Fabook

APPLICABLE FUNCTIONAL ANALYSIS
C.E. Chidume∗
International Centre for Theoretical Physics
Trieste, Italy
°c C.E. Chidume (July, 2006)

All rights reserved. No part of this book may be reproduced, stored in a
retrieval system, or transmitted in any form or by any means, electronic, me-
chanical, photocopying, recording or otherwise, without the prior permission
of the copyright owner.
∗
Permanent address: Department of Mathematics, University of Nigeria, Nsukka,
Nigeria.
To my wife, Ifeoma
and our children
Chukwudi, Uzoamaka, Kenechukwu and Okechukwu.
ii
Preface
This book is based on lectures I gave to the Postgraduate Diploma students

(Mathematics) at the International Centre for Theoretical Physics, Trieste, Italy and
to upper–level undergraduate and graduate students at the University of Nigeria,
Nsukka, Nigeria and the University of Jos, Jos, Nigeria over a period of several years.
Though the book is primarily addressed to upper–level undergraduate and grad-
uate students of pure and applied mathematics, its contents are a valuable resource
for researchers in several areas of mathematics and to physicists and engineers who
need a good understanding of basic principles of functional analysis in order to cope,
in particular, with modern trends in differential equations.
In the first part of the book, we present fundamental theorems of functional
analysis which are of paramount importance from the point of view of applications.
It is well known that “compactness” is central in mathematical analysis and its
numerous applications. In fact, one characterization of finite dimensional spaces is
a classical result known as the Heine–Borel Theorem which states that “a normed
linear space E is finite dimensional if and only if the unit ball in E is compact”.
The importance of this result in mathematics and in several applications is common
knowledge. Another very important classical result connected with compactness
is the well known Bolzano-Weierstrass Theorem which states that any bounded
sequence in IR has a convergent subsequence.
The Heine–Borel Theorem and the Bolzano-Weierstrass Theorem have a sort of
equivalent formulations in certain infinite dimensional spaces. The difficulty in ex-
tending them to all infinite dimensional spaces is that there are too many “open sets”
to contend with when a space is infinite dimensional; more precisely, the “topology”
of infinite dimensional spaces is “too big” to be able to give us “compactness”.
Thus, in order to obtain some form of compactness it is necessary to “cut down”
the number of open sets under consideration, i.e., it is necessary to reduce the size
of the topology of the infinite dimensional space E. This brings us to the idea of
weak topology on E. With respect to this topology we have what can be regarded as
analogues of the Heine–Borel and Bolzano-Weierstrass Theorems in some large class
of infinite dimensional spaces: A Banach space is reflexive if and only if the unit
ball in E is weakly compact ( this is a theorem of Kakutani which is the analogue
of Heine-Borel Theorem); Any bounded sequence in a reflexive space has a weakly
convergence subsequence (a theorem of Eberlin- Smul’yan which is an analogue of
the Bolzano-Wierstrass theorem).
Reflexive spaces are encountered in numerous applications. For example, the
Sobolev spaces W m,p (Ω), 1 < p < ∞, in which most PDE’s are done is reflexive. In
particular, all Hilbert spaces are reflexive. Consequently, these analogues have wide
iii
variety of applications.
The aim of this book is, in summary, to present the above results of Kakutani
and Eberlein-Smul’yan . Furthermore, we shall present two important applications
of these theorems: one - a fundamental theorem in optimization theory, and the
other, the Browder-Ghode-Kirk fixed point theorem for nonexpansive mappings.
In order to understand these theorems of Kakutani and Eberlein-Smul’yan, one
requires a reasonably good knowledge of the weak topology and to understand the
weak topology and prove some of its basic theorems, one requires virtually all the
fundamental theorems of functional analysis.
Most of the existing books in functional analysis cover a lot more than these fun-
damental theorems (including applications in several diverse fields) and the reader
is left to choose what he or she wants to study. In the Preface of his recent excellent
book on functional analysis, J.B. Conway wrote, ”Functional analysis has grown so
much that you can find two mathematicians who claim to be functional analysts but
they do not understand each other”. But whatever direction one wants to pursue,
the following topics covered in this book are fundamental:
Normed Linear Spaces; Hilbert Spaces; Linear Maps; Hahn Banach Theo-
rem (Analytic and Geometric Forms); Uniform Boundedness Principle; Open
Mapping and Closed Graph Theorems; Weak and Weak∗ Topologies; Reflexive
Spaces; the Banach Alaoglu Theorem; the Theorem of Kakutani characterizing
reflexive spaces; Uniformly Convex Spaces, Milman-Pettis Theorem; Eberlein-
Smul’yan Theorem.
A thorough understanding of these topics at the upper–level undergraduate and

graduate level requires about 30 hours of lectures and tutorials. Consequently, the
topics are ample for a one semester course in functional analysis. Thus, there is the
need for a textbook covering just the above topics.
This book which is written in a leisurely literary style, yet covering the above
topics rigorously and in detail is designed to fill this need. There are miscellaneous
exercises at the end of the book covering the materials presented in the book. These
exercises constitute a very important part of the book and adequate time should be
devoted to them. A solutions manual containing complete solutions to all the exer-
cises is available. The book is, therefore, designed to be used as a formal textbook
for a course in functional analysis at the upper–undergraduate and graduate levels.
It is also directed to a reader who may not have easy access to an expert in the field
(in which case, he or she may use the solutions manual as needed). Another special
feature of this book is that there are numerous examples which are worked out in
detail and “pictures”, all designed to assist the reader.
iv
I would like to express my profound gratitude to my students over the last
fifteen years or so for their interest in the subject and their numerous questions and
comments which helped me to improve on the exposition. I would like to thank
specially Dr. Habtu Zegeye, my former student, for reading the manuscript, for his
very useful comments and criticisms and for typesetting the final version of the book.
I am also grateful to my former student, Dr. Ani Udomene for typesetting work
and for helping with drawing the pictures, especially those in chapter 10. My son,
Chu-chu ,who is a mathematician, read the final version of the book, made useful
suggestions produced solutions to many of the exercises. I am grateful to him.
Special thanks go to my wife and our children for their patience, understanding and
encouragement. I would also like to thank the University of Nigeria, Nsukka, Nigeria
for granting me leave of absence during which I worked at the International Centre
for Theoretical Physics, Trieste, Italy and had adequate time to concentrate on the
book. It is my pleasure to thank the International Centre for Theoretical Physics,
for hospitality. In particular, my special thanks go to the Publications Section of
the ICTP for the excellent job of typing the original manuscript.
C.E. Chidume
Trieste, 2006
v
vi
Contents
1 Normed Linear Spaces 1

1.1 Normed Linear Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Incomplete normed linear spaces . . . . . . . . . . . . . . . . . 12
1.3 The Contraction Mapping Principle . . . . . . . . . . . . . . . . . . 16
1.4 Proof of Hölder’s Inequality . . . . . . . . . . . . . . . . . . . . . . 20
2 Linear Maps 23
2.1 Definition and some examples . . . . . . . . . . . . . . . . . . . . . . 23
2.2 A basic result concerning linear maps . . . . . . . . . . . . . . . . . . 24
2.3 Bounded Linear Maps . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Equivalent norms and finite dimensional spaces . . . . . . . . . . . . 32
3 The Hahn Banach Theorem: Analytic and Geometric Forms. 37

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Some Consequences of Hahn Banach Theorem . . . . . . . . . . . . . 40
3.3 Geometric Forms of Hahn Banach Theorem . . . . . . . . . . . . . . 45
3.4 Proof of Theorem 3.7 . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4 Uniform Boundedness Principle 59

4.1 Introduction and Preliminaries . . . . . . . . . . . . . . . . . . . . . . 59
4.2 The Uniform Boundedness Principle . . . . . . . . . . . . . . . . . . . 62
vii
5 Open Mapping and Closed Graph Theorems 67
5.1 The Open Mapping Theorem . . . . . . . . . . . . . . . . . . . . . . 67
5.2 The Closed Graph Theorem . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.1 Closed Maps, Closed Graphs . . . . . . . . . . . . . . . . . . . 69
5.2.2 The Closed Graph Theorem . . . . . . . . . . . . . . . . . . . 72
5.3 Some Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6 Hilbert Spaces 81
6.1 Definition and Examples . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.3 Completeness of an Inner Product Space . . . . . . . . . . . . . . . . 85
6.4 Jordan Von Neumann Theorem . . . . . . . . . . . . . . . . . . . . . 86
6.5 Orthonormal Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.6 The Projection Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.7 The Riesz Representation Theorem . . . . . . . . . . . . . . . . . . . 94
6.8 Complete Orthonormal Sets . . . . . . . . . . . . . . . . . . . . . . . 94
7 Operators on Hilbert Spaces 101

7.1 Self-adjoint, Normal and Unitary Operators . . . . . . . . . . . . . . 103
7.1.1 Self-adjoint operators . . . . . . . . . . . . . . . . . . . . . . . 103
7.1.2 Normal Operators . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.1.3 Unitary Operators . . . . . . . . . . . . . . . . . . . . . . . . 105
8 Weak and Weak∗ Topologies 109

8.1 Notions from general topology . . . . . . . . . . . . . . . . . . . . . . 109
8.2 The weak topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.2.1 Open and closed sets . . . . . . . . . . . . . . . . . . . . . . . 116
8.2.2 Weak Topology and Convex Sets . . . . . . . . . . . . . . . . 118
8.3 The Weak Star Topology . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.4 The Banach Alaoglu Theorem . . . . . . . . . . . . . . . . . . . . . . 120
9 Reflexive and Uniformly Convex Spaces 121

9.1 Reflexive Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
9.1.1 A Theorem of Kakutani . . . . . . . . . . . . . . . . . . . . . 122
9.1.2 Eberlein–Smul’yan Theorem . . . . . . . . . . . . . . . . . . . 128
9.2 Uniformly Convex Banach Spaces . . . . . . . . . . . . . . . . . . . . 129
9.2.1 Milman–Pettis Theorem . . . . . . . . . . . . . . . . . . . . . 133
9.3 Separable spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
9.3.1 A Non–separable Inner Product Space . . . . . . . . . . . . . 139
viii
10 Application in Optimization 141
10.1 Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
10.2 Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
10.2.1 Notations and further definitions . . . . . . . . . . . . . . . . 146
10.3 Strictly Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . 148
10.4 Lower Semi-continuity . . . . . . . . . . . . . . . . . . . . . . . . . . 150
10.5 A fundamental theorem of optimization . . . . . . . . . . . . . . . . . 155
11 Application in Fixed Point Theory 159

11.1 Some Fixed Point Theorems . . . . . . . . . . . . . . . . . . . . . . . 159
11.1.1 A fixed point theorem . . . . . . . . . . . . . . . . . . . . . . 161
11.2 Normal structure in Banach spaces . . . . . . . . . . . . . . . . . . . 162
11.3 A fixed point theorem for nonexpansive mappings . . . . . . . . . . . 165
12 Function Spaces; Weak Derivatives 171

13.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
13.2 Spaces of Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
13.3 Pseudo Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
13.4 Distributional or Weak Derivatives . . . . . . . . . . . . . . . . . . . 176
13 An Introduction to Sobolev Spaces and Applications 179

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
14.2 Abstract Minimization Problem . . . . . . . . . . . . . . . . . . . . . 180
14.2.1 The Symmetric Case . . . . . . . . . . . . . . . . . . . . . . . 180
14.2.2 The Non-symmetric case . . . . . . . . . . . . . . . . . . . . . 183
14.2.3 The Lax-Milgram Theorem . . . . . . . . . . . . . . . . . . . 183
14.3 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
14.4 Applications to Partial Differential Equations . . . . . . . . . . . . . 189
Miscellaneous Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
A1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
A1.2 Some basic lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
A1.3 The metric space (W, ρW ) . . . . . . . . . . . . . . . . . . . . . . . . 205
ix
x
CHAPTER 1
Normed Linear Spaces
1.1 Normed Linear Spaces

This chapter is basically a quick review of normed linear spaces, completeness and
the contraction mapping principle which the reader is assumed to be familiar with.
We begin with the following definition.
Definition 1.1 Let X be a linear space over a field K, where K = IR or CI. A norm
on X is a real-valued function ||.||,
||.|| : X → [0, ∞)
which satisfies the following conditions:

N1: kxk = 0 if and only if x = 0, x ∈ X;
N2: kkxk = |k|.kxk for all k ∈ K, x ∈ X;
N3: kx + yk ≤ kxk + kyk for arbitrary x, y ∈ X.

A linear space with a norm defined on it is called a normed linear space.
Condition N 3 is called the Triangle Inequality.
Example 1.2 Let X = IR2 . For arbitrary x̄ = (x1 , x2 ) in X define
k.k2 : IR2 → [0, ∞) by
1
Fig. 1.1
kx̄k2 = (x21 + x22 )1/2 .

Then, k · k2 is a norm on IR2 (see Fig. 1.1.)
Verification that ||.||2 is a norm on IR2 requires the use of the following inequality
called the Schwartz’s inequalityfor finite sums, i.e., for xi , yi ∈ CI, (where CI denotes
the set of complex numbers) i = 1, 2, ..., n, we have
n
X ³X ´ 12 ³X ´ 12
|xi yi | ≤ |xi |2 |yi |2 .
i=1
(A proof of this inequality is a corollary of a more general inequality which is proved

at the end of this chapter).
Verification. Conditions N 1 and N 2 are clearly satisfied (convince yourself).

For N3: Let x̄ = (x1 , x2 ) and ȳ = (y1 , y2 ) be arbitrary elements of IR2 . Then,
2
X
||x̄ + ȳ||22 = k(x1 + y1 , x2 + y2 )k22 = (xi + yi )2
i=1
2
2
X 2
X 2
X 2
X
= (x2i + 2xi yi + yi2 ) ≤ x2i +2 |xi yi | + yi2 (why?)
i=1 i=1 i=1 i=1
2
Ã 2 !1/2 Ã 2 !1/2 2
X X X X
≤ x2i + 2 x2i yi2 + yi2
i=1 i=1 i=1 i=1
2 2
= A + 2AB + B ,
µ ¶1/2
¡P 2 ¢
2 1/2
P
2
where A = i=1 xi = kx̄k2 and B = yi2 = kȳk2 . Hence
i=1
kx̄ + ȳk22 ≤ (A + B)2 = (kx̄k2 + kȳk2 )2 and so

kx̄ + ȳk2 ≤ kx̄k2 + kȳk2 , verifying N3.
Example 1.3 The space `p , 1 ≤ p < ∞.

For 1 ≤ p < ∞, let `p be defined as follows:
∞
X
`p (K) := {x̄ = (x1 , x2 , x3 , . . .), xi ∈ K : |xi |p < ∞}.
i=1
For x̄ = (x1 , x2 , x3 , . . .), ȳ = (y1 , y2 , y3 , . . .) arbitrary elements of `p , define x̄ + ȳ =

(x1 + y1 , x2 + y2 , x3 + y3 , . . .) and λx̄ = (λx1 , λx2 , λx3 , . . .). Then `p is a linear space.
P
Example 1.4 Consider the space, l1 := {x = (x1 , x2 , ...) : xi ∈ R, |xi | < ∞}.
(a) Let x = (1, 12 , 13 , ...). Is x ∈ l1 ?.
To answer this question, we compute the sum

X 1 1 X1
|xi | = 1 + + , ...) = .
2 3 n
Clearly, this series diverges and we conclude that x is not in l1 .
1
(b) Consider the sequence xn := n2 +1
,n = 1, 2, .... Is xn ∈ l1 ? Again, we compute,
X X 1 X 1
|xn | = < < ∞,
n2 + 1 n2
by the p−series test with p = 2 > 1. Hence xn ∈ l1 .
P
Example 1.5 Let X = l2 := {x = (x1 , x2 , ...) : xi ∈ R, |xi |2 < ∞}.
Let x = (1, 12 , 31 , ...). Is x ∈ l2 ?.
3
To answer this question, we again compute the sum
X X 1
|xi |2 = < ∞.
n2
Hence, x ∈ l2 . Recall that this x is not in l1 .
1
(b)Consider the sequence xn := n2 +1
,n = 1, 2, .... Is xn ∈ l2 ? Again, we compute,
X X 1 X 1
|xn |2 ≤ < < ∞,
(n2 + 1)2 n4
by the p−series test with p = 4 > 1. Hence xn ∈ l2 .
Example 1.6 The space `∞ .

We define `∞ as follows:
`∞ := {x̄ = (x1 , x2 , . . .), xi ∈ IR : x̄ is bounded},
i.e., `∞ is the set of all bounded sequences. If addition and scalar multiplication are
defined componentwise as in Example 1.3, then `∞ is a linear space. More generally,
we have the following proposition.
Proposition 1.7 If 1 ≤ p ≤ q ≤ +∞ then, `p ⊆ `q .
Proof. We consider two cases.

P
(i) q = ∞ and p < ∞. Let {xn }n≥1 ∈ lp . Then |xn |p < ∞. This implies that
|xn | → 0 as n → ∞, so that supn≥1 |xn | < ∞. This, in turn, implies, {xn } ∈ lq (= l∞ ).
P∞
(ii)p < q < ∞. Let {xn } ∈ lp . Then n=1 |xn |p < ∞. As in part (i), M :=
supn≥1 |xn | < ∞. Hence,
∞
X ∞
X
q q−p
|xn | ≤ M |xn |p < ∞,
n=1 n=1
and this implies {xn } ∈ lq . Hence, lp ⊆ lq .

¤
We now give our next example.
4
Example 1.8 The function k.k : `p / [0, ∞) defined by
Ã∞ !1/p
X
kx̄klp = |xi |p ,
i=1
for arbitrary x̄ = (x1 , x2 , . . .) in `p , is a norm on `p .
Proof. The verification that k.klp satisfies N1 to N3 is easily carried out. The details
are left to the reader (see Exercises 1.4, Problem 4). The condition N3 is verified
by using Hölder’s inequality for infinite sums i.e., if 1 ≤ p < ∞ and p1 + 1q = 1, and
if ai , bi , i = 1, 2, . . . , are complex numbers such that {ai }∞ ∞
i=1 ∈ lp and {bi }i=1 ∈ lq ,
then Ã∞ !1/p Ã ∞ !1/q
X∞ X X
|ai bi | ≤ |ai |p |bi |q .
i=1 i=1 i=1
(The proof of Hölder’s inequality is given at the end of this chapter, Section 1.4).
¤
Proposition 1.9 The function k.k`∞ : `∞ → [0, ∞) defined by
kx̄k`∞ = sup |xi |,

i≥1
for arbitrary x̄ = (x1 , x2 , . . .) in `∞ , is a norm on `∞ .
Proof. The obvious verification is left as an easy exercise.

¤
Example 1.10 Let X = (C[a, b]) denote the set of all real-valued functions which
are functions of an independent real variable t and are defined and continuous on
a given closed and bounded interval [a, b]. For arbitrary f, g ∈ C[a, b], let addition
and scalar multiplication be defined as follows:
(f + g)(t) = f (t) + g(t);
(λf )(t) = λf (t), for arbitrary scalar λ.

With these definitions, C[a, b] is a linear space. Now define, for arbitrary f ∈ C[a, b],
the real-valued functions k.k∞ , k.k1 and k.k2 as follows:
kf k∞ = sup |f (t)|;
t∈[a,b]
5
Z b
kf k1 = |f (t)|dt;
a
and µZ ¶1/2
b
2
kf k2 = |f (t)| dt .
a
Then (X, k.k∞ ), (X, k.k1 ), and (X, k.k2 ) are normed linear spaces.
The proofs that k.k∞ , and k.k1 are norms on C[a, b] are trivial and are left as an
easy exercise for the reader. For, k.k2 , see Exercises 1.1, Problem 7 where generous
hints are supplied.
Example 1.11 The space, c.

Let c := {x̄ = (x1 , x2 , ...) : {xi }∞
i=1 converges }. For arbitrary x̄ = (x1 , x2 , ...) in c,
define ||x̄||c = supi≥1 |xi |. Then, ||.||c is a norm on c.
Example 1.12 The space, c0 .

Let c0 := {x̄ = (x1 , x2 , ...) : xn / 0 as n / ∞}. For arbitrary x̄ = (x1 , x2 , ...)
in c0 , define ||x̄||c0 = supi≥1 |xi |. Then, ||.||c0 is a norm on c0 .
Remark 1.13 We first note that c and c0 are clearly proper subspaces of l∞ . Some
other important examples of normed linear spaces are given in Exercises 1.1 and
more will be introduced in what follows.
EXERCISES 1.1
1. Let IRn := {x = (x1 , x2 , ..., xn ), xi ∈ IR} with addition and scalar multiplica-
tion defined componentwise. Let the real-valued functions k.k∞ , k.k1 and k.k2
be defined on IRn as follows:
kxk∞ = max |xi |;

1≤i≤n
n
X
kxk1 = |xi |;
i=1
6
nX
n o 12
2
kxk2 = |xi | .
i=1
(a) Draw a 2-dimensional picture (n = 2) for each of these functions.
(b) Verify that (IRn , k.k∞ ), (IRn , k.k1 ) and (IRn , k.k2 ) are normed linear spaces.
( The space (IRn , k.k2 ) is generally called the Euclidean space and the space
(IRn , k.k∞ ) is sometime denoted by `n∞ .
2. Let IRn be as in Exercises 1.1 (1). Define a real-valued function k.kp (p ≥ 1)
as follows: for arbitrary x = (x1 , x2 , ..., xn ) ∈ IRn
³X
n ´ p1
kxkp = |xi |p .
i=1
Verify that k.kp is a norm on IRn . The space (IRn , k.kp ) is usually denoted by
`np .
3. Let CIn := {z = (z1 , z2 , ..., zn ), zi ∈ CI} where addition and scalar multiplication
are defined componentwise. Let the real-valued function k.k2 be defined on CIn
as follows:
³Xn ´ 12
kzk2 = |zi |2 .
i=1
n
Verify that (CI , k.k2 ) is a normed linear space. This space is usually referred
as the unitary space .
4. Verify that the function ||.||lp defined in Example 1.8 is a norm on lp .
(Hint: For N3, let x = {ai }∞ ∞
i=1 ∈ lp and y = {bi }i=1 ∈ lp . Then,
∞
X ∞
X
||x + y||plp = p
|ai + bi | = |ai + bi |.|ai + bi |p−1
i=1 i=1
X∞ ∞
X
p−1
≤ |ai |.|ai + bi | + |bi |.|ai + bi |p−1 .
i=1 i=1
Then apply Hölder’s inequality. Justify the application).

5. Let X = C 1 [a, b] be the space of all continuous real-valued functions on [a, b]
which also have continuous derivatives on [a, b]. For arbitrary f ∈ C 1 [a, b],
define
df (t)
||f || = max |f (t)| + max | |.
a≤t≤b a≤t≤b dt
7
Verify that X, with addition and scalar multiplication as in Example 1.10, is
a normed linear space.
6. Let X = C[a, b] be the space of all real-valued continuous functions on [a, b].
For arbitrary f ∈ X, define
Z b
||f || = |f (t)|dt.
a
Verify that X, with addition and scalar multiplication as defined in Example

1.8, is a normed linear space.
7. Let X = C[a, b] be the space of all continuous real-valued functions on [a, b]

with
µZ b ¶ 12
2
||f || = |f (t)| dt ,
a
for arbitrary f ∈ X. Verify that X is a normed linear space.

Hint: Use Cauchy–Schwartz’s inequality for integrals :
For arbitrary f, g ∈ C[a, b],
Z b µZ b ¶1/2 µZ b ¶1/2
2 2
|f g|dt ≤ |f | dt |g| dt .
a a a
8. Let X be a vector space. A function f : X → IR is said to be a convex function

if for each x, y ∈ X, and λ ∈ [0, 1], the following inequality is satisfied:
f [λx + (1 − λ)y] ≤ λf (x) + (1 − λ)f (y).
Verify that norm is a convex function. What can you say about ||.||2 ?
9. Let ¯X be a normed¯ linear space. Prove that for arbitrary x, y ∈ X,

¯ ¯
(a) ¯||x|| − ||y||¯ ≤ ||x − y||;
(b) The mapping x → ||x|| is continuous (in the sense that if xn → x
then ||xn || → ||x||);
(c) Addition and scalar multiplication are jointly continuous, i.e., if
xn → x and yn → y then xn + yn → x + y; and if xn → x and
an → a then an xn → ax as n → ∞, where an and a are real scalars.
8
1.2 Completeness
In this section we shall consider the concept of completeness in normed linear spaces.
For this, it is more convenient to work in metric spaces with which the reader is prob-
ably very familiar. (For a quick review of metric spaces the reader may consult [9]).
We first obseve that if (E, k · k) is a normed linear space, the norm k · k always
induces a metric ρ on E given by ρ(x, y) = kx − yk for each x, y ∈ E. With this,
(E, ρ) becomes a metric space. Recall that a sequence {xn } in a metric space (E, ρ)
is called Cauchy if ρ(xn , xm ) / 0, as n, m / ∞, and that a metric space (E, ρ)
is said to be complete if every Cauchy sequence in E converges to an element of
E. Completeness is a very important concept in functional analysis and this will
become more evident in subsequent chapters.
To check or verify that a metric space (X, ρ) is complete, we take an arbitrary

Cauchy sequence {xn }∞ n=1 in X and show that it converges to a point in X. The
general pattern is the following:
(a) Construct an element x∗ which is to be used as the limit of the Cauchy sequence.
(b) Prove that x∗ is in the space under consideration.
(c) Prove that xn / x∗ (in the sense of the metric under consideration).
Step (a) is easily stated but generally requires some trick to handle. To construct
x∗ mentioned in (a), one generally uses the fact that the given sequence {xn }∞ n=1
is Cauchy to generate a Cauchy sequence in a complete space that is normally as-
sociated with the normed linear space under consideration. (This complete space is
usually the reals, IR or the complex numbers, CI ). Once x∗ is constructed, steps (b)
and (c) are generally not too difficult to complete. We give examples to illustrate
the above steps. In the sequel we shall assume the well known fact that IR and CI
are complete.
Theorem 1.14 The normed linear space l2 is complete.
½ ¾
P
∞
2
Proof. Recall that l2 := x = (x1 , x2 , . . . , ...) : |xi | < ∞ .
i=1
9
Let {x(m) }∞
m=1 be a Cauchy sequence in l2 where,
(1) (1)
(1) (1)

x(1) = (x1 , x3 , . . . , xn , . . .) 
x2 , 




(2) (2) (2) (2) 

x = (x1 , x2 , x3 , . . . , xn , . . .) 
(2)
.. .. .. .. .. . (1.1)
. . . . . 

(m) (m) (m) (m) 

x(m) = (x1 , x2 , x3 , . . . , xn , . . .) 


.. .. .. .. .. 
. . . . .
{x(m) }∞m=1 is Cauchy implies, given ε > 0, there exists an integer N > 0 such that
for all m, t, ≥ N,
ρ(x(m) , x(t) ) < ε.
This implies, according to the definition of the norm in l2 , that
P∞ ¯ ¯2
(m) (t) 2 ¯ (m) (t) ¯
||x − x || = ¯ xi − xi ¯ < ε2 . (1.2)
i=1
These inequalities imply that each column in the array (1.1) is Cauchy. Since each
column is a sequence in IR, and IR is complete, it follows that each column converges
(m) / x∗ as m / ∞ (see array
to a point in IR. Without loss of generality, let xi i
(1.3) below) 
(1) (1) (1) (1)
x(1) = (x1 , x2 , x3 , . . . , xn , . . .)  





x = (x1 , x2 , x3 , . . . , xn , . . .) 
(2) (2) (2) (2) (2) 


.. 

. 



x = (x1 , x2 , x3 , . . . , xn , . . .) 
(m) (m) (m) (m) (m)
.. . (1.3)
. 



x(t) = (x1 , x2 , x3 , . . . , xn , . . .) 
(t) (t) (t) (t) 


.. .. .. .. .. 

. . . . . 



↓ ↓ ↓ ↓ ↓ 



∗ ∗ ∗ ∗
(x1 x2 x3 xn , . . .)
Now define x∗ = (x∗1 , x∗2 , x∗3 , . . . , x∗n , . . .). (This completes step (a)).
Observe that we have used the completeness of IR to get x∗j ; j = 1, 2, . . . , n, . . . and

from x∗j ’s we are able to define x∗ .
P∞
For step (b), we need to prove x∗ ∈ l2 , i.e., we need to show that i=1 |xi |2 < ∞.
10
This is not easy to show. So, we employ the following strategy. Using the Cauchy
condition, we show that x(n) − x∗ ∈ l2 . We know x(n) ∈ l2 . Since l2 is a linear
(vector) space, we have x(n) − (x(n) − x∗ ) ≡ x∗ ∈ l2 . We proceed to do this.
Let k ∈ IN be arbitrary. It follows from the Cauchy condition (1.2) that for
m, t ≥ N , we have
P
k
(m) (t) P
∞
(m) (t)
|xi − x i |2 ≤ |xi − x i |2 < ² 2 . (1.4)
i=1 i=1
Let m be fixed, and let t → ∞ in inequality (1.4), (a finite sum). Then, we get for
every m ≥ N ,
X k
(m)
|xi − x∗i |2 ≤ ²2 .
i=1
Since k is an arbitrary integer, we have, for every m ≥ N,

P∞ (m)
i=1 |xi − x∗i |2 ≤ ²2 . (1.5)
Now from (1.5), we have for m = N, x(N ) − x∗ ∈ l2 and since x(N ) ∈ l2 , we have
x∗ = x(N ) − (x(N ) − x∗ ) ∈ l2 (because l2 is a linear space). This completes step (b).
Finally, from inequality (1.5), there exists a positive integer N such that ||x(m) −
x∗ || ≤ ² for all m ≥ N . So, x(m) → x∗ in l2 . The proof is complete.
¤
Theorem 1.15 C[a, b] (endowed with “sup norm” in Example 1.10) is complete.
Proof. Let {fn }∞ n=1 be a Cauchy sequence in C[a, b]. For each t ∈ [a, b], there
exists a positive integer N such that given ε > 0,
kfn − fm kC[a,b] := sup |fn (t) − fm (t)| < ε ∀ n, m ≥ N.

t∈[a,b]
Hence, for any fixed t0 ∈ [a, b], |fn (t0 ) − fm (t0 )| < ε for all n, m ≥ N . This
shows that {fn (t0 )}∞
n=1 is a Cauchy sequence of real numbers. Since IR is complete,
{fn (t0 )}∞ converges to a real number, say, f (t0 ) as n / ∞, i.e.,
n=1
fn (t0 ) / f (t0 ) as n / ∞.
This is the same as saying that the function, fn converges pointwise to the function
f.
We prove now that this pointwise convergence is actually uniform in t ∈ [a, b], i.e.,
given ε > 0, there is an integer N ∗ such that sup |fn (t) − f (t)| < ε for all n ≥ N ∗ .
t∈[a,b]
11
Now, for any ² > 0, there exists an integer N > 0 such that for all n, m ≥ N , we
have
||fn − fm ||C[a,b] < ².
This means, for n, m ≥ N, we have
sup |fn (t) − fm (t)| < ².

t∈[a,b]
So, for n, m ≥ N , we have
|fn (t) − fm (t)| < ²∀t ∈ [a, b].
Fix n ≥ N , and let m → ∞ to get,
|fn (t) − f (t)| < ².
Since t is arbitrary, we have for n ≥ N ,
sup |fn (t) − f (t)| < ², i.e., ||fn − f || < ²,

t∈[a,b]
and so the convergence is uniform. This implies f ∈ C[a, b] and completes the proof.
¤
Definition 1.16 A complete normed linear space is called a Banach space
1.2.1 Incomplete normed linear spaces

Some important normed linear spaces are not complete. We give examples.
Example 1.17 Let X = C[a, b] be the space of all continuous real–valued functions
defined on the closed and bounded interval [a, b] with norm, k · k, given by
Z b
kf k = |f (t)|dt, t ε [a, b], f ε C[a, b].
a
We show that the space X endowed with this norm is NOT complete.
Solution.
It suffices to produce a Cauchy sequence in C[a, b] which does not converge to an
12
element of C[a, b]. Without loss of generality we may take [a, b] = [−1, 1]. Consider
the function fn given by


 0, if − 1 ≤ t ≤ 0;



fn (t) = nt, if 0 ≤ t ≤ n1 ;





1, if n1 ≤ t ≤ 1,
and sketched in Fig. 1.2. Observe that the expression for fn (t) is trivially obtained
as follows. The function is 0 for all values of t between −1 and 0. Then, for values of
t between 0 and n1 , we use the equation of a straight line studied in high school with
the coordinates of O as (0, 0) and the coordinates of P as ( n1 , 1). Thus the slope of
OA is n so that the equation of the line OA is y = nt. For values of t between n1
and 1, the function is constant with value 1. Observe that no inequality is strict.
This guarantees (by Pasting Lemma) that fn is continuous.
Fig. 1.2
It is clear (convince yourself) that {fn }∞

n=1 is a sequence of continuous real-valued
functions defined on [−1, 1]. We show that {fn }∞ n=1 is Cauchy. Let m > n so that
1 1
m
< n
(see Fig. 1.3).
We need to show kfn − fm k / 0 as n, m / ∞. But,
Z 1
kfn − fm k = |fn (t) − fm (t)|dt.
−1
13
Fig. 1.3
But this integral simply represents the area between fn and fm and from Fig. 1.3,
we have
Z 1
1 1 1
||fn − fm || = |fn (t) − fm (t)|dt = area of T riangle OCD = ( − ) → 0,
−1 2 n m
Z 1
as n, m → ∞. Hence {fn } is Cauchy. We can also evaluate |fn (t) − fm (t)|dt
−1
directly as follows:
Z 1
kfn − fm k = |fn (t) − fm (t)|dt
−1
Z 0 Z 1/m Z 1/n Z 1
= |0 − 0|dt + (mt − nt)dt + |1 − nt|dt + |1 − 1|dt
−1 0 1/m 1/n
Z 1/m Z 1/n
= (mt − nt)dt + (1 − nt)dt → 0 as m, n → ∞,
0 1/m
so that kfn − fm k / 0 as n, m / ∞ and so {fn }∞ is Cauchy.

n=1
We next show that {fn }∞ ∞

n=1 converges to an element not in C[−1, 1]. If {fn }n=1
is convergent, in order to find a candidate for its limit we examine fn (t) defined
above. If n / ∞, the interval 0 ≤ t ≤ 1 reduces to t = 0 and the interval
n
1
n
< t ≤ 1 becomes 0 < t ≤ 1. We, therefore, consider a function f : [−1, 1]
14
/ [0, 1] defined by

 0, if − 1 ≤ t ≤ 0;
f (t) =

1, if 0 < t ≤ 1.
Fig. 1.4
Clearly,
Z 1 µ ¶
1 1 /0 / ∞,
kfn −f k = |fn (t)−f (t)|dt = area of Triangle OAB = as n
−1 2 n
(see Fig. 1.4). Thus, the sequence {fn }∞ n=1 defined above is a Cauchy sequence in
C[−1, 1] which converges to f . However, f is not in C[−1, 1] since f is not continuous
at 0 (Verify this!). Thus, C[−1, 1] is not complete.
Remark 1.18 The interval [−1, 1] which is employed in the above example has
been used by several authors in showing that the space C[a, b] is not complete when
endowed with the integral norm. This interval is certainly not crucial. Other intervals
can easily be used.
µZ 3 ¶1/2
2
Example 1.19 Let X = C[−3, 3] with kf k2 = |f (t)| dt . Then X is not
−3
complete.
15
Solution. It is easy to see that
Z µZ ¶1/2
3 √ 3
2
|f (t)|dt ≤ 6 |f (t)| dt .
−3 −3
(Hint: Use the Cauchy-Schwartz inequality given in Exercises 1.7, Problem 7). It fol-
lows from this inequality and the sequence {fn } in Example 1.17 that X = C[−3, 3] is
not complete. Explain how this follows.
Remark 1.20 Incomplete spaces are not at all worthless. In fact every metric space
(complete or not) can be regarded as a dense subset of some complete metric space.
This complete metric space is called the completion of the incomplete space. Given
a metric space (M, ρ) (not necessarily complete) its completion can be constructed.
This process, however, requires the notions of linear maps and isometries which we
have so far not treated. We have therefore treated the question of completion of a
metric space as an Appendix at the end of this book.
1.3 The Contraction Mapping Principle

We end our basic introduction to normed linear spaces with the following classi-
cal and very important theorem-The Contraction Mapping Principle- which is con-
cerned with fixed points of contraction mappings defined on complete metric spaces.
Let (X, ρ) be a metric space and let f : X / X be any map. A point x∗ ∈ X is

called a fixed point of f if f (x∗ ) = x∗ . The map f is called a strict contraction (or
simply a contraction) if there exists a constant k ∈ [0, 1) such that
ρ(f (x), f (y)) ≤ kρ(x, y) for all x, y ∈ X.
With these definitions, we now state the following theorem whose proof the reader
is assumed to be familiar with (see Exercises 1.2, Problem 8).
Theorem 1.21 (Contraction Mapping Principle) . Let (M, ρ) be a complete metric

space, f : M / M a contraction mapping with constant k ∈ [0, 1). Then f has
a unique fixed point in M , i.e., there exists a unique x∗ ∈ M such that f (x∗ ) = x∗ .
Furthermore, for arbitrary x0 ∈ M , the sequence {xn } defined by xn = f (xn−1 ), n =
1, 2, ... converges strongly to x∗ .
EXERCISES 1.2
16
1. Let X = C[−2, 2] be the space of real–valued continuous functions on the
closed and bounded interval [−2, 2], and let X be endowed with the norms
Z 2
(i) kf k1 = |f (t)|dt, and
−2
µZ 2 ¶1/2
2
(ii) kf k2 = |f (t)| dt ,
−2
for arbitrary f in X. Justify the following statements:
(i) (X, k · k1 ) is not complete;
(ii) (X, k · k2 ) is not complete.
2. Let X = C[0, 4], with the usual notation and suppose X is endowed with the
norm,
µZ 4 ¶1/2
2
kf k2 = |f (t)| dt .
0
Consider the sequence {gn }∞

n=1 defined by the following graph (Fig. 1.5).
Fig. 1.5
(a) Write down the explicit expression for gn (t), 0 ≤ t ≤ 4.

(b) Verify that
(i) gn ∈ C[0, 4] for each n;
17
(ii) {gn }∞
n=1 is a Cauchy sequence;
(iii) gn / g as n / ∞, where

 0, 0 ≤ t ≤ 2;
g(t) =

1, 2 < t ≤ 4.
Conclude that C[0, 4] with k · k2 norm is not a complete space. Give another
R4 1
proof (similar to Example 1.19) that C[0, 4] with ||f ||2 := ( 0 (f (t))2 dt) 2 is
not complete.
3. Let X = C[0, 5] be the space of real–valued continuous functions on the closed

and bounded interval [0, 5]. Let X be endowed with norms,
Z 5
kf k1 = |f (t)|dt,
0
µZ 5 ¶1/2
2
kf k2 = |f (t)| dt ,
0
for arbitrary f in X. Show that:

(i) (X, k · k1 ) is not complete; and (ii) (X, k · k2 ) is not complete. (Hint: You
may use Fig. 1.6).
Fig. 1.6
18
4. Let S be the set of real sequences having only a finite
¡ 1 number of nonzero ¢ terms.
1 1
Clearly, S ⊆ l∞ . Take {xn }∞n=1 in S where x n
= 1, ,
2 3
, . . . , n
, 0, 0, . . . . Prove
that:
¡ ¢
(i) {xn } is a Cauchy sequence in S; (ii) xn / x where x := 1, 1 , 1 , . . . .
2 3
Conclude that S is not complete.
*5 Prove the following statements.

³P ´1
n 2 2
(a) IRn , with ||x||IRn := i=1 |x i | is complete.
(b) lp (1 ≤ p < ∞) is complete.
(c) l∞ is complete.
(d) c and c0 are complete.
6. Let X be a normed linear space and let f : X → X be any map. The map f
is called ıLipschitz if there exists k ≥ 0 such that ||tx − T y||
leqk||x − y|| for all x, y ∈ X. The map T is called nonexpansive if k ≤ 1.
Now, let X = c0 (the space of sequences of real numbers which converge to 0).
Define T : c0 → c0 by
T (x1 , x2 , x3 , ...) = (1, x1 , x2 , x3 , ...)
for arbitrary x = (x1 , x2 , x3 , ..) ∈ c0 . Prove that

(i) T is nonexpansive;
(ii) T has no fixed points in X, i.e., there are no points u ∈ X
such that T u = u.
7. Let X = l∞ (the space of sequences of real numbers which are bounded). Let
K = {x ∈ l∞ : ||x||∞ ≤ 1}. Define T : K → K by
T x = (0, x21 , x22 , x23 , ...)
for arbitrary x = (x1 , x2 , x3 , ...) ∈ K. Prove:

(i) T p = p if and only if p = 0;
(ii) ||T x − T p|| ≤ ||x − p||, where p = 0;
(iii) T is not nonexpansive. (Hint: Consider x = ( 43 , 34 , ...), y = ( 21 , 12 , ...)).
8. Prove Theorem 1.20. (Hint: For arbitrary x0 ∈ M , define xn+1 = f (xn ), for
all integers n ≥ 0. Prove {xn } is Cauchy and use the continuity of f to obtain
the existence of an x∗ ∈ M such that f (x∗ ) = x∗ . Establish uniqueness by
contradiction).
19
9. Prove the inequality
Z µZ ¶1/2
3 √ 3
2
|f (t)|dt ≤ 6 |f (t)| dt
−3 −3
stated in Example 1.19.
1.4 Proof of Hölder’s Inequality
Theorem 1.22 (Hölder’s Inequality) Let p ≥ 1. For arbitrary x = {xi }, y = {yi },

x, ∈ `p , y ∈ `q we have,
∞
Ã∞ !1/p Ã ∞ !1/q
X X X
|xi yi | ≤ |xi |p |yi |q , (i)
i=1 i=1 i=1
1 1
where p
+ q
= 1.
Proof. We first prove that if p ≥ 1 and p1 + 1q = 1, then for arbitrary constants

a > 0, b > 0 we have
ap bq
ab ≤ + . (ii)
p q
So, let a > 0, b > 0 be constants. Consider the function y = xp−1 (or, equivalently
1
x = y q−1 , since p−1 = q − 1). This function is sketched in Fig. 1.7 and it is easy to
see that ab is the area of the rectangle in Fig. 1.7.
y = xp−1 y = xp−1
b
b A2
A2
A1 A1
a a
Fig. 1.7
20
By integration we obtain
Z a Z b
p−1 a p bq
ab ≤ x dx + y q−1 dy = A1 + A2 = + .
0 0 p q
Observe that if a = 0 or b = 0, this inequality still holds. Let {ui }, {vi } be such
that ∞ ∞
X X
p
|ui | = 1, |vi |q = 1. (iii)
i=1 i=1
Define a = |ui |, b = |vi | and substitute in inequality (ii) to get
|ui |p |vi |q
|ui vi | ≤ + .
p q
Summing this inequality and using (iii) we obtain,
∞
X ∞ ∞
1 X 1 X 1 1
|ui vi | ≤ |ui |p + |vi |q = + = 1,
i=1
p i=1 q i=1 p q
i.e.,
∞
X
|ui vi | ≤ 1. (iv)
i=1
Now, let x := {xi }, y := {yi }, x ∈ `p , y ∈ lq , x 6= 0, y 6= 0.
Define
xi yi
ui := µ ¶ 1/p
, v i := µ ¶1/q .
P
∞ P
∞
|xi |p |yi |q
i=1 i=1
P
∞ P
∞ P∞
Clearly, |ui |p = 1, |vi |q = 1, so that inequality (iv) yields: i=1 |ui vi | ≤ 1
i=1 i=1
i.e.,
∞
Ã∞ !1/p Ã ∞ !1/q
X X X
|xi yi | ≤ |xi |p |yi |q .
i=1 i=1 i=1
Observe that if either x or y is zero this inequality still holds. This completes the
proof of Hölder’s inequality.
¤
Theorem 1.23 (Cauchy-Schwartz inequality) . For arbitrary x = {xi }∞ i=1 , y =
∞
{yi }i=1 , x, y ∈ l2 , we have,
X∞ ³X
∞ ´ 1 ³X
∞ ´1
2 2 2 2
|xi yi | ≤ |xi | |yi | .
i=1 i=1 i=1
21
Proof. Set p = 2 in Theorem 1.20 and the result follows.
¤
Remark 1.24 It is clear that Schwartz’s inequality is a special case of Hölder’s

inequality in which p = 2. Of course, Schwartz’s inequality is also valid for finite
sums.
22
CHAPTER 2
Linear Maps
Linear maps play a crucial role in Functional Analysis. In this chapter we shall
define linear maps and present some basic results concerning them.
2.1 Definition and some examples

Definition 2.1 Let X and Y be linear spaces over a scalar field K. A mapping
T : X → Y is said to be a linear map if
T (αx + βy) = αT (x) + βT (y), (2.1)
for arbitrary x, y ∈ X and arbitrary scalars α, β ∈ K. Some authors use the term
linear operator or linear transformation instead of linear map. Condition (2.1) is
equivalent to the following two conditions: (i) T (x + y) = T x + T y ∀ x, y ∈ X;
and (ii) T (αx) = αT (x) ∀ x ∈ X, and for each scalar, α.
Definition 2.2 If in Definition 2.1 the linear space Y is replaced by a scalar field
K then the linear map T is called a linear functional on X.
Example 2.3 Differentiation, integration and limits (of convergent sequences) are
examples of linear maps.
23
Example 2.4 Let X = `2 . For each x̄ = (x1 , x2 , x3 , ...) in `2 define
³ x2 x3 ´
T x̄ = 0, x1 , , , ... .
2 3
Then T is a linear map on `2 .
Verification. Let x̄ = (x1 , x2 , ...) and ȳ = (y1 , y2 , ...) be arbitrary elements of `2

and let α, β be scalars. Then
αx̄ + β ȳ = (αx1 , αx2 , αx3 , ...) + (βy1 , βy2 , βy3 , ...)

