Fixed Point Theory for Lipschitzian-type

Mappings with Applications

Topological Fixed Point Theory and Its
Applications
VOLUME 6
 Fixed Point Theory for
Lipschitzian-type Mappings with
Applications
by

Ravi P. Agarwal
Florida Institute of Technology
Melbourne, FL, USA

Donal O’Regan
National University of Ireland
Galway, Ireland

and

D.R. Sahu
Banaras Hindu University
Varanasi, India

123
Ravi P. Agarwal Donal O’Regan
Department of Mathematical Sciences Institute of Mathematics
Florida Insitute of Technology University College Galway
Melbourne, FL 32901 USA National University of Ireland
[email protected] Galway, Ireland
[email protected]

D.R. Sahu
Department of Mathematics
Faculty of Science
Banaras Hindu University
Varanasi, India
[email protected]

ISBN 978-0-387-75817-6 e-ISBN 978-0-387-75818-3

DOI 10.1007/978-0-387-75818-3
Springer Dordrecht Heidelberg London New York
Library of Congress Control Number: 2009927662

AMS Subject Classifications (2000): 47H09, 47H10


c 2009 Springer Science+Business Media, LLC.
All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media LLC, 233 Spring Street, New York, NY
10013, U.S.A.), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection
with any form of information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to
proprietary rights.
Printed on acid-free paper.

Springer is part of Springer Science+Business Media (www.springer.com)

Dedicated to our daughters

Sheba Agarwal
Lorna Emily O’Regan
Gargi Sahu
Preface
Over the past few decades, fixed point theory of Lipschitzian and non-
Lipschitzian mappings has been developed into an important field of study in
both pure and applied mathematics. The main purpose of this book is to present
many of the basic techniques and results of this theory. Of course, not all aspects
of this theory could be included in this exposition.
The book contains eight chapters. The first chapter is devoted to some
of the basic results of nonlinear functional analysis. The final section in this
chapter deals with the classic results of fixed point theory. Our goal is to
study nonlinear problems in Banach spaces. We remark here that it is hard to
study these without the geometric properties of Banach spaces. As a result in
Chapter 2, we discuss elements of convexity and smoothness of Banach spaces
and properties of duality mappings. This chapter also includes many interest-
ing results related to Banach limits, metric projection mappings, and retraction
mappings. In Chapter 3, we consider normal structure coefficient, weak normal
structure coefficient, and related coefficients. This includes the most recent
work in the literature. Our treatment of the main subject in the book begins in
Chapter 4. In this chapter, we consider the problem of existence of fixed points
of Lipschitzian and non-Lipschitzian mappings in metric spaces. Chapter 5
is devoted to problems of existence of fixed points of nonexpansive, asympto-
tically nonexpansive, pseudocontractive mappings in Banach spaces. Most of
the results are discussed in infinite-dimensional Banach spaces. The theory
of iteration processes for computing fixed points of nonexpansive, asympto-
tically nonexpansive, pseudocontractive mappings is developed in Chapter 6.
In Chapter 7, we prove strong convergence theorems for nonexpansive, pseudo-
contractive, and asymptotically pseudocontractive mappings in Banach spaces.
Finally in Chapter 8, we discuss several applicable problems arising in different
fields.
Each chapter in this book contains a brief introduction to describe the topic
that is covered. Also, an exercise section is included in each chapter. Because the
book is self-contained, the book should be of interest to graduate students and
mathematicians interested in learning fundamental theorems about the theory
of Lipschitzian and non-Lipschitzian mappings and fixed points.
We wish to express our deepest appreciation to Professors Y. Alber,
T. D. Benavides, J. S. Jung, W. A. Kirk, S. Prus, S. Reich, B. E. Rhoades,
B. K. Sharma, W. Takahashi, H. K. Xu, and J. C. Yao for their encouragement
and personal support.
viii Preface

Finally the authors are very grateful to Ms. V. Damle, Ms. B. Marcia and the
staff from the Springer Publishers for their indefatigable cooperation, patience
and understanding.

R.P. Agarwal
Donal O’Regan
D.R. Sahu
Contents

Preface vii

1 Fundamentals 1
1.1 Topological spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Normed spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Dense set and separable space . . . . . . . . . . . . . . . . . . . . 20
1.4 Linear operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5 Space of bounded linear operators . . . . . . . . . . . . . . . . . 25
1.6 Hahn-Banach theorem and applications . . . . . . . . . . . . . . 28
1.7 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.8 Reflexivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.9 Weak topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.10 Continuity of mappings . . . . . . . . . . . . . . . . . . . . . . . 43

2 Convexity, Smoothness, and Duality Mappings 49

2.1 Strict convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2 Uniform convexity . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.3 Modulus of convexity . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.4 Duality mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.5 Convex functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2.6 Smoothness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
2.7 Modulus of smoothness . . . . . . . . . . . . . . . . . . . . . . . 94
2.8 Uniform smoothness . . . . . . . . . . . . . . . . . . . . . . . . . 98
2.9 Banach limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
2.10 Metric projection and retraction mappings . . . . . . . . . . . . . 115

3 Geometric Coefficients of Banach Spaces 127

3.1 Asymptotic centers and asymptotic radius . . . . . . . . . . . . . 127
3.2 The Opial and uniform Opial conditions . . . . . . . . . . . . . . 136
3.3 Normal structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
3.4 Normal structure coefficient . . . . . . . . . . . . . . . . . . . . . 153
3.5 Weak normal structure coefficient . . . . . . . . . . . . . . . . . . 162
3.6 Maluta constant . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
3.7 GGLD property . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
x Contents

4 Existence Theorems in Metric Spaces 175

4.1 Contraction mappings and their generalizations . . . . . . . . . . 175
4.2 Multivalued mappings . . . . . . . . . . . . . . . . . . . . . . . . 188
4.3 Convexity structure and fixed points . . . . . . . . . . . . . . . . 197
4.4 Normal structure coefficient and fixed points . . . . . . . . . . . 201
4.5 Lifschitz’s coefficient and fixed points . . . . . . . . . . . . . . . . 206

5 Existence Theorems in Banach Spaces 211

5.1 Non-self contraction mappings . . . . . . . . . . . . . . . . . . . 211
5.2 Nonexpansive mappings . . . . . . . . . . . . . . . . . . . . . . . 222
5.3 Multivalued nonexpansive mappings . . . . . . . . . . . . . . . . 237
5.4 Asymptotically nonexpansive mappings . . . . . . . . . . . . . . 243
5.5 Uniformly L-Lipschitzian mappings . . . . . . . . . . . . . . . . . 250
5.6 Non-Lipschitzian mappings . . . . . . . . . . . . . . . . . . . . . 259
5.7 Pseudocontractive mappings . . . . . . . . . . . . . . . . . . . . . 264

6 Approximation of Fixed Points 279

6.1 Basic properties and lemmas . . . . . . . . . . . . . . . . . . . . 279
6.2 Convergence of successive iterates . . . . . . . . . . . . . . . . . . 286
6.3 Mann iteration process . . . . . . . . . . . . . . . . . . . . . . . . 288
6.4 Nonexpansive and quasi-nonexpansive mappings . . . . . . . . . 292
6.5 The modified Mann iteration process . . . . . . . . . . . . . . . . 300
6.6 The Ishikawa iteration process . . . . . . . . . . . . . . . . . . . 303
6.7 The S-iteration process . . . . . . . . . . . . . . . . . . . . . . . . 307

7 Strong Convergence Theorems 315

7.1 Convergence of approximants of self-mappings . . . . . . . . . . . 315
7.2 Convergence of approximants of non-self mappings . . . . . . . . 324
7.3 Convergence of Halpern iteration process . . . . . . . . . . . . . 327

8 Applications of Fixed Point Theorems 333

8.1 Attractors of the IFS . . . . . . . . . . . . . . . . . . . . . . . . . 333
8.2 Best approximation theory . . . . . . . . . . . . . . . . . . . . . . 335
8.3 Solutions of operator equations . . . . . . . . . . . . . . . . . . . 336
8.4 Differential and integral equations . . . . . . . . . . . . . . . . . 339
8.5 Variational inequality . . . . . . . . . . . . . . . . . . . . . . . . 341
8.6 Variational inclusion problem . . . . . . . . . . . . . . . . . . . . 343

Appendix A 349
A.1 Basic inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
A.2 Partially ordered set . . . . . . . . . . . . . . . . . . . . . . . . . 350
A.3 Ultrapowers of Banach spaces . . . . . . . . . . . . . . . . . . . . 350

Bibliography 353

Index 365
Chapter 1

Fundamentals

The aim of this chapter is to introduce the basic concepts, notations, and
elementary results that are used throughout the book. Moreover, the results
in this chapter may be found in most standard books on functional analysis.

1.1 Topological spaces

Let X be a nonempty set and d : X × X → R+ := [0, ∞) a function. Then d is
called a metric on X if the following properties hold:
(d1 ) d(x, y) = 0 if and only if x = y for some x, y ∈ X;
(d2 ) d(x, y) = d(y, x) for all x, y ∈ X;
(d3 ) d(x, y) ≤ d(x, z) + d(z, y) for all x, y, z ∈ X.

The value of metric d at (x, y) is called distance between x and y, and the
ordered pair (X, d) is called metric space.
Example 1.1.1 The real line R with d(x, y) = |x − y| is a metric space. The
metric d is called the usual metric for R.
For any r > 0 and an element x in a metric space (X, d), we define
Br (x) := {y ∈ X : d(x, y) < r}, the open ball with center x and radius r;
Br [x] := {y ∈ X : d(x, y) ≤ r}, the closed ball with center x and radius r;
∂Br (x) := {y ∈ X : d(x, y) = r}, the boundary of ball with center x and
radius r.

For a subset C of X and a point x ∈ X, the distance between x and C,

denoted by d(x, C), is defined as the smallest distance from x to elements of C.
More precisely,
d(x, C) = inf d(x, y).
x∈C
The number sup{d(x, y) : x, y ∈ C} is referred to as the diameter of set C and
is denoted by diam(C). If diam(C) is finite, then C is said to be bounded, and

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 1,

c Springer Science+Business Media, LLC, 2009
2 1. Fundamentals

if not, then C is said to be unbounded. In other words, C is bounded if there

exists a sufficiently large ball that contains C.

Interior points and open set – Let G be a subset of a metric space (X, d).
Then x ∈ G is said to be an interior of G if there exists an r > 0 such that
Br (x) ⊂ G. The set G is said to be open if all its points are interior or is the
empty set. The interior of set G is denoted by int(G).

Observation
• int(G) ⊂ G for any subset G of metric space X.
• For any open set G ⊂ X, int(G) = G.
• The empty set ∅ and entire space X are open.

Definition 1.1.2 Let X be a nonempty set and τ a collection of subsets of X.

Then τ is said to be a topology on X if the following conditions are satisfied:
(i) ∅ ∈ τ and X ∈ τ ,
(ii) τ is closed under arbitrary unions,
(iii) τ is closed under finite intersections.
The ordered pair (X, τ ) is called topological space.

Observation
• The members of τ are called τ -open sets or simply open sets.

Definition 1.1.3 A topological space is said to be metrizable if its topology can

be obtained from a metric on the underlying space.

Denoting the class of all open sets of a metric space (X, d) by τd , then we
have
(1) ∅ and X are in τd ,
(2) an arbitrary union of open sets is open,
(3) a finite intersection of open sets is open.

The class τd is called a metric topology on X.

Definition 1.1.4 Let C be a subset of a topological space X. Then the interior

of C is the union of all open subsets of C. It is denoted by int(C).

In other words, if {Gi : i ∈ Λ} are all open subsets of C, then int(C) =

∪i∈Λ {Gi : Gi ⊂ C}.

Observation
• int(C) is open, because it is union of open sets.
• int(C) is the largest open set of C.
• If G is an open subset of C, then G ⊂ int(C) ⊂ C.
1.1. Topological spaces 3

Definition 1.1.5 A set F in a topological space X whose complement F c =

X − F is open is called a closed set.

Theorem 1.1.6 Let C be a collection of all closed sets in a topological space

(X, τ ). Then C has the following properties:
(i) ∅ ∈ C and X ∈ C,
(ii) C is closed under arbitrary intersections,
(iii) C is closed under finite unions.

Definition 1.1.7 Let C be a subset of a topological space X. Then the closure

of C is the intersection of all closed supersets of C. The closure of C is denoted
by C.

In other words, if {Fi : i ∈ Λ} is a collection of all closed supersets of C in

X, then C = ∩i∈Λ Fi .

Observation
• C is closed, because it is the intersection of closed sets.
• C is the smallest closed superset of C.
• If F is a closed subset of X containing C, then C ⊂ C ⊂ F .

Theorem 1.1.8 Let C be a subset of a topological space X. Then C is closed

if and only if C = C.
Exterior points and boundary of sets – Let C be a subset of a topological
space X. Then the exterior of C, written by ext(C), is the interior of the
complement of C, i.e., ext(C) = int(C c ). The boundary of C is a set of points
that do not belong to the interior or the exterior of C. The boundary of set C
is denoted by ∂(C). Obviously, ∂(C) = C ∩ (X \ C) is a closed set.

Proposition 1.1.9 Let A and B be two subsets of a topological space X. Then

the following properties hold:

Properties of interiors Properties of closures

int(int(A)) = int(A) (A) = A
int(A ∩ B) = int(A) ∩ int(B) A∩B ⊂A∩B
int(A ∪ B) ⊃ int(A) ∪ int(B) (A ∪ B) = A ∪ B
A ⊂ B ⇒ int(A) ⊂ int(B) A⊂B⇒A⊂B

Definition 1.1.10 Let τ1 and τ2 be two topologies on a topological space X.

Then τ1 is said to be weaker than τ2 if τ1 ⊂ τ2 .

Note that if τ1 and τ2 are two topologies on X such that τ = τ1 ∩ τ2 . Then

the topology τ is weaker than τ1 and τ2 both.

Theorem 1.1.11 Let {τi : i ∈ Λ} be a collection of topologies on a topological

space X. Then the intersection ∩i∈Λ τi is also a topology on X.
4 1. Fundamentals

We now turn to the notion of a base for the topology τ .

Definition 1.1.12 Let (X, τ ) be a topological space. Then a subclass B of τ is

said to be a base for τ if every member of τ can be expressed as the union of
some members of B.

Observation
• Every topology has a base. In fact, we can take B = τ .
• In a metric space (X, d), collection of all open balls Br (x) (x ∈ X, r > 0)
is a base for the metric topology.

Then, we have the following theorem:

Theorem 1.1.13 Let (X, τ ) be a topological space and B ⊂ τ . Then B is a
base for τ if and only if, for every x ∈ X and every open set G containing x,
there exists B ∈ B such that x ∈ B and B ⊂ G.
We now consider a base of open sets at a point.
Definition 1.1.14 Let (X, τ ) be a topological space and x0 ∈ X. Then the
collection Bx0 ⊂ τ is called a base at a point x0 if, for any open set G containing
x0 , there exists B ∈ Bx0 such that x0 ∈ B ⊂ G.
Observation
• In the metric topology of a metric space (X, d), the collection of all Br (x0 ), where
r runs through the positive real numbers, constitutes a base at a point x0 ∈ X.

Neighborhoods – Let X be a topological space and G an open set. Then

G is called an open neighborhood of a point x0 ∈ X if x0 ∈ G. The set G with-
out x0 , i.e., G \ {x0 }, is called a deleted open neighborhood of a point x0 ∈ X.
A subset C of X is said to be a neighborhood of a point x0 ∈ X if there exists
an open set G ∈ τ such that x0 ∈ G ⊂ C.

Let (X, τ ) be a topological space. Then a collection ν of neighborhoods of

x0 ∈ X is said to be a neighborhood base at a point x0 if every neighborhood of
x0 contains a member of ν.
A collection σ of subsets of a topological space (X, τ ) is said to be a subbase
for τ if σ ⊂ τ and every member of τ is a union of finite intersections of sets
from σ. In other words, σ is a subbase for τ if σ ⊂ τ and for all G ∈ τ and
x ∈ G, there are sets U1 , U2 , · · · , Un in σ such that x ∈ ∩ni=1 Ui ⊂ G.
Let (X, τ ) be a topological space. Then X is said to be
1. a T0 -space if x and y are any two distinct points in X, then there exists
an open set that contains one of them, but not the other;
2. a T1 -space if x and y are two distinct points in X, there exists an open set
U containing x and not y, and there exists another open set V containing
y, but not x;
1.1. Topological spaces 5

3. a T2 -space or Hausdorff topological space if x and y are two distinct points

in X, there exist two open sets G and H such that x ∈ G, y ∈ H, and
G ∩ H = ∅.

Observation
• Every Hausdorff space is a T1 -space.
• A topological space X is T1 -space if and only if every subset consisting of a single
point is closed.
• Every metric space is a Hausdorff space.

A topological space (X, τ ) is said to be compact if every open cover has a

finite subcover, i.e., if whenever X = ∪i∈Λ Gi , where Gi is an open set, then
X = ∪i∈Λ0 Gi for some finite subset Λ0 of Λ.

A subset C of a topological space (X, τ ) is said to be compact if every open

cover has finite open subcover, i.e., if whenever C ⊆ ∪i∈Λ Gi , where Gi is an
open set, then C ⊆ ∪i∈Λ0 Gi for some finite subset Λ0 of Λ.

Observation
• Every finite set of a topological space is compact.
• Every closed subset of a compact space is compact.
• In a compact Hausdorff space, a set is compact if and only if it is closed.

Net – Let D be a nonempty set and a relation on D. Then the ordered

pair (D, ) is said to be directed if
(i) is reflexive: α α for all α ∈ D;
(ii) is transitive: whenever α β and β y⇒α y for all α, β, γ ∈ D;
(iii) for any two elements α and β, there exists γ such that α γ and β γ.

Observation
• (N, ≥) is a directed set.
• If X = ∅, then (P (X), ⊆) and (P (X), ⊇) are directed sets, where P (X) is the
power set of X.
• Every lattice is a directed set.

A net, or a generalized sequence in a set X is a mapping S from a directed

set D into X. The net {xα : α ∈ D} is simply written as {xα }.

Let {xα : α ∈ D} be a net in a set X and let E be another directed set.

Then a net {xαβ : β ∈ E} in X is said to be a subnet of {xα : α ∈ D} if it
satisfies the following conditions:
(i) {xαβ : β ∈ E} ⊂ {xα : α ∈ D};
(ii) for any α0 ∈ D, there exists β0 ∈ E such that α0 αβ exists β0 β.
6 1. Fundamentals

A net {xα : α ∈ D} in a topological space X is said to converge to the point

x in X if for every neighborhood U of x, there exists α0 ∈ D such that xα ∈ U
whenever α  α0 . In this case, we write

xα → x, or lim xα = x.
α

A point x in a topological space X is said to be a cluster point of a net

{xα : α ∈ D} if for every neighborhood U of x and every α ∈ D, there exists
β ∈ D such that β  α and xβ ∈ U .

Theorem 1.1.15 Let {xα }α∈D be a net in a topological space X and let x ∈ X.
Then x is a cluster point of the net {xα }α∈D if and only if the net {xα }α∈D has
a subnet converging to x.

In a metric space (X, d), a sequence {xn } in X is convergent to x ∈ X if

lim d(xn , x) = 0, i.e., if given ε > 0, there exists an integer n0 ∈ N such that
n→∞
d(xn , x) < ε for all n ≥ n0 . A sequence {xn } in a metric space (X, d) is said to
be Cauchy if lim d(xn , xm ) = 0. A metric space (X, d) is said to be complete
m,n→∞
if every Cauchy sequence in X is convergent in X.

Observation
• In a Hausdorff topological space, the limit of a net is unique.
• In a metric space, every convergent sequence is Cauchy.
A subset E of a directed set D is said to be eventual if there exists β ∈ D
such that for all α ∈ D, α β implies that α ∈ E. A net S : D → X is said to
be eventually in a subset C of X if the set S −1 (C) is an eventual subset of D.
A net {xα } in a set X is called a universal net if for each subset C of X, either
{xα } is eventually in C or {xα } eventually in X \ C.
The following facts are important:

(a) Every net in a set has a universal subnet.

(b) If f : X1 → X2 is a mapping and if {xα } is a universal net in X1 , then

{f (xα )} is a universal net in X2 .

(c) If X is compact and if {xα } is a universal net in X, then lim xα exists.

α

We now state the following important result:

Theorem 1.1.16 For a topological space (X, τ ), the following statements are
equivalent:
(a) X is compact.
(b) For any collection of closed sets {Fi }i∈Λ having the finite intersection
property (i.e., the intersection
 of any finite number of sets from the collection
is nonempty), then i∈Λ Fi = ∅.
1.1. Topological spaces 7

(c) Every net in X has a limit point (or, equivalently, every net has a
convergent subnet).
(d) Every filter in X has a limit point (or, equivalently, every net has a
convergent subfilter).
(e) Every ultrafilter in X is convergent.

We now turn our attention to the concept of continuity in topological spaces.

Definition 1.1.17 Let (X, τ ) and (Y, τ  ) be two topological spaces. Then a
function f : X → Y is said to be continuous relative to τ and τ  (more precisely,
τ − τ  continuous) or simply continuous at a point x ∈ X if for each V ∈ τ 
with f (x) ∈ V, there exists U ∈ τ such that x ∈ U and f (U ) ⊂ V .

The function f is called continuous if it is continuous at each point of X.

Using the concept of net, we have the following result for continuity of a function
in a topological space.

Theorem 1.1.18 Let X and Y be two topological spaces and let f be a mapping
from X into Y . Then f is continuous at a point x in X if and only if for every
net {xα } in X,
xα → x ⇒ f (xα ) → f (x).

Some other formulations for continuous functions are the following:

Theorem 1.1.19 Let f be a function from a topological space (X, τ ) into

another topological space (Y, τ  ). Then the following statements are equivalent:
(1) f is continuous (i.e., τ − τ  continuous).
(2) For each V ∈ τ  , f −1 (V ) ∈ τ .
(3) For each closed subset A of Y , f −1 (A) is closed in X.
(4) For all A ⊂ X, f (A) ⊂ f (A).
(5) There exists a subbase σ of τ  such that f −1 (V ) ∈ τ f or all V ∈ σ.

The following result shows that continuous image of a compact set is

compact.

Theorem 1.1.20 Let X and Y be two topological spaces and let T : X → Y be

a continuous mapping. If C ⊆ X is compact, then T (C) is compact.

The following result shows that there exists the smallest topology for which
each member of {fi : i ∈ Λ} is continuous.

Theorem 1.1.21 Let {(Xi , τi ) : i ∈ Λ} be an indexed family of topological

spaces, X any set, and {fi : i ∈ Λ} an indexed collection of functions such that
for each i ∈ Λ, fi is a function from X to Xi . Then there exists the smallest
topology τ on X that makes each fi continuous (i.e., τ − τi continuous).
8 1. Fundamentals

Proof. Let σ = {fi−1 (Vi ) : Vi ⊂ Xi is open in τi (i ∈ Λ)} be a subbase for the

topology τ given by

τ = {∪F ∈F ∩C∈F C : F ⊂ σ} ∪ {∅, X}, (1.1)

where σ is the set of all finite subsets of σ. Thus, G ⊂ X is open in τ if and

only if for every x ∈ G, there are i1 , i2 , · · · , in ∈ Λ and Vi1 ∈ τi1 , Vi2 ∈ τi2 , · · · ,
Vin ∈ τin such that x ∈ ∩nk=1 fi−1
k
(Vik ) ⊂ G.

Remark 1.1.22 The topology τ on X defined by (1.1) making each fi continu-

ous (τ −τ  continuous) is called the weak topology generated by F and is denoted
by σ(X, F).

Product space – Let X1 , X2 , · · · , Xn be n arbitrary sets with the Cartesian

product X = X1 × X2 × · · · × Xn . For each i = 1, 2, · · · , n, define πi : X → Xi
by πi (x1 , x2 , · · · , xn ) = xi . Then πi is called the projection on Xi or the ith
projection. If x ∈ X, then πi (x) is called the ith coordinate of x.

Theorem 1.1.23 Let {(Xi , τi ) : i = 1, 2, · · · , n} bea collection oftopological

spaces and (X, τ ) their topological product, i.e., X = i Xi and τ = i τi . Then
each projection πi is continuous. Moreover, if Y is any topological space, then
a function f : Y → X is continuous if and only if the mapping πi of : Y → Xi
is continuous for all i = 1, 2, · · · , n.

Theorem 1.1.24 (Tychonoff ’s theorem) – The Cartesian product X of an

arbitrary collection {Xi }i∈Λ of compact spaces is compact (with respect to product
topology).

1.2 Normed spaces

A linear space or vector space X over the field K (the real field R or the complex
field C) is a set X together with an internal binary operation “+” called addition
and a scalar multiplication carrying (α, x) in K × X to αx in X satisfying the
following for all x, y, z ∈ X and α, β ∈ K:
1. x + y = y + x,
2. (x + y) + z = x + (y + z),
3. there exists an element 0 ∈ X called the zero vector of X such that
x + 0 = x for all x ∈ X,
4. for every element x ∈ X, there exists an element −x ∈ X called the
additive inverse or the negative of x such that x + (−x) = 0,
5. α(x + y) = αx + αy,
6. (α + β)x = αx + βx,
7. (αβ)x = α(βx),
8. 1 · x = x.
1.2. Normed spaces 9

The elements of a vector space X are called vectors, and the elements of
K are called scalars. In the sequel, unless otherwise stated, X denotes a linear
space over field R.

Observation
• With the usual addition and multiplication, R and C are linear spaces over R.
• X = {x = (a1 , a2 , · · · ) : ai ∈ R} is a linear space.
• The set of solutions of a linear differential equation (and linear partial differential
equation) is a linear space.

A subset S of a linear space X is a linear subspace (or a subspace) of X if S

is itself a linear space, i.e., αx + βy ∈ S for all α, β ∈ K and x, y ∈ S.

If S is a subset of a linear space X, then the linear span of S is the intersection

of all linear subspaces containing S. It is the smallest linear subspace of X
containing S. The linear span of set S is denoted by [S].

Given the points x1 , x2 , · · · , xn of a linear space X, then the element

x = a1 x1 + a2 x2 + · · · + an xn , ai ∈ K
is called linear combination of {x1 , x2 , · · · , xn }.
Proposition 1.2.1 Let S be a nonempty subset of a linear space X. Then the
linear span of S is the set of all linear combinations of elements of S.
A linear space X is said to be finite-dimensional if it is generated by the
linear combination of a finite number of points (which are linearly indepen-
dent). Otherwise, it is infinite-dimensional. The dimension of a linear space X
is denoted by dim(X).

Convex set – Let C be a subset of a linear space X. Then C is said to

be convex if (1 − λ)x + λy ∈ C for all x, y ∈ C and all scalar λ ∈ [0, 1].

