. (If (W rn,,) (for m ? 1) is a fundamental system of convex neighbour-
hoods of in E" which are open in En' let Urn" be a convex neighbourhood of in E such that
E" n Urn" = W rn" (II, p. 33, lemma 2); on E consider the topology for which the Urn" (m ? 1,
n ? 1) form a fundamental system of neighbourhoods of 0).
c) Extend Smulian's theorem to a strict inductive limit space (II, p. 33) of Frechet spaces.

3) a) Let E be a Hausdorff locally convex space and E' its dual; suppose that there exists
a countable everywhere dense set in E' for the topology crCE', E). Show that the topology of E
(resp. cr(E, E'») is finer than a metrizable locally convex topology.
b) Let CXn) be a sequence of points of E such that every sequence extracted from (x,,) has a
limit point for the initial topology (resp. the topology cr(E, E'»). Show that there exists a sequence
extracted from (x,,) which converges in E for the initial topology (resp. the topology cr(E, E'»).
(Let (a;) be an everywhere dense sequence in E' for cr(E', E); extract a sequence (y,,) from (x,,)
such that (y", a~») tends to a limit for every index p, and show that the sequence (y,,) has
only one limit P01l1t for the initial topology (resp. for cr(E, E'».)
c) For a subset A of E to be relatively compact for the initial topology (resp. for cr(E, E'),
it is necessary and sufficient that from every sequence (x,,) of points of A, we can extract a
sequence (x"J which converges to a point of E for the initial topology (resp. for cr(E, E'»
(use exerc. 2, a»).

4) Let E be the Banach space CW(N). which does not satisfy the first axiom of countability
(I, p. 25, exerc. 1), and let E' be its dual. For every integer n ? 0, let e;, be the continuous linear
TVS IV.68 DUALITY IN TOPOLOGICAL VECTOR SPACES §5

form on E which associates to each x = (~n) E E the nth term of this sequence. Show that
the sequence (e;) is total in E' for O'(E', E); moreover, every sequence extracted from (e~)
has a limit point in E', for the topology O'(E', E), but there exists no sequence extracted from
(e;.) which converges in E' for this topology.

5) a) Let X be a compact space. H an arbitrary subset of the space '(,(X) of continuous nume-
rical functions on X. Let j~) be a point of '6'(X) which is in the closure of H for the topology
:Ys of simple convergence on (;!i'(X). Show that there exists a countable subset Ha of H such
that fa is in the closure of Ho for ·-:Fs ' (Show that for every pair of integers 111 > 0, n > 0 there
exists a finite subset H(m, n) of H with the following property: for every set of m points tk in
X (1 :;::: k :;::: 111), there exists f E H(m, n) such that Ifa(tk) - f(tk)1 :;::: l/n for I :;::: k :;::: 111.)
b) Let E 'be a locally convex metrizable space, E' its dual and H a subset of E. Show that if
xa is in the closure of H for the weakened topology O'(E, E'), then there exists a countable
subset Ho of H such that Xa is in the closure of Ho for this topology. (Use a), observing that
E' is the union of a countable family of compact sets for O'(E', E).)

"!l 6) a) Let X be a compact space. H a convex subset of the product space R X , consisting
of continuous functions on X. Suppose that every decreasing: sequence of non-empty convex
and closed subsets in H has a non-empty intersection. Show that the closure IT of H in R X is
compact and consists of continuous functions on X. (Argue by reductio ad absurdum: consider
a non continuous function U E IT: show that then there will exist a point a E X, a number
o > 0, a sequence (x,,) points of X and a sequence (j~) offunctiolls in H such that:
1° lu(xn)-u(a)l~o for all n: 2° Ifm(x")-fm(a)I:;:::~ for 111:;:::11: 3() lu(x.,)-j~(xn)I:;:::~ and

lu(a) - .f~(a)1 :;::: ~ for 111 ~ n + 1. Consider a limit point b of the sequence (x,,), and a func-
tion f belonging to the intersection of the Am' where Am is the closed convex envelope in H
of the set of all h for k ~ m.)
b) Let E be a Hausdorff and quasi-complete locally convex space, E' its dual. Let H be a
convex subset of E such that every decreasing sequence of non-empty and closed convex
subsets of H has a non-empty intersection; show that H is relatively compact in E for the
topology O'(E, E'). (Reduce to the case where E is complete; consider E to be embedded in
E'* and use a) and also Ill, p. 21, cor. 1.)
7) Let E be a Frcchet space satisfying the first axiom of countability, E' its dual. Show that
for a convex subset A' of E' to be closed for O'(E', E), it is suiTicient that, if a sequence (x~) of
points of A' has a limit a' in E' for O'(E', E), then a' E A'.

8) Let F and G be two vector spaces in separating duality, Show that the properties a) and ~)
of IV, p. 52, excrc. 6, il) are also equivalent to the following:
y) F, with 1:(F, G) is quasi-complete, and every bounded sequence of points of F has a
limit point for O'(F, G) (ef IV, p. 35, tho I);
0) F, with 1:(F, G) is quasi-complete, and every decreasing sequence of non-empty closed
convex and bounded sets in F has a non-empty intersection (ef exerc. 6, b».
9) Let E be a Hausdorff and quasi-complete locally convex space; for E to be semi-reflexive,
it is necessary and sufficient that every closed vector subspace of E, in which there exists a
countable everywhere dense subset, is semi-reflexive (ef IV, p. 35, tho n

"!l' 10) Let F be a HausdoIif, quasi-complete and non semi-reflexive locally convex space,
and let H be a closed hyperplane in E containing the origin. Let (C,,) be a decreasing sequence
of non-empty, closed convex and bounded sets, contained in H and not containing 0 and
whose intersection is empty (exerc. 8). Let x be a point not belonging to H, and let A be the closed
convex and balanced envelope of the union of the sets (1 - ~) x + e" for 11 > O.
a) Show that there exists no supporting hyperplane of A which is parallel to H (for y E x + H,
observe that there exists an integer n such that y ~ x + C n ).
§5 EXERCISES TVS IV.69

b) Let Z E H be such that z ¢c C 1 ; show that the convex envelope of the union of two closed
convex and bounded sets A and B = x + z + C 1 , is not closed (prove that x + z is in the
closure of this envelope, but does not belong to it).

II) Let A be an infinite set, and let E = J'C'(A) (IV, p. 47, exerc. I), E' = fleA) its dual,
E" = i'Xl(A) its bidual (IV, p. 54, exerc. 14). Show that every subset of E', which is compact for
cr(E', E") is strongly compact (use Smulian's theorem and IV, p. 54, exerc. IS).

12) Let E be a non reflexive Banach space. Prove that there exists a closed, non reflexive
vector subspace M of infinite codimension in E. (Let (x n ) be a bounded sequence in E which
has no limit point for the topology cr(E, E') (IV, p. 68, exerc. 8); by induction construct a
sequence (x n.) extracted from (xn) and a topologically free sequence (Yk) such that
I x nk - Yk I ~ Ilk for k ~ I, and consider the closed vector subspace of E generated by the
Y2k')
~ 13) Let E be a non reflexive Banach space satisfying the first axiom of countability, and let
M be a non reflexive closed vector subspace of E of infinite codimension (exerc. 12). Let (xn)
be an everywhere dense sequence in the unit sphere, C the closed convex balanced envelope
of the sequence (x,jn); C is strongly compact in E and C + M = A is a closed convex set
(GT. III. 9 4. No.1. cor. 1 to prop. 1). Let S be the unit ball in E, and B = An S.
a) Show that 0 is not an interior point of A and deduce that there exists Xo E E such that
Ax o ¢C A for A > O.
b) Show that there exists no supporting hyperplane of B passing through 0 (observe that such
a hyperplane must be a supporting hyperplane of C).
c) Let U o = M n S. and let (Un)n;"! be a decreasing sequence of closed convex, bounded
and non-empty sets. such that U 1 C tUo and such that the intersection of the Un is empty
(IV, p. 68, exerc. 8). Let F be the closed convex envelope of the union of the sets'!' Xo + Un
n
(n ~ O. Show that B n F = 0 but that there exists no closed hyperplane separating B
and F (use b)).
14) Let E be a Banach space satisfying the first axiom of countability. A sequence (en)n;,o
of elements of E is said to be a Banach basis if the following property is satisfied: for every
Xl
x E E, there exists a unique sequence (an) of scalars such that x = S anen, where the series on
n=O
the right hand side is convergent.
a) Show that the family (en) is total and free. Let En be the vector subspace (closed) of E gene-
rated by the em for indices m ~ n, and let P n be the projector from E onto En defined
rx:
I
11

by P n . (S ame m) = amem· Show that the Pn are continuous linear mappings and that
m=O m=O

sup
n
IIPnll < + 00. (Consider the norm on E defined by III
n=O
S aA,111 = sup I
n
t
m=O
amemll;
show that E is complete for this norm and deduce that it is equivalent to the given norm on
E (cf I, p. 17, tho 1).) Show that for every pair of integers p < q, the norm of the projection
P q •P of Eq onto Ep, which is parallel to the vector subspace generated by the em for indices m
such that p + I ~ m ~ q, is bounded by a number independent of p, q.
b) Conversely, let (en) be a total sequence of dements of E, which is a free family and is such
that the norms of the projections Pq •P for 0 ~ p < q are bounded by a number M independent
of p and q. Show that (en) is a Banach basis of E. (First prove that the sequence (en) is topolo-
gically free and define the projectors P n ; then I P n I ~ M for all n. Next observe that if d(x, En)
is the distance of a point x E E from En' then I x- Pn' xii ~ (M + I) d(x, En).) 1

1 It is clear that the existence of a Banach basis in E implies that E satisfies the first axiom
of countability. But there are examples of Banach spaces satisfying the first axiom of counta-
bility in which there does not exist a Banach basis (P. ENFLO, Acta Math., 1. CXXX (1973),
p. 309-317).
TVS IV.70 DUALITY IN TOPOLOGICAL VECTOR SPACES §5

~ 15) Let E be a Banach space with a Banach basis (en )(exerc. 14); then there exists a unique
sequence (c;,) in the dual E' of E such that the expression of every x E E in terms of the Banach
'Y'

basis (en) can be written as x = S (x, e~) en'

n= 0
a) Let Fn be the closed vector subspace of E generated by the em for m ? n. For every x' E E',
let I x'iln denote the norm of the restriction of the linear form x' to Fn' Show that, for (c;,) to be
a Banach basis of E', it is necessary and sufficient that for every x' E E', the sequence (1Ix'lln)
tends to O. (Consider the transpose 'P" and evaluate the norm II'P".x' - x'II.) In this case,
the Banach basis (e") is said to be contracting.
b) Suppose that the Banach basis (e,,) is contracting. Show that for every point x" of the bidual

En of E, the sequence of sums I (e~, x") em is bounded in E (consider the transpose '('p")
"
m=O

in E"). Conversely, for every sequence (il,,) of scalars such that the sequence of sums I" amem
m=O
is bounded in E, there exists a unique x" E E" s lIch that ( c~, x") = an for all n (use the com-
pactness of a closed ball in En for the topology a(E", E'»).
c) A Banach basis (en) in E is said to be complctc if, for every sequence (a") of scalars such that
"
I
~
the sequence of sums amCm is bounded, the series S a"c" converges. !f(e,,) is a contracting
/II--=- 0 ,,= 0
basis ofE, the basis (e;,l ofE' is complete (use the compactness of a closed ball in E' for a(E', E»).
d) In general, the sequence (e;) is a Banach basis of the closed subspace F' of the strong dual
E' of E, generated by the e;,. and there is an injective continuous linear mapping J from E
into the strong dual F' of F' such that (J.x, z') = (x, z') for all x E E and all z' E F'. Show
that there exists a constant K > 0 such that 111. xii ? K.II and that if the basis (e") is
complete, then J is an isomorphism from E onto the topological vector space F".
e) Show that for E to be reflexive, it is necessary and sutricient that the basis (c n ) is contracting
and complete (use b).

16) a) Let E be a Banach space. For an infinite sequence (x,,) of points of E, the following
properties are equivalent:
ex) The series with the general term xn is eommutatively convergent (OT, III, § 5, No.7).
~) For every subset I of N, the series defined by the sequence (Xn)"cl is convergent (OT, III,
~ 5, No.3, prop. 2 and § 5, exerc. 4).
y) For every sequence (E") of numbers equal to 1 or to - 1, the series with the general
term E"X" is convergent.
8) For every 0 > 0, there exists a finite subset J of N such that, for every finite subset H
of N not intersecting J, we have I I :s; E.
nEH
0) Let (e,,) be a Banach basis of E. The following properties are equivalent:
:x) For every permutation TC of N, the sequence (C RI ,,) is a Banach basis of E.
~) For every sequence (E") of numbers equal to 1 or to - 1, the sequence (tV,,) is a Banach
basis of E.
y) For every x = S ~"e" in E and every sequence (l1,,)"EN for which 111,,1 :s; I~nl for all 11,
n= 0
the series with the general term l1"C" converges in E.
00

8) For every x = S ~t1e" in E and every strictly increasing sequence (l1 k )kcN of integers? 0,
n= 0
the series with the general term ~nke"k converges in E.
0) There exists <1 real number M > 0 such that, for every finite subset J of:"l and for every
,
x = S ~"c" in E, we have III ~"e"il :s; M
11- () IIEJ

(To prove that ex) implies E), argue as in IV, p. 69, exerc. 14. To prove that ~) implies y),
q
reduce to the case where the ~" and the 11" are real, and consider the sums I (11"1',,, x')
n=p
Cor .y' E E'.)
§5 EXERCISES TVS IV.71

When these conditions are satisfied, (en) is said to be an unconditional Banach basis of E.
c) Suppose (en) is an unconditional basis. Show that there exists a real number K > Osuch
YO

that, for every sequence (E,) of numbers equal to 1 or - 1 and for all x = S ~nen in E, we
11= 0
x
have I S En~nenll os:; M Ilxll (same method as in IV, p. 69, exerc. 14). Deduce that for every
1/= 0

bounded sequence of scalars (A,) and for all x = S ~nen in X, we have

11=0

00 00

I S An~nen I
11=0
os:; 2K sup
11
IAnl . I S ~nen I
1/=0

(argue as in b). for the proof that ~) implies y».

'If 17) a) Let E be a Banach space with an unconditional Banach basis (en) (exerc. 16). Show
that if the basis (en) is not contracting (IV, p. 70, exerc. 15), then there exists a number a > 0,
a linear form x' E E' such that Ilx'll = I, a strictly increasing sequence of integers (17 k ) and for
each k, an element Yk which is a linear combination of the e} for n k os:; j os:; n k + l ' and is such
that IIYkl1 os:; 1 and (Yk' x') ): a. Deduce that for every finite sequence (1,.)1 "}"o1 of scalars,
m C1 m
we have III AjyJ ): 2K .I IAjl (use exerc. 16, c». Conclude that there exists a topological
J~l J~l

vector space isomorphism from 1'1 (N) onto a closed subspace of E.

b) Deduce from a) that if the strong dual E' of E satisfies the first axiom of countability, then
every unconditional basis (e,) of E is contracting, and hence (e;) is an unconditional basis of
E' (IV, p. 70, exerc. IS, a».
(Observe that if a closed vector subspace ofE is isomorphic to fl(N),
then E' cannot satisfy the first axiom of countability.)
c) Show that if E has an unconditional basis and if the strong bidual EO of E satisfies the first
axiom of countability, then E is reflexive. (Using IV, p. 51, exerc. 25, note that the strong dual
E' of E satisfies the first axiom of countability; then use IV, p. 70, exerc. 15. c) and IV, p. 53,
exerc. 11.)

'If 18) a) In the space RN of all infinite sequences of real numbers, consider the set J of all
sequences x = (~,.) for which the number
Ilxll = sUP(~Pl - ~py + (~P2 - ~PY + ... + (~Pm-l - ~pJz + ~;m' )1/2

is finite, the supremum being taken for all integers m ): 1 and for all strictly increasing sequences
of integers Pi < P2 < ... < Po1 + I' Show that Ilxll is a norm on J and that for this norm 1 is a
Banach space. For every x E E, show that the sequence (~,,) has a finite limit u(x) and that u

tion u(x) = °
is a non zero continuous linear form on E; let 10 denote the closed hyperplane of 1 with equa-
(R. C. James' space).
b) Show that the vectors en = (8 01n )m;;,0 form a Banach basis of 10 and that this basis is con-
tracting (IV, p. 70, exerc. 15). (To prove the latter point, argue as in exerc. 17, a), by showing
that the constructed sequence (Yk), is such that the series with the general term Yk/k is conver-
gent in E; this implies a contradiction.)
c) Show that the identity mapping from 10 onto itself can be extended to a topological vector
space isomorphism from the strong bidual 1~ onto J, in such a way that 1~/Jo is I-dimensional
(use IV, p. 70, exerc. 15, b».
Deduce that no Banach basis of 10 can be unconditional
(exerc. 17, e».
* d) Let Hi' H2 be two closed vector subspaces of 10 generated by the e 2n and the e 2n + 1
respectively, for n ): 0. Show that as topological vector spaces, H I and Hz are isomorphic
to the Hilbert space 1'2(N), and that 10 is not the sum of HI and Hz. *
e) Show that on 10 there exists no complex locally convex space structure having the real
locally convex space structure of Jo as the underlying structure (ef IV, p. 52, exerc. 3).
TVS IV.n DUALITY IN TOPOLOGICAL VECTOR SPACES APP.

APPENDIX

1) Let E be a Hausdorff locally convex space. K a compact convex subset of E, and S a set of
continuous affine linear transformations from K into itself, which is stable under composition.
The set S is said to be distal if, for every pair of distinct points a, b in K, the closure of the set
of pairs (s. a, s. b), where s ranges over S, does not contain any point of the diagonal of K x K.
a) Show that an equicontinuous group of affine transformations of K is distal.
b) Show that if K is non-empty and S is distal, then there exists at least one point of K which
is invariant under every transformation of S. (If M is a non-empty compact convex subset
of K which is stable under S, show that if M contains two distinct points XI' x 2 ' and if A is the
closure of the orbit of X = Xl ~ x 2 , then A cannot contain any extremal point of M. Deduce
that if L is a minimal element of the family of non-empty compact convex subsets of K. which
are stable under S, then L reduces to a point; argue by reductio ad absurdum: with the same
notations, the closed convex envelope of A would be equal to L, which would contradict
the Krein-Milman theorem.)

2) Let E be a Banach space, K a precompact subset of E which is not a single point, and d the
diameter of K. Show that there exists a point Xo E K and a number r such that 0 < r < d,
such that II X - Xo II ~ r for all X E K (choose f: > 0 small enough, and n points Yl' ... , Yn
of K such that every point of K is at a distance < f: from one of the Yj' and put
Xo = 1.n (YI + ... + Yn).) Deduce a new proof of the Ryll-Nardzewski theorem for convex,
strongly compact sets in E.

* 3) Let G be a topological group and TI a continuous unitary representation ofG on a complex

hilbertian space E. A continuous linear mapping c: G --> E which satisfies the relation
c(st) = TI(s). c(t) + c(s)
for every s, tin G, is called a continuous l-cocycle. Let ZI(G; E) denote the complex vector
space of continuous l-cocycles. For every a E E, the mapping 8(a): Sf-> TI(s). a - a is a conti-
nuous l-cocycle, called the cobord of a. Let Bl(G; E) denote the image of the linear mapping
8: E --> Zl(G; E); put Hl(G; E) = Zl(G; E)jBl(G; E) (<< first continuous cohomology
group of G with values in EO).
a) Show that Bl(G; E) is composed of continuous and bounded I-cocycles. (For every conti-
nuous l-cocycle c and every s E G, we define an affine transformation As on E by
As' X = TI(s). X + c(s); then Ast = As' At for all s, t in G. Let K be the closed convex envelope
of c(G); then As(K) = Kfor all s E G, and As induces an isometry ofK onto itself. If c is bounded,
show that the Ryll-Nardzewski theorem applies to As' and if As.a = a for all s E G, then
c = - 8(a).)
b) If G is compact, show that Hl(G; E) = {O}. *

~ 4) Let G be a discrete group. We say that G is a group on which a mean can be defined
ifthereexistsalinearformuone~(G)(I,p.4)suchthatu(x)? Ofor x ? O,u(l) = 1,andsuch
that u(y(s) x) = u(x) for all S E G and all x E e:(G) (where (y(s) x) (t) = X(S-l t) for all t E G)
(invariance under left translations).
a) If we put x = x(t -1) for x E e:(G) and t E G, and if the linear form u is invariant under
left translations, then the linear form v: x f-> u(x) is invariant under right translations, in other
words v(8(s) x) = vex) for all s E G and for all x E e:(G) (where (8(s) x) (t) = x(ts) for all
t E G). Put Fx(s) = u(8(s) x) for s E G and x E e:(G), then Fy(t)x(s) = F xes) and Fo(t)x(s) = F x(st)
for all t E G; deduce that the linear form w on e:(G) defined by w(x) = v(F x) is invariant
under left and right translations, and is such that w(x) ? 0 for x ? 0 and w(l) = I.
* b) Let K be a non-empty compact space and r a subgroup of the group of all homeo-
morphisms from K onto itself. Show that if a mean can be defined on r, then there exists a
measure J.l ? 0 on K of mass I, which is invariant under r. (If a E K, consider the linear mapping
APP. EXERCISES TVS IV.73

which associates to each continuous real function fin K, the function cr f-> f( cr(a)) belonging
toe~(n)
c) Let E be a Hausdorlf topological vector space over R, and K a non-empty compact convex
subset of E. Let r be a group of continuous affine transformations from K onto itself. Show
that if a mean can be defined on r, then there exists a point bE K such that cr(b) = b for all
cr E r. (Use b) and consider the barycenter of fl.)
d) Show that a mean can be defined on a discrete group G if and only if there exists a non
null measure fl which is invariant under G, on every non-empty compact space K on which G
operates continuously. (If E = ff(G), consider the unit ball B in E' = eMG), endowed with
cr(E', E), and associate to each element x E ~(G) = E" its restriction to B.) If G is countable,
it is enough that the above property holds for every compact metrizable K. *

~ * 5) Let G be a discrete group.

a) Show that a mean can be defined on G (exerc. 4) if and only if the closed vector subspace N
of e~(G) generated by the functions yes) x - x, where s E G and x E e~(G), is distinct from
e~(G) (use the Hahn-Banach theorem).
b) Suppose that for every € > 0 and for every finite sequence SI' ... , Sk of elements of G, there
exists a non-empty finite subset F of G such that
Card(F n sjF) ~ (I - €) Card(F) for I ~ j ~ k.
Show that a mean can be defined on G (use a) and show that 1 ¢ N).
c) Suppose a mean can be defined on G. Let € > 0 and let SI' ... , Sk be elements of G. Show
that there exists a vector x ~ 0 in the space e~(G) such that Ilxll = I and that
k
I IIY(s) x-xii ~€. (In the space E=(e~(G)l, consider the set C of points (y(s) X-X)I<;j<;k'
j= I
where x ranges over the set of all vectors ore~(G) such that x ~ 0 and Ilxll = I. Show that
o belongs to the closure of the convex set C for the topology cr(E, E'); for this, use a), observing
that the unit ball of E is dense in the unit ball of E" for cr(E", E').)
d) For every x ~ 0 in eMG), and for every a > 0, let xa denote the characteristic function of
the set of all S E G such that xes) ~ a (i.e. xa(s) = I if xes) ~ a, xa(s) = 0 if xes) < a). Then for

every S E G, {" x/s) dr = xes) and, for two elements x ~ 0, Y ~ 0 ofe~(G),

{" Ix/s) - Yr(s) I dr = Ix(s) - y(s)l·

e) Show that, if a mean can be defined on G, then for every € > 0, and for every finite sequence
of elements SI' ... , Sk of G, there exists a non-empty finite subset F of G such that
Card(F n sjF) ~ (I - €) Card(F) for I ~ j ~ k.
k
(Show that, having chosen x as in c), there exists an a> 0 such that I IIY(s) Xa - xall ~ €;
j= I
use d).) *
~ 6) Let S be a set on which a group r operates on the left (A, I, § 5, No.1). Let E be the real
vector space IE(S) of all bounded numerical functions on S (I, p. 4, Example). Suppose that
the group r (endowed with the discrete topology) has a left invariant mean, and let r operate
on E in such a way that sf(x) = f(S-1 x) for all s E r, fEE and XES.
Let gEE be a positive function; let EI be the vector subspace ofE generated by the functions
sg, where s ranges over r; let E z be the vector subspace of E generated by the positive func-
tions which are bounded by functions of EI .
Show that if there exists a non null positive linear form q, on E I , which is invariant under r,
then there exists a non null positive linear form on E 2 , which is invariant under r. (Using
prop. 1 of II, p. 21, first construct a positive linear form on E 2 , which extends q" and let r
operate on the set of these extensions). In particular consider the case where 9 = I.
TVS IV.74 DUALITY IN TOPOLOGICAL VECTOR SPACES APP.

* 7) a) Let mbe the set of bounded subsets of the plane R2 and let r be the group of dis-
placements of R2 ; let C be the square (0, I) x (0, 1). Show that there exists a positive addi-
tive set function A defined on m, which is invariant under r and is such that A(C) = 1 (apply
exerc. 6 to the case where g is the characteristic function of C; observe that r is solvable,
hence has an invariant mean (IV, p. 41, corollary)). If A is a bounded subset ofRz, whose boun-
dary is negligible for the Lebesgue measure Il on RZ, then A(A) = Il(A) (for every E > 0,
there exists two sets Al and A2 such that Al cAe A z , Il(A z - A l ) < E, and Al and A z are
unions of a finite number of squares).
b) Consider the same question as in a), taking for mthe set of all subsets ofR 2, for r the group
of similarity transformations and C = R2. *

8) Let E be a real vector space and r a solvable group of automorphisms of E. Let p be a

semi-norm on E which is invariant under rand M a vector subspace of E, invariant under r.
Let u be a linear form on M, invariant under r and such that lu(x)1 ~ p(x) for all x EM.
Show that there exists a linear form v on E which is invariant under r and is such that Ivl ~ p
and that v extends u. (Let K be the set of linear forms v on E which extend u, and such that
Ivl ~ p; then K is a convex subset ofE*, stable under r and compact for the topology induced
by cr(E*, E). Apply the corollary of IV, p. 40.)
 TABLES TVS IV.7S

TABLE I. - Principal types of locally convex spaces.

(N.B. - « Dual» is taken in the sense of« strong dual ».)

Finite Frechet- Montel

dimensional - - - - - - - - - - . . ,..~ Montel - - - - - - - - - - - . . space
space space
t
Hilbert
space

I
Reflexive Banach ----7~----~----_ Dual of a
space reflexive
Frechet space

~e~~h~e
~
Frechet space space
satisfying the first /
axiom of coun.tability
Banach
space D·st·n
1 1 gUi·shed Reflexive
Frechet space space
Dual of a
distinguished
Frechet space /
/
UItrabornological
Dual of a
semireflexive
space
space

\
Frechet

Barrelled Dual of a
space Frechet

Semi~reflexive
semi-barrelled

Complete
space ~ space

Quasi-complete Semi-barrelled
space space
TVS IV.76 DUALITY IN TOPOLOGICAL VECTOR SPACES

TABLE II. - Principal homologies on the dual of a locally convex space.

Contained in a convex
balanced and strongly Equicontinuous
compact set

1
strongly

I
bounded

Contained in a convex
balanced and weakly compact set

Bounded for aCE', E")

t (5)
~

~(6~~
Weakly
Precompact

Bounded fo,
(1) the 6-topology

~(7)t
Weakly
bounded

N.B. - We denote by G a family of bounded subsets of E such that every element is con-
tained in a set belonging to 6. A number on the side of an arrow indicates that the corres-
ponding implication holds if the property with the same number is satisfied.

PROPERTIES

1) Whenever E is semi-reflexive;
2) whenever E is bomological (III, p. 22, prop. 10);
3) if and only if E has the Mackey topology T(E, E') ;
4) if and only if E is barrelled;
5) if and only if E' is quasi-complete for cr(E', E) ;
6) whenever E is semi-complete (a fortiori, quasi-complete or complete) (III, p. 27, cor. 1);
7) whenever 6 consists of sets whose closed convex balanced envelope is semi-complete
(III, p. 27, tho 2).
When E is a Monte! space, all the preceding homologies are identical.
 CHAPTER V

Hilbertian spaces!
(elementary theory)
Throughout this chapter, K d£!}!otes the field R or the field C. For every complex
number ~ = cr + i~ (cr, ~ real), ~ denotes the conjugate cr - i~ of ~; in particular,
we have ~ = ~ if and only if ~ is real.

§ 1. PREHILBERTIAN SPACES AND HILBERTIAN SPACES

1. Hermitian forms

We recall the following definition given in Algebra (A, IX, § 3, No. 1) :

DEFINITION 1. - Let E be a vector space over the field K. A hermitian form (on the
left) on E is a map f from E x E into K satisfying the following conditions (for Xl'
xZ' x, Yl' YZ' Y in E and A, ~ in K) :

Jf(x l + xZ' y) = f(xl' y) + j(x z , y)

(1)
1 f(x, Yl + yz) = f(x, Yl) + f(x, yz)

(2) 1 f("Ax, y)
f(x, ~y) =
= 'i,J(x, y)
~f(x, y)
(3) f(x, y) = f(y, x) .

When the field K is R, the notion of hermitian form on E reduces to that of sym-
metric bilinear form on E x E (A, III, § 6, No.3).
We note that the second condition (1) and the second condition (2) follow from the
other three.

1 For the reader specially interested in hilbertian spaces, we point out that only No.7
of§ 1 and No.8 of § 4 depend on results of chapters III and IV. For this the reader can consult
« Summary of some important properties of Banach spaces» which appears at the end of this
volume. The only references to chapters I and II concern the definition of a convex set and of
a semi-norm (II, p. 1 and p. 7), that of a topological direct sum (I, p. 4), of a total family and
a topologically independent family (I, p. 12).
TVS V.2 HlLBERTIAN SPACES §1

From (1) and (2) we deduce immediately that

(4)

In particular, if E is finite dimensional, and if (e)l ~j~1I is a basis of E, then

I I
II II

for x = ~jej and y = T]jej' we have,

j~l j~l

f(x, y) = I C(jk~jT]k
j.k
with the notation C(jk = f(ej' ek); moreover, relation (3) amounts to C(jk = C(kj for
every pair of indices j. k; this implies in particular that the numbers C(jj are reaL
From (3), the number Q(x) = f(x, x) is real for all x E E. Moreover, we imme-
diately establish the following formulas, known as polarization formulas

(5) 4f(x, y) = L sQ(x + sy) if KisR,

62 = 1

(6) 4f(x, y) = L sQ(x + ey) if K is C.

£4= 1,GEe

Remark. - We observe that formula (6) is valid for every sesquilinear form on E x E
(that is to say, for every function f satisfying (1) and (2), but not necessarily (3)). This
remark shows that, when K = C, a sesquilinear form f such that f(x, x) is real for
all x E E is necessarily hermitian: relation (6) then gives fey, x) = f(x, y) since we
have y + eX = e(X + "Ey) and Q(eZ) = Q(z) whenever e4 = 1.

From the polarisation formulas, we have in particular,

PROPOSITION 1. - Iff is a hermitian form on E, and M a vector subspace of E such that

f(x, x) = 0 for all x E M, then we also have f(x, y) = 0 for every pair of points x, y
in M.
Let f be a hermitian form on E; the set N of all x E E such that f(x, y) = 0 for
all y E E is a vector subspace of E. It follows from (3) that, if Xl == X z (mod. N) and
YI == yz (mod. N), we have f(x l , YI) = f(x z , yz); hence, on the quotient space
E/N we define a sesquilinear form f by putting j(x, y) = f(x, y) for all x E X and
all y E y; it is clear that j is hermitian and that the re!ation «j(x, y) = 0 for all
x
y. E E/N » implies = 0 in E/N, in other words (A, IX) f is separating. We say that
f is the separating hermitian form associated with f

2. Positive hermitian forms

DEFINITION 2. - Let E be a vector space over the field K. A hermitian form f on E

is said to be positive if f(x, x) ~ 0 for all x E E.
It is clear that hermitian forms on a vector space E form a vector space over the
field R (but not over the field C, when K is C) : in this space the positive hermitian
forms constitute a pOinted convex proper cone (II, p. 10) as a result of def 2 and prop. 1.
No. 2 PREHILBERTIAN SPACES AND HlLBERTIAN SPACES TVS V.3

PROPOSITION 2. - Iff is a positive hermitian form, we have

(7) If(x, yW ~ f(x, x) fey, y)

for every x and y in E (Cauchy-Schwarz inequality).
First assume that we have fey, y) =f. O. For every S E K, we have
fey, y) f(x + Sy, x + Sy) ~ 0
which can be written as

f(x, x) fey, y) - If(x, yW + (Sf(y, y) + f(x, y)) (V(y, y) + f(x, y)) ~ O.

Replacing S by - f(x, y)lf(y, y) in this inequality, we get (7). If f(x, x) =f. 0, we

argue similarly.
Finally, if f(x, x) = fey, y) = 0, we have f(x + Sy, x + Sy) ~ 0 for all S E K,
which can be written as
Sf(x, y) + Sf(x, y) ~ O.
Replacing S by - f(x, y) in this inequality, we get - 2 If(x, yW ~ 0, and therefore
f(x, y) = 0; we again get (7) in this case.

COROLLARY 1. - Iff is a positive hermitian form, the set N of all x E E such that
f(x, x) = 0 coincides with the vector subspace of all x E E such that f(x, y) = 0 for
all y E E.

COROLLARY 2. - For a positive hermitian form to be separating, it is necessary and

sufficient that the relation x =f. 0 implies f(x, x) > O.
This follows immediately from cor. 1.
For every positive hermitian form f on E, the separating hermitian form asso-
ciated with f (V, p. 2) is evidently a positive hermitian form on E/N.

PROPOSITION 3. - Let f be a positive hermitian form on E. Put

p(x) = f(x, X)1/2

for all x E E. Then p is a semi-norm on E, and is a norm if and only iff is separating.
It is enough to prove the inequality p(x + y) ~ p(x) + p(y). But we have

f(x + y, x + y) = f(x, x) + fey, y) + f(x, y) + f(x, y)

and, by Cauchy-Schwarz inequality

(8) f(x + y, x + y) ~ f(x, x) + fey, y) + 2(f(x, x) fey, y))1/2

= (f(x, X)1/2 + fey, y)1/2)2 .

Remarks. - 1) Suppose f is positive and separating, and let x, y be two vectors =I o.

The proof of Cauchy-Schwarz inequality shows that, if the two members of (7) are
equal, then there exists a scalar ~ such that f(x + ~y, x + ~y) = 0, hence x + ~y = 0,
TVS V.4 HILBERTIAN SPACES §1

in other words, x and yare linearly dependent; the converse is immediate. The proof
of inequality (8) shows that the equality p(x + y) = p(x) + p(y) is possible only if
x and yare linearly dependent; if y = AX, the preceding equality can be written as
11 + AI = 1 + IAI, and implies that A is real and positive.
2) Let I be a positive hermitian form on E, and let E be assigned the semi-norm
X f-> I(x, X)l/Z; if j is the positive, separating hermitian form defined on E/,N asso-
ciated with j, then the normed space obtained by assigning the norm x f-> I(x, X)l/Z
to E/N is the normed space associated with E (II, p. 5).

DEFINITION 3. - Let E be a vector space over the field K. A semi-norm p on E is said

to be prehilbertian if there exists a positive hermitian form f on E such that
p(x) = f(x, X)1/2 for all x E E.
Observe that for a semi-norm p on E, there exists at most one positive hermitian
form f such that p(x) = f(x, X)1/2 for all x E E; this follows from the polarization
formulas (V, p. 2).

3. Prehilbertian spaces

DEFINITION 4. - A prehilbertian space is a set E with the structure of a vector space

over K and with a positive hermitian form. We say that E is a real (resp. complex)
prehilbertian space when K is R (resp. K is C).

Examples. ~ 1) The form (A, I-l) f->):I-l defines a prehilbertian structure on K, said
to be canonical. When K is considered as a prehilbertian space, we shall always mean,
unless otherwise mentioned that it has this structure.
2) Let I be an interval (bounded or not) in R, and let E be the set of regulated func-

f
tions (FVR, II, p. 4) defined on I with values in C, having compact support. It is clear
that E is a vector space over C; let I be the sesquilinear form (x, y) f-> x(t) yet) dt;
1
it is immediate that I is a positive hermitian form on E, and hence defines a prehil-
bertian structure on this space.
3) Let n ~ 0 be an integer. We define a prehilbertian space structure on the space
Kn, by means of the hermitian form

(x, y) f-> I
j~ 1
XjYj

(for x = (xl' ... , x n) andy = (yp ... , Yn)). When K is R, we see that this isjust the scalar
product of two vectors of Rn (GT, VI, § 2, No.2).
* 4) Let £z (or fZ(N)) be the set of sequences x = (xn)neN of elements of K such that
L
00 •

IXnlz is finite. One can show that £z is a vector subspace of KN and define a pre-
n=O

L
00

hilbertian space structure on £z by means of the hermitian form (x, y) f-> xnYn
n=O
(cf V, p. 18). *
5) Let E be a real prehilbertian space, I the corresponding symmetric bilinear form
on E. Let E(C) be the vector space complexification of E; we identify E with a subset
of E(C) by the map x f-> 1 ® x, in such a way that every element of E(C) can be written
uniquely as Xl + iX z with Xl' X z in E. The map I extends uniquely to a hermitian
form i;C) on E(c) ; we have,
i;C)(x l + ixz' Yl + iyz) = I(x p Yl) + I(x z , Yz) + i(f(xp Yz) - I(x z , Yl)) .
No.3 PREHILBERTIAN SPACES AND HILBERTIAN SPACES TVS V.5

In particular, we have
i;C)(x l + ix z , Xl + ix z ) = f(x» Xl) + f(x z , xz) ? 0,
hence i;C) is positive. We say that E(c), with i;C) is the prehilbertian space complexifi-
cation of E.
Whenever only one prehilbertian space structure on a vector space E is under
consideration, the value, for a pair (x, y) of points of E, of the hermitian form which
defines the said structure is denoted by < xIY)E or simply < xIY), if no confusion
is likely to arise. This number is called the scalar product 1 of x and y (scalar square
of x if y = x). Two vectors x, yare said to be orthogonal if < xly) = o. The function
x f--> Ilxll = < xlx)1/2 is a semi-norm on the vector space E (V, p. 3); a prehilbertian
space is always considered with this semi-norm assigned to it (and consequently
also with the corresponding topology and uniform structure).
With these notations, in a prehilbertian space E, the Cauchy-Schwarz inequality
can be written as

(9) l<xly)1 ~ Ilxll·llyll·

Consequently, the scalar product is a continuous sesquilinear form on E x E (II,

p. 5, prop. 4).
In order that E be Hausdorlf, it is necessary and sufficient that x f--> Ilxll is a norm
on E; in other words, that the hermitian form (x, y) f--> < xly) is positive and
separating; this is equivalent to saying that 0 is the only vector of E, which is ortho-
gonal to itself.
According to general definitions (S, IV, ~ 1, No.5), an isomorphism from a prehil-
bertian space E onto a prehilbertian space F is a bijective linear mapping u from E
onto F such that
(10) <u(x)lu(y) = <xly)
for every x and yin E. We deduce from this that Ilu(x)11 = Ilxll for all x E E, and u
is evidently an isomorphism for the topological vector space structures ofE and of F;
if E and Fare Hausdorlf, u is an isometry from E onto F. Conversely, if u is a bijective
linear mapping from E onto F, such that II u(x) II = II x II for all x E E, the polarization
formulas (V, p. 2) show that u is a prehilbertian space isomorphism from E onto F.
Let E be a complex prehilbertian space, and < xly) the scalar product in E. On
the set E, we can define a second vector space structure with respect to C, taking
the same law of the additive group and for the law of external composition (A, x) f--> ~x
(A, II, ~ 1, No. 13) for this vector space structure, (x, y) f--> < ylx) is a positive hermitian

1 It may happen sometimes that we write (xIY) for <ylx). Observe that the formula (4)
of V, p. 2, takes the following equivalent forms:
(4')
i,j

(4") (Z: "'ixilI IljyJ = I

i,j
"'i~j (x.!..)!) .
TVS V.6 HILBERTIAN SPACES §1

form. The prehilbertian space E obtained by assigning this new vector space struc-
ture and the new hermitian form to E, is said to be conjugate to E. An isomorphism
u from E onto E is a semi-linear mapping from E onto itself (with respect to the
automorphism ~ f---+~ of C) such that <u(y)lu(x) = <xly) or <u(x)lu(y) = <xly)
(for x, y in E); such a mapping is said to be a semi-automorphism of the prehilbertian
space E.
If E is a prehilbertian space, M a vector subspace of E, the restriction of the scalar
product <xly) to M x M is a positive hermitian form on M, which then defines
a prehilbertian space structure on M; we say that this structure is induced by the
structure of E, or that M is a prehilbertian subspace of E.