= (αx1 + βy1 , αx2 + βy2 , αx3 + βy3 , ...)
and µ ¶
αx1 + βy1 αx2 + βy2 αx3 + βy3
T (αx̄ + β ȳ) = 0, , , , ...
1 2 µ 3 ¶
³ αx2 αx3 ´ βy2 βy3
= 0, αx1 , , , ... + 0, βy1 , , , ...
³ 2 3 ´ ³ 2 3´
x2 x3 y2 y3
= α 0, x1 , , , ... + β 0, y1 , , , ...
2 3 2 3
= αT (x̄) + βT (ȳ),
and so T is a linear map.
Example 2.5 Let CI denote the linear space of complex numbers over CI , and define
the map T : CI /C
I by T z = z̄, where z̄ denotes the conjugate of z. Then T is not
a linear map. (Justify this).
2.2 A basic result concerning linear maps

We remark first that since linear functionals are special forms of linear maps, any
result proved for linear maps also holds for linear functionals.
Proposition 2.6 Let X and Y be two linear spaces over a scalar field, K, and let
T : X → Y be a linear map. Then
(i) T (0) = 0;
(ii) The range of T , R(T ) = {y ∈ Y : T x = y for some x ∈ X} is a linear subspace
of Y ;
(iii) T is one-to-one if and only if T (x) = 0 implies x = 0;
(iv) If T is one-to-one, then T −1 exists on R(T ) and T −1 : R(T ) → X is also a
linear map.
24
Proof. (i) Since T is linear, we have, T (αx) = αT (x) for each x ∈ X and each
scalar α. Take α = 0 and (i) follows immediately.
(ii) We need to show that for y1 , y2 ∈ R(T ) and α, β scalars, αy1 +βy2 ∈ R(T ). Now,
y1 , y2 ∈ R(T ) implies that there exist x1 , x2 ∈ X such that T (x1 ) = y1 , T (x2 ) = y2 .
Moreover, αx1 + βx2 ∈ X (since X is a linear space). Furthermore, by the linearity
of T ,
T (αx1 + βx2 ) = αT (x1 ) + βT (x2 ) = αy1 + βy2 .
Hence αy1 + βy2 ∈ R(T ), and so R(T ) is a linear subspace of Y .
(iii) (⇒) Assume T is one-to-one. Clearly T x = 0 ⇒ T (x) = T (0) since T is

linear (and so T (0) = 0). But T is one-to-one. So, x = 0.
(⇐) Assume that whenever T u = 0 then u must be 0. We want to prove that

T is one-to-one. So, let T x = T y. Then, T x − T y = 0 and by the linearity of T ,
T (x − y) = 0. By hypothesis, x − y = 0 which implies x = y. Hence T is one-to-one.
(iv) Left as an easy exercise for the reader (see Exercises 2.1, Problem 1).
Example 2.7 Let T : IR → IR be defined by T x = ax + b; a, b ∈ IR, b 6= 0. Then T

is not a linear map. It suffices to observe that T (0) = b 6= 0.
2.3 Bounded Linear Maps
Definition 2.8 Let X and Y be normed linear spaces over a scalar field K, and
let T : X → Y be a linear map. Then T is said to be bounded if there exists some
constant k ≥ 0 such that for each x ∈ X,
||T (x)|| ≤ K||x||,
the constant K is called a bound for T and in this case, T is called a bounded linear
map.
We now turn our attention to linear maps that are continuous. The notion of
continuity can be stated, for linear maps in several useful equivalent forms. We
state these equivalent forms in the following theorem.
Theorem 2.9 Let X and Y be normed linear spaces and let T : X → Y be a linear
map. Then the following statements are equivalent:
(i) T is continuous;
25
(ii) T is continuous at the origin (in the sense that if {xn } is a sequence in X such
that xn → 0 as n → ∞, then T xn → 0 in Y as n → ∞);
(iii) T is Lipschitz, i.e., there exists a constant K ≥ 0 such that, for each x ∈ X,
||T x|| ≤ K||x||;
(iv) If D := {x ∈ X : ||x|| ≤ 1} is the closed unit disc in X, then T (D) is bounded

(in the sense that there exists a constant M ≥ 0 such that ||T x|| ≤ M for all x ∈ D).
Proof. (i) ⇒ (ii). Let T : X → Y be linear and continuous. We want to prove

that T is continuous at 0. So, let {xn } be a sequence in X such that xn → 0. By
continuity of T , we have, T xn → T (0). But T is a linear map so that T (0) = 0.
Hence, the result follows.
(ii) ⇒ (iii) We are given that if {xn } is any sequence in X such that xn → 0
as n → ∞, then T (xn ) → 0. We want to prove that there exists a constant K ≥ 0
such that ||T x|| ≤ K||x|| for each x ∈ X. Suppose for contradiction that there is
no such K. Then, for any positive integer n, there exists some x = x(n) ∈ X, x(n)
different from 0, call it xn (since it depends on n) such that
||T xn || > n||xn ||.
This implies that

||T xn ||
> 1.
n||xn ||
xn
Now, define the sequence un := n||xn ||
. Clearly, un → 0. For,
° x ° 1 ||x || 1
° n ° n
||un − 0|| = ° °= = → 0.
n||xn || n ||xn || n
However, T un 6→ 0. For,
° T (x ) ° ||T (x )||
° n ° n
||T un − 0|| = ° °= > 1,
n||xn || n||xn ||
and so T un 6→ 0 (even though un → 0). This contradicts the hypothesis that T is

continuous at the origin. Thus our supposition is false and so , (ii) ⇒ (iii).
(iii) ⇒ (iv) Given that a linear map T is Lipschitz we want to prove that T (D)
is bounded, (i.e., ||T (x)|| ≤ M for all x ∈ D, and some constant M ≥ 0). Take
26
y ∈ D arbitrary. Then, ||y|| ≤ 1. By (iii), ||T (y)|| ≤ K||y|| for some constant
K ≥ 0. But ||y|| ≤ 1. So,
||T y|| ≤ K||y|| ≤ K.
Since y was arbitrarily chosen in D, it follows that for all y in D, ||T y|| ≤ K. Now,
take K = M and we are done.
(iv) ⇒ (i) Let x1 , x2 be arbitrary elements of X. Assume first that x1 − x2 = 6 0.

Consider the vector u := ||xx11 −x
−x2
2 ||
. Clearly, u ∈ D, and so, by condition (iv), (i.e.,
T (D) is bounded), there exists a constant K ≥ 0 such that ||T (u)|| ≤ K i.e.,
x1 − x2
||T ( )|| ≤ K, or, ||T x1 − T x2 || ≤ K||x1 − x2 ||.
||x1 − x2 ||
If, on the other hand, x1 − x2 = 0, this inequality clearly also holds. Thus, the
²
inequality holds for all x1 , x2 ∈ X. Now, given any ² > 0, choose δ = K+1 . So, if
||x1 − x2 || < δ we obtain,
µ ¶
K
||T x1 − T x2 || ≤ K||x1 − x2 || ≤ Kδ = ² < ²,
K +1
and hence T is (uniformly) continuous on X.

¤
Property (iii) of the last theorem is a very important one. In fact, we have the
following definition:
Remark 2.10 In the light of definition 2.8, (and Theorem 2.9) we have that a linear
map T : X → Y is continuous if and only if it is bounded. Thus, for linear maps,
continuity and boundedness (as given in definition 2.8) are equivalent.
Remark 2.11 If T : X → Y is a bounded linear map, then inequality (??) is

satisfied for some constant K ≥ 0 and for all x ∈ X. It is clear from this inequality
that the constant K is not unique. For, if, for example, ||T x|| ≤ 3||x||, then,
||T x|| ≤ 3||x|| ≤ 4||x|| ≤ 5||x|| ≤ .... However, in general, K has a lower bound
(which is 0). Thus,
inf{K : ||T x|| ≤ K||x||},
for each x ∈ X, exists. This leads to the following definition:
27
Definition 2.12 If T : X → Y is a bounded linear map from a normed linear space
X into a normed linear space Y , we define the norm of T by
||T || := inf{K : ||T x|| ≤ K||x||}, for each x ∈ X.
Note that from definition 2.12 we obtain immediately that ||T x|| ≤ ||T ||||x|| for
each x ∈ X and that for every ² > 0, there exists x² ∈ X, x² 6= 0, such that
||T x² || > (||T || − ²)||x² ||.
Example 2.13 Let X and Y be normed linear spaces and B(X, Y ) denote the
family of all bounded linear maps from X to Y . Define,
(T + L)(x) = T (x) + L(x);
(αT )(x) = αT (x),

for all T, L ∈ B(X, Y ) and scalar α. Then, clearly B(X, Y ) is a vector space.
We now prove the following theorem.
Theorem 2.14 Let B(X, Y ) be the family of all bounded linear maps from X to Y .
Then we have the following: For arbitrary T ∈ B(X, Y ),
kT xk
kT k = sup kT xk = sup kT xk = sup .
kxk≤1 kxk=1 x6=0 kxk
Proof. Since T is bounded and linear, ∃ K ≥ 0 such that for all x ∈ X, kT xk ≤

Kkxk. If kxk ≤ 1 then kT xk ≤ Kkxk ≤ K and so, {kT xk : kxk ≤ 1} is a
bounded set in IR so its “sup” exists and sup kT xk ≤ K for any K such that
kxk≤1
kT xk ≤ Kkxk ∀ x ∈ X. Taking the infimum over all such K 0 s we obtain that
sup kT xk ≤ inf{K ≥ 0 : kT xk ≤ Kkxk ∀ x ∈ K} := kT k.

kxk≤1
Hence,
sup kT xk ≤ kT k. (2.2)
kxk≤1
Conversely, from Definition 2.12, we get that for every ² > 0, ∃ x² in X, xε 6= 0,

such that
kT x² k > (kT k − ²)kx² k.
28
(Otherwise kT x² k ≤ (kT k − ²)kx² k and kT k would no longer be the infimum). Let
u² = kxx∈² k then ku² k = 1 and kT u² k > kT k − ².
Consequently, we obtain from inequality (2.2) that
³ x ´
²
kT k ≥ sup kT xk ≥ sup kT xk ≥ sup kT k > kT k − ².
kxk≤1 kxk=1 kx² k6=0 kxε k
Since ² > 0 is arbitrary, we get that
||T x||
kT k ≥ sup kT xk ≥ sup kT xk ≥ sup ≥ kT k
kxk≤1 kxk=1 kxk6=0 ||x||
so that all these quantities are equal. The proof is complete. ¤
Remark 2.15 Since ||T || is the smallest M such that ||T x|| ≤ M ||x|| for each
x ∈ X, it follows that whatever value M has, we must always have ||T || ≤ M .
Thus, whenever T is a linear map and we have the relation
||T x|| ≤ (∗)||x||, (2.3)
for each x ∈ X, then ||T || must be less than or equal to (∗).
Example 2.16 Let T : l2 → l2 be defined by

x2 x3
T (x1 , x2 , x3 , ..., ) = (0, x1 , , , ...).
2 3
Then ||T x|| ≤ ||x|| so that ||T || ≤ 1. In this example, we actually have ||T || = 1
since T (1, 0, 0, ...) = (0, 1, 0, ...).
Proposition 2.17 Let X and Y be normed linear spaces. Then, ||.|| defined by
||T || = sup ||T x|| for each T in B(X, Y ) is a norm on B(X, Y ).
||x||≤1
(Hence, B(X, Y ) is a normed linear space with this norm).
Proof. Exercise (See Exercises 2.1, Problem 1).
Proposition 2.18 Let X and Y be normed linear spaces. Then, B(X, Y ) is a

Banach space if Y is.
29
Proposition 2.19 Let X, Y, Z be normed linear spaces and let P ∈ B(Y, Z), Q ∈
B(X, Y ). Define (P Q)(x) = P (Qx). Then
(a) P Q ∈ B(X, Z), and
(b) kP Qk ≤ kP k kQk.
Remark 2.20 If Y = IR (the reals) then B(X, Y ) becomes B(X, IR) and is denoted
by X ∗ and is called the conjugate space or the dual space of X. Thus, f ∈ X ∗ means
that f is a mapping of X into IR, which is linear and bounded. Recall that any map
from a linear space into a scalar field is called a functional. The members of X ∗ are,
therefore, bounded linear functionals. For f ∈ X ∗ we shall define
||f || := sup |f (x)|.

||x||≤1
1 1
We now show that for p > 1, the dual space of lp is lq , where q
+ p
= 1.
1 1
Proposition 2.21 For 1 < p < ∞, `∗p = `q where p
+ q
= 1, (i.e., the dual of `p is
`q , where p1 + 1q = 1, 1 < p < ∞).
Proof. The proof is divided into 2 parts.

Part 1. We prove that every element y in `q defines an element fy in `∗p with
||y|| = ||fy ||.
Part 2. We prove that every element h ∈ `∗p defines an element zh in `q with
||h|| = ||zh ||. Hence, lp∗ and lq are isometric.
We begin with Part 1. Let y = {yi }∞ ∞

i=1 ∈ `q . For every x = {xi }i=1 in `p ,
define fy : `p / IR by
∞
X
fy (x) = xi y i .
i=1
Clearly, fy is well defined. For,
∞
Ã∞ !1/p Ã ∞ !1/q
X X X
|fy (x)| ≤ |xi yi | ≤ |xi |p |yi |q = kxkkyk < ∞.
i=1 i=1 i=1
This inequality also implies fy is bounded and (by remark 2.15),
kfy k ≤ kyk.
30
Linearity of fy is obvious. Hence, fy is a bounded linear functional on `p and so
fy ∈ `∗p .
For Part 2. Let h ∈ `∗p . Let {ei }∞

i=1 be the canonical basis in `p , i.e.,
ei = (0, 0, . . . , 0, 1, 0, . . .).
↑ith position
Then, for every x = {xi }∞

i=1 in `p we have that
∞
X n
X
x= xi ei = lim / xi ei .
n ∞
i=1 i=1
Using the linearity and continuity (boundedness) of h we have that

Ã n
! n ∞
X X X
h(x) = h lim / xi ei = lim / xi h(ei ) = xi h(ei ).
n ∞ n ∞
i=1 i=1 i=1
Define,
zh := {h(ei )}∞ ∞
i=1 = {αi }i=1 .
It suffices now to prove that zh ∈ `q . Let xn ∈ `p be defined as follows:

n o
xn := |α1 |q−1 sign α1 , |α2 |q−1 sign α2 , . . . , |αn |q−1 sign αn , 0, 0, . . . .
Then,
n
X n
X
q−1
|h(xn )| = |αi | (sign αi )αi = |αi |q ≤ khk · kxn k
i=1 i=1
( n
)1/p ( n
)1/p
X X
= khk · |αi |(q−1)p = khk · |αi |q .
i=1 i=1
½ ¾1− p1
P
n
q
P
∞
Hence, |αi | ≤ khk ∀ n = 1, 2, 3, . . . . Hence |αi |q < ∞, i.e.,
i=1 i=1
zh := {αi }∞
i=1 ∈ `q and ||zh || ≤ ||h||.
Observe that since h ∈ lp∗ is arbitrary, if we now take h = fy (the functional induced
by y) then zh = y and we get that
∞
X 1
||y|| = ||zh || = ( |αi |q ) q ≤ ||h|| = ||fy ||.
i=1
31
Hence, ||y|| ≤ ||fy ||. But we already have ||fy || ≤ ||y||, so that kfy k = kyk. This
proves that `∗p is isometrically isomorphic to `q . This proof is complete.
¤
2.4 Equivalent norms and finite dimensional spaces
Definition 2.22 Suppose E is a vector space over IR and suppose k·k1 and k·k2 are
two norms on E. Then k·k1 and k·k2 are said to be equivalent (written k·k1 ∼ k·k2 )
if there exist numbers α > 0, β > 0 such that
αkxk1 ≤ kxk2 ≤ βkxk1 ∀ x ∈ E, (2.4)
or equivalently,
αkxk2 ≤ kxk1 ≤ βkxk2 ∀ x∈E .
The following facts are easily verified:
F1 The relation ∼ is an equivalence relation on the set of all norms on E.
F2 If k · k1 and k · k2 are equivalent norms, then a sequence {xn } which is Cauchy

with respect to k · k1 is also Cauchy with respect to k · k2 , and vice versa.
F3 If k · k1 and k · k2 are equivalent norms, the class of open sets defined by k · k1 is

the same as the class of open sets defined by k · k2 , and vice versa. To prove
this it suffices to prove the following two statements:
1. Given an arbitrary open ball induced by k · k1 , Bε1 (x0 ), say, ∃ an open ball
Bδ2 (x0 ), say, induced by k · k2 such that Bδ2 (x0 ) ⊂ Bε1 (x0 ).
2. Given an arbitrary open ball induced by k · k2 , Bt2 (x0 ) say, there exists an
open ball, Br1 (x0 ) say, such that
Bt1 (x0 ) ⊂ Br2 (x0 ).
We now prove the following important theorem which is the main theorem of this
section.
Theorem 2.23 Let E be a finite dimensional normed linear space. Then all norms
on E are equivalent.
32
Proof. Let dim E = n < ∞ and let {ei }ni=1 be a basis for E. For arbitrary
P
n
vector x ∈ E, there exist unique scalars {αi }ni=1 such that x = αi ei . Define
i=1
kxk0 = max |αi |. Clearly, k · k0 is a norm on E (Verify). It suffices now to prove
1≤i≤n
that any norm on E is equivalent to k · k0 ; i.e., that if k · k is an arbitrary norm on
E, there exist constants a > 0, b > 0 such that
akxk0 ≤ kxk ≤ bkxk0 ∀ x ∈ E.
P
n °P ° P
° n ° n
From x = αi ei we obtain, kxk = ° αi ei ° ≤ max |αi | kei k. Since {ei }ni=1
i=1 i=1 1≤i≤n i=1
P
n
is a basis, it follows that kei k is a number. Call it b. The last inequality then
i=1
yields
kxk ≤ bkxk0 .
It now only remains to prove akxk0 ≤ kxk. Let S = {x ∈ E : kxk0 = 1}. Define the
map
ϕ : (E, k · k0 ) / IR by ϕ(x) = kxk ∀x ∈ E .
° °
° x °
Observe that ° kxk0
° = 1. Moreover, ϕ is continuous. To see this, let ² > 0 be
0
given. We want to find a δ > 0 such that for arbitrary x∗ ∈ E, if kx − x∗ k0 < δ then
|ϕ(x) − ϕ(x∗ )| < ². Recall that in part (a) we proved that kxk ≤ bkxk0
∀x ∈ E. Using this we now obtain that
|ϕ(x) − ϕ(x∗ )| = |kxk − kx∗ k| ≤ kx − x∗ k ≤ bkx − x∗ k0 ≤ bδ < ² ,
²
where we have chosen δ = 1+b . So, ϕ is continuous. Since S is compact, ϕ attains its
infimum on S. Let this infimum ° be°denoted by a. Then 0 < a < ϕ(x) := ||x||, ∀x ∈
x ° x °
S. But ||x||0 ∈ S. So, 0 < a ≤ ° kxk0
° ∀ x ∈ E and this implies akxk0 ≤ kxk ∀x ∈ E.
Combining this with kxk ≤ bkx0 k, we obtain that akxk0 ≤ kxk ≤ bkxk0 ∀x ∈ E
as required. ¤
EXERCISES 2.1
1. Prove part (iv) of Proposition 2.6; and also prove Theorem 2.17.
2. Let X P
be a finite dimensional linear space and let x1 , x2 , x3 , ..., xn be a basis.
If x = ni=1 αi xi , define
³X n ¯2 ´ 1
¯ 2
||x|| = |αi ¯ .
i=1
33
If f is a bounded linear functional on X, compute P ||f ||. What will ||f || be if
||x|| = maxi |αi |? What will ||f || be if ||x|| = ni=1 |αi |?
3. Prove that if X is a finite-dimensional space, then all linear functionals are

bounded.
(Hint: You may use the fact that all norms on a finite dimensional space are
equivalent).
4. On the space X = l1 , for x = (α1 , α2 , ...) ∈ l1 , define

∞
X
f (x) = αn ,
n=1
and introduce a new norm ||x||∞ := P∞sup|αn |. Prove that:

(a) f is continuous with ||x||l1 := i=1 |αn |;
(b) ||f || = 1;
(c) f is not continuous with respect to ||x||∞ := sup{αn }. (This shows that
not all linear maps are bounded. Compare part (c) with Problem 3 above
where X is finite dimensional).
5. Let f be a bounded linear functional defined on C[a, b] (with “sup” norm) by

Z b
fx = x(t)dt,
a
for x ∈ C[a, b]. Compute ||f ||.
6. Let f be a continuous linear functional defined on a normed linear space, X.

Prove that, if a sequence {xn } is Cauchy in X, then {f (xn )} is a Cauchy
sequence of complex numbers.
7. Let X and Y be normed linear spaces and T : X → Y be a bounded linear

map. Prove that Ker (T) is a closed subspace of X. (Thus, the kernel or null
space of any continuous linear map is closed).
8. Prove Proposition 2.18.
10. Verify the facts F1, F2 and F3 following Definition 2.22.
11. Prove that any finite dimensional normed linear space is complete.
34
12. Prove that if E is a normed linear space and M is any finite dimensional
subspace, then M is closed. (Hint: A complete subset of a metric space is
closed).
35
36
CHAPTER 3
The Hahn Banach Theorem: Analytic and Geometric Forms.
3.1 Introduction
In this chapter we shall study one of the most important theorems in Linear Func-
tional Analysis – the Hahn Banach Theorem, and shall also study some of its many
striking consequences. From now on, all linear spaces are assumed to be real. We
start with the following definition.
Definition 3.1 Let X be a linear space (not necessarily normed), and consider the
mapping
p : X → IR
with the following properties: (i) p(x) ≥ 0 for all x ∈ X; (ii) p(x + y) ≤ p(x) + p(y)
for all x, y ∈ X; (iii) p(αx) = αp(x) for all x ∈ X and scalar α, α > 0.
A mapping satisfying all the conditions (i)-(iii) is called a convex functional. A map-
ping satisfying condition (ii) is called subadditive and is called positive homogeneous
if it satisfies (iii). If p satisfies (ii) and (iii), then p is called a sublinear functional.
Example 3.2 Every norm is a convex functional.
Definition 3.3 Let M be a proper subset of a real vector space, X. Let f be a map
defined on the subset M into a vector space, W , (see Fig. 3.1).
37
X EE
EE
EE
EE
S EE
EEF
EE
EE
EE
EE
EE
"
M f
/W
Fig.3.1
The mapping f (defined on M ) is said to be extended to X if there exists a map
F : X → W (defined on X into W ) such that F (x) = f (x) for all x ∈ M .
The map F is called an extension of f from M to X.

Example 3.4 Let M = (−∞, 0) ∪ (0, ∞) and X = IR. Observe that 0 6∈ M so that
M is a proper subset of X. Define
f : M → IR
by ½
0, if x ∈ (−∞, 0);
f (x) =
1, x ∈ (0, ∞).
To extend f to X we simply need to define a function F on X which agrees with f
on M . In particular, if we define F : IR → IR by
½
0, x ∈ (−∞, 0];
F (x) =
1, x ∈ (0, ∞);
then F is an extension of f to X. Observe that the extension is not continuous at
x = 0. Define two other extensions of f .
Example 3.5 Let M = (−∞, 0) ∪ (1, ∞) and X = IR. Clearly, M is a proper
subset of IR. Define f : M → IR by
½
0, x ∈ (−∞, 0);
f (x) =
1, x ∈ (1, ∞).
One extension of f to IR is easily defined by

 0, x ∈ (−∞, 0];
F (x) = x, x ∈ [0, 1];

1, x ∈ [1, ∞).
38
Remark 3.6 If f is a map defined on a subset M of a vector space V , to find an
extension of f to V , in general, is not particularly of interest. What is always of
interest is to establish the existence of an extension F which preserves some or all
of the properties of f . For example, let M = [0, 1) and let f : M / [0, 1] be
x
defined by f (x) = x+1 for each x ∈ M . Then, clearly f is continuous on M . If now
we define ½
f (x), if x ∈ M ;
F1 (x) =
1, if x = 1,
it is clear that F1 is an extension of f to [0, 1]. Observe that, F1 is not continuous
at x = 1 so that it does not preserve the continuity property of f . If, however, we
define ½
f (x), if x ∈ M ;
F2 (x) = 1
2
, if x = 1,
then F2 is an extension of f which preserves the continuity property of f .
The Hahn Banach Theorem, in its analytic form (there is also a geometric form
which we shall study later in this chapter) is concerned with the extension of a lin-
ear functional defined on a linear subspace to the whole space and preserving certain
properties of the functional. More precisely, we have the following theorem:
Theorem 3.7 (Hahn Banach Theorem, Analytic Form)
Let M be a subspace of a real linear space, V ; p : V → IR a sublinear functional on
V , and f : M → IR a linear functional defined on M such that
f (x) ≤ p(x)∀x ∈ M.
Then,
(i) there exists a linear functional F defined on V which is an extension of f ;
(ii) F (x) ≤ p(x) for all x ∈ V .
Theorem 3.7 is illustrated pictorially in Fig. 3.4.
p
V E
/ R
E
E
E
E
S E∃F, F (x)≤p(x) ∀x∈V
E
E
E
E
E
E
E"
M f
/R
f (x)≤p(x) ∀x∈M
39
Fig.3.2
Remark 3.8 Observe that in Theorem 3.7, the linear space V need not be normed.
Consequently, f and F are linear functionals (not necessarily bounded).
A proof of Theorem 3.7 is given at the end of this chapter, Section 3.4. We shall
now obtain, as a consequence of Theorem 3.7, a version of Hahn Banach Theorem
which holds in normed linear spaces X and for bounded linear functionals defined on
subspaces of X. Furthermore, in Section 3.3 we shall give the proofs of two geometric
forms of Hahn Banach Theorem. These geometric forms will be very useful in what
follows. Meanwhile, we consider some consequences of Theorem 3.7.
3.2 Some Consequences of Hahn Banach Theo-

rem
In this section we shall obtain some important direct consequences of the Hahn Ba-
nach Extension Theorem. We shall restrict our discussion to normed linear spaces
and consequently we shall confine our discussion of these applications of the theo-
rem to bounded linear maps defined on subsets (not necessarily subspaces) of a real
normed linear space. We begin with the following theorem:
Theorem 3.9 Let f be a bounded linear functional defined on a subspace M of a

real normed linear space, X. Then,
(i) there exists a bounded linear functional F defined on X, which extends f , and,
(ii) ||F || = ||f ||.
Proof. The technique of proof is to apply Theorem 3.7. So, we try to obtain the
hypotheses of Theorem 3.7 from those given in Theorem 3.9. If M = {0}, then
f = 0 and the required extension clearly is F = 0. So, we may assume M 6= {0},
i.e., we may assume that there exists at least one x ∈ M, x 6= 0. By hypothesis, f is
a bounded linear map on M . This implies
|f (x)| ≤ ||f ||||x||
for each x ∈ M . We now define p : X → IR by
p(x) = ||f ||||x||
for each x ∈ X.
Claim:
40
(i) p(x) ≥ 0 for all x ∈ X; (ii) p(x + y) ≤ p(x) + p(y) for all x, y ∈ X; (iii)
p(αx) = αp(x) for all x ∈ X, α ≥ 0.
The claim is easily verified and so, is left as an easy exercise. Thus, p is a convex
functional. In particular, p is a sublinear functional. Moreover, for every x ∈ M ,
f (x) ≤ |f (x)| ≤ p(x),
since |f (x)| ≤ ||f ||||x|| = p(x). Thus, by Theorem 3.7, there exists a linear functional
F (not necessarily bounded) defined on X which is an extension of f and such that
F (x) ≤ p(x) = ||f ||||x||
for each x ∈ X. Also, F (−x) ≤ kf k k − xk = kf k kxk = p(x). Thus −F (x) ≤ p(x)

and this implies,
F (x) ≥ −p(x).
Thus,
−p(x) ≤ F (x) ≤ p(x),
and this yields
|F (x)| ≤ p(x).
This inequality immediately implies (by Remark 2.15) that
(a) F is bounded;
(b)
||F || ≤ ||f ||. (3.1)
Furthermore, for x ∈ M , we have
|f (x)| = |F (x)| ≤ ||F ||||x||,
(since F is bounded) and this inequality implies (again, by Remark 2.15),
||f || ≤ ||F ||. (3.2)
From (3.1) and (3.2), we obtain ||F || = ||f ||, completing the proof.
¤
Theorem 3.9 is sometimes referred to as the Hahn Banach Theorem for normed
linear spaces. Our next result deals with another application of Theorem 3.7 to
a question of the existence of a bounded linear functional having certain special
properties.
41
Theorem 3.10 Let X be a normed linear space and let x0 6= 0 be an arbitrary
element of X. Then, there exists a bounded linear functional f on X such that
||f || = 1 and f (x0 ) = ||x0 ||.
Proof. Our technique of proof here is to apply Theorem 3.9. We start by con-
sidering the subspace M of X consisting of elements of the form x = αx0 , where
α is a scalar, i.e., M = {x : x = αx0 , α is a scalar}. Define f : M → IR by
f (x) ≡ f (αx0 ) = α||x0 ||, for each x ∈ M .
Claim:
(i) f is a linear functional;
(ii) f (x0 ) = ||x0 ||;
(iii) f is bounded;
(iv) ||f || = 1.
Verifications of (i) and (ii) are immediate, so we verify (iii) and (iv). Since x ∈ M ,
we have,
|f (x)| = |α|||x0 || = ||αx0 || = ||x||. (3.3)
It follows from (3.3) that f is bounded, establishing (iii). For (iv), observe that
equation (3.3) also implies,
||f || ≤ 1.
But if there were a real constant k < 1 such that |f (x)| ≤ k||x|| for all x ∈ M , then
equation (3.3) would be contradicted. For then, we would have
|f (x)| ≤ k||x|| < ||x||
for all x ∈ M . Thus, ||f || = 1. We have now established that f is a bounded

linear functional on M and ||f || = 1. By Theorem 3.9, there exists a bounded linear
functional F on X, extending f , with ||F || = ||f || = 1. Moreover, F (x0 ) = f (x0 ) =
||x0 ||. This completes the proof.
¤
Theorem 3.10 is an extremely useful result. We note two immediate applications of
it.
Corollary 3.11 Let X be a normed linear space. If X 6= {0}, then X ∗ = 6 {0}.

(in other words, if a normed linear space contains at least one nonzero vector then
the dual space contains at least one nontrivial element).
42
Proof. Exercise (Try a proof by contradiction, using Theorem 3.10).
Corollary 3.12 If f (x) = 0 for all f ∈ X ∗ , then x = 0.
(i.e., if all bounded linear functionals vanish on a given vector then the vector must
be the zero vector).
Proof. Exercise (Apply Theorem 3.10).
We give another application of Hahn Banach Theorem to the existence of a bounded

linear functional with yet another property.
Theorem 3.13 Let S be a subspace of a normed linear space, X. Let x0 ∈ X\S
be such that
δ := inf ||n − x0 || > 0.
n∈S
∗
Then, there exists f ∈ X such that
(i) f (x0 ) = δ;
(ii) ||f || = 1 and;
(iii) f (n) = 0 for all n ∈ S.
Proof. Consider the subspace spanned by S and x0 denoted by M i.e.,
M := [S ∪ {x0 }] = {m = x + αx0 : x ∈ S},
where α is a scalar. We remark immediately that since x0 6∈ S, the representation
m = x + αx0 for m ∈ M is unique. (Convince yourself by assuming m ∈ M could be
written in two different ways and obtaining that x0 ∈ S, a contradiction). Because
of the uniqueness of the representation, the functional f on M , given by
f (m) ≡ f (x + αx0 ) = αδ
is a well defined map. Furthermore,
(i) f is linear; (ii) f (x0 ) = δ; (iii) f (x) = 0 for all x ∈ S.
For arbitrary m ∈ M , let m = x + αx0 where x ∈ S and α is a scalar. Assume for

now that α > 0. Then,
³ x ´ ° x °
° °
||m|| = ||x + αx0 || = || − α − − x0 || = |α| ° − − x0 °.
α α
Since (− αx ) ∈ S and δ = inf ||x − x0 || we have that
x∈S
x
||m|| = |α||| − − x0 || ≥ |α|δ.
α
43
Hence,
|f (m)| = |α|δ ≤ ||m||,
for all m ∈ M . Observe that this inequality also holds if α is any scalar. Hence, f
is a bounded linear functional on M with (by Remark 2.15)
||f || ≤ 1. (3.4)
For arbitrary ² > 0, by the definition of δ, there exists x ∈ S such that
||x − x0 || < δ + ². (3.5)

x−x0
For such an x consider the vector y := ||x−x0 ||
Clearly, ||y|| = 1 and by the definition
of f , and inequality (3.5),
|f (x − x0 )| |0 − δ| δ
|f (y)| = = > . (3.6)
||x − x0 || ||x − x0 || δ+²
But ||f || = sup |f (x)| so that from (3.6) we obtain

||x||=1
δ
||f || = sup |f (y)| > , i.e.,
||y||=1 δ+²
δ
||f || > . (3.7)
δ+²
Since ² > 0 was arbitrary, it follows from (3.7) that
||f || ≥ 1. (3.8)
From (3.4) and (3.8) we obtain, ||f || = 1. By Theorem 3.9, there exists a bounded
linear functional F defined on X, extending f , such that
||F || = ||f || = 1,
and (by taking α = 1 and x = 0 in the definition of f ) F (x0 ) = δ. This completes

the proof.
¤
44
EXERCISES 3.1
1. (a) Prove that norm is a convex functional.

(b) Find an example of a subadditive functional which is not a convex func-
tional.
(c) Find an example of a sublinear functional which is not linear.
2. In the proof of Theorem 3.9, p : X → IR was defined for each x ∈ X by
p(x) = ||f ||.||x||.
Prove that
(a) p(x) ≥ 0 for each x ∈ X;
(b) p(x + y) ≤ p(x) + p(y) for x, y ∈ X;
(c) p(αx) = αp(x), x ∈ X; α ≥ 0.
3. Let M be as defined in the proof of Theorem 3.10 and f : M → IR be defined

by f (x) ≡ f (αx0 ) = α||x0 || for each x ∈ M . Prove that
(a) f is a linear functional;
(b) f (x0 ) = ||x0 ||.
4. Prove Corollaries 3.11 and 3.12.
5. Let X be a normed linear space. Prove that if f (x) = f (y) for all f ∈ X ∗ then
x = y.
6. Let x be an element of a normed linear space X and let X ∗ denote the dual
space of X. Prove that
||x|| = sup{|f (x)| : f ∈ X ∗ , ||f || = 1}.
(This is a very important result and should be remembered).
3.3 Geometric Forms of Hahn Banach Theorem

In this section we study the geometric forms of Hahn Banach Theorem. We begin
with the notion of hyperplanes. Throughout this section E will denote a normed
linear space, unless otherwise stated.
Definition 3.14 A hyperplane H in E is simply any set of the form
H := {x ∈ E : f (x) = α},
45
where f (not identically zero) is a linear functional on E and α ∈ IR. Following the
notation of Brezis [4] we shall call H a hyperplane of equation [f = α]. We consider
some examples.
Example 3.15 Let E = IR. Define H1 = {x ∈ IR : 3x = −2}. Here f (x) = 3x and
α = −2 and the hyperplane of equation [f = α] is given by
½ ¾ ½ ¾
2 2
H1 = {x ∈ IR : 3x = −2} = x ∈ IR : x = − = − .
3 3
H1 is the singleton {− 23 } , a point in IR. Observe that H1 is a closed subset of IR.
(Why?).
Example 3.16 Let E = IR2 . Define H2 = {(x1 , x2 ) ∈ IR2 : 3x1 + 2x2 = 5}. Here
the hyperplane of equation [f = α] is simply the straight line in IR2 whose equation
is 3x + 2y = 5. Is H2 a closed subset of E? Justify your answer.
Example 3.17 Let E = IR3 and H3 = {(x1 , x2 , x3 ) ∈ IR3 : 3x1 − x2 + 4x3 = −1}.
It is clear that the hyperplane H3 is the plane in IR3 whose equation is given by
3x − y + 4z = −1. Is H3 a closed subset of IR3 ? Justify your answer.
For hyperplanes, we have the following Proposition.

Proposition 3.18 The hyperplane H of equation [f = α] is closed if and only if f
is continuous.
An immediate consequence of this proposition is that a closed hyperplane is simply
an element of the dual space.
In order not to obscure our presentation of the main theorems of this section, we
defer the proof of this proposition. It will be given at the end of this section.
Definition 3.19 Let A ⊂ E and B ⊂ E. We say that the hyperplane H of equation

[f = α] separates A and B in a general sense if f (x) ≤ α for all x ∈ A and f (x) ≥ α
for all x ∈ B, and we say that H separates A and B in a strict sense if there exists
² > 0 such that f (x) ≤ α − ² for all x ∈ A and f (x) ≥ α + ² for all x ∈ B.
Geometrically, separation of A and B by H means that A and B are situated on
opposite sides of H.
We are now ready to state the geometric forms of Hahn Banach Theorem. Recall
that a set K ⊂ E is called convex if λx + (1 − λ)y ∈ K for all x, y ∈ K; λ ∈ [0, 1].
46
Theorem 3.20 (Hahn Banach Theorem, First geometric form)
Let A ⊂ E, B ⊂ E be two convex sets, nonempty and disjoint. Suppose that A is
open. Then there exists a closed hyperplane that separates A and B in the general
sense. (See Fig. 3.3).
closed hyperplane
B
A
111111
000000
000000
111111 open,
000000
111111 convex,
000000
111111
000000
111111
nonempty convex, nonempty
Fig.3.3
Remark 3.21 Recall that the method we used in the proof of the Hahn Banach
Theorem for normed linear spaces (Theorem 3.9) was to apply Theorem 3.7. To
prove the first geometric form of Hahn Banach Theorem (Theorem 3.20) we shall
use the same method, i.e., we shall apply Theorem 3.7. However, to apply Theorem
3.7, we must have a sublinear functional defined on E, we must have a subspace M
of E, and we must have a linear functional defined on M . To prepare the grounds,
we begin with the following definition.
Definition 3.22 Let E be a normed linear space. Let M ⊂ E be an open, convex
set with 0 ∈ M . For all x ∈ E define
p(x) := inf{α > 0 : α−1 x ∈ M }.
Then p is called the gauge of M .
Lemma 3.23 The gauge of M satisfies the following conditions:
G1: p(λx) = λp(x) for all x ∈ E, λ > 0;
G2: There exists a constant K > 0 such that 0 ≤ p(x) ≤ K||x|| for all x ∈ E;
G3: M = {x ∈ E : p(x) < 1};
G4: p(x + y) ≤ p(x) + p(y) for all x, y ∈ E.
47
Proof. G1: This follows from the definition;
G2: Since 0 ∈ M and M is open, there exists r > 0 such that Br (0) ⊂ M . Observe
rx x
that for arbitrary ε > 0 and x ∈ E, (1+ε)||x|| = (1+ε)||x|| ∈ Br (0) ⊂ M . From the
r
definition of p we obtain that
(1 + ε)
p(x) ≤ ||x||,
r
which establishes G2.
G3: Let x ∈ M . Since M is open, (1 + ²)x ∈ M, for sufficiently small ² > 0.

Then, from the definition of p,
1
p(x) ≤ < 1.
1+²
Conversely, if p(x) < 1 then there exists α ∈ (0, 1) such that α−1 x ∈ M . Now, we
have 0 ∈ M, α−1 x ∈ M . By the convexity of M , we obtain, since α ∈ [0, 1],
α(α−1 x) + (1 − α)0 ∈ M
i.e., x ∈ M . Thus,
M = {x ∈ E : p(x) < 1}.
G4: Let x, y ∈ E and ² > 0. By G1 and G3 we know that
x y
u := ∈ M, v := ∈ M,
p(x) + ² p(y) + ²
h i
x 1
(since p(u) = p p(x)+² = p(x)+² p(x) < 1 it follows that u ∈ M ,
similarly, v ∈ M ). Since M is a convex set, we have that
µ ¶ µ ¶
x y
λ + (1 − λ) ∈M
p(x) + ² p(y) + ²
for all λ ∈ [0, 1]. Now choose
p(x) + ²
λ= ,
p(x) + p(y) + 2²
so that µ ¶ µ ¶
x y x+y
λ + (1 − λ) = .
p(x) + ε p(y) + ε p(x) + p(y) + 2ε
48
x+y
Clearly λ ∈ [0, 1] and p(x)+p(y)+2²
∈ M . Again, by G3,
h x+y i
p <1
p(x) + p(y) + 2²
and by G1 this implies

p(x + y)
< 1,
p(x) + p(y) + 2²
so that
p(x + y) ≤ p(x) + p(y) + 2².
Since ² is arbitrary, this yields, for all x, y ∈ E,
p(x + y) ≤ p(x) + p(y),
as required. This completes the proof of the lemma.
Remark 3.24 In order to prove Theorem 3.20 using Theorem 3.7, the gauge func-
tion introduced in Definition 3.22 will be a candidate for a sublinear functional.
Since we still need to find candidates for a subspace M of E and for a linear func-
tional f on M , we first prove the following lemma which is a special case of Theorem
3.20 in which B is a singleton {x0 } where x0 6∈ A. In this case, the required subspace
M and functional f are easily defined as can be seen from the following lemma.
Lemma 3.25 Let U ⊂ E be open, convex and nonempty and let x0 ∈ E, x0 6∈ U .

Then there exists F ∈ E ∗ such that F (x) < F (x0 ) for all x ∈ U . (In particular, the
hyperplane of equation [F = F (x0 )] separates {x0 } and U in the general sense).
Proof. By translation we may assume that 0 ∈ U . Introduce the gauge of U

denoted by p. Consider the subspace M of E generated by x0 , i.e.,
M = {x : x = λx0 : λ ∈ IR}.
Define f : M → IR by
f (x) ≡ f (λx0 ) = λ.
Then,
(i) f is a linear functional on M ; (ii) f (x) ≤ p(x) for all x ∈ M .
By Theorem 3.7, there exists a linear functional F : E → IR extending f and such
that
F (x) ≤ p(x) for all x ∈ E.
49
In particular, by taking λ = 1, we have, since x0 ∈ M , F (x0 ) = f (x0 ) = 1.
Furthermore, from G2, F is continuous (Verify this) and from G3, for all x ∈ U ,
F (x) ≤ p(x) < 1 = F (x0 ).
This completes the proof of the Lemma.

¤
Using Lemma 3.25, the proof of Theorem 3.20 is now immediate.
Proof of Theorem 3.20. Our technique here is to apply Lemma 3.25. Put
U := A − B = {a − b : a ∈ A, b ∈ B}.
Then
(i) U is nonempty;
(ii) 0 6∈ U ;
(iii) U is open;
(iv) U is convex.
By Lemma 3.25 there exists f ∈ E ∗ such that
f (z) < f (0) = 0, ∀z ∈ U.
But z ∈ U ⇒ z = x − y for x ∈ A, y ∈ B. Thus, f (z) < 0 ⇒ f (x − y) < 0 ⇒

f (x) < f (y) for all x ∈ A, y ∈ B. Choose r such that
sup f (x) ≤ r ≤ inf f (y),

x∈A y∈B
and so the hyperplane H of equation [f = r] separates A and B in the general sense.

This completes the proof of Theorem 3.20.
¤
Theorem 3.26 (Hahn Banach Theorem; Second Geometric Form) Let A ⊂ E and
B ⊂ E be two convex sets, nonempty and disjoint. Suppose that A is closed and B
is compact. Then, there exists a closed hyperplane that separates A and B in the
strict sense. (See Fig.3.4).
50
closed hyperplane
closed, convex
nonempty
convex, compact
nonempty
Fig.3.4
Proof. For ² > 0, set

A² := A + B(0, ²).
B² := B + B(0, ²).
Then,
(i) A² T
and B² are convex, nonempty, and open;
(ii) A² B² = ∅ for sufficiently small ² > 0.
The verification of (i) and (ii) is left as an exercise. (See Exercises 3.2, Problem 5.).
By Theorem 3.20 there exists a closed hyperplane of equation [f = α] which sepa-

rates A² and B² in the general sense, so that we have
f (x + ²z) ≤ α ≤ f (y + ²z)
for all x ∈ A, y ∈ B, z ∈ B(0, 1). Consequently, we obtain from this inequality that
(Verify, see Exercises 3.2, Problem 5),
f (x) + ²||f || ≤ α ≤ f (y) − ²||f ||
for all x ∈ A, y ∈ B. Since ||f || 6= 0, this implies that A and B are separated in the
strict sense by the hyperplane of equation [f = α]. This completes the proof.
¤
We now give the proof of Proposition 3.18 which we deferred earlier.
51
Proof of Proposition 3.18 Let f be continuous. We want to prove that the hy-
perplane H of equation [f = α] is closed. So let {xn } be a sequence in H such
that xn → x as n → ∞. Since f is continuous, we have f (xn ) → f (x). But
xn ∈ H ⇒ f (xn ) = α. Hence f (x) = α and so x ∈ H. Therefore, H is closed.
Conversely, let the hyperplane H of equation [f = α] be closed. We want to prove f
is continuous. H is closed implies H c (the complement of H) is open. Let x0 ∈ H c
and suppose that f (x0 ) < α. Without loss of generality we may assume α ≥ 0. Let
r > 0 be such that Br (x0 ) ⊆ H c where Br (x0 ) = {x ∈ E : ||x − x0 || < r}. (This is
possible since H c is open).
Claim: f (x) < α for all x ∈ Br (x0 ).