By definition of convexity, we have the following fact:

Proposition 1.2.2 Let C be a subset of a linear space X. Then C is convex if
and only if λ1 x1 + λ2 x2 + · · · + λn xn ∈ C for any finite set {x1 , x2 , · · · , xn } ⊆ C
and any scalars λi ≥ 0 with λ1 + λ2 + · · · + λn = 1.
Convex hull – Let C be an arbitrary subset (not necessarily convex) of a linear
space X. Then the convex hull of C in X is the intersection of all convex subsets
of X containing C. It is denoted by co(C). Hence
co(C) = ∩{D ⊆ X : C ⊆ D, D is convex}.
Thus, co(C) is the unique smallest convex set containing C. Clearly,
 
n 
co(C) = α1 x1 + α2 x2 + · · · + αn xn : xi ∈ C, αi ≥ 0 and αi = 1
i=1
= the set of all convex combination of elements of C.
10 1. Fundamentals

The closure of convex hull of C is denoted by co(C). Thus,

 
n 
co(C) = α1 x1 + α2 x2 + · · · + αn xn : xi ∈ C, αi ≥ 0 and αi = 1 .
i=1

The closed convex hull of C in X is the intersection of all closed convex

subsets of X containing C. It is denoted by co(C). Thus,

co(C) = ∩{D ⊆ X : C ⊆ D, D is closed and convex}.

One may easily see that closure of convex hull of C is closed convex hull of C,
i.e., co(C) = co(C).

Observation
• The empty set ∅ is convex.
• For two convex subsets C and D in a linear space X, we have
(i) C + D is convex,
(ii) λC is convex for any scalar λ.
• Any translate C + x0 of a convex set C is convex.
• If {Ci : i ∈ Λ} is any family of convex sets in a linear space X, then ∩i Ci is
convex.
• If C is a convex subset of a linear space X, then
(i) the closure C and the interior int(C) are convex,
(ii) co(C) = C.
• If C is a subset of a linear space, co(C) = co(C).
• In general, co(C) = co(C).

The vector space axioms only describe algebraic properties of the elements
of the space: vector addition, scalar multiplication, and other combinations of
these. For the topological concepts such as openness, closure, convergence, and
completeness, we need a measure of distance in a space.

Definition 1.2.3 Let X be a linear space over field K (R or C) and f : X → R+

a function. Then f is said to be a norm if the following properties hold:
(N1 ) f (x) = 0 if and only if x = 0; (strict positivity)
(N2 ) f (λx) = |λ|f (x) for all x ∈ X and λ ∈ K; (absolute homogeneity)
(N3 ) f (x + y) ≤ f (x) + f (y) for all x, y ∈ X.
(triangle inequality or subadditivity)

The ordered pair (X, f ) is called a normed space.

Observation
• f (x) ≥ 0 for all x ∈ X.
1.2. Normed spaces 11

• |f (x) − f (y)| ≤ f (x − y) and |f (x) − f (y)| ≤ f (x + y) for all x, y ∈ X.

• f is a continuous function, i.e., xn → x ⇒ f (xn ) → f (x).
• f is a convex function, i.e., f (λx+(1−λ)y) ≤ λf (x)+(1−λ)f (y) for all x, y ∈ X
and λ ∈ [0, 1].
• Addition and scalar multiplication are jointly continuous, i.e., if xn → x and
xn → y, then xn + yn → x + y and if xn → x and λn → λ, then λn xn → λx.

We use the notation  ·  for norm. Then every normed space (X,  · ) is a
metric space (X, d) with induced metric d(x, y) = x−y and a topological space
with the induced topology. It means that the induced metric d(x, y) = x − y
in turn, defines a topology on X, the norm topology.

Observation
• In every linear space X, we can easily define a function ρ : X × X → R+ by

0 if x = y,
ρ(x, y) = (1.2)
1 if x = y,

which is a metric on X. It shows that every linear space (not necessarily normed
space) is always a metric space.

At this stage, there arises a natural question:

Under what conditions will any metric on a linear space be a normed space?
Such sufficient conditions are given in following proposition:

Proposition 1.2.4 Let d be a metric on a linear space X. Then function

 ·  : X → R+ defined by

x = d(x, 0) f or all x ∈ X

is a norm if d satisfies the following conditions:

(d1 ) d is homogeneous : d(λx, λy) = |λ|d(x, y) for all x, y ∈ X and λ ∈ K;
(d2 ) d is translation invariant : d(x + z, y + z) = d(x, y) f or all x, y, z ∈ X.

Remark 1.2.5 The metric ρ defined by (1.2) is not homogeneous and the linear
space X is a metric space under metric ρ, but not a normed space.

The following example also demonstrates that a metric space is not neces-
sarily a normed space.

Example 1.2.6 Let X be a space of all complex sequences {xi }∞

i=1 and d(·, ·)
a metric on X defined by

∞
1 |xi − yi |
d(x, y) = ·
i 1 + |x − y |
, x = {xi }, y = {yi } ∈ X. (1.3)
i=1
2 i i
12 1. Fundamentals

Then d is not a norm under the relation d(x, y) = x − y. In fact,

∞ ∞
1 |λ||xi − yi | 1 |xi − yi |
d(λx, λy) = · 
= |λ| · = |λ|d(x, y),
i=1
2 1 + |λ||xi − yi |
i
i=1
2 1 + |xi − yi |
i

i.e., d is not homogeneous.

Remark 1.2.7 The metric d defined by (1.3) is bounded, because

∞ ∞
1 |xi − yi | 1
d(x, y) = · ≤ < ∞.
i=1
2 1 + |xi − yi |
i
i=1
2i

This metric is called a Fréchet metric for X.

We now consider some examples of normed spaces:

Example 1.2.8 Let X = Rn , n > 1 be a linear space. Then Rn is a normed
space with the following norms:

n
x1 = |xi | f or all x = (x1 , x2 , · · · , xn ) ∈ Rn ;
i=1

n 1/p
xp = |xi |p f or all x = (x1 , x2 , · · · , xn ) ∈ Rn and p ∈ (1, ∞);
i=1
x∞ = max |xi | f or all x = (x1 , x2 , · · · , xn ) ∈ Rn .
1≤i≤n

n
Remark 1.2.9 (a) Rn equipped with the norm defined by xp = ( i=1 |xi |p )1/p
is denoted by np for all 1 ≤ p < ∞.
(b) Rn equipped with the norm defined by x∞ = max |xi | is denoted by n∞ .
1≤i≤n

Example 1.2.10 Let X = 1 , the linear space whose elements consist of all
absolutely convergent sequences (x1 , x2 , · · · , xi , · · · ) of scalars (real or complex
numbers), i.e.,
 ∞
 
1 = x : x = (x1 , x2 , · · · , xi , · · · ) and |xi | < ∞ .
i=1

∞
Then 1 is a normed space with the norm defined by x1 = i=1 |xi |.

Example 1.2.11 Let X = p (1 < p < ∞), the linear space whose elements
consist of all p-summable sequences (x1 , x2 , · · · , xi , · · · ) of scalars, i.e.,
 ∞
 
p = x : x = (x1 , x2 , · · · , xi , · · · ) and |xi | < ∞ .
p

i=1

∞
Then p is a normed space with the norm defined by xp = ( i=1 |xi |p )1/p .
1.2. Normed spaces 13

Example 1.2.12 Let X = ∞ , the linear space whose elements consist of all
bounded sequences (x1 , x2 , · · · , xi , · · · ) of scalars, i.e.,

∞ = {x : x = (x1 , x2 , · · · , xi , · · · ) and {xi }∞

i=1 is bounded}.

Then ∞ is a normed space with the norm defined by x∞ = sup |xi |.
i∈N

Example 1.2.13 Let X = c, the sequence space of all convergent sequences of

scalars, i.e.,

c = {x : x = (x1 , x2 , · · · , xi , · · · ) and {xi }∞

i=1 is convergent}.

Then c space is a normed space with the norm  · ∞ .

Example 1.2.14 Let X = c0 , the sequence space of all convergent sequences of

scalars with limit zero, i.e.,

c0 = {x = (x1 , x2 , · · · , xi , · · · ) : {xi }∞
i=1 is convergent to zero}.

The c0 space is a normed space with norm  · ∞ .

Example 1.2.15 Let X = c00 , the sequence space defined by

c00 = {x = {xi }∞
i=1 ∈ ∞ : {xi }∞
i=1 has only a finite number of nonzero terms}.

Then c00 space is a normed space with norm  · ∞ .

Observation
• c00 ⊂ p ⊂ c0 ⊂ c ⊂ ∞ for all 1 ≤ p < ∞.
• If 1 ≤ p < q ≤ ∞, then p ⊂ q . In fact, let x = (1, 1/2, · · · , 1/n, · · · ), and we
have

∞ ∞
1 ∞ ∞
1 π2
|xi | = = ∞, and |xi |2 = 2
= < ∞.
i=1 i=1
i i=1 i=1
i 6

Note that x ∈ 2 , but x ∈ / 1 . Hence an element of 2 is not necessarily an

element of 1 . But each element of 1 is an element of 2 .

Example 1.2.16 Let X = Lp [a, b] (1 ≤ p < ∞), the linear space of all equiva-
lence classes of p-integrable functions on [a, b]. Then Lp [a, b] space is a normed
space with the norm defined by
 b 1/p
f p = |f (t)| dt
p
< ∞.
a

Example 1.2.17 Let X = L∞ [a, b], the linear space of all equivalence classes
of essentially bounded functions on [a, b]. Then L∞ [a, b] space is a normed space
with the norm defined by

f ∞ = ess sup|f (t)| < ∞.

14 1. Fundamentals

Example 1.2.18 Let X = C[a, b], the set of all continuous scalar-valued func-
tions and let “ + ” and “ · ” be operations defined by

(f + g)(t) = f (t) + g(t) f or all f, g ∈ C[a, b];

(λf )(t) = λf (t) f or all f ∈ C[a, b] and scalar λ ∈ K.

Then C[a, b] is a linear space and is also a normed space with the norms:
 b 1/p
f p = |f (t)| dt
p
, 1 ≤ p < ∞; (1.4)
a
f ∞ = sup |f (t)|. (1.5)
t∈[a,b]

Observation
• The norm  · p defined by (1.4) on C[a, b] is called a Lp -norm.
• The norm  · ∞ defined by (1.5) on C[a, b] is called a uniform convergence norm.
 
Equivalent norms – Let X be a linear space over K and let  ·  and  · 
 
be two norms on X. Then  ·  is said to be equivalent to  ·  (written as
 
 ·  ∼  ·  ) if there exist positive numbers a and b such that
  
ax ≤ x ≤ bx for all x ∈ X,

or   
ax ≤ x ≤ bx for all x ∈ X.
Observation
• The relation ∼ is an equivalence relation on the set of all norms on X.
• In a finite-dimensional normed space X, all norms on X are equivalent.
 
• If  ·  and  ·  are equivalent norms on a linear space X, then a sequence {xn }

that is convergent (Cauchy) with respect to  ·  is also convergent (Cauchy)

with respect to  ·  and vice versa.
 
• If  ·  and  ·  are equivalent norms on a linear space X, then the class of

open sets with respect to  ·  is same as the class of open sets with respect to

 ·  and vice versa.

Seminorm – Let X be a linear space over field K (R or C). Then a function

p : X → R+ is said to be a seminorm on X if (N2 ) and (N3 ) (see Definition 1.2.3)
are satisfied. The ordered pair (X, p) is called seminormed space. Note that a
seminorm p is a norm if p(x) = 0 ⇒ x = 0.
Example 1.2.19 Let X = R2 and define p : X → R+ by

p(x) = p((x1 , x2 )) = |x1 |, x ∈ X.

Then p is a seminorm, but not a norm, because p(x1 , x2 ) = 0 implies that only
the first component of x is zero, i.e., x1 = 0.
1.2. Normed spaces 15

We now consider the notion of topological linear spaces.

Definition 1.2.20 A linear space X over K is said to be a topological linear

space if on X, there exists a topology τ such that X × X and K × X with the
product topology have the property that vector addition + : X × X → X and
scalar multiplication · : K × X → X are continuous functions.

In this case, τ is called a linear topology on X.

Definition 1.2.21 A linear topology on a topological linear space X is said to

be a locally convex topology if every neighborhood of 0 (the zero of X) includes
a convex neighborhood of 0. Then X is called a locally convex topological space.

Then we have the following interesting result.

Proposition 1.2.22 If X is a locally convex topological linear space over K,

then a topology of X is determined by a family of seminorms {pi }i∈I .

Inner product – Let X be a linear space over field C. An inner product on X

is a function ·, · : X × X → C with the following three properties:
(I1 ) x, x ≥ 0 for all x ∈ X and x, x = 0 if and only if x = 0;
(I2 ) x, y = y, x, where the bar denotes complex conjugation;
(I3 ) αx + βy, z = αx, z + βy, z for all x, y, z ∈ X and α, β ∈ C.

The ordered pair (X, ·, ·) is called an inner product space. Sometimes,
it is called a pre-Hilbert space. x, y is called inner product of two elements
x, y ∈ X.

Example 1.2.23 Let X = Rn , the set of n-tuples of real numbers. Then the
function ·, · : Rn × Rn → R defined by


n
x, y = xi yi f or all x = (x1 , x2 , · · · , xn ), y = (y1 , y2 , · · · , yn ) ∈ Rn
i=1

is an inner product on Rn . Rn with this inner product is called real Euclidean

n-space.

Example 1.2.24 Let X = Cn , the set of n-tuples of complex numbers. Then

the function ·, · : Cn × Cn → C defined by


n
x, y = xi y i f or all x = (x1 , x2 , · · · , xn ), y = (y1 , y2 , · · · , yn ) ∈ Cn
i=1

is an inner product on Cn . Cn with this inner product is called a complex

Euclidean n-space.
16 1. Fundamentals

Example 1.2.25 Let X = 2 , the set of all sequences of complex numbers

∞
(a1 , a2 , · · · , ai , · · · ) with i=1 |ai |2 < ∞. Then the function ·, · : 2 × 2 → C
defined by
∞

x, y = xi y i for all x = {xi }∞ ∞
i=1 , y = {yi }i=1 ∈ X (1.6)
i=1

is an inner product on 2.

We note that the series (1.6) converges by the Cauchy-Schwarz inequality

(see Proposition 1.2.28).

Example 1.2.26 Let X = C[a, b], the linear space of all scalar-valued
continuous functions on [a, b]. Then the function ·, · : C[a, b] × C[a, b] → C
defined by
b
f, g = f (t)g(t)dt for all f, g ∈ C[a, b] (1.7)
a

is an inner product on C[a, b].

We now give some interesting characterizations of linear spaces having inner

products.
Proposition 1.2.27 Let X be an inner product space. Then the function  ·  :
X → R+ defined by
x = x, x, x ∈ X
is a norm on X.

Proposition 1.2.28 (The Cauchy-Schwarz inequality) – Let X be an

inner product space. Then the following holds:

|x, y|2 ≤ x, xy, y for all x, y ∈ X,

i.e.,
|x, y| ≤ x · y for all x, y ∈ X.

Proposition 1.2.29 (The parallelogram law) – Let X be an inner product

space. Then x + y2 + x − y2 = 2x2 + 2y2 f or all x, y ∈ X.

Proposition 1.2.30 The norm on a normed linear space X is given by an

inner product if and only if the norm satisfies the parallelogram law:

x + y2 + x − y2 = 2x2 + 2y2 f or all x, y ∈ X.

Proposition 1.2.31 (The polarization identity) – Let X be an inner prod-

uct space. Then
 
1
x, y = x + y − x − y + ix + iy − ix − iy
2 2 2 2
f or all x, y ∈ X.
4
1.2. Normed spaces 17

Orthogonality of vectors – Let x and y be two vectors in an inner product

space X. Then x and y are said to be orthogonal if x, y = 0.

Remark 1.2.32 If x and y are orthogonal, then we denote x⊥y and we say “x
is perpendicular to y.”

Proposition 1.2.33 Let X be an inner product space and let x, y ∈ X such

that x⊥y. Then x + y2 = x2 + y2 .

Observation
• 0⊥x for all x ∈ X.
• x⊥x if and only if x = 0.
• Every inner product space is a normed space.
• Every normed space is an inner product space if and only if its norm satisfies the
parallelogram law.

Convergent sequence – A sequence {xn } in a normed space X is said to

be convergent to x if lim xn − x = 0. In this case, we write xn → x or
n→∞
lim xn = x.
n→∞

Observation
• xn → x ⇒ xn  → x (this fact can be easily shown by the continuity of norm).
The converse of this fact is not true in general (see Theorem 2.2.13).
• The limit of convergent sequence is unique. To see it, suppose xn → x and
xn → y. Then x − y ≤ xn − x + xn − y → 0.

Cauchy sequence – A sequence {xn } in a normed space X is said to be

Cauchy if lim xm − xn  = 0, i.e., for ε > 0, there exists an integer n0 ∈ N
m, n→∞
such that xm − xn  < ε for all m, n ≥ no .

Observation

• A sequence in (R, | · |) is convergent if and only if it is Cauchy sequence.

• Every convergent sequence is a Cauchy, but the converse need not be true in
general. In fact, if xn → x, then

xm − xn  ≤ xm − x + x − xn  → 0 as m, n → ∞.

Conversely, suppose X = c00 is the linear space of finitely nonzero sequences

(x1 , x2 , · · · , xi , 0, · · · ) with the norm x = sup |xi |. Let {xn = (1, 1/2, 1/3, · · · ,
i∈N
1/n, · · · )} be a sequence in X. Now

xn − xm  = max{1/n, 1/m} → 0 as m, n → ∞,

i.e., {xn } is a Cauchy sequence. Clearly, the limit x has infinitely nonzero
elements. Thus, x ∈/ X. Therefore, a Cauchy sequence is not convergent in X.
18 1. Fundamentals

• Every Cauchy sequence is bounded.

• Every Cauchy sequence is convergent if and only if it has a convergent sub-

sequence.

Hilbert space and Banach space – A normed space (X,  · ) is said to be

complete if it is complete as a metric space (X, d), i.e., every Cauchy sequence
is convergent in X.
A complete normed space (inner product space) is called a Banach space
(Hilbert space).
Example 1.2.34 n
p (1 ≤ p ≤ ∞) are (finite-dimensional) Banach spaces.

Example 1.2.35 p and Lp [0, 1], 1 ≤ p ≤ ∞ are (infinite-dimensional) Banach

spaces.

Example 1.2.36 The linear space C[a, b] of continuous functions on closed and
bounded interval [a, b] is a Banach space with the uniform convergence norm
f ∞ = sup |f (t)|, but an incomplete normed space with the norm
t∈[a,b]

 b 1/p
f p = |f (t)|p dt , 1 ≤ p < ∞.
a

Example 1.2.37 c00 is not complete.

Theorem 1.2.38 Every finite-dimensional normed space is a Banach space.

The topological property closedness has an important role in the construction

of Banach spaces from its subspaces. A point x in a normed space X is said to
be a limit point of a subset C ⊆ X if there exists a sequence {xn } in C such
that lim xn = x. Also a subset C of a normed space is said to be closed if it
n→∞
contains all of its limit points, i.e., C = C.
Theorem 1.2.39 A closed subspace of a Banach space is a Banach space.

Theorem 1.2.40 Let C be a subset of a normed space X and let x ∈ X. Then

x ∈ C if and only if there exists a sequence {xn } in C such that lim xn = x.
n→∞

Observation
• The subspaces c and c0 are closed subspaces of ∞ (and hence are Banach
spaces). The space c00 is only a subspace in c0 , but not closed in c0 (and hence
not in ∞ ). Therefore, c00 is not a Banach space.

• The subspace C[a, b] is not closed in Lp [a, b] for 1 ≤ p < ∞. Hence C[a, b] is not
a Banach space with the Lp -norm  · p (1 ≤ p < ∞) defined by (1.4).

We now give examples of Banach spaces that are not Hilbert spaces.
1.2. Normed spaces 19

Example 1.2.41 np is a finite-dimensional Banach space that is not a Hilbert

space for p = 2. Indeed, for x = (1, 1, 0, 0, · · · ) and y = (1, −1, 0, 0, · · · ), we have
x + y = (2, 0, 0, 0, · · · ) and x − y = (0, 2, 0, 0, · · · ). Hence
 n 1/p
x = |xi |p
= (1p + 1p )1/p = 21/p ,
i=1

y = (1p + 1p ) = 21/p ,

x + y = (2p )1/p = 2,
x − y = (2p )1/p = 2.
If p = 2, then the parallelogram law:
x + y2 + x − y2 = 2x2 + 2y2
is satisfied, which shows that n2 is a Hilbert space. If p = 2, then the paralle-
logram law is not satisfied. Therefore, np is not a Hilbert space for p = 2.
The following example shows that there exists an infinite-dimensional Banach
space that is not a Hilbert space.
Example 1.2.42 Let X = C[0, 2π], the space of all real-valued continuous func-
tions on [0, 2π] with “sup” norm. Then (C[0, 2π],  · ∞ ) is a Banach space, but
 · ∞ does not satisfy the parallelogram law. In fact, for x(t) = max{sin t, 0},
y(t) = min{sin t, 0}, we have
x∞ = 1, y∞ = 1, x + y∞ = 1, x − y∞ = 1,
i.e., the parallelogram law:
x + y2∞ + x − y2∞ = 2x2∞ + 2y2∞
is not satisfied.
Remark 1.2.43 C[a, b] is an inner product space with the inner product defined
by (1.7), but not a Hilbert space.
Observation
• n
2 , 2 , L2 [a, b] are Hilbert spaces.

p , p , Lp [a, b] (p = 2) are not Hilbert spaces.

• n

We conclude this section with some important facts about the completeness
property.
Definition 1.2.44 A subset C of a normed space X is said to be complete if
every Cauchy sequence in C converges to a point in C.
∞
Definition 1.2.45 Let n=1 xn be an infinite series of elements x1 , x2 , · · · ,
∞
xn , · · · in a normed space X. Then the series n=1 xn is said to converge to
an element x ∈ X if lim sn − x = 0, where sn = x1 + x2 + · · · + xn is nth
n→∞
∞
partial sum of series n=1 xn .
20 1. Fundamentals

∞
Definition 1.2.46 The series n=1 xn in a normed space X is said to be
∞
absolutely convergent if n=1 xn  converges.

The following result shows that completeness and closure are equivalent in
a Banach space.
Theorem 1.2.47 In a Banach space, a subset is complete if and only if it is
closed.

Remark 1.2.48 Notice every normed space is closed, but not necessarily
complete.

Theorem 1.2.49 A normed space X is a Banach space if and only if every

absolutely convergent series of elements in X is convergent in X.

Theorem 1.2.50 (Cantor’s intersection theorem) – A normed space X

is a Banach space if and only if given any descending sequence {Fn } of closed
bounded subsets of X,

lim diam(Fn ) = 0 ⇒ ∩∞
n=1 Fn = ∅. (1.8)
n→∞

Proof. Let X be a Banach space and {Fn } a descending sequence of nonempty

closed bounded subsets of X for which lim diam(Fn ) = 0. For each n, select
n→∞
xn ∈ Fn . Then given ε > 0, there exists an integer n0 ∈ N such that n ≥ n0 ⇒
diam(Fn ) < ε. If m, n ≥ n0 , both xn and xm are in Fn0 , then xn − xm  ≤ ε.
Hence {xn } is a Cauchy sequence. Because X is a Banach space, there exists
x ∈ X such that lim xn = x. This shows that x ∈ F n = Fn if n ≥ n0 . Because
n→∞
the sequence {Fn } is descending, x ∈ ∩∞
n=1 Fn .

Conversely, suppose that the condition (1.8) holds. Suppose {xn } ⊆ X

is a Cauchy sequence. For each n ∈ N, let Fn = {xn , xn+1 , · · · }. Then {F n }
is a descending sequence of nonempty closed subsets of  X for which
∞
lim diam(F n ) = 0. By assumption, there exists a point x ∈ n=1 F n . Let
n→∞
ε > 0 and choose N ∈ N so large that

n ≥ N ⇒ diam(F n ) < ε.

Then clearly for such n we have that xn − x ≤ ε. Hence lim xn = x.

n→∞
Therefore, X is complete.

1.3 Dense set and separable space

A sequence {xn } in a normed space X is said to be a (Schauder) basis of X if
∞
each x ∈ X has a unique expansion x = n=1 αn xn for some scalars α1 , α2 , · · · ,
αn , · · · .
1.3. Dense set and separable space 21

Observation
• {xn } is a basis of a normed space X if for each x ∈ X, there exists a unique
n
sequence {αn } of scalars such that lim x − αi xi  = 0.
n→∞
i=1

• The elements
en = (0, 0, 0, · · · , 1, 0, · · · ), n∈N
↑
nth position
from a basis for c00 , c0 and p (1 ≤ p < ∞).
• {en }n∈N is not a Schauder basis of ∞ .
• The sequence (1, e1 , e2 , · · · ) is a basis for c, where 1 = (1, 1, 1, · · · ).
A subset C of a metric space (X, d) is said to be dense in X if C = X. This
means that C is dense in X if and only if C ∩ Br (x) = ∅ for all x ∈ X and r > 0.
A metric space (X, d) is said to be separable if it contains a countable dense
subset, i.e., there exists a countable set C in X such that C = X.

Observation
• If X is a separable metric space, then C ⊂ X is separable in the induced metric.
• A metric space X is separable if and only if there is a countable family {Gi } of
open sets such that for any open set G ⊂ X,
G = ∪Gi ⊂G Gi .
Next, we give some examples of separable and nonseparable spaces.
Example 1.3.1 The space p, 1 ≤ p < ∞ is separable metric space.
Example 1.3.2 The ∞ space is not a separable space.
Example 1.3.3 The linear space X of all infinite sequences of real numbers
with metric d defined by
∞
1 |xi − yi |
d(x, y) = ·
i 1 + |x − y |
,
i=1
2 i i

x = (x1 , x2 , · · · , xi , · · · ), y = (y1 , y2 , · · · , yi , · · · ) ∈ X
is a separable complete metric space.
Theorem 1.3.4 Every normed space with basis is separable.
Theorem 1.3.5 Every subset of a separable normed space is separable.
Theorem 1.3.6 Every finite-dimensional normed space is separable.
Observation
• R, Rn , c, C[0, 1], p , Lp (1 ≤ p < ∞) are separable normed spaces.
• ∞ , L∞ are not separable.
22 1. Fundamentals

1.4 Linear operators

Let X and Y be two linear spaces over the same field K and T : X → Y an
operator with domain Dom(T ) and range R(T ). Then T is said to be a linear
operator if
(i) T is additive: T (x + y) = T x + T y for all x, y ∈ X;
(ii) T is homogeneous: T (αx) = αT x for all x ∈ X, α ∈ K.
One may easily check that T is linear if and only if

T (αx + βy) = αT x + βT y for all x, y ∈ X and α, β ∈ K.

Otherwise, the operator is called nonlinear. The linear operator is called a

linear functional if Y = R.
Example 1.4.1 Let X = Rn , Y = R, and T : X → R an operator defined by

n
Tx = xi yi f or all x = (x1 , x2 , · · · , xn ),
i=1

where y = (y1 , y2 , · · · , yn ) is the fixed element in Rn . Then T is a linear func-

tional on Rn .