4. Hilbertian spaces

DEFINITION 5. - A hilbertian ~pace (or Hilbert space) is a prehilbertian space which

is Hausdorff and complete. We say that a norm on a vector :,pace E (over K) is hilbertian
if it is prehilhertian, and if the normed space E is complete.
If E is a hilbertian space and M a closed vector subspace of E, the prehilbertian
space structure induced on M is in fact a hilbertian space structure. In this case
we say that M, with the induced structure is a hilbertian subspace of E.

Examples. - I) The prehilbertian spaces defined in examples I, 3, 4 of V, p. 4, are

hilbertian spaces. On the other hand, the prehilbertian space defined in example 2
is neither Hausdorff, nor complete. The complexification of a hilbertian space is a hil-
bertian space.
* 2) Let X be a Hausdorff topological space and let 11 be a positive measure on X.
Let L 2(X, 11) be the space consisting of equivalence classes, for 11, of all square Il-inte-
grable functions on X with values in C. This is a complex hilbertian space, whose scalar
product is given by

<fig> = Ix f(x) g(x) d~l(X) . *

* 3) Let n ~ I be an integer and let U be an open set in R". Let 11 be the measure
on U induced by the Lebesgue measure on R", and put £ 0 = U(U, 11). Let,1/' 1 denote
the space of all functions f E £ 0 with the following property; for I :::; i :::; n, there
exists a function gi E £0 such that

L g/x) hex) dll(x) = - L f(x) Dih(X) dll(X)

for every function h of class C I with compact support in U. The function Yi is defined
uniquely up to equivalence with respect to 11), and is denoted by OJ or af/ax. (ith
partial derivative). By induction on the integer s ~ I, we define ,Yt' as the set ~f all
functions fE £1 such that DJE ;t{'s-I for 1 :::; i:::; n. We define a scalar product
on ;t{" by the formula

<fig) = t , I.
k = 0 1:::;; t 1 ~ ... ~ lk ~ n
fDil ... Dikf.Dil ... Oik9dll.

Then £s is a complex hilbertian space, called Sobolev space of index s.

* 4) Let X be a differential variety of class C' (with r ~ 1) pure of finite dimension n.
No.4 PREHILBERTIAN SPACES AND HILBERTIAN SPACES TVS V.7

In the vector fibre space A"T(X), let L be the complement of the zero section. For every
real number A i= 0, the mapping u f-7 Au from A"T(X) into itself leaves L stable.
Let rx be a complex number. A complex valued function W on L such that
w(Au) = IAI" w(u) for u ELand any non-zero real number A is called a density of
order rx on X. We say that a density w of order 1 is locally integrable if there exists an
open cover (V)iEI of X, and for every i E I a system of coordinates ~i = (~f, ... , ~?)
on Vi and a complex valued function J; on ~i(V) satisfying the following conditions:
a) The function J; is locally integrable on the open set ~JV) of Rn with respect
to the Lebesgue measure ~;
b) Let x E Vi; if (al,i,X' ... , an,i,x) is the basis of TxX associated to the system of
coordinates (~;, ... , ~;') in Vi we have
W(al.i,x /\ ... /\ an,i.x) = f;(~f(x), ... , ~7(x)) .

Then, there exists one and only one measure & on X such that for every i E I, the image
under ~i of the restriction of & to Vi is equal to the measure J;. ~ (c.f VAR, R, 10.4.3).
Let 1/ (resp. %) be the vector space of measurable densities w of order 1/2 such
that the measure associated with the density Iwll of order I is bounded (resp. null).
Let WI and W z be in 1/; then w = 00 1 wl is a density of order I, and the measure &
associated with w is bounded; the number Ix & depends only on the classes 00 1 and

< <
ooz of WI and W z modulo % and is denoted by w1 lw z ) or 00 1 100 2 ), Then the mapping
f-7 <00 1 100 1 ) assigns a complex hilbert ian space structure to the vector space
(00 1 , ooz)
QI~Z(X) = 1//%. *
5) Let D be the open disc with centre 0 and radius 1 in C. The Hardy space BZ(D)
consists of all holomorphic functions f: D ---> C for which

sup
O<R<l
IIo
If(R.e(e))i2 de < + 00 .

If fl and fz belong to Bl(D), the limit

<
exists; the mapping (fl' fz) f-7 fllfz) assigns a complex hilbertian space structure
to the vector space BZ(D).
For a function f: D --> C to belong to BZ(D), it is necessary and sufficient that there
I
00

exists a sequence (an)nEN of complex numbers such that lanl l < + 00 and that
n= 0

00

fez) = I anzn
n= 0

I
00

for all ZED. Then we have Ilflll = lanl z which gives an isomorphism from Bl(D)
n=O
onto the hilbertian space £1 (V, p. 4). *

Every Hausdorff prehilbertian space is isomorphic to an everywhere dense sub-

space of a hilbertian space determined up to an isomorphism. Precisely:

PROPOSITION 4. - Let E be a Hausdorff prehilbertian space, E the normed space

completion ofE(GT, IX, § 3, No.3). The scalar product (x, y) f---* xly) extends by conti- <
TVS V.8 HILBERTIAN SPACES §1

nuity to a positive and separating hermitian form on E, and defines a hilbertian space
structure on E.
The existence of the extension of (x, y) ~ <xly) to E x E follows from the conti-
nuity of this sesquilinear form an E x E (GT, Ill, § 6, No.5, tho O. Moreover, this
extension, which will also be denoted by (x, y) ~ <xly) is a hermitian form and satis-
fies the relation <xix) = Ilx11 2 , by virtue of the principle of extension of identities
(11xll being the norm on E obtained by extending the norm on E by continuity);
this proves that the relation <xix) = 0 implies x = 0 in E, hence that the form
(x, y) ~ <xly) is positive and separating, and consequently defines a hilbertian
space structure on E. Q.E.D.
This hilbertian space is said to be the completion of the Hausdorff prehilbertian
space E.

* Example 6. ~ Let U be an open subset of Rn (n ;;:. 1). Let'?? t(U) be the vector space
of all functions of class C 1 with compact support in U. We define a Hausdorff prehil-
bertian space structure on '??5(U) whose scalar product is given by

<fig) = Jl Iv Dd(x).Dig(x) dx.

This prehilbertian space is not complete. Its completion is called the Dirichlet space
associated with U. *

COROLLARY. ~ Let V be a vector space over K and f a positive hermitian form on V.

a) There exists a Hilbert space E and a linear mapping u: V --+ E such that
f(x, y) = <u(x)lu(y) >for x, y in V, and such that u(V) is dense in E.
b) If two pairs (Ei , ui ) satisfy the conditions analogous to a), then there exists a
unique isomorphism q, from the Hilbert space E1 onto the Hilbert space E2 such that
u2 = q, u1 .
0

Let N be the set of all x E V such that f(x, x) = o. We define a positive and sepa-
rating hermitian form on the space VIN by <xl.y) = f(x, y) for x E X and y E y. Let
E be the hilbertian space completion of VIN and u the mapping x ~ x + N from
V into E. Then the conditions of a) are satisfied.
Under the hypotheses of b), N is equal to the kernel of u 1 and to that of u 2 . Hence
there exists a bijective linear mapping q,o from u1(V) onto uiV) such that
uz(x) "" q,o(ut(x») for all x E V. We verify immediately that q,o is an isomorphism
of prehilbertian spaces, hence an isometry. Since ui(V) is dense in E j for i = 1, 2, q,o
extends uniquely to an isometry q, from E1 onto E2 , and b) follows.
We say that the hilbertian space E is the separated completion of V (for the form j).

Example 7. ~ Let G be a group (with unit element 1) and 1t a homomorphism from

G into the group of automorphisms of a complex hilbertian space E; we say that 7t
is a unitary representation of G in E. Let a € E; We put
<j>(x) = <al7t(x).a)
for all x € G. Then <l>: G --+ C is positive definite, in other words satisfies the relation:
No.5 PREHILBERTIAN SPACES AND HILBERTIAN SPACES TVS V.9

(PD) For every 1..1' ... , An in C and xl' ... , xn in G, we have

n
(11) L: J::;Ajq,(x;- 1 x) ::;:, 0 .
i,j= 1

n
In fact, the first member of (11) is precisely II L: A;rc(xJ. a11 2 .
i= 1
Conversely, let q, be a positive definite function on G. Let C(G) be the vector space
of all functions with finite support on G. We define a hermitian form «1> on G by

(12) «1>(u, v) = L: u(x) v(y) q,(x-1y)

X,YEG

and the relation (PD) expresses the fact that «1> is positive. By the corollary of prop. 4,
there exists a hilbertian space E and a linear mapping p : C(G) ..... E, with a dense image,
such that
(13) «1>(u,v) = <p(u)lp(v» for u,v in C(G).
For every x E G, let Yx be the left translation by x in C(G) defined by yxu(y) = u(x-1y)
for u E C(G) and y E G. We have «1>(yxu, Yxv) = «1>(u, v). Now apply assertion b) of the
corollary of prop. 4 to p and po y x : there exists a unique automorphism rc(x) of the
hilbertian space E such that po Yx = rc(x) 0 p. We see immediately that rc is a homo-
morphism from G into the group of automorphisms of E.
Let b be the element of C(G) defined by 6(1) = I, 6(x) = 0 for x i= 1 in G. We have
u = L: u(x). y xb for all u E C(G), and so p(u) = L: u(x) rc(x). a by putting a = pCb).

Formulas (12) and (13) imply that q,(x) = <alrc(x). a) for all x
XEG XEG

E G. We remark that
the set of vectors rc(x).a for all x E G, is total in E.

5. Convex subsets of a prehilbertian space

If we calculate Ilx - yl12 = <x - ylx - y) and Ilx + yl12 = <x + ylx + y) for
any two points x, y of a prehilbertian space E, we immediately get the « identity of
the median»

(14)

From this identity we deduce the following proposition

!(x + y)

B'

FIG. 1.

real number such that °: ;

PROPOSITION 5. ~ Let E be a prehilbertian space. Let d be a real number > 0, 8 a
8 ::;; d. Let Band B' be subsets ofE defined by Ilxll < d,
TVS V.10 HlLBERTlAN SPACES §1

Ilxll ~ d + 8 respectively, and let A be a convex set contained in Bf - B. Then for

every pair of points x, y of A, we have Ilx - yll ~ J12d8 (fig. 1).
In fact, we have !(x + y) E A, hence IWx + y)11 ~ d; hence from (14) we get
the inequality
11!(x - y)112 = !(llxl12 + Ily112) - IWx + y)112 ~ (d + 8)2 - d 2 ~ 3d8

from which the proposition follows.

THEOREM 1. - Let E be a prehilbertian space, and H a non-empty convex subset ofE

such that H is a Hausdorff and complete uniform subspace ofE. For every x E E, there
exists a unique point PH(X) in H such that Ilx - PH(X) I = inf Ilx - YII. The element
YEH
PH(X) of H is also the unique element a of H satisfying the relation 1

(15) f!ll<x - aly - a) ~ 0

for all y E H.

a a).~i~~~
H-

+ A(y -

FIG. 2.

Put d = inf I x - y II, and for every integer n > 0, let Hn be the set of points y
YEH
of H such that I x - y I ~ d + n - 1. The set Hn is closed in H, is convex and non-
empty, and its diameter is bounded by J12 din for all large enough n, by prop. 5.
The sequence (Hn)n;. 1 being decreasing, and the set H being Hausdorff and complete
it follows that the base of the Cauchy filter (Hn)n;. 1 converges to a point PH(X) of H;
we have {pix)} = n H n , hence pix) is the unique point a of H such that
n~ 1
Ilx - all = d.

1 We recall (GT, VIII, § I, No. I) that ~(z) denotes the real part of the complex number z;
we have ~(z) = z if z is real.
No.5 PREHILBERTIAN SPACES AND HILBERTIAN SPACES TVS V.l1

Let Y E H; since H is convex, the point z(A) = PH(X) + A(Y - PH(X)) of E belongs
to H for every real number A such that 0 < A < l. Hence we have

which gives

Conversely, let a be a point of H such that f1ll x - a IY - a) < ~ 0 for all Y E H.

F or every Y E H, we have

Ilx - yI1 2 = Ilx - al1 2 + Ily - al1 2 - 2f1ll<x - aly - a) ;?; Ilx - a11 2 ,

and so II x - a II = d and finally that a = PH(X) follows from the first part of the
proof. Q.E.D.
In what follows the mapping PH of E in H will be called the projection from E
onto H. We remark that PH(X) = x for all x E H.
The first part of tho 1 is valid under more general hypotheses on the space E (Y, p. 67,
exerc. 31).

The proof of tho 1 establishes, among others, the following property :

COROLLARY l. - Let I be a set directed by afilter lJ and let (Y)iEI be afamily of points
of H. Let x E E. Suppose that we have

lim-Ilx - Yill = inf Ilx - zil .

i,15' ZEH

Then Yi tends to pix) with respect to the filter lJ.

COROLLARY 2. - For every x, Y in E, we have

In particular, the mapping PH from E into H is continuous.

Let x, y be two points ofE. Put a = PH(X) - x, b = PH(Y) - PH(X), c = Y - PH(Y)'
<
By formula (15) (V, p. 10) we have f1ll alb) ;?; 0 and f1ll clb) ;?; O. We also have <
a + b + c = Y - x, which gives,

Ilx - yI1 2 = Iia + b + cl1 2 = IIbl1 2 + + cl1 2 + 2f1ll<alb) + 2f1ll<clb)

Iia

;?; IIbl1 2 = IlpH(X) - PH(Y) 112 .

This proves corollary 2.

PROPOSITION 6. - Let E be a prehilbertian space and let <D be a non-empty, directed

decreasing set of non-empty Hausdorff and complete convex subsets of E. For every
x E E and every subset H of E, put d(x, H) = inf II x - z II. In order that the inter-
ZEH
TVS V.12 HILBERTIAN SPACES §1

section M of the sets H belonging to <1> be non-empty, it is necessary and sufficient

that there exists X o in E such that sup d(xo, H) is finite. For every x E E we then have
HEel>
PM(X) = lim PH(X) (limit with respect to the directed set <1».
HE <1>
If M is non-empty, d(x, H) :::;: d(x, M) for all H E <1> and all x E E.
Conversely, suppose that there exists a point Xo in E and a real number C ~ 0
such that d(xo' H) :::;: C for all HE <1>. Let x E E; then

d(x, H) :::;: II x - Xo II +C for all HE <1> ,

hence the number d = sup d(x, H) is finite. Let B be the set of all Z E E such that
HEel>
II x z II :::;: d. Since B is convex and closed in E, the sets H n B, for H ranging
over <1>, are convex, Hausdorff and complete. Let 8 > 0; there exists a set H E <1>
such that d(x, H) ~ d - 8, and if 8 < d12, the diameter of H n B is bounded by
J12 8(d - E) by prop. 5 (Y, p. 9). In other words, for all Ho E <1>, the closed sets
H n B, for HE <I> and H c Ho, form a base of the Cauchy filter on the Hausdorff
and complete space Ho. Hence the intersection of the sets H n B (for H E <1» reduces
to a point y. We get Y E M and II x - y II = d = d(x, M). Since M is closed in H o ,
it is a Hausdorff, convex and complete set in E, and so Y = PM(X). For every HE <1>,
we have PH(X) E H n B, from which we get that PM(X) = limPH(x).
HEel>

PROPOSITION 7. - Let E be a Hausdorffprehilbertian space and let 'P be a non-empty

directed increasing set of non-empty, convex, complete subsets of E. Put A = U H
HEtIJ
and suppose that the closure N of A is complete. Then N is convex and we have
PN(X) = lim PH(X) for all x E E.

It is clear that A is convex, hence its closure N is convex (II, p. \3). With the nota-
tions of prop. 6, d(x, N) = inf d(x, H), and consequently d(x, N) is the limit of
liE'!'
d(x, H) with respect to the section filter of 'P. Since PIlCX-) E Hand

lim II x - PH(X) II = lim d(x, H) = d(x, N) ,

HE'!' HE'!'

it follows from cor. 1 of V, p. 11 that PH(X) tends to the projection PN(X) of x onto
N with respect to the section filter of 'P.

6. Vector subspaces and orthoprojectors

Let E be a prehilbertian space. Recall that two vectors x and y of E are said to
<
be orthogonal if xly) = 0; then

(16)

(<< Pythagoras' theorem »).

No. 6 PREHILBERTIAN SPACES AND HILBERTIAN SPACES TVS V.13

Let A be a subset of E. We say that a vector x in E is orthogonal to A if it is ortho-

gonal to every vector of A. The set of all vectors orthogonal to A is a closed vector
subspace of A, denoted by A° and called (by abuse of language) the orthogonal of A.
Let A and B be two subsets of E. We say that A and B are orthogonal if every
vector of A is orthogonal to every vector ofB. This is equivalent to saying that A c BO,
or that B c AO. If E is Hausdorff and if A and B are orthogonal then A n B is empty
or reduces to 0 since 0 is the only vector of E orthogonal to itself
THEOREM 2. - Let E be a prehilbertian space and M a vector subspace of E, which
is Hausdorff and complete. Then E is the topological direct sum of M and of MO the
subspace orthogonal to M. The projector from E onto M associated with the decom-
position E = M EEl MO is the projection PM from E onto M defined in tho 1 (V, p. 10).
We first show that x - PM(X) belongs to MO for all x E E. Let y E M. For every
scalar Ie E K, the vector PM(X) + ley belongs to M; hence by formula 15 (V, p. 10)
we have,

for all Ie E K. If, in particular we take Ie = <x - PM(x)ly) we conclude that

< >
x - PM(X) Iy = 0, hence our assertion.
Since M is Hausdorff, 0 is the only vector of M, orthogonal to itself, hence
MnMo={O}. For every xEE, we havePM(x)EM and X-PM(X)EMo. Conse-
quently, E is the direct sum of M and M and PM is the projector from E onto M
O
,

with kernel MO. Since PM is a continuous mapping from E into M (V, p. 11, cor. 2),
if follows from GT, III, § 6, No.2 that E is the topological direct sum ofM and MO.

COROLLARY. - Let E be a Hausdorff prehilbertian space and M a finite dimensional

vector subspace of E. Then E is the direct sum of M and MO.
Since E is Hausdorff, so is M; since M is finite dimensional, it is complete (I,
p. 13). It is therefore enough to apply tho 2.
With the notations of tho 2, we say that MO is the orthogonal complement of M
and that PM is the orthoprojector (or the orthogonal projector, or by abuse oflanguage,
the projector) from E onto M; if x is a vector of E, the vector PM(X) of M is also
called the orthogonal projection of x on M. Note that PM is a continuous linear map-
ping from E onto M and that we have IlpM II = 1 by cor. 2 of V, p. 11, except in the
case when M = {O} in which case PM = O.
It follows immediately from Pythagoras theorem that the canonical mapping ljJ
from ElM onto MO deduced from the direct sum decomposition E = M E8 MO is iso-
metric if ElM is assigned the quotient semi-norm from that of E (II, p. 4). We shall
always assign that prehilbertian structure to ElM for which \jJ is an isomorphism
of prehilbertian spaces; the quotient semi-norm on ElM is then deduced from this
prehilbertian structure.
We shall often use the preceding results when E is a hilbertian space and M a
closed vector subspace of E. In this case, MO is a closed vector subspace of E, and
PM' = 1 - PM' and (MO)O = M.
TVS V.14 HILBERTIAN SPACES §1

PROPOSITION 8. - Let E be a hilbertian space, M a closed vector subspace of E,

I a non-empty ordered directed set and (M)iEI a family of closed vector subspaces ofE.
We assume that either the mapping i f--+ Mi is increasing and that M is the closure
of U Mi or that the mapping i f--+ Mi is decreasing and that M = n
Mi. Then
iEI iEI
PM(X) = lim PMi(X) for all x E E.
iEJ
Prop. 8 follows immediately from props. 6 (V, p. 11) and 7 (V, p. 12).

PROPOSITION 9. - Let E be a hi/bertian space and M, N two closed vector subspaces

ofE.
a) The following conditions are equivalent:
(i) PMPN = PNPM;
(ii) if x E M is orthogonal to M n N and if YEN is orthogonal to M n N, then
x and yare orthogonal;
(iii) every vector of M orthogonal to l' n N is orthogonal to N;
(iv) M = (M n N) + (M n N°).
b) If the equivalent conditions of a) are satisfied, we have PMnN = PMPN' the vector
subspace M + N of E is closed and we have PM+N = PM + PN - PMPN.
c) We have PMPN = 0 if and only if M is orthogonal to N. If this is so, then the
vector subspace M + N of E is closed, and PM + N = PM + PN.
Put L = M n N, Ml = M n LO and Nl = N n LO. Condition (ii) implies that
Ml and N 1 are orthogonal, and (iii) implies that Ml and N are orthogonal. Since
we have N = Nl + Land Ml is orthogonal to L, we have proved the equivalence
of (ii) and (iii). If condition (iii) is satisfied, we have Ml = M n N° and since
M = L + M 1 , condition (iv) is satisfied. Conversely, from (iv) we conclude that
Ml = M n N° since the subspaces M n Nand M n N° of M are orthogonal,
and so Ml c N°, that is, the relation (iii).
Assume that condition (iv) is satisfied. It is immediate that PN(y) = hey) for all
y E M and hence PNP~X) = hP~x) for all x E E. But, for every x E E, the vector
hPM(X) belongs to L, and the vector

belongs to MO + U = LO; hence wehavehPM(x) = hex). FinallY,PNPM = PLPM = PL"

Since condition (ii) is equivalent to (iv) and is symmetric in M and N, we also have
PMPN = h· Finally we get PMPN = PNPM = PMnN which gives (i).
Conversely, suppose condition (i) is satisfied. Let x EM; we have

and so pJx) E M. We conclude that x - pJx) E M, hence x is the sum of an ele-

ment pJx) of M n N and an element x - PN(X) of M n N°, which gives (iv).
We have proved a) and the first part of b). Assume now that PM andpN commute
and put q = PM + PN - PMPN; since PM and PN are idem po tents in the algebra
.P(E), so is q; hence (OT, III, § 6, No.2) the image of q is a closed vector subspace of E.
No.7 PREHILBERTIAN SPACES AND HILBERTIAN SPACES TVS V.15

It is clear that the image of q is contained in M + N; however, we have pJx) = x,

hence q(x) = x for all x EN; since we also have q = PM + PN - PNPM' we get
q(x) = x for all x EM. We conclude that the image of q is equal to M + N. The
orthogonal of M + N is equal to MO n N°, and the kernel of q obviously contains
MO n N°, hence q = PM+N. This proves b).
We have PMPN = 0 if and only if the image N of PN is contained in the kernel MO
of PM' that is, if and only if M is orthogonal to N. The rest of the assertion c) is then
a particular case of b).
Remark. - Let E be a hilbertian space and M, N two closed vector subspaces of E.
The relation MeN is equivalent to the orthogonality of M and N°, that is to say,
to the relation PMPN0 = 0 by prop. 9, c). Since we have Pw = 1 - PN' we conclude
that the relations MeN and PM = PMPN are equivalent (<< the three perpendicular
theorem », cf. fig. 3).

7. Dual of a hilbertian space

THEOREM 3. - Let E be a hilbertian space. For every x E E, let x* be the continuous

linear form y f--* < xly-> on E; the mapping x f--* x* is a bijective, semi-linear (for the
automorphism ~ f--*~) mapping from E onto its dual E', and an isometry from the
normed space E onto the normed space E'.
The mapping x f--* x* is semi-linear by (2) (V, p. 1) and by virtue of the Cauchy-
Schwarz inequality, we have Ilx*11 = sup l<xly)1 = Ilxll, hence x f--* x* is an
II yll < 1
isometry from E into E', and in particular, is injective. To complete the proof, we
need to prove that for all x' i= 0 in E', there exists x E E such that x' = x*. But
the hyperplane H = Ker x' is closed in E; its orthogonal is a line D. Let b be a
non-zero element of D; the kernel of the linear form.!?* is equal to H and hence
there exists a scalar A i= 0 such that x' = A. b* = (A. b)*. Q.E.D.
The mapping x f--* x* from E onto its dual E' is said to be canonical. The inverse
mapping from E' onto E is also called canonical and is denoted by x' f--* x'*. We
have
(17) <xly) = <y, x*), <x, x') = <x'*lx)

for x, y in E and x' in E'. Also (x*)* = x for x E E.

TVS V.16 HILBERTIAN SPACES §1

When K is R, the mapping x f-+ x* is linear. We shall transfer the scalar product
of E to E' by this mapping. When K =: C.!.. we can consider the mapping x f-+ x*
as an isomorphism from the vector spa.£e E, the conjugate of E onto E' (V, p. 6).
We shall transfer the scalar product of E to E' by this mapping.
In the two cases considered, E' is a hilbertian space and we have the formulae

<x*IY*) = <xly), <x'lx') = IIx'I1 2

for x, y in E and x' in E'.
To say that the vector x E E is orthogonal to a vector Y E E is equivalent to saying
that the linear form x* E E' is orthogonal to y in the sense defined in II, p. 41 (this
justifies the use of the word « orthogonal» in the two cases). If M is a closed vector
subspace of E, the subspace MO orthogonal to M in E' (II, p. 44) is the image under
x f-+ x* of the orthogonal of M in E, defined in V, p. 13 (this justifies the use of the
notation MO in the two cases).

CoROLLARY l. - In order that the family (Xi)iEI of points of a hilbertian space E

<
be total, it is necessary and sufficient that the relations Xi Iy) = 0 for y E E and for
all indices i E I imply that y = O.
In fact, this says that 0 is the only vector of E' which is orthogonal to all the Xi
(II, p. 43 and IV, p. 1).

CoROLLARY 2. - Let E and F be two hilbertian spaces. For u E Y(E; F), x E E

and y E F, put

(18) <l>Jy, x) = <ylU(x» .

The mapping u f-+ <l>u is an isomorphism from the Banach space Y(E; F) onto the
space of all continuous sesquilinear 1 forms on F x E, endowed with the norm

(19) Ilfll = sup If(y, x)l·

xeE,yeF
Ilxll';; 1,1IY11';; 1

It is clear that <l>u is sesquilinear and continuous for all u E Y(E; F). Conversely,
let f be a continuous sesquilinear form on F x E. For every x E E, the mapping
y f-+ f(y, x) is a continuous linear form on the hilbertian space F. By tho 3, for every
x E E, there exists a unique element u(x) in F such that f(y, x) = u(x) Iy) for all <
y E F. The mapping u: x f-+ u(x) from E into F is linear and we have

Ilfll sup sup If(y, x)1 = sup sup 1< ylU(x» I

Ilxll';;l lIyll';;l IIxll';;l Ilyll';;l

sup II u(x) II ;
IIxll';; 1

hence u belongs to Y(E; F), f = <l>u and Ilull = Ilfll. This proves cor. 2.

1 Recall (A, IX, § 1, No.5) tha.t a sesquilinear form (on the left) f on F x E is a mapping
from F x E into K which satisfies relations (1) and (2) of V, p. 1.
No.1 ORTHOGONAL FAMILIES IN A HlLBERTIAN SPACE TVS V.17

The canonical mapping from E into its bidual E" (IV, p. 14) maps E onto E",
in other words (IV, p. 16), E is a reflexive Banach space. In fact, if E is a real (resp.
complex) hilbertian space, the canonical mapping <jl from E' onto E is an isomor-
phism from the normed ~ace E' onto E (resp. onto the conjugate space E of E);
applying tho 3 to E (resp. E), we see that every continuous linear form on the normed
space E' is of the form x' f---* <<jl(x') Ix >= <x, x' >with x E E, hence our assertion
follows.
As a consequence (IV, p. 17, prop. 6) :
THEOREM 4. ~ In a hilbertian space E, the unit ball is weakly compact.
PROPOSITION 10. ~ If, in a hilbertian space E, a filter ~ converges weakly to x o,
and ifmoreover lim\] Ilxll = Ilxoll, then ~ converges to xofor the initial topology ofE.
In fact, Ilx - x ol1 2 = IIxl12 - 2f2ll <xlxo > + Ilx o l1 2. Since <xlxo > tends to IIxol12

°
with respect to ~ by hypothesis, and Ilxll tends to Ilxoll with respect to ~, Ilx - xoll
tends to with respect to ~, hence the proposition.

Remark. - If E is a Hausdorff prehilbertian space and E the hilbertian space comple-

tion of E, we know (III, p. 16) that the dual E' of E can be identified with the dual of E;
it then follows from tho 3 (V, p. 15) that every continuous linear form on E can be written
in a unique way as x I-> <alx), where a E E.

§ 2. ORTHOGONAL FAMILIES IN A HILBERTIAN SPACE

1. External hilbertian sum of hilbertian spaces

PROPOSITION 1. ~ Let (E)iEI be a family of hilbertian spaces, P the product vector
space n
Ei and E the subset ofP consisting of all families x = (X)iEI such that I I Xi 112
~ ~

is finite.
a) E is a vector subspace of P.
b) For every x = (X)iEI and y = (Y)iEI in E, the family «xiIYi»iE[ is summable.
If we put <xly > = I <Xi IYi >, we define a positive separating hermitian form on E.
iEI
c) For the scalar product so defined, E is a hilbertian space; the direct sum S of the
Ei is dense in E.
For x = (X)iEI and y = (Y)iE[ in E, we have

hence x + y = (Xi + Y)iEI belongs to E. This proves a).

By the Cauchy-Schwarz inequality, we have

l<x;lYi>1 '-S; IlxJ·lIyill '-S; 1(lI x dl 2 + IIYir)

hence L l<xdYi>1 < + 00. If X :f- 0, we have <xix) = I IIxdl2 > 0, hence asser-
~ ~

tion b) follows.
TVS V.18 HILBERTIAN SPACES §2

We recall that S is the subspace of P consisting of all families x = (X)iEI such

that the set of all i E I for which Xi =I- 0 is finite. It follows immediately that S is
dense in E; hence it remains to prove that E is complete for the topology 51 obtained
by the norm Ilxll = <XIX)1/2. Let 52 be the topology induced on E by the product
topology on TI E i · For every r > 0, let Br be the set of all X E E such that Ilxll ~ r.
iEI
This relation implies that we have L IIxil12 ~ r2 for every finite subset J of!, and so
ieJ
Br is a closed subset of TI E i , hence also complete. The fact that E is complete for
iel
51 now follows from GT, III, § 3, No.5, cor. 2 to prop. 10.

DEFINITION 1. - Let (E)iEI be a family of hilbertian spaces. The hilbertian space E

defined in prop. 1 is called the external hilbertian sum of the family (E)iEI and written
as CD Ei or EB Ei 1.
ieI ieI
Let J; be the mapping from Ei into E which transforms Z E Ei into an element
(xk ) E E such that X k = 0 for all k =I- i and Xi = z; it is clear that J; is an isomorphism
from the hilbertian space Ei onto a closed vector subspace of E. We say that J; is
the canonical mapping from Ei into E and we shall generally identify Ei with its
image in E by this isomorphism. With this convention, Ei and Ek are orthogonal
in E for i =I- k, and E is the closed vector subspace generated by the union of the
subspaces E i •
When I is finite, E is the direct sum of the E i ; since the canonical projector from
E onto Ei is continuous for all i E I, E is also the topological direct sum of the Ei
(GT, III, § 6, No.2, prop. 2). If I = (1, n), we also write E1 EB E2 EB ... EB En instead of
n

CD Ei ·
i= 1

Example. - Let E be a hilbertian space and I a set of indices. Let f~(I) denote the
external hilbertian sum of the family (Ei)iEI where Ei = E for all i E I. In other words,
e~(1) is the space of all families x = (X)iEI of elements ofE such that I IIxil12 < +00,
ieI
endowed with the scalar product <xly >= I <X;lYi > (space of square summable
ieI
families of elements of E indexed by I). We put e2(1) = e~{I).

2. Hilbertian sum of orthogonal subspaces of a hilbertian space

DEFINITION 2. - A hilbertian space E is said to be a hilbertian sum of a family (E)iEI

of closed vector subspaces of E when :
1) for two distinct indices i, k in I, the subspaces Ei and Ek are orthogonal in E;
2) the closed vector subspace generated by the union of the Ei is E.

1 Care must be taken not to confuse this notation with that of the « algebraic» direct
sum of the spaces E; (A, II, § 1, No.6).
No.2 ORTHOGONAL FAMILIES IN A HILBERTIAN SPACE TVS V.19

THEOREM 1. - Let E be a hilbertian space which is a hilbertian sum of a family (E)iEI

of closed vector subspaces ofE. There exists an isomorphism f and only one, from E
onto the external hilbertian sum CD Ei = F of the family (E) such that, for all i E I,
the restriction of f to E is the canonical mapping 1; from Ei into F.
Let S c F be the « algebraic» direct sum of the E i , and let 9 be the linear mapping
(X)iEI H L Xi from S into E. We shall show that 9 is an isomorphism from the pre-
hilbertian space S onto the (prehilbertian) subspace g(S) of E, generated by the union
of the Ei : for, for two elements x = (X)iEI' Y = (Y)iEI' we have

<g(x)lg(y) = <I xiiI

iEI iEI
Yi) = L
(i.kjEI x I
<xiIYk)·

But if i =1= k, < Xi IYk) = 0 by hypothesis, hence

<g(x)lg(y) = L <xiIYi) = <xly);

iEI

this proves our assertion. Since S is dense in F and g(S) dense in E, the isomorphism 9
extends to an isomorphism g from F onto E (V, p. 8, cor.). It is clear that the inverse
isomorphism f of g is the required mapping; its uniqueness follows from the fact
that the closed subspace of E generated by the union of the Ei is E itself.
When E is the hilbertian sum of a family (EJiEI of subspaces, we shall often identify E
with the external hilbertian sum F of the Ei by means of the isomorphism .f If the set
I is finite, saying that E is the hilbertian sum of the family (EJiEI means that the Ei are
two by two orthogonal and that the vector space E is the direct sum of the family (EJiEI
of subspaces.

COROLLARY 1. - Let E be a hilbertian space, which is a hilbertian sum of a family

(E)iEI of closed vector subspaces ofE ; for all i E I, let PE, be the orthoprojector (V, p. 13)
from E onto E i .
a) For all x E E, the family (1IpE(X)11 2)iEI is summable in R, the family (PE,(X))iEI
is summable in E, and we have

IIxI12 = L II PEi(X) II 2 , X = LPE/ X).

iEI iEI

b) Conversely, if (X;)iEI is a family of elements of E such that Xi E Ei for all i E I

and I II Xi 112 < + 00, this family is summable, and the sum x is the only point of E
iEI
for which PEi(X) = Xi for all i E I.
c) For every pair of points x, y of E, we have

<Xly) = I
iEI
<PEi(X)lpE/y)·

These properties are in fact obvious for the external hilbertian sum of the E i ,
and can be transferred to E by isomorphism.
TVS V.20 HILBERTIAN SPACES §2

COROLLARY 2. - Let E be a HausdorjJprehilbertian space, (E)iEi afamily of complete

vector subspaces ofE such that, for every pair of distinct indices i, k in I, the subspaces
Ei and Ek are orthogonal. Let V be the closed vector subspace of E generated by the
union of the E i . For every i E I, let PE, be the orthoprojector from E onto E i . Let x E E.
1) We have I IlpE,(x)112 :::; Ilx112.
ieI
2) The following conditions are equivalent : a) x E V; b) I I PE,(X) 112 = I X 112 ;
ieI
c) the family (PE,(X))iEi is summable in E, and we have x = I PE,(X).
ieI
3) Suppose V is complete. Then the family (PE,(X))iEi is summable in E, and

py(X) = IPE'(X) , I py(x)112 = I IIPE,(X) I 2 ,

ieI ieI

where Pv denotes the orthoprojector from E onto V.

Let E be the hilbertian space completion of E; we identify E with a dense sub-
space of E; the Ei , being complete, are closed subspaces of E. The closure V of V
in E is the closed vector subspace ofE generated by the union of the E i , and V = VI i E.
The space E is the hilbertian sum of the Ei and of the subspace W, the orthogonal
complement of V in E; put Xo = pwCx) and Xi = PE(X) for all i E I. By cor. 1, we
have IIxI12 = IIxol12 + I Ilxil12, and x = Xo + I Xi i~ E. This implies assertion 1),
ieI ieI
and the fact that conditions b) and c) of 2) are equivalent to the condition Xo = 0,
hence to the condition x E V. Finally, if V is complete, and if we put x' = py(x);
we have x' - Xi = (x - x) - (x - py(x)), hence x' - Xi is orthogonal to E i ,
and so Xi = PE,(X') for all i E I; it is now enough to apply property 2) to the vector x'.
Remark. - Let E be a Hausdorffprehilbertian space. (V)iEI a family of vector subspaces
of E such that for every pair of distinct indices i. k. the subspaces Vi and V k are ortho-
gonal. Then. for every k E I. the intersection of Vk and of the closed vector subspace W k
generated by the union of the Vi for all i =1= k reduces to 0 for. if x belongs to V k and
also to Wk. then it is orthogonal to all the Vi for i =1= k. hence to Wk' In particular.
x is orthogonal to itself. hence is zero.

PROPOSITION 2. - Let E be a hilbertian space and (V))AEL a family of closed vector

subspaces of E ; for every A E L, let (W A1J IlEM l be a family of closed vector subspaces
of VA such that VA is the closed vector subspace generated by the union of this family.
In order that E is the hilbertian sum of the family (WAIl)AEL.IlEM", it is necessary and
sujJicient that E is the hilbertian sum of the family (VA)AEL and that, for each A E L,
V Ais the hilbertian sum of the family (W A)IlEM l (<< associativity of the hilbertian sum »).
To show that the condition is necessary, it is enough to see that V~ and V pare
orthogonal if CI. =I p. But, every element of W~1l (ll EM,,) is orthogonal to all the
W pv (v E Mp), hence to the closed vector subspace V p which they generate; the same
argument then shows that every element of V p is orthogonal to V~, being ortho-
gonal to all the W~1l (ll EM,,).
To show that the condition is sufficient, it is enough to verify that, if it is satisfied,
E is equal to the closed vector subspace F generated by the union of the WAil (A E L,
No.3 ORTHOGONAL FAMILIES IN A HILBERTIAN SPACE TVS V.21

Il E MI.); but, for each A E L, F contains the closed vector subspace generated by
the union of the WAJl such that Il E M A, that is, F contains VA; hence F is the closed
vector subspace generated by the union of the VA' which is E by hypothesis.

3. Orthonormal families

DEFINITION 3. - In a prehilbertian space, afamily (e);EI of vectors is said to be ortho-

gonal if ei and ek are orthogonals for all i =I- k, and is said to be orthonormal, if in
addition I e; I = I for all i E I.
A subset S of E such that the family defined by the identity mapping from S onto
itself is orthonormal is said to be an orthonormal set. If (e)iEI is an orthonormal
family, the mapping i H ei is injective; we can then talk indifferently of an ortho-
normal family or an orthonormal set.
If (e)iEI is an orthonormal family, the complete one dimensional vector sub-
spaces Di = Ke; are two by two orthogonal. For every x E E, the orthogonal pro-
jection of x on D; is A;e; with <ei Ix - Aiei > = 0, which gives <eiIx >= A;< e; lei> = Ai'
The results of No.2 applied to the subspaces D; imply the following propositions:

PROPOSITION 3. - In a Hausdorff prehilbertian space E, every orthonormal family

is topologically independent.
We note that this property immediately follows from the characterization of topo·
logically independent families (IV, p. 1 and II, p. 43, cor. 2), on account of the identifica·
tion of the dual ofE with the completion ofE or with the space conjugate to E according
as K is equal to R or C (V, p. 17, Remark).

PROPOSITION 4. - Let E be a Hausdorff prehilbertian space, (e)iEI an orthonormal

family in E, V the closed vector subspace of E generated by the ei .
1) For every x E E, we have

(1) L l<e;lx>12 ~ IIxl12

iEI

(Bessel's inequality); here the set of all i E I such that ei Ix =I-

over, the following conditions are equivalent : a) x E V; b)
< > °I xl12
is countable. More-
= I 1< ei Ix >12 ;
ieI
c) the family <e;lx >. ei is summable in E, and x = I <ei Ix >. ei .
< >.
ieI
2) If V is complete, then the family of all e; Ix ei is summable in E for all x E E,
and I
<e;lx>.e i = py(x), L
1< ei lx>12 = IIpy(x)V
ieI ieI
3) Suppose V is complete. For every family (A)iEI of scalars such that I IA; 12 < + 00,
ieI
< >
there exists a unique point x E V such that ei Ix = Ai for all i E I. If(Il);EI is a second
family of scalars such that L IIlil 2 < + 00, and if y E V is such that <ei Iy > = Ili
ieI
for all i E I, then <xly >= L ~illi'
iEI
TVS V.22 HlLBERTIAN SPACES §2

PROPOSITION 5. - Let (e)iEI be an orthonormal family in a Hausdorff prehilbertian

space E. The following properties are equivalent:
a) the family (e) is total;
b) for every x E E, the family <
eJr). e i is summable in E, and we have
x = L <eilx).ei ;
ieI
c) for every x E E,

(2) IIxl12 = L 1< ei lx)12

ieI

(Parseval's relation).
When E is hilbertian these conditions are also equivalent to
d) the relations <ei Ix) = 0 for all i E I imply that x = O.
The equivalence of conditions a), b), c) follows immediately from prop. 4. When
E is hilbertian, the equivalence of conditions a) and d) follows from cor. 1 of V,
p. 16.