To verify the claim. Suppose for some x1 ∈ Br (x0 ),
f (x1 ) > α.
(Note that f (x1 ) = α is not possible since x1 6∈ H). Then, since Br (x0 ) is a convex
set, we have that the set
L = {xλ : xλ = λx0 + (1 − λ)x1 , λ ∈ [0, 1]}
is contained in Br (x0 ) and so (since xλ 6∈ H)
f (xλ ) 6= α, (3.9)
for all λ ∈ [0, 1]. But then,
f (x1 ) − α
λs := ∈ [0, 1]
f (x1 ) − f (x0 )
h i
and f (xλs ) = f λs x0 + (1 − λs )x1 = λs f (x0 ) + (1 − λs )f (x1 ) = α,
contradicting (3.9). Hence, the verification of claim is complete. As a consequence
of this claim, we obtain
f (x0 + ru) < α
for all u ∈ B(0, 1) (since x0 + ru ∈ Br (x0 )). Then,
(i) f is continuous; (ii) ||f || ≤ 1r (α − f (x0 )).
(For a verification of (i) and (ii), see Exercises 3.2, Problem 1). The proof is com-
plete.
¤
EXERCISES 3.2
52
1. With λ, f and the other notations as in the proof of Proposition 3.18, verify
the following assertions:
(i) f (xλs ) = α;
(ii) f is continuous;
(iii) ||f || ≤ 1r (α − f (x0 )).
2. Let E = IR, M = (−1, 1). Let p : E → IR denote the gauge of M . Compute
p(−3), p( −1
2
), p(0), p( 12 ), p(3).
3. Let M = {x : x = λx0 : λ ∈ IR} as in the proof of Lemma 3.25. Define
f : M → IR by
f (x) ≡ f (λx0 ) = λ.
Prove
(i) f is linear;
(ii) f (x) ≤ p(x) for all x ∈ M , where p is the gauge of U (see Lemma 3.25);
(iii) Verify that F obtained in the proof of Lemma 3.25 is continuous.
4. In the proof of Theorem 3.20, we set
U := A − B
and made the following assertions:
(i) U is nonempty;
(ii) 0 ∈/ U;
(iii) U is convex;
(iv) U is open.
Verify these assertions.
5. In the proof of Theorem 3.26, for any ² > 0, we set
A² := A + B(0, ²);
B² := B + B(0, ²).
Prove
(i) A² ,TB² are convex, nonempty, and open;
(ii) A² B² = ∅ for sufficiently small ² > 0, (give a detailed proof);
(iii) Using the notations of Theorem 3.26, prove that
f (x + ²z) ≤ α ≤ f (y + ²z)
for all x ∈ A, y ∈ B, z ∈ B(0, 1) implies
f (x) + ²||f || ≤ α ≤ f (y) − ²||f ||.
53
[Hint: For part (iii), consider first f (x + ²z) ≤ α to obtain
²f (z) ≤ α − f (x).
Then, z ∈ B(0, 1) ⇒ (−z) ∈ B(0, 1) ⇒
²f (z) ≥ −[α − f (x)].
Also, consider α ≤ f (y + ²z)].
We conclude this chapter with a proof of Theorem 3.7.
3.4 Proof of Theorem 3.7
Proof. The proof will be broken into two parts. Fix x0 ∈ V \M and let N :=
M ∪ {x0 }, i.e., N is the subspace of V generated by M and {x0 }.
Part 1. In this part of the proof we shall prove that the map f is extendable
to N and with the desired properties. In short, we first prove Theorem 3.7 in the
case V = N . Once we have done this, we shall then use the result to obtain a proof
of Theorem 3.7.
For the first part, since f (x) ≤ p(x) for x ∈ M and f is linear we have for arbitrary
m1 , m 2 ∈ M
f (m1 − m2 ) = f (m1 ) − f (m2 ) ≤ p(m1 − m2 ),
so that since p is sublinear,
f (m1 ) − f (m2 ) ≤ p(m1 + x0 ) + p(−m2 − x0 ).
This yields
−p(−m2 − x0 ) − f (m2 ) ≤ p(m1 + x0 ) − f (m1 ). (∗)
If m1 is held fixed and m2 is allowed to vary over M , inequality (∗) implies that
the set {−p(−m2 − x0 ) − f (m2 )} of real numbers has an upper bound and hence
has the least upper bound. Let α = sup {−p(−m2 − x0 ) − f (m2 )}. Similarly, by
m2 ∈M
interchanging the roles of m1 and m2 we can set β = inf {p(m1 + x0 ) − f (m1 )}.
m1 ∈M
Then clearly, α ≤ β and so we can choose a real number r such that
α ≤ r ≤ β. (i)
54
We observe immediately that if α = β then we take r to be the common value.
Hence, for all m ∈ M , we obtain
−p(−m − x0 ) − f (m) ≤ r ≤ p(m + x0 ) − f (m). (ii)
From the definition of N , if x ∈ N then x = m + λx0 for some unique m ∈ M where

λ ∈ IR is uniquely determined. Now define F : N / IR by F (x) ≡ F (m + λx0 ) =
f (m) + λr where r is given in inequality (i). Then clearly,
(a) F is well defined;

(b) F is linear;
(c) F (x) = f (x) ∀ x ∈ M ;
(d) F (x) ≤ p(x) ∀ x ∈ N .
Verifications of (a), (b) and (c) are straightforward and are therefore left for the
reader. We verify (d). Here we shall consider separately the three cases λ > 0,
λ = 0 and λ < 0.
Case 1: λ = 0. By (c), F (x) ≡ F (m + 0.x0 ) = F (m) = f (m) = f (m + 0.x0 ) =

f (x) ≤ p(x) ∀ x ∈ N .
Case 2. λ > 0. From inequality (ii), we obtain
r ≤ p(m + x0 ) − f (m). (iii)
Now, (m/λ) ∈ N (since N is a subspace). Thus, replacing m by (m/λ) in inequality

(iii), we have:
³m ´ ³m´ µ ¶ ³m´
1
r≤p + x0 − f so that r ≤ p (m + λx0 ) − f .
λ λ λ λ
But p is a sublinear functional. Hence, the last inequality implies, using also the
linearity of f , that
f (m) + λr ≤ p(m + λx0 ),
and this implies, from the definition of F , that F (x) ≤ p(x) ∀ x ∈ N .
Case 3. λ < 0. From inequality (ii) we have −p(−m − x0 ) − f (m) ≤ r. Replacing

m by (m/λ) in this inequality and using the linearity of f we obtain
µ ¶
−m 1
−p − x0 ≤ r + f (m).
λ λ
55
Multiplying this inequality by λ (noting that λ is negative) we get
·µ ¶ ¸
1
(−λ)p − (m + λx0 ≥ F (x).
λ
Since p is a sublinear functional and (− λ1 ) > 0, we get p(m + λx0 ) ≥ F (x), i.e.,
F (x) ≤ p(x) ∀ x ∈ N. This completes the proof of the first part.
Part 2. We now prove the second part which will complete the proof of the theorem.
We shall make use of Zorn’s lemma with which the reader is assumed to be famil-
iar with (see e.g., Chapter 10, section 10.1.1). Let S denote the set of all linear
functionals G extending f and satisfying G(x) ≤ p(x) ∀ x ∈ D(f ), domain of f .
Clearly S is nonempty for F ∈ S, where F is the functional constructed in the first
part of this proof. Now, partially order S by set-inclusion, i.e., F1 < F2 if and only if
domF1 ⊂ domF2 and F2 |domF1 = F1 . (Here domF means domain of F and F2 |domF1
means F2 restricted to domF1 ). We now show that every chain in S has an upper
bound in S. Let C = {Fα : α ∈ ∆} be a chain in S. Let D = ∪ domFα . Consider
α∈∆
the function F ∗ whose domain is D. If x ∈ D, then x ∈ DomFα0 for some α0 ∈ ∆.
Define F ∗ : D / IR by F ∗ (x) = Fα (x) .
0
Claim 1 (a) D is a subspace of V ; (b) F ∗ is well defined.
Proof.
(a) Recall that a linear functional can only be defined on a subspace (why?). This
follows from the fact that domFα is a subspace for each α and that C is a
chain. The details are left to the reader
(b) Suppose x ∈ domFα and x ∈ domFβ . So x ∈ (domFα ) ∩ (domFβ ). Then, by

the definition of F ∗ , F ∗ (x) = Fβ (x). But C is a chain. So, either Fα extends
Fβ , or vice versa, and so Fα (x) = Fβ (x). Hence F is well-defined.
Claim 2 (i) F ∗ is linear; (ii) F ∗ (x) = f (x) ∀ x ∈ Df = M ; (iii) F ∗ (x) ≤

p(x) ∀ x ∈ Df .
Claim 2 is immediate and is therefore left for the reader. Hence, for each Fα ∈ C,
Fα ≤ F ∗ , so that F ∗ is an upper bound for C. By Zorn’s lemma, there exists a
maximal element F in S. This implies, F is linear, F extends f , F (x) ≤ p(x),
and F ∗ < F for every F ∗ ∈ S. It now only remains to prove that domF = V .
56
Clearly domF ⊆ V . We prove V ⊆ domF . Suppose that this is not the case,
i.e., suppose that domF $ V . Let x0 ∈ V \domF . By the first part of our proof
there exists F̂ such that F̂ is linear, F̂ (x) = F (x) for each x ∈ domF , F̂ (x) ≤
p(x) ∀ x ∈ [dom F ∪ {x0 }], and F < F̂ . This contradicts the maximality of F .
Hence dom F = V . The proof of the theorem is complete.
¤
57
58
CHAPTER 4
Uniform Boundedness Principle
4.1 Introduction and Preliminaries

In this chapter, we consider bounded linear maps from a real Banach space X into
a real Banach space, Y . We shall study one of the celebrated theorems for bounded
linear maps on a Banach space–The Uniform Boundedness Principle. This Theorem
is concerned with the following question: If X and Y are normed linear spaces and
{Tα }α∈I , (I is an arbitrary index set), is a family of bounded linear maps of X into
Y , when is
sup ||Tα || < ∞ ?
α∈I
Recall that since {Tα } is a collection of bounded linear maps, it follows that
||Tα || < ∞ for each α ∈ I. However, of course, there is no guarantee that
sup ||Tα || < ∞. The Uniform Boundedness Principle provides a condition for deter-
α∈I
mining when sup ||Tα || < ∞. As an immediate consequence of the uniform bound-
α∈I
edness principle, we shall prove another important result concerning bounded linear
maps from X into Y – The Banach Steinhaus Theorem. This theorem asserts that
if X is a Banach space and Y is a normed linear space and if {Tn } is a family of
bounded linear maps from X into Y such that for each x ∈ X, {Tn (x)} converges,
then T : X → Y defined by T x := lim Tn (x) is a bounded linear map. For the
n→∞
study of the uniform boundedness principle and the Banach Steinhaus theorem, we
start with the following preliminaries:
59
Lemma 4.1 (Baire’s Lemma) Let X be a nonempty complete metric space. Let
{Fn }∞
n=1 be a sequence of closed sets such that
∞
[
X= Fn .
n=1
Then there exists an integer n0 such that
int(Fn0 ) 6= ∅.
In other
S∞words, Baire’s lemma asserts that if X 6= ∅ is a complete metric space and if
X = n=1 Fn where Fn is closed for each n, then at least one of the Fn ’s must have a
nonempty interior i.e., must contain a nonempty open set (“int” means “interior”).
Proof. (see fig. 4.1)
Fig. 4.1
Assume, for contradiction, that no Fn contains an open ball. Then, the complement
Fnc intersects every open ball, S0 in X, i.e., Fnc ∩ S0 6= ∅ for each n. Observe that
Fnc ∩ S0 is a nonempty open set. Let x1 ∈ F1c ∩ S0 . So, there exists r1 < 12 such that
B1 ≡ B(x1 , r1 ) ⊂ F1c ∩ S0 . Since no Fn contains an open ball, B1 6⊆ F2 . Hence F2c
intersects B1 , i.e., F2c ∩ B1 6= ∅. Let x2 ∈ F2c ∩ B1 . Since F2c ∩ B1 is an open set,
there exists r2 < 41 such that
B2 ≡ B(x2 , r2 ) ⊂ F2c ∩ B(x1 , r1 ) ⊂ B1 .
60
By induction, we obtain a sequence Bn ≡ B(xn , rn ) with
c 1
Bn+1 ⊂ Fn+1 ∩ Bn ⊂ Bn ; rn < .
2n
Furthermore, Bn+1 ⊆ B n for each n.
Moreover, {xn }∞ /
n=1 is a Cauchy sequence in X. By completeness of X, xn
T
∞
x∗ ∈ Bn (since x∗ ∈ B n for each n). Hence x∗ ∈ Fnc for each n so that x∗ 6∈ Fn
n=1
S
∞
for each n. This implies x∗ 6∈ Fn = X. Contradiction. This completes the proof.
n=1
¤
Remark 4.2 Baire’s Lemma is sometimes stated in the following equivalent form:
Let (X, ρ) be a complete metric space. Let {Fn }∞n=1 be a sequence of closed subsets
of X. Assume that
int(Fn ) = ∅ f or all n ≥ 1,
then ∞
[
int( Fn ) = ∅.
n=1
A subset K of a topological space E is said to be dense in B ⊆ E if K ⊆ B. In

particular, K is dense in E if and only if K = E. For instance, the set of rational
numbers, Q, is dense in the reals (since Q = IR).
S
∞
A metric space E is said to be of first category if E = Fn wher Fn is a closed
n=1
set for each n, implies F n does not contain any open ball. The set E is of second
category if it is not of first category. With these definitions, Baire’s lemma can be
stated as follows:
Every complete metric space is of second category in itself.
As a consequence of this statement, Baire’s lemma is, in several books, referred to

as Baire’s Category Theorem.
We give some examples to illustrate the lemma.
Example 4.3 Let X = IR (with the usual metric), {Fn } = {qn }, where, for each n,
qn is a rational number. Clearly, {qn } is closed and int(qn ) = ∅ for each n (Verify).
61
By Baire’s Lemma,
∞
[
int(Q) = int( qn ) = ∅,
n=1
where Q denotes the set of rational numbers.
Example 4.4 Let X = IR2 , Fn = {(x, y) ∈ IR2 : x2 + y 2 = n1 }, n ≥ 1. Clearly, Fn is

closed and int(Fn ) = ∅, for each n ≥ 1. Baire’s Lemma implies that
∞
[
int( Fn ) = ∅.
n=1
4.2 The Uniform Boundedness Principle

We now prove the Uniform Boundedness Principle.
Theorem 4.5 (Uniform Boundedness Principle) Let X, Y be Banach spaces. Sup-

pose that
(i) {Tα }α∈∆ ⊂ B(X, Y );
(ii) ∀x ∈ X, supα∈∆ ||Tα (x)|| < ∞.
Then,
sup ||Tα || < ∞.
α∈∆
Proof. For each integer n, define
Fn := {x ∈ X : for each α ∈ ∆, ||Tα (x)|| ≤ n}.
Then,
(i) Fn is closed for each n,
S
∞
(ii) X = Fn .
n=1
Condition (i) is easy to prove and is left as an exercise (See Exercises 4.1, Prob-
lem 3(i)). Condition (ii) follows from the condition sup ||Tα (x)|| < ∞. (Explain
α∈∆
how this happens. See Exercises 4.1, Problem 3(ii)). Hence, by Lemma 4.1, at least
one of the Fn ’s must have a nonempty interior. Let int(Fn0 ) 6= ∅. Let x0 ∈ int(Fn0 ).
Hence there exists r > 0 such that B̄r (x0 ) ⊂ Fn0 . Then since, for all z ∈ B̄(0, 1),
x0 + rz ∈ B̄r (x0 ) ⊂ Fn0 , we must have
||Tα (x0 + rz)|| ≤ n0 , for all z ∈ B̄(0, 1),
62
and this implies (see Exercises 4.1, Problem 4) that
r||Tα || ≤ n0 + ||Tα (x0 )||,
so that (since x0 ∈ Fn0 , and consequently ||Tα (x0 )|| ≤ n0 ) we obtain:
sup ||Tα || ≤ 2r−1 n0 < ∞.

α∈∆
This completes the proof.

¤
An immediate consequence of the Uniform Boundedness Principle is the following
theorem:
Theorem 4.6 (Banach Steinhaus Theorem) Let X and Y be Banach spaces. Sup-
pose,
(i) Tn ∈ B(X, Y ) for all n ≥ 1.
(ii) ∀x ∈ X, {Tn (x)} converges to a limit denoted by T x.
Then,
(a) sup ||Tn || < ∞;
n≥1
(b) T ∈ B(X, Y );
(c) ||T || ≤ lim inf ||Tn ||.
Proof. {Tn (x)} converges implies it is bounded. Thus, there exists Kx ≥ 0 such
that
sup ||Tn x|| ≤ Kx < ∞.
n≥1
By the Uniform Boundedness Principle,
sup ||Tn || < ∞,

n≥1
which establishes (a).

Let sup ||Tn || ≤ c. Then, using this result, Tn ∈ B(X, Y ) implies for each x ∈ X
n≥1
||Tn x|| ≤ ||Tn ||||x|| ≤ c||x||, n ≥ 1. (4.1)
Furthermore, since
lim||Tn x|| = ||T x||,
taking limits as n / ∞ in inequality (4.1) we obtain
||T x|| ≤ c||x||,
63
for each x ∈ X. Hence T ∈ B(X, Y ), establishing (b). Finally, from the first part
of inequality (4.1) we obtain
lim inf ||Tn x|| ≤ (lim inf ||Tn ||)||x||, i.e., ||T x|| ≤ (lim inf ||Tn ||)||x||
for each x ∈ X. This implies (by Remark 2.15) that
||T || ≤ lim inf ||Tn ||,
establishing (c), and completing the proof.
¤
Corollary 4.7 Let E be a Banach space and let K be a subset of E. Suppose for
each f ∈ E ∗ , the set [
f (K) = hf, ki
k∈K
is bounded in IR. Then, K is a bounded subset of E.
Proof. We apply the Uniform Boundedness Principle. Set
X = E ∗;
Y = IR;
I = K.
Define the map Tk : E ∗ / R as follows. For each k ∈ K, set Tk (f ) = hf, ki, f ∈
∗ ∗
E . Then, Tk ∈ B(X , IR) and by hypothesis,
sup |Tk (f )| = sup |hf, ki| < ∞.
k∈K k∈K
By Uniform Boundedness Principle, sup ||Tk || ≤ c for some constant c ≥ 0. Then,

k∈K
|Tk (f )| = |hf, ki| ≤ c||f ||

for all f ∈ E ∗ , k ∈ K. This implies (by Exercises 3.1, Problem 6) that ||k|| ≤ c for
each k ∈ K. Hence, K is bounded.
¤
Remark 4.8 To establish that a given subset K of a Banach space E is bounded,
it suffices (by Corollary 4.7) to prove that for each f ∈ E ∗ , the set
[
f (K) := hf, xi
x∈K
is a bounded subset of IR. Note that in Theorem 4.5 and Theorem 4.6, it is sufficient
that Y be a real normed linear space (completeness of Y is not necessary). (Convince
yourself). Corollary 4.7 is also sometimes expressed by saying that weak and strong
boundedness are equivalent. The reason for this is the following Corollary.
64
Corollary 4.9 Let E be a Banach space, K ∗ be a subset of E ∗ . Suppose that for
each x ∈ E the set [
hK ∗ , xi = hf, xi
f ∈K ∗
is bounded in IR. Then K ∗ is bounded.
Proof. As in the proof of Corollary 4.7, we employ the Uniform Boundedness Prin-
ciple. Set
X = E;
Y = IR;
I = K ∗.
Define the map Tf ∗ : E / R as follows.
For each f ∗ ∈ K ∗ , put Tf ∗ (x) = hf ∗ , xi, x ∈ E. Then,
sup |Tf ∗ (x)| = sup |hf ∗ , xi| < ∞.

f ∗ ∈K ∗ f ∗ ∈K ∗
By the Uniform Boundedness Principle, there exists a constant c ≥ 0 such that
|hf ∗ , xi| = |Tf ∗ (x)| ≤ c||x||
for all x ∈ E, f ∗ ∈ K ∗ . Hence, ||f ∗ || ≤ c. This completes the proof.

¤
EXERCISES 4.1
1. With the notations of Example 4.3, verify the following assertions:

(a) qn is closed for each n;
(b) int (qn ) = ∅ for each n.
2. With the notations of Example 4.4 verify the following assertions:
(a) Fn is closed;
(b) int(Fn ) = ∅ for each n.
3. Let X, Y be real Banach spaces and Tα ∈ B(X, Y ) for each α ∈ I and suppose
sup kTα (x)k < ∞ for each x ∈ X.
α∈∆
(i) Prove that for each integer n, the set
Fn = {x ∈ X : for all α ∈ I, ||Tα (x)|| ≤ n}
is a closed subset of X.
S
∞
(ii) Verify that X = Fn .
n=1
65
4. With the notations as in the proof of the Uniform Boundedness Principle
(Theorem 4.5), prove that
||Tα (x0 + rz)|| ≤ n0 , for all α ∈ I, z ∈ B(0, 1)
implies
r||Tα || ≤ n0 + ||Tα (x0 )||.
5. Let X = E ∗ , Y = IR, I = K as in Corollary 4.7. For each k ∈ K, define
Tk (f ) := hf, ki, f ∈ E ∗ .
Prove that Tk ∈ B(E ∗ , IR) for each k ∈ K.
6. Let X be a real normed linear space and for r > 0, let
Br (x0 ) := {x ∈ X : ||x − x0 || ≤ r}.
Prove that
Br (x0 ) = {x0 } + Br (0).
Generalize this result to the following equation:
Br (x0 + y0 ) = {x0 } + Br (y0 )
for fixed x0 , y0 ∈ X.
(We shall use this result in the next section. The result is also a useful tool
for Exercises 3.2, Problems 4 and 5).
7. Apply the Baire’s lemma to prove the following version of the uniform bound-
edness principle for functionals: Let ∆ be an arbitrary index set and let E be
a complete metric space and {fα }α∈∆ be a family of real–valued continuous
functionals on E. Assume {fα }α∈∆ is pointwise bounded (i.e., for each x ∈ E,
there exists Mx ≥ 0 such that |fα (x)| ≤ Mx ). Prove that there exists an open
set U in E on which {fα }α∈∆ is uniformly bounded (i.e., there exist an open
set U and a constant M ≥ 0 such that
|fα (x)| ≤ M ∀ x ∈ U, ∀ α ∈ ∆).
66
CHAPTER 5
Open Mapping and Closed Graph Theorems
5.1 The Open Mapping Theorem

We present the following lemma which will be used to prove one of the fundamental
theorems of functional analysis, the Open mapping Theorem.
Lemma 5.1 Let X and Y be Banach spaces and let T ∈ B(X, Y ). Assume that T
is surjective. Then, there exists a constant c > 0 such that
T (BX (0, 1)) ⊃ BY (0, c). (5.1)
A proof of Lemma 5.1 is given at the end of this chapter.
We now give one important consequence of the lemma.
We shall use the notations Br (0) ≡ B(0, r), Br (x0 ) ≡ B(x0 , r).
Remark 5.2 Under the hypothesis of Lemma 5.1, the inclusion (5.1) implies that
T maps open sets in X into open sets in Y . This is, in fact, the Open Mapping
Theorem. Its importance will become clear in what follows. For now, we verify
that T maps open sets to open sets. Let U be an open set in X. We shall show
T (U ) is an open set in Y . So, let y0 ∈ T (U ). Then, there exists x0 ∈ U such
that T x0 = y0 . But U is open and contains x0 . So there exists r > 0 such that
Br (x0 ) ⊂ U , i.e., x0 + Br (0) ⊂ U (by Exercises 4.1(6)). Then, T (x0 + Br (0)) ⊂ T (U )
67
Fig. 5.1
i.e., y0 + T (Br (0)) ⊂ T (U ) which is the same as y0 + T (rB1 (0)) ⊆ T (U ) and by the
linearity of T , y0 +rT (B1 (0)) ⊆ T (U ). But, by property (5.1), BY (0, rc) ⊂ T (Br (0))
since Br (0) = rB1 (0). Hence,
y0 + BY (0, rc) ⊂ T (U )
which implies (again by Exercises 4.1(6)) that BY (y0 , rc) ⊂ T (U ). Since y0 ∈ T (U )

is arbitrary, we have proved that T (U ) is open.
As a result of this remark, we have proved the following fundamental theorem of
functional analysis.
Theorem 5.3 (The Open Mapping Theorem) Let X and Y be Banach spaces. Sup-
pose,
(i) T ∈ B(X, Y );
(ii) T is surjective.
Then T is an open mapping (i.e., T maps open sets into open sets).
Corollary 5.4 Let X and Y be two real Banach spaces. Suppose,

(i) T ∈ B(X, Y );
(ii) T is bijective.
Then, T −1 : Y → X is continuous.
68
Proof. T ∈ B(X, Y ) and T is surjective imply (by Lemma 5.1) there exists c > 0
cT x
such that T (BX (0, 1)) ⊇ BY (0, c). Thus, ∀² > 0 and ∀x ∈ X, (1+²)||T x||
∈ BY (0, c)
cT x
and this implies (1+²)||T x|| = T z for some z ∈ BX (0, 1). Hence, since T −1 exists and is
c||x||
linear, this implies (1+²)||T x||
= ||z|| < 1 ∀² > 0. Hence, c||x|| < (1 + ²)||T x|| ∀x ∈ X.
Since ² > 0 is arbitrary, this implies
1
||x|| ≤ ||T x|| ∀ x ∈ X,
c
which implies T −1 is continuous (Verify this) (See Exercises 5.1, Problem 3).
We shall see other applications of the Open Mapping Theorem later. Meanwhile
we introduce another fundamental theorem of functional analysis.
5.2 The Closed Graph Theorem

So far, we have studied bounded linear maps and have proved some theorems con-
cerning such maps. In applications, therefore, whenever one encounters a map, if
one is able to prove that the map is linear and bounded, then one has those theorems
at ones disposal. It turns out, however, that some of the important maps which are
defined in applications, under certain conditions, are linear but not bounded. In
particular, many maps defined in terms of the ordinary and partial differentiation
are, in general, not bounded. Luckily, these operators often posses another property
which, in a way, makes up for the fact that they are not bounded. These operators
are closed (or have closed graphs (defined below)). We shall begin in the next section
with the definitions and some properties of closed maps, closed graphs and these will
lead to one of the key theorems of functional analysis – the Closed Graph Theorem.
This theorem basically tells us when a closed linear map is bounded (continuous),
thus providing a means of recovering boundedness (continuity) for the class of closed
linear maps.
5.2.1 Closed Maps, Closed Graphs

Definition 5.5 Let X and Y be normed linear spaces and T : X → Y be any map.
Then, the graph of T denoted by G(T ), is defined by
G(T ) := {(x, T x) : x ∈ X}.
Observe that G(T ) is a subset of X × Y and that
(x, y) ∈ G(T ) if and only if T x = y.
69
Example 5.6 Let X = [0, 1], Y = IR, and T : [0, 1] → IR be defined by
T x = x2 , x ∈ [0, 1].
Then, the graph of T , G(T ), is given by
G(T ) = {(x, T x) : x ∈ [0, 1]} = {(x, x2 ) : x ∈ [0, 1]}.
Example 5.7 Let X = [−1, 1], Y = [0, 1] and S : X → Y be defined by

½
0, −1 ≤ x ≤ 0;
Sx =
1, 0 < x ≤ 1.
Then, S
G(S) = {(x, Sx) : x ∈ [0, 1]} = {(x, 0) : x ∈ [−1, 0]} {(x, 1) : x ∈ (0, 1]}.
Definition 5.8 Let X and Y be normed linear spaces and T : X → Y be any map.
Then T is called a closed map (or, is said to be closed or is said to have closed graph)
if the graph of T , G(T ), is a closed subset of X × Y .
Example 5.9 The graph of T given in Example 5.6 is closed. To see this, it suffices
to take an arbitrary convergent sequence in G(T ) and show that its limit is in G(T ).
So, let {(xn , T xn )} be an arbitrary sequence in G(T ) and suppose (xn , T xn ) → (x, y)
as n → ∞. This implies xn → x and T xn → y as n → ∞. Observe that the map T
is continuous on [0, 1]. Thus, xn → x ⇒ T xn → T x as n → ∞. By uniqueness of
limit, T x = y so that (x, y) ∈ G(T ). Thus, G(T ) is a closed subset of X × Y and so
T is a closed map.
Example 5.10 The graph, G(S), of the map S defined in Example 5.7 is not closed.
To see this, it suffices to produce a convergent sequence in G(S) whose limit is not in
G(S). To this end, consider the sequence {xn } in [0, 1] defined by xn = n1 , n = 1, 2, ....
Then, Sxn = 1 for each n. So, (xn , Sxn ) = ( n1 , 1) ∈ G(S) for each n. Moreover,
(xn , Sxn ) → (0, 1) as n → ∞. However, (0, 1) 6∈ G(S) since S(0) 6= 1. Hence, G(S)
is not a closed subset of X × Y and so S is not a closed map.
We now give a characterization of closed maps which is convenient in checking

whether or not a given map is closed.
Proposition 5.11 Let X and Y be normed linear spaces and T : X → Y be any

map. Then, T is closed if and only if for arbitrary {xn } ∈ D(T ), domain of T , with
xn → x, and T xn → y we have (i) x ∈ D(T ), and (ii) T x = y.
70
Proof. Exercise.
To appreciate the importance of our next theorem, we first give an example of a

map which is
(i) linear;
(ii) closed; and
(iii) not bounded.
Example 5.12 Let X = C[0, 1] = Y , where C[0, 1] is endowed with the sup norm.
Let
0
D = {f ∈ C 1 [0, 1] : f exists and is continuous on [0, 1]}
where the prime denotes differentiation. Let T be a map with domain D defined by
0
Tf = f ,
(i.e., T is the differentiation operator). Then,

(i) T is linear;
(ii) T is closed;
(iii) T is not bounded. (i) is obvious, We verify (ii). Let {fn } in D be such that
0
fn → f and T fn := fn → y.
Observe that T fn → y
0
⇒ ||T fn − y|| = sup |(T fn )(t) − y(t)| = sup |fn (t) − y(t)| → 0
t∈[0,1] t∈[0,1]
0
as n → ∞. This convergence is uniform and y(t) = lim fn (t). Since the conver-
n→∞
gence is uniform, we have:
Z t Z t Z t
0 0
y(s)ds = limfn (s)ds = lim fn (s)ds = f (t) − f (0),
0 0 0
so that Z t
f (t) = f (0) + y(s)ds.
0
0 0
It is now easy to see that f (t) exists and f (t) = y(t) for all t ∈ [0, 1]. Thus,
f ∈ D and T f = y so T is closed.
(iii) T is not bounded.
0
Take fn (t) = tn . Then, ||fn || = sup |tn | = 1 and fn (t) = ntn−1 so that
t∈[0,1]
||T fn || = sup |ntn−1 | = n, n = 1, 2, .... Thus, T is not bounded.
t∈[0,1]
71
5.2.2 The Closed Graph Theorem
Theorem 5.13 (The Closed Graph Theorem) Let X and Y be Banach spaces. Let
(i) T : X → Y be a linear map;
(ii) The graph of T, G(T ), be closed.
Then, T is continuous.
(Simply put, Theorem 5.13 says that if any linear map from a Banach space to a
Banach space has closed graph then it is continuous).
72
Proof.
Method 1. (See Fig.5.2)
G(T\ )
π1 π1−1
²
X
Fig.5.2
Consider the space X × Y endowed with norm k(x, y)kX×Y = kxkX + kykY . Then,
since X and Y are Banach spaces, it follows that X × Y is a Banach space with this
norm. G(T ) is closed in X × Y ⇒ G(T ) is a Banach space. Consider the projection
map
π1 : G(T ) /X
defined by
π1 ((x, T x)) = x.
Then, π1 is a bijective, continuous linear map. By Corollary 5.4, π1−1 : X / G(T )

−1
is continuous. Since π1 is also linear, there exists a constant M > 0 such that
kπ1−1 (x)k ≤ M kxk ∀ x ∈ X.
Then
kπ1−1 (x)k ≤ M kxk ⇒ k(x, T x)k ≤ M kxk

⇒ kxkX + kT xkY ≤ M kxkX
⇒ kT xkY ≤ M kxkX , ∀x ∈ X,
which implies T is continuous and completes the proof.

¤
73
Method 2. As in Method 1, consider now the projections π1 and π2 defined in the
usual way (see Fig.5.3), π1 (x, y) = x and π2 (x, y) = y.
G(TR 4 )
44
44
44
44
44
π1 44 π
44 2
44
44
π −1 44
1 44
44
¥ ½
X / Y
Fig.5.3
Now, G(T ) is closed in X × Y ⇒ G(T ) is a Banach space. Moreover, π1 is a

bijective continuous linear map. By Corollary 5.4, π1−1 is continuous. Clearly π2
is also continuous. Hence T = π2 ◦ π1−1 is continuous. The proof is complete.
¤
Note: The converse of Closed Graph Theorem is also true, i.e., if X and Y are
Banach spaces and T : X / Y is such that T ∈ B(X, Y ) then the graph of
T, G(T ), is closed (or equivalently, T is a closed map). This is easily seen. For,
suppose {xn } is a sequence in X such that xn / x and T xn / y. Continuity
of T implies T xn / T x, and uniqueness of limit gives T x = y, completing the
argument.
5.3 Some Applications

Proposition 5.14 Let E be a Banach space and let ||.||1 and ||.||2 be two norms
defined on E such that
(E, ||.||1 ) and (E, ||.||2 )
are Banach spaces. Suppose there exists a constant c > 0 such that
||x||1 ≤ c||x||2 for all x ∈ E.
Then, there exists k > 0 such that
||x||2 ≤ k||x||1 for all x ∈ E.
74
Proof. See Exercises 5.1, Problem 4.
P
∞ P
∞
Example 5.15 Prove that if αk ξk is convergent whenever |ξk | < ∞, then
1 1
sup |αk | < ∞.

k≥1
Method 1. (Application of the closed graph theorem) Consider the map

T : `1 / `∞ defined by
T (ξ1 , ξ2 , . . .) = (α1 ξ1 , α1 ξ1 + α2 ξ2 , α1 ξ1 + α2 ξ2 + α3 ξ3 , . . .)
i.e., ( )∞
n
X
T (ξ1 , ξ2 , ξ3 , . . .) = αk ξk .
k=1 n=1
Then T is: (i) well defined; (ii) linear; (iii) bounded.
Verification.
P∞ P∞
(i) x = {ξk }∞
k=1 ∈ `1 ⇒ k=1 |ξk | < ∞. Then by hypothesis αk ξk < ∞. This
½ n ¾∞ k=1
P
implies the sequence {Sn }∞
n=1 = αk ξk converges and so is bounded
k=1 n=1
P
n
i.e., sup |Sn | := sup | αk ξk | < ∞, i.e., T x ∈ l∞ . Hence T is well defined.
n≥1 n≥1 k=1
(ii) linearity is trivial.

(iii) T is bounded. We prove this by using the closed graph theorem.
Let {xn }∞ n ∞ ∞
n=1 := {ξk }n=1 be an arbitrary sequence in `1 and x = {ξk }k=1 ∈ `1 such
that
xn T xn
↓ in `1 and ↓ in `∞ .
x z = {ηk }∞
k=1
It suffices to show T x = z. Now,
∞
X
xn /x ⇒ kxn − xk`1 := |ξkn − ξk | / 0.
k=1
¯ ¯
¯Xi ¯
¯ ¯
T xn /z ⇒ kT xn − zk`∞ := sup ¯ αk ξkn − ηk ¯ / 0 as n / ∞.
i≥1 ¯ ¯
k=1
75
Therefore, for each i, i = 1, 2, 3, . . . ,
¯ ¯
¯Xi ¯ i
X
¯ ¯
¯ αk (ξkn − ξk )¯ ≤ sup |αk | · |ξkn − ξk | / 0 as n / ∞.
¯ ¯ k=1,2,...,i
k=1 k=1
½ ¾∞ ½ ¾∞
P
i P
i
Thus, for each i, {T xn }∞
n=1 := αk ξkn converges to αk ξk in `∞ . But
k=1 n=1 k=1 i=1
{T xn }∞ ∞
n=1 also converges (by hypothesis) to z = {ηk }k=1 and so, by the uniqueness
of limit,
i
X
ηi = αk ξk , i = 1, 2, . . . ,
k=1
i.e., z = T x, and so T is a closed linear map. Since `1 and `∞ are Banach spaces, it
follows from the closed graph theorem that T is bounded (continuous).
Furthermore,
kT xk`∞ ≤ kT k · kxk`1 ,
i.e., ¯ i ¯
¯X ¯ ∞
X
¯ ¯
sup ¯ αk ξk ¯ ≤ kT k · |ξk |.
i ¯ ¯
k=1 k=1
In particular, for each i = 1, 2, . . ., and ∀ x = {ξk }∞

k=1 in `1 we have,
¯ ¯
¯Xi ¯ ∞
X
¯ ¯
¯ αk ξk ¯ ≤ kT k · |ξk |. (5.2)
¯ ¯
k=1 k=1
We now consider x = {ξk } defined by

½ ½
sign ᾱi , if i = k 0 if ᾱi = 0;
ξk = where sign ᾱi = ᾱ
0, otherwise |ᾱi |
if ᾱi =
6 0.
¯ ¯
¯P i ¯ P∞
Then, ¯¯ αk ξk ¯¯ = |αi | for each i and |ξk | = 1. Using this in (5.2) now yields
k=1 k=1
|αi | ≤ kT k ∀ i and so, sup |αi | ≤ kT k < ∞, as desired.
i≥1
¤
Method 2. (Application of the Uniform Boundedness Principle). For each
n, define
fn : l1 → IR
76
by
n
X
fn (ξ) = αi ξi ,
1
P
∞ P
∞
where ξ = {ξi } ∈ l1 (so that |ξi | < ∞). By hypothesis, for each ξ ∈ l1 , αk ξk
1 k=1
is convergent. This implies, from our definition of fn that the sequence {fn (ξ)} is
convergent and consequently it is bounded. Thus,
sup |fn (ξ)| < ∞ for all ξ ∈ l1 .

n≥1
Since l1 is a Banach space, we can now invoke the uniform boundedness principle to
obtain that
sup ||fn || < ∞.

n≥1
We show ||fn || = max |αi |. For every ξ ∈ l1 we have

1≤i≤n
n
X n
X
|fn (ξ)| = | αi ξi | ≤ max |αi | |ξi | ≤ max |αi |.||ξ||l1 .
1≤i≤n 1≤i≤n
i=1 i=1
Hence, by Remark 2.15,
||fn || ≤ max |αi | = |αk |, (5.3)

1≤i≤n
for some k ∈ {1, 2, ..., n}. Now choose ξ = {ξi } in `1 where

½
sign (αk ), if i = k;
ξi =
0, otherwise.
Consequently, we have
||ξ||l1 = 1.
Moreover, using this equality, we have that
n
X
|fn (ξ)| = | αi ξi | = |αk | = max |αi |.||ξ||l1 . (5.4)
1≤i≤n
i=1
From (5.3) and (5.4) we obtain that
||fn || = max |αi |. (5.5)

1≤i≤n
77
From (5.3) and (5.5) we finally obtain that
sup ||fn || = sup( max |αi |) = sup |αk | < ∞.

n≥1 n≥1 1≤i≤n k≥1
This implies
sup |αn | < ∞,
n≥1
which completes the proof.

¤
EXERCISES 5.1
1. If T is a closed linear map with N (T ) = {0}, prove that T −1 is closed, where

N (T ) denotes the null space of T .
2. Let (X, k · kX ) and (Y, k · kY ) be Banach spaces. By considering X with a new

norm
kxk0 := kxkX + sup kTα xkY
α∈∆
where ∆ is an arbitrary index set, deduce the uniform boundedness principle

from the closed graph theorem.
3. Prove the following fact used in the proof of Corollary 5.4:

if X and Y are Banach spaces and T : X → Y is a linear map such that for
some constant c > 0 and for each x ∈ X we have
1
||x|| ≤ ||T x||,
c
then T −1 exists and is continuous.

(Hint: Consider the identity map Id : (E, ||.||1 ) → (E, ||.||2 ), set T = Id in
Corollary 5.4).
P
∞
5. Prove, that if αk ψk is convergent whenever lim ψk = 0 then
k=1
∞
X
|αk | < ∞.
k=1
78
6. Recall that two norms k · k1 and k · k2 on a linear space are called equivalent if
there exist numbers α > 0, β > 0 such that α||x||1 ≤ ||x||2 ≤ β||x||1 , ∀x ∈ X.
Suppose X is a linear space which is a Banach space for each of these norms.
Suppose that kxk1 ≤ kkxk2 is valid for some k and all x ∈ X. Prove that the
two norms are equivalent.
(Hint: Apply Proposition 5.14).
We conclude this chapter with the following proof which we deferred in Section 5.1.
Proof of Lemma 5.1
Part 1. We first establish that there exists c∗ > 0 such that
T (BX (0, 1)) ⊃ BY (0, c∗ ).
T is surjective implies that Y = ∪∞ ∞

n=1 T (BX (0, n)) and thus Y = Ȳ = ∪n=1 T (BX (0, n)).
Since Y is complete, by Baire Category theorem, there exists n0 ∈ IN such that
int(T (BX (0, n0 ))) 6= ∅. Let y0 ∈ int(T (BX (0, n0 ))). Without loss of generality we
may assume y0 ∈ T (BX (0, n0 )). Define h : Y / Y by
h(y) = y − y0 .
Clearly, h is a homeomorphism of Y onto itself. In particular, h(T (BX (0, n0 )) =

T (BX (0, n0 )) − y0 . Since y0 ∈ int(T (BX (0, n0 ))) and h is a homeomorphism, then
the image of y0 under h must be an interior point of T (BX (0, n0 ))−y0 . But the image
of y0 under h is h(y0 ) = 0, so T (BX (0, n0 )) − y0 has the origin as an interior point.
Moreover, since y0 ∈ T (BX (0, n0 )), there exists x0 ∈ BX (0, n0 ) such that T x0 = y0 .
But x0 ∈ BX (0, n0 ) implies (−x0 ) ∈ BX (0, n0 ) and so, T (−x0 ) = −T x0 = −y0 .
Hence −y0 ∈ T (BX (0, n0 )). Hence,
T (BX (0, n0 )) − y0 ⊂ T (BX (0, n0 )) + T (BX (0, n0 )) = 2T (BX (0, n0 ))

= T (2BX (0, n0 )) = T (BX (0, 2n0 )).
From this we obtain that
T (BX (0, n0 )) − y0 = T (BX (0, n0 )) − y0 ⊆ T (BX (0, 2n0 )),
which implies that the origin is an interior point of T (BX (0, 2n0 )). But T (BX (0, 2n0 ) =
2n0 T (BX (0, 1) and since multiplication by a nonzero scalar is a homeomorphism
79
of Y , it follows that the origin is an interior point of T (BX (0, 1)) , i.e., 0 ∈
int(T (BX (0, n0 ))). Hence there exists c∗ > 0 such that: (a)T (BX (0, 1)) ⊇ BY (0, c∗ ).
Part 2. To conclude the Lemma, it suffices now to show that: (b) T (BX (0, 3)) ⊃
∗
BY (0, c∗ ) which is equivalent to T (BX (0, 1)) ⊃ BY (0, c), where c := c3 . We now
prove (2). Let y ∈ BY (0, c∗ ). By part (1), y ∈ T (BY (0, 1)). So, there exists
c∗
x1 ∈ X such that ||x1 || ≤ 1 and ||y − y1 || < 2
, where y1 = T x1 .
∗
Observe that BY (0, c2 ) ⊆ T (BX (0, 21 )), (from Part 1, by dividing both sides by 2).
∗
So, since (y − y1 ) ∈ BX (0, c2 ), there exists
1 c∗
x2 ∈ X such that ||x2 || ≤ 2
and ||(y − y1 ) − y2 || < 22
, where y2 = T x2 .
Continuing in this way, we obtain a sequence
1 c∗
{xn } ∈ X such that ||xn || < 2n−1
and ||y − (y1 + y2 + ... + yn )|| < 2n
, where
yn = T x n .
1
If we put Sn := x1 + x2 + ... + xn , then it follows from ||xn || < 2n−1 that {Sn } is
a Cauchy sequence in X for which ||Sn || ≤ ||x1 || + ||x2 || + ... + ||xn || < 1 + 12 +
1 / x.
... + 2n−1 < 2. By completeness of X, there exists x ∈ X such that Sn
Moreover, ||x|| = || lim Sn || = lim ||Sn || ≤ 2 < 3. Hence, x ∈ BX (0, 3). By the
continuity of of T , T (x) = T (lim Sn ) = lim T (Sn ) = lim(y1 + y2 + ... + yn ) = y, since
||y − (y1 + y2 + ... + yn )|| < 21n / 0 as n / ∞. Hence, since x ∈ BX (0, 3)
and T x = y, it follows that y ∈ T (BX (0, 3)). Hence T (BX (0, 3)) ⊇ BY (0, c∗ ). Set
∗
c := c3 . Then T (BX (0, 1)) ⊃ BX (0, c), completing proof of the lemma.
¤
80
CHAPTER 6
Hilbert Spaces
6.0 Introduction
In this chapter we introduce a special class of Banach spaces called Hilbert space.
These spaces are extremely useful in applications.
6.1 Definition and Examples

Definition 6.1 Let E be a linear space. An inner product on E is a function,
h, i : E × E / K defined on E × E with values in K (where K = IR or C I) such
that the following three conditions are satisfied: For x, y, z ∈ E, λ, µ ∈ C
I,
I1 : hx, xi ≥ 0 and hx, xi = 0 if and only if x = 0;

I2 : hx, yi = hy, xi, where the “bar” indicates complex conjugation; and
I3 : hλx + µy, zi = λhx, zi + µhy, zi.
The pair, (E, h, i) is called an inner product space. We shall simply write E
for the inner product space (E, h, i) when the inner product h, i is known.
Example 6.2 The linear space IRn , with the function h, i defined, for arbitrary
vectors x = (x1 , x2 , . . . , xn ), y = (y1 , y2 , . . . , yn ) in IRn , by
X n
hx, yi = xi y i
i=1
81
is an inner product space.
Example 6.3 The linear space CI n , with the function h, i defined, for arbitrary
u = (z1 , z2 , . . . , zn ), ω = (w1 , w2 , . . . , wn ) in CI n ; by
n
X
hu, ωi = zi wi
i=1
Example 6.4 The linear space `2 (C I), with the function h, i defined, for arbitrary
vectors x = (x1 , x2 , . . .), y = (y1 , y2 , . . .) in `2 , by
∞
X
hx, yi = xi y i
i=1
is an inner product space. (That the series converges is an immediate consequence

of the Cauchy–Schwartz inequality applied to the partial sums (Proposition 6.7
below)).
Example 6.5 The linear space C[0, 1] with the function h, i defined, for arbitrary
f, g ∈ C[0, 1], by
Z 1
hf, gi = f (t) g(t) dt
0
Example 6.6 The linear space L2 [0, T ] with h, i defined, for arbitrary f, g ∈ L2 [0, T ],
by
Z T
hf, gi = f (t) g(t) dt
0
is an inner product space. (This example is for readers who are familiar with the
space L2 [0, T ]).
The trivial verifications are left as easy exercises for the reader.
82
6.2 Basic Properties
In this section we establish some basic properties of inner product spaces.
Observe first that an immediate consequence of I2 and I3 is that for arbitrary

x, y, z ∈ E and λ, µ ∈ CI,
hz, λx + µyi = λ̄hz, xi + µ̄hz, yi.
For,
hz, λx + µyi = hλx + µy, zi by I2

= λhx, zi + µhy, zi, by I3
= λ̄hx, zi + µ̄hy, zi, property of CI
= λ̄hz, xi + µ̄hz, yi, by I2
as stated.
Proposition 6.7 (Cauchy-Schwartz Inequality) Let E be an inner product space.