Example 1.4.2 Let X = Y = 2 and T : 2 → 2 an operator defined by

 
x2 x3 xn
T x = 0, x1 , , , · · · , , · · · for all x = (x1 , x2 , x3 , · · · , xn , · · · ) ∈ 2.
2 3 n
Then T is a linear operator on 2.

Example 1.4.3 Let X = C[a, b], the linear space of all continuous real-valued
functions on closed bounded interval [a, b]. Then the operator T : C[a, b] →
C[a, b] defined by
t
T (f )t = f (u)du, t ∈ [a, b]
a
is a linear operator.

Example 1.4.4 Let X = L2 [0, 1], Y = R and T : X → R an operator defined

by
1
Tx = x(t)y(t)dt f or all x ∈ L2 [0, 1],
0

where y is a fixed element in L2 [0, 1]. Then T is a linear functional on L2 [0, 1].

The following result is very useful for linear operators:

Proposition 1.4.5 Let X and Y be two linear spaces over the same field K
and T : X → Y a linear operator. Then we have the following:
1.4. Linear operators 23

(a) T (0) = 0.
(b) R(T ) = {y ∈ Y : y = T x for some x ∈ X}, the range of T is a linear
subspace of Y .
(c) T is one-one if and only if T x = 0 ⇒ x = 0.
(d) If T is one-one operator, then T −1 exists on R(T ) and T −1 : R(T ) → X
is also a linear operator.
(e) If dim(Dom(T )) = n < ∞ and T −1 exists, then dim(R(T )) =
dim(Dom(T )).
Recall an operator T from a normed space X into another normed space Y is
continuous if for any sequence {xn } in X with xn → x ∈ X ⇒ T xn → T x. The
following Theorem 1.4.6 is very interesting because the continuity of any linear
operator can be verified by only verifying T xn → 0 for any sequence {xn } ⊆ X
with xn → 0.
Theorem 1.4.6 Let X and Y be two normed spaces and T : X → Y a linear
operator. If T is continuous at a single point in X, then T is continuous through-
out space X.
Proof. Suppose T is continuous at a point x0 ∈ X. Let {xn } be a sequence in
X such that lim xn = x ∈ X. By the linearity of T , we have
n→∞

T xn − T x = T (xn − x + x0 ) − T x0 .
Because T is continuous at x0 ,
lim (xn − x + x0 ) = x0 ⇒ lim T (xn − x + x0 ) = T x0 ,
n→∞ n→∞

it follows that T xn − T x = T (xn − x + x0 ) − T x0  → 0 as n → ∞. Thus, T

is a continuous operator at an arbitrary point x ∈ X.

Boundedness of linear operator – Let X and Y be two normed spaces

and T : X → Y a linear operator. Then T is said be bounded if there exists a
constant M > 0 such that
T x ≤ M x for all x ∈ X.
A linear functional f : X → R is called bounded if there exists a constant M > 0
such that
|f (x)| ≤ M x for all x ∈ X.

We now present an example of a linear operator that is unbounded.

Example 1.4.7 Let X = c00 , the linear space of finitely nonzero real sequences
with “sup” norm and T : X → R a functional defined by

n
Tx = ixi f or all x = (x1 , x2 , · · · , xn , 0, 0, · · · ) ∈ X.
i=1

Then T is clearly a linear functional, but it is unbounded.

24 1. Fundamentals

With this example, we remark that linearity of the operator does not imply
boundedness. Hence we require additional assumption for boundedness of any
linear operator. The following important result shows that such an additional
assumption is continuity of the linear operator.

Theorem 1.4.8 A linear operator on a normed space is bounded if and only if

it is continuous.

Proof. Let T be a bounded linear operator from a normed space X into another
normed space Y . Then there exists a constant M > 0 such that

T x ≤ M x for all x ∈ X.

Then if xn → 0, we have that

T xn  ≤ M xn  → 0 as n → ∞,

and it follows that T is continuous at zero. By Theorem 1.4.6, we conclude that

T is continuous on X.
Conversely, suppose T is continuous. We show that T is bounded. Suppose,
for contradiction, that T is unbounded. Hence there exists a sequence {xn } in
X such that

T xn  > nxn  for all n ∈ N.

Because T 0 = 0, this implies that xn = 0. Set yn := xn /(nxn ), n ∈ N. Then

yn  = xn /(nxn ) = 1/n → 0, which implies that lim yn = 0. Observe
n→∞
that
 
xn 1
T yn  = T = T xn  > 1 for all n ∈ N
nxn  nxn 

and hence {T yn } does not converge to zero. This means that T is not continuous
at zero, a contradiction.

If the dimension of X is finite, it also forces the boundedness of a linear

operator.

Theorem 1.4.9 Let X and Y be two normed spaces. If X is a finite-dimensional

normed space, then all linear operators T : X → Y are continuous (hence
bounded).

Remark 1.4.10 Example 1.4.7 shows that the conclusion of Theorem 1.4.9
is not true in general (in infinite-dimensional normed spaces). Thus, linear
operators may be discontinuous in infinite-dimensional normed spaces.
1.5. Space of bounded linear operators 25

1.5 Space of bounded linear operators

Let X and Y be two normed spaces. Given two bounded linear operators
T1 , T2 : X → Y, we define

(T1 + T2 )x = T1 x + T2 x,
(αT1 )x = αT1 x for all x ∈ X and α ∈ K.

We denote by B(X, Y ), the family of all bounded linear operators from X

into Y . Then B(X, Y ) is a linear space. The space B(X, Y ) becomes a normed
space by assigning a norm as below:

T B = inf{M : T x ≤ M x, x ∈ X}
 
T x
= sup : x = 0, x ∈ X
x
= sup{T x : x ∈ X, x ≤ 1}
= sup{T x : x ∈ X, x = 1}.

Theorem 1.5.1 The normed space B(X, Y ) is a Banach space if Y is a Banach

space.
We now state an important result:
Theorem 1.5.2 (Uniform boundedness principle) – Let X be a Banach
space, Y a normed space, and {Ti }i∈Λ ⊆ B(X, Y ) a family of bounded linear
operators of X into Y such that {Ti x} is bounded set in Y for each x ∈ X, i.e.,
for each x ∈ X, there exists Mx > 0 such that

Ti x ≤ Mx for all i ∈ N.

Then {Ti B } is a bounded set in R+ , i.e., Ti are uniformly bounded.

As an immediate consequence of Theorem 1.5.2 (uniform boundedness
principle), we have
Theorem 1.5.3 Let X and Y be two Banach spaces and {Tn } a sequence in
B(X, Y ). For each x ∈ X, let {Tn x} converges to T x. Then we have the
following:
(a) T is a bounded linear operator, i.e., T ∈ B(X, Y );
(b) T B ≤ lim inf Tn B .
n→∞

Proof. (a) Because each Tn is linear, it follows that

T (αx + βy) = lim Tn (αx + βy) = lim Tn (αx) + lim Tn (βy)

n→∞ n→∞ n→∞
= α lim Tn x + β lim Tn y
n→∞ n→∞
= αT x + βT y
26 1. Fundamentals

for all x, y ∈ X and α, β ∈ K. Further, because the norm is continuous,

lim Tn x = T x for all x ∈ X,
n→∞

it follows that {Tn x} is a bounded set in Y . By the uniform boundedness

principle, there exists a positive constant M > 0 such that sup Tn B ≤ M.
n∈N
Thus,
Tn x ≤ Tn B x ≤ M x.
Taking the limit as n → ∞, we have
T x ≤ M x,
so T is bounded. Therefore, T ∈ B(X, Y ).

(b) Because
Tn x ≤ Tn B x,
this implies that
lim inf Tn x ≤ lim inf Tn B x.
n→∞ n→∞

Hence T x ≤ lim inf Tn B x. Thus, T B ≤ lim inf Tn B .

n→∞ n→∞

Dual space – The space of all bounded linear functionals on a normed space
X is called the dual of X and is denoted by X ∗ . Clearly, X ∗ = B(X, R) and is
a normed space with norm denoted and defined by
f ∗ = sup{|f (x)| : x ∈ SX }.
In view of Theorem 1.5.1, we have the following interesting result, which is
very useful for the construction of Banach spaces from normed spaces.
Corollary 1.5.4 The dual space (X ∗ ,  · ∗ ) of a normed space X is always a
Banach space.
We now give basic dual spaces:
The dual of Rn – Let Rn be a normed space of vectors x = (x1 , x2 , · · · , xn )
n
with norm x2 = ( i=1 x2i )1/2 . Then for y = (y1 , y2 , · · · , yi , · · · , yn ) ∈ Rn , any
functional f : Rn → R of the form

n
f (x) = xi yi , x = (x1 , x2 , · · · , xi , · · · , xn ) ∈ Rn
i=1

is linear. Further, from the Cauchy-Schwarz inequality,


n  n 1/2  
n 1/2  n 1/2
|f (x)| = xi yi ≤ x2i yi2 = yi2 x2 ,
i=1 i=1 i=1 i=1
1.5. Space of bounded linear operators 27

n
which shows that f is bounded with f ∗ ≤ ( i=1 yi2 )1/2 . However, because
for x = (y1 , y2 , · · · , yn ) equality is achieved in the Cauchy-Schwarz inequality,
n
we must in fact have f ∗ = ( i=1 yi2 )1/2 .
Now, let j be any bounded linear functional on X = Rn . Define the basis
vectors ei in Rn by
ei = (0, 0, · · · , 1, 0, · · · , 0).
↑
ith position
Suppose j(ei ) = ai . Then for any x = (x1 , x2 , · · · , xn ), we have x =
n
i=1xi ei . By the linearity of j, we have

n 
n 
n
j(x) = j(ei xi ) = j(ei )xi = ai xi .
i=1 i=1 i=1

Thus, the dual space X ∗ of X = Rn is itself Rn in the sense that the space X ∗
consists of all functionals of the form f (x) = i=1 ai xi and the norm on X ∗ is
n
n
f ∗ = ( i=1 |ai | )
2 1/2
= a, where a = (a1 , a2 , · · · , an ) ∈ Rn .

The dual of p , 1 ≤ p < ∞ – For 1 ≤ p < ∞, the dual space of p is q

(1/p + 1/q = 1) in the sense that there is a one-one correspondence between
elements y ∈ q and bounded linear functionals fy on p such that
∞

fy (x) = xi yi , x = {xi }∞
i=1 ∈ p,
i=1

where
y = {yi }∞
i=1 ∈ q

and ⎧ ∞
⎪
⎨ ( i=1 |yi |q )1/q , if 1 < p < ∞,
fy ∗ = yq =
⎪
⎩ sup |yi | if p = 1.
i∈N
Observation
• The dual of 1 is ∞ .
• The dual of p is q , 1 < p < ∞ and 1/p + 1/q = 1.
• The dual of ∞ is not 1 .

The dual of c0 – The Banach space c0 of all real sequences x = {xi } such
that lim xi = 0 with norm x∞ = sup |xi | is a subspace of ∞ . The dual
i→∞ i∈N
of c0 is 1 in the usual sense that the bounded linear functionals on c0 can be
represented as
∞

fy (x) = xi yi , x = {xi }∞
i=1 ∈ c0 ,
i=1
∞
where y = {yi }∞
i=1 ∈ 1 and fy ∗ = y1 = i=1 |yi |.
28 1. Fundamentals

The dual of Lp [0, 1], 1 ≤ p < ∞ – For 1 ≤ p < ∞, the dual space of Lp [0, 1]
is Lq [0, 1], (1/p + 1/q = 1) in the sense that there is one-one correspondence
between elements y ∈ Lq [0, 1] and bounded linear functionals fy : Lp [0, 1] → R
such that
1
fy (x) = x(t)y(t)dt and fy ∗ = yq .
0

We now state an important theorem in Hilbert space that is called the Riesz
representation theorem. This theorem demonstrates that any bounded linear
functional on a Hilbert space H can be represented as an inner product with a
unique element in H.
Theorem 1.5.5 (Reisz representation theorem) – Let H be a Hilbert space
and f ∈ H ∗ . Then we have the following:
(1) There exists a unique element y0 ∈ H such that f (x) = x, y0  for each
x ∈ H.
(2) Moreover, f ∗ = y0 .
Remark 1.5.6 In a Hilbert space H, (distinct) bounded linear functionals f
on H are generated by (distinct) elements y of the space H itself, i.e., there is
one-one correspondence between f ∈ H ∗ and y ∈ H. Therefore, H ∗ = H.

1.6 Hahn-Banach theorem and applications

The Hahn-Banach theorem is one of the most important theorems in functional
analysis. To state it, we need the following definitions:

Sublinear functional – Let X be a linear space and p : X → R a functional.

Then p is said to be a sublinear functional on X if
(i) p is subadditive: p(x + y) ≤ p(x) + p(y) for all x, y ∈ X,
(ii) p is positive homogeneous: p(αx) = αp(x) for all x ∈ X and α ≥ 0.
It is evident that every norm is a sublinear functional.
The sublinear functional p on X is called convex functional on X if p(x) ≥ 0
for all x ∈ X. Obviously, every norm is a convex functional also.
Example 1.6.1 Let p : ∞ → R be a functional defined by
p(x) = lim sup xn f or all x = (x1 , x2 , · · · , xn , · · · ) ∈ ∞.
n→∞

Then p is a sublinear functional on ∞.

Extension mapping – Let C be a proper subset of a linear space X and

f a mapping from C into another linear space Y . If there exists a mapping
F : X → Y such that
F (x) = f (x), x ∈ C,
then F is called an extension of f .
1.6. Hahn-Banach theorem and applications 29

Example 1.6.2 Let X = [0, 1], C = [0, 1) and f : C → R defined by

f (x) = x2 , x ∈ [0, 1).

Then 
f (x) if x ∈ C,
F1 (x) =
0 if x = 1
and 
f (x) if x ∈ C,
F2 (x) =
1 if x = 1
are two extensions of f , where F2 is continuous, but F1 is not.
Simply, the Hahn-Banach theorem states that a bounded linear functional
f defined only on a subspace C of a normed space X can be extended to a
bounded linear functional F defined on the entire space and with norm equal
to that of f on C, i.e.,
|f (x)|
F X = f C = sup .
x∈C x

We now state the theorem without proof.

Theorem 1.6.3 (Hahn-Banach theorem) – Let C be a subspace of a real
linear space X, p a sublinear functional on X, and f a linear functional defined
on C satisfying the condition:

f (x) ≤ p(x) f or all x ∈ C.

Then there exists a linear extension F of f such that F (x) ≤ p(x) f or all x ∈ X.

Corollary 1.6.4 Let C be a subspace of a real normed space X and f a bounded

linear functional on C. Then there exists a bounded linear functional F defined
on X that is an extension of f such that F ∗ = f C .

Proof. Take p(x) = f C x, x ∈ X.

The following corollary gives the existence of nontrivial bounded linear func-
tionals on an arbitrary normed space.
Corollary 1.6.5 Let x be an element of a normed space X. Then there exists
(nonzero) j ∈ X ∗ such that j(x) = j∗ x and j∗ = x.

Corollary 1.6.6 Let x be a nonzero element of a normed space X. Then there

exists j ∈ X ∗ such that j(x) = x and j∗ = 1.

Corollary 1.6.7 Let X be a normed space. Then for any x ∈ X,

x = sup |j(x)|.

j∗ ≤1
30 1. Fundamentals

Corollary 1.6.8 If X is a normed space and x0 ∈ X such that j(x0 ) = 0 for

all j ∈ X ∗ , then x0 = 0.
Proof. Suppose x0 = 0. By Corollary 1.6.6, there exists a functional j ∈ X ∗
such that
j(x0 ) = x0  and j∗ = 1.
This implies that j(x0 ) = 0, which is a contradiction. Hence j(x0 ) = 0 for all
j ∈ X ∗ ⇒ x0 = 0.

The following theorems are very useful in many applications.

Theorem 1.6.9 Let C be a subspace of a normed space X and x0 an element
in X such that d(x0 , C) = d > 0. Then there exists a bounded linear functional
j ∈ X ∗ with norm 1 such that j(x0 ) = d and j(x) = 0 f or all x ∈ C.
Theorem 1.6.10 (Separability) – If X ∗ is the dual space of a normed space
X and X ∗ is separable, then X is also separable.
Next, we discuss geometric forms of the Hahn-Banach theorem. We need
the following:
Hyperplane – A subset H of a linear space X is said to be a hyperplane if
there exists a linear functional f = 0 on X such that
H = {x ∈ X : f (x) = α}, α ∈ R.
f (x) = α is called the equation of the hyperplane.
Example 1.6.11 Let X = R, f (x) = 3x, α = 2. Then the set
H = {x ∈ X : f (x) = α} = {x ∈ X : 3x = 2} = {2/3}.
Hence H is a hyperplane.
We have the following interesting result.
Proposition 1.6.12 Let X be a topological linear space. Then the hyperplane
{x ∈ X : f (x) = α} is closed if and only if f is continuous.
Let f (x) = α, α ∈ R, be the equation of hyperplane in a linear space X.
Then we have the following:
(i) {x ∈ X : f (x) < α} and {x ∈ X : f (x) > α} are open half-spaces.
(ii) {x ∈ X : f (x) ≤ α} and {x ∈ X : f (x) ≥ α} are closed half-spaces.

It is easy to see that the boundary of each of the four half-spaces is just a
hyperplane.
Remark 1.6.13 In a topological linear space X, we have
(i) open half-spaces are open sets,
(ii) the closed half-spaces are closed sets if and only if f is continuous, i.e.,
the hyperplane {x ∈ X : f (x) = α} is closed.
1.6. Hahn-Banach theorem and applications 31

Let X be a linear space. We say that the hyperplane {x ∈ X : f (x) = α}

separates two sets A ⊂ X and B ⊂ X if f (x) ≤ α for all x ∈ A and f (x) ≥ α
for all x ∈ B. We say that the hyperplane {x ∈ X : f (x) = α} strictly separates
two sets A ⊂ X and B ⊂ X if f (x) < α for all x ∈ A and f (x) > α for all
x ∈ B.

Theorem 1.6.14 (Hahn-Banach separation theorem) – Let X be a normed

space and let A ⊂ X, B ⊂ X be two nonempty disjoint convex sets. Suppose
that A is open. Then there exists a closed hyperplane that separates A and B,
i.e., there exist j ∈ X ∗ and a number α ∈ R such that

j(x) > α if x ∈ A and j(x) ≤ α if x ∈ B.

Proposition 1.6.15 Let C be a nonempty open convex subset of a normed

/ C, there exists f ∈ X ∗ such that
space X. Then for x0 ∈ X, x0 ∈

f (x) < α f or all x ∈ C,

where f (x0 ) = α.

An immediate consequence of the separation theorem shows that co(C) is

the intersection of all closed half-spaces containing C. Indeed,
Theorem 1.6.16 Let C be a nonempty subset of a normed space X. Then

co(C) = {x ∈ X : f (x) ≤ sup f (y) f or all f ∈ X ∗ }.

y∈C

Theorem 1.6.17 Let C be a nonempty closed convex subset of a normed space

X. If x is not an element in C, there exists a continuous linear functional
j ∈ X ∗ such that
j(x) < inf{j(y) : y ∈ C}.

Theorem 1.6.18 (Hahn-Banach strictly separation theorem) – Let A

and B be two nonempty disjoint convex subsets of a normed space X. Suppose
A is closed and B is compact. Then there exists a closed hyperplane that strictly
separates A and B.

Supporting hyperplane – Let C be a convex subset of a normed space X

with int(C) = ∅ and x0 ∈ ∂C. Then a nonzero functional f ∈ X ∗ is said to be
a support functional for C at x0 if f (x) ≤ f (x0 ) for all x ∈ C. The correspond-
ing hyperplane {x ∈ X : f (x) = f (x0 )} is called a supporting hyperplane for
C at x0 .

A point of C through which a supporting hyperplane passes is called a point

of support of C.

Observation
• Any supporting hyperplane of a set C with nonempty interior is closed.
• An interior point of C cannot be a point of support.
32 1. Fundamentals

We give some conditions on C under which a boundary point is a point of

support.

Theorem 1.6.19 Let C be a convex subset of a normed space X with

int(C) = ∅. Then every boundary point of C is a point of support, i.e., for
every x0 ∈ ∂C, there exists an f ∈ X ∗ such that f = 0 and f (x0 ) = sup f (x).
x∈C

1.7 Compactness
Let (X, d) be a metric space. Recall that a subset C of X is called compact if
every open cover of C has a finite subcover. Equivalently, a subset C of X is
compact if every sequence in C contains a convergent subsequence with a limit
in C.

A subset C of X is said to be totally bounded if for each ε > 0, there exists a

finite number of elements x1 , x2 , · · · , xn in X such that C ⊆ ∪ni=1 Bε (xi ). The
set {x1 , x2 , · · · , xn } is called a finite ε-net.

Observation
• Every subset of a totally bounded set is totally bounded.
• Every totally bounded set is bounded, but a bounded set need not be totally
bounded.

Proposition 1.7.1 A subset of a compact metric space is compact if and only

if it closed.

Proposition 1.7.2 Let X be a metric space. Then the following are equivalent:
(a) X is compact.
(b) Every sequence in X has a convergent subsequence.
(c) X is complete and totally bounded.

Proposition 1.7.3 Let C be a subset of a complete metric space X. Then we

have the following:

(a) C is compact if and only if C is closed and totally bounded.

(b) C is compact if and only if C is totally bounded.

Observation
• X = (0, 1) with usual metric is totally bounded, but not compact.
• X = R with usual metric is complete. But it is not totally bounded and hence
not compact.

A subset C of a topological space is said to be relatively compact if its closure

is compact, i.e., C is compact. In particular, we have an interesting result:
1.7. Compactness 33

Proposition 1.7.4 Let C be a closed subset of a complete metric space. Then

C is compact if and only if it is relatively compact.

We now state the following fundamental theorems concerning compactness.

Theorem 1.7.5 (The Heine-Borel theorem) – A subset C of R is compact

if and only if it is closed and bounded.

Corollary 1.7.6 A set C ⊂ Rn is compact if and only if it is closed and

bounded.

Theorem 1.7.7 (Weierstrass theorem) – Let C be a nonempty compact

subset of a metric space (X, d) and f : C → R a continuous function. Then f
attains its maximum and minimum, i.e., there exist x, x ∈ C such that

f (x) = inf f (x) and f (x) = sup f (x).

x∈C x∈C

Theorem 1.7.8 (Mazur’s theorem) – The closed convex hull co(C) of a

compact set C of a Banach space is compact.

Observation
• Rn , n ≥ 1 is not compact. However, every closed bonded subset of Rn is compact.
For example, C = [0, 1] ⊂ R is compact, but R itself is not compact.
• C[0, 1] and 2 are not compact.
• The subset C = {{xn } ∈ 2 : |xn | ≤ 1/n, n ∈ N} of 2 is compact.
• The closed unit ball BX = {x ∈ X : x ≤ 1} in infinite-dimensional normed
space is not compact in the topology induced by norm (see Proposition 1.7.14).

Proposition 1.7.9 A subset C of p space is compact if C is bounded and for

∞
ε > 0, there exists n0 (ε) ∈ N such that i=n+1 |xi |p < εp f or all n ≥ n0 and
x = {xi }∞
i=1 ∈ C.

Proposition 1.7.10 Every compact subset of a normed space X is closed, but

the converse may not be true.

Observation
• Rn is closed.

Proposition 1.7.11 Every compact subset of a normed space X is complete,

but the converse may not be true.

Proposition 1.7.12 Every compact subset of a normed space is bounded, but

the converse may not be true.

Proposition 1.7.13 Every compact subset of a normed space is separable.

Proposition 1.7.14 A closed and bounded subset of a normed space need not
be compact.
34 1. Fundamentals

∞
Proof. Let X = 2 . Then the unit ball BX = {x ∈ 2 : x2 = ( i=1 |xi |2 )1/2 ≤
1} is closed and bounded. We now show that BX is not compact. Let {xn } be
a sequence in BX defined by

xn = (0, 0, · · · , 1, 0, · · · ), n ∈ N.

↑
nth position
Hence for m = n, √
xn − xm 2 = 2,
i.e., there is no convergent subsequence of {xn }. Therefore, BX is not totally
bounded and hence it is not compact.
Remark 1.7.15 B2 is compact in the weak topology (see Theorem 1.9.26).

Proposition 1.7.16 A normed space X is finite-dimensional if and only if

every closed and bounded subset of X is compact.

1.8 Reflexivity
Let X1 , X2 , · · · , Xm be m linear spaces over the same field K. Then a func-
tional f : X1 × X2 × · · · × Xm → R is said to be an m-linear (multilinear)
functional on X = X1 × X2 × · · · × Xm if it is linear with respect to each of the
variables separately. For m = 2, such a functional is called a bilinear functional .

Duality pairing - Given a normed space X and its dual X ∗ , we define the
duality pairing as the functional ·, · : X × X ∗ → K such that

x, j = j(x) for all x ∈ X and j ∈ X ∗ .

The properties of duality pairing can be easily derived from the definition:
Proposition 1.8.1 Let X ∗ be the dual of a normed space X. Then we have
the following:
(a) The duality pairing is a bilinear functional on X × X ∗ :
(i) ax + by, j = ax, j + by, j for all x, y ∈ X; j ∈ X ∗ and a, b ∈ K;
(ii) x, αj1 +βj2  = αx, j1 +βy, j2  for all x ∈ X; j1 , j2 ∈ X ∗ ; α, β ∈ K.
(b) x, j = 0 for all x ∈ X implies j = 0.
(c) x, j = 0 for all j ∈ X ∗ implies x = 0.
Natural embedding mapping - Let (X,  · ) be a normed space. Then
(X ∗ ,  · ∗ ) is a Banach space. Let j ∈ X ∗ . Hence for given x ∈ X, the equation

fx (j) = x, j
1.8. Reflexivity 35

defines a functional fx on the dual space X ∗ . The functional fx is linear by

Proposition 1.8.1. Moreover, for j ∈ X ∗ we have

|fx (j)| = |x, j| ≤ xj∗ . (1.9)

This shows that fx is bounded and hence fx is a bounded linear functional

on X ∗ .
The space of all bounded linear functionals on X ∗ is denoted by X ∗∗ and is
called the second dual of X. Then fx ∈ X ∗∗ . Note that X ∗∗ is a Banach space.
Let  · ∗∗ denote a norm on X ∗∗ . From (1.9), we have

fx ∗∗ ≤ x.

By Corollary 1.6.5, there exists a nonzero functional j ∈ X ∗ such that

x, j = xj∗ and j∗ = x.

This implies that fx ∗∗ = x.

Define a mapping ϕ : X → X ∗∗ by ϕ(x) = fx , x ∈ X. Then ϕ is called the
natural embedding mapping from X into X ∗∗ . It has the following properties:
(i) ϕ is linear: ϕ(αx + βy) = αϕ(x) + βϕ(y) for all x, y ∈ X, α, β ∈ K;
(ii) ϕ(x) is isometry: ϕ(x) = x for all x ∈ X.