DEFINITION 4. - An orthonormal and total family in a Hausdorff prehilbertian space

E is called an orthonormal basis ofE.
An orthonormal basis of a Hausdorff prehilbertian space E is also an orthonormal
basis of the completion of E.
Let (e)iEI be an orthonormal basis of E; for every x E E, the numbers <ei Ix)
are called, by abuse of language, the coordinates of x with respect to the basis (eJ
F or every x and y in E, we have

(3) <xly) = I <eilx) <eiIY)·

ieI

An orthonormal basis of E is not, in general, a basis of E over the field K in the sense
defined in A, II, p. 25; to avoid any confusion we shall always say that a basis of a
prehilbertian space E in the sense of lac. cit. is an algebraic basis of E over K.

Let E and F be two Hausdorff prehilbertian spaces and u a continuous linear

mapping from E into F. Let (e)iEI (resp. (f)jEJ) be an orthonormal basis ofE (resp. F).
Put
uji = <.t;Iu(e)
for i E I, j
E 1. The family (uj)(i.j)EI x J is called the matrix of u with respect to the ortho-
normal bases (e) and (1). Let x E E and y = u(x); if we write ~i = <ei Ix) and

J
<
11· = .l;ly) for the coordinates of x and y respectively, we get 11j = L Uji~i for
~

all j E 1. When (e) is an algebraic basis of E and (.I;) an algebraic basis of F, our
definition is consistent with that of A, II, § 10, No.4.
Example. - Let E be the space of all complex valued continuous functions on R,
such that f(x + n) = f(x) for x E Rand n E Z. We assign to E the scalar product

f
defined by

<fig) = f(t)g(t)dt.
No.4 ORTHOGONAL FAMILIES IN A HILBERTIAN SPACE TVS V.23

Then E is a Hausdorff prehilbertian space, but is not complete. For every integer
nEZ, let e.,(x) G' e(nx). It is immediate that the family (e")"EZ is orthonormal in E.
Moreover, the topology of uniform convergence on E is finer than the topology deduced
from the norm IIfl12 = <flf)1(2. The family (e")"EZ is total in E for the uniform con-
vergence (G T, X, ~ 4, No.4), and a fortiori in the prehilbertian space E. Hence (en).,EZ is
an orthonormal basis of E.

4. Orthonormalisation

THEOREM 2. - For every orthonormal set L in a hilbertian space E, there exists an

orthonormal basis B of E containing L.
In fact, let Tl be the family of all orthonormal subsets of E, linearly ordered by
inclusion; it is immediate that this family has finite character (S, III, § 4, No.5). Hence
there exists a maximal family B in Tl containing L, by tho 1 ofS, III, § 4, No.5. It remains
to prove that B is a total set. If not, there will exist a vector y#-O which is orthogonal
to all the vectors of B (V, p. 22, prop. 5), and multiplying y by a suitable scalar, we
can assume that I y I = 1; then, B u {y} will be an orthonormal set distinct from
B and containing B; this contradicts the definition of B; hence the theorem.

COROLLARY l. - In every hilbertian space, there exists an orthonormal basis.

It is enough to apply tho 2 to the case L = 0.

COROLLARY 2. - Every hilbertian space is isomorphic to a space eZ (1).

More precisely, let (e)iEI be an orthonormal basis of a hilbertian space E. By
props. 4 (V, p. 21) and 5 (V, p. 22), the mapping <p defined by
(4)
is a hilbertian space isomorphism from E onto fZ(I). The inverse isomorphism ~
is defined by
(5) ~((A)iEI) = I Aiei .
iEI

PROPOSITION 6. - Let E be a Hausdorff prehilbertian space, and let (an)nEI (I an

interval ofN with origin 1) be a countable (finite or not) independent family of vectors
ofE. There exists an orthonormal family (en)nEI' and only one, in E, with the following
properties :
1) for every integer pEl, the vector subspace of E generated by el' e z , ... , e p is
identical with the vector subspace of E generated by ai' a z , ... , a p ;
< >
2) for every integer pEl, the number aplep is real and> O.
In fact, let Vn be the subspace (of dimension n) generated by al' a z , ... , an' If
n + 1 E I and bn+1 = an+1 - pVn(an+1) (where PV n is the orthoprojector onto the
complete subspace V n), the line Kb n+1 is the orthogonal of Vn in V n+ l ' If the en
satisfy condition 1) of the proposition, we must have en + 1 = Abn + 1; the condition
Ilen + 1 1 = 1 then implies IAlz Ilbn + 1 1l z = 1 and the condition <an+1Ien+1> > 0
<
implies A an + 11bn+ 1 > > 0; this completely determines A, and we have proved
TVS V.24 HlLBERTIAN SPACES §2

that we can determine, by induction, an orthonormal family (e ll)IIEI' and only one,
so as to satisfy conditions 1) and 2) of the proposition.
The sequence (en )IIEI is said to be obtained by orthonormalisation from the inde-
pendent family (an )IIEI. It is clear that the vector subspace generated by the family
(en) is identical with the vector subspace generated by the family (an). In particular,
if (an) is a total sequence, so is (en)' which is then an orthonormal basis of E; hence
we get:

COROLLARY. - In every Hausdorff prehilbertian space E satisfying the first axiom

of counlability, there exists a countable orthonormal basis.
If E satisfies the first axiom of countability, then there exists a total sequence in E,
and we can always extract an independent total family from such a sequence (A, II,
§ 7, No.1, tho 2).
We can give examples of Hausdorlf prehilbertian spaces not having any orthonormal
basis (V, p. 70, exerc. 2).
Example. - Let I be the interval (- I, 1) of Rand E the vector space of real valued
continuous functions on I. Let x denote the canonical injection from I into R, consi-
dered as an element of E. By the Stone-Weierstrass theorem the sequence (X")nEN is
total in E for the topology of uniform convergence GT, X, § 4, No.2).
Consider E as a real Hausdorlf prehilbertian space in which the scalar product is
given by

<fIg) = fl J(t) get) dt.

The sequence (xn)nEN is then total in the prehilbertian space E. Let (IIn)nEN be the sequence
obtained by the orthonormalisation of the sequence (xn)nEN. We can show that
lIn = (n + t)1/2Pn, where the Legendre polynomial P n is defined by

PROPOSITION 7. - In a hilbertian space E, two orthonormal bases are equipotent.

Let Band C be two orthonormal bases of E. The case when one of the two sets
B, C is finite is trivial, since a finite orthonormal basis is an algebraic basis of the
space. Suppose therefore that Band C are infinite. For every x E B, let C x be the
< >
subset of C consisting of all y E C such that xly =P O. The set C x is countable
(V, p. 21, prop. 4). For every y E C, there exists x E B such that y E Cx' since B is an
orthonormal basis and y =P 0; in other words C is the union of the countable sets
C x as x ranges over B. The cardinality of C is hence less than that of N x B, hence
less than that of B (S, III, § 6, No.3, cor. 4); similarly, the cardinality of B is less
than that of C; this completes the proof
The cardinality of an arbitrary orthonormal basis of a hilbertian space E is called
the hilbertian dimension of E.

. COROLLARY1. - Given two orthonormal bases in a hilbertian space E, there exists

an automorphism of E transforming the first basis into the second
No.1 TENSOR PRODUCT OF HILBERTlAN SPACES TVS V.25

COROLLARY 2. ~ In order that the hilbertian spaces £2(I) and £2(1) be isomorphic,
it is necessary and sufficient that I and J are equipotent.

§ 3. TENSOR PRODUCT OF HILBERTIAN SPACES

1. Tensor product of prehilbertian spaces

Let El and E2 be two prehilbertian spaces and let F = El ® E2 be the tensor

product of the vector spaces E j and E 2 . Let Xl EEl and x 2 E E 2 ; since the mapping
(Yl' Yz) ~ xllYl < ><
x 2 1Y2) from El x E2 into K is bilinear, there exists a linear
form <PXIoX2 on El ® E2 such that

(1)

for Yl EEl and Y2 E E 2· Let z E F. The mapping (Xl' X 2) ~ <PXIoX,(Z) from El x E2

L
/I

into K is bilinear; this can be seen by writing z in the form z = Yi.l ® Yi.2 with
i= j

Yi.l EEl andYi.2 E E2 for 1 '-'S i '-'S n. Then there exists a 1inearform \j!z on F =E l ® E2
such that
(2)

We put <D(z, t) = \j!z(t) for z, t in F. We see immediately that <D is a sesquilinear

form on El ® E2 characterized by

(3)

(cl A, IX, § 1, No. 11).

PROPOSITION 1. ~ The sesquilinear form <I> on El ® E2 is hermitian and positive,
hence assigns the structure ofa prehilbertian space to El ® E 2 • This space is H ausdorJJ
irE l and E z are Hausdorff.
The formu1a<D(z, t) = <D(t, z) follows from (3) when z = Xl ® x 2 and t = Yl ® Yz.
The general case is obtained by linearity, hence <I> is hermitian.
Suppose El and E2 are Hausdorff; we shall prove that the hermitian form <I> is
n
positive and separating. Let z = L Xi ® Yi be a non-zero element of F = El ® E 2 •
i= 1
Let (e l' ... , en) be an orthonormal basis of the subspace of E 1 generated by Xl' ... , Xn
(V, p. 23, cor. 1). There exist elements fl' ... , fm in E 2, not all null, such that
n
Z = L ei @ h, hence
i= 1
m
<D(z, z) = L <D(e; ® J;, ej @ .f)
i.j= 1

= L <e;lej ) OJ!) = L IIhl1 2 > O.

i,j i= 1
TVS V.26 HlLBERTIAN SPACES §3

For the general case, we shall now prove that <I> is positive. Let Ei be the Hausdorlf
prehilbertian space associated with Ei and let 1t i be the canonical mapping from Ei
onto Ei (i = 1,2). Put 1t = 1t1 (8) 1t 2 · Let <I> be the hermitian form on E1 (8) E2
constructed in the same way as <1>. We have

<I>(z, t) = <I>(1t(z),1t(t)) (z E F, t E F) ,

and since <I> is positive, so is <1>. Q.E.D.

The prehilbertian space defined in prop. 1 is called the tensor product of the pre-
hilbertian spaces E1 and E2 and is written as E1 (8) E 2 . Henceforth we shall write
< zl t) for <I>(z, t), and therefore by definition

(4)

we shall also write IIzl12 or Ilzll for <ZIZ)1/2. From (4), we get

(5)

then the bilinear mapping (Xl' X2) H Xl ® x 2 from E1 x E2 into E1 ® E2 is

continuous.
For i = 1,2 let Fi be a vector subspace of E i, endowed with the induced prehil-
bertian structure. Then F1 (8) F2 can be identified with a vector subspace ofE1 (8) E2
(A, II, § 7, No.7). Formula (4) shows that F1 (8) F2 with the prehilbertian space struc-
ture induced by that ofE 1 (8)2 E 2, is exactly F1 ®2 F 2. We shall henceforth identify
F1 ®2 F2 with a prehilbertian subspace of E1 (8)2 E 2·

PROPOSITION 2. - For i = 1,2, let Ei and Fi be two Hausdorff prehilbertian spaces

and let ui E 2(E i ; FJ The linear mapping u1 ® U 2 from E1 (8)2 E2 into F1 (8)2 F2
is continuous and we have

Consider the positive hermitian form on E1 given by

By prop. 1 (V, p. 25), there exists a positive hermitian form <I> on E1 ® E2 such
that

<l>(x1 ® x 2' Y1 (8) Yz) = f(x!> Y1) <x 2IY2) =

= II u l 11 2 <Xl (8) x 21Y1 ® Yz) - «u1 (8) I)(X1 ® x 2)I(u1 (8) I)(Y1 (8) Y2)

for Xl' Y1 in E1 and x 2' Y2 in E 2. By linearity we have

for z, t in E1 (8) E 2. Since <I> is positive, we get <l>(z, z) ;;::, 0, that is, II(u1 ® I), Zll2
No.1 TENSOR PRODUCT OF HILBERTIAN SPACES TVS V.27

~ II ulll·11 zl12 for z E E l @2 E2, or II u l <8> 111 ~ II ulll· Similarly we prove the ine-
quality III <8> u211 ~ II u211, and since u l @ u 2 = (u l <8> 1) 0 (1 @ u 2 ), we get

On the other hand,

sup Ilul(xl)II.lluix2)11
IlxI11 "'1.I1x211 ",1

sup !I(ul @ U 2 ) (Xl @ x 2)112 ~ Ilu l <8> u211 .

II x III '" 1, II X2 II '" 1

This completes the proof of prop. 2. Q.E.D.

Let E l , ... , En be prehilbertian spaces (n ~ 2). We define the tensor product
n
E l @2 ... <8>2 En (also denoted by O2EJ by induction, by
i= 1

Hence, by the definition of scalar product, we have

n
(6) <Xl @ ... <8> xnlYl <8> ... <8> Yn) = TI <xiIYi)'
i= 1

and in particular 1

(7) Ilxl <8> ... @ xnl12 = Ilxlll ... Ilxnll ,

for xi' Yi in Ei (1 ~ i ~ n). If the Ei are Hausdorff, then so is El <8>2'" <8>2 En'
Let F l , ... , Fn be prehilbertian spaces and U i E .!l'(E i ; F) for 1 ~ i ~ n. By
induction on n, prop. 2 implies that U l @ ... @ Un is a continuous linear mapping
from El <8>2'" <8>2 En into F J <8>2'" <8>2 Fn and that

(8)

Let (J E 6 n be a permutation of the set {I, 2, ... , n}. Because of (6), the linear
mapping Per from E l @2 ... @2 En onto E,,-I(1) <8>2 ... @2 E,,-I(n) characterized by

(9)

is a prehilbertian space isomorphism (<< commutativity of tensor product »).

Similarly, consider a partition of {I, 2, ... , n} into m consecutive intervals 11' ... , In
with Ik = (ak,a k+ l - 1) for 1 ~ k ~ m. Put
ak+ 1 - 1
Fk = 02 Ei (1 ~ k ~ m) .
i=ak

1 Here again we put IIzl12 = (zlz)I/2 for z in El ®2 ... ®2 En'

TVS V.28 HILBERTIAN SPACES §3

The canonical isomorphism from F 1 ® ... ® F m onto E1 ® ... ® En which trans-

m ak + 1 - 1
forms ® ® Xi into Xl ® ... ® Xn (A, II, § 3, No.9) is a prehilbertian space iso-
~k=l i=ak
morphism (<< associativity of the tensor product »).

2. Hilbertian tensor product of hilbertian spaces

DEFINITION l. - Let E 1, ... , En be hilbertian spaces. The completion of the Hausdorff

prehilbertian space E1 ®2 ... ®2 En is called the hilbertian tensor product of the E j
and is denoted by E1 O O En (or 2... 2
EJ ®2
1 ~i~n
Let F l' ... , Fn be hilbertian spaces and ui E 2'(E i, F) for 1 ~ i ~ n. The continuous
linear mapping u 1 ® ... ® un then extends to a continuous linear mapping
U 1 O2 ... O2 Un from E1 O2 ... O2 En into F1 O2 ... O2 Fn' We have

(10)

by formula (8) of V, p. 27. Moreover, if lE denotes the identity mapping of any hil-
bertian space E, we have

(11)

Finally, ifG 1, ... , G n are hilbertian spaces and V i E2'(Fi; G;) for 1 ~ i ~ n, we get

(12) (v 1 0 u 1) O2 ,,, O2 (v n 0 un) = (v 1 O2 ,,, O2 vn) 0 (u 1 O2 ,,, O2 un)'

We leave to the reader the task of formulating the «commutativity» and the
« associativity » of the hilbertian tensor product, in analogy with what has been
said above for prehilbertian spaces.
Remark. - Let El' ... , En be Hausdorff prehilbertian spaces, and El' ... , En their
t:espective c0!llpletions. Then E l @2 ... @2 En is a prehilbertian s.ubspace .of
E l @2 ... @2 En' Since the mapping (Xl' ... , Xn) f--> Xl @ ... @ Xn fro!ll E j x ... X Pn
into E l @2'" @2 En is continuous, E l @2'" @2 En is dense in E l @2'" @2 En'
1 f;Jrtiori. th~ completion of E l @2 ... @2 En is precisely the hilbertian ~pace
E j @2 ... @2 En' This completion is sometimes simply written as E l @2'" @2 En
(or ®2
E;).
1 ~i!S;n

PROPOSITION 3. - Let E 1, ... , En be hilbertian spaces. Suppose that for 1 ~ i ~ n

the space E j is ahilbertian sum of a family (Ei,~)~EA(i) of closed vector subspaces.
ThenE 1 ®2'" ®2 En isa hilbertian sumofthefamilyofsubspacesE1'~1 O2 ,,, O2 En,~"
with (C'l1' ... , C'ln) ranging over A(1) x ... x A(n).
By formula (6) of V, p. 27, the subspaces E1'~1 O2 ,,, O2 En,~n of E1 O2 ,,, O2 En
are mutually orthogonal. For every integer i between 1 and n, the set U Ei,~ is
~EA(i)

total in E i, and the multilinear mapping (Xl' ... , Xn) f---4 Xl ® ... ® Xn is continuous.
It follows that the union ofthe subs paces E1 ,~1 O2 ... 0 2 En,~" is total, hence prop. 3.
NO.3 TENSOR PRODUCT OF HILBERTIAN SPACES TVS V.29

COROLLARY 1. - For 1 ~ i ~ n, let (ei,a)aEA(i) be an orthonormal basis of E i . Then

the family of vectors e 1 ,al ® ... ® en,an as (a p ... , an) ranges over A(1) x ... x A(n)
is an orthonormal basis of E 1 @z ... @z En'

COROLLARY 2. - Let E1 and E z be two hilbertian spaces, and (e;)iEI an orthonormal

basis of E 1 . Let (Yi)iEI be a family of elements of E z , such that I I Yi liZ < + 00.
iEI
Then the family (e i ® Y;)iEI is summable in E 1 @z E2 ; moreover, every element of
E 1 @ E z can be written uniquely in the form I ei ® Yi with L IIYi liz < + 00.
iEI iEI
Let Fi be the line in E1 generated by the ei (i E I). Then E1 is the hilbertian sum
of the family of subspaces (F)iEI' By prop. 3, the space E1 0 E z is the hilbertian
sum of the family of subspaces (Fi 0 z EZ);EI' hence cor. 2 follows.

Examples. - 1) By cor. 1, the space eZ(I) <8>z £z(1) is canonically isomorphic to

fZ(I x J), the tensor product x ® y of x = (X)iEI and y = (Y)jEJ can be identified
with the family (XiY)iEI,jEJ' Similarly, by cor. 2, £Z(I) <8>z E can be identified with
£~(I), in such a way that we have (X)iEI ® Y = (XiY)iEI for every y in the hilbertian
space E.
* 2) Let X be a Hausdorff topological space, and ~ a positive measure on X.
Let E be a hil~ertian space. We can identify LZ(X,~) <8>z E with L~(X, 11) in a cano-
nical way : if f is the. class of the square integrable scalar function f on X, and if
a belongs to E, then f ® a is the class of the function x H f(x). a with values in E.
Let Y be a Hausdorff topological space and v a positive measure on Y. In an
analogous manner, we can identify the hilbertian spaces U(X, 11) @z U(Y, v)
and L Z(X x Y, 11 ® v); then j ® g can be identified with the class of the function
(x, y) H f(x) g(y) on X x Y. *

3. Symmetric hilbertian powers

Let E be a hilbertian space, and let n be a positive integer. Let Tn(E) or E®n denote
the tensor product of n hilbertian spaces each equal to E. In other words, Tn(E)
is the completion of the space Tn(E) = E ® ... ® E (n factors) for the Hausdorff
prehilbertian space structure defined by
n
<Xl ® ... ® x nlY1 ® ... ® Yn) = TI
i= 1
<xiIYi)'

If(e)iEI is an orthonormal basis ofE, the family of vectors eh ® ... ® ein for ip ... , in
in I, is an orthonormal basis of Tn(E) (V, p. 29, cor. 1). We get TO(E) = K.
Let cr E 6 n be a permutation of the set { 1, 2, ... , n }. By V, p. 27, there exists an
automorphism Pcr of Tn(E) characterized by
TVS V.30 HlLBERTIAN SPACES §3

We have PaT = PaP, for cr, 't in 6 n, and consequently the endomorphism
II" = -!,
n.
L
aeiS n
Pa of the vector space t"(E) is the orthoprojector onto the subspace

of all elements left invariant by 6". However (A, III, § 6, No.3), II" maps the « alge-
braic» tensor product T"(E) onto the subspace TS"(E) of all symmetric tensors of
order n. In other words, the image of II" is the completion of the space TS"(E) endowed
with a scalar product induced by that of T"(E); this completion will be denoted
by TS"(E).
Let S"(E) be the nth symmetric power of the vector space E (A, III, § 6, No.1). The
canonical mapping from T"(E) onto S"(E) defines by restriction an isomorphism
A" from TS"(E) onto S"(E). We verify immediately that the inverse isomorphism
is given by

(13)

for Xl' ... , X" in E.

We define a Hausdorff prehilbertian space structure on S"(E) by putting

(14)

We then have (compare with formula (29) of A, III,§ 11, No.5)

(15) <xl",X"!Yl"'Y") = L
(JEtSn
n" <X;lYa(i)
i= 1
,

and in particular

(16)

Let S"(E) denote the completion of the pre-Hilbertian space S"(E) and SeE)
the external hilbertian sum of the hilbertian spaces S"(E). We can show (V, p. 73,
exerc. 1) that the multiplication in the algebra SeE) cannot be extended by conti-
nuity to SeE), unless E is just O.

PROPOSITION4. - Let (e;)ieI be an orthonormal basis of the hilbertian space E. For

every r:J. in N(l), put
(17) Za = n e~i/(r:J.i !)1/2 .
ieI

Then (za)aeN(I) is an orthonormal basis of SeE).

Let Eo be the vector subspace ofE generated by the vectors ei for i ranging over I.
Then the za form a basis of the vector space S(E o) (A, III, § 6, No.6). But Eo is dense
in E, and the multilinear mapping (Xl> ... , x") f--+ Xl ... X" from Ex'" x E into
SeE) is continuous for all n ? 1; hence S(Eo) is dense in SeE). It is now enough
No.3 TENSOR PRODUCT OF HILBERTIAN SPACES TVS V.31

to prove that the family of the z" is orthonormal. First observe that S"(E) and sm(E)
are orthogonal for n i= m. Hence it is enough to prove the formula

when IIXI = L lXi and I~I = L ~i are equal to the same integer n.
ie] ie]
Consider a partition (P)iEi of the set { 1, 2, ... , n} such that Card Pi = lXi for all
i E I. Put
X k = e i if k belongs to Pi' then Xl ... X" = n
e~i. Similarly we define (Qi)iEI

n
ieI
and Yk in such a way that Card Qi = ~i and Yl ... y" = er i• Since the ei are mu-
;EI
< >
tually orthogonal, we have xkIYO"(k) = 0 except ifthere exists an indice i E I such
that kEPi and cr(k)EQ;. By formula (15), we then have <xl ... x"IYl ... Y,,> = 0
unless there exists a permutation cr E 6" such that cr(P) = Q i for all i E I, which
implies that IX = ~. Then <z"lz~> = 0 for IX i= ~. The same argument proves that
I Xl··· X" 112 is equal to the number of the cr E 6" such that cr(P) = Pi for all i E I,
hence equal to nIX;!. We get I z'" I = 1, and the proposition is proved.
ieI

COROLLARY. - Suppose that the hilbertian space E is the direct sum of the ortho-
gonal subspaces M and N. The canonical isomorphism 9 from SCM) <8l SeN) onto
SeE) (A, III, § 6, No.6) extends uniqu~ly to a hilbertian space isomorphism h from
SCM) ®2 SeN) onto SeE).
Let (eJiEI (resp. (f)jEJ) be an orthonormal basis of the hi1bertian space M (resp. N)
and let Mo (resp. No) be the vector subspace of E generated by the vectors e; (resp.
f). Put Eo= Mo + No and let 90 be the canonical isomorphism from S(M o) <8l S(No)
onto S(Eo). Put

z" = n
;EI
e~i/(lXi !)1/2, t~ = njl/(~j !)1/2
jEJ

for IX E N(I) and ~ E N(J). By prop. 4, we have thus defined the orthonormal bases
(Z")"EN(I) for SCM), (t~)~EN(J) for SeN) and (Z"t~)"EN(I).~EN(J) for SeE). Since we have
z"t~ = 90(z" <8l t~), and since the elements z" <8l t~ form an orthonormal basis of
SCM) ®2 SeN) (V, p. 29, cor. 1), we see that 90 extends to a hi1bertian space iso-
morphism h : SCM) ®2 SeN) ~ SeE). By the construction, we have

for all vectors xl' ... , xm in Mo and Yl' ... , y" in No. By continuity, the same relation
also holds for the vectors Xl' ... , X" in M and the vectors Yl' ..., y" in N; in other
words, h extends g. The uniqueness of h is clear.
Let E and F be two hi1bertian spaces and u E 5t'(E; F). The linear mapping
t"(u) = u O2... ®2 u (n factors) from t"(E) into t"(F) is continuous with norm
Ilull" (V, p. 28, formula (10)). Moreover, formulas (13) and (14) of V, p. 30, show
TVS V.32 HILBERTIAN SPACES §3

that there exists an isomorphism <l>n,E from sn(E) onto the subspace fSn(E) of
tn(E), and only one, such that

Hence there exists a continuous linear mapping sn(u) from sn(E) into sn(F) and
only one which makes the following diagram commutative

+n,E
S"(E) • f"(E)
sn(u) t t tn(u)

S"(F)
+.,F
.. f"(F)

We now prove the formula

(9)
We clearly have II sn(u) II ~ II tn(u) II = Iluli n. Further, for all x E E, we have
sn(u) (x") = u(x)n, Ilxnll = (n!)I/21Ixlln and Ilu(x)nll = (n!)1/21Iu(x)lln, which gives

Ilsn(u)II Ilxll n ~ II u(x) lin ;

it follows immediately that II sn(u) II ~ Ilull n, hence formula (19).
It is clear that we have the formulas

(20) snOE) = lSn(E)

(21) sn(V 0 U) = sn(V) 0 sn(u) for v E .!t'(F; G) .

Finally, 5 n(u) coincides on 5 n(E) with the linear mapping 5 n(u) : 5 n(E) -+ 5 n(F)
defined in A, III, § 6, No.2 since it transforms Xl'" xn into u(x l ) ... u(xn) for every
Xl' ... , xn in E.

Examples. - * I) Let d ~ I be an integer and co a positive function on Rd, locally

integrable with respect to the Lebesgue measure Ji. Let E be the hilbertian space
L 2(R d, co. Ji), and let S = S(E), Then S can be identified with the space of all sequences
f = (fn)n",O' where each 1. is a function on (R~n which is measurable with respect to
the Lebesgue measure Ji ® ... ® Ji (n factors) and invariant under the permutations
of the n factors in (R~n, and is such that

(22) !lfl12 = In! f

n=O Rd
.. f Rd
1J.(x!, ... , x.)i2 co(x!) ... co(x.) dx! ... dX n

is finite. The norm Ilfll in S is defined by formula (22). The hilbertian space S defined
above is called the symmetric Fock space corresponding to the weight co *
* 2) Let X be a Hausdorff topological space, Ji a positive measure of norm I on X
and E a hilbertian subspace of the real hilbertian space Li(X, Ji), We say that E is a

f
gaussian space if the following equivalent conditions are satisfied :
a) for all fE E, we have eiJ dJi = exp( -llfI12/2);
x
No. 4 TENSOR PRODUCT OF HILBERTIAN SPACES TVS V.33

b) for all fEE with norm 1, the image of the measure Il under f is the measure
(21t)-t/2 e- x2 / 2 dx.
on R.
Suppose E is a gaussian space. Let f t , ... , I. be functions whose classes f. belong to E.
We define a function: f t ... I.: on X (called « Wick's product» of f t , ... , I.) by the for-
mula
p n

(23) :it .. '!n: = L (-

O~2p~n
l)P L TI
O'EIp i==l
<ia(2i-t)lfa(2i» TI
j=2p+l
fa(J)'

where Ip is the set of permutations cr of {I, 2, ... , n} such that we have

cr(1) < cr(2), ... , cr(2p - 1) < cr(2p)
cr(l) < cr(3) < ... < cr(2p - 1)
cr(2p + 1) < cr(2p + 2) < ... < cr(n) .

Then there exists an isomorphism <P from SeE) onto a hilbertian subspace of L~(X, Il)
which transforms the product h ... i, h, ...,in,
of calculated in SeE), into (:/t ".!n :).
Suppose that X is a Souslin space and that there exists a countable family (I.) of func-
tions whose classes belong to E and which separate the points of X. Then <p is an iso-
morphism from SeE) onto L~(X, Il). *

4. Exterior hilbertian powers

Let E be a hilbertian space and n a posItlve integer. For every permutation

0' E 6 n , let c" denote its signature; put an = ~ L c"P" in 2(1'n(E») (V, p. 29).
n. uE6n

It is immediate that an is an orthoprojector, whose image A~(E) is the closure in

Tn(E) of the space A~(E) of all antisymmetrictensors of order n(A, III, § 7, No.4). There
exists an isomorphism 1tn from An(E) onto A~(E) which is characterized by

(24)

for Xl' ... , xn in E. We can now define a Hausdorff prehilbertian space structure
on N(E) by putting

(25)

More explicitly, we have (compare with formula (30) of A, III, § 11, No.5).

(26)

for all Xl' ... , xn and YI' ... , Yn in E.

Let An(E) denote the completion of the prehilbertian space N(E), and A(E)
the external hilbertian sum of the hilbertian spaces AnCE).
Example. - * With the notations of example 1 of V, p. 32, we can identify the hilbertian
space A(E) with the set of all sequences (J;,)n ~ 0 of measurable functions for which the
TVS V.34 HILBERTIAN SPACES §3

number Ilfll defined in (22) is finite, and where each function J" is antisymmetric, that
is, satisfies the relation
InC x O(1)' ... , xo(n)) = Eoln(x!, ... , xn)
for every permutation cr E 6 •. The hilbertian space ACE) is called the antisymmetric
Fock space corresponding to the weight co. *

PROPOSITION 5. - Let (e;)iEI be an orthonormal basis of the hilbertian space E.

Endow I with a totally ordered structure. Then the set of all elements eit /\ ... /\ ei n
for il < ... < in is an orthonormal basis of An(E).
We know (A, III, ~ 7, No.8) that the elements in question form a basis of the vector
space A n(Eo) where Eo is the vector subspace ofE generated by the vectors e i . Further,
<
for il < ... < in' the matrix of scalar products eikle if ) is the unit matrix of order n;
by (26), we have Ileit /\ ... /\ ed = 1. Finally, if (ip ... , in) and Up ... ,in) are two
distinct, strictly increasing sequences of elements of I, then there exists an element if
<
distinct from ip ... , in and so we have eiklej!') = 0 for I ::::; k ::::; n, and by (26),
< eit /\ ... /\ einleh /\ ... /\ ejn ) = O. In other words, the family of elements
e it /\ ... /\ ein , for i l < ... < in' is orthonormal.
But Eo is dense in E, and the mapping (Xl' ... , Xn) H Xl /\ ... /\ xn from Ex··· x E
into A"(E) is continuous. Consequently, A"(Eo) is dense in A"(E), and proposition 5
follows.

COROLLARY. - Suppose that the hilbertian space E is the direct sum of two orthogonal
subspaces M and N. The canonical isomorphism g from A(M) ® A(N) onto A(E)
(A. III, ~ 7, No.7) extends in a unique way to a hilbertian space isomorphism from
A(M)®z A(N) onto A(E).
The proof is analogous to that of the corollary of prop. 4 (V, p. 31).
Let E and F be two hilbertian spaces and u E 2'(E; F). We shall show, as in
the case of symmetric powers S"(E) (V, p. 32) that the linear mapping N(u) from
N(E) into N(F) (A, III, ~ 7, No.4) extends to a continuous linear mapping A"(u) from
AnCE) into keF). We have the relations

(27) An(1E) = lAn(E)'

(28) A"Cv 0 u) = An(v) 0 An(u) if v belongs to 2'(F; G),

(29)

In general we do not have equality in formula (29) (TS, IV, § 6). Finally, we have
an isomorphism "'n = "'",E
from A neE) onto the subspace A~(E) of tn(E) defined by

(30)
No.5 TENSOR PRODUCT OF HlLBERTIAN SPACES TVS V.35

5. Exterior multiplication

Let E be a hilbertian space. For every integer n ~ 0, let en be the canonical mapping
from Tn(E) onto I\. n(E); then

(31)

for Xl' ... , xn in E. Let p and q be two positive integers; on account of formulas (30)
and (31) we have

(32) U 1\ V = ep+q(p !~1/2 "'iu) ® (q !~1/2 "'q(V))

for u E I\.P(E) and v E N(E). Since Ilenll ~ (n !)1/2, we get the inequality

(33) Ilu 1\ vii ~ (pp.q.

~ ~)!)1/21Iull.llvll
for u E I\.P(E) and v E I\.q(E). Consequently, the mapping (u, v) 1--* U 1\ v extends by
continuity to a bilinear mapping from Ap(E) x Aq(E) into Ap+q(E), with a norm
p + q) ')1/2
at most equal to ( " . (cf V, p.73, exerc. 2). We again denote this by
p. q.
(u, v) 1--* U 1\ v.

PROPOSITION 6. - Let E be a hilbertian space. We have

(34) Ilx 1\ ull ~ Ilxll·llull

for x E E and u E A(E).
It is clearly enough to consider the case Ilxll = 1.
Let F be the hilbertian subspace of E consisting of all vectors orthogonal to x.
Since E is the hilbertian sum of F and the line K. x, it follows from the corollary
of V, p. 34 that the mapping (v, w) 1--* V + X 1\ W is a hilbertian space isomorphism
from A(F) EEl A(F) onto A(E). If u = v + X 1\ w with v, w in A(F), we have
x 1\ u = X 1\ v, hence Ilx 1\ ull = Ilvll ~ (11v112 + IlwI12)1/2 = Iluli.

COROLLARY 1. - a) Let Xl' ... , xn be elements of the hilbertian space E. We have

(35)

where equality holds only if one of the Xi is null, or if the sequence (x 1> •.• , x n ) is ortho-
gonal.
b) Let xi' ... , x n ' Y l ' ... , Y n be elements of the hilbertian space E. We have

(36)
TVS V36 HILBERTIAN SPACES §3

if the vectors Xi and Yi are non-null; equality holds in (36) if and only if(xl' ... , x.)
and (y l ' ... , Y.) are each an orthogonal basis of the same vector subspace of E.
The inequality (35) follows from prop. 6 by induction on n; the inequality (36)
can be deduced by applying the Cauchy-Schwarz inequality in An(E) and the for-
mula (26) of V, p. 33.
Suppose that the sequence (Xl' ... , X.) is orthogonal. Then

n IIxil12
n
IIXl /\ ... /\ x.112 = det«xilxj ») =
i= 1

>
since < xilxj = 0 for i #- j.
Now/suppose that the vectors Xl' ... , Xn are not null and do not form an orthogonal
sequence. Since Ilxl /\ ... /\ x.11 depends symmetrically on the vectors xl' ... ,xn '
we can assume that Xl is notorthogonal to the subspace F ofE generated by X2 , ••• , Xn'
and that F is not simply O. We can decompose Xl in the form x~ + Y with y#-O
in F and x~ orthogonal to F; then II x~ II < II xlii. But

and so

IIXl /\ ... /\ xnll ~ Ilx~llllx211 ... Ilxnll

< Ilxlll IIx211 ... Ilxnll·
Suppose that the vectors Xi and the vectors Yi are not null. The equality in relation
(36) is equivalent to the conjunction of the equalities

(37) I<Xl /\ ... /\ XnlYl /\ ... /\ Y.>I = Ilxl /\ ... /\ xnll·IIYl /\ ... /\ Ynll
(38) Ilxl /\ ... /\ xnll = Ilxlll··· Ilxnll, IIYl /\ ... /\ YIIII = IIY111··· IIYnll·

By the first part of the proof, the equalities (38) imply that each of the sequences
(Xl' ... , xn) and (Yl' ... , Yn) is orthogonal, which in tum implies that Xl /\ ... /\ X. #- 0
and that Yl /\ ... /\ Y. #- O. Under these conditions, relation (37) implies that there
exists a scalar A#-O such that Yl /\ ... /\ Y n = AX l /\ ... /\ xn (V, p. 3, Remark 1);
in other words, that (xl' ... , xn) and (Yl' ... , Yn) are bases of the same vector subspace
of E (A, III,§ 11, No. 13).

COROLLARY 2. - Let (ai)Hi,j,;;n be a hermitian matrix, with complex elements, and

with determinant D. Suppose that the inequality
n
(39) L 1 ailizj ~
i,j=
0

holds for all complex numbers Zl' ... , zn' Then

(40)
 No.5 SOME CLASSES OF OPERATORS IN IDLBERTIAN SPACES TVS V.37

Suppose D is non-null; the equality D = all ... ann holds if and only if aij = °
for
all i -# j.
Let <l> be the hermitian form on the vector space en given by
n
<l>(z, z') = I aijzizj
i,j= 1

for z = (zl' ... , zn) and z' = (z~, ... , z~) in en. By hypothesis, <l> is positive.
First assume that <l> is separating, that is, that D is non-null. If (el' ... , en) is the
canonical basis of en, we have <l>(ep e) = aij' and cor. 2 follows immediately from
cor. 1, a) by putting Xi = ei •
Since au = <l>(ep e) ~ 0, we also have inequality (40) if D = 0.

COROLLARY 3 (<< Hadamard's inequalities »). - Let (a i) 1 ~i,j~n be a matrix with

complex elements, and with determinant D. Put
n
ci = (I laijI2)1/2 for 1 ~ i ~ n,
j= 1

and m = sup laijl. Then we have

i.j

(41)

If D -# 0, in order that IDI = c 1... cn it is necessary and sufficient that the rows
Yi = (a i)l ~j~n of the matrix (a;) 1 ~i,j~n are two by two orthogonal vectors.
Let the space en be assigned a scalar product defined by
n
<Zlz') = I ZiZ;,
i=l

Let (Xl' ... , xn) be the canonical basis of en and Yi the vector with components aij
for 1 ~ j ~ n. We have Ilxill = 1 and IIYil1 = cJor 1 ~ i ~ n; also <xiIYj) = aji'
The inequality IDI ~ c1... cn and the condition of equality are then particular
cases of V, p. 35, cor. 1. Obviously we have ci ~ m.n 1 / 2, hence c1 ... cn ~ mn.nn/2.

§ 4. SOME CLASSES OF OPERATORS

IN HILBERTIAN SPACES

Throughout this paragraph, 1E denotes the identity mapping of a hilbertian space E.

The composition v 0 u of two linear mappings will usually be denoted by vu or v. u.
TVS V.38 HILBERTIAN SPACES §4

1. Adjoint

PROPOSITION 1. - Let E and F be two hilbertian spaces. For every mapping

u E 2(E; F), there exists a unique mapping u* E 2(F; E) such that

(1) <u(x)IY)F = <xlu*(Y)E

for all x E E and all Y E F. The mapping u H u* from 2(E, F) into 2(F; E) is bijec-
tive, isometric and semi-linear (with respect to the automorphism ~ H ~ of K).
Let Y'(E, F) be the space of all continuous sesquilinear forms on E x F, endowed
with the norm

(2) 11<1>11 = sup 1<1>(x, Y)I·

Ilxll';;l,llyIIQ

We define the space Y'(F, E) similarly. We defined (V, p. 16, cor. 2) a Banach space
isomorphism from 2(E; F) onto Y'(F, E), denoted by u H<1>u and characterized by

(3) <1>u(Y, x) = <Ylu(x)F (x E E, Y E F) .

In an analogous way we define an isomorphism from 2(F, E) onto Y'(E, F). Finally
we define a mapping <1> H <1>* from Y'(F, E) onto Y'(E, F) by

(4) <1>*(x, y) = <I>(y, x) (x E E, Y E F) .

This mapping is bijective, semi-linear and isometric. But formula (1) translates as
<1>u* = (<1>u)*' hence the proposition.

DEFINITION 1. - Let E and F be two hilbertian spaces. For every continuous linear
mapping u: E ~ F, the continuous linear mapping from F into E defined by formula (1)
is called the adjoint of u and is denoted by u*.
We have
(5) (u + v)* = u* + v*
(6) (AU)* = AU*
(7) (u*)* = u
(8) (1E)* = IE
(9) (wu)* = u*w*;

in all these formulas, u and v belong to 2 (E ; F), A is in K, and w in 2 (F ; G) where

G is a hilbertian space. Formulas (5) and (6) mean that u H u* is semi-linear. For-
mula (8) is obvious. To prove (7), we take the conjugate of the two members of(1),
No. I SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.39

which gives (u*(Y)lx) = (ylU(x), and this proves that u is the adjoint of u*. Finally,
with the notations of (9), we have, for all Z E G

(w(u(x))lz) = (U(x)lw*(z) = (xlu*(w*(z))) ,

hence u*w* is the adjoint of wu.