For arbitrary x, y ∈ E,
|hx, yi|2 ≤ hx, xi · hy, yi.
Equality holds if and only if x and y are linearly dependent.
Proof. Let x, y ∈ E be arbitrary. Choose z ∈ C I with |z| = 1 and such that

zhx, yi = |hx, yi|.
Set a = hx, xi, b = |hx, yi| and c = hy, yi. Then for arbitrary scalar t ∈ IR we obtain,
0 ≤ htzx + y, tzx + yi
= t2 z z̄hx, xi + tzhx, yi + tz̄hy, xi + hy, yi
= at2 + 2t|hx, yi| + c = at2 + 2bt + c.
From the theory of quadratic functions this yields the following inequality: b2 ≤ ac,
i.e., |hx, yi|2 ≤ hx, xi · hy, yi, as required.
¤
p
Lemma 6.8 The function k · k : E / IR defined by kxk = hx, xi is a norm on
E.
83
Proof. That k·k satisfies N1 and N2 follows from the definition of k·k and conditions
I1 and I2 . It now suffices to verify N3 . For arbitrary x, y ∈ E,
kx + yk2 = hx + y, x + yi = hx, x + yi + hy, x + yi = hx, xi + hx, yi + hy, xi + hy, yi
³ ´
= kxk + hx, yi + hx, yi + kyk2
2
= kxk2 + 2Re hx, yi + kyk2

≤ kxk2 + 2|hx, yi| + kyk2
≤ kxk2 + 2kxk · kyk + kyk2 = (kxk + kyk)2 ,
and the result follows immediately.
¤
Remark 6.9 As a consequence of Lemma 6.8, the Cauchy–Schwartz inequality is
generally written as follows:
|hx, yi| ≤ kxk · kyk, for arbitrary x, y ∈ E.
Lemma 6.10 The inner product h, i is a continuous function on E × E.
Proof. This follows from the Cauchy–Schwartz inequality and is left as an easy exer-
cise for the reader. ¤
Proposition 6.11 (The Parallelogram Law) Let E be an inner product space.
Then for arbitrary x, y ∈ E,
kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ).
(This result is generally called the parallelogram law for obvious reasons).
Proof. Exercise (Simply expand LHS to obtain RHS).

Proposition 6.12 (The Polarization Identity) Let E be an inner product space.
Then for arbitrary x, y ∈ E,
1© ª
hx, yi = kx + yk2 − kx − yk2 + ikx + iyk2 − ikx − iyk2 , where i2 = −1.
4
Proof. Exercise (Expand the RHS to obtain the LHS).
With respect to the norm defined in Lemma 6.8 we can define a Cauchy sequence
in an inner product space E. A sequence {xn }∞ n=1 in E is called Cauchy if and only
1/2
if hxn − xm , xn − xm i := kxn − xm k / 0 as n, m / ∞. Consequently, an
inner product space E is called complete if every Cauchy sequence in E converges
to a point of E. A complete inner product space is called a Hilbert space.
84
6.3 Completeness of an Inner Product Space
A Hilbert space will be called a complex Hilbert space or a real Hilbert space accord-
ing as the underlying linear space is complex or real, respectively. In what follows
we shall denote Hilbert space by the letter H. Recall that a set K in a vector space
is called convex if for each x, y ∈ K and λ ∈ [0, 1] we have λx + (1 − λ)y ∈ K. In
other words, K is convex if for any two points x, y ∈ K the line segment joining x
and y lies in K.
The parallelogram law has the following important consequence.
Theorem 6.13 Let H be a Hilbert space and let K be a closed convex subset of H.
Then K contains a unique vector of minimum norm.
Proof. Let δ = inf{kxk : x ∈ K}. Let {xn }∞ / δ (this

n=1 in K be such that kxn k
follows from the definition of inf ). Using the parallelogram law and the convexity
of K we have the following estimates:
¡ ¢
kxn − xm k2 = 2 kxn k2 + kxm k2 − 4k(xn + xm )/2k2 / 0 as m, n / ∞,
since (xn + xm )/2 ∈ K and so k(xn + xm )/2k ≥ δ.

Hence, {xn }∞n=1 is a Cauchy sequence in K and since K is a closed subset of the
complete space H, it is complete. Thus, the sequence {xn }, has a limit x∗ (say) in
K. We first show that x∗ has minimum norm in K. Observe that
kx∗ k = k lim xn k = lim kxn k = δ.
Hence x∗ has minimum norm. We now show x∗ is unique. If u ∈ K, u 6= x∗ and

kuk = δ then
0 < kx∗ − uk2 = 4δ 2 − 4k(x∗ + u)/2k2 ,
so that k(x∗ + u)/2k < δ, and since convexity of K implies (x∗ + u)/2 is in K, this
is a contradiction. Hence u = x∗ , completing the proof.
¤
Remark 6.14 Closed convex subsets in arbitrary Banach spaces need not contain
unique vectors of minimum norm, (see Exercises 6.1, problem 20). Moreover, a
closed convex subset of an arbitrary Banach space may have infinitely many vectors
of minimum norm (see Exercises 6.1, Problem 21). Apart from Hilbert spaces, there
are also other Banach spaces in which closed convex subsets have unique vectors of
minimum norms (see e.g., Miscellaneous Exercises, Problem 25).
85
6.4 Jordan Von Neumann Theorem
Let V be a linear space (not necessarily normed). It is clear that we can always
define a metric on V . For example, the function δ : V × V / IR defined by
½
1, x 6= y
ρ(x, y) = is a metric on V . In case that V is a normed linear space,
0, x = y
the norm on V , k · k, can also induce a metric on V given by
ρ(x, y) = kx − yk.
Thus, given a metric space, it is sometimes of interest to know whether or not the
metric on the space is induced by a norm. It can be shown that a metric ρ on a
linear space (M, ρ) is induced by a norm if the following two conditions are satisfied:
C1 : The metric is homogeneous, i.e.,
ρ(λx, λy) = |λ|ρ(x, y)
for arbitrary x, y ∈ M and scalar λ.
C2 : The metric is translation invariant, i.e.,
ρ(x + z, y + z) = ρ(x, y)
for arbitrary x, y, z ∈ M .
For an example of a metric not induced by a norm, the reader is referred to Problem
13 in the Miscellaneous Exercises (see also the above example where ρ(λx, λy) =
1 ∀λ and ∀x 6= y).
In Lemma 6.8, we defined a norm on an inner product space by kxk = hx, xi1/2 .
The following question is of interest.
Question. When is the norm on a normed linear space given by an inner product?
The answer to this question is given in the following theorem.

Theorem 6.15 (Jordan–Von Neumann, 1935) The norm on a normed linear space
E is given by an inner product if and only if the norm satisfies the parallelogram
law. i.e., if and only if for arbitrary x, y ∈ E,
¡ ¢
kx + yk2 + kx − yk2 = 2 kxk2 + kyk2 .
Remark 6.16 Theorem 6.15 is useful in showing, for example, when a normed
linear space E is not an inner product space with kxk2 = hx, xi. It suffices, in view
of this theorem, to show that the norm defined on the space does not satisfy the
parallelogram law. (See Exercises 6.1, Problems 13, 14 and 15).
86
6.5 Orthonormal Sets
Definition 6.17 Two vectors x and y in an inner product space E are said to be
orthogonal (written x ⊥ y or y ⊥ x and read x ‘perp’ to y or y ‘perp’ to x) if
hx, yi = 0. If M is a subset of E, then we write x ⊥ M if x ⊥ y for every y ∈ M .
Remark 6.18 Since hx, yi = hy, xi, it follows that x ⊥ y if and only if y ⊥ x.
Furthermore, x ⊥ x if and only if x = 0, and x ⊥ 0 for all x ∈ E.
Definition 6.19 A set S in an inner product space E is called an orthogonal set

if hx, yi = 0 for each x, y ∈ S, x 6= y. The set S is called orthonormal if it is an
orthogonal set and kxk = 1 for each x ∈ S.
Example 6.20 In `2 , S = {en }∞

n=1 given by
e1 = (1, 0, 0, . . .)
e2 = (0, 1, 0, . . .)
.. ..
. .
nth position
en = (0, 0, . . . , 0, 1, 0, . . .)
.. ..
. .
is an orthonormal set.
n o∞
Example 6.21 In L2 [0, 2π], the set S = √1 int
e is an orthonormal set.
2π n=1
For orthonormal sets we have the following result.
Proposition 6.22 Let S be an orthonormal set in an inner product space E. Let

{u1 , u2 , . . . , un } be a finite subset of S. Then for any x ∈ E,
n
X
|hx, ui i|2 ≤ kxk2 .
i=1
Proof. Let αi = hx, ui i. Then,

° n
X °2
° °
°x − αi ui ° ≥ 0,
i=1
87
so that, * +
n
X n
X
x− α i ui , x− αj uj ≥ 0,
i=1 j=1
i.e.,
* n
+ * n
+ * n n
+
X X X X
hx, xi − x, αj uj − αi ui , x + αi ui , αj uj ≥ 0,
j=1 i=1 i=1 j=1
or,
n
X n
X n X
X n
2
kxk − ᾱj hx, uj i − αi hui , xi + αi ᾱj hui , uj i ≥ 0,
j=1 i=1 i=1 j=1
By definition, αi = hx, ui i, αj = hx, uj i. Therefore,

n
X n
X n
X
2
kxk − ᾱj αj − αi ᾱi + |αi |2 ≥ 0,
j=1 i=1 i=1
using hui , uj i = 0, i 6= j, so that

n
X n
X n
X
2 2 2
kxk − |αj | + |αi | − |αi |2 ≥ 0
j=1 i=1 i=1
which yields
n
X n
X
|αj |2 := |hx, uj i|2 ≤ kxk2 , as required.
j=1 j=1
¤
Proposition 3.22 can be generalized to yield the following theorem.
Theorem 6.23 (Bessel’s Inequality) If {ui }∞
i=1 is an orthonormal set in an inner
product space E, then for arbitrary x ∈ E,
∞
X
|hx, ui i|2 ≤ kxk2 .
i=1
³ P
∞ ´
Furthermore, x − hx, ui iui is orthogonal to ui for each i.
i=1
Proof. Exercise (see Exercise 6.1, Problem 10).
Note: It is easy to check that in an orthonormal set, Pythagoras theorem holds,

i.e., if x ⊥ y then kx + yk2 = kxk2 + kyk2 .
88
6.6 The Projection Theorem
Using the concept of orthogonality, we shall extend the well known elementary fact
that the shortest distance from a point to a plane is given by the perpendicular
from the point to the plane. This, in fact, gives us a very powerful and important
theorem – the Projection Theorem – used, for example, in Optimization Theory.
Let E be an inner product space and let M be a subspace of E. Given an arbi-

trary vector x ∈ E, we consider the problem of finding a vector m∗ ∈ M which is
closest to x in the sense that kx − mk is minimized i.e.,
kx − m∗ k = inf{kx − yk : y ∈ M } = dist (x, M ).
The basic questions can be listed as follows:
Question 1 Is there actually a vector m∗ ∈ M which minimizes kx − mk over

all m ∈ M ?
Question 2 If the answer to Question 1 is affirmative, is m∗ unique?
Question 3 If a unique m∗ exists that minimizes kx−mk, how is m∗ characterized?
The answers to these questions are contained in the next two theorems.
Theorem 6.24 Let E be an inner product space, M a subspace of E, and x an

arbitrary vector in E. Then, if there exists a vector m∗ in M such that
kx − m∗ k = inf kx − mk,
m∈M
then m∗ is unique. In fact, m∗ ∈ M is a unique minimizing vector if and only if

(x − m∗ ) ⊥ M .
Proof. (⇒) Let m∗ ∈ M be the unique minimizing vector in M . We prove (x −

m∗ ) ⊥ M . Assume for contradiction that this is not the case. Then there exists
0 6= m0 ∈ M which is not orthogonal to (x − m∗ ). Without loss of generality we
may assume km0 k = 1 (otherwise, normalize m0 by dividing by km0 k). Since m0 is
not orthogonal to (x − m∗ ), let hx − m∗ , m0 i = δ 6= 0. Define a vector mδ ∈ M by
mδ = m∗ + δm0 . Then,
kx − mδ k2 = kx − m∗ − δm0 k2
= kx − m∗ k2 − hx − m∗ , δm0 i − hδm0 , x − m∗ i + |δ|2
= kx − m∗ k2 − |δ|2 < kx − m∗ k2 ,
89
contradicting the hypothesis that m∗ is the unique minimizing vector. Hence
(x − m∗ ) ⊥ M .
(⇐) Let (x − m∗ )⊥ M . We want to prove m∗ is the unique minimizing vector.

For arbitrary m ∈ M, m 6= m∗ , we compute:
kx − mk2 = k(x − m∗ ) + (m∗ − m)k2 = kx − m∗ k2 + km∗ − mk2 .
Thus,
kx − mk > kx − m∗ k for m 6= m∗ ,
which shows m∗ is the minimizing vector. Uniqueness follows trivially, completing
the proof of Theorem 6.24.
¤
Remark 6.25 Theorem 6.24 does not guarantee the existence of the minimizing
vector. It only asserts that if it exists, then it is unique and (x − m∗ ) ⊥ M .
If, instead of considering arbitrary inner product spaces we consider Hilbert spaces,
H, and closed subspaces M of H, we obtain the following theorem which guarantees
the existence of the minimizing vector.
Theorem 6.26 (The Projection Theorem) Let H be a Hilbert space and M a closed
subspace of H. For arbitrary vector x in H, there exists a unique vector m∗ ∈ M
such that kx − m∗ k ≤ kx − mk for all m ∈ M . Furthermore, m∗ ∈ M is the unique
vector if and only if (x − m∗ ) ⊥ M .
Proof. We establish only the existence of a minimizing vector m∗ . The uniqueness

follows from Theorem 6.24. If x ∈ M , then choose m∗ = x and there is nothing left
to prove. So, assume x 6∈ M and define
δ := inf{kx − mk : m ∈ M }.
It suffices now to produce an m∗ ∈ M with kx−m∗ k = δ. Let {mj }∞

j=1 be a sequence
of vectors in M such that kx − mj k / δ as j / ∞. (This follows from the
definition of “inf”). By the parallelogram law,
k(mi − x) + (x − mj )k2 + k(mi − x) − (x − mj )k2 = 2kmi − xk2 + 2kx − mj k2 .
Rearrangement yields,
° mi + mj °
2 2 °2 °2
kmi − mj k = 2kmi − xk + 2kx − mj k − 4°x − °.
2
90
For i and j, the vector (mi + mj )/2 is in M since M is a linear subspace. Hence, by
the definition of δ,
kx − (mi + mj )/2k ≥ δ,
so that,
kmi − mj k2 ≤ 2kmi − xk2 + 2kx − mj k2 − 4δ 2 .
Since kmj − xk / δ as j / ∞, we have that kmi − mj k / 0 as i, j / ∞.
Hence {mj }∞j=1 is a Cauchy sequence in M and since M is complete (as a closed
subset of a Hilbert space), it follows that {mj }∞ ∗
j=1 has a limit m in M . This
implies, x − mj / x − m∗ as j / ∞ and so (by continuity of the norm)
kx − mj k / kx − m∗ k as j / ∞. But kx − mj k / δ as j / ∞, so that
by the uniqueness of limit, we obtain,
kx − m∗ k = δ = inf{kx − mk : m ∈ M }.

¤
As an immediate application of the Projection Theorem, now we prove that a Hilbert
space can be represented as a “direct sum” of two of its closed subspaces.
Definition 6.27 Let E be a vector space. E is said to be the direct sum of two
subspaces M and N of E, written E = M ⊕ N if each x ∈ E can be represented
uniquely as x = m + n with m ∈ M and n ∈ N . In this case, N is called the
algebraic complement of M in E (and vice versa). The subspaces M and N are
called complementary pair of subspaces in E.
Definition 6.28 For a Hilbert space H, given a subset S of H, the set of all vectors
orthogonal to S is called the orthogonal complement of S and is denoted by S ⊥ i.e.,
S ⊥ = {x ∈ H : x ⊥ s for each s ∈ S}.
Some of the relationships between a set S in H and its orthogonal complement in

H are summarized in the following proposition.
Proposition 6.29 Let M and N be arbitrary subspaces of a Hilbert space H. Then,

(i) M ⊥ is a closed subspace of H; (ii) M ⊂ M ⊥⊥ ; (iii) If M ⊂ N then N ⊥ ⊂ M ⊥ ;
(iv) M ⊥⊥⊥ = M ⊥ .
Proof.
(i) The proof of this part is left as an easy exercise for the reader.
91
(ii) If x ∈ M then x ⊥ y for all y ∈ M ⊥ . Therefore, x ∈ (M ⊥ )⊥ = M ⊥⊥ .
(iii) y ∈ N ⊥ ⇒ y ⊥ x for each x ∈ M since M ⊂ N . Therefore, y ∈ M ⊥ .
(iv) Using (ii), M ⊥ ⊆ M ⊥⊥⊥ and M ⊆ M ⊥⊥ . From (iii), (M ⊥⊥ )⊥ ⊆ M ⊥ ⊆ M ⊥⊥⊥

and so M ⊥⊥⊥ = M ⊥ .
Theorem 6.30 (Direct Sum Decomposition) Let M be a closed subspace of a

Hilbert space, H. Then H = M ⊕ M ⊥ .
Proof. The proof is an application of the Projection Theorem. Let x ∈ H be

arbitrary. By the Projection Theorem, there exists a unique vector m∗ ∈ M such
that
kx − m∗ k ≤ kx − mk for all m ∈ M and n∗ := x − m∗ ∈ M ⊥ .
Consequently, we can write
x = m∗ + (x − m∗ ) = m∗ + n∗ , where n∗ := x − m∗ ,
with m∗ ∈ M and n∗ ∈ M ⊥ . It remains now to show that this representation

is unique. Suppose that x = m1 + n1 with m1 ∈ M and n1 ∈ M ⊥ is another
representation of x. Then,
m∗ + n∗ = m1 + n1 so that m1 − m∗ + (n1 − n∗ ) = 0.
But (m1 − m∗ ) and (n1 − n∗ ) are orthogonal. Hence, by Pythagoras Theorem,
0 = k(m1 − m∗ ) + (n1 − n∗ )k2 = km1 − m∗ k2 + kn1 − n∗ k2 .
This implies , km1 − m∗ k = 0 and kn1 − n∗ k = 0, i.e., m1 = m∗ and n1 = n∗ ,

establishing the uniqueness of the representation. Thus, H = M ⊕ M ⊥ .
¤
1 1
In Proposition 2.21, we proved that the dual of lp is lq where p + q = 1. By taking
p = 2, it follows that the dual of l2 is l2 i.e., l2∗ = l2 . Since l2 is a Hilbert space, it
follows that the dual of a Hilbert space H is itself. We give a direct proof of this
fact here and obtain from the proof, the celebrated Riesz Representation Theorem.
Proposition 6.31 Let H be a Hilbert space. Then H ∗ = H, where H ∗ denotes the

dual of H.
92
Proof. As in Proposition 2.21, the proof is divided into 2 parts.
Part 1. We prove that every element y ∈ H defines an element fy of H ∗ with

the same norm (i.e., kfy k = kyk).
Part 2. We prove that every f ∈ H ∗ defines a unique vector y ∈ H with the

same norm (kf k = kyk). Hence H and H ∗ are isometric.
We begin with Part 1. Let y ∈ H be given. For arbitrary x ∈ H define fy : H /K

by fy (x) = hx, yi. Clearly, fy is linear. Moreover, |fy (x)| = |hx, yi| ≤ kxk · kyk so
that fy is bounded. Hence fy ∈ H ∗ . Furthermore, the last inequality also yields
that
kfy k ≤ kyk. (6.1)
If y = 0, then from (6.1) we have ||fy || ≤ ||y|| = 0. In this case, fy = 0 and we

y
are done. We may suppose that y 6= 0. Take x = ||y|| , then we have ||x|| = 1, and
fy (x) = ||y||. This shows that
|fy (x)| = ||y|| ≤ ||fy ||.||x|| = ||fy ||, i.e.,||y|| ≤ ||fy ||. (6.2)
From (6.1) and (6.2), we have that ||fy || = ||y||.
Part 2. Let f ∈ H ∗ . Consider the kernel of f , Kerf defined by Kerf = {u ∈ H :

f (u) = 0}. Set K := Kerf . K is a closed subspace of H (Exercises 2.1, Problem 7).
By Theorem 6.30, every element x ∈ H can be written uniquely as x = w + z where
w ∈ K and¡ z¢ ⊥ K. Let x 6∈ K, then x = w∗ + z where z 6= 0 and f (z) = δ 6= 0.
Let x1 = zδ . Then, f (x1 ) = 1. Hence, for arbitrary u ∈ H, u ∈ / K, we have
that f (u) = α ⇒ f (u) = αf (x1 ) i.e., f (u − αx1 ) = 0. Let u − αx1 = w0 . Hence,
u = w0 + αx1 , w0 ∈ K and αx1 ⊥ K so that
hu, x1 i = hw0 + αx1 , x1 i = hw0 , x1 i + αhx1 , x1 i = αkx1 k2 .
Hence, f (u) = α = hu, kxx11k2 i so that f (u) = hu, y0 i ∀ u ∈ H\K where y0 = kxx11k2 .
But this also holds if u ∈ K. Hence it holds for all u ∈ H. Moreover by part 1, we
have that ||f || = ||y0 ||. Finally, it only remains to show that y0 is unique. But, if
f (x) = hx, y ∗ i ∀ x ∈ H then
hx, y0 i = hx, y ∗ i ∀ x ∈ H so hx, y0 − y ∗ i = 0 ∀ x ∈ H.
Take x = y0 − y ∗ . Then hy0 − y ∗ , y − y ∗ i = ky0 − y ∗ k2 = 0 ⇒ y0 = y ∗ . The proof is

complete. ¤
93
6.7 The Riesz Representation Theorem
In part 2 of proposition 6.31, we actually proved an important theorem in Hilbert
spaces called the Riesz Representation Theorem. This theorem states that any
bounded linear functional on a Hilbert space can be represented as an inner product
with a unique vector in H. More precisely, we proved the following theorem.
Theorem 6.32 (Riesz Representation Theorem) Let H be a Hilbert space and let
f be a bounded linear functional on H. Then,
(i) There exists a unique vector y0 ∈ H such that
f (x) = hx, y0 i for each x ∈ H; (6.3)
(ii) Moreover, kf k = ky0 k.
6.8 Complete Orthonormal Sets

Definition 6.33 An orthonormal set {uα }α∈∆ in an inner product space E is called
complete if x ⊥ uα for all α ∈ ∆ implies x = 0 (∆ denotes the arbitrary index set).
The set {uα }α∈∆ is said to form a basis for E if for arbitrary x ∈ E, x can be written
as follows: X
x= hx, uα iuα .
α∈∆
Lemma 6.34 Let E be an inner product space. Then an orthonormal set S =

{uα }α∈∆ is a basis for E if and only if S is complete (in the sense of Definition
6.33).
Proof. (⇒) Let S = {uα }α∈∆ be a basis. LetPx ∈ E and x ⊥ s for each s ∈ S. We
want to show x = 0. But S is a basis ⇒ x = hx, uα iuα and x ⊥ S ⇒ x ⊥ uα for
α∈∆
all uα ∈ S. This implies hx, uα i = 0 for each uα . Hence x = 0.
(⇐) Assume now that S = {uα }α∈s is complete. We want to show {uα }α∈∆ is a
basis. Now, for every x ∈ E,
* +
X
x− hx, uα iuα , uα = 0 (verify, see Exercises 6.1, Problem 18).
α
94
P
Hence, since S is complete, we must have x − α hx, uα iuα = 0, which yields the
desired result.
¤
Lemma 6.35 Let E be an inner product space. Then S = {uα }α∈∆ is a basis for
E if and only if for arbitrary x ∈ E,
X
|hx, uα i|2 = kxk2 .
α∈∆
P
Proof. S is a basis for E if and only if x = hx, uα iuα for arbitrary x ∈ E, if
α∈∆
and only if DX X E
2
kxk = hx, uα i uα , hx, uβ iuβ ,
α∈∆ β∈∆
if and only if Ã !
X X
kxk2 = hx, uα i hx, uβ i huα , uβ i,
α∈∆ β∈∆
if and only if Ã !
X X
kxk2 = hx, uα i hx, uβ i huα , uβ i,
α∈∆ β∈∆
if and only if
X
kxk2 = hx, uα i hx, uα i, since huα , uβ i = 0 for α 6= β,
α∈∆
if and only if X
kxk2 = |hx, uα i|2 , as desired.
α
¤
P
Remark 6.36 The identity: |hx, uα i|2 = kxk2 is called Parseval’s Identity (or
P
α∈∆
equation). The expression x = hx, uα iuα is called the Fourier expansion of x and
α∈∆
the numbers hx, uα i are called the Fourier coefficients.
Example 6.37 Let H = L2 [0, 2π]. Recall that for arbitrary f, g ∈ H,

Z 2π µZ 2π ¶1/2
2
hf, gi = f (x) g(x) dx and kf k = |f (x)| dx .
0 0
95
Consider S = {einx , n = 0, ±1, ±2, . . .}, where i2 = −1. Then,
Z 2π ½
inx −imx 0, m 6= n
e e dx =
0 2π, m = n.
Thus, if we define
1
un = √ einx , n = 1, 2, ...,
2π
then S := {un }∞
n=1 is an orthonormal set in L2 [0, 2π]. For any f ∈ L2 [0, 2π], the
numbers Z 2π
1
αn = hf, un i = √ f (x) e−inx dx
2π 0
are its Fourier coefficients; the Bessel’s inequality takes the form:
∞
X Z 2π
2
|αn | ≤ |f (x)|2 dx.
−∞ 0
Z 2π
P
∞
In fact, S is a complete orthonormal set in L2 [0, 2π] (so that 2
|αn | = |f (x)|2 dx).
−∞ 0
Moreover, each f ∈ L2 [0, 2π] has Fourier expansion given by
∞ Z 2π X
∞ Z 2π
1 X inx 1
f (x) = √ αn e = f (x + y) e−iny dy.
2π −∞ 2π 0 −∞ 0
P
∞
Remark 6.38 The Fourier expansion f (x) = αn un (x) does not mean that the
−∞
RHS converges pointwise to f . The meaning of the expansion is that the partial sum
n
X
of the series fn (x) := αn un (x) converges to f in the sense that kfn − f kL2 [0,2π]
−n
/ 0 as n / ∞.
EXERCISES 6.1
1. Verify that the spaces defined in Examples 6.2 to 6.6 are inner product spaces.
96
2. Compute the `2 inner product of x and y where
½ ¾ ½ ¾
1 1 1 1 1 1
x = 1, , , , . . . and y = , , ,... .
2 3 4 3 4 5
3. Let E = `2 . Check if kx + yk = kxk + kyk where

¡ ¢
(a) x = 0, 12 , 14 , 18 , . . . , y = (1, 0, 0, . . .);
¡ ¢
(b) x = 1, 13 , 19 , 271
, . . . , y = 3x.
© ª∞ © 1 ª∞
4. Consider the sequences x = n1 n=1 and y = n+1 n=1
.
(a) Verify that x and y are in `2 .

(b) Compute `2 inner product of x and y.
5. Consider the functions,

f (t) = sin 2t; g(t) = cos 3t,
h(t) = sin4 t; m(t) = sec2 t.
(a) Verify that f, g, h and m are elements of L2 [0, 2π].

(b) Compute the L2 [0, 2π] inner products:
(i) hf, gi and (ii) hh, mi.
6. Let Pn = {p = p(t), polynomial of degree less than or equal to n over the

interval [a, b]}. Define the function h, i on Pn × Pn by
Z b
hp, qi = p(t) q(t) dt .
a
(a) Prove that h, i is an inner product on Pn .

3
(b) Compute hf, gi where f (t) = 1 − 2
t2 , g(t) = 3t2 .
7. Let (i) x(t) = t2½, y(t) = −t2 , ½

2, 0 ≤ t ≤ 14 0, 0 ≤ t ≤ 41
(ii) x(t) = y(t) =
0, 14 < t ≤ 1 2, 1
4
< t ≤ 1.
(a) Verify that x and y are in L2 [0, 1].
(b) Compute hx, yi in each of (i) and (ii).
(c) Check whether or not kx + yk2 = kxk2 + kyk2 .
97
8. Prove Lemma 6.10, Propositions 6.11 and 6.12.
9. Verify that the sets S defined in Examples 6.20 and 6.21 are orthonormal sets.
10. Prove Theorem 6.23 (Bessel’s inequality).
11. Give an example of an inner product space which is not a Hilbert space. Prove
that your example has the desired properties.
12. Prove that if H is a Hilbert space, for any subspace S of H, S ⊥ is a closed

subspace of H.
Use Theorem 6.15 to solve problems 13, 14 and 15.
13. Verify that C[0, 1] with “sup” norm is not an inner product space.
14. Verify that `p , 1 < p < ∞, p 6= 2, is not an inner product space.
15. Is IRn with k(x1 , x2 , . . . , xn )k∞ = max |xi |, an inner product space? Justify
1≤i≤n
your answer.
Z 1
16. Let E = C[0, 1] with hf, gi = f (t) g(t) dt.
0
Consider the functions f (t) = sin πt + i sin πt, g(t) = − sin 2πt + i sin 3πt. Are
the functions f and g orthogonal? Justify your answer.
17. Prove that if K is a closed convex subset of a Hilbert space H, then for each
x ∈ H there exists a unique x∗ ∈ K such that
kx − x∗ k = inf kx − uk .
u∈K
Hint: Use the parallelogram law and follow the argument of Theorem 6.26.
18. VerifyPthe assertion made in the proof of Lemma 6.34 that for every x ∈ E,
hx − hx, uα iuα , uα i = 0 where {uα }α∈∆ is an orthonormal set.
α
19. (a) Define Ti : `2 / IR by
Ti (x1 , x2 , . . . , xn ) = xi ∀ (x1 , x2 , . . .) ∈ `2 .
(i) Prove that Ti is a bounded linear function on `2 for each i.

(ii) Compute the unique vector y0 guaranteed by the Riesz Representa-
tion Theorem (Theorem 6.32).
98
(b) Define T : L2 [0, 2π] / IR by
Z 2π
(T f )(t) = f (t)dt ∀ f ∈ L2 [0, 2π].
0
(i) Prove that T is a bounded linear map.

(ii) Compute the y0 of Theorem 6.32.
*20. Let E = C[0, 1], with “sup” norm. Let K consist of all f ∈ E such that
Z 1/2 Z 1
f (s)ds − f (s)ds = 1.
0 1/2
Prove that K is a closed convex subset of E which contains no element of

minimum norm.
*21. Let K be the set of all f ∈ L1 ([0, 1]), relative to Lebesgue measure, such that
Z 1
f (s)ds = 1.
0
Prove that K is a closed convex subset of L1 ([0, 1]) which contains infinitely
many elements of minimum norm.
99
100
CHAPTER 7
Operators on Hilbert Spaces
In this chapter we examine some important operators defined on Hilbert spaces.
Definition 7.1 Let H be a Hilbert space. Let A : H → H be a bounded linear

map. We define a map
A∗ : H → H
by the relation
hAx, yi = hx, A∗ yi
for all x, y ∈ H. The map A∗ is called the adjoint of A.
Observe that the existence of A∗ is a consequence of the Riesz Representation The-

orem. To see this, consider the map f : H /C
I defined for each x ∈ H by f (x) =
hAx, yi. Then f is a bounded linear functional on H. By Riesz Representation
Theorem, there exists a unique y0 ∈ H such that hAx, yi = f (x) = hx, y0 i ∀ x ∈ H.
We can now define A∗ : H / H by A∗ y = y0 (which is unique).
The following theorem is fundamental for adjoint operators on Hilbert spaces.
Theorem 7.2 Let A, A1 , A2 : H → H be bounded linear maps with adjoints

A∗ , A∗1 , A∗2 respectively. Then,
(a) (A1 + A2 )∗ = A∗1 + A∗2 ;
(b) (αA)∗ = ᾱA∗ , where α is a complex scalar;
(c) (A1 A2 )∗ = A∗2 A∗1 , (Note that the order is reversed);
101
(d) (A∗ )∗ = A;
(e) ||A|| = ||A∗ ||;
(f) ||A∗ A|| = ||A||2 .
Proof. We verify (e) and (f) and leave the others as easy exercises. For (e), from
the relation hx, A∗ yi = hAx, yi, by setting x = A∗ y we obtain
||A∗ y||2 = hAA∗ y, yi ≤ ||AA∗ y||.||y||,
and using the boundedness of A, this yields for all y ∈ D(A∗ ),
||A∗ y||2 ≤ ||A||.||A∗ y||.||y||,
so that
||A∗ y|| ≤ ||A||.||y||. (7.1)
Observe that if A∗ y = 0 for all y, then A∗ = 0 and (e) follows.

Inequality (7.1) yields that A∗ is bounded and also (by Remark 2.15) that
||A∗ || ≤ ||A||. (7.2)
Applying (7.2) to A∗ , we have ||(A∗ )∗ || ≤ ||A∗ ||. Using (d), we now have
||A|| ≤ ||A∗ ||. (7.3)
Inequalities (7.2) and (7.3) yield (e).

For (f), using (e), we obtain,
||A∗ A|| ≤ ||A∗ ||.||A|| = ||A||2 . (7.4)
Moreover, for each x ∈ D(A),
||Ax||2 = hAx, Axi = hA∗ Ax, xi ≤ ||A∗ A||.||x||2 ,

1
so that ||Ax|| ≤ (||A∗ A||) 2 ||x|| which yields
||A||2 ≤ ||A∗ A||. (7.5)
(7.4) and (7.5) yield the desired result.

¤
102
7.1 Self-adjoint, Normal and Unitary Operators
Definition 7.3 Let H be a Hilbert space. An operator T : H → H is called:
(i) Self-adjoint or Hermitian if T = T ∗ ;
(ii) Normal if T ∗ T = T T ∗ ;
(iii) Unitary if T T ∗ = T ∗ T = I.
7.1.1 Self-adjoint operators

Let B(H) denote the space of all bounded linear operators on H. The following
theorem is fundamental for self-adjoint operators.
Theorem 7.4 The collection of self-adjoint operators on H forms a closed, real
linear subspace of B(H).
Proof. The set of self-adjoint operators in B(H) is closed under addition and scalar
multiplication by real numbers (Verify). Thus, they form a subspace. We prove that
this subspace is closed. Let {Tn } be a sequence of self-adjoint operators on H such
that Tn → T . It suffices to prove T = T ∗ . But
||T − T ∗ || ≤ ||T − Tn || + ||Tn − Tn∗ || + ||Tn∗ − T ∗ ||
= ||T − Tn || + ||Tn∗ − T ∗ ||, since Tn∗ = Tn
= ||T − Tn || + ||(Tn − T )∗ || = ||T − Tn || + ||T − Tn || = 2||Tn − T || → 0
as n → ∞ and so T = T ∗ .
¤
For our next theorem, we shall need the following lemma.
Lemma 7.5 If T is a bounded linear operator on H, then hT x, xi = 0 for all x ∈ H
if and only if T = 0.
Proof. If T = 0, the result is trivial. Conversely, suppose hT x, xi = 0 for all x in
H. Then, for arbitrary scalars α, β; and arbitrary vectors x, y ∈ H,
0 = hT (αx + βy), αx + βyi − |α|2 hT x, xi − |β|2 hT y, yi
= αβ̄hT x, yi + β ᾱhT y, xi. (7.6)
Set α = β = 1 in (7.6) to obtain
0 = hT x, yi + hT y, xi.
Set α = i, β = 1 in (7.6) to obtain, 0 = hT x, yi − hT y, xi. Thus, 2hT x, yi = 0 i.e.,
hT x, yi = 0 for all y. Now set y = T x to get ||T x|| = 0 for all x which implies
T x = 0 for all x, i.e., T = 0, as required.
¤
For self-adjoint operators, we also have the following theorem.
103
Theorem 7.6 Let T : H → H be a bounded linear operator on a complex Hilbert
space, H. Then, T is self-adjoint if and only if hT x, xi is real.
Proof. Let T be self-adjoint. It suffices to prove hT x, xi = hT x, xi where the bar

indicates complex conjugation. Now, since T is self-adjoint, we have:
hT x, xi = hx, T ∗ xi = hT x, xi
and the result follows. Conversely, let hT x, xi be real. Then,
hT x, xi = hT x, xi = hT ∗ x, xi
so that
hT x − T ∗ x, xi = 0 or h(T − T ∗ )x, xi = 0,
for all x ∈ H. By Lemma 7.5, T − T ∗ = 0 i.e., T = T ∗ .

¤
Remark 7.7 If T : H → H is an arbitrary bounded linear operator on H, we can

always obtain a self-adjoint operator from T in the following way
1 1
T = (T + T ∗ ) + i[ (T − T ∗ )] = A + iB,
2 2i
where A = 12 (T + T ∗ ) and B = 1
2i
(T − T ∗ ) are self-adjoint operators.
Remark 7.8 Let T : H → H be a bounded linear map. We have already proved

that ||T || = ||T ∗ ||. This, however, does not imply, in general, that ||T x|| = ||T ∗ x||
for every x ∈ H. To see this, consider the following example:
Example 7.9 Let T : l2 → l2 be defined by
T (x1 , x2 , x3 , ...) = (0, x1 , x2 , x3 , ...).
Then, ||T ∗ || = ||T ||, (already proved for all bounded linear operators on H). Take
x = (1, 0, 0, ...) ∈ l2 . Verify that ||T x|| 6= 0 whereas ||T ∗ x|| = 0. (See Exercises 7.1,
Problem 9).
104
7.1.2 Normal Operators
Proposition 7.10 Let T : H → H be a bounded linear map. If T is normal, then
||T x|| = ||T ∗ x||, for each x ∈ H.
Proof. Let T be normal. Then,
||T x||2 = hT x, T xi = hx, T ∗ T xi = hx, T T ∗ xi = hT ∗ x, T ∗ xi = ||T ∗ x||2
and the result follows.

¤
For normal operators, we also have the following fundamental result.
Theorem 7.11 The set of all normal operators on H is a closed subset of B(H)
which contains the set of all self-adjoint operators, and is closed under scalar mul-
tiplication.
Proof. It is obvious that every self-adjoint operator is normal and that if α is

a scalar, then αT is normal whenever T is. We now show that the limit of any
convergent sequence {Tn } of normal operators is normal. Indeed, if Tn → T then,
Tn∗ → T ∗ . (Verify). Since {Tn } is a sequence of normal operators, we have for each
n, Tn Tn∗ − Tn∗ Tn = 0. Hence, ||T T ∗ − T ∗ T || ≤ ||T T ∗ − Tn Tn∗ || + ||Tn Tn∗ − Tn∗ Tn || +
||Tn∗ Tn −T ∗ T || = ||T T ∗ −Tn Tn∗ ||+||Tn∗ Tn −T ∗ T || → 0 and so T T ∗ = T ∗ T , completing
the proof.
¤
7.1.3 Unitary Operators

Recall that an operator T on H is called unitary if T T ∗ = T ∗ T = I. Thus, a
unitary operator is obviously normal. Furthermore, unitary operators have inverses
and their adjoints are their inverses. We begin with the following theorem.
Theorem 7.12 Let T be an operator on a Hilbert space, H. Then, the following

are equivalent
(i) T ∗ T = I;
(ii) hT x, T yi = hx, yi;
(iii) ||T x|| = ||x|| for all x ∈ H.
Proof. (i) ⇒ (ii) For all x, y ∈ H,
hx, yi = hx, Iyi = hx, T ∗ T yi = hT x, T yi.
105
(ii) ⇒ (iii)
||T x||2 = hT x, T xi = hx, xi = ||x||2 ,
as required.
(iii) ⇒ (i) For all x ∈ H, ||x||2 = ||T x||2 ⇒ hx, xi = hT x, T xi ⇒ hx, xi =
hT ∗ T x, xi ⇒ h(I − T ∗ T )x, xi = 0
⇒ (by Lemma 7.5) (I − T ∗ T ) = 0 ⇒ T ∗ T = I as required.
¤
Remark 7.13 It is easy to see that an operator satisfying any of the conditions
(i)-(iii) is not necessarily unitary. It suffices, for example, to consider the right shift
operator on l2
T : l2 → l2
defined by T (x1 , x2 , x3 , ...) = (0, x1 , x2 , x3 , ..). Clearly ||T x|| = ||x||. Observe that if
y = (y1 , y2 , y3 , ...) ∈ l2 then, hT x, yi = h(0, x1 , x2 , ...), (y1 , y2 , y3 , ...)i = x1 y2 + x2 y3 +
x3 y4 + . . . = h(x1 , x2 , x3 , ...), (y2 , y3 , ...)i,
so that T ∗ y = T ∗ (y1 , y2 , y3 , ...) = (y2 , y3 , y4 , . . .). Now,
(T T ∗ )(y) = T (y2 , y3 , ...) = (0, y2 , y3 , ...) 6= Iy,
so T T ∗ 6= I, and so T is not unitary.
Theorem 7.14 Let T : H → H be a bounded linear operator on a complex Hilbert
space, H. If T is an isometry, then T ∗ T = I.
Proof. This follows from parts (i) and (iii) of Theorem 7.12.
¤
We now give the following characterization of unitary operators:
Corollary 7.15 Let T : H → H be a bounded linear operator on a complex Hilbert
space. Then, T is unitary if and only if T and T ∗ are both isometries.
Proof. Applying Theorem 3.50, the proof follows immediately.
¤
Remark 7.16 In the light of the results of this subsection, we deduce that unitary
operators are those that preserve all the structures of H.
• Set theoretic structure is preserved because they are bijective.
• Linear structure is preserved because they are linear maps.
• Norm structure is preserved because they are isometries.
• Inner product structure is preserved because hT x, T yi = hx, yi for all x, y ∈ H.
106
EXERCISES 7.1
1. Prove that 0∗ = 0 and I ∗ = I. Use the latter to show that if T is nonsingular,

then T ∗ is also nonsingular and in this case
(T ∗ )−1 = (T −1 )∗ .
2. Prove parts (a), (b), (c) and (d) of Theorem 7.2.
3. If T is an arbitrary bounded linear operator on H and if α and β are scalars

such that |α| = |β|, prove that αT + βT ∗ is normal.
4. Prove that the set of unitary operators on H is a group.
5. If T is a normal operator on a Hilbert space, H, prove that (λI − T ) is also a

normal operator, where λ is a scalar and I is the identity operator on H.
6. If T1 , T2 are normal operators and T1 T2∗ = T2∗ T1 , prove

(i) T1∗ T2 = T2 T1∗ , (ii) T1 T2 , T2 T1 , T1 + T2 are normal operators.
7. Let T : H → H be a bounded linear operator. Write T in the form

1 1
T = (T + T ∗ ) + i[ (T − T ∗ )] = A + iB.
2 2i
Prove:
(i) A and B are self-adjoint;
(ii) T is normal if and only if AB = BA.
8. A self-adjoint operator is called positive if hAx, xi ≥ 0 for all x ∈ H. Prove

that T T ∗ and T ∗ T are positive operators.
9. Verify that kT x0 k 6= 0 whereas kT ∗ x0 k = 0, where T is the map defined in

Example 7.9 and x0 = (1, 0, 0, . . .).
107
108
CHAPTER 8
Weak and Weak∗ Topologies
8.1 Notions from general topology

Most of the materials of this chapter are basically adaptations from Brezis [4]. Let
us begin by recalling some notions of general topology which we shall need. Let E
be a set and let {Yi }i∈I be a family of topological spaces. (Here and in the rest of
this chapter I or ∆ denotes an arbitrary index set). For each i ∈ I we shall associate
a map,
φ i : E → Yi .
Our problem of interest is to find how to endow E with a smallest topology such that
the maps φi , i ∈ I are continuous. More precisely, we want to construct a topology
τ on E with minimum open sets which makes {φi }i∈I continuous.
Observe that if E is endowed with the discrete topology (i.e., the topology in which
every subset of E is open) then each φi is continuous (Why ?). Of course, this
topology is far from being the minimum. It is, in fact, the largest.
Let ωi ∈ Yi be open, if we are able to find the topology τ , then φ−1

i (ωi ) is necessarily
open for the topology τ . (Why?). When ωi are open in Yi , i ∈ I, the sets φ−1 i (ωi )
constitute a family of subsets of E which are necessarily in the topology τ ; denote
this family by U = {Ui }i∈I , where Ui = φ−1i (ωi ) for some i ∈ I. Thus, the topology
we are looking for must necessarily contain inverse images (under φi ) of all open sub-
109
sets of Yi ’s. Having noted this fact we now ask the following question: Is the family
U = {Ui }i∈I a topology on E, and if it is, is it the smallest topology for which all φi
are continuous? The answer to this question is implicit in the following construction.
The Construction of τ
To construct the desired topology, we consider first, finite intersections of mem-
bers of U . In this way, we obtain a family Φ of subsets of E closed under finite
intersections. We then consider the family of arbitrary unions of members of Φ.
This family is clearly closed under arbitrary unions. We shall denote it by τ . It is,
however, no longer evident that the family τ is closed under finite intersection. We
state this as a lemma.
Lemma 8.1 The family τ is closed under finite intersections.
Proof. Exercise.
Remark 8.2 We cannot reverse the order of the operations in the above construc-
tion of τ . It was more natural to begin by considering arbitrary unions of subsets of
U and then after that to take finite intersections. If we follow this order, the family
obtained in this way is, of course, closed under finite intersection. It turns out that
it is not closed under arbitrary unions.
Let us now endow E with the topology τ constructed above and prove two ele-
mentary but important properties of this topology. Before we do this, however, we
make the following remark.
Remark 8.3 Recall that the open sets of the topology τ are obtained by considering
first the finite intersection of sets of the form φ−1
i (ωi ), ωi open in Yi and then taking
arbitrary unions of members of the resulting set. Thus, given a point x0 ∈ E we
obtain a neighborhood base at x0 for the topology τ by considering sets of the form
\
φ−1
i (Vi )
finite
where Vi is a neighborhood of φi (x0 ) in Yi .
Proposition 8.4 Let {xn } be a sequence in E (with topology τ ). Then, xn → x in

E if and only if φi (xn ) → φi (x) for each i ∈ I.
Proof. (⇒) Let xn → x. By continuity of φi , the result follows trivially.