Generally, however, the natural embedding mapping ϕ from X into X ∗∗ is

not onto. It means that there may be elements in X ∗∗ that cannot be represented
by elements in X.
In the case when ϕ is onto, we have an important class of normed spaces.
Definition 1.8.2 A normed space X is said to be reflexive if the natural
embedding mapping ϕ : X → X ∗∗ is onto. In this case, we write X ∼
= X ∗∗
∗∗
or X = X .
Observation
• Rn is reflexive. (In fact, every finite-dimensional Banach space is reflexive.)
• p and Lp for 1 < p < ∞ are reflexive Banach spaces.
• Every Hilbert space is a reflexive Banach space, i.e., H ∗∗ = H.
• 1 , ∞ , L1 and L∞ are not reflexive.
• c and c0 are not reflexive Banach spaces.

We now state the following facts for the class of reflexive Banach spaces.
Proposition 1.8.3 (a) Any reflexive normed space must be complete and, hence,
is a Banach space.
(b) A closed subspace of a reflexive Banach space is reflexive.
(c) The Cartesian product of two reflexive spaces is reflexive.
(d) The dual of a reflexive Banach space is reflexive.
36 1. Fundamentals

Theorem 1.8.4 (James theorem) – A Banach space X is reflexive if and

only if for each j ∈ SX ∗ , there exists x ∈ SX such that j(x) = 1.

1.9 Weak topologies

Let X ∗ be the dual space of a Banach space X. The convergence of a sequence
in a Banach space X is the usual norm convergence or strong convergence, i.e.,
{xn } in X converges to x if lim xn − x = 0. This is related to the strong
n→∞
topology on X with neighborhood base Br (0) = {x ∈ X : x < r}, r > 0 at
the origin. There is also a weak topology on X generated by the bounded linear
functionals on X. Indeed, G ⊂ X is open in the weak topology (we say G is
w-open) if and only if for every x ∈ G, there are bounded linear functionals
f1 , f2 , · · · , fn and positive real numbers ε1 , ε2 , · · · , εn such that

{y ∈ X : |fi (x) − fi (y)| < εi , i = 1, 2, · · · , n} ⊂ G.

Hence a subbase σ for the weak topology on X generated by a base of neigh-

borhoods of x0 ∈ X is given by the following sets:

V (f1 , f2 · · · , fn : ε) = {x ∈ X : |x − x0 , fi | < ε, for every i = 1, 2, · · · , n}.

In particular, a sequence {xn } in X converges to x ∈ X for weak topology

σ(X, X ∗ ) if and only if xn , f  → x, f  for all f ∈ X ∗ .

Observation
• The weak topology is not metrizable if X is infinite-dimensional.
• Under the weak topology, the normed space X is a locally convex topological
space.
• The weak topology of a normed space is a Hausdorff topology.

We are now in a position to define convergence, closedness, completeness,

and compactness with respect to the weak topology.

Weakly convergent – A sequence {xn } in a normed space X is said to con-

verge weakly to x ∈ X if f (xn ) → f (x) for all f ∈ X ∗ . In this case, we write
xn  x or weak- lim xn = x.
n→∞

Weakly closed – A subset C of a Banach space X is said to be a weakly

closed if it is closed in the weak topology.

Weak Cauchy sequence – A sequence {xn } in a normed space X is said

to be a weak Cauchy if for each f ∈ X ∗ , {f (xn )} is a Cauchy sequence in K.

Weakly complete – A normed space X is said to be weakly complete if every

weak Cauchy sequence in X converges weakly to some element in X.
1.9. Weak topologies 37

Weakly compact – A subset C of a normed space X is said to be weakly

compact if C is compact in the weak topology.

Schur property – A Banach space is said to satisfy Schur property if there

exist weakly convergent sequences that are norm convergent.
Theorem 1.9.1 (Schur’s theorem) – In 1, weak and norm convergences of
sequences coincide.
We have the following basic properties of weakly convergent sequences in
normed spaces:

Proposition 1.9.2 (Uniqueness of weak limit) – Let {xn } be a sequence

in a normed space X such that xn  x and xn  y. Then x = y.

Proof. Because {f (xn )} is a sequence of scalars such that f (xn ) → f (x) and
f (xn ) → f (y), it follows that f (x) = f (y). This implies that f (x − y) = 0.
Therefore, x = y by Corollary 1.6.8.
Proposition 1.9.3 (Strong convergence implies weak convergence) –
Let {xn } be a sequence in a normed space X such that xn → x. Then xn  x.
Proof. Because xn → x, xn − x → 0. Hence

|f (xn ) − f (x)| ≤ f ∗ xn − x → 0 for all f ∈ X ∗ .

Therefore, xn  x.

The converse of Proposition 1.9.3 is not true in general. It can be seen from
the following example:
Example 1.9.4 Let X = 2 and {xn } be a sequence in 2 such that

xn = (0, 0, 0, · · · , 1, 0, · · · ), n ∈ N.

nth position

For any y = (y1 , y2 , · · · , yn , · · · ) ∈ X ∗ = 2, we have

(xn , y) = yn → 0 as n → ∞.

Hence xn  0 as n → ∞. However, {xn } does not converge strongly because

xn  = 1 for all n ∈ N. Therefore, a weakly convergent sequence need not be
convergent in norm.

Theorem 1.9.5 (Weak convergence in p space, 1 < p < ∞ ∞) – For 1 <

p < ∞, let
(n) (n)
xn = (α1 , α2 , · · · , αin , · · · ) ∈ p , n ∈ N
38 1. Fundamentals

and
x = (α1 , α2 , · · · , αi , · · · ) ∈ p.

Then xn  x if and only if

(i) {xn } is bounded, i.e., xn  ≤ M for all n ∈ N and for some M ≥ 0;
(n)
(ii) for each i, αi → αi as n → ∞.
Theorem 1.9.6 Let X be a finite-dimensional normed space. Then strong con-
vergence is equivalent to weak convergence.
Theorem 1.9.7 Every reflexive normed space is weakly complete.
Convergence of sequences in B(X, Y) – Let X and Y be two normed spaces.
A sequence {Tn } in B(X, Y ) is said to be
(i) uniformly convergent to T ∈ B(X, Y ) in the norm of B(X, Y ) if
Tn − T B → 0 as n → ∞, i.e., for ε > 0, there exists an integer n0 ∈ N
such that sup Tn x − T x < ε for all n ≥ n0 ,
x≤1

[uniform convergence of {Tn }]

(ii) strongly convergent to T ∈ B(X, Y ) if lim Tn x − T x = 0 for all x ∈ X,
n→∞
[strong convergence of {Tn }]
(iii) weakly convergent to T ∈ B(X, Y ) if |f (Tn x) − f (T x)| → 0 for all x ∈ X
and f ∈ Y ∗ .
[weak convergence of {Tn }]
It follows immediately from the inequality
Tn x − T x ≤ Tn − T B x, x ∈ X
that the uniform convergence implies strong convergence. It can be easily ob-
served for the sequence of operators in B(X, Y ) that
uniform convergence ⇒ strong convergence ⇒ weak convergence.
We note that the converse is not true in general.

Weak* topology - We have seen that if τ is the norm topology of a normed

space X, then the weak topology σ(X, X ∗ ) is a subset of the original norm topo-
logy τ . Let τ ∗ be the norm topology of X ∗ generated by the norm  · ∗ (of X ∗ ).
Then there exists a topology denoted by σ(X ∗ , X) on X ∗ such that σ(X ∗ , X) ⊂
τ ∗ . The topology σ(X ∗ , X) is called the weak* topology on X ∗ . Thus, we
can speak about strong neighborhood, strongly closed, strongly bounded, weak
convergence in (X ∗ ,  · ∗ ) and weak* neighborhood, weak*ly closed, weak*ly
bounded, weak*ly convergence in (X ∗ , σ(X ∗ , X)), respectively.
We now study some basic properties of the weak topology and weak* topo-
logy. We begin with a simple characterization for the convergence of sequences
in the weak topologies.
1.9. Weak topologies 39

Proposition 1.9.8 Let X be a normed space and {fn } a sequence in X ∗ . Then

we have the following:

(a) {fn } converges strongly to f in the norm topology on X ∗ (denoted by

fn → f ) if
fn − f ∗ → 0.

(b) {fn } converges to f in the weak topology on X ∗ (denoted by fn  f ) if

fn − f, g → 0 for all g ∈ X ∗∗ .

(c) {fn } converges to f in the weak* topology on X ∗ (denoted by fn → f

weak*ly or fn ∗ f ) if

x, fn − f  → 0 for all x ∈ X.

On the other hand, the following result is an immediate consequence of

Theorem 1.5.3.

Corollary 1.9.9 Let C be a nonempty subset of a Banach space X. For each

f ∈ X ∗ , let f (C) = ∪x∈C x, f  be a bounded set in R. Then C is bounded.

Proof. Set X := X ∗ , Y := R, and Tx (f ) := x, f , x ∈ C. Then Tx ∈ B(X ∗ , R).

Because f (C) is bounded, it follows that

sup |Tx (f )| = sup |x, f | ≤ K,

x∈C x∈C

for some K > 0. By the uniform boundedness principle, there exists a constant
M > 0 such that
Tx  ≤ M for all x ∈ C.
This implies that

|x, f | = |Tx (f )| ≤ Tx f ∗ ≤ M f ∗ .

By Corollary 1.6.7, we have

x ≤ M for all x ∈ C.

Therefore, C is bounded.

Applying Corollary 1.9.9, we have

Theorem 1.9.10 Let {xn } be a sequence in a Banach space X. Then we have

the following:
(a) xn  x (in X) implies {xn } is bounded and x ≤ lim inf xn .
n→∞
(b) xn  x in X and fn → f in X ∗ imply fn (xn ) → f (x) in R.
40 1. Fundamentals

Proof. (a) Because xn  x, then f (xn ) → f (x) for all f ∈ X ∗ . Hence {f (xn )}
is bounded for all f ∈ X ∗ . Thus, by Corollary 1.9.9, {xn } is bounded.
Moreover,
|xn , f | ≤ xn f ∗ .
Taking liminf in the above inequality, we have

|x, f | ≤ lim inf xn f ∗ .

n→∞

By Corollary 1.6.7, we obtain

x = sup |x, f | ≤ sup (lim inf xn f ∗ ) ≤ lim inf xn .
f ∗ ≤1 f ∗ ≤1 n→∞ n→∞

(b) Because xn  x in X, it follows that xn − x, f  = f (xn ) − f (x) → 0 and

{xn } is bounded (by part (a)). Hence

|xn , fn  − x, f | ≤ |xn , fn  − xn , f | + |xn , f  − x, f |

= |xn , fn − f | + |xn − x, f |
≤ xn  fn − f ∗ + |xn − x, f |
≤ M fn − f ∗ + |xn − x, f | → 0

for some constant M > 0. Therefore, fn (xn ) → f (x).

Observation
• Let {xn } be a sequence in a Banach space X with xn  x ∈ X and {αn } a
sequence of scalars such that αn → α. Then {αn xn } converges weakly to αx.

Theorem 1.9.11 Let X be a Banach space and {xn } a sequence in X such

that xn  x ∈ X. Then there exists a sequence of convex combinations of {xn }
that converges strongly to x, i.e., there exists convex combination {yn } such that

m 
m
yn = λi xi , where λi = 1 and λi ≥ 0, n ≤ i ≤ m,
i=n i=n

which converges strongly to x.

Corollary 1.9.12 Let C a nonempty subset of a Banach space X and {xn } a

sequence in C such that xn  x ∈ X. Then x ∈ co(C).

The weak topology is weaker than the norm topology, and every w-closed
set is also norm closed. The following result shows that for convex sets, the
converse is also true.
Proposition 1.9.13 Let C be a convex subset of a normed space X. Then C
is weakly closed if and only if C is closed.
The following proposition is a generalization of Theorem 1.7.8.
1.9. Weak topologies 41

Proposition 1.9.14 Let C be a weakly compact subset of a Banach space X.

Then co(C) is also weakly compact.

The following result is a direct consequence of the uniform bounded principle:

Proposition 1.9.15 Let C be a weakly compact subset of a Banach space X.

Then C is bounded.

Theorem 1.9.16 (Eberlein-Smulian theorem) – Let C be a weakly closed

subset of a Banach space. Then the following are equivalent:

(a) C is weakly compact.

(b) C is weakly sequentially compact, i.e., each sequence {xn } in C has a

subsequence that converges weakly to a point in C.

Corollary 1.9.17 Let C be a closed convex subset of a Banach space. Then

the following are equivalent:

(a) C is weakly compact.

(b) Each sequence {xn } in C has a subsequence that converges weakly to a

point in C.

Proposition 1.9.18 Any closed convex subset of a weakly compact set is itself
weakly compact.

Theorem 1.9.19 (Kakutani’s theorem) – Let X be a Banach space. Then

X is reflexive if and only if the unit closed ball BX := {x ∈ X : x ≤ 1} is
weakly compact (i.e., BX is compact in the weak topology of X).

Using Proposition 1.9.13 and Kakutani’s theorem, we obtain

Theorem 1.9.20 Let X be a Banach space. Then X is reflexive if and only

if every closed convex bounded subset of X is weakly compact (compact in weak
topology).

Theorem 1.9.21 Let C be a subset of a reflexive Banach space. Then

C is weakly compact ⇔ C is bounded
(compactness in weak topology) (boundedness in strong topology)

Theorem 1.9.22 Let {xn } be a sequence in a weakly compact convex subset of

a Banach space X and ωw ({xn }) denote the set of all weak subsequential limits
of {xn }. Then co(ωw ({xn })) = ∩∞
n=1 co({xk }k≥n ).

Proof. Set W := ωw ({xn }), An := co({xk }k≥n ), and A := ∩∞

n=1 An . We now
show that co(W ) = A. The inclusion W ⊂ A (and hence co(W ) ⊂ A) is trivial.
42 1. Fundamentals

Hence it suffices to prove that A ⊂ co(W ). Suppose, for contradiction, that

x ∈ A \ co(W ). Then there exists j ∈ X ∗ such that

x, j > sup{y, j : y ∈ co(W )} = sup{y, j : y ∈ W . (1.10)

Because x ∈ A ⊂ An ,

x, j ≤ sup{y, j : y ∈ An } = sup{xk , j : k ≥ n}.

Therefore,
x, j ≤ lim supxn , j.
n→∞

It follows from the Eberlein-Smulian theorem that there exists a subsequence

{xni } of {xn } such that

xni  x and x, j ≤ x , j.

Because x ∈ W by definition, this is a contradiction of (1.10).

Corollary 1.9.23 Let X be a Banach space and {xn } a sequence in X weakly

convergent to z. Let An = co({xk }k≥n ). Then ∩∞
n=1 An = {z}.

Proposition 1.9.24 Let {xn } be a bounded sequence in reflexive Banach space

X and An = co({xn }k≥n ). If ∩∞ ∞
n=1 An = ∩n=1 co({xn , xn+1 , · · · }) = {x}, then
xn  x.

Proposition 1.9.25 Let {xn } be a weakly null sequence in a Banach space X

and {jn } a bounded sequence in X ∗ . Then for each ε > 0, there exists an
increasing sequence {nk } in N such that |xni , jnk | < ε if i = k.

Proof. Without loss of generality, we may assume that X is a separable space.

We can assume that {jn } converges weak*ly to some j ∈ BX ∗ . Given ε > 0, we
find n1 such that |xn , j| < ε/2 for all n ≥ n1 . Next, having n1 < n2 < · · · <
nk−1 , we pick nk > nk−1 with |xnk , jni | < ε and |xni , jnk − j| < ε/2 for all
i = 1, 2, · · · , k − 1. Then |xni , jnk | < ε.

We now list several properties that characterize reflexivity.

Theorem 1.9.26 Let X be a Banach space. Then following statements are

equivalent:

(a) X is reflexive.

(b) BX is weakly compact.

(c) Every bounded sequence in X in strong topology has a weakly convergent

subsequence.

(d) For any f ∈ X ∗ , there exists x ∈ BX such that f (x) = f ∗ .

1.10. Continuity of mappings 43

(e) X ∗ is reflexive.

(f ) σ(X ∗ , X) = σ(X ∗ , X ∗∗ ), i.e., on X ∗ the weak topology and the weak*

topology coincide.

(g) If {Cn } is any descending sequence of nonempty closed convex bounded

subsets of X, then ∩∞
n=1 Cn = ∅.

(h) For any closed convex bounded subset C of X and any j ∈ X ∗ , there exists
x ∈ C such that x, j = sup{y, j : y ∈ C}.

Finally, we give the fundamental result concerning the weak* topology.

Theorem 1.9.27 (Banach-Alaoglu’s theorem) – The unit ball BX ∗ of the

dual of a normed space X is compact in the weak* topology.

1.10 Continuity of mappings

In this section, we discuss various forms of continuity of mappings with their
properties.

Definition 1.10.1 Let T be a mapping from a metric space (X, d) into another
metric space (Y, ρ). Then T is said to be

(i) continuous at x0 ∈ X if xn → x0 implies T xn → T x0 in Y , i.e., for each

ε > 0, there is a δ = δ(ε, x0 ) > 0 such that ρ(T x0 , T y) < ε whenever
d(x0 , y) < δ for all y ∈ X,

(ii) uniformly continuous on X if for given ε > 0, there exists δ = δ(ε) > 0
such that

ρ(T x, T y) < ε whenever d(x, y) < δ for all x, y ∈ X.

Example 1.10.2 Let X = (0, 1] and Y = R and let X and Y have usual metric
defined by absolute value. Then the mapping T : X → Y defined by T x = 1/x
is continuous, but not uniformly continuous.

Observation

• Every uniformly continuous mapping from X into Y is continuous at each point

of X, but pointwise continuity does not necessary imply uniform continuity.

• Every uniformly continuous mapping T from a metric space X into another

metric space Y maps a Cauchy sequence in X into a Cauchy sequence in Y .

Proposition 1.10.3 Let T be a continuous mapping from a compact metric

space (X, d) into another metric space (Y, ρ). Then T is uniformly continuous.
44 1. Fundamentals

A mapping T from a metric space (X, d) into another metric space (Y, ρ) is
said to satisfy Lipschitz condition on X if there exists a constant L > 0 such
that
ρ(T x, T y) ≤ Ld(x, y) for all x, y ∈ X.

If L is the least number for which Lipschitz condition holds, then L is called
Lipschitz constant. In this case, we say that T is an L-Lipschitz mapping or
simply a Lipschitzian mapping with Lipschitz constant L. Otherwise, it is called
non-Lipschitzian mapping. An L-Lipschitz mapping T is said to be contraction
if L < 1 and nonexpansive if L = 1. The mapping T is said to be contractive if

ρ(T x, T y) < d(x, y) for all x, y ∈ X, x = y.

Remark 1.10.4 Every Lipschitz continuous mapping T from a metric space

X into another metric space Y is uniformly continuous on X. Indeed, choose
δ < ε/L (independent of x), and we get

ρ(T x, T x0 ) ≤ Ld(x, x0 ) < ε.

The following example shows that the distance functional f (x) = d(x, C) is
nonexpansive.
Example 1.10.5 Let C be a nonempty subset of a normed space X. Then for
each pair x, y in X
|d(x, C) − d(y, C)| ≤ x − y.
In particular, the function x → d(x, C) is nonexpansive and hence uniformly
continuous.
The following proposition guarantees the existence of Lipschitzian mappings.

Proposition 1.10.6 Let T : [a, b] ⊂ R → R be a differentiable function on

(a, b). Suppose T  is continuous on [a, b]. Then T is a Lipschitz continuous
function (and hence is uniformly continuous).
Proof. By the Lagrange’s theorem, we have

T y − T x = T  (c)(y − x) for all a ≤ x < y ≤ b,

where c ∈ (x, y) ⊂ [a, b]. Because T  is continuous and interval [a, b] is compact
in R, by Weierstrass theorem, there exists x0 ∈ [a, b] such that

L = |T  (x0 )| = sup |T  (c)|.

c∈[a,b]

Thus, |T x − T y| ≤ L|x − y|, which proves that T is Lipschitz continuous.

The following example shows that there is a Lipschitzian mapping for which
T  does not exist.
1.10. Continuity of mappings 45

Example 1.10.7 The function T x = |x|, x ∈ [−1, 1] satisfies Lipschitz condi-

tion with L = 1, i.e., |T x − T y| ≤ |x − y| for all x, y ∈ [−1, 1]. Note T is not
differentiable at zero.

We now give an example of a non-Lipschitzian mapping that is continuous.

   
Example 1.10.8 Let T : − π1 , π1 → − π1 , π1 be a mapping defined by

0 if x = 0,
Tx = x
2 sin(1/x) if x = 0.

Then T is continuous, but not Lipschitz continuous.

For linear mappings, the continuity condition can be restated in terms of

uniform continuity.

Proposition 1.10.9 Let X and Y be two normed spaces and T : X → Y a

linear mapping. Then the following conditions are equivalent:
(a) T is continuous.
(b) T is Lipschitz function: there exists M > 0 such that T x ≤ M x for
all x ∈ X.
(c) T is uniformly continuous.

Let X and Y be two Banach spaces and let T be a mapping from X into Y .
Then the mapping T is said to be

1. bounded if C is bounded in X implies T (C) is bounded;

2. locally bounded if each point in X has a bounded neighborhood U such

that T (U ) is bounded;

3. weakly continuous if xn  x in X implies T xn  T x in Y ;

4. demicontinuous if xn → x in X implies T xn  T x in Y ;

5. hemicontinuous at x0 ∈ X if for any sequence {xn } converging to x0

along a line implies T xn  T x0 , i.e., T xn = T (x0 + tn x)  T x0 as tn →
0 for all x ∈ X;

6. closed if xn → x in X and T xn → y in Y imply T x = y;

7. weakly closed if xn  x ∈ X and T xn  y in Y imply T x = y;

8. demiclosed if xn  x in X and T xn → y in Y imply T x = y;

9. compact if C is bounded implies T (C) is relatively compact (T (C) is com-

pact), i.e., for every bounded sequence {xn } in X, {T xn } has convergent
subsequence in Y ;
46 1. Fundamentals

10. completely continuous if it is continuous and compact;

11. demicompact if any bounded sequence {xn } in X such that {xn − T xn }
converges strongly has a convergent subsequence.

In the case of linear mappings, the concepts of continuity and boundedness

are equivalent, but it is not true in general.

Proposition 1.10.10 Every continuous linear mapping T : X → Y is weakly

continuous.

Proposition 1.10.11 Let X be a reflexive Banach space and Y a general

Banach space. Then every weakly continuous mapping T : X → Y is bounded.

Proposition 1.10.12 A completely continuous mapping maps a weakly con-

vergent sequence into a strongly convergent.

Proposition 1.10.13 Every linear mapping is hemicontinuous.

Proof. Every linear and demicontinuous mapping is continuous.

It is clear that every demicontinuous mapping is hemicontinuous, but the
converse is not true.
Example 1.10.14 Let X = R2 , Y = R, and T : X → Y a mapping defined by

x2 y
x4 +y 2 , (x, y) = (0, 0),
T (x, y) =
0, (x, y) = (0, 0).

Then T is hemicontinuous at (0, 0), but not demicontinuous at (0, 0).

Let X and Y be two sets. A multivalued T from X to Y , denoted by
T : X → Y , is a subset T ⊆ X × Y . The inverse of T : X → Y is a multivalued
function T −1 : Y → X defined by (y, x) ∈ T −1 if and only if (x, y) ∈ T. The
values of T are the sets T x = {y ∈ Y : (x, y) ∈ T }; the fibers of T are the sets
T −1 (y) = {x ∈ X : (x, y) ∈ T } for y ∈ Y .

For A ⊂ X, the set

T (A) = ∪x∈A T x = {y ∈ Y : T −1 (y) ∩ A = ∅}

is called the image of A under T ; for B ⊂ Y , the set

T −1 (B) = ∪y∈B T −1 (y) = {x ∈ X : T x ∩ B = ∅},

the image of B under T −1 , is called inverse image of B under T . A point of

a set that is invariant under any transformation is called a fixed point of the
transformation. A point x0 ∈ X is said to be a fixed point of T if x0 ∈ T x0 .

Let X and Y be two topological spaces. Then a multivalued function T :

X → Y is said to be upper semicontinuous (lower semicontinuous) if the inverse
1.10. Continuity of mappings 47

image of a closed set (open set) is closed (open). A multivalued function is

continuous if it is both upper and lower semicontinuous.

Finally, we conclude the chapter with the following important fixed point
theorems.

Theorem 1.10.15 (Brouwer’s fixed point theorem) – Every continuous

mapping from the unit ball of Rn into itself has a fixed point.

Theorem 1.10.16 (Schauder’s fixed point theorem) – Let C be a non-

empty closed convex bounded subset of a Banach space X. Then every continu-
ous compact mapping T : C → C has a fixed point.

Theorem 1.10.17 (Tychonoff ’s fixed point theorem) – Let C be a non-

empty compact convex subset of a locally convex topological linear space X and
T : C → C a continuous mapping. Then T has a fixed point.

Exercises

1.1 Let (X, d) be a metric space. Show that ρ(x, y) = min{1, d(x, y)} for all
x, y ∈ X is also a metric space.

1.2 Give an example of a seminorm that is not a norm.

1.3 Let ·, · be an inner product on a linear space X and T : X → X a one-one

linear mapping. Let x, yT = T x, T y for all x, y ∈ X. Show that ·, ·T
is an inner product space.
1.4 Show that the space c0 of all real sequences converging to 0 is a normed
∞
space with norm x = n=1 |xn − xn+1 | < ∞.

1.5 Let c00 be a normed space with p -norm (1 ≤ p ≤ ∞) and {fn } a sequence of
functional on c00 defined by fn (x) = nxn for all x = (x1 , x2 , · · · , xn , · · · ).
Show that fn (x) → 0 for every x ∈ c00 , but fn  = n for all n.

1.6 Show that the space p (1 < p < ∞) is reflexive, but 1 is not reflexive.

1.7 Let C be a nonempty closed convex subset of a normed space X and {xn }
a sequence in C such that xn  x in X. Show that x ∈ C.

1.8 Let {xn } be a sequence in a normed space X such that xn  x. Show that
x ∈ span {xn }.

1.9 Let {xn } be a sequence in normed space X such that xn  x. Show that
{xn } is bounded.

1.10 Let X = c00 or c0 with norm  · ∞ . Show that xn  x in X if and only

if {xn } is bounded in X and xn, i → xi as n → ∞ for each i = 1, 2, · · · .
Chapter 2

Convexity, Smoothness, and

Duality Mappings

Geometric structures such as convexity and smoothness of Banach spaces play

an important role in the existence and approximation of fixed points of nonlinear
mappings. This chapter presents a substantial number of useful properties of
duality mappings and Banach spaces having these geometric structures.

2.1 Strict convexity

Let X be a linear space. The line segment or interval joining the two points
x, y ∈ X is the set [x, y] := {λx + (1 − λ)y : 0 ≤ λ ≤ 1}, i.e., [x, y] = co({x, y})
is convex hull of x and y.

The basic property of a norm of a Banach space X is that it is always convex,

i.e.,

(1 − λ)x + λy ≤ (1 − λ)x + λy for all x, y ∈ X and λ ∈ [0, 1].