Let u: E --+ F be a bijective and continuous linear mapping; then it is also bicon-
tinuous (I, p. 19, cor. I). From (8) and (9) we immediately deduce that u* is bijective
and bicontinuous and that

(10)

PROPOSITION 2. - For every u E J.t'(E; F), we have

(11) Ilu*ull = Iluu*11 = IIul1 2 = Ilu*112.

By prop. I, Ilu*11 = Ilull, hence Ilu*ull :'S; Ilu*II.llull :'S; Ilu11 2 • On the other hand,

IIul1 2 = sup IIu(x)112 = sup (U(x)lU(x) = sup (xlu*U(x):'S; Ilu*ull ,

Ilxll';l Ilxll';l Ilxll';l

hence Ilu*ull = Ilu11 2 . Replacing u by u*, we get Iluu*11 = Ilu*11 2 , hence (11) follows
since II ull = II u* II·
Let E 1 , ... , En and F l' ... , F n be hilbertian spaces, and for every integer i between I
and n, let ui be a continuous linear mapping from Ei into F i. Then

(12)

Let v be the continuous linear mapping u 1 ®2 ... O2 Un from

E = El ®2 ... O2En into F = Fl O2... ®2 Fn

and w the continuous linear mapping ui ®2 ... ®2 u: from F into E. It is enough

to prove the equality (Ylv(x) = (w(Y)lx) for x E E and Y E F. By linearity and
continuity, we reduce to the case when x and Y have the following form

x = Xl (8) ... @ x n ' Y = Yl @ ... @ Yn

with Xi E Ei and Yi E Fi for I :'S; i :'S; n. From the definition of scalar product in a
tensor product (V, p. 27, formula (6)), we then get
n n
(Ylv(x) = f1 (Yilui(X) = f1 (ui(y)lx i ) = (W(Y)lx).
i= 1 i= 1

This proves our assertion.

Let E and F be two hilbertian spaces, u E J.t'(E; F) and n a positive integer. If
we put u1 = ... = un = u in formula (12) we obtain the result that the continuous
TVS V.40 HILBERTIAN SPACES §4

linear mapping Tn(u*) from Tn(F) into Tn(E) is the adjoint of the continuous linear
mapping Tn(u) from Tn(E) into Tn(F). The formulas

(13)

can be established in the same way as formula (12), on account of the definition of
the scalar product in 5 n(E) (V, p. 30, formula(15)) and in An(E) (V, p. 33, formula (26)).

Remark 1. ~ Suppose the hilbertian space E does not reduce to O. We identify it'(K; E)
with E by the mapping u H u(1); in other words, the vector x of E is identified with the
mapping AHA. x from K into E. Then the adjoint of x is the mapping x* : E --+ K
given by x*(y) = (xly). In other words, x H x* is the canonical semi-linear mapping
from E onto its dual (V, p. 15).
Similarly, we identify the number A E K with the endomorphism A. 1E of E. Then A*
is precisely the conjugate of A.
With these identifications, we can define a product 11 ... In where each ti is, either
a number in K, or a vector in E, or a linear form belonging to E', or an element of
it' (E), provided that there are never two consecutive factors Ii and Ii + 1 of one of the
following types :
• xy where x, yare both in E, or both in E' ;
• xA or Ax' with A E it'(E), x E E and x' E E'.
We have the following rules of composition :
a) associativity;
b) every element of K commutes with all the other factors;
c) we have (tl ... In)* =r: ...Ii ; in other words, the adjoint of a product is the product
of the adjoints taken in the reverse order. Also t** = I.
For example, let x, y be in E and let A be in it'(E). Then x*y represents the scalar
product (xly) and x*Ay represents the scalar product (xIAy). We also have
(A*x)* = x*A** = x*A, hence (A*x)*y = x*Ay, which can be interpreted as
(A*xly) = (xIAy)
in conformity with the definition of the adjoint We observe that yx* is the endo-
morphism Z H Y (xlz) of E, since yx*z can be interpreted as y(x*z) by associativity,
or as y.(xlz).
Following Dirac 1, in most works of Mathematical Physics, the elements of E
are represented by the symbol Ix), those of E' by (II. The scalar product is written as
(x Iy) = (xl.ly) and the first rule of interdiction in the products excludes the combi-
nations of the signs) I and 1(, for example Ix) Iy).

PROPOSITION 3. - Let E and F be two hilbertian spaces and u E .P(E; F). The fol-
lowing conditions are equivalent :
(i) u is a topological vector space isomorphism, with an inverse equal to u*;
(ii) u is surjective and u*u = IE;
(iii) u is injective and uu* = 1F ;
(iv) u is an isomorphism of normcd spaces;
(v) u is a hilbertian space isomorphism.
Condition (1) means that we have u*u = IE and uu* = IF. Hence the equivalence
of(i), (ii) and (iii) follows from E, II, § 3, ~o. 8, prop. 8. We have already seen theequiva-
lenee of (iv) and (v) (V, p. 5). Finally, the relation u*u = IE is equivalent to

1 See P. A. M. DIRAC, Quantum Mechanics, Oxford University Press, New York, 1935.
No.2 SOME CLASSES OF OPERATORS IN HlLBERTIAN SPACES TVS V.41

(xlu*u(y» = (xly), that is, to(u(x)lu(y» = (xly) for all x, y in E, and evidently
implies that u is injective; this proves the equivalence of (ii) and (v).
An automorphism of the hilbertian space E is also called a unitary operator,
that is, an operator u E .P(E) satisfying uu* = u*u = IE'
Remark 2. - The relation u*u = IE does not characterize all the automorphisms of
the hilbertian space E. For example, let E = f2(N) and let u be defined by u(xn) = Xn- I
for n ~ 1 and u(x)o = O. We have Ilu(x) II = Ilxll for all x E E, that is, u*u = IE' but u
is not surjective.
Remark 3. - The definition (1) of the adjoint u* can also be written as
<ylu(x) = <u*(y)lx) ,
or, by Y, p. 15, as
<u(X), y*) = <x, (u*(y))*).

But we also have <u(x), y*) = <x, 'u(y*), hence we can express the adjoint in terms
of the transpose,
(u*(y»)* = tu(y*) .

2. Partially isometric linear mappings

DEFINITION 2. - Let E and F be two hitbertian spaces and u E .P(E; F). The ortho-
gonal of the kernel of u in E is said to be the initial subspace of u and the closure of
the image ofu in F is called the final subspace ofu. The orthoprojector from E (resp. F)
onto the initial (resp. final) subspace of u is called the initial (resp. final) orthoprojector
ofu.
Let P be the initial subspace of u. Since E is the direct sum of P and of the kernel
of u, we have u(P) = u(E).

PROPOSITION 4. - (i) The initial (resp. final) subspace of u* is equal to the final
(resp. initial) subspace of u.
(ii) Suppose that E = F. Let M be a closed vector subspace ofE and MO its ortho-
gonal. The relations u(M) c M and u*(MO) c MO are equivalent.
Let Q = u(E) be the final subspace of u. The orthogonal QO of Q in F consists
of all vectors y such that (u(x)ly ) = 0 for all x E E; this is equivalent to : (xlu*(y» = 0
for all x E E, or to u*(y) = O. Hence we have QO = Ker u*, and Q is the initial
subspace of u*. Since u is the adjoint of u*, the final subspace of u* is also the initial
subspace of u. This proves (i).
The relation u(M) c M implies that u(M) is orthogonal to MO, and the rela-
tion u*(MO) c MO implies that u*(MO) is orthogonal to M. But we have
(u(x)ly) = (u*(Y)lx) for all x EM and y E MO; hence (ii) follows.

We remark that prop. 4 can be deduced from the general properties of the transpose
(II, p. 51, cor. 2) in view of remark 3, Y, p. 41.

DEFINITION 3. - Let E and F be two hitbertian spaces. A mapping u E .P(E; F) is

TVS V.42 HlLBERTIAN SPACES §4

said to be partially isometric if II U(X) II = .llxll for all x belonging to the initial subspace
ofu.
Let u E .,;t(E; F) and let N be its kernel and I its image. To say that u is partially
isometric is the same as saying that the linear mapping i1: E/N --+ I deduced from u
is isometric (V, p. l3). Then the subspace I of F is complete, hence closed, and is
the final subspace of u. Consequently, u induces a hilbertian space isomorphism
from the initial subspace of u onto its final subspace.

PROPOSITION 5. - Let u E .,;t(E; F), let P be its initial subspace and Q be the final
subspace. Let p (resp. q) denote the initial (resp. final) orthoprojector of u. Assume that
u is partially isometric.
(i) The mapping u* E .,;t(F; E) is partially isometric, with initial subspace Q and
final subspace P. The isomorphism from Ponto Q induced by u is then the inverse of
the isomorphism from Q onto P induced by u*.
(ii) We have u*u = p and uu* = q.
On account of prop. 4 (i), assertion (i) is a consequence of (ii).
We now prove (ii). Since P contains the image of u*, the mapping u*u maps E
into P. Let x E E and YEP, then

<u*u(x)ly) = <u(x)lu(y) .

If x belongs to P, then <u(x)lu(y) = <xly) by the definition of a partially isometric

mapping; if x belongs to the kernel N of u, then u(x) = 0, hence <u(x)lu(y) = 0
and <xly) = 0 since Nand P are orthogonal. Since E = P EB N, we have
<u*u(x) - xly) = 0 in all the cases, and so u*u is the orthoprojector p from E
onto P. That uu* = q follows by interchanging u and u* in the above.

PROPOSITION 6. - For every u E .,;t(E; F), the following conditions are equivalent:
(i) u is partially isometric;
(ii) u* is partially isometric;
(iii) u*u is an orthoprojector;
(iv) uu* is an orthoprojector;
(v) uu*u = u;
(vi) u*uu* = u*.
By prop. 5, (i) is equivalent to (ii).
(i) => (v) : Suppose u is partially isometric. Then u*u is the initial orthoprojector
of u by prop. 5. Hence for every x E E, u*u(x) - x belongs to the kernel of u, that is,
uu*u(x) = u(x).
(v) => (iii) : Suppose that uu*u = u and let p = u*u; then p = p* and p2 = p.
Let M (resp. N) be the image (resp. the kernel) of p. For x EM and YEN, we have
<xly) = <p(x)ly) = <xlp*(y) = <xlp(y) = O. Since M and N are orthogonal,
p is the orthoprojector from E onto M.
(iii) => (i) : Suppose p = u*u is an orthoprojector with image M and kernel N.
No.3 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.43

For all x E E, we have

Ilu(x)11 2 = <u*u(x)lx) = <p(x)lx).

Hence u(x) = 0 for x E Nand Ilu(x) I = Ilxll for x EM, and so u is partially isometric
with kernel N and initial subspace M.
We have proved the equivalence of (i), (iii) and (v). Replacing u by u*, we can
deduce the equivalence of (ii), (iv) and (vi). This proves prop. 6.

3. Normal endomorphisms

DEFINITION 4. - Let E be a hilbertian space and u E 2(E). We say that u is normal

if it commutes with its adjoint u*.
For example, every automorphism u of the hilbertian space E is normal since
we have uu* = u*u = IE'

PROPOSITION 7. -
For u E 2(E) to be normal, it is necessary and sufficient that
IIu(x)11 = I u*(x) I
for all x E E.
We define a hermitian form <l> on E by

<l>(x,y) = <uu*(x)IY) - <u*u(x)ly)·

For u to be normal, it is necessary and sufficient that <l> = O. By the polarization

formulas (V, p. 2), this is equivalent to <I>(x, x) = 0 for all x E E. The proposition
now follows since
<l>(x, x) = Ilu*(x)112 - Ilu(x)11 2.
PROPOSITION 8. - Suppose that u E 2(E) is normal. Let N be the kernel ofu and M
the orthogonal of N in E; let m and n be two positive integers such that m + n ~ l.
Then N is the kernel of If''(u*)n and M is both the initial and the final subspace of
If''(u*)n. In particular, M is both the initial and the final subspace of u and of u*, and
is stable under u and u*.
Prop. 7 shows that u and u* have the same kernel N. By prop. 4, (ii) of V, p. 41,
the subspace M of E is stable under u and u* since this is so for N = MD, since
M n N = {O}, the endomorphisms of M induced by u and u* are injective. Let
v = um(u*)n; the preceding argument shows that the restriction of v to M (resp. N)
is injective (resp. null), hence N is the kernel of v. Consequently, M = N° is the
initial subspace of v. By prop. 4, (i) of V, p. 41, the final subspace of v is equal to
the'initial subspace of v*. But v* = u"(u*)m and so the initial subspace of v* is equal
to M by the preceding.

COROLLARY. - Let Ie E K. The following subspaces ofE are equal:

a) the eigen subspace of u relative to Ie;
b) the eigen subspace of u* relative to )::;
TVS V.44 HILBERTIAN SPACES §4

c) the primary subspace of u relative to A (in other words, by LIE, VII, § 1, No.1,
the set of all vectors x of E for which there exists an integer n ~ 0 such that
(u - A.IEt (x) = 0);
d) the primary subspace of u* relative to ~.
It is clear that w = u - ~.lE is a normal endomorphism of E, hence the endo-
morphisms w, w* = u* - A.I E , wn and (w*)n ofE have the same kernel by prop. 8.

4. Hermitian endomorphisms

DEFINITION 5. - Let E be a hilbertian space and let u E 2(E). We say that u is her-
mitian if u* = u.
Let J'l'(E) denote the set of all hermitian elements of 2(E); this is a vector sub-
space of the vector space 2(E)[R] over R which is deduced from 2(E) by restricting
the scalars.
To each u E 2(E), we associated (V, p. 16, cor. 2) a sesquilinear form
<1>u :(x, y) ~ <xlu(y) on E x E. We have

(14) <1>u.(x, y) = <1>U<y, x) (x, y in E);

consequently, u is hermitian if and only if the form <1>u is hermitian. When K is C,

it is enough to assume that <1>u(x, x) = <xlU(x) is real for all x E E (V, p. 2, Remark).
Let u E 2(E). We have seen (V, p. 16, cor. 2) that the norm of u can be calculated
by the formula

(15) lIuli = sup l<1>u(x, Y)I·

IlxIIG,llyll"'l

When u is hermitian, we have the following result :

PROPOSITION 9. - For every hermitian endomorphism u of E, we have

(16) lIuli = sup I<xlu(x) I·

Ilx II'" 1

Put <1> = <1>u and c = sup 1<1>(x, x)l, then evidently c :oS; lIuli. Let x, y be in E
Ilxll '" 1
such that IIxll :oS; 1, lIyll:oS; 1. Then

<1>(x + y, x + y)= <1>(x, x) + <1>(y, y) + 2~<1>(x, y) ,

hence
4~<1>(x, y) = <1>(x + y, x + y) - <1>(x - y, x- y);

but 1<1>(t, t) 1 :oS; cll t 112 for all tEE, thus

41~<1>(x, y)1 :oS; c(lIx + yll2 + IIx - Y1l2) = 2c(lIxll2 + IIYIl2) :oS; 4c.

Let a = <1>(x, y); there exists a complex number A with absolute value 1 such that
No.5 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS VA5

'Aa = lal. Replacing y by 'Ay in the preceding inequality, we get 1<l>(x,y)1 :( c. By

(15), Ilull :( c and the proposition follows. Q.E.D.
Evidently every hermitian endomorphism is normal. Conversely :

PROPOSITION 10. - Suppose K is C. Let u E 2(E). Then there exists a unique pair
(h 1, h z ) of hermitian endomorphisms of E, such that u = h1 + ih z . In order that u
is normal, it is necessary and sufficient that h1 and h z commute.
For, the relation « u = h1 + ihz, hi = h1' hi = h2 » is equivalent to

«h1 = 21 (u + u*) and h2 = 2i (u* - u) » .

In addition, we have h1 h2 - h zh 1 = 4(uu* - u*u). This proves prop. 10.

PROPOSITION 11. - Let p E 2(E). In order that p is the orthoprojector from E onto
a closed vector subspace of E, it is necessary and sufficient that p2 = P = p*.
Suppose pZ = p. Let M be the image of p and N its kernel. E is the topological
direct sum of M and N. In order thap p is an orthoprojector, it is necessary and
sufficient that M is orthogonal to N, that is to say that we have (p(x)ly - p(y) = 0 >
for all x, y in E. This latter relation is equivalent to p = p*p, and implies that
p* = (p*p)* = p*p = p; conversely if p* = p, we have p = p2 = p*p.

5. Positive endomorphisms

DEFINITION 6. - Let E be a hilbertian space and u E 2(E). We say that u is positive,

>
and write u ? 0, if u is hermitian and if (xju(x) ? 0 for all x E E.
When K is equal to C, the relation

(xlu(x) >? 0 for all x E E

implies that u is hermitian (V, p. 2, Remark), hence positive.

Let 2 +(E) denote the set of all positive elements of 2(E); this is a proper pointed
convex cone in the real vector space 2(E)[R) underlying 2(E). In order that u is
positive, it is necessary and sufficient that the sesquilinear form <l>u on E x E asso-
ciated with u is positive hermitian. Given u and v in 2(E), the relation u - v ~ 0
can also be written as u ? v or v :( u; this is an order relation on 2(E)[R] compa-
tible with its real vector space structure.

PROPOSITION 12. - Let u be a hermitian (resp. positive) element of 2(E) and let v
be a continuous linear mapping from E into a hilbertian space F. Then vuv* is a her-
mitian (resp. positive) element of 2(F).
For, we have (vuv*)* = v**u*v* = vuv*. On the other hand, if u ? 0, we have
(Ylvuv*(y) >= (v*(y)lu(v*(y» >~ 0

for all y E F, hence vuv* ? O.

TVS V.46 HILBERTIAN SPACES §4

Prop. 12 shows, in particular, that vv* is positive for all v E 2'(E; F). Since, in
particular, an orthoprojector p satisfies p = p2 = pp*, it is positive.

Remarks. - 1) For every hermitian u in £'(E), put m(u) = inf <xlu(x»,

Ilxll ~ 1

M(u) = sup <xlu(x». If E is not just 0, m(u) and M(u) are finite; moreover, M(u)
Ilx!1 ~ 1
is the smallest real number A such that u :;;:; A. 1E and m(u) the largest real number ~
such that u ~ ~.IE' Clearly we have m( - u) = - M(u) and M( - u) = - m(u). It
is clear that

sup (Im(u)l, IM(u)l) = sup l<xlu(x»1

Ilxll ~ 1

and prop. 9 (Y, p. 44) implies (for E i= {O}) that

(17) !iull = sup (lm(u)l. IM(u)I)·

* For another proof of this

formula when K is C, see prop. 14 of TS. I. § 6, No.8. *
2) Let M and N be two closed vector subspaces of E, and PM (resp. PN) the ortho-
projector from E onto M (resp. N). Then M c: N if and only if PM :;;:; PN' For, we have
P~PM = PM. hence

for all x E E. The relation PM :;;:; PN is therefore equivalent to « IlpM(x) I :; :; IIp:-.;(x) I for
all x E E ». If M c: N. we have P. = PMPN. hence I PM(X) I :; :; I PN(X) I since II PM II :;;:; 1.
Conversely, if IlpM(x) I :; :; I PN(X)] for all x E E, the kernel of PM contains the kernel
of PN. that is, that MO :::J N which implies that M c: N.
C
•

PROPOSITION 13. - Let £,(E) be the set of all continuous hermitian endomorph isms
of the hilbertian space E. Let .fF be a non-empty, directed increasing and bounded
subset of £' (E).
(i) The set ff has an upper bound U o in £,(E); we have

(18) <xluo(x» = sup <xlu(x» for all x E E.

UE§

(ii) The filter of sections of.fF converges to U o in the space 2'(E) endowed with the
topology of simple convergence.
Let L be the filter of sections of .'!F; for every u E £'(E), let <l>u be the continuous
hermitian form on E defined by

<l>u(X, y) = <xlu(y» .
Let

for u E £'(E) and x E E. By the polarization formulas (V, p. 2), we have

(19) 4<1>Jx, y) = \fu(x+y)-\fu(x-y) if K = R

(20) 4<1>u(x,y) = \fJx+y)-\fu(x-y)-i\fJx+iy)+i\fJx-iy) if K=C.
No.5 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.47

For every x E E, the mapping u I----> 'PuCx) from into R is increasing and bound-
ed, hence has a limit with respect to L. By the preceding formulas, the limit

lim <l>u(x, y) = <I>(x, y)

u,1:

exists for every pair (x, y) of elements ofE. It is clear that <I> is a hermitian form on E.
If Vi E:?F and V2 isa bound of :?F, the hermitian forms jl = <I> - <I> v1 andj2 = <l>v2-<I>
are positive; there exists a real number M ~ 0 such that

hence

consequently the semi-norms XI----> hex, X)1/2 are continuous on E. Since

we conclude that x I----> <l>(x, x) is a continuous function on E, and by formulas (19)

and (20), that <I> is continuous on E x E. Therefore there exists (V, p. 16, cor. 2) an
element U o of J'l'(E) such that <I> = <l>uo' Formula (18) is evidently satisfied, hence
U o is the upper bound of :?F in J'l'(E). This proves (i).
We have, by construction

(21) lim
u,1:
<xl(u o - u) (x) = 0 for all x E E.

Let Vi E :?F; given a u E :?F such that u ~ Vi' let V = U o - u. If we apply the Cauchy-
Schwarz inequality to the positive hermitian form <l>v on E, we get

I vex) 114 = l<I>vCv(x), x)j2 ~ <l>v(V(X), v(x)).<I>v(x, x)

= <v(x)lv 2 (x) <xlv(x) ~ IIvl1 3 IIxl1 2 <xlv(x)
~ Iluo - vl 1 3 IIxl1 2 <xlv(x) ,

since Ilvll ~ Iluo - viii by V, p. 44, prop. 9. Then by (21) we get lim
u,1:
II(uo-u)(x)11 =0
for all x E E; which proves assertion (ii). Q.E.D.
In particular, prop. 13 can be applied to the case of an increasing and bounded
sequence (Un)nEN of elements of J'l'(E). Then there exists an element V of J'l'(E) cha-
racterized by
<xlv(x) = lim
n~oo
<xlunCx) = sup <xlunCx) (x E E),

and we have vex) = lim u/x) for all x E E. Moreover, v is the upper bound of the
n~oo

set of the Un in J'l'(E).

TVS V.48 HILBERTIAN SPACES §4

6. Trace of an endomorphism

Let E and F be two hilbertian spaces. Conforming to the conventions of V, p. 40,

we let ba*, for a in E and b in F, denote the continuous linear mapping x H b alx) <
from E into F.

Lemma 1. - There exists an isomorphism e from the vector space F (8) E' onto the
space 2 IE; F) of all finite rank continuous linear mappings from E into F, charac-
terized by e(b ® a*) = ba* for a E E, bE F.
By A, II, § 4, No.2, there exists an injective linear mapping e from F ® E' into
2(E; F) and only one such, which transforms b (8) a' into the linear mapping
x H ba'(x) for a' E E', bE F. Evidently e(b (8) a*) = ba*, and the image ofe is contain-
ed in 2 j (E; F). However, let u E 2 j (E; F) and let (e l , ... , en) be an orthonormal basis
of the image of u in F. Let I; = u*(e) for 1 :( i :( n. For every x E E, we have
n n
u(x) = I <eilu(x).ei = I <l;lx).e i ,
i= 1 i= 1

n n
hence u = L eil;* = e( I ei ® 1;*). Therefore the image of e is equal to 2 IE; F).
i= 1 i= 1
We shall henceforth assume that E = F, and we set 2 IE) = 2 IE; E). By
lemma 1, there exists a unique linear form, on 2 IE), such that ,(e(a (8) a')) = a'(a)
for a E E, a' E E'; in other words, we have

(22) ,(ba*) = <alb) for a, b in E.

When E is finite dimensional, we have 2 IE) = 2(E) and ,(u) is the trace of the
endomorphism u of E (A, II, § 4, No.3).

Lemma 2. - Let (e)iE[ be an orthonormal basis of E. Then

,(u) = I <eilu(e)
iEI

for all u E 2 IE).

It is enough to consider the case when u = ba* with a, b in E. Then

and lemma 2 follows from formula (22) and formula (3) of V, p. 22.

Lemma 3. - Let u be a continuous and positive endomorphism of E, and :F the set

of all finite !!!nk orthoprojectors on E. Then for every orthonormal basis (e)iE[ of E,
we have (in R +) the equality

L <eilu(ei) = sup ,(pup) .

iEI PE:F
NO.6 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.49

For every finite subset J of I, put PJ = I eief ; this is the orthoprojector from E
iEJ
onto the vector subspace generated by the vectors ei , where i ranges over J. We have

PJuPJ = I <eilu(e) eiej ,

iEJ,jEJ

hence ,(PJuPJ) = L <eilu(e;). Since PJ E ff',

ieJ

I <eilu(e;) ~ sup 'r(pup);

ieJ PE~

and so we conclude that

I <eilu(ei) = sup I <e;lu(ei) ~ sup ,(pup) .

iEI J iEJ PE~

Let v be a finite rank continuous and positive endomorphism of E and let P E ff'.
By tho 2 of V, p. 23 there exists an orthonormal basis U;,)a.EA of E and a finite subset
B of A such that (J;,)a.EB is an orthonormal basis of the image of p. Then we have
P = L fafa*, and so, as above, the relation ,(pvp) = I <.I;, 1v([a.». By lemma 2
!XEB (XED

(V, p. 48) we have ,(v) = I <fa 1v(fa.) ), which gives the formula
a.EA

I <falv(fa.) ~ 't(v) .
<XEB

Applying this inequality to the case where v = PJ. uPJ and where J is a finite subset
of I, we get

(23) I <pi fa) lupifa) ) ~ I <edu(e;) .

a.EB iEJ

For every x E E, we have pix) = I <eilx) ep and so x = lim pix) with respect
iEJ J
to the ordered directed set of finite subsets J of I. Passing to the limit over J in (23),
we get
,(pup) = I <.I;, u(fa.) )
1 ~ I <eilu(e) ,
(XeB ieI

and this completes the proof of lemma 3.

DEFINITION 7. - Let u be a continuous and positive endomorphism of the hilbertian

space E. Let

(24) Tr(u) = sup ,(pup)

pefii

(upper bound in R+), where ff' is the set of all finite rank orthoprojectors on E. We
say that Tr(u) is the trace of u.
TVS V.50 HILBERTIAN SPACES §4

Let p be the orthoprojector from E onto a finite dimensional vector subspace of E,

and let (Xl' ... , xm) be an orthonormal basis of F. We have established the relation
m
.(pup) = I <xilu(x). Consequently, we can define the trace by the formula
i= 1

m
(24') Tr(u) = sup I <xilu(x) ,
x b ... ,Xnt i= 1

where (Xl' ... , Xm) ranges over the set of all finite orthonormal sequences of vectors
ofE.
By lemma 3 (V, p. 48), we have

(25) Tr(u) = I <eilu(eJ)

iEI

for every orthonormal basis (e)iEI of E. From this, we deduce

(26) Tr(u + v) = Tr(u) + Tr(v)

(27) Tr(Au) = A. Tr(u)

for all continuous and positive endomorphisms u and v of E and for every real
number A ? 0 (we make the convention 0.( + ex)) = 0 in (27)). Let <I> be an isomor-
phism from E onto a hilbertian space F; since <I> transforms every orthonormal
basis of E into an orthonormal basis of F, we get from (25) that

(28) Tr(<I>u<l>-l) = Tr(u).

Let (U~)~EA be a non-empty directed increasing and bounded famity of continuous

and positive endomorphisms of E; let u = sup ua ' then <xlu(x) = sup <xluix)
~ 0

for all x E E (V, p. 46, prop. 13). We have Tr(u) = sup I <eilu(e;», where J ranges
J cI iEJ
over all finite subsets of I, hence

(29) Tr(u) = sup Tr(uJ for u = sup ua •

~ ,

Let PF be the orthoprojector from E onto the hilbertian subspace F; there exists
an orthonormal basis (eJiEI of E and a subset J of I, such that (ei)iEJ is an orthonor-
mal basis of F. We have Tr(PFuPF) = I <eilu(eJ). This formula has two conse-
iEJ
quences : firstly. we have Tr(pF uPF) ~ Tr(u); secondly, taking u = I E, we get

dim F if F is finite dimensional

(30) Tr(PF) = { .
+x if not.
No.6 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.51

DEFINITION 8. - Let E be a complex hilbertian space. We write 2 l(E) for the vector
subspace of 2(E) generated by all continuous, positive endomorph isms of E with
finite trace.
By formula (25) of V, p. 50, the trace extends to a linear form on 2 1 (E), again
denoted by Tr, and satisfying the relation Tr(u) = I <edu(e) for all u in 21(E)
iEI
and for every orthonormal basis (ei)iEI of E. For every u E 2 1(E), we ha ve u* E 21 (E)
and Tr(u*) = Tr(u). Formula (28) of V, p. 50 extends to the case where u belongs
to 2 1(E). Let F be a hilbertian subspace of E; by formula (30), the orthoprojector
PF belongs to 21(E) if and only if F is finite dimensional. For every a and b in E,
we have 4ab* = I c(a + cb) (a + cb)* and cc* is a positive operator with finite
£4= 1
trace for all c E E; consequently, if u is a finite rank, continuous endomorphism of E,
then u E 21(E) and Tr(u) = 't(u).
Let E be a real hilbertian space, and let E(C) be its complexification (V, p. 5). We
identify E with a subset of E(C). Then 2(E) can be identified with a real vector sub-
space of 2(E(C» consisting of all continuous linear mappings u from E(C) into E(C)
such that u(E) c E. In this case we write 21(E) = 2(E) ( l 2 1(E(C». For every
u E 2 1(E), the trace Tr(u) is real and is equal to Tr(u*). Formulas (25) and (28)
are again valid, 2/(E) c 21(E) and Tr(u) = 't(u) for all u E 2 f (E). Finally, a closed
vector subspace F of E is finite dimensional if and only if PF belongs to 2' 1 (E).
* Remark 1. - We shall later define the notion of a nuclear mapping from a Banach
space E into a Banach space F. We shall show that when !£ll(E) consists of all nuclear
mappings from E into E, then E is a real or complex hilbertian space. *

PROPOSITION 14. - Let E 1 , ... , En be hilbertian spaces, E = E1 O2 •.• O2 En' and

ui a continuous endomorphism of Ei for I ::::; i ::::; n. If u 1 , .•. , Un are positive, then so
is u = U 1 O2 •.. O2 Un' and

(31) Tr(u) = n Tr(u).

n

i= 1

If Ui E 2 l(Ei) for all 1 ::::; i ::::; n, then u E 21(E) and formula (31) is again valid in
this case.
Proceeding by induction on n, we immediately reduce to the case n = 2.
For i = I, 2, we define a sesquilinear form <l>i on Ei by the formula <l>;(x, y) = <xlu;(y»
for x, Y in E i . If ul and U 2 are positive, the forms <1>1 and <1>2 are hermitian and posi-
tive. By prop. 1 of V, p. 25 there exists a positive hermitian form <1> on the vector
space El ® E2 such that

<l>(Xl ® x 2' Yl ® Y2) = <1>l(Xl' Yl)·<1>i x 2' Y2)

°
for Xl' Yl in E1 and x 2' Y2 in E 2 · We verify immediately the relation <l>(z, t)=<zlu(t»
for z and tin E1 ® E 2 . Since <1> is positive, we have <zlu(z» ~ for all z in E1 ® E 2 .
Since u is continuous and E1 ® E2 is dense in the hilbertian space E = El O2 E 2 ,
we conclude that u is a continuous and positive endomorphism of E.
TVS V.52 HILBERTIAN SPACES §4

Let (e)ieI be an orthonormal basis of El and Uj)jeJ an orthonormal basis of E2 ;

then the family (e i ® J;)ieI,jeJ is an orthonormal basis of E and we have

Tr(u) = L L <ei ® fjlu(ei ® fj)

ieI jeJ
= L L <edul(e). <fjluifj)
ieI jeJ

In particular, if U l and U 2 are positive endomorphisms with finite trace, then so

is u. By linearity, we deduce that u belongs to .!l'l(E) when K = C and that the
U i belong to .!l'l(Ei) for i = 1, 2; formula (31) extends to this case by linearity.
Finally, the case when K = R and the ui E .!l'l(Ei) reduces to the complex case by
extension of the scalars.

Remark 2. - Let E be a hilbertian space, which is the hilbertian sum of a family (EJiEI
of hilbertian subspaces. Let u be an element of 2(E) such that u(Ej) c Ei for all i E I;
let ui be the element of 2(E) which coincides with u on E i . Then Tr(u) = L Tr(u)
ieI
when u is positive, or belongs to 21(E); this relation follows from formula (25) of V,
p. 50 applied to an orthonormal basis of E which is the union of orthonormal bases
of each of the E i .

7. Hilbert-Schmidt mappings

DEFINITION 9. - Let E and F be two hilbertian spaces. A continuous linear mapping

u from E into F is called a Hilbert-Schmidt mapping if the trace of the positive endo-
morphism u*u of E is finite. The set of all Hilbert-Schmidt mappings from E into F
is denoted by .!l' 2(E, F).
When E = F, we write .!l' 2(E) for .!l' 2(E; E).
For every uE.!l'(E,F), let IIull2 = Tr(u*u)l!2, so that u belongs to .!l'2(E;F)
if and only if II u 112 is finite. By the definition of the trace, we get
m
(32) Ilull~ = sup
Xl, ... ,Xm
L
i=l
Ilu(x;)112

where (Xl' ... , Xm) range over the set of finite orthonormal sequences in E. In parti-
cular, taking m = 1 in formula (32), we have

(33)

Let (e)ieI be an orthonormal basis of E and (fj)jeJ an orthonormal basis of F.

By formula (25) of V, p. 50 and the Parseval's relation (V, p. 22), we have

(34) Ilull~ = L
ieI
II u(ei) 112 = L
i,j
I<fjlu(ei) 12 .
No.7 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.53

Since 1<fjl u(ei» 1= 1< e;lu*(fj» I, formula (34) implies that

(35) IIu*II 2 = IIull ;
hence, the adjoint of a Hilbert-Schmidt mapping is a Hilbert-Schmidt mapping.
Let E 1 , F 1 be hilbertian spaces and v: El -> E, w: F -> F 1 continuous linear map-
pings. From (32), we deduce immediately that

(36)

By (35), (36) and the relation uv = (v*u*)*, we get

(37)

In particular, if u belongs to !£ 2(E, F) then wuv belongs to !£2(E 1 , F 1).

THEOREM 1. - Let E and F be two hilbertian spaces.

(i) The set !£2(E, F) is a vector subspace of !£(E; F) and u f-> I ull z is a hilbertian
norm (V, p. 6) on !£Z(E; F).
(ii) The isomorphism efrom F ® E' onto !£f(E; F) characterized by e(y ® x*) = yx*
e
extends to an isomorphism from F ®z E' onto !£2(E ; F). In particular, 2 f (E; F)
is dense in 2 2(E; F).

Let (e;)iEI (resp. (fj) jEJ) be an orthonormal basis ofE (resp. F). For every U E 2 (E ; F),
let A(u) be the matrix of u with respect to chosen orthonormal bases for E and F
(V, p. 22). Let I aliz denote the norm of an element a of the hilbertian space e2 (J x I).
By formula (34), A is a mapping from 2 2(E; F) into fZ(J x I) such that IIA(u) 112 = I ull ;
it is clear that A is injective. To prove (i), it is enough to prove that A is surjective.
Let a = (a ji ) be an element of e2 (J x I); by Cauchy-Schwarz inequality, we have

for every; = (0 in f2(I) and 11 = (11) in e2 (1). Then there exists a continuous ses-
quilinear form <l> on F x E such that <l>(y, x) = L
lljaji~i for x =
j,i
L
~iei in E and
i
Y = I lljj in F. Let u E 2(E; F) be such that <l>(y, x) = <ylu(x) >(V, p. 16, cor. 2).
j
We get

hence a = A(u).
Since A is a hilbertian space isomorphism from 22(E; F) onto e2 (J x I) and
since (fj ® en
is an orthonormal basis of F 0 z E', there exists an isomorphism {)
from F O2 E' onto 2Z(E; F) such that

<fjI8(t) e) = <fj ® e1lt)

TVS V.54 HlLBERTIAN SPACES §4

for every i E I, j E I and tE F <8>2 E'. In particular, for t = Y ® x*, we find

hence e(y ® x*) = yx*. This proves (ii). Q.E.D.

Examples. - 1) Let I and J be two sets. By the proof given above, in order that a
mapping u from £2(1) into f2(J) be a Hilbert-Schmidt mapping, it is necessary and
sufficient that there exists a matrix (a j ) in f2(J x I) such that U@j = L aji~i for
iEI
all ~ = (0
in £2(1).
* 2) Let X and Y be two Hausdorff topological spaces, endowed respectively
with positive measures ~ and v. We can show that the Hilbert-Schmidt mappings
from 22(X) into 22(y) correspond bijectively to classes of square integrable
functions on Y x X; to the class of a function N E 22 (Y X X, v ® ~) corresponds
the mapping UN given by

(38) (UN!) (y) = Ix N(y, x)f(x) d~(x)

for v-almost all y E Y and f E 22(X, ~). We have

(39) I uNII~ = Ix Iv IN(y, x)i2 d~(x) dv(y) . *

Remarks. -1) Suppose K = C. Let U and v be in 22(E; F). We have the relation
4 u*v = L e(u + EV)* (u + EV), hence u*v belongs to 21(E). The scalar product
£4= 1
in the hilbertian space 2 2(E ; F) is given by
(40) <ulv) = Tr(u*v)
since this formula defines a hermitian form on 22 (E ; F) and we get u Iu) = I u I ~ . <
If u E 22(E; F) and v E 22(F; E), then vu belongs to 21(E) and uv to 21(F)
by the preceding; moreover, we have
(41) Tr(uv) = Tr(vu) .
By linearity and continuity, it is enough to verify this formula when u = y 1 xi
and v = x 2yi (with Xl' x 2 in E, Y1' Y2 in F); but then uv is the mapping
YHY1<x 1Ix 2> <Y2Iy) and vu the mapping xHx2<Y2IY1)<x1Ix), and (41)
follows from formula (22) of V, p. 48.
Consequently, if u 1 ' u 2 are two elements of 2 2(E ; F), we have, in the hilbertian
space 22(F; E),

(42) <uilui) = Tr(u 1ui) = Tr(uiu1) = <u 2 Iu 1 ) = <u1Iu2);

in other words, u H u* is an isomorphism from the hilbertian space 22(E; F)

onto the conjugate (V, p. 6) of the hilbertian space 2 2(F; E). If we identify this
No.8 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.55

conjugate with the dual of ~2(F; E) (V, p. 15), we see that ~2(E; F) can be iden-
tified with the dual of ~ 2(F ; E), the canonical bilinear form (v, u) H < v, u> being
identified with (v, u) H Tr(vu).
2) Suppose K = R. We leave it to the reader to verify that formulas (40) and
(41) are again valid, and to show that ~ 2(E; F) can be identified with the dual of
~ 2(F; E) by means of the bilinear form (u, v) H Tr(uv).

8. Diagonalization of Hilbert-Schmidt mappings

THEOREM 2. - Let E and F be two hilbertian spaces and u a Hilbert-Schmidt map-

ping from E into F. There exists an orthonormal basis (e)iEI of E which is transformed
by u into an orthogonal family in F.
Let B denote the (closed) unit ball of E, with the weakened topology assigned
to it; this is a compact space (V, p. 17). We put Q(x) = I u(x) 112 for all x E B. Finally
let P denote the set of all vectors x in E satisfying the following property ;
(H) For every y E E orthogonal to x, the element u(y) of E is orthogonal to u(x).

Lemma 4. - The function Q: B --+ R is continuous.

Let Uj)jEJ be an orthonormal basis of F. Put Aj = I u*(fj) 112 for all j E J. Since
u belongs to ~ 2(E; F) we have u* E ~ 2(F ; E), hence I Aj < + 00. Further, we
j
have

(43) Q(x) = Ilu(x)112 = II<u*(fj)lx>1 2

j

by Parseval's formula (V, p. 22) and the definition of the adjoint (V, p. 38). For
every xEB, l<u*(fj)lx>1 2 :::; Aj by Cauchy-Schwarz inequality; consequently, the
convergence of the sum in formula (43) is uniform on B, hence lemma 4 (GT, X,
§ 1, No.6).

Lemma 5. - Let El be a closed vector subspace of E, stable under u*u. If El #- {O},

then there exists a vector of norm 1 in El n P.
Since B is weakly compact, so is the weakly closed subspace B n El of B. Hence

° °
there exists (GT, IV, § 6, No.1, tho 1) a point Xo in B n El such that Q(x o ~ Q(x) for all
x E B n E 1 . If Q(x o) = 0, we have Q(x) = and so u(x) = for all x E B n E 1 .
Thus El c P and lemma 5 follows in this case.
Suppose now that Q(x o) > 0, then Xo #- 0. Since the vector I Xo 11- 1 .x o belongs
to B n E 1 , we have

i.e. II Xo I = 1. We shall prove that Xo belongs to P; let y E E be orthogonal to xo.