(⇐) Let φi (xn ) / φi (x) as n / ∞, ∀ i ∈ I. We want to prove xn / x. So,
110
let U be a neighborhood of x. We want to produce an N such that for all n ≥ N,
we have xn ∈ U . Since E is endowed with the topology τ we may suppose that U is
of the form \
U= φ−1
i (Vi ), J ⊂ I,
i∈J
J finite, where Vi is open in Yi and contains φi (x). For each i ∈ J, since

φi (xn ) / φi (x) as n / ∞, there exists an integer Ni such that for all n ≥ Ni ,
φi (xn ) ∈ Vi .
Let
N = max Ni .
i∈J
We then have, for all n ≥ N , and for each i ∈ J, φi (xn ) ∈ Vi , i.e., for all i ∈ J and
for all n ≥ N we have xn ∈ φ−1
i (Vi ). This implies
\
xn ∈ φ−1
i (Vi ) = U
i∈J
for all n ≥ N . Since U was arbitrary, it follows that xn → x. The proof is complete.
¤
Proposition 8.5 Let E be a topological space with the topology τ and let Z be an
arbitrary topological space. Let ψ be a mapping of Z into E (i.e., ψ : Z → E).
Then, ψ is continuous if and only if φi ◦ ψ is continuous from Z onto Yi , i ∈ I.
ψ
(Z, any Top.) / (E, ω)
NN
NN
NN
NN
NN φi
φi ◦ψ N N
NN
NN
NN ²
&
Yi
Fig.8.1
Proof. (See Fig. 8.1) (⇒) If ψ is continuous, then clearly φi ◦ ψ is also continuous
for all i ∈ I.
(⇐) Conversely, we want to prove ψ is continuous. Let U be an arbitrary open
111
subset of E. It suffices to show that ψ −1 (U ) is open in Z. By hypothesis, φi ◦ ψ is
continuous. We note that U is of the form
[ ³ \ ´
U= φ−1
i (ωi )
arbitrary finite
³ST ´
−1 −1 −1
with ωi open in Yi . Consequently, ψ (U ) = ψ φi (ωi )
S T −1
= (φi ◦ ψ) (ωi ) which is open in Z since φi ◦ ψ is continuous and ωi
arbitrary finite
is open in Yi . Hence ψ is continuous.
¤
8.2 The weak topology

Let E be a Banach space. For each f ∈ E ∗ , we associate a map
φf : E → IR
defined by φf (x) = hf, xi = f (x) for all x ∈ E. As f ranges over E ∗ , we obtain a

family {φf }f ∈E ∗ of maps of E into IR.
Definition 8.6 The weak topology on E (denoted by ω) is the smallest topology

on E which makes the maps φf continuous.
Remark 8.7 We shall use the notation hf, xi and f (x) interchangeably to denote
the action of f on x.
We observe that if we take Yi = IR for each i, and I = E ∗ , then the weak topology
ω on E becomes a special case of the topology τ defined earlier.
We now restate Proposition 8.4 when E is endowed with the weak topology as
follows:
Proposition 8.8 Let (E, ω) denote a Banach space endowed with the weak topology.
Then xn / x in (E, ω) if and only if φf (xn ) / φf (x), i.e., if and only if f (xn )
/ f (x) for each f ∈ E ∗ .
Proposition 8.8 is used in several books on functional analysis as the definition of

weak topology.
Proposition 8.9 The weak topology is Hausdorff.
112
Proof. Let x1 , x2 , ∈ E, x1 6= x2 . We shall
T construct U1 , U2 open in the weak topology
such that x1 ∈ U1 , x2 ∈ U2 and U1 U2 = ∅. Observe that x1 6= x2 implies that
the sets {x1 } and {x2 } are disjoint. Clearly, these two sets are also convex, closed
and nonempty. Furthermore, they are compact. By Hahn Banach Theorem (Second
Geometric Form) there exists a closed hyperplane which separates {x1 } and {x2 } in
the strict sense. Hence, there exist f ∈ E ∗ and α ∈ IR such that
f (x1 ) = hf, x1 i < α < hf, x2 i = f (x2 ).
We now set
U1 = {x ∈ E : hf, xi < α} = φ−1
f {(−∞, α)}.
U2 = {x ∈ E : hf, xi > α} = φ−1

f {(α, ∞)}.
T
Clearly, U1 , U2 are open in the weak topology; x1 ∈ U1 , x2 ∈ U2 and U1
U2 = ∅.
¤
Notation. For a sequence {xn } in E we shall denote by xn * x the convergence of
xn to x in the weak topology.
Proposition 8.10 Let x0 ∈ E; we obtain a neighborhood base at x0 for the weak
topology by considering sets of the form
V = {x ∈ E : |hfi , x − x0 i| < ², i ∈ J}
where J is finite, fi ∈ E ∗ and ² > 0.

T −1
Proof. Claim: V = φfi (ai − ², ai + ²) with ai = hfi , x0 i, is an open set
i∈J
in the weak topology and contains x0 . The proof of the claim follows easily since
u ∈ V iff |hfi , u − x0 i| < ² for all i ∈ J, iff ai − ² < fi (u) < ai + ² for all
i ∈ J, iff fi (u) ∈ T(ai − ², ai + ²) for all i ∈ J, iff φfi (u) ∈ (ai − ², ai + ²) for
all i ∈ J, iff u ∈ i∈J φ−1 fi (ai − ², ai + ²). Clearly, x0 ∈ V and V is open in the
weak topology. Let U be a neighborhood of x0 for the weak topology. We know
that there exists a neighborhood U1 of x0 , U1 ⊂ U , U1 is of the form
\
U1 = φ−1
f (ωj ),
j∈J
where J is finite and ωj is a neighborhood of hfi , x0 i = ai in IR. Hence, there exists

² > 0 such that (ai − ², ai + ²) ⊂ ωj for each i ∈ J so that
x0 ∈ V ⊂ U1 ⊂ U.
113
Proposition 8.11 Let {xn } be a sequence in E. We have the following results.
(i) xn * x ⇔ f (xn ) → f (x) for each f ∈ E ∗ ;
(ii) xn → x ⇒ xn * x;
(iii) xn * x ⇒ {xn } is bounded and ||x|| ≤ lim inf ||xn ||;
(iv) xn * x (in E), fn → f (in E ∗ ) ⇒ fn (xn ) → f (x) (in IR).
Proof. (i) This is Proposition 8.8.
(ii) To prove that xn * x it suffices to prove that hf, xn i → hf, xi for each
f ∈ E ∗ . But for each f ∈ E ∗ we have:
|hf, xn i − hf, xi| = |hf, xn − xi| ≤ ||f ||.||xn − x|| → 0
since xn → x as n → ∞. Hence, f (xn ) → f (x) for each f ∈ E ∗ .
(iii) Here we shall make use of Corollary 4.7. So, it suffices to prove that for
each f ∈ E ∗ , the sequence {f (xn )} is bounded. Since xn * x we have (by Part (i)):
for each f ∈ E ∗
hf, xn i → hf, xi
as n → ∞. Hence, {f (xn )} is bounded for each f ∈ E ∗ and by Corollary 4.7, this
implies {xn } is bounded. Now, since f ∈ E ∗ we have
|hf, xn i| ≤ ||f ||.||xn || . (8.1)
Taking “lim inf” in (8.1) we obtain: |hf, xi| ≤ ||f ||.lim inf ||xn ||. By Exercises 3.1,
Problem 6, we obtain
||x|| = sup |hf, xi| ≤ sup (||f ||.lim inf ||xn ||) ≤ lim inf ||xn ||,
||f ||≤1, f ∈E ∗ ||f ||≤1, f ∈E ∗
which completes proof of part (iii).
(iv) We have: |hfn , xn i − hf, xi| ≤ |hfn − f, xn i| + |hf, xn − xi| ≤ ||fn − f ||.||xn || +
|hf, xn − xi|. Since xn * x we have (by (iii)) that there exists a > 0 such that
||xn || ≤ a for all n ≥ 0 and since xn * x, we also have hf, xn − xi → 0. Hence
|hfn , xn i − hf, xi| ≤ ||fn − f ||a + |hf, xn − xi| → 0
as n → ∞, so that fn (xn ) → f (x) as n → ∞.

¤
114
Remark 8.12 Proposition 8.11 (ii) shows that strong convergence implies weak
convergence. The converse, however, is false, i.e., weak convergence does not always
imply strong convergence. We give the following example.
Example 8.13 Let H be a real Hilbert space and define the sequence {en } in H
by
en = (0, 0, 0, ...0, 1, 0, ...)
with 1 in the nth position and zero in every other position. By Riesz Representation
Theorem, for each f ∈ H ∗ ,
f (en ) = hz, en i
for some unique z ∈ H, and by Bessel’s inequality,
∞
X
|hz, en i|2 ≤ ||z||2 ,
n=1
P∞
and so n=1 |hz, en i| converges. Hence, limhz, en i = 0. But f (en ) := hz, en i → 0 for
each f ∈ H ∗ . Hence, en * 0 as n → ∞. However, {en } does not converge strongly
to zero since, for n 6= m, √
||en − em || = 2
and so {en } is not convergent (not even Cauchy!).
We now give an example in which strong and weak topologies coincide.
Proposition 8.14 Let E be a finite dimensional normed linear space. Then strong
and weak topologies coincide. (In particular, a sequence {xn } in E converges weakly
if and only if it converges strongly).
Proof. Let s denote the strong topology of E. The weak topology always has less
open sets than the strong topology. So, if a set is open in the weak topology then
it is open in the strong. Hence ω ⊂ s. Conversely, we must show that (in this case
where E is finite dimensional) a set which is open in the strong topology is also
open in the weak topology. Let the dimension of E be n. Let x0 ∈ E and let U be
a neighborhood of x0 in the strong topology. We shall construct a neighborhood V
of x0 in the weak topology such that x0 ∈ V ⊂ U . In particular, we shall find a set
{fi }i∈J , with J finite, in E ∗ and an ² > 0 such that
V = {x ∈ E : |hfi , x − x0 i| < ², for all i ∈ J}, (8.2)
is contained in U . Since U is a neighborhood of x0 in the strong topology, there

exists r > 0 such that B(x0 , r) ⊂ U . In fact, we shall construct V such that
115
x0 ∈ V ⊂ B(x0 , r) ⊂ U . We choose a basisP{ei }ni=1 of E with ||ei || = 1 for all i.
Each x ∈ E can be written uniquely as x = ni=1 αi ei . Define the maps
fi : E → IR
by
³X
n ´
fi (x) = fi αi ei = αi . i = 1, 2, ..., n.
i=1
Clearly, (i) fi is linear for each i; (ii) fi is bounded for each i. For, |fi (x)| = |αi | ≤
max1≤i≤n |αi | = ||x|| ∀ x ∈ E. So, {fi }ni=1 ∈ E ∗ and J = {1, 2, ..., n}. We take
² := nr (where, you recall, r is the radius of B(x0 , r)). We now show that with the
P
n
set {fi }ni=1 and this ², the set V defined in (8.2) is in B(x0 , r) ⊂ U . Let x0 = βi ei .
i=1
P
n
Then, for arbitrary x ∈ V , x = αi ei , we have
n=1
n
X n
X
||x − x0 || = || (αi − βi )ei || ≤ |αi − βi |.||ei ||
i=1 i=1
n
X Xn
= |fi (x) − fi (x0 )| = |hfi , x − x0 i|
i=1 i=1
n
X
< ² = ²n = r.
i=1
Hence, x ∈ B(x0 , r) ⊂ U . This completes the proof.

¤
8.2.1 Open and closed sets

Open (respectively closed) sets in the weak topology are also open (resp. closed) in
the strong topology. This is easy to see since the strong topology contains more open
sets than the weak. The converse, however, in not always true. In many infinite
dimensional spaces, there are open (respectively closed) sets in the strong topology
which are not open (resp. not closed) in the weak topology. We give some examples.
Example 8.15 The set S = {x ∈ E : ||x|| = 1} is clearly closed in the strong

topology (Prove this). We show it is not closed in the weak topology. To do this,
it suffices to show that S is not equal to its closure in the weak topology. Let the
closure of S in the weak topology be denoted by cl Sw . Thus, it suffices to produce
116
an x0 ∈ E, ||x0 || < 1, such that x0 ∈ cl Sw . Since such an x0 is not in S this will
complete the proof. So, let x0 ∈ E be such that ||x0 || < 1. Let TU be an arbitrary
neighborhood of x0 in the weak topology. It suffices to show U S 6= ∅. We can
always suppose that U is of the form
U = {x ∈ E : |hfi , x − x0 i| < ², i = 1, 2, 3, ..., n}
with ² > 0 and fi ∈ E ∗ . Fix y0 ∈ E, y0 6= 0 such that hfi , y0 i = 0, i = 1, 2, ..., n.

(Note that such a y0 exists. Otherwise, we can construct a map T : E → IRn by
T (u) = {hfi , ui} which is an isomorphism of E onto T (E) which will then imply
that dim(E) ≤ n). Define the function
h : [0, ∞) → IR
by h(t) = ||x0 + ty0 ||.
Then (see Fig. 8.2),

(i) h is continuous on [0, ∞);
(ii) h(0) < 1;
(iii) lim h(t) = +∞.
t→+∞
1010 h
1010 h(t)
010
101
1001
(0, 1)
010
101 t
0
10 Fig. 8.2
t0
Then there exists t0 > 0 such that ||x0 + t0 y0 || = 1. This implies x0 + t0 y0 ∈ S.

Observe also that if x := x0 + t0 y0 , then hfi , x −Tx0 i = hfi , t0 y0 i = t0 hfi , y0 i = 0
so that x = x0 + t0 y0 ∈ U . Hence, x0 + t0 y0 ∈ U S so that, since U is an arbitrary
117
neighborhood of x0 in the weak topology, we have proved that x0 ∈ cl Sw . This
completes the example.
Example 8.16 The set A = {x ∈ E : ||x|| < 1} is clearly open in the strong
topology (Prove this). It is not open in the weak topology. It suffices to prove that
A has empty interior. We do this by contradiction. Assume that the interior of A
is nonempty. Let x0 ∈ intA and U be a neighborhood (in the weak topology) of
x0 . As in Example 8.15 x = x0 + t0 y0 ∈ U ∩ S ⊂ A ∩ S, where S is as defined in
Example 8.15, contradiction.
Remark 8.17 If E is infinite dimensional, the weak topology on E is not metriz-

able. If, however, E is separable, one can construct a metric defined on B := {x ∈
E : ||x|| ≤ 1} and induce on B the same topology as the weak topology on E.
Remark 8.18 We have seen (Proposition 8.11(ii)) that if a sequence converges

strongly then it converges weakly. We also saw (Example 8.13) that a sequence
which converges weakly need not converge strongly. But if a Banach space is finite
dimensional, then the strong and weak topologies coincide (Proposition 8.14). This
fact is not confined only to finite dimensional spaces. In fact, there exists an infinite
dimensional Banach space on which weak convergence implies strong convergence.
In particular, l1 has this property.
8.2.2 Weak Topology and Convex Sets

We have seen that every closed set in the weak topology is also a closed set in the
strong topology i.e., a set which is weakly closed is also strongly closed. We also
saw that in several infinite dimensional spaces, the converse is false. We shall now
show that if a set is convex and strongly closed then it is weakly closed.
Theorem 8.19 Let K ⊂ E be convex and closed (in the strong topology). Then, K
is closed in the weak topology , (i.e., if K is strongly closed and convex, then it is
weakly closed).
Proof. The proof is another application of the Second Geometric Form of Hahn
Banach Theorem. It suffices to prove that K c (the complement of K) is open in
the weak topology. Let x0 ∈ K c . Then, x0 ∈
/ K. Since {x0 } is convex, nonempty
and compact and K is nonempty, convex and closed, it follows from Hahn Banach
Theorem (second geometric form) that there exist f ∈ E ∗ , α ∈ IR such that
f (x0 ) = hf, x0 i < α < hf, xi
118
for all x ∈ K. Set
U = {x ∈ E : hf, xi < α}.
T
Then x0 ∈ U, U K = ∅ so that U ⊂ K c . Moreover, U is open in the weak
topology. Hence, K c is open in the weak topology and so K is weakly closed.
¤
We now turn our attention to the weak star topology.
8.3 The Weak Star Topology

For each x ∈ E we consider the map
φx : E ∗ → IR
defined by
φx (f ) = hf, xi.
As x ranges over E we obtain a family of maps {φx }x∈E of E ∗ into IR.
The weak star topology is the smallest topology on E ∗ for which all the maps φx
are continuous.
Notation. We shall write ω–topology for weak topology and ω ∗ (or weak∗ )–topology
for weak star topology.
Proposition 8.20 The weak∗ topology is Hausdorff.
Proof. Let E ∗ be endowed with the weak star topology and let f1 , f2 be arbitrary
elements of E ∗ with f1 6= f2 . Then, there exists x ∈ E such that hf1 , xi 6= hf2 , xi.
Without loss of generality, let hf1 , xi < hf2 , xi. Introduce α ∈ IR such that
hf1 , xi < α < hf2 , xi.
Set
U1 = {f ∈ E ∗ : hf, xi < α} = φ−1
x ((−∞, α)).
U2 = {f ∈ E ∗ : hf, xi > α} = φ−1

x ((α, ∞)).
T
Then, U1 , U2 are open in the weak star topology, and f1 ∈ U1 , f2 ∈ U2 , U1 U2 =
∅.
¤
119
Proposition 8.21 Let f0 ∈ E ∗ . We obtain a neighborhood base at f0 for the w∗ −
topology by considering sets of the form:
V = {f ∈ E ∗ : |hf − f0 , xi i| < ²}
for all i ∈ J where J is finite , xi ∈ E, ² > 0.
Proof. Exercise (Use an argument similar to that used in the proof of Proposition
8.10).
ω∗
Notation. Given a sequence {fn } in E ∗ , we shall denote by fn −→ f the con-
vergence of fn to f in the weak star topology.
Proposition 8.22 Let {fn } be a sequence in E ∗ . Then,
ω∗
(a) fn −→ f iff hfn , xi → hf, xi for all x ∈ E;
ω∗
(b) fn * f ⇒ fn −→ f ;
/ f in E ∗ ⇒ fn −→ω∗
(c) fn f in E ∗ ;
ω∗
(d) fn −→ f ⇒ {fn } is bounded and ||f || ≤ lim inf ||fn ||;
ω∗ / x ⇒ hfn , xn i / hf, xi in IR.
(e) fn −→ f, xn
Proof. Exercise (Follow the argument similar to that used in the proof of Proposi-
tion 8.11).
8.4 The Banach Alaoglu Theorem

Theorem 8.23 (Banach Alaoglu Theorem) The set
BE ∗ = {f ∈ E ∗ : ||f || ≤ 1}
is compact in the w∗ − topology.
For a proof of this result (which is not our focus in this book) the reader may consult
Brézis [4]. We shall, however, make use of this theorem to prove one of our main
theorems in the next chapter.
EXERCISES 8.1
1. Prove Lemma 8.1.

120
CHAPTER 9
Reflexive and Uniformly Convex Spaces
We have already seen that given any normed linear space E, the space E ∗ of all
bounded linear functionals on E is a Banach space. As a Banach space, E ∗ has its
own dual space which we denote by (E ∗ )∗ or simply by E ∗∗ and often refer to as the
second conjugate space of E or the double dual or bidual of E.
9.1 Reflexive Spaces

There exists a natural mapping J : E → E ∗∗ of E into E ∗∗ defined, for each x ∈ E
by
J(x) = φx
where
φx : E ∗ → IR
is given by
φx (f ) = hf, xi,
for each f ∈ E ∗ . Thus,
hJ(x), f i ≡ hf, xi for each f ∈ E ∗ . (9.1)
We verify the following properties of J:

(i) J is linear (This is trivial and is left for the reader);
121
(ii) ||Jx|| = ||x|| for all x ∈ E; i.e., J is an isometry. In fact, for each f ∈ E ∗ ,
||Jx|| = sup |hJx, f i| = sup |hf, xi| = ||x||.

||f ||=1, f ∈E ∗ ||f ||=1, f ∈E ∗
(Explain why the last equality follows). In general, the map J need not be onto.
Consequently, we always identify E as a subspace of E ∗∗ . Since an isometry is always
injective, it follows that J is an isomorphism onto J(E) ⊂ E ∗∗ . The mapping J
defined above is called the canonical map (or canonical embedding) of E into E ∗∗ ,
and the space E is said to be embedded in E ∗∗ . This leads to the following definition.
Definition 9.1 Let E be a normed linear space and let J be the canonical embed-
ding of E into E ∗∗ . If J is onto, then E is called reflexive. Thus, a reflexive Banach
space is one in which the canonical embedding is onto.
Proposition 9.2 Let E be a finite dimensional normed linear space. Then the
strong, weak and weak star topologies coincide
Proof. The canonical map J is surjective in this case (since dim E = dim E ∗ =
dim E ∗∗ ). (See also Exercises 9.1, Problem 1).
¤
We now prove one of the main theorems of this chapter. We shall use the following
notation: for an arbitrary normed linear space X, we shall use BX to denote the
closed unit ball in X, i.e., BX = {x ∈ X : kxk ≤ 1}. In particular, E is a reflexive
normed linear space if and only if J(BE ) = BE ∗∗ .
9.1.1 A Theorem of Kakutani

We shall need the following lemmas whose proofs are given later in this section.
Lemma 9.3 (Goldstein’s Theorem) Let E be a Banach space. Then, J(BE ) is dense
in (BE ∗∗ , ω ∗ ).
Lemma 9.4 Let X and Y be Banach spaces and let T : (X, s) → (Y, s) be a linear
continuous map. Then,
T : (X, ω) → (Y, ω)
is continuous, and conversely, where s denotes strong topology and w denotes weak
topology.
122
The following theorem is an important characterization of reflexive spaces. It is, in
some sense, the infinite dimensional spaces analogue of the Heine Borel Theorem.
Theorem 9.5 (Kakutani’s Theorem) Let E be a Banach space. Then, E is reflexive

if and only if
BE = {x ∈ E : ||x|| ≤ 1}
is weakly compact (i.e., if and only if BE is compact in the weak topology of E).
Proof. (⇒) Suppose first that E is reflexive. Then, J(BE ) = BE ∗∗ . By the

Banach Alaoglu Theorem, BE ∗∗ is weak∗ compact. Thus, it now suffices to prove
that J −1 : (E ∗∗ , ω ∗ ) → (E, ω) is continuous, where E ∗∗ is endowed with the weak∗
topology and E is endowed with the weak topology. If this is done then BE would
be the image of a (weak∗ ) compact set under a continuous map. We shall use
Proposition 8.5 (see Fig. 9.1).
J −1 / (E, ω)
(E ∗∗ , ω ∗N)
NN
NN
NN
NN
N f
f ◦J −1 N N
NN
NN
NN ²
&
R
Fig.9.1
It suffices now to prove that for any f ∈ E ∗ , the map
f ◦ J −1 : (E ∗∗ , ω ∗ ) → IR
(defined by (f ◦ J −1 )(ψ) = hf, J −1 ψi) is continuous. But, for arbitrary ψ ∈ E ∗∗ ,
(f ◦ J −1 )(ψ) = hf, J −1 ψi = hψ, f i,
the last equality follows from equation (9.1), so that (f ◦J −1 ) is continuous. (Justify.
See Exercises 9.1, Problem 2).
(⇐) Suppose now BE is weakly compact. We prove E is reflexive. It suffices to prove
J(BE ) = BE ∗∗ . Since, J : (E, s) → (E ∗∗ , s) is an isometry, it is continuous, where
s denotes strong topology. By Lemma 9.4, this implies J : (E, w) → (E ∗∗ , w) is
continuous. This, in turn, implies J : (E, w) → (E ∗∗ , w∗ ) is continuous (Prove this.
123
See Exercises 9.1, Problem 3). Hence, J(BE ) is w∗ −compact in E ∗∗ . Since a compact
set in a Hausdorff topological space is always closed (Prove this. See Exercises 9.1,
Problem 4), it follows that J(BE ) is w∗ −closed in E ∗∗ . But by Lemma 9.3, J(BE ) is
dense in BE ∗∗ . Hence, in the w∗ −topology of E ∗∗ , we have J(BE ) = J(BE ) = BE ∗∗ .
Hence, E is reflexive.
¤
We now examine some elementary properties of reflexive spaces.
Proposition 9.6 Let E be a reflexive Banach space and let K be a closed subspace
of E. Then K is reflexive.
Proof. Exercise.
Corollary 9.7 Let E be a Banach space. Then, E is reflexive if and only if E ∗ is.
Proof. (⇒) Let E be reflexive. We prove E ∗ is reflexive. By Kakutani’s Theorem,
it suffices to prove that BE ∗ is compact in the weak topology of E ∗ . By Banach
Alaoglu Theorem, BE ∗ is compact in the weak* topology of E ∗ . Since E is reflexive,
the weak* topology of E ∗ = the weak topology of E ∗ . (Verify). Hence, BE ∗ is
compact in the weak topology of E ∗ . By Kakutani’s theorem, E ∗ is reflexive.
(⇐) Let E ∗ be reflexive. We prove E is reflexive. By the first part we know that
E ∗∗ is reflexive. But J(E) is a closed subspace of E ∗∗ and hence, by Proposition 9.6,
J(E) is reflexive. Since J is an isometry, E is reflexive (see Exercises 9.1, Problem 5).
¤
Corollary 9.8 Let E be a reflexive Banach space, K a closed bounded convex
nonempty subset of E. Then K is weakly compact.
Proof. E reflexive ⇒ (by Kakutani’s Theorem) BE is weakly compact. K bounded
⇒ K ⊂ tBE for some t > 0. Furthermore, tBE is weakly compact (Why?). K
closed and convex ⇒ (by Theorem 8.19) K is weakly closed. Hence, K is a weakly
closed subset of a weakly compact set, and so K is weakly compact.
¤
We now give the proofs of Lemma 9.3 and Lemma 9.4. For a proof of Lemma 9.3,
we need the following lemma.
Lemma 9.9 (Helly’s Theorem) Let E be a Banach space, {fi }ni=1 ∈ E ∗ and {αi }ni=1 ∈
IR fixed. Then the following properties are equivalent:
(i) ∀² > 0, ∃ x² ∈ E s.t kx² k ≤ 1 and

|hfi , x² i − αi | < ² ∀ i = 1, 2, . . . , n;
124
(ii)
¯Xn ¯ n
X
¯ ¯
¯ βi αi ¯ ≤ k β i fi k ∀β1 , β2 , . . . , βn ∈ IR.
i=1 i=1
Proof. (i)⇒(ii) Fix β1 , . . . , βn ∈ IR and set

n
X
S= |βj |.
j=1
Then, for arbitrary ∈ > 0 using (i) we obtain

¯ ¯ ¯ ¯
¯Pn Pn ¯ ¯P n ¯
¯ βi hfi , x² i − β α ¯ = ¯ βi {hfi , x² i − αi }¯
¯ i i ¯ ¯ ¯
i=1 i=1 i=1
P n
≤ |βi |.|hfi , x² i − αi |
i=1
Pn
≤ |βi |² = S²,
i=1
and hence
¯n ¯ ¯ n ¯
¯P ¯ ¯ P P n Pn ¯
¯ βi αi ¯ = ¯− βi hfi , x² i + β α + β hf , x i ¯
¯ ¯ ¯ i i i i ² ¯
i=1 i=1¯ ¯
i=1 i=1
¯P n ¯
≤ S² + ¯¯ βi hfi , x² i¯¯
i=1
Pn
≤ S² + k βi fi k · kx² k
i=1
Pn
≤ S² + k βi fi k (since kx² k ≤ 1), ∀² > 0.
i=1
Since ² > 0 is arbitrary, we obtain,

¯ n ¯
¯X ¯ n
X
¯ ¯
¯ βi αi ¯ ≤ k βi fi k, as required.
¯ ¯
i=1 i=1
(ii)⇒(i).
We establish this by contradiction.
Put α~ = (α1 , . . . , αn ) ∈ IRn and consider the map
φ : E → IRn
125
defined by φ(x) = (hf1 , xi, . . . , hfn , xi). It suffices to prove α
~ ∈ φ(BE ). Suppose α ~ 6∈
φ(B¯ E ). Clearly, {ᾱ}, as a singleton in IRn , is convex, nonempty and compact, φ(B̄E )
is convex, nonempty and closed. By Hahn-Banach theorem (second geometric form)
we can separate {~ α} and φ(B ¯ E ) in the strict sense; i.e., there exist β=(β
~ 1 , . . . , βn ) ∈
n
IR and δ ∈ IR such that
φ(x).β~ < δ < α

~ .β~ ∀x ∈ BE .
Observe that
φ(x) · β~ < δ ⇒ φ(−x) · β~ < δ ⇒ −φ(x) · β~ < δ ⇒ φ(x) · β~ > −δ.
Consequently, ¯ ¯
¯X n ¯ n
X
¯ ¯
¯h βi fi , xi¯ ≤ δ < βi αi ∀x ∈ BE ,
¯ ¯
i=1 i=1
so that taking ‘sup’ over all x with kxk ≤ 1 we obtain,

n
X ¯Xn ¯
¯ ¯
k β i fi k ≤ δ < ¯ βi αi ¯
i=1 i=1
which contradicts (ii). The proof of the lemma is complete.

¤
Proof of Lemma 9.3. Let ψ ∈ BE ∗∗ be arbitrary and let T U be a neighbourhood

∗ ∗∗
of ψ in the w topology of E . It suffices to prove J(BE ) U 6= ∅. We can always
suppose that U is of the form
U = {ρ ∈ E ∗∗ : |hρ − ψ, fi i| < ², fi ∈ E ∗ , i = 1, 2, ..., n}.
It now suffices to find x ∈ BE such that J(x) ∈ U (or, equivalently),
|hJx − ψ, fi i| < ²,
i.e., |hfi , xi − hψ, fi i| < ², (9.2)
since hJx, fi i = hfi , xi. (At this point think of the Condition (ii) of Helly’s lemma.
This may help us to find x² with ||x² || ≤ 1 which satisfies (9.2). So, we try to con-
struct Condition (ii) of the lemma which, by Lemma 9.9, is equivalent to condition
(i)).
Put αi = hψ, fi i and note that for arbitrary β1 , β2 , ..., βn ∈ IR, we have
126
Pn Pn Pn Pn
| i=1 βi αi | = |hψ, i=1 βi fi i| ≤ ||ψ||.|| i=1 βi fi || ≤ || i=1 βi fi ||,
since ψ ∈ BE ∗∗ . Hence, by Helly’s lemma, there exists x² ∈ BE such that
|hfi , x² i − αi | < ²
for all i = 1, 2, ..., n. This implies |hJ(x² ), fi i − hψ,

T fi i| <∈ i.e, |hJ(x² ) − ψ, fi i| < ∈
which implies J(x² ) ∈ U . Hence, J(x² ) ∈ J(BE ) U . This completes the proof.
¤
Proof of Lemma 9.4. (⇒) Suppose T : (X, s) / (Y, s) is linear and continuous.
We want to prove T : (X, ω) / (Y, ω) is continuous. By Proposition 8.5, it suffices
∗
to verify that for each f ∈ Y , the map f ◦ T : (X, ω) → IR is continuous (See Fig.
9.2).
T / (Y, s)
(X, s)
NN
NN
NN
NN
NN f
f ◦T N N
NN
NN
NN ²
&
R
Fig. 9.2
We show first that the map f ◦ T : (X, s) → IR is linear and continuous. Linearity
is obvious. Clearly, f ◦ T is continuous as a composition of two continuous maps.

We now show f ◦ T : (X, ω) → IR is continuous. Let {xn } be an arbitrary sequence
in X such that xn converges weakly to y. Since (f ◦ T ) ∈ X ∗ , it follows that
(f ◦ T )(xn ) → (f ◦ T )(y),
i.e., f ◦ T : (X, ω) → IR is continuous. By Proposition 8.5 it follows that
T : (X, ω) → (Y, ω)
is continuous.
(⇐) Suppose now that T : (X, ω) → (Y, ω) is linear and continuous. Then,
G(T ) = {(x, T x) : x ∈ X}
127
is a weakly closed subset of X × Y . Consequently, G(T ) is closed in the strong
topology of X × Y (since weakly closed sets are also strongly closed). Hence, by the
Closed Graph Theorem, T : (X, s) → (Y, s) is continuous. This completes the proof.
¤
EXERCISES 9.1
1. Give a detailed proof of Proposition 9.2.
2. Give the concluding argument why the map f ◦ J −1 defined in the first part
of the proof of Theorem 9.5 is continuous.
3. Let ω, ω ∗ denote the weak and weak∗ topology respectively on a Banach space.
Prove that if
J : (E, ω) → (E ∗∗ , ω)
is continuous, then
J : (E, ω) → (E ∗∗ , ω ∗ )
is also continuous. (Here J is the canonical map of E into E ∗∗ ).
4. Prove that a compact subset of a Hausdorff topological space is closed.
5. Let X and Y be two Banach spaces and let T : X → Y be a linear surjective

isometry of X onto Y . Prove that X is reflexive if and only if Y is.
6. Let X, Y be Banach spaces and let T : (X, s) → (Y, ω) be a linear and

continuous map. Prove that
T : (X, s) → (Y, s)
is also continuous.
(Hint: Use Lemma 9.4).
9.1.2 Eberlein–Smul’yan Theorem

In this section we introduce a very important and useful theorem concerning re-
flexive spaces – the Eberlein–Smulyan Theorem. We shall be contented here with
a statement of the theorem. A proof can be found in any more advanced text in
functional analysis.
128
Recall (Corollary 9.8) that every nonempty closed convex and bounded subset of a
reflexive Banach space is weakly compact. The following result is, in fact, true (see
e.g., [9]).
Proposition 9.10 A subset of a reflexive Banach space is weakly compact (or

weakly sequentially compact) if and only if it is (norm) bounded.
An immediate consequence of this proposition is that the unit ball (open or closed) in
a reflexive Banach space is always weakly compact. Moreover, every (norm) bounded
sequence in a reflexive Banach space has a weakly convergent subsequence. These
results are embodied in the following very important theorem:
Theorem 9.11 (Eberlein–Smul’yan Theorem) A Banach space E is reflexive if and

only if every (norm) bounded sequence in E has a subsequence which converges
weakly to an element of E.
9.2 Uniformly Convex Banach Spaces

Let E be a Banach space, and Sr (x0 ), Br (x0 ) denote the sphere and the open ball
respectively centred at x0 and with radius r > 0, i.e.,
Sr (x0 ) = {x ∈ E : kx − x0 k = r}.
Br (x0 ) = {x ∈ E : kx − x0 k < r}.
A Banach space E is called uniformly convex if for any ε ∈ (0, 2], there exists a
δ = δ(ε) > 0 such that if x, y ∈ E with kxk ≤ 1, kyk ≤ 1 and kx − yk ≥ ε, then
k 21 (x + y)k ≤ 1 − δ.
Roughly speaking, E is uniformly convex if and only if for each two points x, y on
the unit ball of E centred at the origin (i.e., kxk ≤ 1, kyk ≤ 1) which are such that
they are at least distance ε apart (i.e., kx − yk ≥ ε) then the midpoint of x and y
is inside the ball and is at least a distance of δ from the boundary of the ball (see
e.g., Fig. 9.3).
129
X 1
0
0
1
δ
Y
1−δ
1
00
1
2 (X +Y)
Fig. 9.3
Remark 9.12 Since x, y ∈ B1 (0), the maximum value of ε > 0 is 2. Hence, the
restriction ε ∈ (0, 2]. Geometrically, a Banach space E is uniformly convex if and
only if the unit ball centred at the origin is “uniformly round”. To fix ideas we
consider the following example:
Example 9.13 Let E = IR2 , the plane. Endow IR2 with the following three norms.
For x = (x1 , x2 ) ∈ IR2 ,
kxk1 = k(x1 , x2 )k1 = |x1 | + |x2 |.

kxk2 = k(x1 , x2 )k2 = (x21 + x22 )1/2 .
kxk∞ = k(x1 , x2 )k∞ = max{|x1 |, |x2 |}.
(We know that k · k1 , k · k2 and k · k∞ are norms on IR2 ). Let E1 = (IR2 , k · k1 ), E2 =

(IR2 , k · k2 ) and E∞ = (IR2 , k · k∞ ). E1 , E2 and E∞ are then three distinct normed
linear spaces, with unit balls centred at the origin drawn respectively in Fig. 9.4(a),
(b) and (c).
(0, 1) (0, 1) (−1, 1) (1, 1)
(−1, 0) (1, 0) (−1, 0) (1, 0)
(0, −1) (0, −1) (−1, −1) (1, −1)
130
Fig. 9.4
It then follows that IR2 with norm k · k2 is uniformly convex whereas the same IR2
with either the norm k · k1 or the norm k · k∞ is not uniformly convex. These as-
sertions can be established rigorously. Consequently, uniform convexity of a Banach
space E is actually a property of the norm on E. However, we shall follow the abuse
of language (or convention) and call a Banach space uniformly convex whenever the
norm on it is a uniformly convex one.
The following theorem shows that a large class of Banach spaces is uniformly convex.
Theorem 9.14 Every inner product space H is uniformly convex.
Proof. Recall the following identity (the parallelogram law Proposition 6.11) which
is valid in any inner product space. For each x, y ∈ H,
kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ). (9.3)
Let ε ∈ (0, 2] be given and let x, y ∈ H be such that kxk ≤ 1, kyk ≤ 1 and
kx − yk ≥ ε. Then equation (9.3) yields:
° °2 ° °2
°1 ° °1 °
° (x + y)° ≤ [2(2) − kx − yk ] = 1 − ° (x − y)° ≤ 1 − 1 ε2 ,
1 2
°2 ° 4 °2 ° 4
so that ° ° r
°1 °
° (x + y)° ≤ 1 − 1 ε2 .
°2 ° 4
q
We can now choose δ = 1 − 1 − 14 ε2 > 0 to complete the proof.
¤
More generally, we have the following theorem.
Theorem 9.15 `p (1 < p < ∞) is uniformly convex.
For a proof of this theorem we need the following lemma.
Lemma 9.16 (R.C. James) Let E = `p . Then for p, q > 1 such that p1 + 1q = 1 and
for each pair x, y ∈ E, the following inequalities hold:
° °q ° °q
°1 ° °1 ° £ ¤
(i) ° (x + y)° + ° (x − y)° ≤ 2−1 (kxkp + kykp ) q−1 , for 1 < p ≤ 2,
°2 ° °2 °
and,
(ii) kx + ykp + kx − ykp ≤ 2p−1 (kxkp + kykp ), for 2 ≤ p < ∞ .
131
Proof of Theorem 9.15. Given ε ∈ (0, 2], let x, y ∈ `p be such that kxk ≤ 1, kyk ≤
1 and kx − yk ≥ ε. Two cases arise:
Case 1: 1 < p ≤ 2. In this case, Lemma 9.16 (i) yields:

° °q ° °q
°1 ° °1 °
° (x + y)° + ° (x − y)° ≤ 2−(q−1) (kxkp + kykp )q−1
°2 ° °2 °
≤ 2−(q−1) 2(q−1) = 1.
Thus, ° °q ° °q
°1 ° ° ° ³ ´q
° (x + y)° ≤ 1 − ° x − y ° ≤ 1 − ε ,
°2 ° ° 2 ° 2
or, ° ° h
°1 ° ³ ´q i1/q
° (x + y)° ≤ 1 − ε ,
°2 ° 2
£ ¡ ¢q ¤1/q ° °
so that by choosing δ = 1 − 1 − 2ε > 0, we obtain, ° 12 (x + y)° ≤ 1 − δ, and
so `p (1 < p ≤ 2) is uniformly convex.
Case 2: 2 ≤ p < ∞ The result follows as in Case 1 by using Lemma 9.16 (ii).
¤
Although Theorems 9.14 and 9.15 provide large classes of spaces which are uniformly
convex, some well–known spaces are not uniformly convex.
Example 9.17 The space `1 is not uniformly convex. To see this, take ε = 1 and
choose x = (1, 0, 0, 0, . . .), y = (0, −1, 0, 0, . . .). Clearly x, y ∈ `1 and kxk`1 = 1 =
kyk`1 , kx − yk`1 = 2 > ε. However, k 21 (x + y)k = 1 so that k 12 (x + y)k ≤ 1 − δ, δ > 0
is not satisfied, showing that `1 is not uniformly convex.
Example 9.18 The space `∞ is not uniformly convex. Consider u = (1, 1, 0, 0, . . .), v =
(−1, 1, 0, 0, . . .) and take ε = 1. Then
kuk`∞ = kvk`∞ = 1, ku − vk`∞ = 2 > ε. However, k 21 (u + v)k`∞ = 1, and so `∞ is
not uniformly convex.
Example 9.19 The space C[a, b] of all real–valued continuous functions on the
compact interval [a, b] endowed with the “sup norm” is not uniformly convex.
To see this, choose two functions f and g as follows:
f (t) = 1 for all t ∈ [a, b].
b−t
g(t) = for each t ∈ [a, b].
b−a
132
Take ε = 12 . Clearly, f, g ∈ C[a, b], kf k = kgk = 1 and kf − gk = 1 > ε. Also
k 21 (f + g)k = 1 and so C[a, b] is not uniformly convex.
Proposition 9.20 Let X be a uniformly convex Banach space. Then for any d >
0, ε > 0 and arbitrary vectors x, y ∈ X with kxk ≤ d, kyk ≤ d, kx − yk ≥ ε there
exists a δ > 0 such that
°1 ° h ³ ε í
° °
° (x + y)° ≤ 1 − δ d.
2 d
Proof. Let ε > 0 be given and let z1 = xd , z2 = yd , and suppose we set ε̄ = dε .
Obviously ε̄ > 0. Moreover, kz1 k ≤ 1, kz2 k ≤ 1 and kz1 − z2 k = d1 kx − yk ≥ dε = ε̄.
Now, by uniform convexity we have
°1 °
° °
° (z1 + z2 )° ≤ 1 − δ(ε̄),
2
i.e., °1 ° ³ε´
° °
° (x + y)° ≤ 1 − δ ,
2d d
which implies °1 ° h ³ ε í
° °
° (x + y)° ≤ 1 − δ d.
2 d
¤
9.2.1 Milman–Pettis Theorem

We now prove an important result concerning uniformly convex spaces.
Theorem 9.21 (Milman–Pettis Theorem) Every uniformly convex Banach space E
is reflexive.
Proof. It suffices to prove that J : BE / BE ∗∗ is surjective. So let ξ ∈ E ∗∗ such
that kξk = 1. Then ξ ∈ BE ∗∗ . It suffices now to prove that ξ ∈ J(BE ). Recall that
J(BE ) is strongly closed in E ∗∗ (prove this (see Exercises 9.2, Problem 7)) so that
J(BE ) = J(BE ). Hence it suffices to prove ξ ∈ J(BE ). To prove this, it suffices
to show that any open ball with centre ξ intersects J(BE ) i.e., given any ε > 0,
Bε (ξ) ∩ J(BE ) 6= ∅, or better still,
∀ ε > 0, ∃ x ∈ BE such that kξ − J(x)k < ε. (9.4)
We now prove (9.4). So let ε > 0 be given. By the uniform convexity of E, there
exists δ > 0 such that for all x, y ∈ E with kxk ≤ 1, kyk ≤ 1, kx − yk ≥ ε, we have
° °
°1 °
° (x + y)° < 1 − δ .
°2 °
133
Fix this δ > 0 (Corresponding to the given ε > 0). Since kξk = 1, it follows that
ξ 6= 0. Hence, by Theorem 3.10, we can choose f ∈ E ∗ such that kf k = 1 and
f (ξ) = hξ, f i = kξk = 1 > 1 − δ/2. (9.5)
Set U := {u ∈ E ∗∗ : |hu − ξ, f i| < δ/2}. Then U is a neighbourhood of ξ in

the ω ∗ –topology of E ∗∗ . But by the lemma of Goldstein (Lemma 9.3), J(BE ) is
dense in BE∗∗ with respect to the ω ∗ –topology of E ∗∗ . Hence, any neighbourhood (in
the ω ∗ –topology of E ∗∗ ) of an arbitrary element of BE ∗∗ must intersect J(BE ). In
particular,
U ∩ J(BE ) 6= ∅.
Let y ∈ U ∩ J(BE ). Then, clearly y ∈ J(BE ). Let x0 ∈ BE be such that J(x0 ) = y.
Then J(x0 ) ∈ U . We have thus picked x0 ∈ BE such that Jx0 ∈ U . Thus, to
complete the proof of (9.4) it now suffices to prove that kξ − J(x0 )k < ε; i.e.,
ξ ∈ B(J(x0 ), ε) or, equivalently, (by Exercises 7.6, Problem 6),
ξ ∈ Jx0 + ε BE ∗∗ . (9.6)
We prove (9.6) by contradiction. So, suppose ξ 6∈ Jx0 + ε BE ∗∗ . Then ξ ∈ (Jx0 +

ε BE ∗∗ )c := V , the complement of (Jx0 + εBE ∗∗ ). Observe that since BE ∗∗ is ω ∗ –
closed it follows that V is a neighbourhood of ξ in the ω ∗ –topology of E ∗∗ . Again,
by the lemma of Goldstein (Lemma 9.3),
(U ∩ V ) ∩ J(BE ) 6= ∅.
This implies there exists x̂ ∈ BE such that J(x̂) ∈ U ∩ V . Consequently, we obtain,
|hJ(x0 ) − ξ, f i| = |hJ(x0 ), f i − hξ, f i|

= |hf, x0 i − hξ, f i| < δ/2 (since we already noted that J(x0 ) ∈ U )
and,
|hf, x̂i − hξ, f i| < δ/2, (since J(x̂) ∈ U and hJ(x̂), f i = hf, x̂i).
These inequalities imply
|2hξ, f i − hf, x0 + x̂i| ≤ |hf, x0 i − hξ, f i| + |hf, x̂i − hξ, f i| ≤ δ,
so that
2hξ, f i ≤ |hf, x0 + x̂i| + δ ≤ kf k · kx0 + x̂k + δ ≤ kx0 + x̂k + δ
134
and consequently, using inequality (9.5), we obtain the following estimate:
° °
°1 °
° (x0 + x̂)° ≥ hξ, f i − δ/2 > 1 − δ/2 − δ/2 = 1 − δ.
°2 °
We now have the following situation: kx0 k ≤ 1, kx̂k ≤ 1 and k 12 (x0 + x̂)k > 1 − δ.
Then by uniform convexity (contrapositive) we have: kx0 − x̂k ≤ ε. But then,
kx0 − x̂k > ε, since J(x̂) ∈ V and J(x0 ) 6∈ V , which implies
kx0 − x̂k = kJ(x0 − x̂)k

= kJx0 − J x̂k > ε.
(The last inequality following from the fact that V = (B(Jx0 , ε))c ). Contradiction.
The proof is complete.
¤
9.3 Separable spaces

We have defined reflexivity of a normed linear space and have given examples of
some important reflexive spaces. It turns out that some of the well known im-
portant spaces (e.g., `∞ , C[0, 1] with ‘sup’ norm) are not reflexive. How can one
establish this fact? One method of doing this is to use the idea of separable spaces.
We explore this idea next. A Banach space is called separable if it contains a count-
able dense subset, i.e., if and only if there is a countable set S, S ⊂ E such that
S̄ = E. Clearly, a subset of a separable space is separable. (Convince yourself). We
prove the following theorem.
Theorem 9.22 Let E be a normed linear space such that E ∗ is separable. Then E
is separable.
Proof. Let E ∗ be separable. Then the unit sphere in E ∗ centred at 0 ∈ E ∗ ,

SE ∗ (0, 1) := {f ∈ E ∗ : kf k = 1} is separable (as a subset of a separable metric
space). Hence, SE ∗ (0, 1) contains a set Ω which is countable and dense. Let Ω =
{fn ∈ SE ∗ (0, 1), i = 1, 2, . . .} and Ω̄ = SE ∗ (0, 1). But fn ∈ SE ∗ (0, 1) ⇒ kfn k =
sup |fn (x)| = 1. By the definition of ‘sup’, given any ε > 0, there exists xn with
kxk=1
kxn k = 1 such that |fn (xn )| ≥ ||fn || − ε. In particular, by choosing ε = 12 ||fn ||, we
obtain that
1 1
|fn (xn )| > kfn k = .
2 2
135
Define F := span{xn }. Observe that the subset S of F consisting of linear com-
binations of xn ’s whose real and imaginary parts are rational numbers is clearly
countable and dense in F (Why?). Hence F is separable. Moreover, F is a closed
subspace of E. We now prove F = E. Proof of this equality will be obtained by
contradiction. So, assume F 6= E. Then there exists x0 ∈ E\F . Since F is a closed
subspace of E, Theorem 3.13 now yields that there exists f ∗ ∈ E ∗ with kf ∗ k = 1
and f ∗ (x0 ) = δ, f ∗ (y) = 0 for all y ∈ F , where δ = inf kx0 − mk > 0. In particular,
m∈F
f ∗ (xn ) = 0 ∀ n. So,
1
≤ |fn (xn )| = |fn (xn ) − f ∗ (xn )|
2
= |(fn − f ∗ )xn | ≤ kfn − f ∗ k · kxn k = kfn − f ∗ k,
so that kfn − f ∗ k ≥ 12 . Since f ∗ ∈ E ∗ , the last inequality contradicts the fact that
{fn } is dense in E ∗ . Hence F = E. So, E is separable.
¤
Remark 9.23 The converse of Theorem 9.22 is not true. There are separable spaces
E with duals E ∗ which are not separable (e.g., `1 is separable but `∗1 = `∞ is not).
However, if E is reflexive and separable then E ∗ is separable. In fact, we have the

following theorem.
Theorem 9.24 A Banach space E is reflexive and separable if and only if E ∗ is

reflexive and separable.
Proof. (Trick of proof: we shall use the hypothesis that E is reflexive and separable
to obtain that E ∗∗ is separable; then we employ Theorem 9.22 to conclude that E ∗
is separable). Now, E is separable ⇒ E has a countable dense subset. Call it
{xk }. Let x∗∗ ∈ E ∗∗ be arbitrary. Since E is reflexive, there exists x ∈ E such that
Jx = x∗∗ . For any ε > 0, by the density of {xk } in E, there exists xk such that
kxk − xk < ε. Since J is an isometry, we obtain,
kJxk − x∗∗ k = kJxk − Jxk = kxk − xk < ε.
Thus, {Jxk } is a countable dense subset of E ∗∗ . Hence E ∗∗ is separable. By Theorem

9.22 we obtain that E ∗ is separable. By Corollary 9.7, E is reflexive if and only if
E ∗ is reflexive. The proof is complete.
¤
136
Remark 9.25 Theorem 9.24 gives us a criterion for showing that a given space is
not reflexive. For, if E is a normed linear space such that (i) E is separable, and
(ii) E ∗ is not separable, then E is not reflexive. Thus, to show that a normed linear
space E is not reflexive, it suffices to verify that (i) and (ii) are satisfied.
We now prove yet another theorem which will also be useful in showing that certain
spaces are not reflexive.
Theorem 9.26 Let E be a reflexive Banach space. Then for each f ∗ ∈ E ∗ there
exists some x0 ∈ E, with kx0 k = 1 such that
f ∗ (x0 ) = kf ∗ k.
Proof. If f ∗ = 0, the assertion is evident. If f ∗ 6= 0, then by Theorem 3.9, there

exists some f0∗∗ ∈ E ∗∗ such that
f0∗∗ (f ∗ ) = kf ∗ k and kf0∗∗ k = 1. (9.7)
By reflexivity, there exists x0 ∈ E such that Jx0 = f0∗∗ . Consequently,

kJx0 k = kf0∗∗ k = 1. Also, kJx0 k = kx0 k. Hence kx0 k = 1. Moreover, by (9.7),
kf ∗ k = f0∗∗ (f ∗ ) = hf0∗∗ , f ∗ i = hJx0 , f ∗ i = f ∗ (x0 ), completing the proof.
¤
Remark 9.27 Theorem 9.26 asserts that all bounded linear functionals on a reflex-
ive Banach space attain their maximum absolute value on the closed unit ball. In
general, the point x0 ∈ E in Theorem 9.26 need not be unique. However, if E is say,
uniformly convex, then each f ∗ 6= 0 attains its maximum absolute value at a unique
point on the unit ball.
Remark 9.28 The converse of Theorem 9.26 is also true, i.e., if E is a Banach
space such that whenever f ∗ ∈ E ∗ , there exists some x0 ∈ E with kx0 k = 1 such
that f ∗ (x0 ) = kf ∗ k, then E is reflexive.
We give some examples.