A number of Banach spaces do not have equality when x = y, i.e.,

(1 − λ)x + λy < (1 − λ)x + λy

for all x, y ∈ X with x = y and λ ∈ (0, 1). (2.1)

We use SX to denote the unit sphere SX = {x ∈ X : x = 1} on Banach

space X. If x, y ∈ SX with x = y, then (2.1) reduces to

(1 − λ)x + λy < 1 for all λ ∈ (0, 1),

which says that the unit sphere SX contains no line segments. This suggests
strict convexity of norm.

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 2,

c Springer Science+Business Media, LLC, 2009
50 2. Convexity, Smoothness, Duality Mappings

Definition 2.1.1 A Banach space X is said to be strictly convex if

x, y ∈ SX with x = y ⇒ (1 − λ)x + λy < 1 for all λ ∈ (0, 1).

This says that the midpoint (x + y)/2 of two distinct points x and y in the
unit sphere SX of X does not lie on SX . In other words, if x, y ∈ SX with
x = y = (x + y)/2, then x = y.
Example 2.1.2 Consider X = Rn , n ≥ 2 with norm x2 defined by
n 1/2
x2 = 2
xi , x = (x1 , x2 , · · · , xn ) ∈ Rn .
i=1

Then X is strictly convex.

Example 2.1.3 Consider X = Rn , n ≥ 2 with norm  · 1 defined by

x1 = |x1 | + |x2 | + · · · + |xn |, x = (x1 , x2 , · · · , xn ) ∈ Rn .

Then X is not strictly convex. To see it, let

x = (1, 0, 0, · · · , 0) and y = (0, 1, 0, · · · , 0).

It is easy to see that x = y, x1 = 1 = y1 , but x + y1 = 2.

Example 2.1.4 Consider X = Rn , n ≥ 2 with norm  · ∞ defined by

x∞ = max |xi |, x = (x1 , x2 , · · · , xn ) ∈ Rn .

1≤i≤n

Then X is not strictly convex. Indeed, for x = (1, 0, 0, · · · , 0) and y = (1, 1,

0, · · · , 0), we have, x = y, x∞ = 1 = y∞ , but x + y∞ = 2.

The other equivalent conditions of strict convexity are given in the following:

Proposition 2.1.5 Let X be a Banach space. Then the following are equiva-
lent:
(a) X is strictly convex.
(b) For each nonzero f ∈ X ∗ , there exists at most one point x in X with
x = 1 such that x, f  = f (x) = f ∗ .
Proof. (a) ⇒ (b). Let X be a strictly convex Banach space and f an element
in X ∗ . Suppose there exist two distinct points x, y in X with x = y = 1
such that f (x) = f (y) = f ∗ . If t ∈ (0, 1), then

f ∗ = tf (x) + (1 − t)f (y) (as f (x) = f (y) = f ∗ )

= f (tx + (1 − f )y)
≤ f ∗ tx + (1 − t)y
< f ∗ , (as tx + (1 − t)y < 1)
2.1. Strict convexity 51

which is a contradiction. Therefore, there exists at most one point x in X with

x = 1 such that f (x) = f ∗ .
(b) ⇒ (a). Suppose x, y ∈ SX with x = y such that (x + y)/2 = 1.
By Corollary 1.6.6, there exists a functional j ∈ SX ∗ such that
j∗ = 1 and (x + y)/2, j = (x + y)/2.
Because x, j ≤ 1 and y, j ≤ 1, we have x, j = y, j. This implies, by
hypothesis, that x = y. Therefore, (b) ⇒ (a) is proved.
Proposition 2.1.6 Let X be a Banach space. Then the following statements
are equivalent:
(a) X is strictly convex.
(b) For every 1 < p < ∞,
tx+(1−t)yp < txp +(1−t)yp f or all x, y ∈ X, x = y and t ∈ (0, 1).

Proof. (a) ⇒ (b). Let X be strictly convex. Suppose x, y ∈ X with x = y.

By strict convexity of X,
tx + (1 − t)yp < (tx + (1 − t)y)p for all t ∈ (0, 1). (2.2)
If x = y, then
tx + (1 − t)yp < xp = txp + (1 − t)yp .
We now assume that x = y. Consider the function t → tp for 1 < p < ∞.
Then it is a convex function and
 p
a+b ap + bp
< for all a, b ≥ 0 and a = b.
2 2
Hence from (2.2) with t = 1/2, we have
   p
 x + y p
  ≤ x + y 1
< (xp + yp ). (2.3)
 2  2 2
If t ∈ (0, 1/2], then from (2.2), we have
 p
 x+y 
tx + (1 − t)yp =  2t + (1 − 2t)y 

2
   p
x + y 
≤ 2t
 2 
 + (1 − 2t)y
 
 x + y p
< 2t 
 2  + (1 − 2t)y
p

≤ txp + (1 − t)yp . (by (2.3))

The proof is similar if t ∈ (1/2, 1).
(b) ⇒ (a). It is obvious.
52 2. Convexity, Smoothness, Duality Mappings

Proposition 2.1.7 Let X be a strictly convex Banach space. If x + y =

x + y for 0 = x ∈ X and y ∈ X, then there exists t ≥ 0 such that y = tx.

Proof. Let x, y ∈ X\{0} be such that x+y = x+y. From Corollary 1.6.6,
there exists j ∈ X ∗ such that

x + y, j = x + y and j∗ = 1.

Because x, j ≤ x and y, j ≤ y, we must have x, j = x and
y, j = y. This means that x/x, j = y/y, j = 1. By strict con-
vexity of X, it follows from Proposition 2.1.5 that x/x = y/y. Therefore,
result holds.

We now present the existence and uniqueness of elements of minimal norm

in convex subsets of strictly convex Banach spaces.

Proposition 2.1.8 Let X be a strictly convex Banach space and C a nonempty

convex subset of X. Then there is at most one point x in C such that x =
inf {z : z ∈ C}.

Proof. Suppose, there exist two points x, y ∈ C, x = y such that

x = y = inf{z : z ∈ C} = d (say).

If t ∈ (0, 1), then by strict convexity of X we have that

tx + (1 − t)y < d,

which is a contradiction, as tx + (1 − y) ∈ C by the convexity of C.

Proposition 2.1.9 Let C be a nonempty closed convex subset of a reflexive

strictly convex Banach space X. Then there exists a unique point x ∈ C such
that x = inf{z : z ∈ C}.

Proof. Existence: Let d := inf{z : z ∈ C}. Then there exists a sequence

{xn } in C such that lim xn  = d. By the reflexivity of X, there exists a
n→∞
subsequence {xni } of {xn } that converges weakly to an element x in C. The
weak lower semicontinuity (w-lsc) of the norm (see Theorem 1.9.10) gives

x ≤ lim xn  = d.

n→∞

Therefore, d = x.

Uniqueness: It follows from Proposition 2.1.8 .

The following result has important applications in the existence and unique-
ness of best approximations.
2.2. Uniform convexity 53

Proposition 2.1.10 Let C be a nonempty closed convex subset of a reflexive

strictly convex Banach space X. Then for x ∈ X, there exists a unique point
zx ∈ C such that x − zx  = d(x, C).
Proof. Let x ∈ C. Because C is a nonempty closed convex subset Banach
space X, then D = {y − x : y ∈ C} is a nonempty closed convex subset
of X. By Proposition 2.1.9, there exists a unique point ux ∈ D such that
ux  = inf{y − x : y ∈ C}. For ux ∈ D, there exists a point zx ∈ C such
that ux = zx − x. Thus, there exists a unique point zx ∈ C such that
zx − x = d(x, C).

2.2 Uniform convexity

The strict convexity of a normed space X says that the midpoint (x + y)/2 of
the segment joining two distinct points x, y ∈ SX with x − y ≥ ε > 0 does
not lie on SX , i.e.,  
x + y 
 
 2  < 1.
In such spaces, we have no information about 1 − (x + y)/2, the distance of
midpoints from the unit sphere SX . A stronger property than strict convexity
that provides information about the distance 1−(x+y)/2 is uniform convexity.

Definition 2.2.1 A Banach space X is said to be uniformly convex 1 if for any

ε, 0 < ε ≤ 2, the inequalities x ≤ 1, y ≤ 1 and x − y ≥ ε imply there
exists a δ = δ(ε) > 0 such that (x + y)/2 ≤ 1 − δ.
This says that if x and y are in the closed unit ball BX := {x ∈ X : x ≤ 1}
with x − y ≥ ε > 0, the midpoint of x and y lies inside the unit ball BX at a
distance of at least δ from the unit sphere SX .
Example 2.2.2 Every Hilbert space H is a uniformly convex space. In fact,
the parallelogram law gives us
x + y2 = 2(x2 + y2 ) − x − y2 f or all x, y ∈ H.
Suppose x, y ∈ BH with x = y and x − y ≥ ε. Then
x + y2 ≤ 4 − ε2 ,
so it follows that
(x + y)/2 ≤ 1 − δ(ε),
where δ(ε) = 1 − 1− ε2 /4. Therefore, H is uniformly convex.
Example 2.2.3 The spaces 1 and ∞ are not uniformly convex. To see it,
take x = (1, 0, 0, 0 · · · ), y = (0, −1, 0, 0, · · · ) ∈ 1 and ε = 1. Then
x1 = 1, y1 = 1, x − y1 = 2 > 1 = ε.

1 The concept of uniform convexity was introduced by Clarkson in 1936.

54 2. Convexity, Smoothness, Duality Mappings

However, (x + y)/21 = 1 and there is no δ > 0 such that (x + y)/21 ≤ 1 − δ.
Thus, 1 is not uniformly convex.
Similarly, if we take x = (1, 1, 1, 0, 0, · · · ), y = (1, 1, −1, 0, 0 · · · ) ∈ ∞ and
ε = 1, then
x∞ = 1, y∞ = 1, x − y∞ = 2 > 1 = ε.
Because (x + y)/2∞ = 1, ∞ is not uniformly convex.

Observation
• The Banach spaces p , n
p (whenever n is a nonnegative integer), and Lp [a, b]
with 1 < p < ∞ are uniformly convex.
• The Banach spaces 1 , c, c0 , ∞ , L1 [a, b], C[a, b] and L∞ [a, b] are not strictly
convex.

We derive some consequences from the definition of uniform convexity.

Theorem 2.2.4 Every uniformly convex Banach space is strictly convex.

Proof. Let X be a uniformly convex Banach space. It easily follows from

Definition 2.2.1 that X is strictly convex.

Remark 2.2.5 The converse of Theorem 2.2.4 is not true in general. Let β > 0
and let X = co with the norm  · β defined by

∞  2 1/2
xi
xβ = xco + β , x = {xi } ∈ co .
i=1
i

The spaces (co ,  · β ) for β > 0 are strictly convex, but not uniformly convex,
while c0 with its usual norm is not strictly convex.

Theorem 2.2.6 Let X be a uniformly convex Banach space. Then we have the
following:

(a) For any r and ε with r ≥ ε > 0 and elements x, y ∈ X with x ≤ r,
y ≤ r, x − y ≥ ε, there exists a δ = δ(ε/r) > 0 such that

(x + y)/2 ≤ r[1 − δ(ε/r)].

(b) For any r and ε with r ≥ ε > 0 and elements x, y ∈ X with x ≤ r,
y ≤ r, x − y ≥ ε, there exists a δ = δ(ε/r) > 0 such that

tx + (1 − t)y ≤ r[1 − 2 min{t, 1 − t}δ(ε/r)] for all t ∈ (0, 1).

Proof. (a) Suppose that x ≤ r, y ≤ r and x − y ≥ ε > 0. Then we have
that      
x    
  ≤ 1,  y  ≤ 1 and  x − y  ≥ ε > 0.
r r r r r
2.2. Uniform convexity 55

By the definition of uniform convexity, there exists δ = δ(ε/r) > 0 such that
 
x + y 
 
 2r  ≤ 1 − δ,

which yields  
x + y 
 
 2  ≤ r(1 − δ).
(b) When t = 1/2, we are done by Part (a). If t ∈ (0, 1/2], we have
x+y
tx + (1 − t)y = t(x + y) + (1 − 2t)y ≤ 2t  + (1 − 2t)y. (2.4)
2
From part (a), there exists a δ = δ(ε/r) > 0 such that
    
x + y 
 ≤r 1−δ ε .
 2  r

From (2.4), we have

  
ε
tx + (1 − t)y ≤ 2t 1 − δ r + (1 − 2t)r (as y ≤ r)
r
  
ε
≤ r 1 − 2tδ .
r

Now by the choice of t ∈ [1/2, 1), we have

tx + (1 − t)y = (2t − 1)x + (1 − t)(x + y)

 
x + y 
≤ (2t − 1)x + 2(1 − t) 2 

  
ε
≤ (2t − 1)r + 2(1 − t)r 1 − δ
r
  
ε
= r 1 − 2(1 − t)δ .
r
Therefore,
  
ε
tx + (1 − t)y ≤ r 1 − 2 min{t, 1 − t}δ .
r

Theorem 2.2.7 Let X be a Banach space. Then the following are equivalent:
(a) X is uniformly convex.
(b) For two sequences {xn } and {yn } in X,

xn  ≤ 1, yn  ≤ 1 and lim xn + yn  = 2 ⇒ lim xn − yn  = 0. (2.5)

n→∞ n→∞
56 2. Convexity, Smoothness, Duality Mappings

Proof. (a) ⇒ (b). Suppose X is uniformly convex. Let {xn } and {yn } be
two sequences in X such that xn  ≤ 1, yn  ≤ 1 for all n ∈ N and
lim xn + yn  = 2. Suppose, for contradiction, that lim xn − yn  = 0. Then
n→∞ n→∞
for some ε > 0, there exists a subsequence {ni } of {n} such that
xni − yni  ≥ ε.
Because X is uniformly convex, there exists δ(ε) > 0 such that
xni + yni  ≤ 2(1 − δ(ε)). (2.6)
Because lim xn + yn  = 2, it follows from (2.6) that
n→∞

2 ≤ 2(1 − δ(ε)),
a contradiction.
(b) ⇒ (a). Suppose condition (2.5) is satisfied. If X is not uniformly convex,
for ε > 0, there is no δ(ε) such that
x ≤ 1, y ≤ 1, x − y ≥ ε ⇒ x + y ≤ 2(1 − δ(ε)),
and then we can find sequences {xn } and {yn } in X such that
(i) xn  ≤ 1, yn  ≤ 1,
(ii) xn + yn  ≥ 2(1 − 1/n),
(iii) xn − yn  ≥ ε.
Clearly xn − yn  ≥ ε, which contradicts the hypothesis, as (ii) gives
lim xn + yn  = 2. Thus, X must be uniformly convex.
n→∞

For the class of uniform convex Banach spaces, we have the following
important results:
Theorem 2.2.8 Every uniformly convex Banach space is reflexive.
Proof. Let X be a uniformly convex Banach space. Let SX ∗ := {j ∈ X ∗ :
j∗ = 1} be the unit sphere in X ∗ and f ∈ SX ∗ . Suppose {xn } is a sequence in
SX such that f (xn ) → 1. We show that {xn } is a Cauchy sequence. Suppose,
for contradiction, that there exist ε > 0 and two subsequences {xni } and {xnj }
of {xn } such that xni − xnj  ≥ ε. The uniform convexity of X guarantees that
there exists δ(ε) > 0 such that (xni + xnj )/2 < 1 − δ. Observe that
|f ((xni + xnj )/2)| ≤ f ∗ (xni + xnj )/2 < f ∗ (1 − δ) = 1 − δ
and f (xn ) → 1, yield a contradiction. Hence {xn } is a Cauchy sequence and
there exists a point x in X such that xn → x. Clearly, x ∈ SX . In fact,
x =  lim xn  = lim xn  = 1.
n→∞ n→∞

Using the James theorem (which states that a Banach space is reflexive if and
only if for each f ∈ SX ∗ , there exists x ∈ SX such that f (x) = 1), we conclude
that X is reflexive.
2.2. Uniform convexity 57

Remark 2.2.9 Every finite-dimensional Banach space is reflexive, but it need

not be uniformly convex, for example, X = Rn , n ≥ 2 with the norm x1 =
n
|xi |.
i=1

Combining Proposition 2.1.9 and Theorems 2.2.4 and 2.2.8, we obtain the
following interesting result:

Theorem 2.2.10 Let C be a nonempty closed convex subset of a uniformly

convex Banach space X. Then C has a unique element of minimum norm, i.e.,
there exists a unique element x ∈ C such that x = inf{z : z ∈ C}.

We now introduce a useful property.

Definition 2.2.11 A Banach space X is said to have the Kadec-Klee property if

for every sequence {xn } in X that converges weakly to x where also xn  → x,
then {xn } converges strongly to x.

Remark 2.2.12 In Definition 2.2.11, the sequence {xn } can be replaced by the
net {xα } for the definition of the Kadec property.

The following result has a very useful property:

Theorem 2.2.13 Every uniformly convex Banach space has the Kadec-Klee
property.

Proof. Let X be a uniformly convex Banach space. Let {xn } be a sequence

in X such that xn  x ∈ X and xn  → x. If x = 0, then lim xn  = 0,
n→∞
which yields that lim xn = 0.
n→∞
Suppose x = 0. Then we show that xn → x. Suppose, for contradiction, that
lim xn = x, i.e., xn /xn   x/x. Then for ε > 0, there exists a subsequence
n→∞
{xni /xni } of {xn /xn } such that
 
 xni x 
 
 xn  − x  ≥ ε > 0.
i

Because X is uniformly convex, there exists δ(ε) > 0 such that

 
1 
 xni + x  ≤ 1 − δ.
2 xni  x 


Because xn  x and xn  → x imply xn /xn   x/x, it follows that

   
 x   
  ≤ lim inf 1  xni + x  ≤ 1 − δ,
 x  
n→∞ 2 xni  x 

a contradiction. Therefore, {xn } converges strongly to x ∈ X.

58 2. Convexity, Smoothness, Duality Mappings

2.3 Modulus of convexity

Definition 2.3.1 Let X be a Banach space. Then a function δX : [0, 2] → [0, 1]
is said to be the modulus of convexity of X if
   
x + y 

δX (ε) = inf 1 −   : x ≤ 1, y ≤ 1, x − y ≥ ε .
2 

It is easy to see that δX (0) = 0 and δX (t) ≥ 0 for all t ≥ 0.

Example 2.3.2 For the case of a Hilbert space H (see Example 2.2.2),

ε2
δH (ε) = 1 − 1 − , ε ∈ (0, 2].
4
We now give the modulus of convexity for p (2 ≤ p < ∞) spaces. The
following result gives an analogue of the parallelogram law in p (2 ≤ p < ∞)
spaces.
Proposition 2.3.3 In p (2 ≤ p < ∞) spaces,

x + yp + x − yp ≤ 2p−1 (xp + yp ) for all x, y ∈ p. (2.7)

Proof. We observe from Lemma A.1.1 of Appendix A that for a, b ∈ R and

p ∈ [2, ∞)

|a + b|p + |a − b|p ≤ [|a + b|2 + |a − b|2 ]p/2

≤ [2|a|2 + 2|b|2 ]p/2
= 2p/2 (|a|2 + |b|2 )p/2
≤ 2p/2 2(p−2)/2 (|a|p + |b|p )
= 2p−1 (|a|p + |b|p ).

Hence for x = {xi }∞ ∞

i=1 , y = {yi }i=1 ∈ p, we have
∞
 ∞
 
∞ ∞
 
|xi + yi |p + |xi − yi |p ≤ 2p−1 |xi |p + |yi |p ,
i=1 i=1 i=1 i=1

which implies that points x, y ∈ p (2 ≤ p < ∞) satisfy the following analogue

of the parallelogram law:

x + yp + x − yp ≤ 2p−1 (xp + yp ).

Example 2.3.4 For the p (2 ≤ p < ∞) space,

  p 1/p
ε
δp (ε) = 1 − 1 − , ε ∈ (0, 2).
2
2.3. Modulus of convexity 59

To see this, let ε ∈ (0, 2) and x, y ∈ p such that x ≤ 1, y ≤ 1 and
x − y ≥ ε. Then from (2.7), we have

x + yp ≤ 2p − x − yp ,

which implies that

    p 1/p    p 1/p 
x + y  ε ε
  ≤ 1− =1− 1− 1−
 2  2 2
≤ 1 − δp (ε),
  p 1/p
where δp (ε) ≥ 1 − 1 − ε
2 .

Observation
• δH (ε) = 1 − 1 − (ε/2)2 .

• δp (ε) = 1 − (1 − (ε/2)p )p/2 .

• δp (ε), the modulus of convexity for p (1 < p ≤ 2) satisfies the following implicit
formula: p p
ε ε
1 − δp (ε) + + 1 − δp (ε) − = 2.
2 2
• δp (ε) > 0 for all ε > 0 (1 < p < ∞).
• δX (ε) ≤ δH (ε) for any Banach spaces X and any Hilbert space H, i.e., a Hilbert
space is the most convex Banach space.

We now give some important properties of the modulus of convexity of

Banach spaces.
Theorem 2.3.5 A Banach space X is strictly convex if and only if δX (2) = 1.
Proof. Let X be a strictly convex Banach space with modulus of convexity δX .
Suppose x = y = 1 and x − y = 2 with x = −y. By strict convexity of
X, we have    
 x − y   x + (−y) 
1=  2  
=  < 1,

2
a contradiction. Hence x = −y. Therefore, δX (2) = 1.
Conversely, suppose δX (2) = 1. Let x, y ∈ X such that x = y =
(x + y)/2 = 1. Then
   
x − y   
  =  x + (−y)  ≤ 1 − δX (x − (−y)) = 1 − δX (2) = 0,
 2   2 

which implies that x = y. Thus, x = y and x + y = 2 = x + y imply
that x = y. Therefore, X is strictly convex.
60 2. Convexity, Smoothness, Duality Mappings

Theorem 2.3.6 A Banach space X is uniformly convex if and only if δX (ε) > 0
for all ε ∈ (0, 2].
Proof. Let X be a uniformly convex Banach space. Then for ε > 0, there
exists δ(ε) > 0 such that
 
x + y 
0 < δ(ε) ≤ 1 − 
 2 


for all x, y ∈ X with x ≤ 1, y ≤ 1 and x − y ≥ ε. Therefore, from the

definition of modulus of convexity, we have δX (ε) > 0.
Conversely, suppose X is a Banach space with modulus of convexity δX such
that δX (ε) > 0 for all ε ∈ (0, 2]. Let x, y ∈ X such that x = 1, y = 1 with
x − y ≥ ε for fixed ε ∈ (0, 2]. By the modulus of convexity δX (ε), we have
 
x + y 
0 < δX (ε) ≤ 1 −   ,
2 
which implies that  
x + y 
 
 2  ≤ 1 − δ(ε),
where δ(ε) = δX (ε), which is independent of x and y. Therefore, X is uniformly
convex.

Theorem 2.3.7 Let X be a Banach space with modulus of convexity δX . Then

we have the following:
(a) For all ε1 and ε2 with 0 ≤ ε1 < ε2 ≤ 2,
ε2 − ε 1 ε2 − ε 1
δX (ε2 ) − δX (ε1 ) ≤ (1 − δX (ε1 )) ≤ .
2 − ε1 2 − ε1
In particular, δX is a continuous function on [0, 2).
(b) δX (s)/s is a nondecreasing function on (0, 2].
(c) δX is a strictly increasing function if X is uniformly convex.

Proof. (a) We define the set

Su,v = {(x, y) : x, y ∈ BX ; x−y = au, x+y = bv for some u, v ∈ X and a, b ≥ 0}

and the function

   
x + y 

δu,v (ε) = inf 1 −   : x, y ∈ Su,v , x − y ≥ ε .
2 

Note that δu,v (0) = 0. For given ε1 and ε2 in (0, 2] and η > 0, we can choose
(xi , yi ) in Su,v such that
 
 xi + yi 
xi − yi  ≥ εi and δu,v (εi ) + η ≥ 1 −  
 2  for i = 1, 2.
2.3. Modulus of convexity 61

Now for t ∈ [0, 1], let x3 = tx1 + (1 − t)x2 and y3 = ty1 + (1 − t)y2 . Because
xi , yi ∈ BX for i = 1, 2, it follows that

x3  ≤ tx1  + (1 − t)x2  ≤ 1

and
y3  ≤ ty1  + (1 − t)y2  ≤ 1.
Because (xi , yi ) ∈ Su,v , there exist positive constants ai , bi ≥ 0 with i = 1, 2 such
that xi −yi = ai u and xi +yi = bi v. Set α := ta1 +(1−t)a2 and β := tb1 +(1−t)b2 .
Then

x3 − y3 = t(x1 − y1 ) + (1 − t)(x2 − y2 )
= ta1 u + (1 − t)a2 u
= (ta1 + (1 − t)a2 )u
= αu.

Similarly, x3 + y3 = βv. Thus, (x3 , y3 ) is in Su,v .

Observe that

x3 − y3  = (ta1 + (1 − t)a2 )u

= tx1 − y1  + (1 − t)x2 − y2 
≥ tε1 + (1 − t)ε2 by the choice of xi , yi ,

and x3 + y3  = tx1 + y1  + (1 − t)x2 + y2 .

By the definition of the function δu,v (·), we have
 
 x3 + y3 
δu,v (tε1 + (1 − t)ε2 ) ≤ 1 −   
2 
   
 x1 + y1   x2 + y2 
≤ 1 − t   − (1 − t)  
2  2 
     
 x1 + y1   x2 + y2 
= t 1−   + (1 − t) 1 −   
2  2 
   
η η
≤ t δu,v (ε1 ) + + (1 − t) δu,v (ε2 ) +
2 2
η
= tδu,v (ε1 ) + (1 − t)δu,v (ε2 ) + .
2
Because η is arbitrary, it follows that δu,v (ε) is a convex function of ε.
Note that
δX (ε) ≤ δu,v (ε) for all u, v
and
(x, y) ∈ Su,v with x ≤ 1 and y ≤ 1 for some u, v ∈ X;
and hence
δX (ε) = inf{δu,v (ε) : u, v ∈ X \ {0}}.
62 2. Convexity, Smoothness, Duality Mappings

Now for any real number ε > 0, there exist u, v ∈ X such that
δu,v (ε1 ) ≤ δX (ε1 ) + ε.
Hence
   
ε2 − ε1 ε2 − ε1
δu,v (ε2 ) = δu,v 2 + 1− ε1
2 − ε1 2 − ε1
 
ε2 − ε 1 ε2 − ε1
≤ δu,v (2) + 1 − δu,v (ε1 ),
2 − ε1 2 − ε1
which implies that
ε2 − ε 1
δu,v (ε2 ) − δu,v (ε1 ) ≤ (δu,v (2) − δu,v (ε1 ))
2 − ε1
ε2 − ε 1
≤ (1 − δX (ε1 )).
2 − ε1
Then we have
δX (ε2 ) − δX (ε1 ) ≤ δu,v (ε2 ) − δu,v (ε1 ) + ε
ε2 − ε 1
≤ (1 − δX (ε1 )) + ε.
2 − ε1
Because ε > 0 is arbitrary, we have
 
ε2 − ε 1
δX (ε2 ) − δX (ε1 ) ≤ 1 − δX (ε1 ) .
2 − ε1
Because δX (ε1 ) ≥ 0, we have
ε2 − ε 1
δX (ε2 ) − δX (ε1 ) ≤ ,
2 − ε1
which implies that δX (·) is continuous on [0, 2).