It is enough to prove that u(y) is orthogonal to u(x o). But since y is the sum of a
vector ofE I and a vector orthogonal to E 1, and both orthogonal to Xo (since Xo EEl)'
it is enough to consider the following two cases :
TVS V.56 HlLBERTIAN SPACES §4

a) Y is orthogonal to E1 : since E1 is stable under u*u, u*u(x o) E E 1, hence

o= <ylu*u(x o) = <u(Y)lu(xo).
b) Y belongs to E1 : for all teR, the vector x(t) = (xo + tY)/llx o + tyll belongs
to B (\ E 1 • We have Q(xt) = f(t)/g(t) with

f(t) = Ilu(xo)112 + 2tBl<u(x o)lu(y) + t2I1u(y)1I2

get) = 1 + t211yI12 .

In view of the definition of x o ' we have Q(x(O)) ~ Q(x(t)) for all real t, hence
dl Q(x(t)) is zero for t
ct
= O. But f(O) = II u(x o) 11 2 , g(O) = 1, 1'(0) = 2Bl <u(xo)lu(y),
g'(O) = O. Since

~ Q(x(t)) = 1'(t) get) - f(t) g'(t)

dt g(t)2'

<
we conclude that 1'(0) = 0, that is, Bl u(xo)lu(y) = 0. When K = R, u(x o) is
orthogonal to u(y), when K = C, the vector iy belongs to E1 and is orthogonal
to x o , hence <~ <u(xo)lu(y) = -:!ll <u(xo)lu(iy) = 0, and finally u(x o) is ortho-
gonal to u(y). This proves lemma 5.
Now we prove tho 2. Applyingth. 1 ofS, III, §4, No.5 we see, as in V, p. 23, that there
exists a set S which is maximal among the orthonormal subsets of E contained
in P. Let E1 be the set of all vectors orthogonal to S. Let y E E1 ; if XES, the vectors
x and yare orthogonal, and since S c P, we conclude that u(x) and u(y) are ortho-
gonal; then
<xlu*u(y) = <u(x)lu(y) = °
and u*u(y) is orthogonal to S. Hence E1 is stable under u*u. If we had E1 =F {O},
there would exist a vector x of norm 1 in E1 (\ P (lemma 5) and S u {x} would
be an orthonormal subset of E contained in P. This would contradict the maximal
character of S. Hence E1 = {O} and S is an orthonormal basis of E. Q.E.D.

COROLLARY 1. - Let v be a continuous, positive endomorphism with finite trace of

the hilbertian space E. There exists an orthonormal basis (e)iEI of E and a summable
family of positive real numbers (A.)iEI such that vee) = A.iei for all i E 1.
<
Let <l>(x, y) = xl v(y) for x, y in E. Then <l> is a positive hermitian form on E.
There exists (V, p. 8, corollary) a hilbertian space F and a continuous linear mapping
<
u from E into F such that <l>(x, y) = u(x)lu(y) for x, y in E. In other words, we
have v = u*u. By virtue of def. 9 (V, p. 52), u is a Hilbert-Schmidt mapping from E
into F. By tho 2, there exists an orthonormal basis (e)ieI of E such that the vectors
u(ei ) are two by two orthogonal. Let i E I; for every j =F i in I, we have
No.9 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.57

hence vee) is proportional to ei and is of the form Aie i, where Ai = <edv(e) ;

then

Ai ;;:, 0 and I Ai = Tr(v) < + 00.

iEI

COROLLARY 2. - Let E be a hilbertian space. Then 2'1(E) c 2'2(E).

The real case reduces to the complex case by the extension of scalars; we can
therefore assume that K = C.
Since 2' 2(E) is a vector subspace of 2' (E), it is enough to prove that every conti-
nuous and positive endomorphism v of E with finite trace belongs to 2' 2(E). With
the notations of cor. 1, we have

I Ilv(eJl12
iEI
= IAf
iEI
~ (IAY
i
< + 00.

COROLLARY 3. - Let v be a continuous positive endomorphism of the hilbertian space

E with a finite trace. There exists a positive Hilbert-Schmidt endomorphism w of E
such that v = w 2 and such that v commutes with w.
With the notations of cor. 1, it is enough to consider the endomorphism w which
transforms the vector I ~iei into the vector I AlI2~iei·
iEI i

Remark. - With the notations of tho 2, let J be the set of all i E I such that u(eJ # o.
For all i El, letA i = II u(eJ II and J; = le i- 1 u(eJ Then (eJiEJ (resp. (J;)iEJ) is an orthonormal
basis of the initial (resp. final) subspace of u, we have u(eJ = Ie); tor all i E J and
I Ie? = II u II ~ is finite.
iEJ

9. Trace of a quadratic form with respect to another

In this section, E will denote a real vector space and Q, H two positive quadratic
forms on E. There exist two symmetric bilinear forms (x, Y) H xIY)Q and <
(x, y) H <XIY)H on E x E, characterized by

for all x E E.
We call the trace of Q with respect to H, and write Tr(Q/H), a real positive num-
ber, finite or not, defined as follows :
a) If there exists x E E with H(x) = °
and Q(x) #- 0, we put Tr(Q/H) = + 00.
b) Otherwise, Tr(Q/H) is the upper bound of the set of all numbers of the form
m
I Q(xJ where (Xl' ... , X m) range over the set of finite sequences of elements of E
i= I
such that <XdXj)H = 8ij (Kronecker's symbol).

Remarks. - 1) For every subspace F of E, let QF denote the restriction of Q to F

and HF that of H. We have Tr(QF/HF) !( Tr(Q/H) and Tr(Q/H) is the upper bound
TVS V.58 HILBERTIAN SPACES §4

of the set of all numbers Tr(QF/HF) where F ranges over the family of all finite dimen-
sional vector subspaces of E.
2) Let E1 be a real vector space, Q1 and H1 two positive quadratic forms on E1
and 1t: E -> E1 a surjective linear mapping. If Q = Q1 o1t and H = H1 0 1t, then
Tr(Q/H) = Tr(QdH1)'

PROPOSITION 15. - Suppose that there exists a real hilbertian space structure on E
such that H(x) = I X 112 for all x E E. For Tr(Q/H) to be finite, it is necessary and
sufficient that there exists a continuous and positive endomorphism u of E with finite
trace, such that Q(x) = <xlu(x) for all x E E; this endomorphism u is unique, and
we have
(44) Tr(u) = Tr(Q/H) = L Q(e)
iEI

for every orthonormal bases (ei)iEI of E.

Suppose that Tr(Q/H) is finite. For every x E E of norm 1, we have H(x) = 1,
hence Q(x) :::;; Tr(Q/H). Therefore, Q(x) :::;; Tr(Q/H).llxI1 2 for all x E E, and
1< xly )QI :::;; Q(x)1/2Q(y)1/2 :::;; Tr(Q/H)·llxll·IIYII

by the Cauchy-Schwarz inequality. Consequently, the bilinear form (x, y) ~ <xIY)Q

on E x E is continuous. There exists (V, p. 16, cor. 2) a mapping u E 2(E) such
that <xIY)Q = <xlu(y). We have <xIY)Q = <ylx)Q for x, Y in E, hence u is hermi-
tian; and <xlu(x) = Q(x) ~ 0, hence u is positive.
Conversely, let u be a continuous and positive endomorphism of E such that
Q(x) = <xlu(x) for all x E E. Then
<xlu(y) = !(Q(x + y) - Q(x) - Q(y)) = <xIY)Q'
which gives the uniqueness of u. By formula (24') (V, p. 50), we get
m m
Tr(u) = sup L <xdU(x i) = sup L Q(x i) ,
Xl,,,.,Xm i=l Xl, ... ,Xm i=l

where (Xl' ... , X m ) range over the set of all finite orthonormal sequences of elements
ofE. By the definition of Tr(Q/H), we get Tr(u) = Tr(Q/H). Finally, for every ortho-
normal basis (e)iEI of E, we have Tr(u) = L <eilu(e) by formula (25) of V, p. 50,
iEI
hence Tr(u) = L Q(eJ Q.E.D.
iEI

Remarks. - 3) Let E and F be two hilbertian spaces and v a linear, not necessarily
continuous mapping from E into F. Let H(x) = II X112 and Q(x) = II v(x) 112 for all
x E E. It follows from prop. 15 that v is a Hilbert-Schmidt mapping ifand only ifTr(Q/H)
is finite, and then Tr(Q/H) = II v II ~.
4) Suppose E is finite dimensional. When the quadratic form H is invertible, prop. 15
< <
applies. Let (el' ... , en) be a basis of E. Put qij = eilej)Q and h ij = eilej)H and intro-
duce the matrices q = (qij) and h = (hij)' Let U be an endomorphism of E such that
Q(x) = <xlu(x»H for all XEE. We have

<xIY)Q = <xlu(Y»H (x,YEE),

No.9 SOME CLASSES OF OPERATORS IN HILBERTIAN SPACES TVS V.59

and hence the matrix of u with respect to the basis (el' ... , en) of E is equal ~o h-1q.
By prop. 15, we have
(45)
If the basis (e l , ... , en) is orthonormal for H, then h is the unit matrix of order n, and
we get
n
Tr(Q/H) = Tr(q) = L: Q(eJ;
i= 1

so that we get formula (44) in this case.

Now suppose that the quadratic form H is not invertible. Let N be the kernel of H,
and let 1t be the canonical mapping from E onto E/N. Therc cxists an invertible qua-
dratic form HI on E/N such that H = HI 0 1t. Let (el' ... , en) be a sequence of elements
of E such that the sequence (n(e l ), ... , 1t(em ») is a basis of E/N, which is orthonormal
for HI' Let (e l , ... , en) be a basis of N. Then (e l , ... , en) is a basis of E and we have

(46)

for all real numbers SI' ... , Sn'

Suppose that for all x E E, the relation H(x) = 0 implies Q(x) = 0: in other words,
suppose that there exists a quadratic form QI on E/N such that Q = Q I 01t. By
remark 2 and prop. 15, we have,

(47)
 Exercises

§ 1

1) Let E be a complex normed space and f a symmetric bilinear form on the underlying
real vector space Eo, such that f(x, x) = II X 112 for all x E E. Show that there exists one and
only one hermitian sesquilinear form 9 on E such that f(x, y) = ~g(x, y) (prove that
f(x, iy) = - f(ix, y) by using formula (5) of V, p. 2), hence g(x, x) = II X 112.

2) Let E be a real or complex normed space. Suppose that for every 2-dimensional vector
subspace P (over R) in E, there exists a symmetric bilinear form fp defined on P x P, such
that fp(x, x) = II X 112 for all x E P. Show that fp is defined unambiguously and that there
exists a hermitian sesquilinear form 9 on E x E such that, for every real plane PeE, we
have fp(x, y) = ~g(x, y), hence g(x, x) = Ilx112. (If E is a real vector space, observe that we
have II x - y 112 + II x + y 112 = 2(11 X 112 + I y 112) for every pair of points of E, and deduce the
identity
Ilx + y + zl12 - Ilx + yl12 - Ily + zl12 - liz + xl12 + IIxl12 + IIyl12 + IIzl12 = 0;

if E is a complex vector space, apply exerc. 1.)

~ 3) Let E be a real finite dimensional vector space with dimension n,f a positive and sepa-
rating symmetric bilinear form on E, and Bf the bounded convex set defined by the relation
f(x, x) ~ 1. If a = (a 1 , ... , an) is a basis of E and A the discriminant of f with respect to this
basis, we call the volume of B f with respect to a the number va(f) = YnIAI-1/2, where
Yn = nnl2/rG + 1). If b = (bi' ... , bn) is a second basis of E, and if

we have vb(f) = 181 va(f).

a) Show that, if f and 9 are two positive, separating symmetric bilinear forms such that
Bf c Bg (which is equivalent to 9 ~ f), then va(f) ~ va(g) (consider a basis for E which is
orthogonal for both f and for g).
b) Let A be a symmetric compact convex set in E, with 0 as an interior point. Show that
§1 EXERCISES TVS V.61

among all positive, separating symmetric bilinear forms f on E such that A c BJ' there
exists one and only one for which the volume of BJ (with respect to the given basis of E) is
the smallest possible. (To show uniqueness observe that if A is contained in BJ and B g , it is
in BU+g)/2 and that we have v.(U + g)/2) :( !(v.U) + v.(g») for every basis a of E which is
orthogonal for both f and g.)
c) Let A be a symmetric compact convex set in E, with 0 as an interior point, and let f be
the positive, separating, symmetric bilinear form such that A c BJ and that BJ has the
smallest possible volume with respect to a given basis of E. Show that there exist points
Xl' ... , X n • U 1 ' ... , Un in E with the following properties :
IX) For every k. we have Xk E A and f(x k, x k) = 1.
~) The sequence (u 1 , .... un) is an orthonormal basis of E for f
n
y) If we put x k = I akpj for 1 :( k :( n, we have akj = 0 for k < j and a~k ? (n - k + l)/n.
j~ 1
(Argue by induction on k. Suppose Xl' ... , X k , Up ... , Uk have been constructed, let P k be the
orthoprojector (for f) from E onto the subspace generated by u 1 ' ... , Uk; for every E > 0,
consider the bilinear form J; defined by

and prove that A ¢ Bf,' using b). For every integer p ? I choose a point Yp in A not belonging
to BJ liP; take for X k + 1 a limit point of the sequence (Yp) such that
k J (Xk+l - P k X k+l' Xk+l - P k X k + l ) ? (n - k)f(P k X k+l' P k X k + 1 ) ;

next choose uk+ 1')

d) Prove the analogues of b) and c) for the positive separating symmetric bilinear forms such
that BJ c A and for which the volume of BJ (with respect to a given bases of E) is the largest
possible.

~ 4) a) Let E be a real or complex normed space. of dimension? 2, having the following

property: the relation II X II = II y II implies the inequality

Show that the norm on E is prehilbertian. (Reduce to the case when E is real and of dimension 2,
by means of exerc. 1 and 2 of V. p. 60. In this case, let A be the unit ball of E, and let f be
the positive, separating, symmetric bilinear form such that A c BJ and such that the volume
of BJ with respect to a given basis is the smallest possible. Let Xl. x 2 be the two points
constructed in exerc. 3, c). Show that the points of intersection of the circle f(z, z) = 1 and
the bisectors of the two vectors Xl' x 2 also belong to A, and conclude, by iteration, that
A = Bf')
b) Let E be a real or complex normed space, of dimension? 2, having the following property:
the relation II X + y II = II X - y II implies II X + y 112 = II X 112 + II y 112. Show that the norm on
E is prehilbertian (reduce to a».
c) Prove the analogue of a) when we assume that the relation II X II = II y II implies the inequality

(use exerc. 3, d».

5) Let E be a real or complex vector space, of dimension ? 2. Suppose a mapping X ~ II X II
from E into R+ is given, satisfying: II Ax II = 1)"'1.11 X II for every scalar )"" that II X II = 0 implies
X = 0 and that we have the « ptolemaic inequality»

II a - c 11.11 b - d II :( II a - b 11.11 c - d II + II b - c 11.11 a - d II

for every a, b, c, d in E.
a) Show that II X II is a norm on E (replace d by 0 and b by - a).
TVS V.62 HILBERTIAN SPACES §1

b) Show that the norm on E is prehilbertian. (From the ptolemaic inequality deduce the
inequality II x + y liz + II x - y 112 ? 411 x 11.11 y II and use exerc. 4, c).) Prove the converse
(show that in a hilbertian space, if we put d = al II a liz, b' = bl II b 112, we have the equality
lid - b'll = Iia - bll/llall·llbll)· If Iiall = Ilbll = Ilcll = Ildll and if the four vectors a, b, c, d
are in the same plane, the two members of the ptolemaic inequality are equal.

~ 6) a) Let E be a real or complex normed space, of dimension ? 2. Show that for every
x oF 0 in E and every real number rt > 0, there exists an element y in E such that II y II = rt
and Ilx + yliZ = Ilxlll + Ilyllz.
b) Suppose that if the vectors x, y in E satisfy the relation Ilx + yliZ = IlxliZ + Ily112, then
we also have Ilx - yliZ = IlxliZ + Ily112. Show that the norm on E is prehilbertian. (Using a),
reduce to the case in exerc. 4 : restricting to the case where E is 2-bidimensional, we prove
that if IlxI11 = Ilxzll = 1, y = ~(XI - xz) and if ZE E is such that IlyliZ + Ilzllz = Ily + zllz= 1,
then z = ~(XI + Xl) or z = - ~(Xl + x z)')
c) Suppose that, for every vector x # 0 in E, the set H of all vectors y satisfying
Ilx - yl12 = Ilxllz + IlyliZ is stable under addition. Show that the conclusion of b) holds.
(Reduce to the case where E is real and of dimension 2. Using a) and the compactness of the
unit ball in E, show that H is a closed set containing at least two distinct half-lines with ori-
gin 0; prove that these two half lines are opposite to each other by observing that, if not,
the convex set which they generate would be contained in H and will contain either x or
- x.)

7) Let E be a real or complex normed space, with dimension? 2, having the following pro-
perty : there exists a real number y distinct from 0 and from ± 1, such that the relation
Ilx + yll = Ilx - yll implies Ilx + yyll = Ilx - yyll·
a) Show that if Ilx + yll = Ilx - yll, the convex mapping ¢:~ H Ilx + ~yll from R into R
is not constant on any interval. (Argue by reductio ad absurdum; let (rt. ~) be the largest inter-
val on which ¢ is constant, show that there exists 8 > P close enough to ~ and such that
¢(8) = ¢(~). by observing that if Ilu + vii = Ilu - vii. then Ilu + ynvll = Ilu - ynvll for
every rational integer n.)
b) Show that if Ilx + yll = Ilx- yll, then Ilx + ~yll = Ilx - ~yll for every real number ~.
(With the notations of a), observe that ¢ has a relative minimum at the point ~ = O. using
the fact that ¢(yn) = ¢( - yn) for every rational integer n; deduce that we have ¢(~) = ¢( - ~)
identically, for, otherwise, we get ¢(A) = ¢(I1) for two numbers A, 11 such that A + 11 oF 0
and that in this case ¢ has a relative minimum at the point ~(A + 11).)
c) Deduce from b) that the norm on E is prehilbertian (show first that if Ilxll = IIYII, we have
II rtX + Pyll = II px + 2yll for every pair of real numbers rt, P and that the equality II x + yll = Ilx - yll
implies IlrtX + Pyll = IlrtX - Pyll for every pair of real numbers rt, p. Next, show that if
II xii = IIYII = 1 and II x + yll = II x - yll, we have the relation II(rt Z - PZ) x + 2rtpyll = rt Z + pz
by using the preceding results, and deduce the conclusion).

8) Let E be a real or complex normed space, having the following property : if x, y, x', y'
are four vectors in E such that

Ilxll = Ilx'll, Ilyll = 11y'11, Ilx + yll = Ilx' + y'11 ,

then Ilx - yll = Ilx' - y'll· Show that the norm on E is prehilbertian (use exerc. 7).

9) Let E be a real hilbertian space, f a continuous linear form on E. Show that on every closed
convex subset A ofE, the function x H II xll z - f(x) is bounded below and attains its minimum
at a unique point of A.

10) Let E be a real hilbertian space, B a bilinear form on E x E, c I ' Cz two numbers> 0 such
that
IB(x, y)1 :S; c11lxll.llyll for every x, yin E;
IB(x,x)l? czllxllz foreveryxEE.
§1 EXERCISES TVS V.63

Show that for every continuous linear form f on E, there exists a unique element xfEE (resp.
YfE E) such that f(x) = B(x f , x) (resp. f(x) = B(x, Yf)) for every x E E.

11) Let E be a hilbertian space, and (x n ) be a sequence of points of E which converges weakly
to a point a. For every Y E E, we put
d(y) = lim.inf Ilxn - yll and D(y) = lim.sup Ilxn - yll .
n----t n-t Cf)

Show that d(y)Z = d(a)Z + Ily - al1 2 and D(y)2 = D(a)2 + Ily - all z . If Cl and ~ are two
real numbers such that 0 ::,; Cl ::,; ~, give examples where d(a) = Cl and D(a) = ~.

d\T 12) a) Show that there exists a number Co > 0 such that, for every real normed vector
space E of dimension n and every integer k ::,; can, there exists a hilbertian norm x I--> Ilxllz
on E such that IIxl12 ::,; Ilxll for all x E E, as well as an orthonormal system (X)I"'Vk ofk ele-
ments of E (for the hilbertian structure) with norms Ilxjll ::,; 2. (Use exerc. 3 of V, p. 60.)
b) Let n, m be two integers > 0 such that n ::,; com. Let E be a real normed vector space of
dimension m. Show that there exists a vector subspace F of E, of dimension n, a positive and
separating symmetric bilinear form (x, y) I--> <xly) on F and an orthonormal basis
{al' a 2 , ... , an} of F such that
t sup I<ajlx)
]
::,; Ilxll ::,; IIxl12

(where I xii ~ = <x Ix») for all x E F. (Apply a) to the dual E' of E.)

13) a) Let (xn)nEN be an infinite sequence in a Banach space E. Show that, in order that the
family (x n) be summable, it is necessary and sufficient that, for every sequence (en) of numbers
equal to I or to - I, the series with the general term (enxn) is convergent (use GT, III, § 5,
exerc. 4).
b) Let (X)I "'j"'n be a finite sequence of points in a hilbertian space E show that

2- n I (II
(£j)
I
j~1
ejXJ2) = I
j~1
IIXjI12,

where (e) ranges over the set of 2n sequences of numbers equal to 1 or to - 1 (use the identity
of the median, cf V, p. 9, formula (14)).
c) Deduce from b) that if (X)iEI is a summable family in a hilbertian space E, the family (II Xi 112)iEI
is summable in R.

d\T 14) Let E be an infinite dimensional Banach space.

a) Show that for every integer N, there exists a sequence (b) 1 "'j"'N ofN vectors in E, of norm 1,
such that, for every sequence (~)I "'j"'N of N scalars, we have
N N
I I ~jbJ2 ::,; 4 I l~jlZ
j~1 j~1

(use exerc. 12, b)).

b) For every sequence (An)n;, 1 of numbers ~ 0 such that I A; < + 00, show that there exists
n
a sequence (xn)n;, 1 of points of E such that Ilxnll = An for all n, and that the series (x n) is sum-
mabie. (Use a) of exerc. 13, a).)
c) Deduce from b) that in every infinite dimensional Banach space, there exists a commuta-
tively convergent series, that is not absolutely convergent (Dvoretzky-Rogers th.).

15) Let E be a complex hilbertian space, El' E z two closed vector subspaces of E, PI' P Z
the orthoprojectors from E onto E 1 , E2 respectively.
a) Show that, in order that PI and P z commute, it is necessary and sufficient that E is the
hilbertian sum of the four subspaces El n E 2 , E~ n E~, E~ n E 2, EI n E~ (where MO denotes
the orthogonal complement of a vector subspace M of E).
TVS V.64 HlLBERTIAN SPACES §1

b) Show that ifEI is finite dimensional and IIP 1 - P 2 11 < 1, then E2 has the same dimension
as El (consider the intersection E~ 1\ E 2)·
c) Show that the endomorphism T = (PI - P 2)2 of E commutes with PI and P 2' and that
the eigen subspace of T corresponding to the eigenvalue 0 (resp. 1) is the direct sum of the
orthogonal subspaces El 1\ E2 and E~ 1\ E~ (resp. E~ 1\ E2 and El 1\ E~).
d) Suppose E is finite dimensional and that T = AI with A # o. Then A > 0 and E is the
hilbertian sum of subs paces of dimension ~ 2, each of which is stable under PI and P 2 (observe
that PI - P 2 is hermitian and deduce that E is the hilbertian sum oftwo subspaces E +, E-
such that P 1 .x - P 2 .x = ,j).x in E+ and P 1 .x - P 2 .x = - ,j).x in E-; then show that
1+,j). 1-,j).._
for xEE+, we have P 1 .x = 2 x + z and P 2 ·x = 2 x + z, with zEE ).
* e) Suppose that El and E2 are finite dimensional. Show that there exists a family (F')'EA
of subspaces of E of dimension ~ 2 such that E, El and E2 are respectively the hilbertian
sums of the families (F.).EA' (F. 1\ E 1).EA and (F. 1\ E 2).EA (use c) to reduce to the case when E
is finite dimensional, then apply d». *

16) Let E be a hilbertian space, and P be a continuous projector on E, i.e. a continuous endo-
morphism of E such that p 2 = P. Show that for P to be an orthoprojector, it is necessary
and sufficient that liP II ~ 1. (To see that the condition is sufficient, consider a vector x ortho-
gonal to the kernel of I - P.)
If P has finite rank, show that there exists a closed subspace F of E, with finite codimension,
containing P(E) and such that the restriction of P to F is an orthoprojector.

17) a) Let E be a real hilbertian space of dimension 2, PI, P 2 two orthoprojectors from E
onto the lines D I , D2 respectively, assumed distinct. Show that for every x E E such that
Ilxll = 1, we have II(P I - P 2 ).xll = sin e, where e is the angle between DI and D2 lying
between 0 and n12, and that for every y # 0 in E, there exists x # 0 such that (PI - P 2 ).x
is collinear with y.
b) Let E be a real hilbertian space, PI' P 2 two orthoprojectors on E, with respective images
E 1, E 2. Show that liP I - P 211 is the lower bound of the numbers sin e, where e is the angle,
between 0 and nl2 of the two lines D 1 , D2 such that DI eEl' D2 c E 2 , DI and D2 being
orthogonal to EI n E 2.
c) Let QI' Q2 be two continuous projectors on E, with images El' E 2 , and let PI' P 2 be the
orthoprojectors onto EI and E2 respectively. Show that IIP 2 - Pili ~ IIQ 2 - QIII. (Observe
that (Q2 - QI) P 2 = (l - QI) (P 2 - PI) and use a) and b).)

~ 18) Let E be a real normed space of dimension;;:. 3. Suppose that there exists a decreasing
bijective mapping ro from the set m of closed vector subspaces of E onto itself, such that
ro(ro(M») = M and M 1\ ro(M) = {O} for every ME m.
a) Show that there exists a linear mapping u from E onto its dual E' defined upto a scalar
factor and such that u(M) = (ro(MW for all M E m. (Considering the case where M is 1 dimen-
sional, apply the fundamental tho of projective geometry (A, II, § 9, exerc. 16) by observing
that the only automorphism of the field R is the identity mapping (GT, IV, § 3, exerc. 3).
b) If we put < xly) = < x, u(y», show that < xix) # 0 for all x # 0 and that the relations
<xly) = 0 and <ylx) = 0 are equivalent. Deduce that <ylx) = <xly) for every pair of
points x, y of E (consider a number A E R such that <AX + ylx) = 0).
<
c) Show that x Ix) has the same sign on the set of all x # 0; replacing u by - u if necessary,
we can then assume that <xly) is a positive, separating symmetric bilinear form on E x E.
d) Let:Yo be the initial topology of E. Show that the topology :Y on E, defined by the norm
< xix )1/2 is finer than the topology :Yo (observe that the dual of E for :Y contains the dual E'
of E for :Yo).
e) Show that u is a continuous mapping from E onto its dual E', for the topologies cr(E, E')
and cr(E', E). Deduce that if E is complete for the initial topology :Yo, then u is continuous
for :Yo and for the strong topology /3(E', E) (observe that u transforms every set bounded
for cr(E, E') into a bounded set for cr(E', E»). Deduce that then the topologies :Y and :Yo are
§1 EXERCISES TVS V.65

identical, and w(M) is the orthogonal complement of M for the hilbertian space structure
defined on E by the form < x Iy > (cf I, p. 17, tho 1).
f) Show that in the space fl(N), with the norm induced by that of fOO(N) assigned to it, there
exists a bijective mapping M H w(M) from 9Jl onto itself, having the properties mentioned
above (IV, p. 47, exerc. 1).

~ 19) Let E be an infinite dimensional complex normed space. Suppose that there exists a
bijective mapping w from the set 9Jl of closed vector subspaces of E onto itself, having the
properties listed in exerc. 18.
a) Show that there exists a semi-linear mapping u from E onto its dual E' (for the automor-
phism ~ H ~ of C) defined up to a scalar factor and such that u(M) = (w(M)t for every M E 9Jl.
(Proceed as in exerc. 18; using IV, p. 65, exerc. 16, show that u is a semi-linear mapping relative
to the identity automorphism of C or to the automorphism ~ H ~; finally prove that the
first case cannot occur since < x, u(x) > =1= 0 for x =1= 0.)
b) If we put <ylx> = <x, u(y», show that <xly> = <ylx> and that <xix> has the same
sign on the set of all x =1= 0 (same method as in exerc. 18).
c) Finally show that the topology defined by the norm < x Ix> 1 (2 is finer than the initial topo-
logy :Yo on E, and that these two topologies are identical when E is complete for :Yo; in the
latter case, w(M) is the orthogonal complement of M in the hilbertian space E.

20) Let E be a real finite dimensional vector space, and <p be a bijective linear mapping from E
onto its dual E*. Let A be a symmetric compact convex set in E, with 0 as an interior point;
assume that for every x in the boundary of A, the hyperplane with equation < y - x, <p(x) > = 0
is a support hyperplane for A.
a) Let f(x) = 1< x, <p(x) > I, and let a be a boundary point of A where f(x) attains its minimum.
Show that for every point b such that < b, <p(a) > = 0, we also have < a, <p(b) > = O. (Observe
that < x, <p(x) > =1= 0 for x =1= 0, and so we can assume that f(x) = < x, <p(x) > ~ 0 ; use the
fact that every support hyperplane of A at the point a is also a support hyperplane of the set
defined by f(x) :( f(a).)
b) Show that (x, y) H < x, <p(y) > is a symmetric bilinear form, and that A is identical with the
set of all points x such that f(x) :( y for a suitable constant y. (Argue by induction on the
dimension of E; with the notations of a), consider the hyperplane with the equation
<x, <p(a) > = 0.)
21) Let E be a finite dimensional complex vector space, and let <p be a bijective semi-linear
(relative to the automorphism ~ H ~ of C) mapping from E onto its dual E*. Let Ilxll be a
norm on E such that, for all xEE, we have I<x, <p(x) >1 = Ilxll.II<p(x)ll. Show that
(y, x) H < x, <p(y) > is, up to a constant factor, a positive separating hermitian form and that
<x, <p(x) > = yllxl1 2 (y constant). (Argue as in exerc. 20.)

~ 22) Let E be a real normed space of dimension ~ 3, such that, for every homogeneous
plane P in E, there exists a continuous projector from E onto P, of norm 1. Show that the
norm on E is prehilbertian. With the help of V, p. 60, exerc. 2, reduce to the case where E
is of dimension 3, and establish successively the following propositions.
a) For every homogeneous plane P in E, there exists a unique continuous projector from E
onto P, with norm 1, and the kernel of this projection is a homogeneous line D(P) such that
P H D(P) is a continuous bijection from the space of homogeneous planes of E into the
space of homogeneous lines of E (GT, VI, ~ 3, No.5).
b) Every point ofthe sphere D: Ilxll = 1 in E is extremal in the ball B ofE defined by Ilxll :( 1.
(First show that if XES is not extremal, its section F x in B (II, p. 87, exerc. 3) will be 2-dimen-
sional, by considering all the homogeneous planes P passing through x, next prove that this
hypothesis is contradictory, by proceeding similarly at a point in F x where there exists only
one support line of F x in the plane generated by F x; the existence of such a point can be esta-
blished by using II, p. 88, exerc. 7 and p. 88, exerc. 8).
c) Every point of the sphere S' with equation Ilx'll = 1 in the dual E' of E is extremal in the
ball B' of E' defined by Ilx'll :( 1. (First observe that for every homogeneous line D' of E',
there exists a unique homogeneous plane P'(D') in E' such that for every point in S' n P'(D'),
TVS V.66 HlLBERTIAN SPACES §1

the support plane of B' at this point (unique by a)) is parallel to D'; moreover, the mapping
D' I-> P'(D') is continuous. Deduce that if x' E S' were not an extremal point in B', its section
F x' in B' would be of dimension at least 2; for this consider all the homogeneous lines D'
parallel to the support plane of B' at the point x'. Next show that this hypothesis implies a
contradiction, by considering a point of strict convexity y' of F x' (II, p. 88, exerc. 8), and the
unique homogeneous line D~ parallel to the support line of F x' at the point y' in the plane
generated by F x' , and prove that the function D' I-> P'(D') would not be continuous for D' = D~.)
d) Show that, if three homogeneous planes P l ' P 2' P 3 in E contain the same line ~, then
the three lines D(P 1)' D(P 2)' D(P 3) are in the same homogeneous plane 1t(~) (consider the
unique support plane of B at a point of intersection of ~ and of S). Applying the fundamental
theorem of projective geometry (A, II, § 9, exerc. 16) deduce that there exists a bijective linear
mapping <j> from E' onto E such that, for all x' E E', the point <j>(x') belongs to the line D(P),
<
where P is the plane with the equation y, x') = O. Show that for every point x' E S', the
<
plane with the equation <j>(x'), y' - x') = 0 is the support plane of B' at this point, and
conclude by applying Y, p. 65, exerc. 20.
~ 23) Let E be a complex normed vector space of dimension? 3, such that for every (com-
plex) homogeneous plane P in E, there exists a continuous projector from E onto P with
norm 1. Show that the norm on E is prehilbertian. Using Y, p. 60, exerc. 2 reduce to the case
where E is of dimension 3 over C, and proceed as in exerc. 22. (For part b) of the proof, consider,
for every x' E E' such that Ilx'll = 1, the convex set G x ' of all XES such that <x, x') = 1,
show that if G x ' is not simply 0, it would have dimension at least 3 over R; then in the real
affine linear variety generated by G x" consider a boundary point of G x', where there exists
only one support hyperplane (real) of G x " Similarly, for part c) of the proof, consider, for
<
all XES, the set G~ of all x' E S' such that x, x') = 1 and show that G~ reduces to a point;
for this, prove that, if not, the real affine linear variety generated by G~ will have dimension
at least 2 over R, and will contain two linearly independent vectors over C. Conclude using Y,
p. 65, exerc. 21.)

-r 24) In a real normed space E of dimension ? 3, we say that a vector y is quaSi-normal

to a vector x, if for every scalar Ie, we have I x + ley I ? I x II.
a) Show that, if the relation «y is quasi-normal to x }) is symmetric in x, y, then the norm
on E is prehilbertian (show that the condition of Y, p. 65, exerc. 22 is satisfied).
b) Show that the same conclusion holds, if for every closed homogeneous hyperplane H
in E, there exists a vector =1= 0 which is quasi-normal to all the vectors of H. (Same method,
apply tho 2 ofE, III, § 2, No.4 to the continuous projectors of norm 1 from the vector subspaces
containing P onto a homogeneous plane P, these projections being linearly ordered by the
relation of extension.)
c) Show that the same conclusion holds if for every vector x =1= 0 in E, there exists a closed
hyperplane H such that x is quasi-normal to all the vectors of H. (Reduce to the case where E
is of dimension 3, and apply Y, p. 65, exerc. 22 to the dual of E).
d) Show that the same conclusion holds if, when z is quasi-normal to x and y, then z is quasi-
normal to x + y (apply Y, p. 65, exerc. 22).

25) a) Let E be a real normed space and x' =1= 0 a vector in the dual E' of E. Show that for
every vector y in the hyperplane X,-I(O) to be quasi-normal to x (exerc. 24), it is necessary
<
and sufficient that x, x') = I x 11.11 x' II.
b) Deduce from a) that for all x =1= 0 in E, there exists a closed homogeneous hyperplane H
of E such that every vector y E H is quasi-normal to x.
c) If x, yare two points in E and x =1= 0, then there exists a scalar ct. such that ct.x + y is quasi-
normal to X.

26) A real normed space E is said to be smooth if all the points of the unit sphere in E are
points of smoothness (II, p. 87, exerc. 6) of the unit ball. For this to be so, it is necessary and
sufficient that there exists a unique positively homogeneous mapping f from E - {O} into
E' - {O}, such that Ilf(x)11 = 1 for Ilxll = 1, and that <x, f(x) = Ilxll.llf(x)ll. Show that
the following properties are equivalent
ct.) E is smooth.
§1 EXERCISES TVS V.67

[3) For all x oft 0 in E and all y E E, there exists a unique scalar a such that ax + y is quasi-
normal to x.
y) For every x E E, if y and z are quasi-normal to x, y + z is quasi-normal to x.
(To see that y) implies [3), observe that if ax + Y and [3x + yare quasi-normal to x, then
(a - [3) x is quasi-normal to x.)
27) A real normed space E is said to be strictly convex if all the points of the unit sphere are
points of strict convexity (II, p. 87, exerc. 6) of the unit ball. Show that, for E to be strictly
convex, it is necessary and sufficient that for all x oft 0 in E and for all y E E, there exists a
unique scalar a such that x is quasi-normal to x + y. (Observe that the mapping t I--> II tx + yll
is convex in R.)
28) Let E be a normed space, E' its dual.
a) Show that if E' is smooth (V, p. 66, exerc. 26), E is strictly convex (if x and yare such that
x oft y, Ilxll = Ilyll = IIt{x + y)11 = 1, consider an x' E E' such that Ilx'll = 1 and
<t(x + y), x') = 1).
b) Show that if E' is strictly convex, E is smooth.

~ 29) Let E be a normed space, E' its dual. A mapping f from E - {O} into E' - {O} is
said to be a support mapping if it is positively homogeneous, and if for every x E E such that
Ilxll = 1, we have II f(x) II = 1 and <x, f(x» = 1. For E to be smooth (Y, p. 66, exerc. 26),
it is necessary and sufficient that there exists a unique support mapping from E - {O} into
E' - {O}.
Let S be the unit sphere in E, S' the unit sphere in E', and let Xo E S. The following con-
ditions are equivalent :
a) Xo is a point of smoothness of the unit ball in E,
[3) There exists a support mapping f whose restriction to S is continuous at the point Xo
when S is assigned the norm topology, and S' the weak topology a(E', E).
y) For every y E E, the mapping t I--> Ilxo + tyll has a derivative at the point t = O.
(To see that a) implies [3), argue by reductio ad absurdum, using the weak compactness of the
unit ball in E'. To see that [3) implies y), reduce to the case where E is 2-dimensional and use
the fact that t I--> Ilxo + tyll is convex.)
Then every support mapping is continuous at the point xo'

30) Let E be a Banach space, E' its strong dual, E" the strong dual of E', E'" the strong dual
of E", E IV the strong dual of E"'.
a) Suppose that E is non-reflexive; then there exists x' E E' such that Ilx'll = 1, but that we
<
do not have x, x') = 1 for any x E E with Ilxll = 1 (IV, p. 57, exerc. 25). On the other hand,
there exists a sequence (x~) of points of E' such that Ilx~11 = 1, tending strongly to x', and
<
a sequence (x n ) of points ofE such that Ilxnll = 1 and x n ' x~) = 1 for all n (II, p. 77, exerc. 4).
Show that in E", the sequence (x n ) does not converge to any point for the topology a(E", E"')
<
(observe that otherwise it would converge to a point x E E for which x, x') = 1).
b) Show that it is not possible that x' and x~ are points of smoothness ofthe unit sphere in E'"
(observe that x n , considered as an element of E IV will be the unique element x~v E E IV such
<
that Ilx~vll = 1 and x~, x~v) = 1 and use exerc. 29).
c) Conclude that if E'" is smooth, or if E IV is strictly convex, then E is necessarily reflexive.

31) A normed space E (real or complex) is said to be uniformly convex if, for every E such
that 0 < E < 2, there exists 0 > 0 such that the relations Ilxll :( 1, Ilyll :( 1, Ilx - yll ~ E
in E imply Ilt(x + y) II :( 1 - o. A uniformly convex space is strictly convex (Y, p. 67, exerc. 27).
We say that E is uniformly smooth if, for every E > 0, there exists 11 > 0 such that the rela-
tions Ilxll ~ 1, lIyll ~ 1, Ilx - yll :( 11 imply the inequality Ilx+yll ~ Ilxll + lIyll-Ellx- YII·
This is equivalent to : for every E > 0, there exists p > 0 such that the relations Ilxll = 1,
Ilyll :( p imply the inequality
Ilx + yll + IIx - yll :( 2 + Ellyll .
A uniformly smooth space is smooth (Y, p.66, exerc. 26).
TVS V.68 HILBERTIAN SPACES §1

a) Show that if E is uniformly convex, its strong dual E' is uniformly smooth, and that if E
is uniformly smooth, then E' is uniformly convex; the restriction ofthe unique support mapping
(V, p. 67. exerc. 29) to the unit sphere S of E is a mapping from S into the unit sphere S' of E',
which is continuous for the norm topologies of E and E'.
b) Show that if E is uniformly convex, and, if a filter 3' on E converges to Xo for the topology
cr(E, E') and is such that limw I x I = I Xo II, then 3' converges to Xo for the initial topology
of E.
c) Show that a Banach space which is uniformly convex or uniformly smooth is reflexive
(use b) and c) and also IV, p. 60. exerc. 12). (ei V, p.72, exerc. 14.)
d) Generalize the first part of tho 1 of V, p. 10, and also cor. 1 and 2 of V, p. 11 to uniformly
convex Banach spaces.