Example 9.29 `1 is not reflexive.
Solution. `1 is separable (see Exercises 9.2, Problem 4).

Assume for contradiction that `1 is also reflexive. By Theorem 9.24, `∗1 is separable.
But `∗1 = `∞ which is not separable (by Exercises 9.2, Problem 8). Contradiction.
Hence `1 is not reflexive.
137
Example 9.30 c0 is not reflexive.
Solution. In view of Theorem 9.26, it suffices to exhibit a continuous linear func-

tional on c0 that does not attain its maximum absolute value on the closed unit ball,
i.e., it suffices to construct some f ∗ ∈ c∗0 such that
|f ∗ (x)| < kf ∗ k
for all x ∈ c0 such that kxkc0 ≤ 1. We now proceed to do this. For each
x = {αk } ∈ c0 , define
X∞
αk
f (x) = .
k=1
k!
Then f is a linear functional on c0 . Furthermore, since |αk | ≤ k{αk }k,
k = 1, 2, 3, . . ., it follows
µ ∞ that¶ if x ∈ c0 with kxk ≤ 1 then
P
∞
αk P 1
|f (x)| = k!
≤ k!
, so that f is a bounded linear functional on c0 with
k=1 k=1
P
∞
1
P
∞
1
kf k ≤ k!
. In fact, kf k = k!
since for each positive integer n, if
k=1 k=1
xn = {αkn }
∈ c0 is such that αkn =
1, k = 1, 2, . . . , n and αkn = 0, k > n then
P 1
n
kxn k = 1 and f (xn ) = k!
. Hence,
k=1
Xn ∞
X
1 n 1
kf k = sup |f (x)| ≥ sup f (x ) ≥ sup = ,
kxk=1 n n
k=1
k! k=1 k!
P
∞
1
so that kf k = k!
, as claimed.
k=1
However, there exists no x ∈ c0 with kxk ≤ 1 for which |f (x)| = kf k as can be seen
from the following argument. Suppose x = {αk } ∈ c0 such that kxk ≤ 1. Since αk
/ 0, there exists an integer N0 > 0 such that |αN | < 1. Hence,
0
¯X ∞
αk ¯¯ X |αk | |αn | X 1
∞ ∞
1 X∞
1
¯
|f (x)| = ¯ ¯≤ + < + = = kf k.
k=1
k! k=1
k! n! k=1
k! n! k=1
k!
k6=n
Hence c0 is not reflexive.

¤
EXERCISES 9.2
138
1. Prove that a subset of a separable space is separable.
2. Prove that the set F defined in the proof of Theorem 9.22 is: (i) separable;
(ii) a closed subspace of E.
3. Prove that every reflexive normed linear space is a Banach space.
4. Prove that `p , 1 ≤ p < ∞ is separable.
5. (a) Is every Hilbert space reflexive? Justify your answer.

(b) What can you say about the reflexivity of `p , 1 < p < ∞?
6. Let E be a Banach space and K be a subset of E.

(a) Prove that if K is compact then K is weakly compact.
(b) Using (a), or otherwise, prove that every finite dimensional normed linear
space is reflexive.
7. Prove that, with our usual notations, J(BE ) is always strongly closed in E ∗∗ ,
where E is an arbitrary Banach space and J is the canonical embedding.
8. Show that `∞ is not separable.
9.3.1 A Non–separable Inner Product Space

The Hilbert spaces which are usually encountered in applications are separable.
However, we give here an example of a non–separable inner product space which is
also a very important space in mathematics. This is the space of almost periodic
functions. We begin with the following definition.
Definition 9.31 A complex–valued function which is continuous on (−∞, ∞) is
called almost periodic if it is the uniform limit on (−∞, ∞) of a sequence of trigono-
P
n
metric polynomials of the form ak eiλk t , λk real.
k=1
Let E denote the set of almost periodic functions. With the usual operations of
addition and scalar multiplication, E is a vector space over CI. If f, g ∈ E are
arbitrary, it can be shown that
Z
1 T
hf, gi = lim / f (s) g(s) ds
T ∞ T 0
exists and defines an inner product on E. Since {eiλt : λ real} is an uncountable

orthonormal set, E is not separable.
139
140
CHAPTER 10
An Application of Weak Topology in Optimization
10.1 Convex Sets

Definition 10.1 Let X be a real linear space and C ⊂ X. The set C is called
convex if for each x1 , x2 ∈ C and for each t ∈ [0, 1], we have tx1 + (1 − t)x2 ∈ C
(see Fig 10.1).
Notation. Let x, y ∈ X. We write [x, y] for the geometric segment from x to

y, i.e.,
[x, y] := {z ∈ X : ∃ t ∈ [0, 1] such that z = tx + (1 − t)y}.
Hence, z = tx + (1 − t)y = y + t(x − y).
Exercise. Prove [x, y] = [y, x].
Remark 10.2 C is convex if and only if ∀ x, y ∈ C, [x, y] ⊂ C.
Notation. Let n ∈ N. Define

n
X
Λn := {(α1 , α2 , ..., αn ) ∈ Rn+ : αi = 1}.
i=1
Proposition 10.3 C ⊂ X is convex if and only if for each n ∈ N, for each

P
n
(α1 , α2 , ..., αn ) ∈ Λn and for each (x1 , x2 , ..., xn ) ∈ C n , we have αi xi ∈ C.
i=1
141
Fig. 10.1
Proof. (⇒) Suppose that C is convex. We prove by induction. For n = 2, let

P
2
(α1 , α2 ) ∈ Λ2 , (x1 , x2 ) ∈ C 2 . Then αi xi = α1 x1 + α2 x2 and α1 + α2 = 1 (so that
i=1
P
2
α2 = 1 − α1 ), i.e., αi xi = α1 x1 + (1 − α1 )x2 ∈ C. So, the result holds for n = 2.
i=1
Assume now it holds for n = k for some k > 2. Let (α1 , α2 , ..., αk+1 ) ∈ Λk+1 (so that
P
k+1
αi = 1), and (x1 , x2 , ..., xk , xk+1 ) ∈ C k+1 . Then,
i=1
k+1
X k
X
yk := αi xi = αi xi + αk+1 xk+1 .
i=1 i=1
P
k+1 P
k
Since αi = 1, we have αk+1 = 1 − αi .
i=1 i=1
Case 1. If αk+1 = 1, then α1 = α2 = αk = 0 (why?), and so yk = xk+1 ∈ C

(by hypothesis).
P
k
Case 2. If αk+1 < 1, then αi = 1 − αk+1 > 0. Hence,
i=1
X k
1
yk = (1 − αk+1 ). αi xi + αk+1 xk+1
(1 − αk+1 ) i=1
142
k
X αi
= (1 − αk+1 ) xi + αk+1 xk+1 .
i=1
(1 − αk+1 )
α1 αk
Observe that ( (1−αk+1 )
, ..., (1−αk+1 )
) ∈ Λk (why?). Thus,
k
X αi
x := xi ∈ C (by induction hypothesis).
i=1
(1 − αi+1 )
Now, yk = (1 − αk+1 )x + αk+1 xk+1 and αk+1 ∈ [0, 1]. Hence, yk ∈ C.
(⇐) Let n = 2, t ∈ [0, 1], (x1 , x2 ) ∈ C 2 . Put α1 = t ≥ 0, α2 = 1 − t ≥

0, α1 + α2 = 1, (α1 , α2 ) ∈ Λ2 . But α1 x1 + α2 x2 ∈ C. Thus, tx1 + (1 − t)x2 ∈ C
P
2
since α1 x1 + α2 x2 = tx1 + (1 − t)x2 . Hence, αi xi ∈ C and this implies that C is
i=1
convex. Hence, the proposition holds. ¤
Example 10.4 Every affine space is convex.
Recall that an affine space is simply a translation of a linear space by a point, e.g.,
if V is a linear space, then X := V + x0 := {v + x0 : v ∈ V } is an affine space (see
Fig. 1.2). Let (α1 , α2 , ..., αn ) ∈ Λn , vi ∈ V, i = 1, 2, ..., n.. Then xi ∈ X implies that
xi = x0 + vi , where vi ∈ V , and so for xi ∈ X,
n
X n
X n
X n
X n
X
α i xi = αi (x0 + vi ) = αi x0 + αi vi = x0 + αi vi ∈ X. ¤
i=1 i=1 i=1 i=1 i=1
Example 10.5 Every Hyperplane is convex.

Let X be a real normed linear space. Recall
X ∗ = {f : X −→ R : f is linear and bounded (continuous)}.
For α ∈ R, define
Hf,α := {x ∈ X : f (x) = α, f ∈ X ∗ } = {x ∈ X : hx, f i = α}
where hx, f i ≡ f (x). Then Hf,α is called a hyperplane of X. // Geometrically, Hf,α
is affine. For y, x0 ∈ Hf,α implies f (x0 ) ≡ hx0 , f i = α and f (y) ≡ hy, f i = α. Thus,
hy − x0 , f i = 0, i.e., y − x0 ∈ N (f ) (where N (f ) denotes the null space or kernel
of f ). This implies y ∈ x0 + N (f ). Hence, Hf,α ⊆ x0 + N (f ) = x0 + Ker(f ).
Clearly, x0 + N (f ) ⊂ Hf,α . For, hx0 + z, f i = hx0 , f i + hz, f i = α + 0 = α. Thus,
x0 + z ∈ Hf,α . Hence, Hf,α = x0 + N (f ) = x0 + Ker(f ). Thus, Hf,α is an affine
space and, by Example 10.4, is a convex set. ¤
143
Fig. 10.2
Fig. 10.3
Example 10.6 Hyperplanes in R, R2 and R3 (see Fig. 1.4)
Example 10.7 Every Half space is convex.
Define the following half-spaces.
144
+
Df,α := {x ∈ X : hx, f i ≥ α}.
+∗
Df,α := {x ∈ X : hx, f i > α}.
−
Df,α := {x ∈ X : hx, f i ≤ α}.
−∗
Df,α := {x ∈ X : hx, f i < α}.
All half spaces are convex sets.

Proof. Exercise.
Proposition 10.8 (a) Let C1 , C2 , ..., Cn be convex subsets of X and α1 , α2 , ..., αn ≥
Pn
0, then αi Ci is convex.
i=1 T
(b) Let {Ci }i∈∆ be a family of convex subsets in X, then i∈∆ Ci is convex.
Proof. Exercise.
10.2 Convex Functions

Definition 10.9 Let D be a subset of a real vector space X and f : D −→ R ∪
{+∞}. Then,
(i) f is said to be convex if:
(a) D is convex, and
(b) For each t ∈ [0, 1] and for each x1 , x2 ∈ D, we have
f (tx1 + (1 − t)x2 ) ≤ tf (x1 ) + (1 − t)f (x2 ).
(ii) f is concave if
(a) D is convex, and
(b) For each t ∈ [0, 1] and for each x1 , x2 ∈ D, we have
f (tx1 + (1 − t)x2 ) ≥ tf (x1 ) + (1 − t)f (x2 ).
Remark 10.10 f is convex if and only if (−f ) is concave, and f is concave if and
only if (−f ) is convex. (See Fig. 1.4).
145
Fig. 10.4
10.2.1 Notations and further definitions

[A] If f : D −→ R ∪ {+∞} is convex, we define f˜ : D −→ R ∪ {+∞} by
½
f (x), if x ∈ D,
f˜(x) =
+∞, if x ∈ XD.
(See Fig. 1.5).

Then f is convex if and only if f˜ is convex.
[B] Let f : X → R ∪ {+∞} be a map. The domain of f is the set defined by
D(f ) := {x ∈ X : f (x) < +∞}.
Domain of f is sometimes called the effective domain of f . The map f is called

proper if D(f ) 6= ∅. This means there exists at least one x0 ∈ D(f ) such that
f (x0 ) ∈ R or that f is not identically +∞.
[C] Epigraph.
The epigraph of f is the set defined by
epi(f ) := {(x, α) ∈ X × R : x ∈ D(f ) and f (x) ≤ α}.
(See Fig. 1.6).
146
Fig. 10.5
Fig. 10.6
[D] Section of f .
Let α ∈ R. We have the following definitions.
Sf,α := {x ∈ X : f (x) ≤ α}
= {x ∈ D(f ) : f (x) ≤ α}.
147
Fig. 10.7
Proposition 10.11 f : X −→ R ∪ {+∞} is convex if and only if epi(f ) is convex.

Proof. (⇒) Suppose that f is convex. Let (x1 , α1 ) ∈ epi(f ), (x2 , α2 ) ∈ epi(f ), t ∈
[0, 1]. Now, f (x1 ) ≤ α1 , f (x2 ) ≤ α2 . Hence, f (tx1 + (1 − t)x2 ) ≤ tf (x1 ) + (1 −
t)f (x2 ) ≤ tα1 + (1 − t)α2 . Thus (tx1 + (1 − t)x2 , tα1 + (1 − t)α2 ) ∈ epi(f ). But
(tx1 + (1 − t)x2 , tα1 + (1 − t)α2 ) = t(x1 , α1 ) + (1 − t)(x2 , α2 ). Hence, epi(f ) is convex.
(⇐). Suppose epi(f ) is convex. We show that f is convex. Let x1 , x2 ∈ D(f ), t ∈

[0, 1]. Then x1 , x2 ∈ D(f ) implies that f (x1 ) ∈ R, f (x2 ) ∈ R. Thus, (x1 , f (x1 )) ∈
epi(f ) and (x2 , f (x2 )) ∈ epi(f ). But epi(f ) is convex. Thus, t(x1 , f (x1 )) + (1 −
t)(x2 , f (x2 )) ∈ epi(f ) if and only if
(tx1 + (1 − t)x2 , tf (x1 ) + (1 − t)f (x2 )) ∈ epi(f )
if and only if
f (tx1 + (1 − t)x2 ) ≤ tf (x1 ) + (1 − t)f (x2 )
if and only if f is convex. ¤
10.3 Strictly Convex Functions

Definition 10.12 f : X −→ R ∪ {+∞} is strictly convex if for each x1 , x2 ∈
D(f ), x1 6= x2 , and for each t ∈ (0, 1) we have
f (tx1 + (1 − t)x2 ) < tf (x1 ) + (1 − t)f (x2 ).
148
f is strictly concave if (−f ) is strictly convex.
(See Fig. 1.9).
Fig. 10.8
Example 10.13 (a) f : X −→ R ∪ {+∞} defined by f (x) = k (constant map).

Then f is convex, but not strictly convex.
(b) Take g ∗ ∈ X ∗ and define f : X −→ R ∪ {+∞} by
f (x) = hx, g ∗ i ∈ R.
Then f is convex but not strictly convex. (Justify).
(c) a : X × X −→ R bilinear map. Suppose a is nonnegative, i.e., for each

x ∈ X, a(x, x) ≥ 0. Then f : X −→ R ∪ {+∞} defined by
f (x) = a(x, x)
is convex.
Proof. Let x1 , x2 ∈ X, t ∈ [0, 1]. Put
∆ := f (tx1 + (1 − t)x2 ) − tf (x1 ) − (1 − t)f (x2 ).
149
Then
∆ = a(tx1 + (1 − t)x2 , tx1 + (1 − t)x2 ) − ta(x1 , x1 ) − (1 − t)a(x2 , x2 )

= a(tx1 , tx1 + (1 − t)x2 ) + a((1 − t)x2 , tx1 + (1 − t)x2 ) − ta(x1 , x1 )
− (1 − t)a(x2 , x2 )
= a(tx1 , tx1 ) + a(tx1 , (1 − t)x2 ) + a((1 − t)x2 , tx1 )
+ a((1 − t)x2 , (1 − t)x2 ) − ta(x1 , x1 ) − (1 − t)a(x2 , x2 )
= t2 a(x1 , x1 ) + t(1 − t)a(x1 , x2 ) + (1 − t)ta(x2 , x1 )
+ (1 − t)2 a(x2 , x2 ) − ta(x1 , x1 ) − (1 − t)a(x2 , x2 )
= (t2 − t)a(x1 , x1 ) + t(1 − t)[a(x1 , x2 )
+ a(x2 , x1 )] + ((1 − t)2 − (1 − t))a(x2 , x2 )
= (t2 − t)a(x1 , x1 ) + t(1 − t)[a(x1 , x1 ) + a(x2 , x2 )
− a(x1 − x2 , x1 − x2 )] + ((1 − t)2 − (1 − t))a(x2 , x2 )
= −t(1 − t)a(x1 − x2 , x1 − x2 ) ≤ 0
This completes the proof. ¤
Corollary 10.14 If a is strictly positive, i.e., for each x 6= 0, a(x, x) > 0, then f
is strictly convex.
10.4 Lower Semi-continuity

Let f : X −→ R ∪ {+∞} and let x̄ ∈ X be arbitrary. Define
U(x̄) := {V ⊂ X : V is a neighbourhood of x̄}.
For V ∈ U , take inf f (x). (See Fig. 10.9(a) and Fig 10.9(b)).
x∈V
This ”inf” may not exist (in which case we say it is −∞). As V ranges over
U(x̄), take
sup { inf f (x)}.
V ∈U x∈V² {x̄}
It is clear that
sup { inf f (x)} = lim inf/ f (x).
V ∈U x∈V² {x̄} x x̄
We now have the following definition.
150
Fig. 10.9(a) Fig. 10.9(b)
Definition 10.15 Let f : X −→ R ∪ {+∞}. Then f is called lower semi-continuous

at x̄ if
sup { inf f (x)} = lim inf/ f (x) ≥ f (x̄), or lim inf/ f (x) ≥ f (x̄).
V ∈U x∈V² {x̄} x x̄ x x̄
If f is lower semi-continuous at each point of X, then f is lower semi-continuous

on X.
Remark 10.16 One easily observes that the function sketched in Fig. 1.10(a) is
lower semi-continuous at x̄, whereas the one sketched in Fig 1.10(b) is not. Observe
also that the function f sketched in Fig 1.10(a) is not continuous at x̄. This brings
us to the following ...
Proposition 10.17 If f : X −→ R ∪ {+∞} is lower semi-continuous at x̄ ∈ X

and {xn } is a sequence in X which converges strongly to x̄ then,
lim inf f (xn ) ≥ f (x̄).
Proof. (See Fig. 1.11 for motivation of proof)

{xn } converges strongly to x̄ implies that, for arbitrary V ∈ U(x̄), ∃ N ∈ N such
that for each n ≥ N, xn ∈ V . Fix n ≥ N, k ≥ n. Then, xk ∈ V . Thus,
{xk : k ≥ n} ⊂ V.
151
Fig. 10.10
This inclusion implies (why?) that
inf f (xk ) ≥ inf f (x). (10.1)

k≥n x∈V {x̄}
Now since f is lower semi-continuous at x̄, we have
sup ( inf f (x) ≥ f (x̄)).

V ∈U (x̄) x∈V {x̄}
Now, if sup ( inf f (x)) = m, say, then either m ∈ R or m = +∞.

V ∈U (x̄) x∈V {x̄}
Case 1. If m ∈ R, then for arbitrary ² > 0, ∃ V² ∈ U(x̄) such that
m−²< inf f (x) ≤ m.

x∈V² {x̄}
So, we obtain that
f (x̄) − ² ≤ m − ² < inf f (x) ≤ inf f (xk ).

x∈V² {x̄} k≥n
(How does the third inequality arise?) This implies,
f (x̄) − ² < inf f (xk ).

k≥n
152
Now taking limit as n / ∞ (why can we do this?) we obtain
f (x̄) − ² < lim / inf f (xk ) = lim inf/ f (xn ).

n ∞ k≥n n ∞
Since ² > 0 is arbitrary, we obtain that

f (x̄) ≤ lim inf/ f (xn ).
n ∞
Case 2 If m = +∞, the proof is easy and is left as an exercise. ¤

Theorem 10.18 Let f : X −→ R∪{+∞} be a map. Then following are equivalent.
(1) f is lower semi-continuous,
(2) epi(f ) is closed,
(3) Sf,α is closed for each α ∈ R.
Proof. (1) ⇒ (2).
We want to prove that epi(f ) is closed. So, let {(xn , λn )}n∈N be a sequence in
epi(f ) such that (xn , λn ) / (x, λ). We show that (x, λ) ∈ epi(f ). Now (xn , λn ) ∈
epi(f ) ⇒ f (xn ) ≤ λn . Hence, lim inf f (xn ) ≤ lim inf λn = lim λn = λ. Since {xn }
converges to x and f is lower semi-continuous, we have
f (x) ≤ lim inf f (xn ).
Hence, f (x) ≤ λ and this implies that (x, λ) ∈ epi(f ).
(2) ⇒ (1). Let x̄ ∈ X be arbitrary. We prove that f is lower semi-continuous

at x̄. So, let ² > 0 be given. Then (x̄, f (x̄) − ²) ∈ / epi(f ) (why?). Hence,
(x̄, f (x̄) − ²) ∈ (epi(f ))c , the complement of epi(f ). Since epi(f ) is closed (by
(2)), we have (epi(f ))c is open. Hence, ∃ U open in X and α > 0 such that x̄ ∈ U
and B := Ux (f (x̄) − ² − α, f (x̄) − ² + α) ⊂ (epi(f
T ))c (see Fig. 10.11).
Hence, Ux (f (x̄) − ² − α, f (x̄) − ² + α) epi(f ) = ∅. Let x ∈ U and define
λ := f (x̄) − ² + α2 . Then (x, λ) ∈
/ epi(f ) since (x, λ) ∈ Ux (f (x̄) − ² − α, f (x̄) − ² + α).
Hence, λ < f (x), i.e., f (x̄) − ² + α2 < f (x), so that
α
f (x̄) − ² + < inf f (x).
2 x∈U {x̄}
This implies,
f (x̄) − ² < inf f (x) (why?)
x∈U {x̄}
< sup ( inf f (x)) (why?).

V ∈U(x̄) x∈V {x̄}
153
Fig. 10.11
Thus,
f (x̄) − ² < lim inf/ f (x).
x x0
since ² > 0 is arbitrary, we have

f (x̄) ≤ lim inf/ f (x),
x x0
and so, f is lower semi-continuous.

(1) ⇒ (3) is left as an easy exercise. ¤
Proposition 10.19 Let f : X −→ R ∪ {+∞} be any map. Then
f is convex and lower semi-continuous ⇔ f is convex and weakly lower semi-
continuous.
Proof.
f is convex and lower semi-continuous ⇔ epi(f ) is convex and closed
⇔ epi(f ) is convex and weakly closed
⇔ f is convex and weakly lower
semi-continuous. ¤
Corollary 10.20 Let f : X −→ R ∪ {+∞} be a convex and lower semi-continuous
mapping. Suppose {xn } is a sequence in X which converges weakly to x̄. Then
lim inf f (xn ) ≥ f (x̄).
154
10.5 A fundamental theorem of optimization
We now prove the main theorem of this chapter.
Theorem 10.21 Let X be a real reflexive Banach space and let K be a closed
convex bounded and nonempty subset of X. Let f : X −→ R ∪ {+∞} be lower
semi-continuous and convex. Then, ∃ x̄ ∈ K such that f (x̄) ≤ f (x) ∀ x ∈ K, i.e.,
f (x̄) = inf f (x) = min f (x).

x∈K x∈K
(See Fig 1.13)
Proof. f is lower semi-continuous and convex ⇒ f is weakly lower semi-continuous.

Put m := inf f (x).
x∈K
Fig. 10.12
First suppose m = −∞. Then for n ∈ N, ∃ xn ∈ K such that
f (xn ) < −n (10.2)
Boundedness of K implies {xn } is bounded ad Eberlein-Smul’yan Theorem implies

∃ {xnk } subsequence of {xn } such that xnk * x ∈ X. But K is convex and closed
⇒ K is weakly closed. Hence x ∈ K. By lower semi-continuity of f , we have
f (x) ≤ lim inf f (xnk ) < −∞ by inequality (10.2)
155
and this is impossible. Hence m ∈ R.
We now use the definition of ”inf”. Let n ∈ N and take ²n = n1 . Then ∃ xn ∈ K
such that m ≤ f (xn ) < m+ n1 . {xn } in K implies {xn } is bounded and so ∃{xnk }k∈N ,
subsequence of {xn } and x̄ ∈ K such that xnk * x̄. Since f is weakly lower semi-
continuous, we have
f (x̄) ≤ lim inf
/ f (xnk )
k +∞
1
≤ lim inf
/ (m + )
k +∞ nk
1
= lim/ (m + ) = m.
k +∞ nk
Thus, f (x̄) ≤ m = inf f (x). But m ≤ f (x̄). Hence,
x∈K
f (x̄) = m = inf f (x). ¤

x∈K
Next, we prove the second important theorem of this chapter.
Theorem 10.22 Let X be a real reflexive Banach space and f : X −→ R ∪ {+∞}

be a proper lower semi-continuous function. Suppose
lim / f (x) = +∞.

kxk ∞
Then, ∃ x̄ ∈ X such that f (x̄) ≤ f (x) x ∈ X, i.e.,

f (x̄) = inf f (x).
x∈X
Proof. We shall apply Theorem 1.21. Since f is proper, ∃ x0 ∈ X such that

f (x0 ) ∈ R. Let
K := {x ∈ X : f (x) ≤ f (x0 )}.
We now show K is closed convex nonempty and bounded, so that we can apply
Theorem 10.21. But K is convex and closed since f is convex and lower semi-
continuous. This implies K is a section with f (x0 ) = α.
CLAIM: K is bounded. Suppose this is not the case. Then for each n ∈ N, ∃ xn ∈ K
such that kxn k > n. Thus,
f (xn ) ≤ f (x0 ), kxn k > n. (10.3)
This implies lim / kxn k = +∞ and so (by hypothesis),
n ∞
lim / f (xn ) = +∞, (10.4)

n ∞
156
Fig. 10.13
contradicting inequality (10.3). Hence K is bounded. Theorem 1.21 then implies

∃ x̄ ∈ K ⊂ X such that ∀ x ∈ K,
f (x̄) ≤ f (x).
Now, let x ∈ XK. Then f (x) > f (x0 ). But x0 ∈ K. So, f (x̄) ≤ f (x0 ). Hence,
f (x) > f (x̄) ∀ x ∈ X, i.e.,
f (x̄) ≤ f (x), ∀ x ∈ X. ¤
Remark 10.23 If lim / f (x) = +∞, we say that f is coercive. This condition
kxk +∞
actually implies that
for each A > 0, ∃ B > 0 such that ∀ x ∈ X, kxk > B ⇒ f (x) > A.
Equivalently, considering the contra-positive statement, this means that f (x) ≤ A ⇒

||x|| ≤ B, i.e., if the range of A is bounded, then the domain of A is bounded.
Exercises 10.1
1. Let X be a real normed linear space. Prove that f : X −→ R ∪ {+∞} is

lower semi-continuous at x̄ ∈ X if and only if ∀ ² > 0, ∃ V ∈ U (x̄) such that
∀ x ∈ V {x̄},
f (x) > f (x̄) − ².
157
2. Let X be a real normed linear space. Prove that f : X −→ R ∪ {+∞} is lower
semi-continuous if and only if ∀ α ∈ R,
f −1 ((α, +∞)) is open,
or alternatively,
if and only if f −1 ((α, +∞]) is closed for each α ∈ R.
158
CHAPTER 11
An Application of Weak Topology in Fixed Point Theory
11.1 Some Fixed Point Theorems

Definition 11.1 Let (M, ρ) be a metric space. A mapping T : M / M is called
a Lipschitz map (or is said to be Lipschitzian) if there exists a number L ≥ 0 such
that
ρ(T x, T y) ≤ Lρ(x, y)
for all x, y ∈ M . Recall (Section 1.3) that the mapping T is called a strict contraction
(or simply a contraction) if L < 1, and is called nonexpansive if L ≤ 1. A point
x∗ ∈ M is called a fixed point of T if T x∗ = x∗ .
Problems concerning the existence of fixed points for Lipschitz maps have been of
considerable interest in Nonlinear Operator Theory. The study of nonlinear opera-
tors had its beginning about the start of the twentieth century with investigations
into the existence properties of solutions to certain boundary value problems aris-
ing in ordinary and partial differential equations. The earliest techniques, largely
devised by E. Picard [12], involved the iteration of an integral operator to obtain
solutions to such problems. In 1922, these techniques of Picard were given precise
abstract formulation by S. Banach [2] and R. Cacciopoli [8] in what is now generally
referred to as the Contraction Mapping Principle (Theorem 1.9).
The classical importance of fixed point theory in functional analysis is due to its
usefulness in the theory of ordinary and partial differential equations. The existence
159
or construction of a solution to a differential equation is often reduced to the ex-
istence or location of a fixed point for an operator defined on a subset of a space
of functions. Fixed point theorems have also been used to determine the existence
of periodic solutions for functional differential equations when solutions are already
known to exist. Apart from this deep involvement in the theory of differential equa-
tions, fixed point theorems have also been extremely useful in such problems as
finding zeros of nonlinear equations and proving surjectivity theorems. Partly as a
consequence of the importance of its applications, fixed point theory has developed
into an area of independent research.
The contraction mapping principle (Theorem 1.19) (which is involved in many of

the existence and uniqueness proofs of differential equations) is perhaps the most
useful fixed point theorem.
An earlier fixed point theorem, called the Brouwer Fixed Point Theorem, concerns
continuous mappings and has an advantage over Theorem 1.19 in that it applies to a
much larger class of functions. It is, however, in a sense, weaker than Theorem 1.19
because the sequence of iterates of the function at a given point need not converge
to a fixed point. Furthermore, it is confined to finite dimensional spaces.
Theorem 11.2 (Brouwer Fixed Point Theorem (1910)) Let B be the closed unit
ball of any finite dimensional Euclidean space and f : B / B be continuous.
Then f has a fixed point.
The first analytic attempt at generalizing Theorem 10.2 to infinite dimensional
spaces was made by Birkoff and Kellog [3]. They were able to show that a con-
tinuous operator defined from a compact, convex subset of C n [0, 1] into itself has a
fixed point. This result was then applied in solving certain differential and integral
equations. Further generalizations resulted in the following theorem.
Theorem 11.3 (Schauder–Tychonov Theorem) Let K be a compact convex subset
of a Banach space E. If T : K / K is continuous, then T has a fixed point.
Despite the fact that there is no known constructive technique for determining a
fixed point of T , the Schauder–Tychonov fixed point theorem is extremely impor-
tant in the proofs of many existence theorems of differential equations.
We have indicated that generalizations of the Brouwer Fixed Point Theorem to

infinite dimensional spaces have involved additional conditions either on the mapping
or on its domain. Observe that in Theorem 10.2 the closed unit ball B is compact.
Furthermore, since the compactness assumption of Theorem 10.3 is often difficult
160
to obtain in applications, considerable research has been done concerning weaker
conditions for the domain which guarantee the existence of a fixed point. In this
connection, the following example shows that if domains which are only bounded,
closed and convex are considered even a Lipschitz map with Lipschitz constant L = 1
may fail to have a fixed point.
Example 11.4 Let E = c0 and B = {x = {xn } ∈ c0 : 0 ≤ xn ≤ 1}.

Define T : B / E by
T (x1 , x2 , x3 . . .) = (1, x1 , x2 , x3 , . . .).
Then, (i) B is closed, bounded and convex; (ii) T maps B into B; (iii) kT x − T yk =
kx − yk for x, y ∈ B; (iv) T has no fixed point in B.
The easy verification of (i)–(iv) is left for the reader.
Remark 11.5 Observe that c0 is not uniformly convex and that the map T in
Example 8.4 is Lipschitzian with Lipschitz constant L = 1. In some situations,
Lipschitz maps with L = 1 may have fixed point (which may even be unique!).
Example 11.6 Let B ⊆ `2 be defined by B = {x = (x1 , x2 , . . .) ∈ `2 : kxk ≤ 1}.

Define T : B / `2 by T (x1 , x2 , x3 , . . .) = (0, x1 , x2 , x3 , . . .). Then, (i) T maps B
into B; (ii) kT x − T yk = kx − yk for all x, y ∈ B; (iii) T x∗ = x∗ if and only if
x∗ = (0, 0, 0, . . .).
11.1.1 A fixed point theorem

In this section, to illustrate some applications of some of the theorems proved in
Chapter 9, we shall establish the existence of a fixed point for a nonexpansive map
in reflexive Banach spaces. Theorem 9.5 (Kakutani’s characterization of reflexive
Banach spaces) will play a central role in our development. A corollary of our main
theorem will make non–trivial use of Theorem 9.21 (Milman–Pettis Theorem). We
now begin with some classical results with which the reader is probably already
familiar. In particular, we shall make use of Zorn’s Lemma. To state this lemma,
we need the following definitions.
Definition 11.7 Let S be a nonempty set. Suppose there is a linear relation be-
tween some (not necessarily all) pairs of elements of S and suppose this relation is
expressed symbolically by x ≤ y, x, y ∈ S such that the following properties hold:
(i) x ≤ x, x ∈ S; (ii) x ≤ y and y ≤ x, ⇒ x = y x, y ∈ S; (iii) x ≤ y and
y ≤ z ⇒ x ≤ z; x, y, z ∈ S. Then, the set S is said to be partially ordered by “≤”.
161
If, in addition, (S, ≤) is such that any pair of elements of S are comparable (i.e., for
any pair x, y ∈ S either x ≤ y or y ≤ x holds) then (S, ≤) is said to be completely
(or totally) ordered, or is called a chain.
Definition 11.8 Let (S, ≤) be a partially ordered set and B be a subset of S. An

element m ∈ S is called an upper bound of B if x ≤ m for each x ∈ B, and
a lower bound of B if m ≤ x for each x ∈ B. An element u ∈ S is said to be
maximal (or minimal) if x ∈ S, u ≤ x (or x ≤ u) ⇒ u = x.
Lemma 11.9 (Zorn’s Lemma) Let (S, ≤) be a partially ordered set and suppose that
every chain in S has an upper (lower) bound. Then S contains at least one maximal
(minimal) element.
The following well known lemmas will also be needed in what follows.
Lemma 11.10 (Principle of Nested Sequences) Let (M, ρ) be any metric space.
Then the following are equivalent:
(i) (M, ρ) is complete;
(ii) For any nested sequence of nonempty closed subsets of M , F1 ⊃ F2 ⊃ . . ., such

that
diam Fn / 0 as n / ∞,
we have, \
Fn 6= ∅.
n≥1
Lemma 11.11 Let (M, ρ) be a metric space. The following are equivalent:
(i) M is compact;
(ii) Every collection of closed nonempty subsets of M with the finite intersection
property has a nonempty intersection.
11.2 Normal structure in Banach spaces

We shall now introduce a geometric property of Banach spaces which will be needed
in our study of the existence of fixed points for nonexpansive mappings in reflexive
Banach spaces. Let C be a bounded subset of a Banach space, E. The diameter d
of C is defined by d := sup{kx − yk : x, y ∈ C}. A point x0 ∈ E is called a diametral
point of C if sup{kx0 −xk : x ∈ C} = d , and a point x0 ∈ E is called a non-diametral
162
point of C if sup{kx0 − xk : x ∈ C} < d. Let E be a Banach space and K a bounded
convex subset of E. Let d(K) denote the diameter of K. The set K will be called
non–trivial if d(K) > 0.
Definition 11.12 A bounded convex subset K of E is said to have normal structure

if every non–trivial convex subset C of K contains at least one nondiametral point,
i.e., there exists x0 ∈ E such that
sup{kx0 − xk : x ∈ C} < sup{kx − yk : x, y ∈ C} = d(C).
The Banach space E is said to have normal structure if every bounded convex subset
of E has normal structure. Geometrically, K is said to have normal structure if for
every non–trivial convex subset C of K there exists a ball centred at a point of C
and whose radius is less than the diameter of C such that the ball contains C, i.e.,
if C is an arbitrary non–trivial convex subset of K then
C ⊆ B(x0 , r)
for some x0 ∈ C and r < d(C).
We give some examples.
Proposition 11.13 Every uniformly convex Banach space E has normal structure.
Proof. Let K be an arbitrary bounded convex subset of E and let C be a convex

subset of K which has at least two distinct points, x1 , x2 ; x1 6= x2 (this implies
d(C) > 0). Set x0 = 21 (x1 + x2 ). Let diameter of C ≡ d. For any x ∈ C, set
u = x − x1 , v = x − x2 . Then given any ∈> 0, there exists δ > 0 (by uniform
convexity of E (see Proposition 9.20)) such that kuk ≤ d, kvk ≤ d and ku − vk ≥∈
imply that
°1 ° h ³ ∈ í
° °
° (u + v)° ≤ 1 − δ d,
2 d
i.e., h ³ ∈ í
kx − x0 k ≤ 1 − δ d < d.
d
Consequently,
sup{kx − x0 k : x ∈ C} < d ,
and so C has a nondiametral point xo and since C is arbitrary, it follows that K has
normal structure.
¤
163
Proposition 11.14 (Brodskii and Mil’man Theorem [5]) Every compact convex
subset K of a Banach space E has normal structure.
Proof. The proof is by contradiction. So, assume K does not have normal structure.
The method of proof is to construct a sequence in K which does not have a conver-
gent subsequence. This, of course, will contradict the compactness of K. Let C be
a convex subset of K which has at least two points. Since K does not have normal
structure, all points of C are diametral. Let d > 0 denote the diameter of C. We
shall construct a sequence {xi }∞
i=1 in C such that kxi − xj k = d; i, j = 1, 2, . . . , i 6= j.
( Since C contains at least two distinct points x1 , x2 , then ||x1 − x2 || = d). To do
this we choose x1 ∈ C arbitrary and assume that x2 , x3 , . . . , xn have already been
chosen such that kxi − xj k = d; i, j = 1, 2, . . . , n. By the convexity of C,
1
(x1 + x2 + . . . + xn ) ∈ C,
n
and so it is a diametral point (by assumption). By compactness, sup{kx−yk : x, y ∈
K} is achieved. So, there exists xn+1 ∈ C such that
° °
° x + x + . . . + x °
°xn+1 − 1 2 n ° = d.
° n °
Hence,
1
d = k(xn+1 − x1 ) + (xn+1 − x2 ) + . . . + (xn+1 − xn )k
n
n
1 X 1
≤ kxn+1 − xj k ≤ · nd = d.
n j=1 n
Consequently, kxn+1 − xj k = d for j = 1, 2, . . .. This implies that the sequence

{xn }∞n=1 has no convergent subsequence, contradicting the compactness of K and
completing the proof.
¤
It is interesting to note that certain Banach spaces do not have normal structure.
We give an example.
Example 11.15 The space C[0, 1] with “sup” norm does not have normal structure.
To see this, consider the subset K of E defined by
K = {f ∈ C[0, 1] : f (0) = 0, f (1) = 1, 0 ≤ f (t) ≤ 1 for all t ∈ [0, 1]}.
Then, (i) K is bounded and convex; (ii) K is clearly a convex subset of itself; (iii)
K does not have normal structure.
164
The verifications are trivial. Observe that for arbitrary f ∈ K, kf k = 1. Moreover,
for f1 , f2 ∈ K arbitrary, and λ ∈ [0, 1], let F = λf1 + (1 − λ)f2 . Then, F (0) =
0, F (1) = 1 and 0 ≤ F (t) ≤ 1 ∀ t ∈ [0, 1]. Each point of K is diametral since if
f0 ∈ K,
sup{kf0 − f k : f ∈ K} = 1 = diameter of K.
Hence, K does not have normal structure.
We shall also need the following definition.