(b) Fix s ∈ (0, 2] with s ≤ ε and x, y ∈ SX and x − y = ε.

Set
s x+y x+y
t := , u := tx + (1 − t) and v := ty + (1 − t) .
ε x + y x + y
Then
 
u+v x+y t
u − v = t(x − y), u − v = s and = x + y + 1 − t .
2 x + y 2
Thus,
   
 x+y u + v  
 −  = t − t x + y 
 x + y 2   2 
  
x + y 
= 1 − 1 − t + t  2 

 
u + v 
= 1−  2 .

2.3. Modulus of convexity 63

Observe that
     
 x+y x + y 1 1 x + y 
   
 x + y − 2  = x + y − 2 x + y = 1 −  2 

and
    
 x+y u + v u + v 
   
 x + y − 2  u − v = 1−
2 
s
  
x + y 
= 
1 − (1 − t) − t  s
2 
  
x + y 
= 1−  2 
 x − y.

Hence
δX (s)
≤ (1 − (u + v)/2)/u − v
s
= ((x + y)/x + y − (u + v)/2)/u − v = (1 − (x + y)/2)/ε.
By taking the infimum over all possible x and y with ε = x − y and x, y ∈ SX ,
we obtain
δX (s) δX (ε)
≤ .
s ε
(c) Observe that
δX (s) δX (t)
≤ for s < t ≤ 2 and δX (t) > 0.
s t
Hence
tδX (s) ≤ sδX (t) < tδX (t),
which implies that
δX (s) < δX (t).
Therefore, δX is a strictly increasing function.
Remark 2.3.8 The modulus of convexity δX need not be convex on [0,2] and
need not be continuous at t = 2.
Theorem 2.3.9 Let X be a Banach space with modulus of convexity δX . Then
tx + (1 − t)y ≤ 1 − 2 min{t, 1 − t}δX (x − y)
for all x, y ∈ X with x ≤ 1, y ≤ 1 and all t ∈ [0, 1].

Proof. The result follows from Theorem 2.2.6(b).

Corollary 2.3.10 Let X be a Banach space with modulus convexity δX . Then
(1 − t)x + ty ≤ 1 − 2t(1 − t)δX (x − y)
for all x, y ∈ X with x ≤ 1, y ≤ 1 and all t ∈ [0, 1].
64 2. Convexity, Smoothness, Duality Mappings

Proof. Because t(1 − t) ≤ min{t, 1 − t} for all t ∈ [0, 1], the result follows
Theorem 2.3.9.

Corollary 2.3.11 Let X be a uniformly convex Banach space with modulus of

convexity δX . If r > 0 and x, y ∈ X with x ≤ r, y ≤ r, then
  
x − y
tx + (1 − t)y ≤ r 1 − 2 min{t, 1 − t}δX f or all t ∈ (0, 1).
r

Theorem 2.3.12 Let X be a uniformly convex Banach space X. Then there

exists a strictly increasing continuous convex function g : R+ → R+ with
g(0) = 0 such that

2t(1 − t)g(x − y) ≤ 1 − (1 − t)x + ty

for all x, y ∈ X with x ≤ 1, y ≤ 1 and all t ∈ [0, 1].

Proof. Let δX be the modulus of convexity of X. Define a function g : R+ →

R+ by
 λ
1
2 0 δX (s)ds if 0 ≤ λ ≤ 2,
g(λ) =
g(2) + 12 δX (2)(λ − 2) if λ > 2.

For t ∈ (0, 2], we have

t
1 t
0 < g(t) = δX (s)ds ≤ δX (t) ≤ δX (t). (as δX (s) ≤ δX (t))
2 0 2
From the definition of g, we have
1
g  (t) = δX (t) for all t ∈ [0, 2].
2
Hence g  is increasing with g  (2) = δX (2)/2 = 1/2, and it follows that g is
convex.
Now, let x ≤ 1, y ≤ 1 and t ∈ [0, 1]. Then, we have (see Coroll-
ary 2.3.10)

(1 − t)x + ty ≤ 1 − 2t(1 − t)δX (x − y). (2.8)

Hence from (2.8) we have

x−y
2t(1 − t)g(x − y) = t(1 − t) δX (s)ds
0
≤ t(1 − t)δX (x − y)x − y
≤ 2t(1 − t)δX (x − y)
≤ 1 − (1 − t)x + ty.
2.3. Modulus of convexity 65

Moreover, for rs < 2, the function s → g(rs)/s is increasing (as (g(rs)/s) =

[rsδX (rs)/2 − g(rs)]/s2 ≥ 0). Therefore, g is a strictly increasing continuous
convex function.
Using Corollary 2.3.11, we obtain the following, which has important
applications in approximation of fixed points of nonlinear mappings in Banach
spaces.

Theorem 2.3.13 Let X be a uniformly convex Banach space and let {tn } be
a sequence of real numbers in (0,1) bounded away from 0 and 1. Let {xn } and
{yn } be two sequences in X such that

lim sup xn  ≤ a, lim sup yn  ≤ a and lim sup tn xn + (1 − tn )yn  = a
n→∞ n→∞ n→∞

for some a ≥ 0. Then lim xn − yn  = 0.

n→∞

Proof. The case a = 0 is trivial. So, let a > 0. Suppose, for contradiction, that
{xn −yn } does not converge to 0. Then there exists a subsequence {xni −yni } of
{xn − yn } such that inf i xni − yni  > 0. Note {tn } is bounded away from 0 and
1, and there exist two positive numbers α and β such that 0 < α ≤ tn ≤ β < 1
for all n ∈ N. Because lim sup xn  ≤ a and lim sup yn  ≤ a, we may assume
n→∞ n→∞
an r ∈ (a, a + 1) for a subsequence {ni } such that xni  ≤ r, yni  ≤ r, a < r.
Choose r ≥ ε > 0 such that

2α(1 − β)δX (ε/r) < 1 and xni − yni  ≥ ε > 0 for all i ∈ N.

From Corollary 2.3.11, we have

tni xni + (1 − tni )yni  ≤ r[1 − 2tni (1 − tni )δX (ε/r)]

≤ r[1 − 2α(1 − β)δX (ε/r)] < a for all i ∈ N,

which contradicts the hypothesis.

We now present the following intersection theorem:

Theorem 2.3.14 (Intersection theorem) – Let {Cn } be a decreasing

sequence of nonempty closed convex bounded subsets of a uniformly convex
Banach space X. Then ∩n∈N Cn is a nonempty closed convex subset of X.

Proof. Let z ∈ X be a point such that z ∈

/ C1 , rn = d(z, Cn ) and r = lim rn .
n→∞
Let {εn } be a sequence of positive numbers that decreases to zero. Set

Dn := Br+εn [z] = {x ∈ Cn : z − x ≤ r + εn },
dn : = diam(Dn ),
d: = lim dn .
n→∞
66 2. Convexity, Smoothness, Duality Mappings

Suppose x and y are two elements in Dn such that x − y ≥ dn − εn . Then

Corollary 2.3.11 gives
    
 
z − x + y  ≤ 1 − δX x − y (r + εn )
 2  r + εn
and hence   
d n − εn
rn ≤ 1 − δX (r + εn ).
r + εn
This yields a contradiction unless d = 0. This in turn implies that ∩n∈N Dn is
nonempty, and so is ∩n∈N Cn .
Remark 2.3.15 Theorem 2.3.14 remains valid if the sequence {Cn } is replaced
by an arbitrary decreasing net of nonempty closed convex bounded subsets of X.
We now study a weaker type convexity of Banach spaces that is called locally
uniform convexity.

Definition 2.3.16 A Banach space X is said to be locally uniformly convex if

for any ε > 0 and x ∈ SX , there exists δ = δ(x, ε) > 0 such that
 
x + y 

x − y ≥ ε implies that   ≤ 1 − δ for all y ∈ SX .
2 
The modulus of local convexity of the Banach space X is
 
x+y
δX (x, ε) = inf 1− : y ∈ SX , x−y ≥ ε for each x ∈ SX and 0 < ε ≤ 2.
2
One may easily see that the Banach space X is locally uniformly convex if
δX (x, ε) > 0 for all x ∈ SX and ε > 0.

Observation
• Every uniformly convex Banach space is locally uniformly convex.
• By Definition 2.3.16, every locally uniformly convex Banach space is strictly
convex.
We now give interesting properties of locally uniformly convex Banach spaces:
Proposition 2.3.17 Let X be a Banach space. Then the following are equiva-
lent:
(a) X is locally uniformly convex.
(b) Every sequence {xn } in SX and x ∈ SX with xn + x → 2 implies that
xn → x.
Proof. (a) ⇒ (b). By locally uniformly convexity of X, δX (x, ε) > 0 for all
ε > 0. Therefore,
xn + x
1− → 0 implies that xn − x → 0.
2
2.4. Duality mappings 67

(b) ⇒ (a). Let {xn } be a sequence in SX such that xn + x → 2 implies that
xn → x. Then
 
 xn + x 
xn − x ≥ ε > 0 implies that  
 2  < 1.

Hence, by the definition of modulus of locally uniform convexity, δX (x, ε) > 0.

Therefore, X is locally uniformly convex.
The following theorem is a generalization of Theorem 2.2.13.
Theorem 2.3.18 Every locally uniformly convex Banach space has the Kadec-
Klee property.
Proof. Let X be a locally uniformly convex Banach space. Let {xn } be
a sequence in X such that xn  x ∈ X and xn  → x. For x = 0, xn  → 0
implies that xn → 0. Suppose x = 0. Then
xn x xn x x
 ⇒ + 2 .
xn  x xn  x x
By w-lsc of the norm, we have
   
 x   
2 = 2  ≤ lim inf  xn + x 
 x  n→∞  xn  x 
 
xn  x
≤ lim sup + = 2,
n→∞ xn  x

which implies that xn /(xn ) + x/(x) → 2. By Proposition 2.3.17, we

conclude that xn /xn  → x/x. Therefore, xn → x.

2.4 Duality mappings

Definition 2.4.1 Let X ∗ be the dual of a Banach space X. Then a multivalued
∗
mapping J : X → 2X is said to be a (normalized) duality mapping if

Jx = {j ∈ X ∗ : x, j = x2 = j2∗ }.

Example 2.4.2 In a Hilbert space H, the normalized duality mapping is the

identity. To see this, let x ∈ H with x = 0. Note that H = H ∗ and

x, x = x · x implies x ∈ Jx.

Suppose y ∈ Jx. Then by the definition of J, we have x, y = xy and

x = y. Because

x − y2 = x2 + y2 − 2x, y,

it follows that x = y. Therefore, Jx = {x}.

68 2. Convexity, Smoothness, Duality Mappings

For a complex number, we define the “sign” function by


0 if α = 0,
sgn α =
α/|α| if α = 0.
Observation 
0 if α = 0,
• |sgn α| =
1  0.
if α =

0 if α = 0,
• α sgn α =
αα/|α| = |α|  0.
if α =

Example 2.4.3 In the 2 space,

Jx = (|x1 |sgn(x1 ), |x2 |sgn(x2 ), · · · , |xi |sgn(xi ), · · · ), x = {xi } ∈ 2.

Example 2.4.4 In the L2 [0, 1] (1 < p < ∞) space, the duality mapping is given
by 
|x| sgn(x)/x, if x = 0,
Jx =
0 if x = 0.
Before giving fundamental properties of duality mappings, we need the
following notations and definitions:
∗
Let T : X → 2X a multivalued mapping. The domain Dom(T ), range
R(T ), inverse T −1 , and graph G(T ) of T are defined as

Dom(T ) = {x ∈ X : T x = ∅},
R(T ) = ∪x∈Dom(T ) T x,
T −1 (y) = {x ∈ X : y ∈ T x},
G(T ) = {(x, y) ∈ X × X ∗ : y ∈ T x, x ∈ Dom(T )}.

The graph G(T ) of T is a subset of X × X ∗ .

The mapping T is said to be

(i) monotone if x−y, jx −jy  ≥ 0 for all x, y ∈ Dom(T ) and jx ∈ T x, jy ∈ T y.
(ii) strictly monotone if x − y, jx − jy  > 0 for all x, y ∈ Dom(T ) with x = y
and jx ∈ T x, jy ∈ T y.
(iii) α-monotone if there exists a continuous strictly increasing function
α : [0, ∞) → [0, ∞) with α(0) = 0 and α(t) → ∞ as t → ∞ such that

x − y, jx − jy  ≥ α(x − y)x − y

for all x, y ∈ Dom(T ), jx ∈ T x, jy ∈ T y.

(iv) strongly monotone if T is α-monotone with α(t) = kt for some constant
k > 0.
(v) injective if T x ∩ T y = ∅ for x = y.
2.4. Duality mappings 69

∗
The monotone operator T : Dom(T ) ⊂ X → 2X is said to be maximal
monotone if it has no proper monotone extensions, i.e., if for (x, y) ∈ X × X ∗

x − z, y − jz  ≥ 0 for all z ∈ Dom(T ) and jz ∈ T z implies y ∈ T x.

The mapping T : Dom(T ) ⊂ X → X ∗ is said to be coercive on a subset C

of Dom(T ) if there exists a function c : (0, ∞) → [−∞, ∞] with c(t) → ∞ as
t → ∞ such that x, T x ≥ c(x)x for all x ∈ C.
x,T x
In other words, T is coercive on C if x → ∞ as x → ∞, x ∈ C.

Observation
• Every monotonically increasing mapping is monotone.

• If H is a Hilbert space and T : H → H is nonexpansive, then I − T is monotone.

We are now in a position to establish fundamental properties of duality

mappings in Banach spaces.
∗
Proposition 2.4.5 Let X be a Banach space and let J : X → 2X be the
normalized duality mapping. Then we have the following:
(a) J(0) = {0}.
(b) For each x ∈ X, Jx is nonempty closed convex and bounded subset of
X ∗.
(c) J(λx) = λJx for all x ∈ X and real λ, i.e., J is homogeneous.
(d) J is multivalued monotone, i.e., x − y, jx − jy  ≥ 0 for all x, y ∈ X,
jx ∈ Jx and jy ∈ J(y).
(e) x2 − y2 ≥ 2x − y, j for all x, y ∈ X and j ∈ Jy.
(f ) If X ∗ is strictly convex, J is single-valued.
(g) If X is strictly convex, then J is one-one, i.e., x = y ⇒ Jx ∩ Jy = ∅.
(h) If X is reflexive with strictly convex dual X ∗ , then J is demicontinuous.
(i) If X is uniformly convex, then for x, y ∈ Br [0], jx ∈ Jx, jy ∈ Jy

x − y, jx − jy  ≥ wr (x − y)x − y,

where wr : R+ → R+ is a function satisfies the conditions:

wr (0) = 0, wr (t) > 0 for all t > 0 and t ≤ s ⇒ wr (t) ≤ wr (s).

Proof. (a) It is obvious.

(b) If x = 0, we are done by Part(a). If x is a nonzero element in X, then
by the Hahn-Banach theorem (see Corollary 1.6.6), there exists f ∈ X ∗ such
that x, f  = x and f ∗ = 1. Set j := xf. Then x, j = xx, f  =
x2 and j∗ = x, and it follows that Jx is nonempty for each x = 0.
Now suppose f1 , f2 ∈ Jx and t ∈ (0, 1). Because

x, f1  = xf1 ∗ , x = f1 ∗

70 2. Convexity, Smoothness, Duality Mappings

and
x, f2  = xf2 ∗ , x = f2 ∗ ,
we obtain
x, tf1 + (1 − t)f2  = x(tf1 ∗ + (1 − t)f2 ∗ ) = x2 .
Observe that
x, tf1 + (1 − t)f2  ≤ tf1 + (1 − t)f2 ∗ x
≤ (tf1 ∗ + (1 − t)f2 ∗ )x
= x2 .
Then
x2 ≤ xtf1 + (1 − t)f2 ∗ ≤ x2 ,
which gives us
x2 = xtf1 + (1 − t)f2 ∗ ,
i.e.,
tf1 + (1 − t)f2 ∗ = x.
Thus,
x, tf1 + (1 − t)f2  = xtf1 + (1 − t)f2 ∗ and x = tf1 + (1 − t)f2 ∗ ,
and this means that tf1 + (1 − t)f2 ∈ Jx, i.e., Jx is a convex set.
Similarly, one can show that Jx is a closed and bounded set in X ∗ .
(c) For λ = 0, it is obvious that J(0x) = 0Jx. Assume that j ∈ J(λx) for
λ = 0. First, we show that J(λx) ⊆ λJx. Because j ∈ J(λx), we have
λx, j = λxj∗ and λx = j∗ ,
and it follows that λx, j = j2∗ . Hence
x, λ−1 j = λ−1 λx, λ−1 j = λ−2 λx, j = λ−2 λxj∗ = λ−1 j2∗ = x2 .
This shows that λ−1 j ∈ Jx, i.e., j ∈ λJx. Thus, we have J(λx) ⊆ λJx.
Similarly, one can show that λJx ⊆ J(λx). Therefore, J(λx) = λJx.
(d) Let jx ∈ Jx and jy ∈ Jy for x, y ∈ X. Hence
x − y, jx − jy  = x, jx  − x, jy  − y, jx  + y, jy 
≥ x2 + y2 − xjy ∗ − yjx ∗
≥ x2 + y2 − 2xy
= (x − y)2 ≥ 0. (2.9)

(e) Let j ∈ Jx, x ∈ X. Then

y2 − x2 − 2y − x, j
= x2 + y2 − 2y, j
≥ x2 + y2 − 2xy
= (x − y)2 ≥ 0. (2.10)
2.4. Duality mappings 71

(f) Let j1 , j2 ∈ Jx for x ∈ X. Then

x, j1  = j1 2∗ = x2

and
x, j2  = j2 2∗ = x2 .
Adding the above identities, we have

x, j1 + j2  = 2x2 .

Because
2x2 = x, j1 + j2  ≤ xj1 + j2 ∗ ,
this implies that
j1 ∗ + j2 ∗ = 2x ≤ j1 + j2 ∗ .
It now follows from the fact j1 + j2 ∗ ≤ j1 ∗ + j2 ∗ that

j1 + j2 ∗ = j1 ∗ + j2 ∗ .

Because X ∗ is strictly convex and j1 + j2 ∗ = j1 ∗ + j2 ∗ , then there exists
λ ∈ R such that j1 = λj2 . Because

x, j2  = x, j1  = x, λj2  = λx, j2 ,

this implies that λ = 1 and hence j1 = j2 . Therefore, J is single-valued.

(g) Suppose that j ∈ Jx ∩ Jy for x, y ∈ X. Because j ∈ Jx and j ∈ Jy, it
follows from j2∗ = x2 = y2 = x, j = y, j that

x2 = (x + y)/2, j ≤ (x + y)/2x,

which gives that

x = y ≤ (x + y)/2 ≤ x.
Hence x = y = (x + y)/2. Because X is strictly convex and x = y =
(x + y)/2, we have x = y. Therefore, J is one-one.
(h) It suffices to prove demicontinuity of J on the unit sphere SX . For this,
let {xn } be a sequence in SX such that xn → z in X. Then Jxn ∗ = xn  = 1
for all n ∈ N, i.e., {Jxn } is bounded. Because X is reflexive and hence X ∗ is
also reflexive. Then there exists a subsequence {Jxnk } of {Jxn } in X ∗ such that
{Jxnk } converges weakly to some j in X ∗ . Because xnk → z and Jxnk  j,
then we have

z, j = lim xnk , Jxnk  = lim xnk 2 = 1.

k→∞ k→∞

Moreover,

j∗ ≤ lim Jxnk ∗ = lim (Jxnk ∗ xnk )

k→∞ k→∞
= lim xnk , Jxnk  = z, j = j∗ .
k→∞
72 2. Convexity, Smoothness, Duality Mappings

This shows that

z, j = j∗ z and j∗ = z.
This implies that j = Jz. Thus, every subsequence {Jxni } converging weakly
to j ∈ X ∗ . This gives Jxn  Jz. Therefore, J is demicontinuous.
(i) Let r > 0 and wr : R+ → R+ a function defined by
⎧
⎪
⎪
wr (0) = 0;
⎨ x−y,jx −jy
wr (t) = inf{ x−y : x, y ∈ Br [0], x − y ≥ t, jx ∈ Jx, jy ∈ Jy}
⎪
⎪ if t ∈ (0, 2r];
⎩
wr (t) = wr (2r); if t ∈ (2r, ∞).
By (d), we have
x − y, jx − jy  ≥ 0,
and it follows that wr (t) ≥ 0 for all t ∈ R+ . It can be readily seen that wr is
nondecreasing. So it remains to prove that wr (t) > 0 for all t > 0.
Suppose, for contradiction, that there exists λ ∈ (0, 2r] such that wr (λ) = 0.
Then there exist sequences {xn }, {yn } in Br [0] such that
xn − yn  ≥ λ > 0 and xn − yn , jxn − jyn  → 0,
where jxn ∈ Jxn , jyn ∈ Jyn . We know from (2.9) that
(xn  − yn )2 ≤ xn − yn , jxn − jyn .
We may assume that
lim xn  = lim yn  = a > 0 (say).
n→∞ n→∞

Notice
xn + yn , jxn + jyn  = 2xn 2 + 2yn 2 − xn − yn , jxn − jyn 
→ 4a2 (2.11)
and
lim sup xn + yn  ≤ lim sup(xn  + yn ) = 2a.
n→∞ n→∞
Moreover, from (2.11), we have
4a2 = lim xn + yn , jxn + jyn 
n→∞
≤ lim inf xn + yn (xn  + yn ) = 2a lim inf xn + yn ,
n→∞ n→∞

which implies that

2a ≤ lim inf xn + yn .
n→∞
Thus, we have that lim xn + yn  = 2a. By the uniform convexity of X (see
n→∞
Theorem 2.3.13), we obtain that lim xn − yn  = 0, which contradicts our
n→∞
assumption that xn − yn  ≥ λ > 0.
The inequalities given in the following results are very useful in many
applications.
2.4. Duality mappings 73

∗
Proposition 2.4.6 Let X be a Banach space and J : X → 2X the duality
mapping. Then we have the following:
(a) x + y2 ≥ x2 + 2y, jx  for all x, y ∈ X, where jx ∈ Jx.
(b) x + y2 ≤ y2 + 2x, jx+y  for all x, y ∈ X, where jx+y ∈ J(x + y).
Proof. (a) Replacing y by x + y in (2.10), we get the inequality.
(b) Replacing x by x + y in (2.10), we get the result.
∗
Proposition 2.4.7 Let X be a Banach and J : X → 2X a normalized duality
mapping. Then for x, y ∈ X, the following are equivalent:
(a) x ≤ x + ty for all t > 0.
(b) There exists j ∈ Jx such that y, j ≥ 0.
Proof. (a) ⇒ (b). For t > 0, let ft ∈ J(x + ty) and define gt = ft /ft ∗ . Hence
gt ∗ = 1. Because gt ∈ ft −1
∗ J(x + ty), it follows that

x ≤ x + ty = ft −1∗ x + ty, ft 

= x + ty, gt  = x, gt  + ty, gt 
≤ x + ty, gt . (as gt ∗ = 1)
By the Banach-Alaoglu theorem (which states that the unit ball in X ∗ is
weak*ly-compact), the net {gt } has a limit point g ∈ X ∗ such that
g∗ ≤ 1, x, g ≥ x and y, g ≥ 0.
Observe that
x ≤ x, g ≤ xg∗ = x,
which gives that
x, g = x and g∗ = 1.
Set j = gx, then j ∈ Jx and y, j ≥ 0.
(b) ⇒ (a). Suppose for x, y ∈ X with x = 0 there exists j ∈ Jx such that
y, j ≥ 0. Hence for t > 0,
x2 = x, j ≤ x, j + ty, j = x + ty, j ≤ x + tyx,
which implies that
x ≤ x + ty.
Observation
• Dom(J) = X.
• J is odd, i.e., J(−x) = −Jx.
• J is homogeneous (hence J is positive homogeneous, i.e., J(λx) = λJx for all
λ > 0).
• J is bounded.

We now consider the duality mappings that are more general than the
normalized duality mappings.
74 2. Convexity, Smoothness, Duality Mappings

Definition 2.4.8 A continuous strictly increasing function μ : R+ → R+ is

said to be gauge function if μ(0) = 0 and lim μ(t) = ∞.
t→∞

Definition 2.4.9 Let X be a normed space and μ a gauge function. Then the
∗
mapping Jμ : X → 2X defined by

Jμ (x) = {j ∈ X ∗ : x, j = xj∗ , j∗ = μ(x)}, x∈X

is called the duality mapping with gauge function μ.

In the particular case μ(t) = t, the duality mapping Jμ = J is called the

normalized duality mapping .

In the case μ(t) = tp−1 , p > 1, the duality mapping Jμ = Jp is called the
generalized duality mapping and it is given by

Jp (x) := {j ∈ X ∗ : x, j = xj∗ , j∗ = xp−1 }, x ∈ X.

Note that if p = 2, then Jp = J2 = J is the normalized duality mapping.

Remark 2.4.10 For the gauge function μ, the function Φ : R+ → R+
defined by
t
Φ(t) = μ(s)ds
0

is a continuous convex strictly increasing function on R+ . Therefore, Φ has a

continuous inverse function Φ−1 .

Example 2.4.11 Let x = (x1 , x2 , · · · ) ∈ p (1 < p < ∞), set

Jμ (x) = (|x1 |p−1 sgn(x1 ), |x2 |p−1 sgn(x2 ), · · · )

and let μ(t) = tp−1 = tp/q , where 1/p + 1/q = 1. Observe that

∞ 1/q 
∞ 1/q
|xi |
(p−1)q
= |xi | p
and Jμ (x) ∈ q.
i=1 i=1

Moreover,
μ(x) = xp/q = Jμ (x)∗
and
∞
 ∞

x, Jμ (x) = xi |xi |p−1 sgn(xi ) = |xi |p = xp
i=1 i=1
= xxp−1 = xμ(x) = xJμ (x)∗ .

Thus, Jμ is a duality mapping with gauge function μ. Therefore, the generalized

duality mapping Jp in p space is given by

Jp (x) = (|x1 |p−1 sgn(x1 ), |x2 |p−1 sgn(x2 ), · · · ), x∈ p.

2.4. Duality mappings 75

One can easily see the following facts:

(i) Jμ (x) is a nonempty closed convex set in X ∗ for each x ∈ X,

(ii) Jμ is a function when X ∗ is strictly convex.

(iii) If Jμ (x) is single-valued, then

sign(λ)μ(λx)
Jμ (λx) = Jμ (x) for all x ∈ X and λ ∈ R
μ(x)

and

x − y, Jμ (x) − Jμ (y) ≥ (μ(x) − μ(y))(x − y) for all x, y ∈ X.

We now give other interesting properties of the duality mappings Jμ in

reflexive Banach spaces.