32) Let E be a normed space (real or complex) of dimension ~ 2, such that, for every 8 such
that 0 < 8 < 2, the relations Ilxll = 1, Ilyll = 1, Ilx - yll ~ 8 in E imply the inequality
8211/2 . Show that the norm on E is prehilbertian. (Reduce to the case
lit(x + y)ll:::.; ( 1 - 4
when E is real and 2- imensional, and argue as in V, p. 61 ; exerc. 4, a).)

-If 33) Let E be a uniformly convex Banach space (V, p. 67. exerc. 31). Then there exists a
number 8 such that ~:::.; 8 < 1 and such that the relation Ilx-yll ~ tsup(llxll, Ilyll) in E
implies lit(x + y)11 :::.; 8 sup (IIxll, Ilyll)·
a) Let (x n) be a sequence of points of E such that II xn I :::.; M and such that the sequence tends
to 0 for cr(E, E'). Show that, if for an index p, II xp I ~ t M, then there exists q > p such that
II xp - Xq II > t M, and consequently I t(xp + xq) I :::.; 8M (argue by reductio ad absurdum,
<
by observing that for all x' EO E' such that II x' I = 1, we have x p' x' = lim xp - x n, x' > n~oc
< ».
Deduce that there exists a strictly increasing mapping e from N into itself such that
Ilt(XI'(2n) + xl'(2n+ 1)11 :::.; M8, and such that if x~1) = t(XI'(2n) + xl'(2n+ I»)' the sequence (x~l))
tends to 0 for cr(E, E') and that I x~1) II :::.; M8 for all n.
b) Show that there exists a sequence (xnJ extracted from (xn) such that, if we put y(k) = x nk '
we have the following property: for every integer p > 1, every integer q < p and every inte-
ger i such that 1 :::.; i :::.; 2 P - q ,
Ily((i - 1) 2q + 1) + y((i - 1) 2q + 2) + ... + y(i2 q ) I :::.; M8 q •

(Iterate the procedure of a) by constructing a sequence (X~k+I») from the sequence (X~k») in the
same way as (X~l») is constructed from (xn); then use a suitable « method of diagonalization ».)
c) Let rand q be two integers > 1. Deduce from b) that if r2 q :::.; k :::.; (r + 1) 2q , then
Ilxn! + x n2 + ... + xnJ :::.; (2 q - 1) M + 2q M + (r - 1) 2q M8 q
(decompose the sum on the left into several subsets, by varying h from 1 to 2 q , then from
(j - 1) 2q + 1 to j2 q for 2 :::.; j :::.; r, then from r2 q + 1 to k).
d) Show that for every bounded sequence (xn) in E, there exists an extracted sequence (xnJ
such that the sequence of the averages (x n! + ... + xnJ/k converges for the initial topology
of E (the Banach-Saks-Kakutani theorem). (Using the fact that E is reflexive, reduce to the
case when the sequence (xn) converges to 0 for cr(E, E'), and use c) for q and r large enough.)

34) Let E be a Banach space and K be a convex, bounded subset that is closed for cr(E, E').
Suppose that for every sequence (xn) in K, t~ere exists an extracted sequence (xnJ such that
the sequence of averages (x n! + ... + xnJ/k IS convergent for cr(E, E'). Show that for every
continuous linear form x' on E, there exists an element x of K such that x, x' = sup y, x' < > < >
<y, x' ».
YEK
(apply the hypothesis to a sequence (xn) of points of K such that <x n, x' >tends to sup
YEK
Deduce that if E has the property of exerc. 33, d), then E is a reflexive space (cf IV, p. 57,
exerc. 25).
* 35) Let E denote a real, finite dimensional hilbertian space of dimension n, S the unit sphere
of E and m the unique positive measure of norm 1 on S which is invariant under the group
§1 EXERCISES TVS V.69

of automorphisms of E. Consider S as a metric space in which the distance is defined by

d(x, y) = Arc cos < xIY). For every XES and r ~ 0, let B(x, r) denote the set of all points y
in S such that d(x, y) (; r; for every subset A of S and every real number r ~ 0, let A, be the
set of all points x in S such that d(x, A) (; r.
a) Given two closed subsets A and B of S, let 8(A, B) be the lower bound of the set of all real
numbers r ~ 0 such that A c B, and B c A,. Show that 8 is a distance on the set g; of all
closed subsets of S, and that g; is a compact metric space for this distance. Show that the
mapping A f-> meA) from .'F into R is upper semi-continuous.
b) Let Xo be a point in S, let H be the hyperplane in E orthogonal to x o, Xl a point in Hand
y the arc of the circle joining Xo to - x o , passing through Xl' i.e. the set of all points in S of
the form Xo sin 8 + Xl cos 8 with 181 (; rc/2. For every y E y, let Hy = H + Y and Sy = S n Hy ;
let m denote the unique positive measure of norm 1 on Sy which is invariant under the group
of al{ automorphisms of E which leave Xo fixed.
Let A be a closed subset of Sand y' the set of all points in y such that A n Sp is non-empty.
For every y E y', there exists a unique real number r(y) such that 0 (; r(y) (; rc and such that
m/A n S) = my(B(y, r(y» n S); let siA) be the union of the sets B(y, r(y») n Sy as y ranges
over y'. Prove that siA) is closed and that meA) = m(siA»).
c) For every closed subset A of S, the infimum rCA) of the set of all real numbers r ~ 0 for
which there exists an XES with A c B(x, r) is called the radius of A. Let M(A) denote the
set of all closed subsets C of S such that m(C) = meA) and m(C E) (; m(A E) for every e > O.
Show that the following conditions are equivalent for every pair (A, B) of closed subsets
of S :
(i) meA) = m(B) and B is of the form B(x, r) with XES and r ~ 0;
(ii) BE M(A) and r(B) (; r(C) for every subset C of A belonging to M(A). (Arguing by
induction on n, we deduce from b) that sy(A) belongs to M(A) for every closed subset A of S ;
if r > 0 is such that A c B(xp r) show that every point of the boundary ofB(x p r) in S which
belongs to s/A) also belongs to A.) *
* 36) The notations are the same as in exerc. 35.
a) Let a be a vector of norm 1 in E, KE the set of all XES such that 1< xla) 1 :( sin e and LE
the set of all XES such that d(x, Sa) ~ e (where Sa is the set of all points of S orthogonal to a).
Show that for e > 0 small enough, we have m(KJ ~ 4e-nE2/2 and m(LE) (; 4e-nE2/2 (we observe
that the image ofthe measure m under the mapping X f-> < xla) from S into the interval (-1, 1)
of R is of the form cn (1 - t 2 )(n- 3) /2 dt with a suitable constant Cn > 0).
b) Let f be a continuous mapping from S into Rand M(f) a real number such that the set
of all XES for which f(x) (; M(f) (resp. f(x) ~ M(f») has a measure ~ for m. Let B be t
the set of all XES such that f(x) = M(f). Deduce from a) that, for every e > 0 small enough,
the set of all XES such that d(x, B) ~ e, has a measure for m at most equal to 4e-nE2/2.
c) For every e > 0, let hen, e) be the smallest integer h ~ I for which there exist points Xl' ... , x h
h
in S such that S = U B(xi' e) show that lim (log hen, e»)/llog el = n.
i= 1 E----j.O
d) Recall that E is a real hilbertian space of dimension n. Let k be a positive integer and e, e'
two strictly positive numbers such that 4h(k, e) < enE·2/2. Let f be a mapping from S into R
such that If(x) - f(y)1 (; Ilx - yll for all x, y in S, and M(f) a real number satisfying the
relation stated in b). Show that there exists a vector subspace F of E, of dimension k, satisfying
the following condition: for every x E F n S, there exists a point yin F n S such that Ilx - yll < e
and If(y) - M(.f)1 (; e'. *
* 37)
f
Let E be a real hilbertian space of dimension n and let y be a positive measure on E

such that ei(xIY)dy(y) = exp( - rcllx112/2) for all x E E (INT, IX, § 6, No.5). Let m be the
E
unique positive measure of norm I on the unit sphere S of E which is invariant under the
group of automorphisms of E.

f f
a) Let p be a continuous function on E, satisfying p(t.x) = t.p(x) for all XES and all real

positive t. Show that pdy = Cn pdm with cn = rc I/2 r(n/2)/r«n + 1)12).

E S
TVS V.70 HILBERTIAN SPACES §2

b) Show that there exists a constant C > 0, independent of n, such that

f ~~~n 1<
E
1 >
xle i Idy(x) ~ c.(log n)1/2
for every orthonormal basis (e1' ... , en) ofE.
c) Let E > 0 and let k be a positive integer. Deduce from b) and exercises 12 (V, p. (3) and
36, d) that if n is large enough, then for even real normed space V of dimension n, there exists
a vector subspace W of V, of dimension k, satisfying the following property: there exists a
real hilbertian space W t' of dimension k, and a bijective linear mapping u from W onto W 1
such that sup (Ilull, Ilu- II):S; 1+ E.

§ 2

1) Let B be an orthonormal basis in an infinite dimensional hilbertian space E.

a) Show that every everywhere dense subset in E has a cardinality at least equal to that of B,
and that there exists an everywhere dense set iri E which is equipotent to B.
b) Show that Card(E) = Card(B N ) (to see that Card(E) :S; Card(B N ) use a»).
c) Show that if Card(B) :S; Card(R), then the cardinality of every algebraic basis of E is equal
to Card(R) = 2~o (use II, p. 80, exerc. 24, c»; however, if Card(B) > Card(R), then every
algebraic basis ofE is equipotent to BN (use b) and A, II, ~ 7, exere. 3, d».

~ 2) a) Let E J , E z be two hilbertian spaces whose respective hilbertian dimensions are

two cardinals 111 and n such that 11t < n :S; m~o. Let E = E J EB E2 be the hilbertian sum
of E l ' E z and let (b))'EL be an orthonormal basis of E 2. Show that there exists an algebraically
independent system (a,hEL in E1 (cf exere. I, c»; let H be the subspace of E generated (alge-
braically) by the family (a, + b,hEL' Show that the hilbertian dimension of II is equal to H
(observe that the orthogonal projection from H onto E2 is everywhere dense, and use exerc. 1,
a». If S is an orthonormal subset of H, show that S n E z = 0 ; and deduce that Card(S) :S; m
(observe that every element of an orthonormal basis of E1 is orthogonal to every element of
S except at most to a countably infinite subset).
b) Let E3 be a hilbertian space with hilbertian dimension p ~ n, and let F be the hilbertian
sum E EB Ey Let G be the subspace H + E3 of F. Show that the hilbertian dimension of
G is p. If T is an orthonormal subset of G, show that T n (E z + E 3) c E 3 ; deduce that
the cardinality of the orthogonal projection from Tonto E z is at most m (argue as in a».
Conclude from this that G does not have an orthonormal basis, by observing that the ortho-
gonal projection from G onto E2 is everywhere dense in E 2.

3) Show that, in every Hausdorff and non complete prehilbertian space E, there exists a
closed hyperplane whose orthogonal subspace in E reduces to O. Deduce that if E satisfies
the first axiom of countability, then there exists a non total orthonormal family in E which
is not contained in any orthonormal basis.

4) Let E be a Hausdorff prehilbertian space, (EJiEI a family of complete vector subspaces

of E, well-ordered by inclusion, such that the union of all Ei is everywhere dense in E. Show
that there exists an orthonormal basis (eJ'EA in E with the following property : for-every
i E 1, the set of all e. belonging to E; is an orthonormal basis of E i. (Consider the set of all
orthonormal subsets S in E such that, for every i E T, every vector of S not belonging to E;,
is orthogonal to Ep and take a maximal element of this set.) From this, deduce a new proof
of the corollary of V, p. 24.

5) Show that for a hilbertian space E with an infinite hilbertian dimension, there exists an
isomorphism from E onto a closed vector subspace of E which is distinct from E.
§2 EXERCISES TVS V.71

6) Let E be a hilbertian space and (e)iEI an orthonormal basis of E. Show that if (a)iEI is a
topologically independent family in E such that l: I ei - a i 112 < + 00, then the family (a)
iEI
is total. (Let J be a finite subset of I; show that there is a continuous linear mapping u from E
into itself such that u(eJ = e i for i E J, u(eJ = a i for i rf J, and that the norm of u - IE can
be made arbitrarily small by a suitable choice of J; then use IV, p. 65, exerc. 17.)

7) Given n points Xi (1 :S; i:S; n) in a Hausdorffprehilbertian space E, we mean by the Gram's

determinant of these n points the determinant

G(X l , ... , xn) = det«xilxj »).

a) Show that G(Xl' ... , xn) ;;0, 0 and that for Xl' ... , Xn to form an independent system, it is
necessary and sufficient that G(X l , ... , Xn) # 0 (assuming that dim (E) ;;0, n, consider an ortho-
normal basis of an n-dimensional subspace containing Xl' ... , xn)'
b) Show that if Xl' ... , xn is an independent system in E, the distance from a point X E E to
the vector subspace V generated by Xl' ... , xn is equal to (G(x, Xl' ... , XJ/G(Xl' ... , x nW/ 2
(find the expression for the orthogonal projection of X on V).
c) Let (xn) be an infinite sequence of points in E. For the family (xn) to be topologically inde-
pendent, it is necessary and sufficient that, for every integer p > 0,

(use b».
8) Let E be a hilbert ian space which has a countably infinite orthonormal basis (en)n;, l'
Let A be the closed convex envelope in E of the set consisting of the points (1 - ~) en for

all n ;;0, 1. Show that there does not exist any pair of points x, y in A whose distance is equal
to the diameter of A (compare with IV, p. 54, exerc. 12).

9) a) Let E be an infinite dimensional real hilbertian space satisfying the first axiom of counta-
bility. Let (an)n;'O be a free family of points in E, such that each of the two families (a 2n ) and
(a 2n + 1) is total in E (II, p. 80, exerc. 26, a». Let F and G be two vector subspaces of E for which
(a 2n ) and (a 2n + 1) are respectively (algebraic) bases. The spaces F and G are put in separating
duality by the bilinearform <ylz). Show that ifB denotes the unit ball in E, then in the space F,
endowed with the topology cr(F, G), the convex set F II B is closed, but does not have any
closed support hyperplane.
b) Let (bn)n,q be an everywhere dense set in B, and for every X E E, let u(x) be the sequence
« b kIX)/k)k;' 1 . Show that u is an injective, continuous linear mapping from E into the hilbertian
space i'~(N) and that u(B) is compact. Show that, in the normed subspace L = u(F) of i'~(N).
the set u(B II F) is closed, convex and precompact, but does not have any closed support
hyperplane (observe that if f is a continuous linear form on L, then f 0 u is a continuous
linear form on F for the topology cr(F, G»).

10) Let E be an infinite dimensional real hilbertian space satisfying the first axiom of counta-
bility, and (en)n;'l an orthonormal basis of E.
a) Let A be the closed convex balanced envelope of the set of all points en/n in E. Show that
A is compact and that there does not exist any closed supporting hyperplane of A at the point 0,
but that there exist lines D passing through 0 such that D II A = {O}.
b) Let F be the hilbertian sum E EB R, eo a vector which, with the en (for n ;;0, 1) forms an
orthonormal basis of F. If B is the closed convex envelope of { eo} u A, show that there exists
a closed segment L with mid-point 0 in F, such that L II B = {O}, but there does not exist
any closed hyperplane passing through 0 which separates Land B (even though there exists
a closed supporting hyperplane of B at 0).

11) Let E l , E2 be two infinite dimensional real hilbertian spaces satisfying the first axiom
TVS v.n HILBERTIAN SPACES §2

of countability, and E the hilbertian sum El EB E2 (which we shall identify with the product
E1 x E 2 ), Let (en)n" 1 be an orthonormal basis of El ; in E 2 , let A be a compact convex set
containing 0, and D a line passing through 0, such that DnA = {O} and suppose there
exists no closed supporting hyperplane of A at the point 0 (exerc, 10), Let (an)' (Pn) be two
sequences of numbers;;:> 0 such that lim Pn = 0 and I
an-I < L Let P be the set of all points
n---> 'ij

I ~"en of EI such that 0 ,;::; ~n ,;::; a" for every n ;;:> 1, Finally, let Q be the closed convex envelope
in E of the set of all points (a,/,,,. x + Pna), where n ~ 1. a # 0 is a fixed point in 0 and x
ranges over A.
a) Show that P n Q = 0 and that there exists no closed hyperplane in E separating P
and Q,
b) Let F be the hilbertian sum E EB R, and let c be an arbitrary point of F not contained
in E. Show that the pointed convex cones PI' Ql with vertex c, generated by P and Q respec-
tively, are closed in F and that there exists no elosed hyperplane in F separating PI and Ql
(to see that PI and 0 1 are closed, prove that neither P nor Q contain the half-line),

12) Let E be an infinite dimensional real hilbertian space. Show that there exist infinitely
many complex hilbertian space structures on E for which E is the real locally convex space
underlying these complex hilbertian spaces (11, p. 61), (To prove the existence of the auto-
morphisms u of the topological vector space structure of E. such that u 2 (x) = - X, use an
orthonormal basis of E; then apply V, p. 60, exerc, 1.) Give an example to show that the pro-
position does not extend to Hausdorff non complete prehilbertian spaces (consider an every-
where dense hyperplane in such a space).

13) Let E be an infinite dimensional hilbertian space satisfying the first axiom of counta-
bility and (en)nEZ an orthonormal basis of E whose set of indices is the set of rational integers.
Let u denote the isometry from E onto itself such that u(e n) = en+ 1 for all nEZ, and put
f(x) = W - Ilxll) "0 + u(x).
a) Let B be th;c unit ball and S the unit sphere in E. Show that the restriction of f to B is a
homeomorphism from B onto itself(observe that the restriction of u to S is a homeomorphism
from S onto itself), and that there does not exist any point Xo E B such that f(x o) = Xo (express
Xo in terms of its coordinates with respect to (en»)'
b) For every x E B, let g(x) be the point of intersection of S with the half-line with origin I(x)
passing through x. Show that 9 is a continuous mapping from B onto S, such that g(x) = x
for all XES (compare with GT. VI, ~ 2, exerc. 8), Deduce that there exists Xo E S and a
continuous mapping h from S x (0, 1) onto S such that /z(x, 0) = Xo and hex, 1) = x for
all XES.

14) a) Let (En)n"O be an infinite sequence of real Banach spaces, E the vector subspace of
the product F = n", En consisting of all sequences x = (x
n=Q
n) such that I I Xn 112 < + 00.

Show that the function Ilxll = (I IlxnI12)t/2 on E, is a norm, and that E is complete for this
norm: we say that E is the hifbertilln sum of the Banach spaces En'
b) Show that the strong dual E' of E can be identified with the hilbertian sum of the strong
duals E~ of the spaces En. and that if x' = (x~) E E', then <x, x') = I <Xn , x~) (ifuis a conti-

nuous linear form on E, Un its restriction to En considered as a subspace of E, and an a point

of En such that II an ',I = 1, show that, for every sequence (An) of real numbers such that
A;
I < + CD the series with the general term Anun(an) is convergent, and deduce that
n
I (u n (a n»)2 < + 00, using Banach-Steinhaus theorem for E2(N), for example).

c) Deduce from b) that when each of the En is reflexive, then E is reflexive. In particular, if
for En we take the space Rn endowed with the norm Ilxll = sup I~il for x = (~Jt"i"n' show
1 ~i~n
§3 EXERCISES TVS V.73

that E is reflexive, but there does not exist any norm compatible with the topology of E and
for which E is uniformly convex (V, p. 67, exerc. 31) .

.,,-r 15) a) For every integer n > 0, let d n ) be the double sequence defined in IV, p. 63, exerc. 8.
Let E be the vector space of double sequences x = (xij) of real numbers such that, for every
integern > O,wehavePn(x) = i (Ia?plx l)1/2
< + 00. ShowthattheP n are semi-norms on E,
i,j
and that E endowed with the topology defined by these semi-norms is a Frechet space and a
Montel space (argue as in IV, p. 60, exerc. 11).
b) Show that the dual of E can be identified with the space E' of all double sequences x' = (x;)
such that, for at least one index n, we have I (a~j)-1 Ix;l < + 00.
i,j

I (I
CD CfJ

c) For every x = (Xi) E E, show that Ixil) < + 00 (use Cauchy-Schwarz inequa-
j::= 1 i= 1
en
lity); for every j ;? 1, put Yj = I xij; the sequence u(x) = (y) then belongs to the hilbertian
i= 1
space [2(N). Show that u is a strict surjective morphism from E onto £2(N); deduce that
there exists weakly compact sets in f2(N) which are not the images under u of a bounded set
in E (argue as in IV, p. 63, exerc. 8).

~ 16) a) Let A be the set of increasing mappings A:N -> R!; for every integer n;? 0, and
every A E A, let <Pn(A) = A(n). Let E be the set of all mappings x: A -> C such that, for every
n E N, we have Pn(x) = (I
IX(A)1 2<pn(A»)1/2 < + 00. Show that E is a vector space on which
A.EA
the Pn are the semi-norms defining a reflexive Frechet space structure.
b) Let B be a bounded set in E, and let rxn = sup Pn(x); let Ao be an element of A such that
XEB

lim Ao(n)-lrx; = O. Show that X(Ao) = 0 for all x E B, and hence that the set B is not total
in E.
c) Let (Un) be a countable fundamental system of convex and balanced neighbourhoods
of 0 in E; ifU~ is metrizable for the strong topology on E', then there exists a sequence (Bnm)m;:,O
of bounded sets in E such that the sets B~m n U~ form a fundamental system ofneigt'lbourhoods
of 0 in U~ for the strong topology. Deduce from b) that there exists an integer n such that
U~ is not metrizable for the strong topology (use exerc. 5 of III, p. 38).

§3

1) Let E be a hilbertian space. Show that the bilinear mapping (u, v) f-> uv from sm(E) x sn(E)
into sm+n(E) is continuous and that its norm is equal to ((mm.;-n.~)!)1/2. (To see that this norm
is bounded by (
(m + n) ')1/2
, , ' , argue as in the case of the exterior algebra (V, p. 35). Deduce
m. n.
that the multiplication in S(E) cannot be extended to S(E) by continuity when E is not simply O.

2) Let E be an infinite dimensional hilbertian space, and let P, q be two integers;? 1; let
p' = [~] q' = [~J (integral parts). Show that the norm of the bilinear mapping (u, v) f-> u 1\ V

from JV(E) x J\q(E) into J\p+q(E) is at least equal to

A A A

P'+ q. ') ')1/2 . (When P = 2p' and

((

p'! q'!
q = 2q' are even, consider a 2n-dimensional subspace En in E, with an orthonormal basis
(e)l "'jOn; let ej = e 2j - 1 1\ e2j for 1 ,;;; j ,;;; n; consider the product u 1\ v, where u = I eiJ,
H
TVS V.74 HILBERTIAN SPACES §4

v = I ei<: where H (resp. K) range over the set of all subsets ofp' (resp. q') elements of {1, 2, ... , n}
K
~nd e~ = e;,. J\ ... J\ e;p (resp. ei<: = e'" J\ ... J\ e"q) if i1 < ... < ip' (re.sp .. jl. < ... < jq.)
ISthe mcreasmg sequence, of elements of H (resp. K). Deduce that the muluplicatlOn in A(E)
cannot be extended to A(E) by continuity.)

§4

1) Let E and F be two infinite dimensional Hilbert spaces satisfying the first axiom of counta-
bility, (an) an orthonormal basis of E, and (b n) an orthonormal basis of F.
a) Let u be a continuous linear mapping from E into F; let u(an) = I ctmnbm. Show that

I Ict mn l2 ~ II ul1 2 and I Ictmn l2 ~ II ul1 2 for every m and n.

m
b) Give an example of a double sequence (ct mn ) such that I Ictmn l2 ~ 1 for all n and I Ictm.1 2 ~ 1
•
for all m, but such that there does not exist any continuous linear mapping u from E into F
<
such that u(an)lbm ) = ctmn for every pair of integers (m, n). (Show that if leN is a set of p
integers, and if V p (resp. W p) is the subspace of E (resp. F) generated by the an (resp. b.) such
that n E I, then there exists a linear mapping up from V ponto W p such that <up(an)lb m ) = )p
for mEl and n E I, and that Ilupll ;;;, .)p.)
~ 2) Let A = (ctmn)(m,n)ENxN be a double sequence of complex numbers, which we also call
an infinite matrix. For every point x = (x n) ofthe direct sum space e(N), the sums Ym = I ctm.x.
n
are defined; let A.x be the point (Ym) ofthe product space eN, then x I-> A.x is a linear mapping
from e(N) into eN, and every linear mapping from e(N) into eN is of this form. Let E. denote
the subspace of e(N) generated by the first n vectors of the canonical basis, P n ·the canonical
projection from eN onto En ; when En is assigned the norm induced by that of the space f~(N),
II ull denotes the norm of the linear mapping u from the finite dimensional hilbertian space En
into itself.
a) In order that the image ofe(N) under the mapping x I-> A.x be contained inf~(N) and that
this mapping extend to a continuous linear mapping from f~(N) into itself, it is necessary
and sufficient that the norms II P nAPnII are bounded. This implies that the rows and the columns
of A belong to f~(N) (exerc. 1).
b) Let A * denote the infinite matrix (ct;"n), where ct;"n = ~.m' If the columns of A belong to
f~(N) (in other words, if x I-> A.x maps e(N) into e~(N»), then the series 13mn = I ~pmctp. are
p
absolutely convergent, and we put A *A = (13mn)' Show that for x I-> A.x to extend to a con-
tinuous linear mapping u from f~(N) into itself, it is necessary and sufficient that the norms
IIPn(A*A)Pnll are bounded (we have <P.(A*A)Pn.xlx) = IIAPf x l12 for all xEE.). Then
x I-> A *A.x extends to a positive hermitian mapping u*u from fdN) into itself.
c) For two infinite matrices X = (~mn)' y = (11mn), we say that the product XY is defined
if the series ~mn = I ~mp11pn are absolutely convergent, and then we put XY = (~m.)' We say
p
that a power X k (k integer> 1) is defined if X k - 1 and X k - 1 X are defined and then we put
X k = X k- 1 X; in this case, xpxq = X k for every pair of integers p, q such that p + q = k.

If A is an infinite matrix whose columns are in f~(N) and if the product (A *A)2 is defined,
< <
show that for all x E En' we have (PnA *AP.)2 .xlx) ~ Pn(A *A)2 Pn.xlx), and deduce that
IIP.A*APnI1 2 ~ IIP.(A*A)2P.II·
d) In order that an infinite matrix A be such that the image of e(N) under x I-> A.x is con-
tained in f~(N) and that x I-> A.x extends to a continuous linear mapping from f~(N) into
itself, it is necessary and sufficient that the following three conditions are satisfied :
(i) the rows and columns of A are in e~(N) ;
§4 EXERCISES TVS V.75

(ii) the powers (A *A)k are defined for every integer k > 1 ;
(iii) we have
sup(supl«A *A);;'m)l/nl) < + OCJ
n m

where (A *A);;'m denotes the term with indices (m, m) of the matrix (A * A)n. (Observe that if
C is the matrix, with respect to the canonical basis of En' ofa positive hermitian endomorphism
of En' then IICII ~ n sup IC;;!, by considering the trace of C and by diagonalizing C. Using
1 ~i~n

the inequality proved in c), show that

IlPnA*APnll ~ n2 - k sup 1«A*A)~kW-k
1 ~i::Sn

for every integer k > 1, if conditions (i), (ii) and (iii) are satisfied.)

~ 3) Let (ai)(i.j)EI xl be a countably infinite double family of complex numbers. Assume that
there exist two numbers p > 0, y > 0, and a family (P)iEI of numbers> 0 satisfying the
relations

for all i, j in 1.
a) Show that there exists a continuous endomorphism u from f~(I), with norm ~ (py)1/2,
such that for all x = (X)iEI in f~(I), we have u(x) = y, where y = (y) is given by Yi = L aUxj
j
(for x = (x) and y = (y) in e(l), put vij = IXil (p)a;)/py/2, wij = Iy) (Pilai)/pYI2, and find
a bound for L V;jW;).
i,j
b) For I, take the set of all integers ? 1, and let aij = (i + j) - I. Show that the conditions (*)
are satisfied withp; = i -1/2 and P = y = 11: (these being the best possible constants) (compare
the series in (*) with an integral). In an analogous manner, treat the case where J = Nand
aij = (i + j + 1)-1 (<< Hilbert's matrix »).
oc
* c) Let Yf' = H2(D) be the Hardy space, consisting of all functions fez) = L a"z", ana-
n=O
lytic in the open disc D : Izl < 1 and such that IIfl12 = L
+ OCJ; then II fll is a norm
la,,1 2 <
"

r
on Yf', for which Yf' is isomorphic to e~(N). Given two functions f; g in Yf', show that the
function t f-4 f(t) get) of the real variable t is integrable on (0, 1) with respect to the Lebesgue

r
measure and that the formula B(f, g) = f(t) get) dt defines a continuous bilinear form

on Yf' x Yf' (consider B(f, f) = f(t)2 dt for a function f E Yf' of the form fez) = ,,~o anz"
with an ? 0 for 0 ~ n ~ N; use Cauchy's theorem to establish the relation

I- 1
I f(t)2dt = - i In f(e
0
ie )2 e i9 d8

from which we get BU; f) ~t fn If(e i8Wd8 = 11: Ilf112. In this way, get the result of b),

according to which the Hilbert's matrix defines an endomorphism of norm ~ 11: of the hil-
bertian space e~(N).) *

4) Let E be a complex hilbertian space of finite dimension d.

a) For every u ? 0 in £l(E) (V, p. 45), prove that there exists a unique v ? 0 in £l(E) such
that u = v2 (diagonalize u); we write v = U l/2 .
TVS V.76 HILBERTIAN SPACES §4

b) For every u E 2'(E), put abs(u) = (u*u)1/2. Show that u and abs(u) have the same norm
and that we have abs(J\"(u») = J\"(abs(u») for every integer n ~ d.
c) Let Sl (u) ~ '2(U) ~ ... ~ siu) ~ 0 be the sequence of eigenvalues of abs(u) counted
with their order of multiplicity. Show that II ull = Sl (u) and that for every integer n ~ d,
II J\"(u) II = Sl (u) '2(U) ... sn(u). In order that II J\"(u) II = Ilull" for all n such that 1 ~ n ~ d,
it is necessary and sufficient that u*u is a homothety, in other words, that u is a scalar multiple
of a unitary operator.
5) Let E, F be two Hilbert spaces, and u a continuous linear mapping from E into F. Let
feu) denote the set of all x E E such that Ilu(x)11 = Ilull.!lxIi-
a) Show that feu) is the closed vector subspace of E which is the kernel of u*u - iiu1121E'
and is orthogonal to the kernel of u. '
b) Show that the restriction of u to feu) is a bijection from feu) onto £(u*), whose inverse bijec-
tion is the restriction of II ull- 2. u* to f(u*) ; moreover the image of the orthogonal complement
(e(u)t under u is contained in (£(u*)r. If u 1 is the restriction of u to (C(u»)O, considered as a
mapping from (f(u)t into (f(u*)t, the adjoint ui is the restriction of u* to (e(u*)t ; if feu) # E,
< <
let u), called the sub-norm of u, denote the norm II u 1 11 ; if feu) = E, we put u) = O. Then
<u*) = <u).

-0 6) Let E be a hilbertian space, M,N two closed subspaces of E and Me, N° their respective
orthogonal complements: let PM' PN denote the orthogonal projections on M and N respective-
ly, such that 1j - PM' IE - fiN are the orthogonal projections on M S and N respectively.
Put UNM = (IE - P'I) PM and oeM, N) = IluNMII; we have oeM, N) = o(N, M O ) ~ 1; the
relation oeM, N) < 1 implies M n N° = {O}.
a) Let M denote the orthogonal complement in M ofM n N and let elM, N) = oeM. N). Show
C

<
,

that (with the notations of exerc. 5) f(u NM ) = M n N S and deduce that e(M, N) = UNM ) ~
oeM, N); moreover, ifM n N° = {O} (and in particular if oeM, N)< 1), then e(M, N) = oeM, N).
b) For a continuous linear mapping u from E into itself, we designate by conorm of u the
number c(u) = inf Ilu(x)ll/llxll, where x ranges over the set of all vectors # 0 orthogonal
to u- 1(0) (if u = 0, put c(u). = 1). For u(E) to be closed in E, it is necessary and sufficient that
c(u) > 0 (I, p. 17, tho 1). We have c(u*) = c(u).
c) Let VNM = Pr;PM' Show that
e(M, N)2 + C(VNM )2 =
(observethatlluNM(x)!!2 + Ilv NM (x)11 2 = IlpM.xI12, and that the kernel ofvNMis M +(MnN°)
<
and deduce that UI'M)2 ~ 1 - c(VNM )2).
d) Deduce from h) and c) that e(N, M) = e(M, N) and, using a), that dM. N°) = e(M, N).
e) Put geM, N) = IlpM - P'III (Cl V, p. 64, exerc. 17). Show that
geM, N) = sup(o(M, N), o(N. M»)

(observe that PM - PN = (IE - PN ) PM - PN(lE - PM»); deduce that e(M, N) ~ geM, N).
If M n N" = N n MO = {O}, we have the relation elM, N) = oeM, N) = Ii(N, M) = geM, N).
If q(M, N) < 1, then M n N° = N n MO = {O}.
f) Let QM' QN be two continuous projections in E, with respective images M and N ; give
another proof of the relation geM, N) ~ Ii Qr:/ - QNII (V, p. 64, exerc. 17). (Note that for all
x E E, we have II(1E - QM).xI1 2 + 11 Qr.";.xll = Ilxl:2 + II(QM - Q~).xI12, and apply this
relation to x = (PM - PN)'Y' observing the relations (1E - QM) (PM - PN) = (QM - Q'I) PN
and (PM - P N) Q\1 = (lE - PN) (QM - QN)·)

-0 7) a) With the notations of ex ere. 6, show that, for M + N S to be closed. it is necessary

and sufficient that M + N° = (M O n Nt.
b) Show that the following properties are equivalent:
ex) elM, N) < 1 :
B) M + N° is closed in E;
y) If M is the orthogonal complement of M n N° in M and (N°)- that of M n N° in N°,
then E is the direct sum of MO n N, M n N°, M and (N°) - . Moreover, if Rand S are respective-
ly the projections from E onto M and (N corresponding to this decomposition, then
G
)-
§4 EXERCISES TVS V.77

IIRII = IISII = (1 - c2(M, N))1/2. (To see that IX) implies ~) observe first that if x EM and
yE (N G
then l<xly)1 ~ c(M, N) Ilxll.llyll; then, let u = x + Y + t be the decom-
)-,

position of an element u E M + N° with x E M, Y E (N°) - and t E M n N°, deduce that

Ilxll ~ (1 - c 2(M, N)tl/2 Ilull, Ilyll ~ (1 - c 2 (M, N)t 1/2 Ilull and Iltll ~ Iluli. To prove that
y) implies IX), consider the decomposition v = VI + V 2 for a v EM, where v, is the orthogonal
projection of v onto N, the orthogonal complement of N n MO in N; we have R. VI = v, and if
we had E(M, N) = 1, there would exist a sequence (v n) E M such that I vnll = 1 and such that II(vn )111
tends to O. Next show that the restriction RI of R to N is a bijection from N onto M; to cal-
culate IIRII, show that IIR1-III ~ (1 - cl(M, N))1/2.)
c) Deduce from b) that if M + N° is closed, then so is MO + N.

8) a) Let E be a hilbertian space, and let T be a continuous linear mapping from E into itself
such that I TIl ~ 1. Show that the relations T.x = x, <T.xlx) = Ilxll, T*.x = x are equi-
valent, and that the kernel of IE - T and the closure of the image of IE - T are the ortho-
gonal complements.
b) Let T be a continuous linear mapping from E into itself, satisfying the inequality
(1)

for all x E E. Then liT. x I < I x I for all x such that T. x i= x; for every x E E, the sequence
(Tn .x) converges to a point P .x, and P is the orthogonal projector onto the kernel of 1E - T.
c) Let PI' ... , P, be orthoprojectors on E. Show that the product T = P I P 2 ... P, satisfies
relation (1) (argue by induction on r); hence the orthoprojector P is the orthogonal projector
onto the intersection of the images of the projectors Pj (note that if liPrxll < Ilxll for an indexj,
then I T .xll < Ilxll)·
~ 9) Let E be a hilbertian space and (Pj)jEN be a sequence of orthoprojectors an E, such that,
for all j E N, there exists an nj E N such that for all kEN, at least one of the orthoprojectors
Pk , Pk+ I ' ... , Pk+n is equal to Pj . Put Rs = P Ps- I ... Po for all sEN.
a) For every x E E, let Xs = R,.x. Show that 2= Ilxs - I - Xsll2 ~ Ilx11 2 , and deduce that for
s
every integer r ~ 1, x s +, - Xs tends to 0 as s tends to + 00.
b) Let (x,) be a sequence extracted from (x,) which tends to a limit Y weakly; then every
sequence (x sk + r ) also tends to y weakly. Deduce that y belongs to each of the subspaces
M j = Pj(E). (For each j, there exists r k such that 0 ~ r k ~ nj and Sk + r k = j; show that
the sequence (X Sk +,) tends to y weakly.)
c) Show that the sequence (x s ) converges weakly to the orthogonal projection from x onto
the intersection M of the M j . (Reduce to the case where M = { 0 }, and use b) and the weak
compactness of every closed ball in E.)
~ 10) Let E be a hilbertian space, u a positive endomorphism of E.
a) Show that for all x E E, we have

Ilu(x)i12 ~ Ilull·<u(x)lx)
(observe that <u(x)lu(x)2 ~ <u(x)lx) <u2(x)lu(x», by Y, p. 3, prop. 2).
b) Let M be a closed vector subspace of E, MO its orthogonal complement. Let x E M, and let
f(x) be the lower bound of <u(x + y)lx + y) as y ranges over MO. For every c > 0, let
E(x, c) be the set of all y E MO such that <u(x + y)lx + y) ~ f(x) + c. Show that E(x, E)
is convex and that for all z E M we have
O
,

(*) <u(x+Y)lz)l~c<u(z)lz) for YE E(x,c)

<
(consider the function g : t f-> u(x + y + tz)lx + y + tz) of the real variable t, which attains
its minimum at a point to and note that g(to) ~ f(x) and g(O) ~ f(x) + E).
c) For every integer n ~ 1, let Yn E E(x, lin); show that the sequence (u(x + Yn)) tends to
a limit XI belonging to M, and that the sequence «u(Yn)lyn») is bounded (find a bound for
the numbers <u(Yn - Ym)IY n - Ym) for m ~ nand 1< u(x + yn)lz)1 for z E M using ine- G

quality (*)).
d) Let (y~) be a sequence of points of MO such that <u(y~ - Y~)IY~ - y~) is arbitrarily small
when m and n are large enough, and such that the sequence (u(x + y~)) has a limit x~ EM;
TVS V.7,!) HILBERTIAN SPACES §4

show that X'I = Xl' (Let Q(z) = (u(z)lz); first show that the number Q((Yp-Y~)-(Yq-y'»)
is arbitrarily small as soon as P and q are large enough, and deduce tliat the sequen~e
(Q(Yn - y~») is bounded; using the fact that (x~ - xllYp - Y~) = 0 for all P, show that the
sequence (Q(Yn - y~» tends to 0 and use a).)
e) Deduce from d) that the point Xl does not depend on the choice of the Yn E E(x, lin),
and that if we put u l (x) = Xl' then u l is a linear mapping from M into itself. Show that
o :( (u l (x)lx) :( (u(x)lx) for all X E M and consequently that u l is continuous and is an
endomorphism of M which is?: O. (Observe that (u(x + Yn)IYn ) tends to 0 and (u(x + yn)lx + Yn )
tends to f(x).)
.f) Let PM be the orthoprojector with image M, and let U o = U I 0 PM' We have 0 :( U o :( u,
and M is stable under uo' and the restriction of U o to MO is null. Show that U o is the largest
element in the family of all endomorphisms v ?: 0 such that v :( u, that M is stable under
v and is such that the restriction of v to MO is null.

11) Let E and F be two hilbertian spaces. Show that for every element u in E @2 F, there
exists an orthonormal sequence (en) in E, an orthonormal sequence (J,,) in F and a sequence
(An) of numbers?: 0 such that I A~ < + 00 and that u = I Anen @ J,,; then Ilull~ = I A~
(cf Y, p. 55, tho 2 and p. 53, tho 1).