Definition 11.16 Let S be a nonempty subset (not necessarily convex) of a normed
linear space, E. The set
( n n
)
X X
coS = λi xi : xi ∈ S, λi = 1, n = 0, 1, 2, . . .
i=1 i=1
which consists of all the convex combinations of elements of S is called the convex
hull of S (and is generally denoted by coS). Clearly, coS is the smallest convex set
that contains S and, in fact, it is the intersection of all convex sets containing S.
Observe that if S is convex then coS = S. The closure of coS, generally denoted by
coS, is called the closed convex hull of S and it is the intersection of all closed convex
sets containing S (i.e., the smallest closed convex set that contains S). Again if S
is a closed convex set, we have,
S = coS = coS.
11.3 A fixed point theorem for nonexpansive map-

pings
We are now ready to prove the main theorem of this chapter.
Theorem 11.17 (W.A. Kirk [11]) Let E be a reflexive Banach space and K a closed
convex bounded and nonempty subset of E with normal structure. Let U : K /K
be nonexpansive. Then U has a fixed point in K.
Proof. Let S denote the family of all closed convex bounded and nonempty subsets
of E which are invariant under U , i.e., S = {Kα : Kα is closed convex bounded and
nonemtpy, and U (Kα ) ⊆ Kα for each α ∈ ∆, where ∆ is an arbitrary index set }.
Clearly S 6= ∅ since K ∈ S. Partially order S by set inclusion in an obvious way,
i.e.,
Kα ≤ Kβ ; Kα , Kβ ∈ S if Kα ⊆ Kβ .
165
Claim S is inductive (i.e., every chain in S has a lower bound). To verify this, let
Z be an arbitrary chain in S. Consider the intersection
\
K∗ = Kα .
Kα ∈Z
Then, clearly, K ∗ is: (i) closed; (ii) convex; (iii) bounded; (iv) nonempty.
Moreover, (v) U (K ∗ ) ⊂ K ∗ .
166
Verifications
(i) K ∗ is closed, (as an arbitrary intersection of closed sets).
(ii) Let x, y ∈ K ∗ be arbitrary and let λ ∈ [0, 1]. Then x, y ∈ Kα for each Kα ∈ Z.
Since each Kα is convex (by hypothesis),
T it follows that λx + (1 − λ)y ∈ Kα
for each α, and so λx + (1 − λ)y ∈ Kα = K ∗ , establishing (ii).
Kα ∈S
(iii) K ∗ ⊆ Kα for each Kα ∈ Z, and so the boundedness of K ∗ follows from the

boundedness of Kα .
(iv) Since E is reflexive and K is closed convex bounded and nonempty we have
(by a consequence of Kakutani’s theorem (Corollary 9.8)) that K is weakly
compact. Each Kα is closed and convex and so is weakly closed (Theorem
8.19). Furthermore Z is a chain and each Kα is nonempty so Z has finite
intersection property. It now follows (from Lemma 11.10) that K ∗ is nonempty.
T
(v) Let x ∈ Kα = K ∗ be arbitrary. It suffices to show U x ∈ K ∗ . But
T Kα ∈Z
x∈ Kα ⇒ x ∈ Kα for each Kα ∈ Z. By hypothesis, U x ∈ Kα for each
Kα ∈S T
Kα ∈ S. Hence U x ∈ Kα = K ∗ .
Kα ∈Z
Hence K ∗ ∈ S. Since K ∗ ⊂ Kα for each Kα ∈ Z, K ∗ is a lower bound for Z. By

Zorn’s lemma, S has a minimal element. Call it K 0 . Now, by the definition of S, K 0
is closed convex bounded, nonempty and is invariant under U . Hence, U (K 0 ) ⊂ K 0
so that co U (K 0 ) ⊂ co K 0 = K 0 (since K 0 is closed, see Definition 11.16). But K 0
is minimal. So,
co U (K 0 ) = K 0 .
Observe that so far we have not made use of the hypotheses that (i) K has normal
structure, and (ii) U is nonexpansive. Both conditions will now be used in proving
the following claim.
Claim K 0 is a singleton. We establish this claim by contradiction. So assume

that the diameter d of K 0 is positive (i.e., K 0 contains at least two distinct points).
Since K has normal structure, there exists at least one point x∗ in K 0 such that
K 0 ⊆ B(x∗ , d1 ) for some d1 < d. Define
K 1 = {x ∈ K 0 : K 0 ⊂ B(x, d1 )}
\
= K0 ∩ ( B(ω, d1 )).
ω∈K 0
167
Then, K 1 is a closed convex subset of K 0 . Furthermore, K 1 6= ∅ (since x∗ ∈ K 1 ).
The condition d1 < d implies K 1 6= K 0 (verify). We now show U (K 1 ) ⊆ K 1 . Let
y0 ∈ K 1 be arbitrary. We want to show U y0 ∈ K 1 . It suffices to show
(a) U y0 ∈ K 0 and , (b) K 0 ⊆ B(U y0 , d1 ).
For (a), y0 ∈ K 1 ⇒ y0 ∈ K 0 ⇒ U y0 ∈ U (K 0 ) ⊂ K 0 , establishing (a). For (b),

y0 ∈ K 1 ⇒ y0 ∈ K 0 and so U y0 ⊂ U (K 0 ) ⊂ K 0 since K 0 is invariant under U .
Moreover, for arbitrary v ∈ K 0 , (we now use the fact that U is nonexpansive),
kU y0 − U vk ≤ ky0 − vk ≤ d1 .
This implies, U (K 0 ) ⊆ B(U y0 , d1 ) since v ∈ K 0 is arbitrary. This implies coU (K 0 ) ⊆

co B(U y0 , d1 ) = B(U y0 , d1 ), since B(U y0 , d1 ) is closed, and so, K 0 = coU (K 0 ) ⊆
B(U y0 , d1 ), verifying (b). Thus, K 1 is invariant under U . The boundedness of
K 1 follows trivially from that of K 0 . We have now proved that K 1 is a closed
convex nonempty and bounded proper subset of K 0 which is invariant under U .
This contradicts the minimality of K 0 in S. So, K 0 is a singleton, say {x∗ }. Since
it is invariant under U , it is a fixed point of U , i.e., U (x∗ ) = x∗ . This completes the
proof. ¤
Corollary 11.18 (F.E. Browder [6]; Göhde [10]) Let E be a uniformly convex Ba-
nach space and let K be a closed convex bounded nonempty subset of E. Suppose
that T : K / K is nonexpansive. Then T has a fixed point in K.
Proof. The corollary follows immediately from Theorem 10.17 on observing that
K has normal structure (Proposition 10.13) and E is reflexive (Theorem 9.21).
¤
Remark 11.19 We had observed that the Contraction Mapping Principle is, per-
haps, the most important fixed point theorem. Certainly, it is now one of the most
durable and fruitful methods in mathematical analysis. Recall that a contraction
map T : (M, ρ) / (M, ρ) of a metric space into itself is one that satisfies the
following condition: for each x, y ∈ M , there exists k ∈ [0, 1) such that
ρ(T x, T y) ≤ kρ(x, y).
It is then clear that nonexpansive maps, (those satisfying this inequality with k ∈
[0, 1]), can be considered as those maps lying at the boundary of the contraction
maps. Apart from being an obvious generalization of contraction mappings, interest
in such mappings stems mainly from the following two facts (see e.g., Bruck [7]):
168
Fact 1: Nonexpansive mappings appear in many applications as the transition op-
erators for initial–value problems of differential inclusions of the form
du
f∈ + A(t)u,
dt
where the operators {A(t)} are generally minimally “continuous” and are, in some
sense “positive”.
Fact 2: Nonexpansive mappings are intimately connected with “monotonicity meth-

ods”, and they constitute one of the first classes of mappings for which the geometric
structure of the underlying Banach space was central in obtaining fixed point results
instead of using only compactness properties.
Research on nonexpansive mappings has continued to be a flourishing area for many

mathematicians. For a good exposition of fixed point theory for nonexpansive map-
pings and some applications the reader may consult [7].
169
170
CHAPTER 12
Function Spaces; Weak Derivatives
13.1 Notations
Throughout this Chapter, Ω will represent some nonempty open subset of IRn (n ≥ 1
integer) with boundary ∂Ω. We will write K ⊂⊂ Ω if K is a compact subset of Ω.
A generic point in IRn is denoted by x = (x1 , x2 , ..., xn ), its Euclidean norm being
³P ´1
2 2
||x|| := |xi | . By a multi-index α, we mean an n-tuple α = (α1 , α2 , ..., αn ) of
nonnegative integers αj , i.e., α ∈ IN n . To the multi-index α, we associate the integer
P |α|
|α| := nj=1 αj ; and a differential operator Dα := ∂xα1 ∂x∂α2 ...∂xαn .
1 2 n
13.2 Spaces of Functions

With Ω, we associate various spaces of real-valued functions. Among these spaces
the “smallest” one is the so-called spaces of test functions , denoted by D(Ω), and
defined by D(Ω):=C0∞ (Ω) := {ϕ : Ω / IR such that ϕ is infinitely differentiable
on Ω and ϕ has (with respect to the topology of IRn ) compact support on Ω}.
Furthermore, the “biggest” space of interesting functions is the space of locally
integrable functions, defined as follows:
L1,loc (Ω) := {v : Ω / IR such that ϕv ∈ L1 (Ω)∀ϕ ∈ D(Ω)}
= {v such that vXK ∈ L1 (Ω)∀K ⊂ Ω, K compact inIRn }, (12.1)
171
where XK is the characteristic function of the set K and L1 (Ω) is the space of
class of measurable functions on Ω which are Lebesgue integrable. In other words,
L1,loc (Ω) := {f : Ω / IR such that f is a measurable function on Ω and f is in-
tegrable on every K ⊂⊂ Ω}. To emphasize the “smallest” and “biggest” size of the
spaces D (Ω) and L1,log (Ω) respectively, we have the following embeddings between
the usual function spaces.
D(Ω) ≡ C0∞ (Ω) ⊂ C0m (Ω) ⊂ C m (Ω̄) ⊂ C m (Ω)

∩ ∩
Lp (Ω) ⊂ L1,loc (Ω) ⊆ D0 (Ω).
Fig 11.1. The smallest and the biggest classical spaces.
We now recall the definitions of the spaces in Fig 11.1. We first remark that it is
conventional to write C 0 (Ω) :≡ C(Ω). Then for a fixed integer m ≥ 1,
(i) C m (Ω) := {f : Ω / IR such that D α f is continuous on Ω, ∀α ∈ IN n , |α| ≤

m}.
(ii) C0m (Ω) := {f ∈ C m (Ω) such that suppf ⊂⊂ Ω}, where suppf =support of f :=
the closure of the set of points at which f is nonzero.
(iii) C m (Ω̄) := {f : C m (Ω) such that Dα f has continuous extension to Ω̄, ∀α ∈

IN n , |α| ≤ m}.
(v) C ∞ (Ω) := ∩k≥0 C k (Ω) and C ∞ (Ω̄) := ∩k≥0 C k (Ω̄).
For 1 ≤ p ≤ ∞, Lp (Ω) is the RLebesgue space defined as the set of class of measurable
functions f on Ω such that Ω |f (x)|p dx < ∞, if p < ∞, and there exists A such
that µ(Ω \ A) = 0 and sup |f (x)| < ∞, if p = ∞, where µ denotes the measure of
x∈A
Ω \ A.
Remarks
1. The space C m (Ω̄) is a Banach space under the norm
||f ||m,∞,Ω := max supx∈Ω |Dα f (x)|.

|α|≤m
172
2. The space LP (Ω) is a Banach space under the norm
 ³ ´ p1
 R p
||f ||0,p,Ω := Ω |f (x) dx , if p < ∞
 ess sup |f (x)|, if p = ∞,
x∈Ω
where ess supx∈Ω |f (x)| := inf{K ≥ 0 such that |f (x)| ≤ K a.e. on Ω} < ∞,
if p = ∞.
We simply write ||f ||0,Ω where p = 2. In this particular case, L2 (Ω) is a Hilbert
R
space, the inner product being defined by hf, gi0,Ω := Ω f (x).g(x)dx.
Example 12.1 Let Ω ⊂ IR3 . We examine what the elements of C 2 (Ω) look like. Re-
call, C 2 (Ω) := {f : Ω / IR such that D α f is continuous on Ω, ∀α ∈ IN n , |α| ≤ 2}.
3
Now x ∈ Ω ⊂ IR so x = (x1 , , x2 , x3 ) for x1 , x2 , x3 ∈ IR. Hence, a generic α is given
by α = (α1 , α2 , α3 ) with |α| = α1 + α2 + α3 ≤ 2. With this restriction the possible
value of α are:
(i) (0, 0, 0) (ii) (1, 1, 0), (iii) (0, 1, 1) (iv) (1, 0, 1) (v) (2, 0, 0), (vi) (0, 2, 0)
(vii) (0, 0, 2) (viii) (0, 1, 0) (ix) (1, 0, 0) (x) (0, 0, 1).
For
(i) α = (0, 0, 0), |α| = 0 and Dα f = f .
∂2f ∂2f
(ii) α = (1, 1, 0), |α| = 2 and Dα f = α α α
∂x1 1 ∂x2 2 ∂x3 3
= ∂x1 ∂x2
.
∂2f ∂2f
(iii) α = (0, 1, 1), |α| = 2 and Dα f = ∂x01 ∂x2 ∂x3
= ∂x2 ∂x3
.
∂2f
(iv) α = (1, 0, 1), |α| = 2 and Dα f = ∂x1 ∂x3
.
∂2f
(v) α = (2, 0, 0), |α| = 2 and Dα f = ∂x21
.
∂2f
(vi) α = (0, 2, 0), |α| = 2 and Dα f = ∂x22
.
173
∂2f
(vii) α = (0, 0, 2), |α| = 2 and Dα f = ∂x23
.
∂f
(viii) α = (0, 1, 0), |α| = 1 and Dα f = ∂x2
.
∂f
(ix) α = (1, 0, 0), |α| = 1 and Dα f = ∂x1
.
∂f
(x) α = (0, 0, 1), |α| = 1 and Dα f = ∂x3
.
Thus C 2 (Ω) consists of all real-valued functions defined on Ω which are such that
all their partial derivatives up to and including order 2 are continuous on Ω.
Example 12.2 Let Ω ⊆ IR2 . We examine the elements of C 3 (Ω) := {f : Ω / IR

such that Dα f is continuous on Ω ∀α ∈ IN n such that |α| ≤ 3}. Now, x ∈ Ω ⊆ IR2 ,
so x = (x1 , x2 ) for x1 , x2 ∈ IR. Furthermore, α ∈ IN n ≡ IN 2 and so a generic α
is given by α = (α1 , α2 ), αi ∈ IN , i = 1, 2. Since |α| ≤ 3, the possible values of α are:
(i) (0, 0) (ii) (1, 0), (0, 1) (iii) (1, 1), (2, 0), (0, 2) (iv)(3, 0), (0, 3), (1, 2), (2, 1)
For
(i) α = (0, 0), |α| = 0, Dα f = f .
(ii) α = (1, 0), |α| = 1, Dα f = ∂f

∂1
α ∂f
α = (0, 1), |α| = 1, D f = ∂x 2
∂2f
(iii) α = (1, 1), |α| = 2, Dα f = ∂x1 ∂x2
∂2f
α = (2, 0), |α| = 2, Dα f = ∂x21
∂2f
α = (0, 2), |α| = 2, Dα f = ∂x22
∂3f
(iv) α = (3, 0), |α| = 3, Dα f = ∂x31
∂3f
α = (0, 3), |α| = 3, Dα f = ∂x32
∂3f
α = (1, 2), |α| = 3, Dα f = ∂x1 ∂x22
∂3f
α = (2, 1), |α| = 3, Dα f = ∂x21 ∂x2
.
174
Again, C 3 (Ω), Ω ⊆ IR2 , consists of all real-valued functions whose partial derivatives
of all orders up to and including order 3 are continuous on Ω.
Remark 12.3 Functions in the smallest space D(Ω) have several nice properties.
They are integrable, differentiable, bounded, .... However, they are not suitable
conditions for solutions of most partial differential equations. For example, in the
electrostatic model
½
− 4 u = f in Ω
u=0 on ∂Ω,
it is not realistic to assume that the solution u is in D(Ω), i.e., that the solution u
is infinitely differentiable on Ω and has compact support in Ω. On the other hand,
the biggest space L1,loc (Ω) contains “pathological” functions which, nonetheless, are
good candidates for solutions of P DEs. To appreciate this fact, let us consider the
beam problem,
½
−u00 = f, on (−1, 1)
u(−1) = u(1) = 0,
where the force f is the famous Heaviside function

½
1, if x > 0
f (x) =
0, if x ≤ 0.
The function f , which clearly belongs to Lp,loc (−1, −1), 1 < p < ∞, is pathological
as a non-continuous and hence non differentiable function.
To overcome these difficulties associated with L1,loc (Ω), we shall define a new space
D0 (Ω) which contains L1,loc (Ω) and for which elements enjoy, in a weak sense , sev-
eral of the nice properties of the functions in D(Ω).
To introduce what is meant by the weak sense, we first endow L1,loc (Ω) and D(Ω)
with the structure of reflexive, Hausdorff, complete, locally convex topological vector
spaces. For D(Ω), the topology in question is referred to as the canonical topology of
L-Schwartz. Actually, we do not need here the explicit definition of these topologies;
only the sense of convergence of sequences in L1,loc (Ω) and D(Ω) is of interest to us
here.
175
13.3 Pseudo Topologies
Definition 12.4 (Pseudo topology of C ∞ (Ω)). We say that a sequence {ϕn } in
C ∞ (Ω) converges to ϕ ∈ C ∞ (Ω) if for any K ⊂⊂ Ω, Dα ϕn / D α ϕ uniformly on
K for all multi-indices α.
Definition 12.5 (Pseudo topology of D (Ω) = C0∞ (Ω)). A sequence {ϕn } in D (Ω)
converges to ϕ ∈ D (Ω) if there exists K ⊂⊂ Ω, such that (a) suppϕn ⊂ K, ∀n ≥ 1,
suppϕ ⊂ K; and (b) Dα ϕn / D α ϕ uniformly on K for all multi-indices α.
Definition 12.6 (Pseudo topology of L1,loc (Ω)). A sequence {fn } in L1,loc (Ω) con-
verges to f ∈ L1,loc (Ω) if for any K ⊂⊂ Ω, ||fn − f ||L1 / 0 as n / ∞ or
equivalently, if for every ϕ ∈ D (Ω), ||fn ϕ − f ϕ||L1 / 0 in K as n / ∞.
Example 12.7 (a) If {ϕn } converges to ϕ in D(Ω), then {ϕn } converges to ϕ in

L1,loc (Ω). Thus, the embedding D (Ω) ⊂ L1,loc (Ω) is continuous.
(b) Consider ϕ ∈ D (Ω). Define ϕn ∈ D (Ω) by ϕn = n1 ϕ. Then ϕn converges to
zero in D (Ω) = C0∞ (Ω).
13.4 Distributional or Weak Derivatives

There are several ways to generalize the idea of the usual derivatives. In this section,
we study the concept of distributional (or weak) derivatives . This concept had been
used in Physics before L. Schwartz gave it a mathematical structure. The definition
of distributional derivative given by Schwartz came from the following simple idea.
Let f : IR / IR be differentiable and ϕ be any function in C ∞ (IR). Since ϕ has

0
compact support, we can choose some open interval (a, b) where −∞ < a, b < ∞
such that suppϕ ⊂ (a, b). Then,
Z Z Z b
0 0
f ϕdx = f (x)ϕ(x)dx = f 0 (x)ϕ(x)dx
IR IR a
Z b
= f (x)ϕ(x)|ba − f (x)ϕ0 (x)dx
a
Z b
= − f (x)ϕ0 (x)dx, since ϕ(a) = ϕ(b) = 0.
a
Rb Rb
So, a f 0 ϕdx = − a f ϕ0 dx. Following this rule, if f has higher derivatives then we
have
Z Z
(n) n
f ϕdx = (−1) f ϕ(n) dx ∀ϕ ∈ D(IR).
IR IR
176
More generally, if f ∈ C |α| (Ω), then by integration by parts we obtain the following
relation:
Z Z
α |α|
D f ϕdx = (−1) f Dα ϕdx ∀ϕ ∈ D(Ω). (12.2)
Ω Ω
We observe immediately that the right side of the above equation is defined even
when f is not differentiable but is just locally integrable. This gives us a method of
generalizing the concept of derivative and leads to the following definition.
Definition 12.8 Let f, g ∈ L1,loc (Ω). Then g is called the distributional (or weak)
derivative of f , of order α, denoted by Dwα f if
Z Z
|α|
gϕdx = (−1) f Dα ϕdx ∀ϕ ∈ D(Ω). (12.3)
Ω Ω
Compare formulas (12.2) and (12.3) to see that these two notions of derivative
coincide when derivative exists in the usual sense.
Example 12.9 Let f (x) = |x| ∀x ∈ IR. In the usual sense, f 0 does not exist at
x = 0. Define the function,

 1, if x > 0
sgn(x) := 0, if x = 0

−1, if x < 0.
Let ϕ ∈ D(IR) and suppϕ ⊂ (−n, n) where n is some positive integer. Then,
Z Z Z n
0 0
f (x)ϕ (x)dx = |x|ϕ (x)dx = |x|ϕ0 (x)dx
IR IR −n
Z 0 Z n
= −xϕ0 (x)dx + xϕ0 (x)dx
−n 0
Z 0 Z n
= ϕ(x)dx − ϕ(x)dx, (integration by parts)
−n 0
Z n Z
= − sgn(x)ϕ(x)dx = − sgn(x)ϕ(x)dx.
−n IR
Hence, Dw1 |x| = sgn(x).
Definition 12.10 Consider the Heaviside function H defined by

½
1, if x > 0
H(x) =
0, if x ≤ 0.
177
Let ϕ ∈ D(R) and suppϕ ⊂ (−n, n). Then,
Z Z n
0
Hϕ dx = 1.ϕ0 (x)dx
IR 0
= ϕ(n) − ϕ(0), (integration by parts)
= −ϕ(0), since ϕ(n) = 0
:= −hδ, ϕi,
where δ, known as the Dirac delta function is defined by δ(ϕ) = ϕ(0).
We observe that the distributional (or weak) derivative of the Heaviside function
does not exist. However, to the Heaviside function H corresponds a distribution TH .
The derivative of this distribution exists and is equal to δ. We do not explore this
here. For more on distributions the reader may consult [1].
178
CHAPTER 13
An Introduction to Sobolev Spaces and Applications
14.1 Introduction
In the classical theory of Partial Differential Equations (PDEs), we seek the solution
u which satisfies some relation among its partial derivatives and u in some open sub-
set Ω of IRn . The existence and uniqueness of the solution of well-known problems
of PDEs can be proved if the given boundary conditions are smooth. This classical
theory fails if the boundary conditions fail to be smooth or if the right hand side
is not smooth enough, even on some finite points. To handle this situation, it is
necessary to generalize the idea of the usual differentiation. This generalization is
the so-called distributional (or weak) derivatives, which we introduce in this chapter.
Next we show that solutions of PDEs lie in a special class of Banach spaces called
Sobolev spaces. There are several Sobolev spaces associated with any open subset
Ω ⊂ IRn . Two special kinds that appear in most problems are H m (Ω) and H0m (Ω)
and both are separable Hilbert spaces. This gives a way to convert our original
problem related to the existence and uniqueness of solution, into a problem that
can be solved from the theory of Hilbert spaces, e.g., by means of a theorem called
Lax- Migram theorem. In this chapter, we first state and prove this theorem. Then
we introduce Sobolev spaces and use some of their properties, in conjunction with
Lax-Migram Theorem to solve the Dirichlet problem; Neumann problem and the
Bi-harmonic equation.
179
14.2 Abstract Minimization Problem
We begin with the notion of a bilinear functional.
Definition 13.1 Let E be a complex vector space. A mapping a(., .) : E ×E /CI
is called a bilinear form if
(a) a(αx + βy, z) = αa(x, z) + βa(y, z); for all x, y, z ∈ E and α, β ∈ C I; and
(b) a(x, λu + µv) = λ̄a(x, u) + µ̄a(x, v), for all x, u, v ∈ E and for all scalars λ, µ ∈ CI.
Note that a bilinear form is linear with respect to the first variable and anti-linear
with respect to the second variable. Of course, in a real vector space, bilinear
functionals are linear in both variables.
Example 13.2 (a) Inner product is a bilinear form.
(b) Let f, g : E / IR be linear functionals on a vector space E. Then a(x, y) =
f (x)g(x) is a bilinear form on E.
(c) Let P, Q : H / H be linear operators on a Hilbert space. Then a(., .) : H × H
/C
I defined by
a(u, v) = (P u, v)
or = (u, Qv)
or = (P u, Qv),
is a bilinear form.
Definition 13.3 Let E be a vector space. A bilinear form a(., .) : E × E /C

I is
called symmetric if a(x, y) = a(y, x) for all x, y ∈ E. If E is a normed space, then
a(., .) is called bounded (or continuous) if there exists some constant M > 0 such
that |a(x, y)| ≤ M ||x||.||y|| for all x, y ∈ E.
14.2.1 The Symmetric Case

Many problems in elasticity, for example, are mathematically represented by the
following minimization problem: Find an unknown u (which is the displacement of
a mechanical system) such that,
u ∈ K and Ju = inf Jv,
v∈K
where the set K of admissible displacements is a closed convex subset of a Hilbert

space H, and the energy J of the system takes the form
1
J(v) = a(v, v) − f (v),
2
180
where a(., .) is a symmetric bilinear form defined and continuous on H × H and f
is a linear form, defined and continuous over the space H.
Let X be a real normed linear space with norm ||.||. Let a(., .) : X × X / IR
be a continuous bilinear form and f : X / IR be a continuous linear form. For
∅ 6= K ⊂ X, we, in general, associate an abstract minimization problem as follows:
Find an element x∗ ∈ X such that:
(i) x∗ ∈ K;
(ii) Jx∗ = infx∈K Jx, where J : X / IR is defined by
1
J(x) := a(x, x) − f (x) ∈ IR.
2
For a solution of this abstract minimization problem we have the following theorem.
Theorem 13.4 Assume in addition that the following conditions hold:
(a) X is complete;
(b) K is closed and convex;
(c) the bilinear form a(., .) is symmetric and;
(d) there exists α > 0 such that a(x, x) ≥ α||x||2 ∀x ∈ X. Then, the abstract
minimization problem has one and only one solution.
Proof. Condition (c) implies the bilinearpform a(., .) is an inner product on X with
the associated norm given by ||x||X := a(x, x), which is equivalent to the given
norm ||.||. Hence X is a complete inner product space and so is a Hilbert space. We
now denote it by H. By Riesz representation theorem, for each x ∈ H, there exists
a unique σf ∈ H such that f (x) = a(σf, x). This implies that
1 1
Jx = a(x, x) − f (x) = a(x, x) − a(σf, x)
2 2
1 1
= a(x − σf, x − σf ) − a(σf, σf ),
2 2
so that
1 1
min Jx = min{a(x − σf, x − σf )} − a(σf, σf ).
x∈K 2 x∈K 2
Hence, solving the abstract minimization problem amounts to minimizing the dis-
tance
p between the unique element σf and the set K, with respect to the norm
a(., .) = ||.||. Consequently, the solution is simply the projection of the element
σf onto K with respect to the inner product a(., .). By the projection theorem, such
a projection exists and is unique, since K is a closed and convex subset of H.
¤
181
Definition 13.5 A linear form a(., .) : X × X /CI which satisfies the condition
2
a(x, x) ≥ α||x|| ∀x ∈ X and for some α > 0 is said to be X-elliptic on X.
Variational formulations.
The variational forms of the abstract minimization problem are summarized in the
following theorem.
Theorem 13.6 An element x∗ is the solution of the abstract minimization problem:

x∗ ∈ K and Jx∗ := infv∈K J(v) if and only if x∗ satisfies the following relations:
(a) x∗ ∈ K and ∀v ∈ K,
a(x∗ , v − x∗ ) ≥ f (v − x∗ ),
where K is a closed convex subset of H, or,

(b) x∗ ∈ K and ∀v ∈ K, a(x∗ , v) ≥ f (v), and a(x∗ , x∗ ) = f (x∗ ), where K is a
closed convex cone with vertex 0, or,
(c) x∗ ∈ K and ∀v ∈ K,
a(x∗ , v) = f (v),
where K is a closed subspace of H.
Proof. (a) The projection x∗ is characterized by the relation x∗ ∈ K and ∀v ∈ K,
a(σf − x∗ , v − x∗ ) ≤ 0.
This implies, a(x∗ , v−x∗ ) ≥ a(σf, v−x∗ ) = f (v−x∗ ), since ∀v ∈ H, f (v) = a(σf, v),
by Riesz representation theorem, and clearly, v − x∗ ∈ H. This establishes (a).
(b) K is a closed convex cone with vertex 0 implies that x∗ ∈ K and if v ∈ K

then x∗ + v ∈ K. From part (a), a(x∗ , v − x∗ ) ≥ f (v − x∗ ). Replacing v by (x∗ + v)
in this inequality we obtain that a(x∗ , v) ≥ f (v) ∀v ∈ K, establishing the first part
of (b). In particular, if v = x∗ , then a(x∗ , x∗ ) ≥ f (x∗ ). Also, by setting v = 0 in
a(x∗ , v − x∗ ) ≥ f (v − x∗ ) we get a(x∗ , −x∗ ) ≥ f (−x∗ ) which yields f (x∗ ) ≥ a(x∗ , x∗ ).
Hence we obtain that a(x∗ , x∗ ) = f (x∗ ), establishing the second part of (b).
(c) Since K is a subspace of H, replacing v by −v in the relation a(x∗ , v) ≥ f (v),

and using the linearity of f we obtain that a(x∗ , v) ≤ f (v) for all v ∈ K, so that
a(x∗ , v) ≥ f (v) ≥ a(x∗ , v), i.e., a(x∗ , v) = f (v) ∀v ∈ H, establishing (c). This
completes the proof.
¤
182
14.2.2 The Non-symmetric case
The abstract minimization problem can also be formulated without making explicit
reference to the functional J. In this case, the problem is called an abstract varia-
tional problem and it is formulated as follows:
(a) Find an element x∗ such that
x∗ ∈ K and ∀v ∈ K, a(x∗ , v − x∗ ) ≥ f (v − x∗ ),
where K is a closed convex subset of H, or, (b) find an element x∗ such that
x∗ ∈ K and ∀v ∈ K, a(x∗ , v) ≥ f (v) and a(x∗ , x∗ ) ≥ f (x∗ ),
if K is a cone with vertex 0,

or, (c) find an element x∗ such that
x∗ ∈ K and ∀v ∈ K, a(x∗ , v) = f (v),
if K is a subspace of X.
We have seen from Theorem 12.4 that each such problem has one and only one
solution if X is complete, the subset K is closed and convex, and the bilinear form
is X-elliptic, continuous and symmetric.
If the assumption of symmetry of the bilinear form is dropped, it has been proved
that the above variational problem still has one and one solution, but in this case,
there is necessarily no longer an associated minimization problem.
14.2.3 The Lax-Milgram Theorem

Confining ourselves here to the case in which K = X, we have the following impor-
tant theorem.
Theorem 13.7 (Lax-Milgram lemma) . Let H be a real Hilbert space and

a(., .) : H × H / IR be a continuous H-elliptic bilinear form, and let f : H
/ IR be a continuous linear form. Then,
(a) there exists a unique u ∈ H such that
a(v, u) = f (v) ∀v ∈ H;
183
(b) if, in addition, a(., .) is symmetric and a(v, v) ≥ 0 ∀v ∈ H, then the function
J :H / IR defined by
1
J(v) := a(v, v) − f (v)
2
has minimum value at u given in part (a).
Proof. By the H-ellipticity and continuity of a, there exist constants M ≥ 0, α > 0

such that
|a(u, v)| ≤ M ||u||.||v||∀u, v ∈ H, (13.1)

|a(u, v)| ≥ α||v||2 ∀v ∈ H. (13.2)
From (13.2), it follows that α||v||2 ≤ a(u, v) ≤ M ||u||2 and thus
a(v, v) = 0 if and only if v = 0. (13.3)
Let w0 ∈ H be arbitrary but fixed. Define the map
Aw0 :H / IR by
Aw0 (v) = a(v, w0 )∀v ∈ H. (13.4)
Claim 1. Aw0 is a bounded linear functional on H. Hence, by Riesz representation

theorem, there exists a unique z0 ∈ H such that
Aw0 (v) (= a(v, w0 )) = hv, z0 i∀v ∈ H.
Hence, for arbitrary w ∈ H, there exists a unique v ∈ H such that
a(v, w) = hv, zi∀v ∈ H. (13.5)
Now, define the map A : H / H by
A(w) = z.
Combining this definition with equation (13.5) we obtain the following relation,
a(v, w) = hv, zi = hv, A(w)i = ∀w, v ∈ H. (13.6)
Claim 2.
(i) A is linear;
184
(ii) A is one-to-one;
(iii) A is continuous (use α||v||2 ≤ a(v, u) ≤ M ||u||2 );
(iv) A−1 exists and is continuous on R(A), the range of A;
(v) R (A) = H.
Now, f ∈ H ∗ implies, by Riesz representation theorem, that there exists a unique
z ∈ H such that
f (v) = hv, zi∀v ∈ H. (13.7)
Since R (A) = H, there exists u ∈ H such that A(u) = z. Using (13.7) and (13.6),
we obtain the following equation:
f (v) = hv, zi = hv, A(u)i = a(v, u)∀v ∈ H.
This completes the proof of existence. We now prove uniqueness. Assume there
exist u and u∗ in H, u 6= u∗ such that
f (v) = a(v, u) = a(v, u∗ )∀v ∈ H.
Then, a(v, u − u∗ ) = 0 ∀v ∈ H. Take v := u − u∗ , so that 0 = a(u − u∗ , u − u∗ ) ≥

α||u − u∗ ||2 , which yields u = u∗ , a contradiction, completing proof of part (a).
For part (b), assume now that a(., .) is symmetric and a(v, v) ≥ 0 ∀v ∈ H.
Then, using the fact that f (u) = a(u, u) we have that
1 1 1
0 ≤ a(u − v, u − v) = a(u, u) − a(u, v) + a(v, v)
2 2 2
1 i 1 i
= −[ a(u, u) − a(u, u) + [ a(v, v) − a(u, v)
2 2
1 i 1 i
= −[ a(u, u) − f (u) + [ a(v, v) − f (v) ,
2 2
so that J(u) ≤ J(v) ∀v ∈ H. Hence, J has minimum value at u, i.e., u = min J(v).
v∈H
The proof is complete.
¤
EXERCISES 12.1
1. Using the notations of the proof of Theorem 12.7, prove the following asser-
tions:
185
(i) Aw0 is a bounded linear functional on H;
(ii) A is linear;
(iii) A is continuous;
(iv) A is one-to-one;
(v) A−1 exists and is continuous on R (A);
(vi) R (A) = H.
³
Hint ||A(v)||2 = hA(v), A(v)i = a(A(v), v) ≤ M ||A(v)||.||v||∀v ∈ H, also,
α||v||2 ≤ |a(v, v)| = |hv, A(v)i| ≤ ||v||.||A(v)|| ⇒ ||A(v)|| ≥ α||v|| ∀v ∈ H.
For (vi), first prove R (A) is a closed subspace of H. Let {A(vn )} ⊂ R

(A) be such that A(vn ) / z ∈ H. Then, ||Avm − Avn || = ||A(vm − vn )|| ≥
α||vm − vn || which implies that ||vn − vm || ≤ α−1 ||Avm − Avn || ⇒ {vn }
is Cauchy in H ⇒ there exists v ∈ H such that vn / v ⇒ Avn
/ Av. Hence, Av = z so z ∈ R (A) and R (A) is closed. Next, assume R
(A) 6= H. Then there exists 0 6= z ∈ R (A)⊥ , i.e., hz, A(v)i = 0∀v ∈ H ⇒
a(z, v) = hz, A(v)i = 0∀v ∈ H. ´Take v = z, then a(z, z) = 0 ⇒ z = 0.
Contradiction. Hence R (A) = H .
14.3 Sobolev Spaces

We begin with the following definitions.
Definition 13.8 Let Ω be a nonempty open subset of IRn . Then, for any nonneg-
ative integer m and for any real number 1 ≤ p < ∞, we define a Sobolev space of
order m, denoted by W m,p (Ω) as follows:
W m,p (Ω) := {u ∈ Lp (Ω) : Dwα u ∈ Lp (Ω)∀|α| ≤ m}, (13.8)
where Dwα u are the distributional (or weak) derivatives.
In the sequel we shall use the same notation for distributional and usual derivatives.
In the special case p = 2, we denote W m,2 (Ω) by H m (Ω).
Definition 13.9 For u ∈ W m,p (Ω) we define a norm, denoted by ||.||m,p as follows:
³X ´1
α p p
||u||m,p := ||D u||p , for 1 ≤ p < ∞,
|α|≤m
||u||m,∞ := max ||Dα u||∞ , for p = ∞.

|α|≤m
186
By using the properties of norm of Lp (Ω), it follows that W m,p (Ω) is a normed space.
Definition 13.10 For any u ∈ W m,p (Ω) we define the semi-norms, denoted by |.|m,p
as follows:
³X ´ p1
||u||m,p := ||Dα u||pp , for 1 ≤ p < ∞,
|α|=m
||u||m,∞ := max ||Dα u||∞ , for p = ∞.

|α|=m
Remark 13.11 H m (Ω) is an inner product space. For, given u and v ∈ H m (Ω) we
define an inner product as follows:
X Z
hu, vi := Dα u(x).Dα v(x)d(x).
|α|≤m Ω
We also have the following definition.

Definition 13.12 W0m,p (Ω):=Closure of D (Ω) in W m,p (Ω).
In particular, W0m,2 (Ω) := H0m (Ω) := Closure of D (Ω) in H m (Ω).

We now state the following important theorems without proof. Proofs can be found
in any of the references [1] and [4].
Theorem 13.13 W m,p (Ω) is a Banach space with respect to the norm given in
Definition 10.2.
Theorem 13.14 W m,p (Ω) is separable for 1 ≤ p < ∞, and it is uniformly convex
(and hence also reflexive) for 1 < p < ∞.
Corollary 13.15 H m (Ω) and H0m (Ω) are separable Hilbert spaces. (H0m (Ω) is sep-
arable, since a closed subspace of a separable space is separable).
Remark 13.16 H0m (Ω) as the origin of BV P . The distinction between H0m (Ω) and
H m (Ω)X
is the origin of boundary value problems (BVP). In fact,
f := (−1)|α| D2α u with u ∈ H m (Ω) := H0m (Ω) ⊕ H0m (Ω)⊥ , then there exists
|α|≤m
X
u ∈ H0m (Ω), solution of (−1)|α| D2α v = f in D0 (Ω).
|α|≤m
∂v ∂ m−1 v
Formally, v ∈ H0m (Ω) implies v = ∂n
= ... = ∂nm−1
= 0 on ∂Ω.
187
Remark 13.17 The notation H0m (IR) is useless because D(IRn ) is dense in H m (IR).
We remark also that D (Ω) is not dense in H m (Ω) if Ω is bounded.
We now introduce the important notion of m-extension property .
Definition 13.18 Let Ω be a nonempty open subset of IRn . We say that Ω has the
m-extension property if there exists an operator T : W m,p (Ω) / W m,p (IRn ) such
that
(i) T is linear;
(ii) T (u)(x) = u(x) a.e. in Ω ∀u ∈ W m,p (Ω);
(iii) ||T u||m,p,IRn ≤ M ||u||m,p,Ω ∀u ∈ W m,p (Ω) for some M ≥ 0. The map T is called
an m-extension operator.
n
Example 13.19 The half space IR+ := {(x1 , x2 , ..., xn ) = (x0 , xn ) ∈ IRn , xn > 0}
has the m-extension property, the m-extension operator T being defined as follows:
½
0 u(x0 , xn ), if xn > 0
T u(x , xn ) := 0
u(x , −xn ), if xn < 0.
More generally, regular (smooth) domains have the m-extension property. A very
important result on m-extensions property is the following theorem.
Theorem 13.20 If Ω has the m-extension property, then C ∞ (Ω̄)∩W m,p (Ω) is dense
in W m,p (Ω).
We now conclude this section with the following important embedding theorem.
Theorem 13.21 (Sobolev Embedding Theorem) Let Ω ⊆ IRn be an open set having
the m-extension property. Then, the following inclusions are continuous.
(a) If 1 ≤ p < n then W 1,p (Ω) ,→ Lp∗ (Ω), p1∗ = 1

p
− n1 ;
(b) If p = n, then W 1,n (Ω) / Lα (Ω), ∀α ∈ [n, ∞);
(c) If p > n, then W 1,p (Ω) / L∞ (Ω).
In particular, if p = 2 we have the following corollary.
188
Corollary 13.22 Assume that the open set Ω has the m-extension property. Con-
sider the number p∗ defied by p1∗ = 12 − m
n
. Then, three situations are possible.
1
(a) If p∗
> 0, i.e., n > 2m then H m (Ω) ,→ Lp∗ (Ω).
(b) If p1∗ = 0, i.e., n = 2m, then H m (Ω) ,→ Lp (Ω0 ) for any 1 ≤ p < ∞ and for
any bounded open subset Ω0 ⊂ Ω. Moreover, we have the global embedding
H m (Ω) ,→ Lp (Ω) if 2 ≤ p < ∞.
(c) If p1∗ < 0, i.e., n < 2m, then H m (Ω) ,→ C(Ω̄). More precisely, H m (Ω) ,→
C s (Ω̄), where s ≥ 0 is the biggest nonnegative integer such that s < m − n2 .
Remark 13.23 In this our short introduction to Sobolev Spaces and in the exam-
ples that will follow we will not require the use of the embedding theorem (Theorem
12.21) and Corollary 12.22.
14.4 Applications to Partial Differential Equations

In this section we discuss the existence and uniqueness of weak solutions of some clas-
sical partial differential equations: the Dirichlet problem; Neumann problem; and
the Bi-harmonic equation. We will see that solutions of the problems lie in Sobolev
spaces. Moreover, to solve these problems by using results of Sobolev spaces is much
easier than to solve by other methods. One important advantage is that none of the
above problems has a smooth classical solution if the boundary conditions are not
smooth. But even in such cases solutions exist in the weak sense.
In what follows we shall need the following two lemmas.
Lemma 13.24 (Poincare inequality) Let Ω ⊂ IRn be an open set. Assume Ω is

bounded along some axis (say along xn ). Then there exists a constant C > 0 such
that
|u|0,p ≤ C|u|1,p ∀u ∈ W01,p (Ω).
In particular, for p = 2,
∂u
||u||0,Ω ≤ C|| || ∀u ∈ H01 (Ω).
∂xn 0,Ω
189
Corollary 13.25 Let Ω ⊂ IRn be an open set, bounded along some axis. |.|1,p
becomes a norm and is equivalent to the norm ||.||1,p on W 1,p (Ω). Consequently, the
Poincare inequality, in this case can be written as:
||u||0,p ≤ C||u||1,p ∀u ∈ W01,p (Ω).
∂u
||u||0,Ω ≤ C|| || ∀u ∈ H0 (Ω).
∂xn 0,Ω
Lemma 13.26 (Green’s Inequality) Let Ω ⊂ IRn be an open set with smooth bound-
ary ∂Ω, u ∈ C 2 (Ω̄) and v ∈ C 1 (Ω̄). Then,
Z Z Z
∂u
5u 5 vdx = − (4u)vdx + v ds,
Ω Ω ∂ Ω ∂n
P 2
∂u
where 5u := ( ∂x 1
∂u
, ..., ∂x n
), 4u := ni=1 ∂∂xu2 , ∂n
∂u
:= 5u.n, and u is the unit normal
i
vector on ∂Ω.
With these lemmas, we now consider the following examples.
Example 13.27 (Drichlet Problem for the Laplace Operator) Let Ω ⊆ IRn be a
bounded smooth domain. Let f be a real-valued continuous function on it. The
Drichlet problem is to find a function u ∈ C 2 (Ω̄) such that
½
− 4 u = f, in Ω
u = 0, on ∂Ω.
Let us assume that the solution of the above problem exists. Then, multiplying the
above equation with arbitrary ϕ ∈ D(Ω) and integrating by parts we obtain,
Z Z
− 4uϕdx = f ϕdx.
Ω Ω
This implies, by using the above Green’s identity and the fact that ϕ = 0 on ∂Ω,
that
Z Z
5u 5 vdx = f ϕdx.
Ω Ω
This observation gives us a clue on how to define a weak solution of the Drichlet
problem. We say that u ∈ H01 (Ω) is a weak solution of the Drichlet problem if
190
Z Z
5u 5 vdx = f vdx∀v ∈ H01 (Ω). (13.9)
Ω Ω
We now prove the existence and uniqueness of this weak solution for a suitable f .
Let f ∈ L2 (Ω) and H = H01 (Ω). Define a bilinear form a(., .) on H × H and a
continuous F on H respectively, as follows:
Z
a(u, v) := 5u 5 vdx ∀v ∈ H01 (Ω);
Ω
Z
F (v) := f vdx ∀v ∈ H01 (Ω).
Ω
Then, |a(u, v)| ≤ |u|1 |v|1 ≤ ||u||1 ||v||1 , by Cauchy-Schwartz inequality and the defi-
nitions of norm and semi-norm on H 1 (Ω). Furthermore,
Z
|a(v, v)| ≥ a(v, v) := 5v 5 vdx ≥ α||v||21 ,
Ω
for some α > 0, by Poincare inequality. So, by Lax-Migram theorem, equation
(13.9) has a unique solution u ∈ H := H01 (Ω) and is characterized as
u= min Jv,
v∈H01 (Ω)
1
R R
where Jv := 2 Ω 5v 5 vdx − Ω f vdx.
Example 13.28 We now consider a generalized form of the problem considered

in Example 12.27. Let Ω ⊂ IRn be as an Example 12.27 and aij ∈ C 1 (Ω̄), for
1 ≤ i, j ≤ n. Suppose there exists M ≥ 0 such that
X
aij (x)ξi ξj ≥ M ||ξ||2 ∀ξ ∈ IRn , x ∈ Ω.
1≤i,j≤n
Consider the problem

 X
 − ∂ ∂u
(aij (x) ) + a0 (x)u(x) = f (x), in Ω
1≤i,j≤n
∂x i ∂x j (13.10)

u = 0, on ∂Ω.
191
If we assume that the solution of (13.10) exists, then just as an Example 12.27, we
can define a weak solution. We say that u ∈ H01 (Ω) is a weak solution of (13.10) if
Z X Z Z
∂u ∂v
aij (x) + a0 uv = f v, ∀v ∈ H01 (Ω), (13.11)
Ω 1≤i,j≤n ∂xj ∂xi Ω Ω
where f ∈ L2 (Ω) and a0 is some nonnegative integrable function on Ω. Now, define

a bilinear form a(., .) on H × H, H := H01 (Ω) as follows:
Z X Z
∂u ∂v
a(u, v) := aij (x) + a0 uv.
Ω 1≤i,j≤n ∂xi ∂xi Ω
By using the fact that a0 is nonnegative, and Poincare inequality, it can be shown
that a(., .) is an H-elliptic bilinear form. Hence, by Lax-migram theorem, there
exists a unique solution of equation (13.11). Furthermore, if aij = aji , so a(., .) is
symmetric, then we can characterize the weak solution u of (13.11) as follows:
u= min J(v),
v∈H01 (Ω)
Z X Z Z
1 ∂v ∂v 1 2
where Jv := aij (x) + a0 v − f v.
2 Ω 1≤i,j≤n ∂xi ∂xj 2 Ω Ω
Example 13.29 (The Neumann Problem) Let Ω ⊂ IRn be an open set with smooth
boundary ∂Ω, f be a real-valued continuous function on Ω. The Neumann problem
is to find a function u ∈ C 2 (Ω̄) such that
½
− 4 u + u = f, in Ω
∂u (13.12)
∂n
= 0, on ∂Ω.
Let ϕ ∈ D (Ω). Multiplying the above equation by ϕ, integrating and using Green’s
formula, it can be shown that
Z Z Z
5u 5 ϕ + uϕ = f ϕ.
Ω Ω Ω
Let f ∈ L2 (Ω). We say u ∈ H 1 (Ω) is a weak solution of (13.12) if
Z Z Z
5u 5 v + uv = f v, ∀v ∈ H 1 (Ω). (13.13)
Ω Ω Ω
Now, define a symmetric bilinear form a(., .) on H × H, where H := H 1 (Ω) by,
Z Z
a(u, v) := 5u 5 v + uv.
Ω Ω
192
Then, a(u, v) := ||u||21 . It is trivially H 1 (Ω)-elliptic. So, (13.13) has a unique solution
u ∈ H 1 (Ω) characterized by
u = min J(v),
v∈H 1 (Ω)
R R R
where Jv := 12 Ω 5v 5 v + 12 Ω v 2 − Ω f v.
We note that to prove the existence and uniqueness of the solution of equation
(13.13), Poincare’s inequality has not been used. Hence we require the weak solu-
tion to lie in H 1 (Ω) instead of in H01 (Ω).
Example 13.30 (The Bi-harmonic Equation) The Bi-harmonic equation is given

by
½ 2
4 u = f, in Ω
∂u (13.14)
u = ∂n = 0, on ∂Ω.
By repeated use of Green’s identity, it can be shown that if a solution of (13.14)

exists, then it must satisfy the equation:
Z Z
4u 4 ϕ = f ϕ, ∀ϕ ∈ D(Ω).
Ω Ω
So, we say that u ∈ H02 (Ω) is a weak solution of (13.14) if
Z Z
4u 4 v = f v, ∀v ∈ H02 (Ω).
Ω Ω
Now, define a symmetric Bi-linear form a(., .) on H × H, where H := H02 (Ω) as
follows: Z
a(u, v) := 4u 4 v.
Ω
Then, |a(u, v)| ≤ | 4 u|0 | 4 v|0 ≤ ||u||2 ||v||2 , ∀u, v ∈ H02 (Ω). So, a(., .) is continuous.
By using Poincare’s inequality, we can show that a(u, v) = | 4 u|20 ≥ α||u||22 for some
α > 0. So, a(., .) is H02 (Ω)-elliptic. By Lax-Migram theorem, there exists a unique
weak solution of the Bi-harmonic equation, and the solution u is characterized by:
u= min J(v),
v∈H02 (Ω)
1
R 2
R 2
where Jv := 2 Ω | 4 v| − Ω f v, ∀v ∈ H0 (Ω).
193
Miscellaneous exercises
1. Let ρ : IR × IR / IR be defined by
¯ x y ¯¯
¯
ρ(x, y) = ¯ − ¯, x, y ∈ IR.
1 + |x| 1 + |y|
(a) Prove that ρ is a metric on IR;

(b) Is IR complete with this metric? Justify your answer.
2. Let E = C 1 [0, 1] denote the space of real–valued continuous functions defined

on [0, 1] which are such that their first derivatives are also continuous. Define
a function k · k : E / IR by kf k = maxt∈[0,1] |f 0 (t)|. Is k · k a norm on E?
Justify your answer.
3. Let E be a normed linear space and let M be a closed linear subspace of E.