Theorem 2.4.12 Let X be a Banach space and J μ a duality mapping with

gauge function μ. Then X is reflexive if and only if x∈X Jμ (x) = X ∗ , i.e., Jμ
is onto.

Proof. Let X be reflexive and let j ∈ X ∗ . By the Hahn-Banach theorem, there

is an x ∈ SX such that x, j = x.
Because μ has the property of Darboux, there exists a constant t ≥ 0 such
that
μ(||tx||) = μ(t) = j∗ .
Because tx, j = txj∗ , it follows that j ∈ Jμ (tx).
Conversely, suppose that for each j ∈ X ∗ , there is x ∈ X such that j ∈ Jμ (x).
Set y := x/x. Then y = 1 and y, j = j∗ . Hence each continuous
functional attains its supremum on the unit ball. By the James theorem, X is
reflexive.

Theorem 2.4.13 Let X be a reflexive Banach space and J a duality mapping

with gauge function μ. Then J −1 is the duality mapping with gauge μ−1 .

Proof. From Theorem 2.4.12, we obtain

J −1 (j) = {x ∈ X : j ∈ Jμ (x)} = ∅ for all j ∈ X ∗ .

Let J ∗ be the duality mapping on X ∗ with gauge μ−1 . Observe that x ∈ J −1 (j)
if and only if x, j = xj∗ and x = μ−1 (j∗ ) or equivalently if and only
if x ∈ J ∗ (j). Thus,

J ∗ (j) = J −1 (j) = {x ∈ X : x, j = xj∗ , x = μ−1 (j∗ )}.

76 2. Convexity, Smoothness, Duality Mappings

Corollary 2.4.14 Let X be a reflexive Banach space and J ∗ : X ∗ → X the

inverse of the normalized duality mapping J : X → X ∗ . Then
J ∗ J = I and JJ ∗ = I ∗ (identity mappings on X and X ∗ , respectively).
Theorem 2.4.15 Let X be a Banach space and let Jμ be the duality mapping
with gauge function μ. If X ∗ is uniformly convex, then Jμ is uniformly continu-
ous on each bounded set in X, i.e., for ε > 0 and K > 0, there is a δ > 0 such
that
x ≤ K, y ≤ K and x − y < δ ⇒ Jμ (x) − Jμ (y)∗ < ε.
Proof. Because X ∗ is strictly convex, Jμ is single-valued. Suppose {xn } and
{yn } are sequences in X such that xn  ≤ K, yn  ≤ K and xn − yn  → 0.
Assume that xn → 0, then yn → 0. Moreover,
Jμ (xn )∗ = μ(xn ) → 0 and Jμ (yn )∗ = μ(yn ) → 0.
Hence Jμ (xn ) − Jμ (yn )∗ → 0 and we are done.
Suppose {xn } does not converge strongly to zero. There exist α > 0 and a
subsequence {xnk } of {xn } such that xnk  ≥ α. Because xn − yn  → 0, one
can assume that ynk  ≥ α/2. Without loss of generality, we may assume that
xn  ≥ β and yn  ≥ β for some β > 0.
Set un := xn /xn  and vn := yn /yn  so that un  = vn  = 1 and
 
 xn yn  − xn yn 
un − vn  =   
xn yn  
 
1 

≤ 2 xn yn  − xn xn  + xn xn  − xn yn 

β
 
1
≤ yn  − xn  xn  + xn xn − yn 
β2
1
≤ (yn − xn K + xn − yn K) → 0 as n → ∞.
β2
Because Jμ (un )∗ = μ(un ) = μ(1) and Jμ (vn )∗ = μ(vn ) = μ(1), we have
μ(1) + μ(1) − μ(1)un − vn  ≤ un , Jμ (un )+vn , Jμ (vn )+un − vn , Jμ (vn )
= un , Jμ (un )+un , Jμ (vn )
= un , Jμ (un )+Jμ (vn )
≤ Jμ (un ) + Jμ (vn )∗ ≤ 2μ(1).
This shows that lim Jμ (un ) + Jμ (vn )∗ = 2μ(1). Because X ∗ is uniformly
n→∞
convex, we have Jμ (un ) − Jμ (vn )∗ → 0 as n → ∞. Hence
Jμ (xn ) − Jμ (yn )
= [μ(xn )(Jμ (un ) − Jμ (vn )) + (μ(xn ) − μ(yn ))Jμ (vn )]/μ(1),

and it follows that Jμ (xn ) − Jμ (yn )∗ → 0 as n → ∞.

2.4. Duality mappings 77

Observation
∗
• If Jμ : X → 2X is a duality mapping with gauge function μ then
(i) Jμ is norm to weak* upper semicontinuous.
(ii) for each x ∈ X, the set Jμ (x) is convex and weakly closed in X ∗ ;
(iii) Jμ (−x) = −Jμ (x) and Jμ (λx) = μ(λx)
μ(x) μ
J (x) for all x ∈ X, λ > 0;
(iv) each selection of Jμ is a homogeneous single-valued mapping j : X → X ∗
satisfying j(x) ∈ Jμ (x) for all x ∈ X,
(v) Jμ is monotone, i.e., x − y, jx − jy  ≥ 0 for all x, y ∈ X and jx ∈ Jμ (x),
jy ∈ Jμ (y);
(vi) the strict convexity of X implies that Jμ is strictly monotone, i.e.,

x − y, jx − jy  > 0 for all x, y ∈ X and jx ∈ Jμ (x), jy ∈ Jμ (y);

(vii) the reflexivity of X and strict convexity of X ∗ imply that Jμ is single-valued

monotone and demicontinuous.

One can easily see that the following are reflexive Kadec-Klee Banach spaces:
(a) a Banach space of finite-dimension,
(b) a reflexive Banach space that is locally uniformly convex,
(c) a uniformly convex Banach space.
We now conclude this section with an interesting result concerning a Banach
space whose dual has the Kadec-Klee property.
Theorem 2.4.16 Let X be a reflexive Banach space such that X ∗ has the
Kadec-Klee property. Let {xα}α∈D be a bounded net in X and x, y ∈ ww ({xα}α∈D ).
Suppose lim txα + (1 − t)x − y exists for all t ∈ [0, 1]. Then x = y.
α∈D

Proof. Because lim txα + (1 − t)x − y exists (say, r), for each ε > 0, there
α∈D
exists α0 ∈ D such that

txα + (1 − t)x − y ≤ r + ε for all α  α0 .

It follows that for all α  α0 and j(x − y) ∈ J(x − y),

txα + (1 − t)x − y, j(x − y) ≤ (r + ε)x − y.

Because x ∈ ωw ({xα }α∈D ), we obtain

x − y2 = tx + (1 − t)x − y, j(x − y)

≤ x − y( lim txα + (1 − t)x − y + ε),
α∈D
= (r + ε)x − y.

Taking the limit as ε → 0, we obtain

x − y ≤ r. (2.12)
78 2. Convexity, Smoothness, Duality Mappings

By Proposition 2.4.6 (b), we have

txα + (1 − t)x − y2 ≤ x − y2 + 2txα − x, j(txα + (1 − t)x − y)

for all t ∈ (0, 1] and j(txα + (1 − t)x − y) ∈ J(txα + (1 − t)x − y). By (2.12), we
have
lim inf xα − x, j(txα + (1 − t)x − y) ≥ 0.
α∈D

Hence there exists a sequence {αn }n∈N such that αn  αm for n ≥ m and
    
1 1 1
xα − x, j xα + 1 − x−y ≥ − for all n ∈ N and α  αn . (2.13)
n n n

Set D1 = {α : α  α1 }. Without loss of generality, we may assume that

D = D1 ,
ωw ({xα }α∈D ) = ωw {xα }α∈D1
and

lim txα + (1 − t)x − y = lim txα + (1 − t)x − y for all t ∈ [0, 1].
α∈D α∈D1

Set tα = inf{1/n : α  αn } for all α ∈ D.

We now consider two cases:
Case 1. α ∈ D and tα > 0.
Set jα := j(tα xα + (1 − tα )x − y). Then

x − y, jα  = tα xα + (1 − tα )x − y2 − tα xα − x, jα  (2.14)

and

jα  = tα xα + (1 − tα )x − y. (2.15)

By (2.13), we have

xα − x, jα  ≥ −tα . (2.16)

Case 2. α ∈ D and tα = 0.
In this case, we can choose a subsequence {j((1/nk )xα +(1−1/nk )x−y)}k∈N
which is weakly convergent to j, and set jα := j. It follows from (2.13) that

xα − x, jα  ≥ 0. (2.17)

Observe that
    
 1 1 
jα  ≤ lim inf 
j xα + 1 − x − y 

k→∞ nk nk
   
 1 1 
= lim   xα + 1 − x − y = x − y.
k→∞ nk nk
2.5. Convex functions 79

On the other hand, we have

    
1 1
x − y, jα  = limx − y, j xα + 1 − x−y
k→∞ nk nk
  
1 1
= lim  xα + 1 − x − y2
k→∞ nk nk
    
1 1 1
− xα − x, j xα + 1 − x−y
nk nk nk
= x − y2 . (2.18)

Therefore,

jα  = x − y (2.19)

and jα ∈ J(x − y).

We note that by the Kadec-Klee property of X ∗ , the sequence {j((1/nk )xα +
(1 − 1/nk )x − y)}k∈N converges strongly to jα .
Now from the net {xα }α∈D , we choose a subset {αβ }β∈D such that {xαβ }β∈D
converges weakly to y ∈ ww ({xα }α∈D ) and {jαβ }β∈D converges weakly to j.
Then by (2.15) and (2.19) we get

j∗ ≤ x − y

and by (2.14) and (2.18), we get

x − y, j = x − y2 .

Hence j ∈ J(x − y). Because X is reflexive and X ∗ has the Kadec-Klee pro-
perty, the space X ∗ has also the Kadec property and this implies that {jαβ }β∈D
converges strongly to j. It follows from (2.16) and (2.17) that

y − x, j ≥ 0,

i.e., x − y2 ≤ 0. Therefore, x = y.

Corollary 2.4.17 Let X be a reflexive Banach space such that its dual X ∗
has the Kadec-Klee property. Let {xn } be a bounded sequence in X and p, q ∈
ωw ({xn }). Suppose lim txn + (1 − t)p − q exists for all t ∈ [0, 1]. Then p = q.
n→∞

2.5 Convex functions

Let X be a linear space and f : X → (−∞, ∞] a function. Then
(i) f is said to be convex if f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) for
all x, y ∈ X and λ ∈ [0, 1];
(ii) f is said to be strictly convex if f (λx+(1−λ)y) < λf (x)+(1−λ)f (y) for
all λ ∈ (0, 1) and x, y ∈ X with x = y, f (x) < ∞, f (y) < ∞;
80 2. Convexity, Smoothness, Duality Mappings

(iii) f is said to be proper if there exists x ∈ X such that f (x) < ∞;

(iv) Dom(f ) = {x ∈ X : f (x) < ∞} is called domain or effective domain;
(v) f is said to be bounded below if there exists a real number α such that
α ≤ f (x) for all x ∈ X;
(vi) the set epif = {(x, α) : x ∈ X, α ∈ R, f (x) ≤ α} is called the
epigraph of f .

Let C be a subset of X. Then the function iC on X defined by


0 if x ∈ C,
iC (x) =
∞ if x ∈ /C
is called the indicator function.

Observation
• iC is proper if and only if C is nonempty.
• dom(iC ) = C.
• The set C is convex if and only if its indicator function iC is convex.
• The domain of each convex function is convex.

Let X be a topological space and f : X → (−∞, ∞] a proper function. Then

f is said to be lower semicontinuous (l.s.c.) at x0 ∈ X if
f (x0 ) ≤ lim inf f (x0 ) = sup inf f (x),
x→x0 V ∈Ux0 x∈V

where Ux0 is a base of neighborhoods of the point x0 ∈ X. f is said to be lower

semicontinuous on X if it is lower semicontinuous on each point of X, i.e., for
each x ∈ X
xn → x ⇒ f (x) ≤ lim inf f (xn ).
n→∞
We now discuss some elementary properties of convex functions:
Proposition 2.5.1 Let X be a linear space and f : X → (−∞, ∞] a function.
Then f is convex if and only if its epigraph is a convex subset of X × R.
Proof. Suppose f is convex. Then for (x, α), (y, β) in epif , we have
f ((1 − t)x + ty) ≤ (1 − t)f (x) + tf (y) ≤ (1 − t)α + tβ for all t ∈ [0, 1].
This implies that ((1 − t)x + ty, (1 − t)α + tβ) ∈ epif .
Conversely, suppose that epif is convex. Then Dom(f ) is also convex.
Because for x, y ∈ Dom(f ) and (x, f (x)), (y, f (y)) ∈ epif , we have
((1 − t)x + ty, (1 − t)f (x) + tf (y)) ∈ epif for all t ∈ [0, 1].
Thus, by the definition of epif ,

f ((1 − t)x + ty) ≤ (1 − t)f (x) + tf (y).

2.5. Convex functions 81

Proposition 2.5.2 Let X be a topological space and f : X → (−∞, ∞] a

function. Then the following statements are equivalent:
(a) f is lower semicontinuous.
(b) For each α ∈ R, the level set {x ∈ X : f (x) ≤ α} is closed.
(c) The epigraph of the function f , {(x, α) ∈ X × R : f (x) ≤ α} is closed.

Proof. We recall that

lim inf f (x) = sup inf f (x).

x→x0 V ∈Ux0 x∈V

(a) ⇒ (b). Let α ∈ R and let x0 ∈ X with f (x0 ) > α. Because f is lower
semicontinuous, there exists V0 ∈ Ux0 such that inf f (x) > α. Hence V0 ⊂
x∈V0
{x ∈ X : f (x) > α}. Consequently, {x ∈ X : f (x) > α} is open and hence
{x ∈ X : f (x) ≤ α} is closed.
(b) ⇒ (a). Let x0 ∈ Dom(f ), ε > 0 and Vε = {x ∈ X : f (x) > f (x0 ) − ε}.
Because each level set of f is closed, it follows that Vε ∈ U (x0 ). Because
inf f (x) ≥ f (x0 ) − ε, it follows that lim inf f (x) ≥ f (x0 ) − ε. As ε is arbitrarily
x∈Vε x→x0
chosen, we conclude that (a) holds.
(a) ⇔ (c). Define ϕ : X × R → (−∞, ∞] by ϕ(x, α) = f (x) − α. Then, f is
l.s.c. on X ⇔ ϕ is l.s.c. on X × R. Because epif is a level set of ϕ, therefore,
the conclusion holds.

Proposition 2.5.3 Let C be a nonempty closed convex subset of a Banach

space X and f : C → (−∞, ∞] a convex function. Then f is lower semicontin-
uous in the norm topology if and only if f is lower semicontinuous in the weak
topology.

Proof. Set Fα := {x ∈ C : f (x) ≤ α}, α ∈ R. Then Fα is convex. Indeed, for

x, y ∈ Fα

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

≤ λα + (1 − λ)α = a for all λ ∈ [0, 1].

It follows from Proposition 1.9.13 (which states that for a convex subset C in a
normed space X, C is closed if and only if C is weakly closed) that Fα is closed
if and only if Fα is weakly closed, i.e., Fα is closed in the weak topology.

Before presenting an important result, we first establish a preliminary result:

Theorem 2.5.4 Let X be a compact topological space and f : X → (−∞, ∞] a

lower semicontinuous function. Then there exists an element x0 ∈ X such that

f (x0 ) = inf{f (x) : x ∈ X}.

82 2. Convexity, Smoothness, Duality Mappings

Proof. Set Gα := {x ∈ X : f (x) > α}, α ∈ R. One may easily see that each Gα
is open and X = α∈R Gα . By compactness of X, there exists a finite family
{Gαi }ni=1 of {Gα }α∈R such that

n
X= Gαi .
i=1

Suppose α0 = min{α1 , α2 , · · · , αn }. This gives f (x) > α0 for all x ∈ X.

It follows that inf{f (x) : x ∈ X} exists. Let m = inf{f (x) : x ∈ X}.
Let β be a number such that β > m. Set Fβ := {x ∈ X : f (x) ≤ β}. Then Fβ
is a nonempty closed subset of X; and hence, by the intersection property, we
have 
Fβ = ∅.
β>m

Therefore, for any point x0 of this intersection, we have m = f (x0 ).

Theorem 2.5.5 Let C be a weakly compact convex subset of a Banach space
and f : C → (−∞, ∞] a proper lower semicontinuous convex function. Then
there exists x0 ∈ Dom(f ) such that f (x0 ) = inf{f (x) : x ∈ C}.
Proof. Because f is proper, there exists u ∈ C such that f (u) < ∞. Then
the set C0 = {x ∈ C : f (x) ≤ f (u)} is nonempty. Because the set C0 is
closed and convex subset of C, it follows that C0 is weakly compact. Applying
Proposition 2.5.3, we have that f is lower semicontinuous in the weak topology.
By Theorem 2.5.4, there exists x0 ∈ C0 ⊂ C such that
f (x0 ) = inf{f (x) : x ∈ C0 } = inf{f (x) : x ∈ C}.
Remark 2.5.6 If f is strictly convex function in Theorem 2.5.5, then x0 ∈ C
is the unique point such that f (x0 ) = inf f (x).
x∈C

Recall that every closed convex bounded subset of a reflexive Banach space
is weakly compact. Using this fact, we have
Theorem 2.5.7 Let X be a reflexive Banach space and f : X → (−∞, ∞] a
proper lower semicontinuous convex function. Then for every nonempty closed
convex bounded subset C of X, there exists a point x0 ∈ Dom(f ) such that
f (x0 ) = inf f (x).
x∈C

In Theorem 2.5.7, the boundedness of C may be replaced by the weaker

assumption
lim f (x) = ∞.
x∈C,x→∞

Theorem 2.5.8 Let C be a nonempty closed convex subset of a reflexive Banach

space X and f : C → (−∞, ∞] a proper lower semicontinuous convex function
such that f (xn ) → ∞ as xn  → ∞. Then there exists x0 ∈ Dom(f ) such that
f (x0 ) = inf{f (x) : x ∈ C}.
2.5. Convex functions 83

Proof. Let m = inf{f (x) : x ∈ C}. Choose a minimizing sequence {xn } in

C, i.e., f (xn ) → m. If {xn } is not bounded, there exists a subsequence {xni }
of {xn } such that xni  → ∞. From the hypothesis, we have f (xni ) → ∞,
which contradicts m = ∞. Hence {xn } is bounded. By the reflexivity X, there
exists a subsequence {xnj } of {xn } such that xnj  x0 ∈ C. Because f is lower
semicontinuous in the weak topology, we have
m ≤ f (x0 ) ≤ lim inf f (xnj ) = lim f (xn ) = m.
j→∞ n→∞

Therefore, f (x0 ) = m.

Differentiation of convex functions – Let X be a normed space and

ϕ : X → (−∞, ∞] a function. Then the limit
ϕ(x + ty) − ϕ(x) ϕ(x + ty) − ϕ(x)
lim = inf
t→0 t t>0 t
is said to be the directional derivative of ϕ at the point x ∈ X in the direction
y ∈ X. If it exists, it is denoted by ϕ (x, y).
The function ϕ is said to be Gâteaux differentiable at a point x ∈ X if there
exists a continuous linear functional j on X such that y, j = ϕ (x, y) for all
y ∈ X. The element j, denoted by ϕ (x) or ϕ(x) (i.e., gradϕ(x)) is called the
Gâteaux derivative of ϕ at x.

One can easily see from the definition of Gâteaux derivative of ϕ that
(i) ϕ (x)(0) = 0,
ϕ(x + tλy) − ϕ(x)
(ii) ϕ (x)(λy) = λ lim = λϕ (x)(y) for all λ ∈ R, i.e.,
t→0 t
ϕ (x)(·) is homogeneous over R.
Remark 2.5.9 If the function ϕ is Gâteaux differentiable at x ∈ X, then there
exists j = ϕ (x) ∈ X ∗ such that
d
ϕ(x + ty) = y, ϕ (x) = y, j for all y ∈ X.
dt t=0

Let X be a normed space and ϕ : X → (−∞, ∞] a function. The function ϕ

is said to be Fréchet differentiable at a point x ∈ X if there exists a continuous
linear functional j on X such that
|ϕ(x + y) − ϕ(x) − y, j|
lim = 0.
y→0 y
In this case, the element j denoted by dϕ(x) is called the Fréchet derivative
of ϕ at the point x.
Proposition 2.5.10 Let X be a normed space and ϕ : X → (−∞, ∞] a
function. If ϕ is Fréchet differentiable at x, then ϕ is Gâteaux differentiable
at x.
84 2. Convexity, Smoothness, Duality Mappings

Proof. Because ϕ is Fréchet differentiable at x,

|ϕ(x + y) − ϕ(x) − dϕ(x)y|

lim = 0. (2.20)
y→0 y

Set y = ty0 for t > 0 and for any fixed y0 = 0. From (2.20), we obtain

|ϕ(x + ty0 ) − ϕ(x) − tdϕ(x)y0 |

lim = 0,
t→0 ty0 

which implies that

ϕ(x + ty0 ) − ϕ(x)

lim = dϕ(x)y0 .
t→0 t

Hence dϕ ∈ X ∗ and ϕ is Gâteaux differentiable at x.

The following example shows that the converse of Proposition 2.5.10 is not
true.

Example 2.5.11 Let X = R2 be a normed space with norm ·2 and ϕ : X → R

a function defined by
 3
x y/(x4 + y 2 ) if (x, y) = (0, 0),
ϕ(x, y) =
0 if (x, y) = (0, 0).

One may easily see that ϕ is Gâteaux differentiable at 0 with Gâteaux derivative
ϕ (0) = 0. Because for (h, k) ∈ X, we have

|ϕ(h, k)| |h3 k| 1

= = for k = h2 .
(h, k)2 4 2 2 2
(h + k )(h + k ) 1/2 2(1 + h2 )1/2

Therefore, ϕ is not Fréchet differentiable.

Observation
• Every Fréchet differentiable function is Gâteaux differentiable.

• If ϕ is Fréchet differentiable at x, then ϕ is continuous at x.

• If ϕ is Gâteaux differentiable at x, then ϕ is not necessarily continuous at x

(e.g., the function ϕ : R2 → R defined by

2y exp(−x−2 )
ϕ(x, y) = , x = 0 and ϕ(x, y) = 0, x = 0
y 2 + exp(−2x−2 )

is Gâteaux differentiable at zero, but not continuous at zero).

• If ϕ is Gâteaux differentiable at x, then ϕ(x + ty) → ϕ(x) as t → 0 (i.e., if

xn → x along a line, then ϕ(xn ) → ϕ(x)).
2.5. Convex functions 85

Let X be a Banach space and ϕ : X → (−∞, ∞] a proper convex function.

Then an element j ∈ X ∗ is said to be a subgradient of ϕ at the point x ∈ X if

ϕ(x) − ϕ(y) ≤ x − y, j for all y ∈ X.

The set (possibly nonempty)

{j ∈ X ∗ : ϕ(x) − ϕ(y) ≤ x − y, j for all y ∈ X},

of subgradients of ϕ at x ∈ X is called the subdifferential of ϕ at x ∈ X.

Thus, the subdifferential of a proper convex function ϕ is a mapping ∂ϕ : X →
∗
2X (generally multivalued) defined by

∂ϕ(x) = {j ∈ X ∗ : ϕ(x) − ϕ(y) ≤ x − y, j for all y ∈ X}.

The domain of the subdifferential ∂ϕ is denoted and defined by

Dom(∂ϕ) = {x ∈ X : ∂ϕ(x) = ∅}.

Remark 2.5.12 If ϕ is not the constant ∞, then Dom(∂ϕ) is a subset of

Dom(ϕ).
Observation
• ∂ϕ(x) is always for every x ∈ X nonempty if ϕ is continuous.
• ∂ϕ(x) is always a closed convex set in X ∗ .
• ∂(λϕ(x)) = λ∂ϕ(x), i.e., ∂ϕ(x) is homogeneous.
• ϕ has a minimum value at x0 ∈ Dom(∂ϕ) if and only if 0 ∈ ∂ϕ(x0 ).
• Dom(∂ϕ) = Dom(ϕ) if ϕ is lower semicontinuous on X.
• For a lower semicontinuous proper convex function ϕ on a reflexive Banach space
X, ∂ϕ is maximal monotone.

The following results are of fundamental importance in the study of convex

functions. We begin with a basic result.
Proposition 2.5.13 Let C be a nonempty closed convex subset of a Banach
space X and iC the indicator function of C, i.e.,

0 if x ∈ C,
iC (x) =
∞ otherwise.

Then ∂iC (x) = {j ∈ X ∗ : x − y, j ≥ 0 for all y ∈ C}, x ∈ C.

Proof. Because the indicator function is convex and lower semicontinuous
function on X, by the subdifferentiability of iC , we have

∂iC (x) = {j ∈ X ∗ : iC (x) − iC (y) ≤ x − y, j for all y ∈ C}.

86 2. Convexity, Smoothness, Duality Mappings

Remark 2.5.14 Dom(iC ) = Dom(∂iC ) = C and ∂iC (x) = {0} for each x ∈
int(C).

We now give a relation between Gâteaux differentiability and subdiffer-

entiability.

Theorem 2.5.15 Let X be a Banach space and ϕ : X → (−∞, ∞] a proper

convex function. If ϕ is Gâteaux differentiable at a point x0 ∈ X, then ∂ϕ(x0 ) =
{ϕ (x0 )}, i.e., the subdifferential of ϕ at x0 ∈ X is a singleton set {ϕ (x0 )} in
X ∗.
Conversely, if ϕ is continuous at x0 and ∂ϕ(x0 ) contains a singleton element,
then ϕ is Gâteaux differentiable at x0 and ϕ (x0 ) = ∂ϕ(x0 ).

Proof. Let ϕ be Gâteaux differentiable at x0 ∈ X. Then

ϕ(x0 + ty) − ϕ(x0 )

y, ϕ (x0 ) = lim for all y ∈ X.
t→0 t
Notice

ϕ(x0 + λ(z − x0 )) = ϕ((1 − λ)x0 + λz) ≤ (1 − λ)ϕ(x0 ) + λϕ(z) for all λ ∈ (0, 1).

Set y := z − x0 . Then, we have

ϕ(x0 + λy) ≤ ϕ(x0 ) + λ[ϕ(x0 + y) − ϕ(x0 )].

Thus,
ϕ(x0 + λy) − ϕ(x0 )
≤ ϕ(x0 + y) − ϕ(x0 ),
λ
which implies that

ϕ(x0 ) − ϕ(x0 + y) ≤ −y, ϕ (x0 ) = x0 − (x0 + y), ϕ (x0 ) for all y ∈ X,

i.e., ϕ (x0 ) ∈ ∂ϕ(x0 ).

Now, let jx0 ∈ ∂ϕ(x0 ). Then, we have

ϕ(x0 ) − ϕ(u) ≤ x0 − u, jx0  for all u ∈ X.