~ 12) Let E be a real hilbertian space and Y be a closed convex cone in E, with vertex 0,
let yo be the polar cone of Y (in E, identified canonically with its dual).
a) Show that every point X E E can be written uniquely in the form X = x+ - x_ where
x+ E Y and x_ E yo, and (x+lx_) = O.
b) For every facet F ofY (II, p. 87, exerc. 3), F is either the point 0 or is a convex cone with
vertex 0; the set of all Y E yo which are orthogonal to F is a closed facet F' of yo (but this is
not the « dual facet» of F in the sense of II, p. 87, exerc. 6, the latter being empty).
c) Take for E the set of all Hilbert-Schmidt endomorphisms of a real hilbertian space, and
for Y the set of positive elements in E. Show that we have yo = Y and in this case interpret
the result of a) (to see that Y c yo, use cor. 1 of Y, p. 56).
d) Under the hypothesis of c), let v E Y; the set L of all X E H such that (v(x)lx) = 0, or,
which is the same, such that vex) = 0 (Y, p. 77, exerc. 10) is a closed vector subspace of H,
and the facet F of v in Y is the closed set of all u E Y such that u(x) = 0 for all X E L; it can
be identified with the cone of all positive Hilbert-Schmidt endomorphisms of the hilbertian
space L". Deduce that the projection from E onto the convex set F (Y, p. 11) is identical with
the orthogonal projector from E onto the closed vector subspace of E generated by F.

~ 13) Let E be a hilbertian space, G a subgroup of the group of all automorphisms of the
hilbertian space structure of E. Let EG be the closed vector subspace of E consisting of all
vectors invariant under G, and P the orthoprojector from E onto E G .
a) Show that the orthogonal complement of EG in E is the closed vector subspace generated
by the vectors S.X - x, where s ranges over G and X E E.
b) Let H be a non-empty closed convex subset of E which is stable under G. Show that the
projection of 0 onto H belongs to EG.
c) Suppose that H is the closed convex envelope of the orbit of a point x of E, and let a be
the projection of 0 onto H. Show that a = p(x) and that H n EG reduces to the point a (<< Bir-

jt
khojf~Alaoglu tho »). (Observe that x - a is contained in the orthogonal complement of EG.)

d) Suppose G is generated by an automorphism u of E. Show that p(x) = !~~ n ~ 1 uj(x)

for all x E E (if Yn

n + j~O
f
uj(x), note that the sequence (Yn) has a weak limit point a,
= _1-1

and that u(a) = a, then use c»).

e) Suppose G is the image of a homomorphism t I--> u, from R onto the group of automor-
phisms of E, such that for all x E E, t I--> U,.X is a continuous mapping from R into E. Show

that p(x) = lim T

1 IT u,.x dt for all x E E.
T-4 OCJ 0
§4 EXERCISES TVSV.79

f) Suppose that there exists an element x =1= 0 of E and a number a such that 0 < a < 1
and Ils.x - xii,,;: allxll for all s E G. Show that EG =1= {O} (use c)).

14) Let E be a complex hilbertian space. and T a Hilbert-Schmidt endomorphism of E.

a) Let Rand L be positive Hilbert-Schmidt endomorphisms such that R 2 = T * T and
U = TT* (V, p. 57, cor. 3); let R = abs(T), and call this the « absolute value» of T (cf V,
p. 75. exerc. 4); we have L = abs(T*). Show that Ker(T) = Ker(R) and L(E) = T(E).
There exists one, and only one isometry V from R(E) onto T(E) such that T = VR; if we
extend V by continuity to R(E), then to an operator U E £-,(E) by taking U.x = 0 on the
orthogonal complement of R(E), we also have T = UR (polar decomposition of T). Then
R = U*T = U*UR = RU*U and L = URU*, T = LU*. If T belongs to £-,I(E), then
so does R = abs(T), and T is the product of two Hilbert-Schmidt endomorphisms.
b) If T belongs to £-,I(E), show that Tr(abs(T)) = sup(I l<adT.bi)l) where, on the right
hand side, (aJ and (bJ range over the set of orthonormal bases of E (use the polar decompo-
sition of T). Show that if we put I Till = Tr(abs(T)), then II Till is a norm on the space £-' I (E),
such that IITI12 ,,;: IITilI'
e) Conversely, if T E £-,(E) is such that, for every pair (aJ, (bJ) of orthonormal bases of E,
the sum I l<aiIT.bi>1 is finite, then T E £-,I(E) (first observe that T is a Hilbert-Schmidt
i
endomorphism, then use the polar decomposition of T).
d) Let (TJ be a sequence of Hilbert-Schmidt endomorphisms (resp. of elements of £-,I(E))
such that for every pair of points x, y of E, the sequence «xITv'Y») converges to <xIT.y),
where T is a linear mapping from E into itself; in addition, assume that the sequence of norms
IITvl12 (resp. IITvlll) is bounded. Show that T is a Hilbert-Schmidt endomorphism (resp. an
element of £-,I(E)) (use b) and c)).
e) Deduce from d) that the space £-,I(E) is a Banach space for the norm IITk
f) For an endomorphism T E £-,(E) to belong to £-,1(E) it is necessary and sufficient that,
for at least one orthonormal basis (eJ of E, the sum I II T .eill is finite (with the notations
i

x 1
g) In the space f~, let (en) be the canonical orthonormal basis, and let a = n~o n + 1 en;

if F is the I-dimensional subspace C.a, the orthoprojector PF has finite trace, but the series
I IIPF·enll is not convergent.

15) Let E be a complex hilbertian space; let !18 = £-'(E) denote the algebra of continuous
endomorphisms of E, endowed with the usual norm I Til = sup I T .xll. For every pair of
IlxiiO
points x, y in E, let O)x,y denote the continuous linear form T f-4 < xl T.y) on !18, and let .OJc be
the closed subspa~e of the strong dual !18' of the Banach space !18, ge~erated by the O)x,y'
a) Show that the llllear mapplllg whIch assocIates to every T E !18 the llllear form 0) f-4 < 0), T)
on !18o ' is an isometry from .OJ onto the strong dual of !18o ; in other words, !18o is a predual (IV,
p. 56, exerc. 23) of (!jJ.
The topology cr(!18, !18o ) on (!jJ is called the ultraweak topology.
b) For every element T of £-,I(E), we define a linear form <PT on (!jJ by the formula
<PT(S) = Tr(ST) for every operator S E !18. Show that <PT is continuous and that the mapping
T f-4 <PT is an isometry from the Banach space £-,1(E) (exerc. 14, e)) onto the Banach space!18o
(first consider the case when T has finite rank).
e) Let (!jJoo be the vector subspace of !18o generated by the O)x,y (in such a way that!18o = ~o).
Show that (!jJoo is barrelled (note that a subset of!18 which is bounded for cr(!18, !18oo ) is bounded
for the norm topology).
d) Let Fn be the subspace of !18co which is the image of the set of all endomorphisms of E
of rank ,,;: n under the isometry defined in b). Show that Fn is nowhere dense in !18co and deduce
that !18oo is not a Baire space.
 Historical notes

(chapters I to V)

(N.B. - The roman letters refer to the bibliography at the end of this note.)
The general theory of topological vector spaces was founded in the period around
the years 1920 to 1930. But the ground work had been under preparation since long
by the study of numerous problems of functional Analysis; we cannot retrace the
history of the subject without indicating, at least briefly, how the study of these pro-
blems slowly (particularly since the beginning of the 20th century) led the mathe-
maticians to an awareness of the relationship between the questions being considered
and the possibility offormulating them in a much more general manner, and applying
to them uniform methods of solutions.
It can be said that the analogies between Algebra and Analysis, and the idea of
considering functional equations (i.e. where the unknown is a function) as « limiting
cases» of algebraic equations have their origins in the infinitesimal Calculus, which
in some sense was invented to generalize «from the finite to the infinite ». But the
direct algebraic ancestor of the infinitesimal Calculus is the Calculus of finite diffe-
rences (cf FVR, Historical note of chapters I, II, III, p. 54-58) and not the solution
of general linear systems; it was only after the middle of the 18th century that the
first analogies between the latter and the problems of differential Calculus made their
appearance in the study of the equations of vibrating strings. We shall not enter
into the details of the history of this problem here; but the constant reappearance
of two fundamental ideas stands out, both of which are apparently due to D. Ber-
noulli. The first consists in considering the oscillation of the string as a « limiting
case» of the oscillation of a system of n point masses as n increases indefinitely;
we know that, later, this problem, for n finite, was the first example in the search
for the eigenvalues of a linear transformation (cf A, Historical Notes of chapters VI-
VII); these numbers correspond in the limiting case, to the frequencies of the « eigen
oscillations» of the string, which where observed experimentally long before, and
whose theoretical existence had been established (notably by Taylor) at the beginning
of the century. This formal analogy, although hardly ever mentioned later ((I, b),
p. 390) never seems to have been lost of sight of during the 19th century; but as we
shall see, it acquired its full importance only around the years 1890-1900.
The other idea of D. Bernoulli (perhaps inspired by experimental facts) is the
« superposition principle» according to which the most general oscillation of the
string should be « decomposable» by superposition of the « eigen oscillations»;
mathematically speaking, this means that the general solution of the equation of
vibrating strings should have a series development as I c,,<p,,(x, t), where the <p,,(x, t)
"
 HISTORICAL NOTES TVS V.81

represent the eigen oscillations. We know that this principle was the starting point of a
long battle on the possibility of developing an « arbitrary» function in a trigonome-
tric series, a battle which was settled by the works of Fourier and of Dirichlet only in
the first third of the 19th century. But even before this result was obtained, there
were other examples of series development in « orthogonal» functions * : spherical
functions, Legendre polynomials, and also various systems of the form (e iAnX ), where
the 'A" are no longer multiples by the same number; these had already been intro-
duced in the 18th century in oscillation problems, as also by Fourier and Poisson
in the course of their researches on the theory of heat. Around 1830, Sturm (I) and
Liouville (II) systematized all the phenomena observed in these various particular
cases into a general theory of oscillations for functions of one variable; they con-
sidered the differential equation

(1) ix (p(X) Ix) + 'Ap(x) y = 0 (p(x) > 0, p(x) > 0)

with the boundary conditions

(2)

and proved the following fundamental results

1) the problem has a non-zero solution only if 'A takes one of the values of a
sequence ('An) of numbers > 0, tending to + 00 ;

f
2) for each 'A", the solutions are multiples of the same function V n ' which may

be assumed « normalized» by the condition pv;dx = 1, and for m #- n we have

f pvmvndx = 0;

3) every twice differentiable function f on [a, b] which satisfies the boundary

conditions (2), can be developed in a uniformly convergent series as f(x) = L: cnv,.(x),

f
n

where Cn = pfvndx;

4) the equality f pPdx = L: c; holds (this equality had already been proved

by Parseval in 1799, though in a purely formal manner, for the system of trigono-
metric functions; and from it « Bessel's inequality» follows immediately; the latter
inequality was announced by Bessel (again for trigonometric series) in 1828).
Half a century later, these properties were completed by the work of Gram (III)
who, following the researches of Tchebichef, threw light on the relationship between
* This term however does not appear before the work of Hilbert.
TVS V.82 TOPOLOGICAL VECTOR SPACES

the development in series of orthogonal functions and the problem of« best quadratic
approximation» (a direct outcome of the «method of least squares» of Gauss
in the theory of errors); the latter consists of the following: given a finite sequence
of functions (\)I)lo(io(n' and a functionj, to find the linear combination L
Gi\)li for
i

which the integral fb p(f - L Gi\)lJ 2dx attains its minimum. In principle, this only
a '
suggests a trivial linear algebraic problem, but Gram solved it in an original way,
by applying the method of« orthonormalization » to the \)Ii' as described in chap. V,
p. 23 (and generally known under the name of Erhard Schmidt). Next, in the case
of an infinite orthonormal system, the question arises of finding out when the « best
quadratic approximation» Il n of a function j, by linear combinations of the first
n functions of the sequence, tends to 0 as n increases indefinitely * ; Gram was thus
led to the definition of the notion of a complete orthonormal system, and recognized
that this property is equivalent to the non-existence of non-zero functions which are
orthogonal to all the <Pn' He even attempted to elucidate the concept of « mean
quadratic convergence », but before the introduction of the fundamental ideas
of measure theory, he could hardly obtain any general results in this direction.
In the second half of the 19th century, the major effort of analysts was mainly
directed towards the extension of the Sturm-Liouville theory to functions of several
variables. This theory was prompted by the study of elliptic partial differential
equations arising in Mathematical Physics, and the boundary value problems
which are naturally associated with these. The main interest primarily centered
on the equation of «vibrating membranes»
(3)

where solutions vanishing on the boundary of a sufficiently regular domain G were

sought; the methods which had worked successfully for functions of one variable
were no longer appropriate for this problem, and the considerable analytic diffi-
culties that presented themselves were overcome little by little. We recall the main
steps towards the solution: the introduction of the « Green's function» of G, whose
existence was proved by Schwarz; the proof, again due to Schwarz, of the existence
of the smallest eigenvalue; and finally, in 1894, H. Poincare, in a celebrated memoir
(V G) succeeded in proving the existence and the essential properties of all the eigen-
values. He considered the solution of the equation L).(u) = 1, for a« second member»
j given; the solution being such as to vanish on the boundary; then by a skillful
generalization of Schwarz's method, he proved that UJ,. is a merom orphic function
of the complex variable A, having only real simple poles An' and these are precisely
the eigenvalues being sought.

f
* It must be pointed out that in this study, Gram did not restrict himself to considering only

continuous functions, but emphasised the importance of the condition pf 2 dx < + 00.
 HISTORICAL NOTES TVS V.83

These researches are directly related to the beginnings of the theory of linear
integral equations, which must have certainly contributed the maximum to the advent
of modern ideas. We shall here limit ourselves to giving a brief outline of the deve-
lopment of this theory (for fuller details, we refer to the Historical Notes which will
follow the chapters of this Treatise dedicated to spectral theory). This kind offunc-
tional equations, which first made a modest appearance in the first half of the 19th
century (Abel, Liouville), had already acquired some importance since Beer and
C. Neumann reduced the solution of« Dirichlet's problem» for a sufficiently regular
domain G to the solution of an « integral equation of second kind»

(4) u(x) + f K(x, y) u(y) dy = f(x)

for the unknown function u; C. Neumann succeeded in solving this equation in

1877 by a method of «successive approximations ». Prompted as much by the
algebraic analogies mentioned above as by the results he had obtained for the equa-
tion of vibrating membranes, H. Poincare, in 1896 (V b) introduced a variable para-
meter A in front of the integral in the preceding equation, and asserted that, just
as in the case of the equation of vibrating membranes, the solution is once again
a meromorphic function of A; but he was unable to prove this result. This was
established seven years later by I. Fredholm (VI) (for a continuous « kernel» K and
a finite interval ( a, b)). The last mentioned author, perhaps with a greater awareness
than his predecessors, let himself be guided by the analogy of( 4) with the linear system

(5)

to obtain the solution of (4) as the quotient of two expressions, based on the model
of determinants, which arise in Cramer's formulas. This, however was not a new
idea : since the beginning of the 19th century, the method of « indeterminate coeffi-
cients» (which consists of obtaining an unknown function, assumed to have a
series development I en<l>n' where the <l>n are known functions, by calculating the
n
coefficients en) had led to « linear systems with infinitely many unknowns»
00

(6) L aijxj = bi (i = 1,2, ... ) .

j= 1

Fourier, who encountered such a system, still solved it like an 18th century mathe-
matician : he suppressed all the terms with indices i or j greater than n, explicitly
solved the finite system so obtained by Cramer's formulas, then passed to the limit by
letting n tend to + 00 in the solution! Much later, when this jugglery was no longer
acceptable, it was again by the theory of determinants that the problem was attacked;
Since 1886 (following the work of Hill), H. Poincare, then H. von Koch, had set up
a theory of « infinite determinants », which permits the resolution of certain kinds
of systems (6) by following the classical model; and in spite of the fact that these
TVS V.84 TOPOLOGICAL VECTOR SPACES

results were not directly applicable to the problem tackled by Fredholm, it is beyond
doubt that von Koch's theory in particular, served as a model for the construction
of Fredholm's « determinants ».
It was at this moment that Hilbert entered the scene and gave a new impetus to
the theory (VII). To begin with, he completed the 'work of Fredholm by effectively
carrying out the passage to the limit, which leads to the solution of (4) from that of
(5); but he immediately brought in the corresponding passage to the limit in the
theory of real quadratic forms, which arose automatically from integral equations
with a symmetric kernel (i.e. such that K(y, x) = K(x, y)). These are the equations
by far the most frequent in Mathematical Physics. He thus succeeded in obtaining
the fundamental formula which directly generalizes the reduction of a quadratic
form to its axes

(7) f bfb K(s, t) xes) x(t) dsdt = ~ ;;:

CJ) 1 (fb <Pis) xes) ds )2 ,
a a n 1 n a

r
where the An are the eigenvalues (necessarily real) of the kernel K, the <Pn forming
the orthonormal system of the corresponding eigen functions, and the second

r
member of formula (7) is a convergent series if X2(S) ds :::;; 1. He also showed

r
how every function which is «representable» as f(x) = K(x, y) g(y) dy has a

« development» n~l <p/x) <Pn(Y) fey) dy, and, following the analogy with the

classical theory of quadratic forms, he indicated a procedure for determining the

An by a variational method. This is precisely the extension of the well-known extremal
properties of the principal axes of a quadric surface ((VII), p. 1-38).
These preliminary results of Hilbert were almost immediately taken up by
E. Schmidt, under a simpler and more general form, avoiding the introduction of
« Fredholm's determinants» and also the passage from the finite to the infinite.
The presentation was already very close to being abstract, the fundamental pro-
perties of linearity and of positivity of the integral being clearly the only facts used
in the proof (VII a). But by then Hilbert had developed much more general con-
cepts. All the earlier works brought out the importance of square integrable func-
tions, and Parseval's formula established a direct link between these functions and
sequences (c n ) such that I c;< 00. It is certainly this idea which guided Hilbert
n
in his 1906 memoirs ((VII), chap. XI-XIII), where, taking up the old method of
« indeterminate coefficients» once again, he showed that the solution of the integral
equation (4) is equivalent to the solution of an infinite system of linear equations
CJ)

(8) xp + I kpqXq = bp (p = 1, 2, ... )

q=l
 HISTORICAL NOTES TVS V.85

for the « Fourier coefficients)} xp = f u(t) rop(t) dt of the unknown function u

with respect to a given complete orthonormal system (ron) (with bp= f f(t) rolt) dt

and kpq = ff K(s, t) rop(s) roit) dsdt). Moreover, from this point of view, the

only solutions of (8) of interest are those for which I

n
x; < + 00; also it was to
this kind of solution that Hilbert systematically restricted himself; but on the other
hand, he extended the conditions imposed on the « infinite matrix)} kpq (which in (8)
is such that I k;q < + 00). Thenceforth, it was clear that the « Elbert space)}
p,q
of all sequences x = (xn ) of real numbers such that I x; < + 00, while not expli-
n
citly introduced was the space underlying the entire theory, and appears as a « pas-
sage to the limit)} from a finite dimensional Euclidean space. In addition, and this
was particularly important for later developments, Hilbert was led to introduce,
not just one, but two distinct notions of convergence in this space (corresponding
to what has since been called the weak topology and the strong topology *), as also
a « principle of choice)} which is precisely the property of weak compactness of
the unit ball. The new linear algebra that he developed in connection with the solu-
tion of the system (8) depended entirely on these topological ideas: linear mappings,
linear forms and bilinear forms (associated with linear mappings) were classified
and studied with respect to their « continuity)} properties **. In particular, Hilbert
discovered that the success of Fredholm's method depended on the notion of « com-
plete continuity)}, which he redeemed by formulating it for bilinear forms *** and
by studying it profoundly; for more details we refer to the part of this Treatise
where this important notion shall be developed, and also to the admirable and pro-
found works of Hilbert, where he inaugurated the spectral theory of symmetric
bilinear forms (bounded or not).
The language of Hilbert still remained classical, and throughout the « Grundzuge )},
he never lost sight of the applica tions of the theory which he developed from numerous
examples (taking up almost half of the volume). The next generation already adopted
a much more abstract point of view. Under the influence of the ideas of Frechet
and ofF. Riesz on general topology (see Historical Notes of GT, chap. I), E. Schmidt

* The Calculus of variations had naturally led to different notions of convergence on the
same set of functions (according to the requirement of uniform convergence of functions,
or of uniform convergence of functions and of a certain number of their derivatives); but
the modes of convergence defined by Hilbert were entirely new at that time.
** It must be pointed out that until around 1935, by a« continuous» function it was gene-
rally meant that this was a mapping wich transformed every convergent sequence into a
convergent sequence.
*** For Hilbert, a bilinear form B(x, y) was completely continuous if, whenever the
sequences (x n), (Yn) tended weakly to x and y respectively, B(x n, Yn) tended to B(x, y).
TVS V.86 TOPOLOGICAL VECTOR SPACES

(VII b) and Frechet himself, in 1907-1908 deliberately introduced the language of

Euclidean geometry into the « Hilbert space» (real or complex); it is in these works
that we find the first mention of the norm (with its present notation Ilxll), the triangle
inequality that it satisfies, and the fact that a Hilbert space is « separable» and com-
plete; in addition, E. Schmidt proved the existence of the orthogonal projection
onto a closed linear variety, which allowed him to give a simpler and more general
form to Hilbert's theory oflinear systems. Also in 1907, Frechet and F. Riesz observed
that the space of square integrable functions has an analogous « geometry», an
analogy which was perfectly explained when, a few months later, F. Riesz and
E. Fischer proved that this space is complete and is isomorphic to a « Hilbert space »,
and at the same time displayed in a striking manner the value of the tool newly
created by Lebe~gue. From this moment onwards, the essential points of the theory
of hilbertian spaces could be considered to have been achieved. Among the later
developments the axiomatic presentation of the theory by M. H. Stone and 1. von
Neumann around 1930 must be mentioned, and also the removal of the restrictions
of « separability» which was the result of the work of Rellich, L6wig and F. Riesz
(IX e) around the year 1934.
Meanwhile, in the first few years of the 20th century, other streams of ideas came
and reinforced the trend which led to the theory of normed spaces. The general
idea of « functional» (i.e. a numerical function defined on a set whose elements are
themselves numerical functions of one or of several real variables) was redeemed in
the last decades of the 19th century in connection with the calculus of variations
on the one hand, and on the other, with the theory of integral equations. But it was
primarily from the Italian school, around Pincherle, and above all Volterra, that
the general idea of « operator» arose. The works of this school often stayed at a
rather formal level and were related to particular problems, for lack of a sufficiently
deep analysis of the underlying topological concepts. In 1903, Hadamard inau-
gurated the modern theory of « topological» duality, in his search for the most
general continuous linear « functionals » on the space ~(I) of continuous numerical

i
functions on a compact interval (endowed with the topology of uniform conver-
gence), and he characterized these as limits of sequences of integrals x f--* k n ( t) x( t) dt.

In 1907, Frechet and F. Riesz proved similarly that the continuous linear forms on a
Hilbert space are the « bounded» linear forms introduced by Hilbert; then in 1909,
F. Riesz put Hadamard's theorem in a definitive form by expressing every conti-
nuous linear form on ~(I) as a Stieltjes integral, a theorem which much later served
as the starting point for the modern theory of integration (see Historical Notes of
INT, chap. II-V).
The following year, F. Riesz (IX a) again made new and important progress in
the theory by introducing and studying (modelled on the theory of the Hilbert
space) the space U(I) of functions on an interval I whose p-th power is integrable
(for an exponent p such that 1 < p < +:D); three years later, this study was
followed by analogous work on the sequence spaces fP(N) (IX c). These researches,
 HISTORICAL NOTES TVS V.87

as we shall see later, made a great contribution towards the classification of ideas
on duality, in the sense that for the first time we encountered two spaces in duality
which were not naturally isomorphic *.
Then onwards, F. Riesz thought of an axiomatic study which would encompass
all these results ((IX a), p. 452), and it seems that only the scruples of an analyst
anxious not to deviate from classical mathematics restrained him from writing
his celebrated memoire of 1918 on Fredholm's theory (IX d) in this form. There he
mainly considered the space 'ti(I) of continuous functions on a compact interval;
but after defining the norm of this space, and having remarked that 'ti(l) endowed
with this norm is complete, he did not use anything other than the axioms of complete
normed spaces in his arguments **. Without entering into a detailed examination
of this work, we mention that the notion of a completely continuous mapping was
defined (by the property of transforming a neighbourhood into a relatively compact
set) in a general way for the first time in this work *** ; by a masterpiece of axiomatic
analysis, the entire theory of Fredholm (with respect to its qualitative aspect) was
reduced to a single fundamental theorem, that every locally compact normed space
is finite dimensional.
The general definition of normed spaces was given in 1920-1922 by S. Banach,
H. Hahn and E. Helly (the latter considered only sequence spaces of real or complex
numbers). In the ten years that followed, the theory of these spaces developed mainly
around two questions of fundamental importance for applications : the theory of
duality and the theorems linked with the notion of Baire « category ».
We have seen that the idea of duality (in the topological sense) originated in the
beginning of the 20th century; it was the underlying notion in Hilbert's theory and
occupied a central place in the work of F. Riesz. The latter, for example, observed
in 1911 ((IX b), p. 41-42), that the relation If(x) I :( M Ilxll (taken as the definition
of « bounded» linear functionals in a Hilbert space) is equivalent to the continuity
of f in the case of the space 'ti(I), and this was proved by fairly general arguments.
Concerning the characterization of continuous linear functionals on 'ti(I), he further
observed that the condition for a set A to be total in 'ti(l) is that there exist no Stieltjes
measure J.t =1= 0 on I which is « orthogonal» to all the functions in A (thus genera-
lizing Gram's condition for complete orthonormal systems); finally, in the same

* In spite of the fact that the duality between L 1 and L was implicit in most of the works
00

i
of this epoch on the Lebesgue integral, it was only in 1918 that H. Steinhaus proved that every
continuous linear form on L 1 (I) (I a finite interval) is of the form x f--> J(t) x(t) dt, where

J E L 00(1).
** F. Riesz however, explicitly noted that the applications of his theorems to continuous
functions is only a « touchstone» of much more general concepts (IX d), p. 71).
*** In his work on LP spaces, F. Riesz had defined completely continuous mappings as
those which transform every weakly convergent sequence into a strongly convergent sequence;
this (on account of the weak compactness of the unit ball in the U for 1 < P < + CXJ) is
equivalent to the above definition in this case; in addition, F. Riesz indicated that for the L 2
spaces, his definition was equivalent to that of Hilbert (by translating from the language
of linear mappings to that of bilinear forms «IX, a), p. 487)).
TVS V.88 TOPOLOGICAL VECTOR SPACES

work, he established that the dual of the space L 00 is « bigger» than the space of
Stieltjes measures ((IX b), p. 62).
On the other hand, F. Riesz, in his work on the spaces U(I) and fP(N) succeeded
in modifying the method of the solution of linear systems in a Hilbert space, as
given by E. Schmidt (VIII b) so as to be applicable in more general cases. E. Schmidt's
idea consisted in determining an « extremal» solution of (6) by finding a point in
the closed linear variety represented by the equations (6), whose distance from the
origin is the minimum. Using the same idea, F. Riesz showed that a necessary and
sufficient condition such that there exists a function x E U(a, b) satisfying the equa-
tions

(9) f (Xi(t) x(t) dt = bi (i = 1,2, ... )

(where the (Xi belong to U (with ~+ i = l)), and such that in addition

f Ix(tWdt ::::; MP, is that, for every finite sequence ("-)1 ";i"'n of real numbers, we

have

(10)

In 1911 (IX b), he treated, in an analogous manner, the « problem of generalized

moments », which consists of the solution of the system

(11) f (X;(t) d~(t) = bi (i = 1,2, ...)

where the (Xi are continuous and the unknown is a Stieltjes measure ~ * ; it was clear
in this case that the problem can be restated by saying that it consists in determining
a continuous linear functional on ~(I) from its values on a given sequence of points
in this space. It was in this form that Helly treated the problem in 1912 - obtaining
F. Riesz's conditions by a rather different method of much wider scope * - and
which he again took up in 1921, with much more general conditions. Introducing
the notion of a norm (on the sequence spaces), as we have seen above, he observed

* The classical « problem of moments» corresponds to the case where the interval )a, b(
is )0, + oo( or )- 00, + 00(, and where Cli(t) = ti; moreover, one assumes that the mea-
sure ~ is positive (in his 1911 memoire, F. Riesz indicated how his general conditions must be
modified when solutions of this nature are sought). Among the various methods for the solu-
tion of the classical problem of moments, we particularly mention that of F. Riesz, who very
elegantly combined the general ideas of functional Calculus and the theory of functions of
one complex variable to obtain explicit conditions on the bi. (Sur Ie probU:me des moments,
3, Ark. for Math., t. XVII (1922-1923), no 16,52 p.)
 HISTORICAL NOTES TVS V.89

that this notion generalizes that of the « gauge» of a convex body in an n-dimensional
space, as used by Minkowski in his celebrated work on the « geometry of numbers»
(IV). In the course of his researches, Minkowski also defined (in Rn) the notions of
a support hyperplane and of a « supj)ort function» (IV b), and proved the existence
of a support hyperplane at every point of the boundary of a convex body ((IVa),
p. 33-35). Helly extended these notions to a space of sequences E, endowed with an
arbitrary norm; he established a duality between E and the space E' of sequences
u = (un) such that for all x = (x n) E E, the series (unx n) is convergent; letting <u, x>
denote the sum of this series, he defined a norm in E' by the formula sup I<u, x>l/llxll,
x*D
which gives the support function in finite dimensional spaces **. Then Helly proved
that the solution of a system (6) in E, where each sequence ui = (ai)j?: 1 is assumed
to belong to E', reduces to the successive resolution of the following two problems:
1. to find a continuous linear form L on the normed space E', such that L(u)=bi
for every index i; this, as he pointed out, leads to conditions of the type (IO); 2. to
find if such a linear form can be written as u ~ <u, x> for some x E E. The latter
problem, he observed, does not necessarily have a solution even if L exists, and
he gave some sufficient conditions which imply the existence of the solution x E E
in some particular cases (X).
In 1927, these ideas were given their definitive form in a fundamental memoire of
H. Hahn (XI), whose results were rediscovered (independently) by S. Banach two
years later (XII b). Hahn applied the methods of Minkowski-Helly to an arbitrary
normed space, and thus defined the structure of a (complete) normed space on the
dual space; this immediately allowed Hahn to consider successive duals of a normed
space, and to pose the problem of reflexive spaces in a general way, as already foreseen
by Helly. But above all, the principal problem of the extension of a continuous
linear functional without increasing its norm was definitely solved by Hahn in
general, by an argument of transfinite induction on the dimension - thus giving
one of the first examples of an important application of the axiom of choice to
functional Analysis ***. To these results, Banach added a detailed study of the
relations between a continuous linear mapping and its transpose, extending to
general normed spaces results previously known in the case ofLP spaces only (IX a),
by means of a deep theorem on weakly closed subsets of the dual (cf IV, p. 25, cor. 2);
these results can be expressed in a more striking way using the notion of the quotient
space of a normed space, which was introduced a few years later by Hausdorff
and by Banach himself Finally, it was once again Banach who discovered the rela-
tion between the weak compactness of the unit ball (observed in several particular

* Like F. Riesz (IX b), p. 49-50), Helly used a « principle of choice» in his proof, which is
precisely the weak compactness of the unit ball in the space of Stieltjes measures; F. Riesz
had also used the analogous property in the U spaces (1 < p < + co).
** To obtain a norm in this way, we must assume that the relation <u, x> = for all
x E E implies u = 0, as is explicitly remarked by Helly.
°
*** Banach had already given an analogous argument in 1923 for defining an invariant
measure in the plane (defined for every bounded subset) (XII a).
TVS V.90 TOPOLOGICAL VECTOR SPACES

cases, as mentioned above) and reflexivity, at least for the spaces satisfying the first
axiom of countability. Since then, the broad outlines of the theory of duality of
normed spaces could be considered to have been fixed.
During the same epoch some seemingly paradoxical theorems, whose first exam-
ples originated in the years around 1910, were clarified. In that year, Hellinger and
Toeplitz had essentially proved that a sequence of bounded bilinear forms Bn(x, y)
on a Hilbert space, whose values Bn(a, b) for every pair (a, b) are bounded (by a
number depending a priori on a and b) is in fact uniformly bounded on every ball.
Their proof was based on an argument of reductio ad absurdum, by inductively
constructing a particular pair (a, b) violating the hypothesis; this method is since
then known under the name « gliding bump », and is still useful in many analogous
questions (cf IV, p. 54, exerc. 15). In 1905, Lebesgue had used a similar method to
prove the existence of continuous functions whose Fourier series diverges at some
points; and in the same year as Hellinger and Toeplitz he used the method again,
to prove that a weakly convergent sequence in L 1 is bounded in norm *. These
examples multiplied in the following years, but without the introduction of any
new ideas until 1927, when Banach and Steinhaus (with the partial collaboration
of Saks) related these phenomena to the notion of a thin set and to Baire's theorem
in complete metric spaces, thus obtaining a general assertion which encompassed
all the previous particular cases (XIII). During the same epoch, the study of questions
of « category» in complete normed spaces led Banach to several other results on
continuous linear mappings; the most remarkable and certainly the deepest being
the « closed graph» theorem which, like the Banach-Steinhaus theorem, has become
a vital tool in modern functional Analysis (XII b).
The publication of Banach's treatise on « Linear Operators» (XII c) marks the
coming of age for the theory of normed spaces. All the above mentioned results
and many others can be found in this volume, though in a somewhat disorganized
manner, but with many striking examples drawn from various domains of Analysis,
and which seemed to forecast a brilliant future for the theory. The work had consi-
derable success, and one of the immediate effects was the almost universal adoption
of the language and notations used by Banach. But in spite of the great number of
researches undertaken during the past 40 years on Banach spaces (XVII), if the
theory of Banach algebras and its applications to commutative and non-commu-
tative harmonic analysis are excluded, then the almost complete absence of new
applications of the theory to the great problems of classical Analysis somewhat
undermines the hopes based on it.
It was more in the sense of widening and of a deeper axiomatic analysis related
to the concepts of normed spaces that the most fruitful developments took place.

* Observe also the analogous (easier) theorem proved by Landau in 1907 and which served
as a starting point for F. Riesz in his theory of the LV spaces : if the series with the general
term unxn converges for every sequence (xn) E fP(N), then the sequence (un) belongs to fq(N)
(with.!.
p
+ .!.q = 1).
 HISTORICAL NOTES TVS V.91

In spite of the fact that the functional spaces encountered since the beginning of the
20th century generally appear to be endowed with a « natural» norm, there are
certain exceptions. Around 1910, E. R. Moore proposed a generalization of the

°
notion of uniform convergence by replacing it with a notion of « relative uniform
convergence », where a neighbourhood of consists of functions J satisfying a rela-
tion IJ(t)1 ~ Eg(t), g being a function which is everywhere > ° and which could
vary with the neighbourhood. On the other hand, before 1930, it was observed that
notions such as simple convergence, convergence in measure for measurable func-
tions, or compact convergence for entire functions, could not be defined by means
of a norm; and in ) 926, Frechet observed that vector spaces of this kind could be
metrizable and complete. But the theory of these more general spaces could be
fruitfully developed only in relation with the idea of convexity. The latter (which
already appeared in ReIly's work) was the subject of the studies carried out by
Banach and his students, who recognized the possibility of interpreting several
results of the theory of normed spac-es geometrically, thus preparing the road for
a general definition of locally convex spaces, given by Kolmogoroff and J. von
Neumann in 1935. The theory of these spaces and notably the questions related to
duality, were mostly developed in the years 1950, and in this Book we have presented
the essential results of this study. In this connection we must point out, on the one
hand, the progress in simplicity and in generality, made possible by the focus on
the fundamental concepts of general Topology developed between 1930 and 1940;
and secondly, the importance of the notion of a bounded set, introduced by Kol-
mogoroff and von Neumann in 1935, and whose fundamental role in the theory
of duality was brought to light by the work of Mackey (XIV) and of Grothendieck
(XVIII). Finally, it is certain that the main impetus which motivated these researches
came from the new possibilities of applications to Analysis in domains where Banacr:
theory did not work : in this connection, we mention the theory of sequence spaces
developed since 1934, by Kothe, Toeplitz and their students in a series of memoirs
(XV), the focus on the theory of « analytic functionals» of Fantappie, and above
all, the theory of distributions by L. Schwartz (XVI), where the modern theory of
locally convex spaces found a field of applications, which is certainly far from being
exhausted.
 Bibliography

(I) C. STURM: a) Sur les equations differentielles lineaires du second ordre, Journ. de Math. (I),
t. I (1836), p. 106-186; b) Sur une classe d'operations a differences partielles, ibid.,
p. 373-444.
(II) J. LIOUVILLE: a) Sur Ie developpement des fonctions ou parties de fonctions en series
dont les divers termes sont assujettis a satisfaire a une meme equation differentielle du
second ordre contenant un parametre variable, Journ. de Math. (1), t. I (1836), p. 253-265,
t. II (1837), p. 16-35 et 418-436; b) D'un theoreme dfi a M. Sturm et relatif a une classe
de fonctions transcendantes, ibid., t. 1(1836), p. 269-277.
(III) J. P. GRAM. Ueber die Entwickelung ree1ler Functionen in Reihen mittelst der Methode
der kleinsten Quadrate, J. de Crelle, t. XCIV (1883), p. 41-73.
(IV) H. MINKOWSKI : a) Geometrie der Zahlen, 1re ed., Leipzig (Teubner), 1896; b) Theorie
der konvexen Korper, Gesammelte Abhandlungen, t. II, p. 131-229, Leipzig-Berlin (Teub-
ner), 1911. (Reimpression, New York (Chelsea), 1967.)
(V) H. POINCARE: a) Sur les equations de la Physique mathematique, Rend. Palermo,
t. VIII (1894), p. 57-156 (= (Euvres, t. IX, p. 123-196, Paris (Gauthier-Villars), 1954);
b) La methode de Neumann et Ie probleme de Dirichlet, Acta Mathematica, t. XX
(1896), p. 59-142 (= (Euvres, t. IX, p. 202-272, Paris (Gauthier-Villars), 1954).
(VI) I. FREDHOLM, Sur une classe d'equations fonctionnelles, Acta Mathematica, t. XXVII
(1903), p. 365-390.
(VII) D. HILBERT, Grundzuge einer allgemeinen Theorie der linearen Integralgleichungen,
New York (Chelsea), 1953 (= Gott. Nachr., J904, 1905, 1906, 1910).
(VIII) E. SCHMIDT: a) Zur Theorie der linearen und nichtlinearen Integralgleichungen.
I. Teil : Entwickelung willkiirlicher Funktionen nach Systemen vorgeschriebener,
Math. Ann., t. LXIII (1907), p. 433-476; b) Ueber die Auflosung linearer Gleichungen
mit unendlich vielen Unbekannten, Rend. Palermo, t. XXV (1908), p. 53-77.
(IX) F. RIESZ : a) Untersuchungen iiber Systeme integrierbarer Funktionen, Math. Ann.,
t. LXIX (1910), p. 449-497; b) Sur certains systemes singuliers d'equations integrales,
Ann. Ec. Norm. Sup. (3), t. XXVIII (1911), p. 33-62; c) Les systemes d'equations lineaires
dune infinite d'inconnues, Paris (Gauthier-Villars), 1913; d) Ueber lineare Funktional-
gleichungen, Acta Mathematica, t. XLI (1918), p. 71-98; e) Zur Theorie des Hilbertschen
Raumes, Acta litt, ac scient, (Szeged), t. VII (1934-35), p, 34-38.
(X) E. HELLY, Ueber Systeme linearer Gleichungen mit unendlich vielen Unbekannten,
Monatsheftefor Math. und Phys., t. XXXI (1921), p. 60-91.
(XI) H. HAHN, Ueber lineare Gleichungssysteme in linearen Riiumen, J. de Crelle, t. CL VII
(1927), p. 214-229.
(XII) S. BANACH: a) Sur Ie probleme de la mesure, Fund. Math., t. IV (1923), p. 7-33;
b) Sur les fonctionnelles lineaires, Studia Math., t. I (1929), p. 211-216 et 223-239;
c) Theorie des operations linea ires, Warszawa, 1932. (Reimpression, New York (Chel-
sea), 1963.)
(XIII) S. BANACH et H. STEINHAUS, Sur Ie principe de condensation des singularites, Fund.
Math., t. IX (1927), p. 50-61.
(XIV) G. W. MACKEY: a) On infinite-dimensional linear spaces, Trans. Amer. Math. Soc.,
t. LVII (1945), p. 155-207; b) On convex topological spaces, Trans. Amer. Math. Soc.,
t. LX (1946), p. 519-537.
(XV) G. KOTHE, Neubegriindung der Theorie der vollkommenen Raume, Math. Nachr.,
t. IV (1951), p. 70-80.
(XVI) L. SCHWARTZ, Theorie des distributions, 2" edition, Paris (Hermann), 1966.
(XVII) J. LINDENSTRAUSS and L. TZAFRIRI, Classical Banach spaces, t. I, Berlin-Heidelberg-
New York (Springer), 1977.
(XVIII) A. GROTHENDIECK : a) Produits tensoriels topologiques et espaces nucleaires, Mem.
Amer. Math. Soc., no 16 (1955) ; b) Espaces vectorie1s topologiques, 3e ed., Sao Paulo
(Pub!. Soc. Mat. sao Paulo), 1964.
 Index of notation

The reference numbers indicate the chapter and page (and, occasionally, exercise).