For arbitrary x + M ∈ E/M define k · k : E/M / IR by
kx + M k = inf kx + mk.
m∈M
Let T : E / E/M be the natural map defined, for arbitrary x ∈ E, by
T x = x + M.
Prove: (a) k·k is a norm on E/M ; (b) T is a bounded linear map; (c) kT k ≤ 1.
4. Let X and Y be Banach spaces and T : X / Y be a bounded linear map

with null space N (T ). Prove that T induces a natural map G : X/N (T )
/ Y such that kGk = kT k. (Hint: Use Problem 3).
5. Say whether or not the following statement is true.

“All Hilbert spaces are isometrically isomorphic to each other”.
If the statement is true, prove it. If it is false, give a counter–example.
6. Let E be a normed linear space and J : E / E ∗∗ be the canonical embed-

ding. Prove that the range of J, R(J), is closed in E ∗∗ if and only if E is a
Banach space.
7. Let H be a Hilbert space and T : H / H be a linear map such that

hT x, yi = hx, T yi ∀ x, y ∈ H. Prove that T is bounded. Hint: Use Closed
Graph Theorem.
194
8. Let {0} 6= E be a normed linear space. Prove that E is a Banach space if and
only if B = {x ∈ E : kxk ≤ 1} is a complete metric space.
9. Let E be a Banach space and M, N be subspaces of E. Suppose E = M ⊕ N .

If x = m + n is the unique representation of a vector x ∈ E, define a new
“norm” k · k∗ on E by kxk∗ = kmk + knk. Prove:
(i) k · k∗ is actually a norm on E;
(ii) (E, k · k∗ ) is a Banach space if M and N are closed.
10. Prove the following special case of Brouwer fixed point theorem. Let f : [0, 1]
/ [0, 1] be continuous. Then f has a fixed point.
(Hint: Consider the function h : [0, 1] / [0, 1] defined by h(x) = f (x) − x
and use the Intermediate Value Theorem).
11. Let E be an arbitrary normed linear space and let T : `np / E, 1≤p<
∞, be an arbitrary linear map. Prove that T is continuous.
Hint: Use the fact that all norms on `np are equivalent. Let {ei }ni=1 be the
canonical basis in `np where
ei = (0, 0, . . . , 0, 1, 0, . . . , 0), i = 1, 2, . . . , n.
↑ ith position
P
n
For arbitrary x ∈ `np , x = αi ei for some unique αi ∈ IR. Recall that
i=1
kxk0 = max |αi | is a norm on `np .
1≤i≤n
12. Let E be a normed linear space with dim E = n < ∞ and let {e1 , e2 , . . . , en }
be a basis for E. Then, for each x ∈ E, there exist unique scalars αi ∈ IR such
that n
X
x= αi ei .
i=1
Let T : E / `n be defined by
1
T (x) = (α1 , α2 , . . . , αn ).
Clearly,
(i) T is 1 − 1 and linear;
(ii) T −1 : `n1 / E is continuous;
(ii) follows trivially from the previous exercise where p = 1. Now prove:
(a) T : E / `n is continuous;
1
(b) If Y is an arbitrary normed linear space, prove that every linear map
U :E / Y is continuous.
195
Hint: For (a), if T is not continuous, then for some ε > 0 there exists a se-
quence {xn } in E such that xn / 0 and kT xn k ≥ ε. If we now set un = xn
kT xn k
we obtain Un / 0 whereas kT Un k = 1. The subset of `n consisting of all
1
vectors of norm 1 is compact, so {T (Un )} has a convergent subsequence which
converges to a vector of norm 1. Now apply the continuity of T −1 .
Remark Part (b) shows that any linear map from a finite dimensional
normed linear space into an arbitrary normed liner space is always continuous.
13. Let E denote the space of all complex sequences {xi }∞
i=1 . Define a function
ρ:E×E / IR by
X∞
1 |xi − yi |
ρ(x, y) = i
,
i=1
2 1 + |xi − yi |
where x = {xi }∞ ∞
i=1 , y = {yi }i=1 . Prove:
(a) ρ is well defined i.e., ρ(x, y) ∈ IR;

(b) ρ is a metric on E;
(c) The metric space (E, ρ) cannot be a normed space (i.e., the metric is not
induced by a norm – it suffices to verify that ρ(x, y) = kx − yk is not
valid).
Hint: If ρ(x, y) = kx − yk then
X∞
1 |α| · |xi − yi |
kαx − αyk = ρ(αx, αy) = i
.
i=1
2 1 + |α| · |xi − yi |
Remark. Part (c) shows that while every normed space is a metric space
(by ρ(x − y) = kx − yk) there are metric spaces which are not normed spaces.
14. Let E = IRn . Consider the function k · k∞ : IRn / IR defined, for arbitrary
n
x = (x1 , x2 , . . . , xn ) ∈ IR by kxk∞ = max1≤i≤n |xi |. (a) Prove k · k∞ is a norm
on IRn . With this norm, IRn is usually denoted by `n∞ . Recall that IRn with
µn ¶1/p
P p
norm kxkp = |xi | is denoted by `np . (b) Prove kxk∞ = lim / kxkp .
i=1 p ∞
Projections Let E be an arbitrary linear space. A projection on E is a

linear operator P (not necessarily bounded) such that P 2 = P . If, however,
E is a normed linear space, we require that a projection, in addition to being
linear, be also bounded. Thus, for normed linear spaces E, a projection P is a
bounded linear map defined on E such that P 2 = P .
A map T such that T 2 = T is called an idempotent map.
196
Remark If P is a projection on a normed linear space E then each x ∈ E
can be written as x = x1 + x2 where P x1 = x and P x2 = 0. In particular,
take x1 ≡ P x and x2 ≡ (I − P )x.
Conversely, given two closed linear subspaces M and N of E such that every
x ∈ E can be expressed uniquely as x = n + m, n ∈ N, m ∈ M then the map
P :E / E defined by P x = n is a closed linear map and hence (by the
Closed Graph Theorem) is bounded, with P 2 = P .
It is then clear that there is a 1 − 1 correspondence between projections on E
and direct sum decomposition of E into two closed subspaces M and N . The
subspaces corresponding to a projection P are N := P X and M := (I − P )X,
characterized respectively by P x = x and P x = 0.
Projection on Hilbert Spaces On Banach spaces E we defined a pro-

jection P on E as a bounded linear operator on E with P 2 = P . On Hilbert
spaces, by a projection we shall mean an orthogonal projection (i.e., P will
be such that the null space of P, N (P ), and the range space of P , R(P), are
orthogonal).
Example Let E = IR2 and P : IR2 / IR be defined by P (x, y) = (x, 0).

Here,
N (P ) = {(x, y) : P (x, y) = {(0, 0)}}

= {(0, y) : y ∈ IR}(= the y − axis).
R(P ) = {(x, 0) : x ∈ IR}(= the x − axis).
Clearly N (P )⊥ R(P ) so that P is an orthogonal projection. In general, pro-

jections need not be orthogonal as can be seen from the following example.
Example Let P : µ IR2 / IR2 be defined by P (x, y) = (x + 0y, x tan π/4 +

³ ´ ¶³ ´ µ ¶
x 1 0 x 1 0
0y). Then, y 7→ . Set P = and observe that
tan π/4 0 y 1 0
P 2 = P and adjoint of P = P , i.e., P ∗ = P . N (P ) = {(0, y) : y ∈ IR}, R(P ) =
{(x, x) : x ∈ IR} and so N (P ) is not orthogonal to R(P ).
Now, solve Problems 15 to 18.
15. Let M be a closed subspace of a Hilbert space, H. For any x ∈ H let P x be

the projection of x on M . The operator P is called the projection of H on M .
Prove that P is a positive operator (and, therefore, is hermitian). Prove also
that if P 6= 0 then kP k = 1.
16. If P is an idempotent hermitian operator on a Hilbert space, prove that P is

a projection (i.e., orthogonal projection).
197
17. Prove that if P1 and P2 are hermitian operators such that P1 P2 = 0 then
P2 P1 = 0.
18. Let P1 and P2 be projections on a Hilbert space. Prove:

(a) P1 P2 is a projection if and only if P1 P2 = P2 P1 .
(b) P1 + P2 is a projection if and only if P1 P2 = 0(= P2 P1 ).
(c) If P1 P2 is a projection then PM1 PM2 = PM1 ∩M2 , where P1 is a projection
onto M1 and P2 is a projection onto M2 .
19. Let E be a finite–dimensional normed linear space of dimension n. Prove

that E ∗ is also finite dimensional. Hence, or otherwise, prove that every finite
dimensional normed linear space is reflexive.
20. Prove that a linear subspace of a normed linear space is closed if and only if it
is weakly closed.
21. Let E be a separable Banach space. Prove that every bounded sequence in E ∗
has a weak∗ convergent subsequence.
22. Let E be a reflexive Banach space. If f ∗ ∈ E ∗ is arbitrary, prove that there

eixsts some x0 ∈ E with kx0 k = 1 such that
f ∗ (x0 ) = kf ∗ k.
Theorem(Weierstrass approximation theorem): If f ∈ C[a, b] and ε > 0 is

given, then there exists a polynomial p such that
sup |f (t) − p(t)| < ε.

t∈[a,b]
With the aid of this theorem solve Problem 23.
23. Let C[0, 1] be endowed with the metric ρ∞ given by
ρ∞ (f, g) = sup |f (t) − g(t)| ∀ f, g ∈ C[0, 1].

t∈[0,1]
Prove that with this metric, C[0, 1] is separable.
24. Let E be a uniformly convex Banach space. If {xn } is a sequence in E such

that
(a) kxn k / 1 as n / ∞; (b) kxn + xm k / 2 as n, m / ∞; prove
that {xn } is a Cauchy sequence in E.
25. Let E be a uniformly convex Banach space. If K ⊂ E is a nonempty closed

convex set, prove that K has a unique vector of minimum norm.
Hint: Use problem 24.
198
26. Let E be a uniformly convex Banach space. If K ⊂ E is a nonempty closed
convex set and x0 ∈ E\K, prove there exists a unique y0 ∈ K such that
kx0 − y0 k = inf kx0 − xk.

x∈K
27. Let E be a uniformly convex Banach space with dual space E ∗ . Suppose
f ∗ ∈ E ∗ , f ∗ 6= 0. Prove there exists a unique x0 ∈ E, kx0 k = 1 such that
f ∗ (x0 ) = kf ∗ k.
28. Let E be a uniformly convex Banach space and let α ∈ (0, 1) and d > 0. Prove
that for any ε > 0, if x, y ∈ E are such that ¡ ¢
kxk ≤ d, kyk ≤ d, kx − yk ≥ ε, then there exists a δ = δ dε > 0 such that
h ³ε´ i
kαx + (1 − α)yk ≤ 1 − 2δ min{α, 1 − α} d.
d
(Hint: Without loss of generality, take α ∈ (0, 1/2] and apply Proposition
9.20).
Definition A Banach space E is called strictly convex if for any pair of

vector x, y ∈ E the equation with ||x|| = ||y|| = 1, we have ||λx+(1−λ)y|| < 1
for all λ ∈ (0, 1).
29. Prove that every uniformly convex Banach space is strictly convex.
Definition Let E be a normed linear space over IR and let E ∗ denote the
dual of E. We define the single–valued normalized duality map of E as the
map j : E / E ∗ such that for each x ∈ E, j(x) ∈ E ∗ and satisfies
hx, j(x)i = kxk · kj(x)k, kj(x)k = kxk,
where h, i is called the duality pairing between x ∈ E and j(x) ∈ E ∗ so that

hx, j(y)i := f ∗ (x) where f ∗ := j(y) ∈ E ∗ .
30. Prove that the duality map j : E / E ∗ is a monotone map in the sense that
for each x, y ∈ E, we have
hx − y, j(x) − j(y)i ≥ 0.
31. (a) Find an example of a complete metric space (E, ρ) and a mapping
f :E / E such that
ρ(f (x), f (y)) < ρ(x, y) ∀ x, y ∈ E,
and f has no fixed points.

Hint: Consider (E, ρ) ≡ (IR, ρ) where ρ is the usual metric and
f (x) = `n (1 + ex ).
199
(b) Prove that if the Contraction Mapping Principle applies to f n where n is
a positive integer, then f has a unique fixed point.
(c) Let (E, ρ) be a complete metric space and f, g : E / E be functions.
Suppose f is a contraction and f (g(x)) = g(f (x))
∀ x ∈ E. Prove that g has a fixed point but that such a fixed point need
not be unique.
32. Let E = C[0, 1] with “sup” norm and let
K = {f ∈ C[0, 1] : f (0) = 0, f (1) = 1, 0 ≤ f (x) ≤ 1}.
For each f ∈ K define ϕ : K / C[0, 1] by
(ϕf )(x) = xf (x).
Prove: (a) K is nonempty, closed, convex and bounded; (b) ϕ maps K into
K; (c) ϕ is nonexpansive; (d) ϕ has no fixed points.
33. (a) State the Contraction Mapping Principle.

(b) Let (E, ρ) be a complete metric space and T : E / E a contraction
map with constant k < 1. Define the sequence {xn } inductively by
xn+1 = T xn , n = 1, 2, . . . , x0 ∈ E.
If x∗ is the unique fixed point of T , prove (i) xn / x∗ as n / ∞; (ii)

kn
ρ(xn , x∗ ) ≤ ρ(x1 , x0 ).
1−k
34. Prove that if f : X / Y is locally Lipschitzian then f is continuous.
35. Let B denote the rectangle [x0 − α, x0 + α] × [y0 − β, y0 + β] in IR2 . Suppose

the function
f :B / IR
is continuous and satisfies the Lipschitz condition
|f (x, y1 ) − f (x, y2 )| ≤ L|y1 − y2 | ∀ (x, y1 ), (x, y2 ) ∈ B
and for some constant L ≥ 0. Suppose further that
|f (x, y)| ≤ M in B.
Let ε ∈ [0, α] be such that

½ ¾
1 β
ε < min , .
L M
200
Prove that the differential equation
dy
= f (x, y),
dx
with the initial condition y(x0 ) = y0 has a unique solution in the interval
[x0 − ε, x0 + ε].
36. Let C[0, 1] be endowed with the “sup” metric. Define
T : C[0, 1] / C[0, 1]
by Z t
(T f )(t) = f (s)ds, f ∈ C[0, 1], t ∈ [0, 1].
0
Prove:
(a) T is not a contraction map;
(b) T 2 is a contraction map.
(Note: “sup” metric ρ is given by ρ(f, g) = sup |f (t) − g(t)|).
0≤t≤1
(c) Does T have a fixed point? (Refer to Problem 32(b)).
Definition If X and Y are normed linear spaces and T is a linear map with
domain D(T ) in X and range R(T ) in Y then T is called compact (or com-
pletely continuous) if for each bounded sequence {xn } in D(T ), we have that
{T xn } contains a subsequence converging to some element in Y .
Consequently, to prove that a linear map T from X into Y is compact it suffices
to prove that if B is the unit open ball in X we have T (B) is relatively compact
in Y . Now prove the following basic facts about compact maps (Problems 38
to 42).
37. Prove that every compact linear operator is continuous.
38. Let X and Y be Banach spaces. Prove,
(a) if T : X / Y is linear continuous and dim X < ∞ then T is compact
(i.e., every linear map on a finite dimensional Banach space is compact).
(b) If (i) T : X / Y is compact, and (ii) Range of T, R(T ), is closed then
dim R(T ) < ∞.
(c) If T : X / Y is linear and dim R(T ) < ∞ then T is compact.
39. If T1 , T2 : X / X are linear compact operators on a Banach space X, prove:

(i) λT1 , (ii) T1 + T2 , (iii) T1 T2 are compact operators.
40. Let X and Y be Banach spaces and Tn : X / Y be a sequence of linear
compact operators. Let Tn / T as n / ∞. Prove that T is compact.
201
41. Let X be a Banach space with dim X = ∞ and let T : X / X be a linear
−1 −1
compact map. Suppose T exists. Is T compact? Justify your answer.
202
Appendix: Completion of a metric space
A1.1 Introduction
Let (M, ρ) be a metric space. A metric space (W, ρW ) is said to be a a completion
of (M, ρ) if and only if the following two conditions hold:
C1: there exists a subspace (W ∗ , ρW ) of (W, ρW ), such that
(a) (W ∗ , ρW ) is dense in (W, ρW ), i.e., W ∗ = W ;

(b) (W ∗ , ρW ) is isometric to (M, ρ),
and
C2: (W, ρW ) is complete.
Remark A1 The metric space (M, ρ) need not be a subspace of a completion, but
it is at least isometric to a dense subspace. This situation is illustrated in Fig. A1
(W, ρW )
S
|
(M, ρ) ∼
= (W ∗ , ρW )
isometric
Fig. A1
In Fig.A1, W ∗ is a subspace of W such that W ∗ = W , and (W, ρW ) is complete.

Thus to construct a completion of an incomplete metric space (M, ρ) we must con-
struct first a metric space (W, ρW ) which is complete, and then we must construct
another metric space (W ∗ , ρW ) which is a subspace of (W, ρW ) and which is dense in
(W.ρW ). Finally we must show that (M, ρ) is isometric to (W ∗ , ρW ). Once this has
been done, the metric space (W, ρW ) is then a completion of the incomplete space
(M, ρ).
In what follows we shall show that every incomplete metric space has a completion.
Moreover, we shall show that all completions of a given metric space are isometric.
This implies that completions are essentially unique.
203
Remark A2
We know that if (X, ρX ) and (Y, ρY ) are two metric spaces that are isometric to
each other, then (X, ρX ) is complete if and only if (Y, ρY ) is. Consequently, since
the incomplete metric space (M, ρ) is isometric to a dense subspace (W ∗ , ρW ) of a
complete metric space (W, ρW ), for questions of continuity and convergence, (M, ρ)
may essentially be regarded as a subspace of its completion.
A1.2 Some basic lemmas

Let (M, ρ) be an incomplete metric space. We shall now construct a completion of
(M, ρ). To do this we shall make use of the following lemmas.
Lemma A2 If {xn } and {yn } are Cauchy sequences in (M, ρ) then
lim / ρ(xn , yn ) exists.

n ∞
Proof. Set δn := ρ(xn , yn ). Clearly, {δn }∞

n=1 is a sequence of nonnegative real
numbers. If we show that {δn } is Cauchy, then lim / δn exists and we are done.
n ∞
So let ∈ > 0 be given. Since {xn } and {yn } are Cauchy sequences, there exists an
integer N0 > 0 such that
ρ(xr , xs ) ≤ ∈ /2 for all r, s ≥ N0 .

ρ(yr , ys ) ≤ ∈ /2 for all r, s ≥ N0 .
Hence,
δr = ρ(xr , yr ) ≤ ρ(xr , xs ) + ρ(xs , ys ) + ρ(ys , yr )

≤ δs + ∈ .
Similarly,
δs ≤ δr + ∈,
so that
δs − ∈ ≤ δr ≤ δs + ∈,
i.e.,
|δr − δs | < ∈ for all r, s ≥ N0 ,
and hence, {δn }∞
n=1 is a Cauchy sequence in IR.
¤
Lemma A3 Let S denote the set of all Cauchy sequences in (M, ρ), i.e.,
S = {{xn } : xn ∈ M and {xn } is Cauchy} .
204
Define a relation ∼ in S as follows:
{xn } ∼ {yn } if and only if lim / ρ(xn , yn ) = 0.

n ∞
Then ∼ is an equivalence relation.
Proof. This is left as an easy exercise for the reader.
Remark A4 The equivalence relation ∼ partitions S into equivalence classes. Let

W denote the set of these equivalence classes.
Lemma A5 If {xn } and {x0n } are in the same equivalence class and if xn / x∗
then x0n / x∗ .
Proof. {xn } and {x0n } in the same equivalence class implies
lim / ρ(xn , x0n ) = 0.

n ∞
xn / x∗ ⇒ lim / ρ(xn , x∗ ) = 0.
n ∞
Hence,
ρ(x0n , x∗ ) ≤ ρ(x0n , xn ) + ρ(xn , x∗ ) /0 as n / ∞.
¤
Remark A6 Lemma A4 asserts that if an element in any class converges, then any
other element in that class must converge, and must converge to the same limit.
Lemma A7 Suppose {xn } and {x0m } are each equivalent to {yn }, then {xn } ∼ {x0n }.
Proof.
{xn } ∼ {yn } ⇒ ρ(xn , yn ) /0 as n / ∞.
{x0n } ∼ {yn } ⇒ ρ(x0n , yn ) /0 as n / ∞.
Hence,
ρ(xn , x0n ) ≤ ρ(xn , yn ) + ρ(yn , x0n ) /0 as n / ∞.
¤
Remark A8 A consequence of Lemma A7 is that a sequence cannot belong to two
distinct classes.
A1.3 The metric space (W, ρW )

Let Γ1 and Γ2 be two arbitrary equivalence classes in W . Choose {xn } in Γ1 and {yn }
in Γ2 arbitrary. Then, clearly {xn } and {yn } are Cauchy sequences in (M, ρ). By
205
Lemma A2, limn /∞ ρ(xn , yn ) exists. We now define the function ρW : W × W
/ IR by
ρW (Γ1 , Γ2 ) = lim / ρ(xn , yn ).
n ∞
Lemma A9 The function ρW is well defined, i.e., if {xn } ∼ {x0n } and {yn } ∼ {yn0 }
then
lim / ρ(xn , yn ) = lim / ρ(x0n , yn0 ).
n ∞ n ∞
0
Proof. Let k = lim / ρ(xn , yn ), k = lim / ρ(x0n , yn0 ). By Lemma A2 these
n ∞ n ∞
limits exist. Let ∈ > 0 be given. {xn } ∼ {x0n } ⇒ ∃ and integer N1 > 0 such that
ρ(xn , x0n ) < ∈ /3 for all n ≥ N1 .
{yn } ∼ {yn0 } ⇒ ∃ an integer N2 > 0 such that
ρ(yn , yn0 ) < ∈ /3 for all n ≥ N2 .
ρ(x0n , yn0 ) / k 0 ⇒ ∃ an integer N3 > 0 such that
|ρ(x0n , yn0 ) − k 0 | < ∈ /3 for all n ≥ N3 .
Choose N = max{N1 , N2 , N3 }. Then for all n ≥ N ,
ρ(xn , yn ) ≤ ρ(xn , x0n ) + ρ(x0n , yn0 ) + ρ(yn0 , yn ) + k 0

≤ ∈ +k 0 ,
so that, since ρ(xn , yn ) / k we have, k ≤ ∈ +k 0 . But this inequality holds for

every ∈ > 0. Hence k ≤ k 0 . Similarly, it can be proved that k 0 ≤ k.
¤
Lemma A10 (W, ρW ) is a metric space.
Proof. Let Γ, γ be arbitrary elements of (W, ρW ). Clearly, ρW (Γ, γ) = lim / ρ(xn , yn ) ≥

n ∞
0, for arbitrary {xn } ∈ Γ, {yn } ∈ γ. We now prove
ρW (Γ, γ) = 0 ⇔ Γ = γ.
But,
ρW (Γ, γ) = 0 ⇔ lim / ρ(xn , yn ) = 0

n ∞
⇔ {xn }, {yn } belong to the same
equivalence class
⇔ Γ = γ.
Furthermore,
ρW (Γ, γ) = lim / ρ(xn , yn ) = lim / ρ(yn , xn ) = ρW (γ, Γ).

n ∞ n ∞
206
To prove the triangle inequality, let β ∈ W and let {zn } ∈ β be arbitrary. Then,
ρW (Γ, γ) = lim / ρ(xn , yn )
n ∞
≤ lim / ρ(xn , zn ) + lim / ρ(zn , yn )
n ∞ n ∞
= ρW (Γ, β) + ρW (β, γ).
¤
With reference to Fig. A1, we have so far constructed the metric space (W, ρW ).
We are yet to prove that it is complete. Meanwhile let us construct the metric space
(W ∗ , ρW ) as a subspace of W , show that it is isometric to (M, ρ) and that it is dense
in (W, ρW ). For the construction of (W ∗ , ρW ), suppose x∗ is an arbitrary element of
M . We immediately construct the constant sequence {x∗ , x∗ , x∗ , . . .}. This sequence
is clearly a Cauchy sequence in M (it is, in fact, a convergent sequence with limit
x∗ ). Since W consists of equivalence classes of all Cauchy sequences in M , this
sequence belongs to some equivalence class in W . Denote this equivalence class by
Γx∗ . Consequently, for all points x in M , we can construct such constant Cauchy
sequences in M and each such sequence belongs to some equivalence class Γx in W .
Let W ∗ denote the set of all such equivalence classes constructed in this way, i.e.,
n
W∗ = Γx : x ∈ M and Γx contains the constant
o
sequence {x, x, x, . . .} .
Clearly, W ∗ is a subset of W . Moreover, we have the following results:
Lemma A11. (M, ρ) is isometric to (W ∗ , ρW ).
Proof. For each x ∈ M , there exists Γx in W ∗ . Define ϕ : (M, ρ) / (W ∗ , ρW ) by
ϕ(x) = Γx ,
for each x ∈ M . We show ϕ is an isometry. Let x, y ∈ M . Then ϕ(x) = Γx ,
ϕ(y) = Γy and ¡ ¢
ρW (Γx , Γy ) = lim / ρ x(n) , y (n) = ρ(x, y),
n ∞
i.e., ρW (ϕ(x), ϕ(y)) = ρ(x, y), and so ϕ is an isometry. This completes the proof.
¤
Lemma A12. (W ∗ , ρW ) is dense in (W, ρW ).
Proof. Let Γ be an arbitrary element of (W, ρW ). It suffices to construct a se-

quence in W ∗ which converges to Γ. Let {xn }∞
n=1 be an arbitrary element in Γ. This
∞
means that {xn }n=1 is a Cauchy sequence in (M, ρ), i.e.,
{x1 , x2 , x3 , . . .} ⊆ M
207
and
ρ(xn , xm ) /0 as n, m / ∞.
Now, for x1 there is a corresponding equivalence class Γx1 (consisting of sequences

{x1 , x1 , x1 , . . .}) in W ∗ . Similarly, for x2 , there is a corresponding Γx2 ∈ W ∗ , for
x3 , there is a corresponding Γx3 ∈ W ∗ . Consequently, for each xn there is a corre-
sponding equivalence class Γxn in W ∗ , n = 1, 2, . . .. Consider now the sequence of
equivalence classes {Γxj }∞ ∗
j=1 in W . Clearly,
ρM (Γ, Γxj ) = lim / ρ(xn , xm ) = 0,

m,n ∞
so that Γxj / Γ in (W, ρW ). Hence (W ∗ , ρW ) is dense in (W, ρW ).

¤
Remark A13 Let us once more return to Fig. A1 We have so far succeeded in
constructing the metric space (W, ρW ), a metric space (W ∗ , ρW ) which is a dense
subspace of (W, ρW ) and which is isometric to the incomplete metric space (M, ρ).
However, before we can conclude that (W, ρW ) is a completion of (M, ρ) we must
show that (W, ρW ) is complete. To do this, we need the following lemma.
Lemma A14. Let {xn } be a Cauchy sequence in a metric space (M, ρ) and let
{yn } be any sequence in (M, ρ) such that ρ(xn , yn ) < n1 for each n. Then,
(a) {yn } is also a Cauchy sequence in (M, ρ);
(b) yn / y ∗ in M if and only if xn / y∗.
Proof.
(a) We want to prove {yn } is Cauchy. So let ∈ > 0 be given. Since {xn } is Cauchy,
there exists an integer N1 > 0 such that
ρ(xn , xm ) < ∈ /3
for all m, n ≥ N1 . Choose N ≥ N1 sufficiently large such that for all n, m ≥ N ,
1 1
< ∈ /3 and ≤ ∈ /3.
m n
Then, for all m, n ≥ N ,
ρ(yn , ym ) ≤ ρ(ym , xm ) + ρ(xm , xn ) + ρ(xn , yn )

1 1
≤ + ρ(xm , xn ) + < ∈ .
m n
Hence {yn } is Cauchy.
208
(b) Assume first that yn / y ∗ as n / ∞. From the inequality
ρ(xn , y ∗ ) ≤ ρ(xn , yn ) + ρ(yn , y ∗ ),
it follows that
µ ¶
∗ 1
lim / ρ(xn , y ) ≤ lim / + lim / ρ(yn , y ∗ ) = 0.
n ∞ n ∞ n n ∞
Hence xn / y ∗ as n / ∞. The converse follows by interchanging the

roles of xn and yn . This completes the proof.
¤
We are now able to prove the following lemma.
Lemma A15. (W, ρW ) is complete.
Proof. Let {Γ1 , Γ2 , . . .} = {Γn }∞

n=1 be a Cauchy sequence in (W, ρW ). Since
(W , ρW ) is dense in (W, ρW ), for every integer n there is an element Γ∗n ∈ W ∗ such
∗
that
1
ρW (Γn , Γ∗n ) < .
n
∗ ∞
By Lemma A14, {Γn }n=1 is also Cauchy. However, we determined earlier that {Γ∗n }
has a limit say Γ∗ in (W, ρW ). By part (b) of Lemma A14, Γ∗ is also the limit of Γn .
Hence (W, ρW ) is complete.
¤
Comment Let us now once more return to Fig. A1 We have shown that for any
metric space (M, ρ) we can construct a metric space (W, ρW ) of equivalence classes
of Cauchy sequences that is complete. Furthermore, we have constructed a subspace
(W ∗ , ρW ) of (W, ρW ) which is dense and which is isometric to (M, ρ). Hence (W, ρW )
is a completion of (M, ρ).
For the rest of this section we shall prove that all completions of (M, ρ) are isomet-
ric to (W, ρW ). This will then imply that the completion of (M, ρ) is unique. To
accomplish this, we shall make use of an important lemma in analysis. This lemma
concerns the extension of uniformly continuous maps defined on metric spaces.
Lemma A16. Let (M, ρ) be a metric space, not necessarily complete, and let
f : (M, ρ) / (E, ρE )
be a uniformly continuous mapping from M into a complete metric space, (E, ρE ).

Let (W, ρW ) be the completion of (M, ρ). Then, there exists a unique continuous
extension F of f from (W, ρW ) to (E, ρE ).
Proof. We start by observing that if {xn } is a Cauchy sequence in (M, ρ), uni-
form continuity of f implies {f (xn )} is also a Cauchy sequence in (E, ρE ). Since
209
(E, ρE ) is complete, f (xn ) / e∗ for some e∗ ∈ E. Furthermore {xn } ⊆ (M, ρ)
and (M, ρ) is isometric to (W , ρW ) ⊆ (W, ρW ), where (W ∗ , ρW ) is the subspace of
∗
(W, ρW ) which is dense in (W, ρW ). Since (W, ρW ) is complete, xn / ω ∗ in (W, ρW )

∗ ∗
for some ω ∈ W . If U is the isometry between (M, ρ) and (W , ρW ) we have U xn
/ ω ∗ . Now given ω ∗ ∈ (W, ρW ) there is a Cauchy sequence {xn } in (M, ρ) such
that xn / ω ∗ ∈ W and f (xn ) / e∗ in E. Define F : (W, ρW ) / (E, ρE ) by
F (ω ) = e . The map F is well defined. For, given ω ∈ (W, ρW ) if {xn } and {x0n } are
∗ ∗ ∗
two Cauchy sequences in (M, ρ) such that xn / ω ∗ and x0 / ω ∗ , it is easy

n
to see, by the uniform continuity of f , that e := f (lim xn ) and (e ) := f (lim x0n )
∗ ∗ 0
are identical (verify this). In particular, if xn / x∗ ∈ M then f (x∗ ) = F (x∗ ).

Hence F is an extension of f . Moreover, if xn / ω ∗ then F (xn ) / F (ω ∗ ) so
that F is also continuous. This completes the proof.
¤
We are now able to prove the following uniqueness lemma.
Lemma A17. Let (W 0 , ρ0W ) be a completion of (M, ρ). Then (W, ρW ) is isometric
to (W 0 , ρ0W ).
Proof. (W, ρW ) is a completion of (M, ρ) implies there exists an isometry h1 ,
h1 : (M, ρ) / (h1 (M ), ρW ) and h1 (M ) = W.
If (W 0 , ρ0W ) is another completion of (M, ρ), then there exists an isometry h2 ,
h2 : (M, ρ) / (h2 (M ), ρ0 ) and h2 (M ) = W 0 .

W
Thus, for each x ∈ M , there is an element h1 (x) ∈ W and an element h2 (x) ∈ W 0 .

Define for each x ∈ M , h : (h2 (M ), ρ0W ) / (h1 (M ), ρW ) by h(h2 (x)) = h1 (x).
−1
Since h1 and h2 are isometries we have that h1 and h2 are continuous and invertible.
So h is invertible. Consequently,
ρ0W (h2 (x), h2 (y)) = ρ(x, y) = ρW (h1 (x), h1 (y)).
This implies that if u∗ ∈ h2 (x) and v ∗ ∈ h2 (y), and if h1 (x) = h(u∗ ) and
h1 (y) = h(v ∗ ) then
ρW (h(u∗ ), h(v ∗ )) = ρ0W (u∗ , v ∗ ),
and so (h2 (M ), ρ0W ) is isometric to (h1 (M ), ρW ). Now, using Lemma A16, we can
construct continuous extensions H1 and H2 of h1 and h2 respectively. Set H =
H1 ◦ H2 , the composition of H1 and H2 . Then, it is not difficult to see that
ρW (H(x∗ ), H(y ∗ )) = ρ0W (x∗ , y ∗ ) ∀ x∗ , y ∗ ∈ W 0 ,
so that (W 0 , ρ0W ) is isometric to (W, ρW ). The proof is complete.

¤
210
As a consequence of Lemma A2 through to Lemma A17 we have proved the follow-
ing theorem:
Theorem A18 Every metric space (M, ρ) has a completion and all of its com-
pletions are isometric (i.e., completions are essentially unique).
Example R 1 A19 We saw in Example 1.15 that the metric space C[0, 1] with metric
kf k1 = 0 |f (t)|dt ∀ f ∈ C[0, 1] is not complete. With this metric, the completion
of C[0, 1] is the space (L1 [0, 1], k · k1 ), (see e.g., [1],[2],[4],[10] or [13]).
211
212
Bibliography
[1] R. A. Adams, Sobolev Spaces, New York, Academic Press, 1975.

[2] S. Banach, Théorie des operations lineires, Warsaw, 1932.
[3] D. G. Birkhoff and O. D. Kellog, Invariant points in Function spaces, Trans.
AMS, 23 (1922), 96–115.
[4] H. Brézis, Analyse functionnelle: théorie et applications, Paris, Masson (1983).
[5] M. S. Broadskii and D.P. Mil’man, On the centre of a convex set, Dokl. Akad.
Nauk. SSSR, 59 (1948), 837–840. (Russian).
[6] F. E. Browder, Nonexpansive nonlinear operators in a Banach space, Proc. Nat.
Acad. Sci. USA, 54 (1965), 1041–1044.
[7] R. E. Bruck, Asymptotic behaviour of nonexpansive mappings, in “Fixed Points
and Nonexpansive Mappings”, Contempory Math. 18 (1983), 1–47.
[8] R. Cacciopoli, Un teorem generale sull’esistenza di elementi uniti in una trans-
formazione funzionale, Rend. Accad. Naz. Lincei, 13 (1931), 498–502.
[9] N. Dunford and J. T. Schwartz, Linear Operators, Vol.I, Interscience, New
York, 1958.
[10] D. Göhde, Zum prin zip der kontraktiven abblidung, Math. Nachr. 30 (1965),
251–258.
[11] W. A. Kirk, A fixed point theorem for mappings which do not increase distance,
Amer. Math. Monthly 72 (1965), 1004–1006.
[12] E. Picard, Sur l’application des methodes d’approximations successive a l’etude
de certaines equations differentielles ordinaires, J. Math. 9 (1893), 217–271.
213
Index
a non-separable inner product space, 131 diametral point, 136

normal, 94 Dirac delta function, 149
direct sum, 83
abstract minimization problem, 153 distributional (or weak) derivatives, 148
abstract variational problem, 155 distributional (or weak) derivative, 149
adjoint, 93 double dual, 113
algebraic complement, 83 Drichlet Problem, 162
dual space, 25
Baire’s Category Theorem, 53
Baire’s lemma, 52 Eberlein–Smulyan Theorem, 120, 121
Banach Alaoglu Theorem, 112 equivalent, 27
Banach Steinhaus Theorem, 54
Bessel’s Inequality, 80 first category, 53
fixed point, 133
bidual, 113
Fourier coefficients, 87
bilinear form, 152
Fourier expansion, 87
bounded, 21
Brodskii and Mil’man Theorem, 137 gauge, 41
Brouwer Fixed Point Theorem, 134 Goldstein’s Theorem, 114
Green’s Inequality, 162
canonical embedding, 114
Canonical Map, 113 Hölder’s Inequality, 16
Cauchy Schwartz Inequality, 75 Hahn Banach Theorem, Analytic Form,
Cauchy sequence, 8 33
completeness, 7 Hahn Banach Theorem, First geometric
Completion, 175 form, 41
completion, 12 Hahn Banach Theorem; Second Geomet-
completion of a metric space, 12 ric Form, 44
conjugate space, 25 Heaviside function, 149
contraction, 15 Helly’s Theorem, 116
convex function, 7 Hermitian, 94
convex functional, 31 Hilbert spaces, 73
214
hyperplanes, 39 Sobolev space, 158
space of almost periodic functions, 131
incomplete normed linear spaces, 10 spaces of test functions, 143
inner product, 73 strict contraction, 133
Kakutani’s Theorem, 115 sublinear functional, 31
The Banach Steinhaus Theorem, 51

Lax-Milgram lemma, 155
The Bi-harmonic Equation, 165
linear functional, 19
the canonical map, 114
linear map, 19
The Closed Graph Theorem, 63
linear operator, 19
The Neumann Problem, 164
linear transformation, 19
The Parallelogram Law, 76
Lipschitz map, 133
The Uniform Boundedness Principle, 51
lower bound, 135
The weak star topology, 111
m-extension property, 160 The weak topology, 104
Milman–Pettis Theorem, 125
Uniform Boundedness Principle, 53
multi-index, 143
Uniformly Convex Banach Spaces, 121
nonexpansive, 15, 133 unitary operators, 94, 97
norm, 1 unitary space, 6
normal operators, 96 upper bound, 135
normal structure, 136
Weierstrass approximation theorem, 170
normed linear space, 2
Zorn’s Lemma, 135
orthogonal, 79
Zorn’s lemma, 49
orthogonal complement, 83
orthogonal set, 79
orthonormal, 79
Parseval’s Identity, 87
Poincare inequality, 161
Polarization Identity, 76
Principle of Nested Sequences, 136
projection, 168
Projection Theorem, 81
Pseudo Topologies, 147
Reflexive Spaces, 113

Riesz Representation Theorem, 86
Schauder–Tychonov Theorem, 134

second category, 53
second conjugate space, 113
Self-adjoint, 94
Separable spaces, 127
215

Fabook

Uploaded by

Document Informationclick to expand document information

Document Informationclick to expand document information

Copyright:

Available Formats

Fabook

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fabook

Uploaded by

Copyright:

Available Formats

APPLICABLE FUNCTIONAL ANALYSIS

°c C.E. Chidume (July, 2006)

This book is based on lectures I gave to the Postgraduate Diploma students

A thorough understanding of these topics at the upper–level undergraduate and

1 Normed Linear Spaces 1

3 The Hahn Banach Theorem: Analytic and Geometric Forms. 37

4 Uniform Boundedness Principle 59

7 Operators on Hilbert Spaces 101

8 Weak and Weak∗ Topologies 109

9 Reflexive and Uniformly Convex Spaces 121

11 Application in Fixed Point Theory 159

12 Function Spaces; Weak Derivatives 171

13 An Introduction to Sobolev Spaces and Applications 179

Normed Linear Spaces

1.1 Normed Linear Spaces

which satisfies the following conditions:

N2: kkxk = |k|.kxk for all k ∈ K, x ∈ X;

N3: kx + yk ≤ kxk + kyk for arbitrary x, y ∈ X.

k.k2 : IR2 → [0, ∞) by

kx̄k2 = (x21 + x22 )1/2 .

(A proof of this inequality is a corollary of a more general inequality which is proved

Verification. Conditions N 1 and N 2 are clearly satisfied (convince yourself).

kx̄ + ȳk22 ≤ (A + B)2 = (kx̄k2 + kȳk2 )2 and so

Example 1.3 The space `p , 1 ≤ p < ∞.

For x̄ = (x1 , x2 , x3 , . . .), ȳ = (y1 , y2 , y3 , . . .) arbitrary elements of `p , define x̄ + ȳ =

(a) Let x = (1, 12 , 13 , ...). Is x ∈ l1 ?.

To answer this question, we compute the sum

by the p−series test with p = 4 > 1. Hence xn ∈ l2 .

Example 1.6 The space `∞ .

`∞ := {x̄ = (x1 , x2 , . . .), xi ∈ IR : x̄ is bounded},

Proposition 1.7 If 1 ≤ p ≤ q ≤ +∞ then, `p ⊆ `q .

Proof. We consider two cases.

and this implies {xn } ∈ lq . Hence, lp ⊆ lq .

for arbitrary x̄ = (x1 , x2 , . . .) in `p , is a norm on `p .

Proposition 1.9 The function k.k`∞ : `∞ → [0, ∞) defined by

kx̄k`∞ = sup |xi |,

for arbitrary x̄ = (x1 , x2 , . . .) in `∞ , is a norm on `∞ .

Proof. The obvious verification is left as an easy exercise.

(f + g)(t) = f (t) + g(t);

(λf )(t) = λf (t), for arbitrary scalar λ.

Example 1.11 The space, c.

Example 1.12 The space, c0 .

kxk∞ = max |xi |;

(a) Draw a 2-dimensional picture (n = 2) for each of these functions.

Then apply Hölder’s inequality. Justify the application).

Verify that X, with addition and scalar multiplication as defined in Example

7. Let X = C[a, b] be the space of all continuous real-valued functions on [a, b]

for arbitrary f ∈ X. Verify that X is a normed linear space.

8. Let X be a vector space. A function f : X → IR is said to be a convex function

f [λx + (1 − λ)y] ≤ λf (x) + (1 − λ)f (y).

9. Let ¯X be a normed¯ linear space. Prove that for arbitrary x, y ∈ X,

To check or verify that a metric space (X, ρ) is complete, we take an arbitrary

Theorem 1.14 The normed linear space l2 is complete.

Observe that we have used the completeness of IR to get x∗j ; j = 1, 2, . . . , n, . . . and

Since k is an arbitrary integer, we have, for every m ≥ N,

kfn − fm kC[a,b] := sup |fn (t) − fm (t)| < ε ∀ n, m ≥ N.

sup |fn (t) − fm (t)| < ².