I~I, Ilxll : I, p. 3.
:?l(I; K), :?lK(I), f ~(I), f~(I), :?l(I), f1(I) : I, p. 4.
EA (A a convex symmetric set in a real vector space E) : II, p. 26.
< x, y) : II, p. 42.
cr(F, G) : II, p. 42.
MO, MOo: II, p. 44.
'u (u a linear mapping) : II, p. 46.
~(X) : III, p. 9.
'(i "'(U) : III, p. 9.
'(iH"'(U), '(i O"(U) : III, p. 9.
~ M(I), ~sCI), '(i (I) : III, p. 10.
Jfi(U), Jr(L) (U an open subset of en, L a compact subset of en) : III, p. 10.
~ (E; F) : III, p. 13.
~2 (E; F) : III, p. 13.
~s(E; F), ~c(E; F), ~ pc(E; F), ~cc(E; F), ~b(E; F) : III, p. 14.
E', E's, E~, E~, E~c' E~c' E~ : III, p. 14.
~(E), ~2 (E), ~s(E), ~cCE), ~pc(E), ~cc(E), ~b(E) : III, p. 14.
PM (p a semi-norm, M a bounded subset) : III, p. 14.
'(io(R) : III, p. 18.
,(E, F) : IV, p. 2.
~(E, F) : IV, p. 4.
CE : IV, p. 14.
P7(N) : IV, p. 17
co(N), pl (N) : IV, p. 18.
SeE) : IV, p. 26.
Hp : IV, p. 26.
Eo : IV, p. 32.
'??,(X) : IV, p. 33.
'(ibeX), '(i(X) : IV, p. 36.
:?l(X; R) : IV, p. 40.
Ind(u) (u a Fredholm operator) : IV, p. 66, exerc. 21.
~ : V, p. 1.
fl, C2 (N) : V, p. 4.
E(C) : V, p. 4.
<xIY), Ilxll = <XIX)1/2, (xIY) = <ylx) : V, p. 5.
E (E a complex prehilbertian space) : V, p. 6.
Jrs (Sobolev space) : V, p. 6.
H2(D) : V, p. 7.
'(i~(U) : V, p. 8.
PH (H convex separated and complete set in a prehilbertian space) : V, p. 10.
x* (x a vector of a Hilbert space) : V, p. 15 et p. 40.
(])Ei,EBEi : V, p. 18.
iEI iEI
E1 EB E2 EB ... EB En (Ei Hilbert spaces) : V, p. 18.
f~(I), f2(I) (E a Hilbert space) : V, p. 18.
348 TOPOLOGICAL VECTOR SPACES

E j O 2 E z,
Ilzllz (z E E 0 z E z) : V, p. 26.
j

E 0 z E z O 2... O 2E", ~7 Ei, Ilzllz (Z E~i Ei) : V, p. 27.

j

E j @2 E Z@2"'@2 E ", ®2 Ei:V,p.28

l.:Si':::;n

U1 ®z U z ® 2 ... ®
2 U" (U i linear mappings) : V, p. 28.
t"(E), E®" : V, p. 29.
S"(E), SeE) : V, p. 30.
t"(u), S"(u) (u a linear mapping) : V, p. 31 and p. 32.
A"(E), A(E) : V, p. 33.
A"(u) (u a linear mapping) : V, p. 34.
v. u, l'U (u, u linear mappings) : V, p. 37.
u* (u a linear mapping) : V, p. 38.
Jf'(E) (E a Hilbert space) : V, p. 44.
u ~ 0 (u an endomorphism of a Hilbert space): V, p. 45.
2' +(E) : V, p. 45.
u ~ v (u, u in 2'(E). E a Hilbert space) : V, p. 45.
,(u) (u an endomorphism of finite rank) : V, p. 48.
Tr(u) (u ~ 0 in £)(E») : V, p. 49.
2'1(E) (E a Hilbert space) : V, p. 51.
2' 2(E; F), .'fJ2 (E) (E, F Hilbert spaces) : V, p. 52.
I u liz (u E 2'(E; F), E, F Hilbert spaces) : V, p. 52.
Tr(Q/H) (Q, H positive quadratic forms) : V, p. 57.
 Index of terminology
Absorbent set, absorption of one set by another: I, p. 7.
Adapted bomology : !II, p. 3.
Adjoint: V, p. 38.
Affine transformation : IV, p. 39.
Associated (Hausdorff vector space) with a topological vector space: I, p. 4.

Balanced convex closed envelope of a set: II, p. 13 and p. 62.

Balanced core of a set: I, p. 7.
Balanced envelope of a set: I, p. 7.
Balanced set : r. p. 6.
Banach-Dieudonne theorem : IV, p. 24.
Banach-Saks-Kakutani theorem: V, p. 68, exerc. 33.
Banach space : I, p. 5.
Banach-Steinhaus theorem : III. p. 25.
Banach's theorem: t p. 17.
Barrel : III, p. 24.
Barrelled space : III, p. 24.
Base of a bomology : Ill, p. 1.
Basis (algebraic) of a Hilbert space: V, p. 22.
Basis (Banach) : IV, p. 69, exerc. 14.
Basis (complcte Banach, contracting Banach) : IV, p. 70, exerc. IS.
Basis (orthonormal) : V, p. 22.
Basis (unconditional Banach) : IV, p. 71, exerc. 16.
Bessel's inequality : V, p. 21.
Bidual : IV, p. 14.
Bipolar theorem : II, p. 44.
Birkhoff-Alaoglu theorem: V. p. 78. exerc. 13.
Bishop-Phelps theorem : II, p. 77. exerc. 4.
Bornologicallocally convex space: III, p. 12.
Bornology : III, p. 1.
Bornology (adapted) : Ill. p. 3.
Bornology (canonical) : III, p. 3.
Bomology (convex) : III, p. 2.
Bomology generated by a family of sets : III, p. 1.
Bomology (product) : III, p. 2.
Bomivorous set: Ill, p. 39, exerc. II.
Bounded set: III, p. 2 and p. 37, exerc. I.

Canonical bomology : III, p. 3.

Canonical mapping of ~ E; in (D Ei)' : IV, p. 13.
Canonical mapping of E in En : IV, p. 14.
Canonical mapping of E onto E' (E a Hilbert space) : V, p. 15.
Canonical topology on a finite dimensional vector space: I, p. 2.
Cap: II, p. 57.
Cauchy-Schwarz inequality: V, p. 3.
Closed graph theorem: I, p. 19.
Closed half-spaces defined by a closed hyperplane: II, p. 15.
Cobord : IV, p. 72, exerc. 3.
l-cocycle (continuous) : IV, p. 72, exerc. 3.
350 TOPOLOGICAL VECTOR SPACES

Compact linear mapping III, p. 6.

Compatible topology and structure of an ordered vector space: II, p. 15.
Compatible vector space structure and preorder : II, p. 12.
Compatible vector space structure and topology : III, p. I.
Compatible with the duality (locally convex topology) : IV, p. I.
Complement (orthogonal) : V, p. 13.
Complete topological vector space: I, p. 5.
Completion of a Hausdorff prehilbertian space: V, p. 8.
Completion of a Hausdorff topological vector space: I, p. 6.
Complex linear form: II, p. 6I.
Complex linear variety: II, p. 6I.
Complex locally convex space: II, p. 62.
Complexification (prehilbertian space) : V, p. 5.
Complexified topological vector space : II, p. 62.
Concave function : II, p. 17.
Cone (asymptotic) : II, p. 67.
Cone (convex) generated by a set: II, p. II.
Cone (pointed and non-pointed) : II, p. 10.
Cone (polyhedral) : 11, p. 9I.
Cone (proper pointed convex) : II, p. II.
Conjugate of a complex prehilbertian space: V, p. 6.
Convex balanced envelope of a set: II, p. 10 and p. 62.
Convex bomology : III, p. 2.
Convex closed envelope of a set: II, p. 13.
Convex function : II, p. 17.
Convex set : II, p. 7 and p. 62.
Convex (symmetric) envelope of a set: II, p. 16.
Coordinates with respect to an orthonormal base : V, p. 22.
Core (balanced) of a set : I, p. 7.

Density of order : V, p. 7.
Dimension (hilbertian) : V, p. 24.
Dimension of a convex set: II, p. 10.
Dirichlet space : V, p. 8.
Distal set: IV, p. 72, exerc. I.
Distinguished space : IV, p. 52, exerc. 4.
Dual (algebraic) of a real topological vector space: II, p. 42.
Dual of a locally convex space (real or complex) : III, p. 14.
Dual of a real topological vector space : II, p. 42.
Dual (weak, strong) : III, p. 14.
Duality separating in F, separating duality: II, p. 4I.
Duality (vector spaces in) : II, p. 40.
Dvoretzky-Rogers theorem: V, p. 63, exerc. 14.

Eberlein's theorem: IV, p. 35.

D. Edwards' theorem: II, p. 94, exerc. 4I.
Endomorphism (hermitian) : V, p. 44.
Endomorphism (normal) : V, p. 43.
Endomorphism (positive) : V, p. 45.
Envelope of a set (balanced) : I, p. 7.
Envelope of a set (balanced convex) : II, p. 10.
Envelope of a set (balanced convex closed) : II, p. 13 and p. 62.
Envelope of a set (closed convex) : II, p. 13.
Envelope of a set (convex) : II, p. 9.
Envelope of a set (symmetric convex) : II, p. 10.
Envelope of a set (symmetric convex closed) : II, p. 13.
6-equihypocontinuous, :t-equihypocontinuous, (6, :r )-equihypocontinuous set III, p. 47,
exerc. 7.
 INDEX OF TERMINOLOGY 351

Exhaustion of a Hausdorff locally convex space, exhaustible space: III, p. 49, exerc. 1.
Extremal generator of a convex cone : II, p. 57.
Extremal point of a convex set : II, p. 54.

Facet: II, p. 87, exerc. 3.

Facet (dual) : II, p. 87, exerc. 6.
Family (orthonormal) : V, p. 21.
Family (topologically independent) : I, p. 11.
Filtered set of semi-norms : II, p. 3.
Final locally convex topology : II, p. 27.
Final subspace: V, p. 41.
Fock spaces: V, p. 32 and p. 34.
Form (bilinear) putting two spaces in duality: II, p. 40.
Form (complex linear, real linear) : II, p. 61.
Form (hermitian) : V, p. l.
Form (positive hermitian) : V, p. 2.
Form (positive linear) : II, p. 13.
Form (separating hermitian) associated with a hermitian form: V, p. 2.
Frechet space: II, p. 24 and p. 63.
Function (concave, convex, strictly concave, strictly convex) : II, p. 16-17.
Function (positive definite) : V, p. 8.
Function (positively homogeneous), sub-linear function: II, p. 19-20.
Fundamental system of semi-norms : II, p. 3.

Gauge of a convex set: II, p. 20.

Gaussian space : V, p. 32.
Generated (bornology) by a family of sets: III, p. l.
Generator (extremal) of a convex cone: II, p. 57.
Gevrey's spaces : III, p. 10.
Gram's determinant: V, p. 71, exerc. 7.
Grothendieck's theorem : III, p. 20.
Group on which a mean can be defined: IV, p. 72, exerc. 4.

Haar's theorem : II, p. 83, exerc. 8.

Hadamard's inequalities: V, p. 37.
Hahn- Banach theorem : II, p. 22, p. 36 and p. 63.
Half-spaces (closed, open) determined by a closed hyperplane: II, p. 15.
Hardy space : V, p. 7.
Hausdorff completion of a topological vector space: I, p. 6.
Hausdorff vector space associated with a topological vector space: I, p. 4.
Helly's theorem: II, p. 68, exerc. 21.
Hermitian endomorphism : V, p. 44.
Hilbert-Schmidt mapping : V, p. 52.
Hilbert space, hilbertian space: V, p. 6.
Hyperplane (support) of a set: II, p. 37.
6-hypocontinuous, 'I-hypocontinuous, (6, 'I )-hypocontinuous bilinear mapping III,
p.30.

Index of a Fredholm operator: IV, p. 66, exerc. 2l.

Induced prehilbertian structure on a vector subspace: V, p. 6.
Inductive limit of locally convex spaces or topologies : II, p. 29.
Inequalities (Hadamard's) : V, p. 37.
Inequality (Bessel's) : V, p. 21.
Inequality (Cauchy-Schwarz) : V, p. 3.
Infra-barrelled space : III, p. 44, exerc. 7.
Initial subspace : V, p. 41.
Initial topology : IV, p. 4.
Internal point of a convex set: II, p. 26.
352 TOPOLOGICAL VECTOR SPACES

James-Klee theorem: IV, p. 57, exerc. 25.

R. C. James' space : IV, p. 71, exerc. 18.

Krein-Milman theorem: II, p. 55.

Krein's theorem: IV, p. 37.

Legendre polynomial : V, p. 24.

Locally convex complex space : II, p. 62.
Locally convex real space : II, p. 23.
Locally convex topology : II, p. 23 and p. 62.

Mackey theorem : IV, p. 2.

Mackey topology : IV, p. 2.
Mapping (Hilbert-Schmidt) : V, p. 52.
Markoff-Kakutani theorem: IV, p. 39.
Matrix (Hilbert's) : V, p. 75, exerc. 3.
Matrix with respect to orthonormal bases : V, p. 22.
Mean : IV, p. 40.
Metrisable topological vector space: I, p. 16.
Minimal type (space of) : II, p. 85, exerc. 13.
Montel space: IV, p. 18.

Normal endomorphism : V, p. 43.

Open half-spaces defined by a closed hyperplane : II, p. 15.

Operator (Fredholm) : IV, p. 66, exerc. 21.
Operator (unitary) : V, p. 41.
Ordered vector space : II, p. 12.
Orthogonal of a subspace for spaces in duality: II, p. 44.
Orthogonal projector, orthoprojector : V, p. 13.
Orthogonal sets in a prehilbertian space: V, p. 13.
Orthogonal sets in spaces in duality: II, p. 41.
Orthogonal to a subset in a prehilbertian space: V, p. 13.
Orthogoanl vectors for a duality: II, p. 41.
Orthogonal vectors in a prehilbertian space : V, p. 5.
Orthonormal basis : V, p. 22.
Orthonormal set, family: V, p. 21.
Orthonormalisation : V, p. 24.
Orthoprojector (initial, final) : V, p. 41.

Parabolic convex set: II, p. 67, exerc. 17.

Parseval's relation : V, p. 22.
Partially isometric mapping: V, p. 42.
Pointed cone : II, p. 10.
Points (exposed, of strict convexity) : II, p. 88, exerc. 6.
Points on the same side, strictly on the same side, of a hyperplane: II, p. 9.
Polar decomposition of a Hilbert-Schmidt endomorphism: V, p. 79, exerc. 14.
Polar of a set : II, p. 44 and 64.
Polarization formulas : V, p. 2.
Polyhedron : II, p. 90, exerc. 24.
Positive endomorphism : V, p. 45.
Positive hermitian form : V, p. 2.
Positively homogeneous function: II, p. 19.
Predual : IV, p. 56, exerc. 23.
Prehilbertian semi-norm : V, p. 4.
Prehilbertian space : V, p. 4.
Preordered vector space: II, p. 12.
Prlllclple of condensation of singularities: III, p. 42, exerc. 10.
 INDEX OF TERMINOLOGY 353

Product bornology : Ill, p. 2.

Projection on a convex set: V, p. 11.
Projection (orthogonal) : V, p. 13.
Proper pointed convex cone: II, p. 11.
Pythagoras' theorem: V, p. 12.

Quasi-complete space : III, p. 8.

Real linear form, real linear variety: II, p. 61.

Real locally convex space : II, p. 23.
Reflexive space: IV, p. 16.
Relatively bounded space.: III, p. 43, exerc. 6.
Representation (unitary) : IV, p. 44.
Ryll-Nardzewski theorem : IV, p. 43.

Scalar product : V, p. 5.
Scalar square : V, p. 5.
Segment (closed, open, open at x and closed at y) : II, p. 7.
Semi-automorphism of prehilbertian spaces : V, p. 6.
Semi-barrelled space: IV, p. 21.
Semi~complete space : III, p. 7.
Semi-norm: II, p. 1.
Semi-normed space : II, p. 2.
Semi~norm (prehilbertian) : V, p. 4.
Semi-reflexive space : IV, p. 15.
Separated completion of a prehilbertian space: V, p. 8.
Separated completion of a topological vector space : I, p. 6.
Separated (sets) by a closed hyperplane: II, p. 37.
Separately continuous bilinear mapping: III, p. 28.
Separately equicontinuous : III, p. 47.
Separating duality: II, p. 41.
Side of a hyperplane (points on the same, strictly on the same) : II, p. 8.
Simplex: II, p. 71, ex ere. 41.
Smulian theorem : IV, p. 36.
Sobolev space : V, p. 6.
Sole of a cone : II, p. 60.
Space (DF) : IV, p. 57, exerc. 2.
Space (topological vector) : I, p. 1.
Space (weak) : II, p. 42.
Square (scalar) : V, p. 5.
Starshaped set : II, p. 65, exerc. 1.
Strict inductive limit of a sequence of locally convex spaces, topologies: II, p. 33.
Strictly concave, strictly convex function : II, p. 16-17.
Strictly separated (sets) by a closed hyperplane: II, p. 37.
Strong dual: III, p. 14.
Strongly bounded subset of E' : III, p. 14.
Sub-linear function : II, p. 20.
Subspace (final, initial) of a continuous linear mapping: V, p. 41.
Subspace (prehilbertian) : V, p. 6.
Sum (external hilbertian) of Hilbert spaces: V, p. 18.
Sum (hilbertian) of vector subspaces : V, p. 18.
Sum (topological direct) of locally convex spaces or topologies: II, p. 30.
Support variety: II, p. 87, exerc. 3.
Symmetric convex closed envelope of a set: II, p. 13.

Tchebycheffs theorem: II, p. 84, exerc. 8.

Tensor product (hilbertian) : V, p. 28.
Tensor product of prehilbertian spaces: V. p. 26-27.
354 TOPOLOGICAL VECTOR SPACES

Topologically independent set, family: I, p. 12 and p. 11.

6-topology : III, p. 13 and IV, p. 2.
Topology compatible with an ordered vector space structure: II, p. IS.
Topology compatible with a vector space structure: I, p. I.
Topology defined by a semi-norm, by a set of semi-norms: II, p. 2-3.
Topology (initial) : IV, p. 4.
Topology (locally convex) : II, p. 23 and p. 62.
Topology (Mackey) : IV, p. 2.
Topology of simple, compact, precompact, compact convex, bounded Gonvergence III,
p. 14.
Topology (weak) : II, p. 42.
Topology (weakened) : IV, p. 4.
Total set: I, p. 11.
Trace of a continuous positive endomorphism: V, p. 49.
Trace of a positive quadratic form with respect to another: V, p. 57.
Transformation (affine) : IV, p. 39.
Transpose of a continuous linear mapping: II, p. 46 and IV, p. 6.

Ultrabornological space: III, p. 45, exerc. 19.

Ultranorm : I, p. 26, exerc. 12.
Ultra-semi-norm : II, p. 2.
Unitary operator : V, p. 41.
Unitary representation: IV, p. 44.
Unpointed cone: II, p. 10.

Variety (complex linear, real linear) : II, p. 61.

Variety (support) : II, p. 87, exerc. 3.

Weak dual: III, p. 14.

Weak topology, weak space: II, p. 42.
Weakened topology: IV, p. 4.
Weakly bounded: III, p. 14.
 Summary of some important properties
of Banach spaces

For the reader's convenience, the principal results of normed spaces and, more
particularly, of Banach spaces are collected here. The field of scalars K is either R or
(.

Linear mapping spaces " dual

1) Let E and F be two normed spaces. A linear mapping u ofE in F is continuous,
if and only if

(1) Ilull = sup I u(x) I

II x II ,,1
is finite. The mapping u ~ I ull is a norm on the vector space £,,(E; F) of continuous
linear mappings ofE in F.
Let F be a Banach space. Then £,,(E; F) is a Banach space. The completion £ of
E is a Banach space and the mapping u ~ ulE is a bijective isometry of £"(£; F) on
£,,(E; F).
2) Let E be a normed space. Write E' = £,,(E; K) where K carries the norm
A~ IAI. The Banach space E' is called the dual ofE, and the dual En ofE' is called the
bidualofE.
Denote by cr(E, E') the coarsest topology on E for which all the linear forms
x' E E' are continuous; it is called the weakened topology of E. Denote by cr(E', E)
>
the coarsest topology on E' for which the linear forms x' ~ <x', x on E' where x
varies in E, are continuous; then cr(E', E) is called the weak topology on E'. The
topology on E' deduced from the norm is called the strong topology.
3) Let E be a normed space and M be a closed vector subspace of E. Let 1t be the
canonical mapping ofE on ElM. A norm on the vector space ElM is defined by

(2) 111;11 = inf Ilxll·

n(x)=~

When E is a Banach space, then so also are M and ElM. For every normed space F,
the linear mapping u ~ u 0 1t of £"(E/M; F) in £,,(E; F) is isometric.
4) Let E be a normed space. For every x' E E', we have by definition
(3) Ilx'll = sup I<x', x>l·
IlxliG
XEE
356 TOPOLOGICAL VECTOR SPACES

Further (Hahn-Banach theorem), we have

(4) Ilxll = sup i<x', x)i

Ilx'IIU
x'EE'

for all x E E. In other words, the canonical mapping ofE in its bidual Eft is isometric.

Polars and orthogonals

5) Let E be a normed space. For every subset A of E (resp. B of E'),the polar of
A (resp. B) denoted by A (resp. BO) is the set of x' E E' (resp. x E E) for which
0

(5) fYl <x', x) ~ - I

for all x E A (resp. x' E B). When A (resp. B) is a vector subspace, the relation (5) is
<
equivalent to x', x) = 0, and we then say that A (resp. BO) is the orthogonal of
0

A (resp. B).
6) (<< The Bipolar Theorem »). Let E be a normed space. Let A (resp. B) be a sub-
set of E (resp. E') which contains O. Then the bipolar ADO of A (resp. BOO of B) is the
closure for the topology cr(E, E') (resp. cr(E', E)) ofthe convex envelope of A (resp. B).
7) Let A be a subset of a normed space E. Let x be a point in the closure of A with
respect to the topology cr(E, E'). Then x is the limit (in the norm sense) of a sequence
of points of the convex envelope of A. In particular, the convex subsets of E that are
closed in the normed space E are the same as those that are closed for cr(E, E').
8) Let E be a normed space and M be a vector subspace ofE. For every linear form
U o EM', there exists a linear from u E E' extending U o and such that Ilull = Iluoli.

Let H be the orthogonal of M in E' ; then the orthogonal HO of H is the closure of M

in E.

Transposition
9) Let E and F be two normed spaces and u E 2(E; F). The transpose
tu E 2(F'; E') of u is defined by the relation

(6) <u(y'), x) = <y', u(x) for all x E E, y' E F'.

We have Iitull = Iluli. The kernel of u is the orthogonal in E of the image of tu.
The kernel oftu is the orthogonal in F' of the image of u.
10) Let E be a normed space, M be a closed vector subspace of E and F = ElM.
Let,i be the canonical injection of M in E and let re be the canonical surjection of E
on F. Then ti has as its kernel the orthogonal MO of M and induces, on passing to the
quotient, an isometry ofE'/Mo on M'. Further tre is an isometry ofF' on MO.

Conditions for continuity of a linear mapping

11) Let E and F be two Banach spaces and u be a linear mapping ofE in F. Suppose
that for every sequence (xn)n",O of points ofE tending to 0 and for which the sequence
(u(xn))n",o has a limit yin F, theny is necessarily O. Then u is continuous.
 SUMMARY OF SOME PROPERTIES 357

* Suppose that for every compact subset K ofE, for every positive measure Jl on K
and for every linear continuous form y' on F, the restriction of y' 0 u to K is Jl-measu-
rable. Then u is continuous.*
12) Let E and F be two Banach spaces and u E iE (E; F). Then either u(E) is meagre,
or u is surjective.
Suppose that u is surjective. Then there exists a number C > 0 such that, for all
y E F, there exists x E E with u(x) = y and Ilxll ::::; C.llyll.lfN is the kernel ofu, then u
induces on passing to the quotient a homeomorphism ofE/N on F.
13) Let E and F be two Banach spaces. If u is a continuous linear mapping of E
in F that is bijective, then u- 1 is continuous.
14) Let E and F be two Banach spaces, let u E iE(E; F) and x' E E'. For x' to
belong to the image of t u, it is necessary and sufficient that there exists a number
C > 0 such that

(7) I<x', x)1 ::::; c.llu(x)11

for all x E E.
(15) Let E and F be two Banach spaces and u E iE(E; F). In order that u be sur-
jective, it is necessary and sufficient that there exists a number C > 0 such that
I tu(y') I ;?! C.II y' I for all y' E F.

The Banach-Steinhaus Theorem

16) (<< The Banach-Steinhaus Theorem »). Let E be a Banach space; F a normed
space and let (UJiEI be a family of elements of iE(E; F). Let A be the subset of x E E
such that sup I u;(x) II < + 00. Then either A is meagre and its complementis dense
iEI
in E, or alternatively sup II uill < + 00. In particular, if A = E, then sup I uill < + 00.
iEI

17) Let E and F be two Banach spaces and let (un)n"O be a sequence of elements
of iE(E; F). Suppose that the limit u(x) = lim un(x) exists for all x E E. Then
sup Ilunll < + 00, u is continuous and the sequence (un) tends to u uniformly on
n
every compact subset of E.
Properties of the weak topology on a dual

18) Let E be a Banach space and B' be a subset of E'. The following conditions are
equivalent:
(i) B' is contained in a ball of E'.
(ii) B' is relatively compact for the topology cr(E', E).
(iii) For all x E E, we have sup 1< x', x) I < + 00.
x'EB

19) Let E be a Banach space and let B' be the (closed) unit ball of E'. Then B' is
compact for cr(E', E). Suppose that there exists a countable total subset of E; then
B' is metrisable for cr(E', E), and there exists a countable subset ofE' that is dense for
cr(E', E).
358 TOPOLOGICAL VECTOR SPACES

20) Let E be a Banach space, u be a linear form on E' and B' be the unit ball of E'.
The following conditions are equivalent:
(i) There exists x E E such that u(x') = < x', x> for all x' E E'.
(ii) The restriction of u to B' is continuous for the topology aCE', E).
(iii) For every sequence (x~) of elements of E' that tends to 0 for aCE', E), we have
lim u(x~) = o.
21) Let E be a Banach space, B' be the unit ball ofE' and C be a convex subset of
E' (in particular a vector subspace). In order that C be closed for aCE', E), it is neces-
sary and sufficient that the intersection C n rB' be closed for aCE', E) for every real
number r > o.
Reflexive spaces
22) Let E be a normed space, E" be its bidual and i be the canonical mapping of E
in E". The unit ball of E" is the closure for a(E", E') of the image under i of the unit
ball of E.
The following conditions are equivalent :
(i) The isometric mapping i : E f--* E" is surjective.
(ii) The unit ball in E is compact for aCE, E').
When these conditions are satisfied, we say that E is reflexive.
Topologies compatible with the duality
23) Let E be a Banach space and !T a locally convex topology on E. The following
conditions are equivalent:
(i) The topology !T is finer than aCE, E') and coarser than the topology defined on
E by the norm.
(ii) E' is the set oflinear forms on E that are continuous for!T.
Suppose that these conditions are satisfied. Let A be a subset of E. Then A is rela-
tively compact for :Y if and only if every sequence of points of A has a cluster point
for!T in E. If this is so then the balanced convex envelope of A is relatively compact
for !T.
 Contents

CHAPTER I. - TOPOLOGICAL VECTOR SPACES OVER A VALUED DIVISION RING. I. 1

§ 1. Topological vector spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. I. 1

1. Definition of a topological vector space. . . . . . . . . . . . . . . . . .. I. 1
2. Normed spaces on a valued division ring ................. I. 3
3. Vector subspaces and quotient spaces of a topological vector
space; products of topological vector spaces; topological
direct sums of subs paces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1.4
4. Uniform structure and completion of a topological vector
space. . . . . . . . .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . .. . . . . . . .. 1.5
5. Neighbourhoods of the origin in a topological vector space
over a valued division ring. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1.6
6. Criteria of continuity and equicontinuity. . . . . . . . . . . . . . . . .. 1.8
7. Initial topologies of vector spaces . . . . . . . . . . . . . . . . . . . . . .. I. 9
§ 2. Linear varieties in a topological vector space. . . . . . . . . . . . . . . . . . .. I. 11
1. The closure of a linear variety. . . . . . . . . . . . . . . . . . . . . . . . . . .. I. 11
2. Lines and closed hyperplanes. . . . . . . . . . . . . . . . . . . . . . . . . . .. I. 12
3. Vector subspaces of finite dimension .................... I. l3
4. Locally compacttopological vector spaces. . . . . . . . . . . . . . . .. I. 15

§ 3. Metrisable topological vector spaces. . . . . . . . . . . . . . . . . . . . . . . . .. I. 16

1. Neighbourhoods of 0 in a metrisable topological vector space. I. 16
2. Properties of metrisable vector spaces. . . . . . . . . . . . . . . . . . . .. I. 17
3. Continuous linear functions in a metrisable vector space. . . .. 1.17
Exercises of § 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1.22
Exercises of § 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1.25
Exercises of § 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1.28

CHAPTER II. - CONVEX SETS AND LOCALLY CONVEX SPACES.............. 11.1

§ 1. Semi-norms... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 11.1
1. Definition of semi-norms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 11.1
2. Topologies defined by semi-norms. . . . . . . . . . . . . . . . . . . . . .. 11.2
3. Semi-norms in quotient spaces and in product spaces. . . . . . .. 11.4
360 TOPOLOGICAL VECTOR SPACES

4. Equicontinuity criteria of multilinear mappings for topologies

defined by semi-norms ................................ . II.5

§ 2. Convex sets . ............................................ . II.7

1. Definition of a convex set ............................. . II.7
2. Intersections ofconvex sets. Products of convex sets ........ . II.9
3. Convex envelope of a set. .............................. . II.9
4. Convex cones ........................................ . II.lO
5. Ordered vector spaces ................................ . n.12
6. Convex cones in topological vector spaces ............... . II.13
7. Topologies on ordered vector spaces .................... . II. 15
8. Convex functions .................................... . II .16
9. Operations on convex functions ........................ . II.18
10. Convex functions over an open convex set. ............... . II.18
11. Semi-norms and convex sets ........................... . II .19

§ 3. The Hahn-Banach Theorem (analytic form) .................. . II.21

l. Extension of positive linear forms ....................... . II.21
2. The Hahn-Banach theorem (analytic form) ............... . II. 22

§ 4. Locally convex spaces .................................... . II. 23

1. Definition of a locally convex space ..................... . II .23
2. Examples oflocally convex spaces ....................... . II.25
3. Locally convex initial topologies ....................... . II.26
4. Locally convex final topologies ......................... . II.27
5. The direct topological sum of a family oflocally convex spaces. II.29
6. Inductive limits of sequences oflocally convex spaces ....... . II.31
7. Remarks on Frechet spaces ............................ . II.34

§ 5. Separation of convex sets . ................................. . II.36

l. The Hahn-Banach theorem (geometric form) .............. . II.36
2. Separation of convex sets in a topological vector space ...... . II.37
3. Separation of convex sets in a locally convex space .......... . II.38
4. Approximation to convex functions .................... . II.39

§ 6. Weak topologies . ........................................ . II.40

1. Dual vector spaces ................................... . II.40
2. Weak topologies ..................................... . II.42
3. Polar sets and orthogonal subspaces .................... . II.44
4. Transposition of a continuous linear mapping ............. . II.46
5. Quotient spaces and subspaces of a weak space ............ . II.48
6. Products of weak topologies ............................ . II.50
7. Weakly complete spaces .............................. . II.51
8. Complete convex cones in weak spaces ................... . II.52
 CONTENTS 361

§ 7. Extremal points and extremal generators. . . . . . . . . . . . . . . . . . . . . .. II. 54

1. Extremal points of compact convex sets. . . . . . . . . . . . . . . . . .. II. 54
2. Extremal generators of convex cones. . . . . . . . . . . . . . . . . . . . .. II. 57
3. Convex cones with compact sole. . . . . . . . . . . . . . . . . . . . . . . .. II. 59

§ 8. Complex locally convex spaces . ............................ . II.60

1. Topological vector spaces over C ........................ . II. 60
2. Complex locally convex spaces .......................... . II.62
3. The Hahn-Banach theorem and its applications ............ . II.63
4. Weak topologies on complex vector spaces ................ . II.64

Exercises on §2 ................................................. . II.65

Exercises on §3 . . . . . . . . . . . . . . . . . . . . . ............................ . II. 72
Exercises on §4 ................................................. . II.74
Exercises on §5 : ................................................ . II. 76
Exercises on §6 . . . . . . . . . . . . . .................................... . II.81
Exercises on §7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... . II.87
Exercises on §8 ................................................. . II.95

CHAPTER III. - SPACES OF CONTINUOUS LINEAR MAPPINGS .............. . III. 1

§ 1. Bornology in a topological vector space. . . . . . . . . . . . . . . . . . . . . . .. III. 1

1. Bomologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. III. 1
2. Bounded subsets of a topological vector space. . . . . . . . . . . . .. III. 2
3. Image under a continuous mapping. . . . . . . . . . . . . . . . . . . . . .. III.4
4. Bounded subsets in certain inductive limits. . . . . . . . . . . . . . . .. III. 5
5. The spaces EA (A bounded). . . . . . . . . . . . . . . . . . . . . . . . . . . .. III. 7
6. Complete bounded sets and quasi-complete spaces. . . . . . . .. III. 8
7. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. III . 9

§ 2. Bornological spaces ...................................... . III.ll

§ 3. Spaces of continuous linear mappings . ....................... . III. 13

1. The spaces 26(E; F) .................................. . III. 13
2. Condition for 26 (E ; F) to be Hausdorff. ................ . III. 15
3. Relations between 2 (E; F) and 2 (E ; F) ................ . III. 15
4. Equicontinuous subsets of 2 (E ; F) ..................... . III. 16
5. Equicontinuous subsets ofE' ........................... . III. 19
6. The completion of a locally convex space ................. . III. 20
7. 6-bomologies on 2(E; F) ............................. . III. 21
8. Complete subsets of 26 (E ; F) ......................... . III. 22

§ 4. The Banach-Steinhaus theorem . ............................ . III. 23

1. Barrels and barrelled spaces ............................ . III. 24
362 TOPOLOGICAL VECTOR SPACES

2. The Banach-Steinhaus theorem ......................... III. 25

3. Bounded subsets of 2(E; F)(quasi-complete case). . . . . . . .. III. 27
§ 5. Hypocontinuous bilinear mappings . .......................... . III. 28
1. Separately continuous bilinear mappings ................. . III. 28
2. Separately continuous bilinear mappings on a product of Fre-
chet spaces .......................................... . III. 29
3. H ypocontinuous bilinear mappings ...................... . III. 30
4. Extension of a hypocontinuous bilinear mapping .......... . III. 32
5. Hypocontinuity of the mapping (u,v) ~ v 0 u ............. . III. 32
§ 6. Borel's graph theorem . ................................... . III. 34
1. Borel's graph theorem ................................. . III. 34
2. Locally convex Lusin spaces ............................ . III. 34
3. Measurablelinear mappings on a Banach space ............ . III. 36
Exercises on §1 .................. . . . . . . . . . . . . . . . . . . . . ........... . III. 37
Exercises on §2 ................................................. . III. 40
Exercises on §3 ................................................. . III .41
Exercises on §4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ......... . III.43
Exercises on §5 ................................................. . III. 46
Exercises on §6 ................................................. . III.49

CHAPTER IV. - DUALITY IN TOPOLOGICAL VECTOR SPACES . . . . . . . . . . . . . . . IV.l

§ 1. Duality .......................................... ....... . IV.l

1. Topologies compatible with a duality .................... . IV.l
2. Mackey topology and weakened topology on a locally convex
space ............................................... . IV.4
3. Transpose of a continuous linear mapping ................ . IV.6
4. Dual of a quotient space and of a subspace ................ . IV.8
5. Dual of a direct sum and of a product. ................... . IV. 11
§ 2. Bidual. Reflexive spaces . .................................. . IV.14
1. Bidual .............................................. . IV.14
2. Semi-reflexive spaces ................................. . IV.lS
3. Reflexive spaces ..................................... . IV.16
4. The case of normed spaces ............................ . IV.17
5. Montel spaces ....................................... . IV.18
§ 3. Dual of a Fn?chet space . .................................. . IV.21
1. Semi-barrelled spaces ................................. . IV.21
2. Dual of a locally convex metrizable space ................. . IV. 22
3. Bidual of a locally convex metrizable space ................ . IV.23
4. Dual of a reflexive Frechet space ....................... . IV.23
 CONTENTS 363

5. The topology of compact convergence on the dual of a Frechet

space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 24
6. Separately continuous bilinear mappings. . . . . . . . . . . . . . . . .. IV. 26
§ 4. Strict morphisms of Frechet spaces. . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 26
1. Characterizations of strict morphisms. . . . . . . . . . . . . . . . . . . .. IV. 27
2. Strict morphisms of Frechet spaces. . . . . . . . . . . . . . . . . . . . .. IV. 28
3. Criteria for surjectivity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 31
§ 5. Compactness criteria ...................................... IV. 32
1. General remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 32
2. Simple compactness of sets of continuous functions. . . . . . . . .. IV. 33
3. The Eberlein and Smulian theorems. . . . . . . . . . . . . . . . . . . . .. IV. 35
4. The case of spaces of bounded continuous functions. . . . . . . .. IV. 36
5. Convex envelope ofa weakly compact set. . . . . . . . . . . . . . . . .. IV. 37
Appendix. - Fixed points of groups of affine transformations. . . . . . .. IV. 39
1. The case of solvable groups. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 39
2. Invariant means ...................................... IV. 40
3. Ryll-Nardzewski theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV.41
4. Applications...... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV.44

Exercises on § 1 .................................................. IV. 47

Exercises on § 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 52
Exercises on § 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 57
Exercises on § 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 62
Exercises on § 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 67
Exercises on Appendix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. IV. 72
Table I. - Principal types oflocally convex spaces. . . . . . . . . . . . . . . . . . . .. IV. 75
Table II. - Principal bomologies on the dual of a locally convex space. . .. IV. 76

CHAPTER V. - HILBERTIAN SPACES (ELEMENTARY THEORY) . . . . . . . . . . . .. V. 1

§ 1. Prehilbertian spaces and hilbertian spaces. . . . . . . . . . . . . . . . . . . .. V. 1

1. Hermitian forms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. V. 1
2. Positive hermitian forms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. V. 2
3. Prehilbertian spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. V. 4
4. Hilbertian spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. V. 6
5. Convex subsets of a prehilbertian space. . . . . . . . . . . . . . . . .. V. 9
6. Vector subspaces and orthoprojectors. . . . . . . . . . . . . . . . . . .. V.12
7. Dual of a hilbertian space. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. V. 15
§ 2. Orthogonal families in a hilbertian space. . . . . . . . . . . . . . . . . . . . .. V. 17
1. External hilbertian sum of hilbert ian spaces. . . . . . . . . . . . . . .. V. 17
2. Hilbertian sum of orthogonal subspaces of a hilbertian space.. V. 18
364 TOPOLOGICAL VECTOR SPACES

3. Orthonormal families. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. V. 21
4. Orthonormalisation ...... . . . . . . . . . . . . . . . . . . . . . . . . . . . .. V. 23
§ 3. Tensor product of hilhertian spaces ......................... . V.25
1. Tensor product of prehilbertian spaces .................. . V.25
2. Hilbertian tensor product of hilbertian spaces ............ . V.28
3. Symmetric hilbertian powers ........................... . V.29
4. Exterior hilbertian powers ............................. . V.33
5. Exterior Multiplication ............................... . V.35
§ 4. Some classes of operators in hilbertian spaces . ................. . V.37
1. Adjoint ............................................. . V.38
2. Partially isometric linear mappings ..................... . V.4l
3. Normal endomorphisms .............................. . V.43
4. Hermitian endomorphisms ............................ . V.44
5. Positive endomorphisms .............................. . V.45
6. Trace of an endomorphism ............................ . V.48
7. Hilbert-Schmidt mappings ............................. . V.52
8. Diagonalization of Hilbert-Schmidt mappings ............. . V.55
9. Trace of a quadratic form with respect to another .......... . V.57
Exercises on § 1 ................................................. . V.60
Exercises on § 2 ................................................. . V.70
Exercises on § 3 ................................................. . V.73
Exetcises on § 4 ................................................. . V.74
Historical notes ................................................. . V.80
Bibliography .................................................... . V.92
Index of notation ................................................ . 347
Index of terminology ............................................. . 349
Summary of some important properties of Banach spaces ............. . 355
Contents ....................................................... . 359