Therefore,
ϕ(x0 + λh) − ϕ(x0 )
≥ h, jx0  for all λ > 0,
λ
it follows that
h, ϕ (x0 ) − jx0  ≥ 0 for all h ∈ X,
i.e., jx0 = ϕ (x0 ). Thus, ϕ is Gâteaux differentiable at x0 and ϕ (x0 ) =
∂ϕ(x0 ).
2.5. Convex functions 87

Corollary 2.5.16 Let X be a Banach space and ϕ : X → (−∞, ∞] a proper

convex function. Then ϕ is Gâteaux differentiable at x ∈ int(dom(ϕ)) if and
only if it has a unique subgradient ∂ϕ(x) = {ϕ (x)}, i.e., the subdifferential of
ϕ at x is a singleton set in X ∗ . In this case
d
ϕ(x + ty) = y, ∂ϕ(x) = y, ϕ (x) for all y ∈ X.
dt t=0
∗
Theorem 2.5.17 Let X be a Banach space, Jμ : X → 2X a duality mapping
 x
with gauge function μ, and Φ(x) = 0 μ(s)ds, 0 = x ∈ X. Then
Jμ (x) = ∂Φ(x).
Proof. Because μ is a strictly increasing and continuous function, it follows
that Φ is differentiable and hence Φ (t) = μ(t), t ≥ 0. Then Φ is a convex
function.
First, we show Jμ (x) ⊆ ∂Φ(x). Let x = 0, and j ∈ Jμ (x). Then x, j =
xj∗ , j∗ = μ(x). In order to prove j ∈ ∂Φ(x), i.e., Φ(x) − Φ(y) ≤
x − y, j for all y ∈ X, we assume that y > x. Then
Φ(y) − Φ(x)
j∗ = μ(x) = Φ (x) ≤ ,
y − x
which yields
Φ(x) − Φ(y) ≤ j∗ (x − y)
≤ x, j − y, j
= x − y, j.
In a similar way, if x > y, we have
Φ(x) − Φ(y) ≤ x − y, j.
In the case when x = y, we have
y − x, j = y, j − x, j
≤ yj∗ − xj∗ (as x, j = xj∗ )
≤ j∗ (y − x),
and it follows that
Φ(x) − Φ(y) = 0 = j∗ (x − y) ≤ x − y, j.
Hence j ∈ ∂Φ(x). Thus, Jμ (x) ⊆ ∂Φ(x) for all x = 0.
We now prove ∂Φ(x) ⊆ Jμ (x) for all x = 0. Suppose j ∈ ∂Φ(x) for
0 = x ∈ X. Then
xj∗ = sup{y, jx : y = 1}
= sup{y, j : x = y = 1}
≤ sup{y, j : x = y}
≤ sup{x, j + Φ(y) − Φ(x) : x = y}
≤ xj∗ . (as y, j ≤ x, j + Φ(y) − Φ(x)).
88 2. Convexity, Smoothness, Duality Mappings

Thus, x, j = xj∗ . To see j ∈ Jμ (x), we show that j∗ = μ(x) =

Φ (x). Because

Φ(x) − Φ(tx) ≤ x − tx, j = (1 − t)xj∗ for all t > 0,

this implies that

Φ(tx) − Φ(x)
j∗ ≤ . (2.21)
tx − x

It follows from (2.21) that

Φ(tx) − Φ(x)
j∗ ≤ if t > 1
tx − x

and
Φ(x) − Φ(tx)
≤ j∗ if t < 1.
x − tx
Taking the limit as t → 1, we get

j∗ = Φ (x) = μ(x).

Thus, ∂Φ(x) ⊆ Jμ (x). Therefore, Jμ (x) = ∂Φ(x) for all x = 0.

Remark 2.5.18 Both the sets Jμ (x) and ∂Φ(x) are equal to {0} if x = 0.

Corollary 2.5.19 For p ∈ (1, ∞), the generalized duality mapping Jp is the
subdifferential of the functional  · p /p.

Proof. Define μ(t) = tp−1 , p > 1. Hence

t t
tp
Φ(t) = μ(s)ds = sp−1 ds = .
0 0 p

Therefore, Jp (·) = ∂( · p /p).

Corollary 2.5.20 Let X be a Banach space and ϕ(x) = x2 /2. Then the
∗
subdifferential ∂ϕ coincides with the normalized duality mapping J : X → 2X
defined by

Jx = {j ∈ X ∗ : x, j = xj∗ , j∗ = x}, x ∈ X.

Theorem 2.5.21 Let X be a Banach space. Then

∂x = {j ∈ X ∗ : x, j = xj∗ , j∗ = 1} for all x ∈ X \ {0}.

2.5. Convex functions 89

Proof. Let j ∈ ∂x. Then

y − x, j ≤ y − x ≤ y − x for all y ∈ X. (2.22)

It follows that j ∈ X ∗ and j∗ ≤ 1. It is clear from (2.22) that x ≤ x, j,
which gives
x, j = x and j∗ = 1.
Thus,
∂x ⊆ {j ∈ X ∗ : x, j = x and j∗ = 1}.
Now suppose j ∈ X ∗ such that j ∈ {f ∈ X ∗ : x, f  = x and f ∗ = 1}. Then
x, j = x and j∗ = 1. Thus,

y − x, j = y, j − x ≤ y − x for all y ∈ X,

i.e., j ∈ ∂x. It follows that

{j ∈ X ∗ : x, j = x and j∗ = 1} ⊆ ∂x.

Therefore, ∂x = {j ∈ X ∗ : x, j and j∗ = 1}.

Using Corollary 2.5.19, we establish an inequality in a general Banach space

that is a generalization of the inequality given in Proposition 2.4.6(b).
∗
Theorem 2.5.22 Let X be a Banach space and let Jp : X → 2X , 1 <
p < ∞ be the generalized duality mapping. Then for any x, y ∈ X, there exists
jp (x + y) ∈ Jp (x + y) such that x + yp ≤ xp + py, jp (x + y).
p
Proof. By Corollary 2.5.19, Jp is the subdifferential of the functional  ·  /p.
p
By the subdifferentiability of  ·  /p, for x, y ∈ X, there exists jp (x + y) ∈
Jp (x + y) such that x + yp ≤ xp + py, jp (x + y).

The following result is very useful in the approximation of solution of non-

linear operator equations.
∗
Theorem 2.5.23 Let X be a Banach space and Jμ : X → 2X a duality map-
ping with gauge function μ. If Jμ is single-valued, then
1
Φ(x + y) = Φ(x) + y, Jμ (x + ty)dt for all x, y ∈ X.
0

Proof. Because Jμ is single-valued, it follows from Theorem 2.5.17 that

∂Φ(x) = {Jμ (x)}. Hence Corollary 2.5.16 implies that Jμ is the Gâteaux
gradient of Φ(x), i.e.,

d
Φ(x + ty) = y, Jμ (x).
dt t=0
90 2. Convexity, Smoothness, Duality Mappings

Hence
d d
Φ(x + ty) = Φ(x + ry + sy) = y, Jμ (x + ry), r ∈ R.
dt t=r ds s=0

Because the function t → y, Jμ (x + ty) is continuous, hence

1 1
d
y, Jμ (x + ry)dr = Φ(x + ty) dr = Φ(x + y) − Φ(x).
0 0 dt t=r

Corollary 2.5.24 Let X be a Banach space. If X ∗ is strictly convex, then we

have the following:
1
(a) Φ(x + y) = Φ(x) + y, Jμ (x + ty)dt for all x, y ∈ X;
0
1
(b) x + yp = xp + p y, Jp (x + ty)dt for all x, y ∈ X and p > 1;
0
1
(c) x + y2 = x2 + 2 y, J(x + ty)dt for all x, y ∈ X.
0

Proposition 2.5.25 Let X be a Banach space with strictly convex dual and C
a nonempty convex subset of X. Let x0 be an element in C and Jμ : X → X ∗
a duality mapping with gauge function μ. Then
x0  = inf x if and only if x0 − x, Jμ (x0 ) ≤ 0 f or all x ∈ C.
x∈C

Proof. Let x0 be a point in C such that x0 − x, Jμ (x0 ) ≤ 0 for all x ∈ C.

Then
x0 Jμ (x0 )∗ = x0 , Jμ (x0 ) ≤ xJμ (x0 )∗ for all x ∈ C.
Therefore, x0  = inf x.
x∈C
Conversely, suppose that x0 ∈ C such that x0  = inf x. Then
x∈C

x0  ≤ x0 + t(x − x0 ) for all x ∈ C and t ∈ [0, 1],

which implies that
Φ(x0 ) − Φ(x0 + t(x − x0 )) ≤ 0.
Because Jμ (z) = ∂Φ(z), it follows that
Φ(x0 + t(x − x0 )) − Φ(x0 ) ≤ x0 + t(x − x0 ) − x0 , Jμ (x0 + t(x − x0 )),
which implies that
tx0 − x, Jμ (x0 + t(x − x0 )) ≤ Φ(x0 ) − Φ(x0 + t(x − x0 )) ≤ 0.
Thus,
x0 − x, Jμ (x0 + t(x − x0 )) ≤ 0.
Letting t → 0, we obtain x0 − x, Jμ (x0 ) ≤ 0.
2.6. Smoothness 91

2.6 Smoothness
Let C be a nonempty closed convex subset of a normed space X such that
the origin belongs to the interior of C. A linear functional j ∈ X ∗ is said
to be tangent to C at a point x0 ∈ ∂C if j(x0 ) = sup{j(x) : x ∈ C}.
If H = {x ∈ X : j(x) = 0} is the hyperplane, then the set H + x0 is called a
tangent hyperplane to C at x0 .
Definition 2.6.1 A Banach space X is said to be smooth if for each x ∈ SX ,
there exists a unique functional jx ∈ X ∗ such that x, jx  = x and jx  = 1.
Geometrically, the smoothness condition means that at each point x of the
unit sphere, there is exactly one supporting hyperplane {jx = 1}. This means
that the hyperplane {jx = 1} is tangent at x to the unit ball, and this unit ball
is contained in the half space {jx ≤ 1}.

Observation
• p , Lp (1 < p < ∞) are smooth Banach spaces.
• c0 , 1 , L1 , ∞ , L∞ are not smooth.

Differentiability of norms of Banach spaces – Let X be a normed space

and SX = {x ∈ X : x = 1}, the unit sphere of X. Then the norm of X is
Gâteaux differentiable at point x0 ∈ SX if for y ∈ SX
d x0 + ty − x0 
(x0 + ty) = lim
dt t=0
t→0 t
exists (say, y, x0 ). x0  is called the gradient of the norm ϕ(x) = x
at x = x0 . The norm of X is said to Gâteaux differentiable if it is Gâteaux
differentiable at each point of SX . The norm of X is said to be uniformly
Gâteaux differentiable if for each y ∈ SX , the limit is approached uniformly for
x ∈ SX .
Example 2.6.2 Let H be a Hilbert space. Then the norm of H is Gâteaux
differentiable with x = x/x, x = 0. Indeed, for each x ∈ X with x = 0,
we have
x + ty − x x + ty2 − x2
lim = lim
t→0 t t→0 t(x + ty + x)

2ty, x + t2 y2
= lim = y, x/x.
t→0 t(x + ty + x)

Therefore, the norm of H is Gâteaux differentiable with x = x/x.

Remark 2.6.3 In view of Example 2.6.2, we have the following:
(i) at x = 0, ϕ(x) = x is Gâteaux differentiable with x = x/x,
(ii) at x = 0, ϕ(x) = x is not differentiable, but it is subdifferentiable. Indeed,
∂ϕ(0) = ∂0 = {j ∈ H : x, j ≤ x f or all x ∈ H}
= {j ∈ H : j∗ ≤ 1}.
92 2. Convexity, Smoothness, Duality Mappings

Theorem 2.6.4 Let X be a Banach space. Then we have the following:

(a) If X ∗ is strictly convex, then X is smooth.
(b) If X ∗ is smooth, then X is strictly convex.

Proof. (a) Suppose X is not smooth. There exist x0 ∈ SX and j1 , j2 ∈ SX ∗

with j1 = j2 such that x0 , j1  = x0 , j2  = 1. This means that x0 determines
a continuous linear functional on X ∗ that takes its maximum value on BX ∗ at
two distinct points j1 and j1 . Hence X ∗ is not strictly convex.
(b) Suppose X is not strictly convex. There exist j ∈ SX ∗ and x, y ∈ SX
with x = y such that x, j = y, j = 1. Thus, two supporting hyperplanes pass
through j ∈ SX ∗ such that

x, f  = y, f  = 1, f ∈ X ∗ .

Therefore, X ∗ is not smooth.

It is well-known that for a reflexive Banach space X, the dual spaces X and
X ∗ can be equivalently renormed as strictly convex spaces such that the duality
is preserved. Using the above fact, we have

Theorem 2.6.5 Let X be a reflexive Banach space. Then we have the follow-
ing:
(a) X is smooth if and only if X ∗ is strictly convex.
(b) X is strictly convex if and only if X ∗ is smooth.

The following theorem establishes a relation between smoothness and Gâteaux

differentiability of the norm.

Theorem 2.6.6 A Banach space X is smooth if and only if the norm is Gâteaux
differentiable on X\{0}.

Proof. Because the proper convex continuous function ϕ is Gâteaux differen-

tiable if and only if it has a unique subgradient, we have
norm is Gâteaux differentiable at x
⇔ ∂x = {j ∈ X ∗ : x, j = x, j∗ = 1} is singleton
⇔ there exists a unique j ∈ X ∗ such that x, j = x and j∗ = 1
⇔ smooth.

Next, we establish a relation between smoothness of a Banach space and a

property of the duality mapping with gauge function μ.

Theorem 2.6.7 Let X be a Banach space. Then X is smooth if and only if

each duality mapping Jμ with gauge function μ is single-valued; in this case

d
Φ(x + ty) = y, Jμ (x) for all x, y ∈ X. (2.23)
dt t=0
2.6. Smoothness 93

Proof. The Banach space X is smooth if and only if there exists a unique
j ∈ X ∗ satisfying

xμ(||x||), j = xμ(x) andj∗ = 1;

in this case μ(x)j = Jμ (x) = ∂Φ(x), and hence by Corollary 2.5.16, we

obtain the formula (2.23).
∗
Corollary 2.6.8 Let X be a Banach space and Jμ : X → 2X a duality mapping
with gauge function μ. Then j ∈ Jμ (x), x ∈ X if and only if H = {y ∈ X :
y, j = xμ(x)} is a supporting hyperplane for the closed ball Bx [0] at x.
∗
Corollary 2.6.9 Let X be a Banach space and J : X → 2X a duality mapping.
Then the following are equivalent:
(a) X is smooth.
(b) J is single-valued.
(c) The norm of X is Gâteaux differentiable with x = x−1 Jx.

We now study the continuity property of duality mappings.

Theorem 2.6.10 Let X be a Banach space and J : X → X ∗ a single-valued

duality mapping. Then J is norm to weak* continuous.

Proof. We show that xn → x ⇒ Jxn → Jx in the weak* topology. Let xn → x

and set fn := Jxn . Then

xn , fn  = xn fn ∗ , xn  = fn ∗ .

Because {xn } is bounded, {fn } is bounded in X ∗ . Then there exists a subse-

quence {fnk } of {fn } such that fnk → f ∈ X ∗ in the weak* topology. Because
the norm of X ∗ is lower semicontinuous in weak* topology, we have

f ∗ ≤ lim inf fnk ∗ = lim inf xnk  = x.

k→∞ k→∞

Because x, f − fnk  → 0 and x − xnk , fnk  → 0, it follows from the fact

|x, f  − xnk 2 | = |x, f  − xnk , fnk |

≤ |x, f − fnk | + |x − xnk , fnk | → 0

that
x, f  = x2 .
As a result
x2 = x, f  ≤ f ∗ x.

Thus, we have x, f  = x2 , x = f ∗ . Therefore, f = Jx.

94 2. Convexity, Smoothness, Duality Mappings

Theorem 2.6.11 Let X be a Banach space with a uniformly Gâteaux differ-

entiable norm. Then the duality mapping J : X → X ∗ is uniformly demicon-
tinuous on bounded sets, i.e., J is uniformly continuous from X with its norm
topology to X ∗ with the weak* topology.
Proof. Suppose the result is not true. Then there exist sequences {xn } and
{zn }, a point y0 and a positive ε such that
xn  = zn  = y0  = 1, zn − xn → 0 and y0 , Jzn − Jxn  ≥ ε for all n ∈ N.
Set
an := t−1 (xn + ty0  − xn  − ty0 , Jxn )
and
bn := t−1 (zn − ty0  − zn  + ty0 , Jzn ).
If t > 0 is sufficiently small, then both an and bn are less than ε/2. On the
other hand, we have
an ≥ t−1 (xn + ty0 , Jzn  − xn + ty0 , Jxn )
= y0 , Jzn − Jxn  + t−1 xn , Jzn − Jxn 
and
bn ≥ t−1 (zn − ty0 , Jxn  − zn − ty0 , Jzn )
= y0 , Jzn − Jxn  − t−1 zn , Jzn − Jxn .
Thus,
an + bn ≥ 2y0 , Jzn − Jxn  + t−1 xn − zn , Jzn − Jxn 
≥ 2ε − 2t−1 xn − zn ,

a contradiction by choosing t = 2xn − zn /ε for sufficiently large n.

2.7 Modulus of smoothness

Recall that the modulus of convexity of a Banach space X is a function
δX : [0, 2] → [0, 1] defined by
δX (t) = inf{1 − (x + y)/2 : x, y ∈ X, x ≤ 1, y ≤ 1, x − y ≥ t}.
We now introduce the modulus of smoothness of a Banach space.
Definition 2.7.1 Let X be a Banach space. Then a function ρX : R+ → R+
is said to be the modulus of smoothness of X if
 
x + y + x − y
ρX (t) = sup − 1 : x = 1, y = t
2
 
x + ty + x − ty
= sup − 1 : x = y = 1 , t ≥ 0.
2
It is easy to check that ρX (0) = 0 and ρX (t) ≥ 0 for all t ≥ 0.
2.7. Modulus of smoothness 95

The following result contains important properties of the modulus of

smoothness.

Proposition 2.7.2 Let ρX be the modulus of smoothness of a Banach space X.

Then ρX is an increasing continuous convex function.

Proof. Because for fixed x, y ∈ X with x = 1, y = 1, the function

x + ty + x − ty
f (t) = − 1, t ∈ R
2
is convex and continuous on R, it follows that the modulus of smoothness ρX is
also continuous and a convex function.
Moreover, f (−t) = f (t) for each t ∈ R, f is nondecreasing on R+ . Hence ρX
is nondecreasing.

The following theorem gives us an important relation between the modulus

of convexity of X (respectively, X ∗ ) and that of smoothness of X ∗
(respectively, X).

Theorem 2.7.3 Let X be a Banach space. Then we have the following:

 
(a) ρX ∗ (t) = sup 2 − δX (ε) : 0 ≤ ε ≤ 2 for all t > 0.
tε

 
(b) ρX (t) = sup 2 − δX ∗ (ε) : 0 ≤ ε ≤ 2 for all t > 0.
tε

Proof. (a) By the definition of modulus of smoothness of X ∗ , we have

2ρX ∗ (t) = sup{x∗ + ty ∗ ∗ + x∗ − ty ∗ ∗ − 2 : x∗ , y ∗ ∈ SX ∗ }

= sup{x, x∗ +tx, y ∗ +y, x∗ −ty, y ∗ −2 : x, y ∈ SX , x∗ , y ∗ ∈ SX ∗ }
= sup{x + y + tx − y − 2 : x, y ∈ SX }
= sup{x + y + tε − 2 : x, y ∈ SX , x − y = ε, 0 ≤ ε ≤ 2}
= sup{tε − 2δX (ε) : 0 ≤ ε ≤ 2}.

Part (b) can be obtained in the same manner.

As an immediate consequence of Theorem 2.7.3 (b), we have

Corollary 2.7.4 Let X be a Banach space. Then ρX (t)/t is increasing function

and ρX (t) ≤ t for all t > 0.

Theorem 2.7.3 allows us to estimate ρX for Hilbert spaces. Indeed, we have

Proposition 2.7.5 Let H be a Hilbert space. Then for t > 0


ρH (t) = sup tε/2 − 1 + (1 − ε2 /4)1/2 : 0 < ε ≤ 2} = (1 + t2 )1/2 − 1.
96 2. Convexity, Smoothness, Duality Mappings

Observation
• √
If X is a Banach space and H is a Hilbert space, then ρX (t) ≥ ρH (t) =
1 + t2 − 1 for all t ≥ 0.

Let X be a Banach space. Then the characteristic of convexity or the

coefficient of convexity of the Banach space X is the number

0 (X) = sup{ε ∈ [0, 2] : δX (ε) = 0}.

The Banach space X is said to be uniformly convex if 0 (X) = 0 and uniformly

nonsquare if 0 (X) < 2. One may easily see that the modulus of convexity δX
is strictly increasing on [0 , 2].

Example 2.7.6 Let X = R2 with norm  · ∞ defined by

x∞ = (x1 , x2 )∞ = max{|x1 |, |x2 |}.

Then X has a square-shaped unit ball for which δX (ε) = 0 for ε ∈ [0, 2]. Hence
0 (X) = 2.

The following theorem gives an important relation between the modulus of

smoothness of a Banach space and the characteristic of convexity of its dual
space.

Theorem 2.7.7 Let X be a Banach space. Then the following statements are
equivalent:
ρX (t)
(a) lim < ε/2 for all ε ≤ 2.
t→0 t
(b) 0 (X ∗ ) < ε for all ε ≤ 2.

Proof. (a) ⇒ (b). Let ε ∈ [0, 2]. Suppose, for contradiction, that 0 (X ∗ ) ≥ ε.
Then there exist {fn } and {gn } in SX ∗ such that

fn − gn ∗ ≥ ε and lim fn + gn ∗ = 2. (2.24)

n→∞

From the definition of ρX , we get

   
 x + ty   x − ty 

ρX (t) ≥   +   − 1 for all t > 0 and x, y ∈ SX .
2   2 

Therefore,

f (x) + g(x) f (y) − g(y)

ρX (t) ≥ +t − 1 for all f, g ∈ SX ∗ .
2 2

Because x and y were arbitrary, we get

   
f + g  f − g 

ρX (t) ≥   
+ t  − 1.
2 ∗ 2 ∗
2.7. Modulus of smoothness 97

In particular, we have
   
 fn + gn   fn − gn 

ρX (t) ≥     − 1 for all n ∈ N.
2  + t 2 
∗ ∗

It follows from (2.24) that

tε
ρX (t) ≥ .
2
(b) ⇒ (a). Assume that 0 (X ∗ ) < ε and let ε ∈ (0 (X ∗ ), ε). Set t = δX ∗ (ε )
and consider t ∈ [0, 2]. There are two possibilities :
(i) Assume that t < ε . Then tλ/2 < λε /2 and so tλ/2 − δX ∗ (t) < λε /2.
(ii) Assume that ε ≤ t. Then δX ∗ (t) ≥ δX ∗ (ε ) = t , because the
modulus of convexity is an increasing function. Therefore,

λt
≤ λ < t < δX ∗ (t) for any λ < t .
2
This implies that
tλ
− δX ∗ (t) < 0.
2
Therefore, in any case we have for λ < t
 
tλ λε
sup − δX ∗ (t) : t ∈ [0, 2] ≤ .
2 2

Using Theorem 2.7.3, we get ρX (λ) ≤ λ /2, which gives that lim ρX (λ)/λ ≤
λ→0
ε /2. Our choice of ε implies that (b) is true.

Let X be a Banach space. Then the characteristic of smoothness of X is the

number
ρX (t)
ρ0 (X) = lim .
t→0 t
The following theorem allows us to estimate ρ0 (X) for Banach spaces X.

Theorem 2.7.8 Let X be a Banach space. Then

ρX (t) 0 (X ∗ )
ρ0 (X) = ρX (0) = lim = .
t→0 t 2
Proof. Assume first that 0 (X ∗ ) = 2. Then δX ∗ (ε) = 0 for every ε ∈ [0, 2].
Therefore, using Theorem 2.7.3, we get ρX (t) = t for every t > 0. Hence

ρX (t) 0 (X ∗ )
lim =1= .
t→0 t 2
Now if we assume that 0 (X ∗ ) < 2, then from Theorem 2.7.7 we get the desired
conclusion.
98 2. Convexity, Smoothness, Duality Mappings

Using Theorem 2.7.3 and 2.7.8, we have

Theorem 2.7.9 Let X be a Banach space. Then we have the following:
(a) ρ0 (X) = 0 (X ∗ )/2.
(b) ρ0 (X ∗ ) = 0 (X)/2.

2.8 Uniform smoothness

Recall that the Banach space X is uniformly convex if δX (ε) > 0 for all ε ∈ (0, 2].

We now define uniform smoothness of a Banach space.

Definition 2.8.1 A Banach space X is said to be uniformly smooth if

ρX (t)
ρX (0) = lim = 0.
t→0 t
Example 2.8.2 The p spaces (1 < p ≤ 2) are uniformly smooth. In fact,

ρp (t) (1 + tp )1/p − 1

lim = lim = 0.
t→0 t t→0 t
Uniform smoothness has a close relation with differentiability of norm.
Theorem 2.8.3 Every uniformly smooth Banach space X is smooth.
Proof. Suppose, for contradiction, that X is not smooth. Then there exist
x ∈ X\{0}, and i, j ∈ X ∗ such that i = j, i = j = 1 and x, i = x, j =
x. Let y ∈ X such that y = 1 and y, i − j > 0. For each t > 0, we have

0 < ty, i − j
= ty, i − ty, j
x + ty, i + x − ty, j
= −1
2
x + ty + x − ty
≤ − 1,
2
and it follows that
ρX (t)
0 < y, i − j ≤ for each t > 0.
t

Hence X is not uniformly smooth.

Next, we establish the duality between uniform convexity and uniform smooth-
ness.
Theorem 2.8.4 Let X be a Banach space. Then X is uniformly smooth if and
only if X ∗ is uniformly convex.
2.8. Uniform smoothness 99

Proof. Recall that

 
tε
ρX (t) = sup − δX ∗ (ε) : 0 < ε ≤ 2 for all t > 0. (2.25)
2

Suppose, for contradiction, that X ∗ is not uniformly convex. Then there exists
ε0 ∈ (0, 2] with δX ∗ (ε0 ) = 0. From (2.25), we have

tε0
− δX ∗ (ε0 ) ≤ ρX (t)
2
which gives us that
ε0 ρX (t)
0< ≤ for all t > 0,
2 t
and this means that X is not uniformly smooth.
Conversely, assume that X is not uniformly smooth. Then ρX (0) =
ρX (t) ρX (t)
lim = 0. Hence for ε > 0 with lim = ε, there exists a sequence
t→0 t t→0 t
{tn } in (0, 1) such that

ρX (tn )
tn → 0 and lim = ε.
n→∞ tn

From (2.25), there exists a sequence {εn } in (0, 2] such that

ε t n εn
tn ≤ − δX ∗ (εn ),
2 2
which implies that
tn
0 < δX ∗ (εn ) ≤ (εn − ε).
2
It follows from the condition tn < 1 that ε < εn . Because δX ∗ is a nondecreasing
function, we have δX ∗ (ε) ≤ δX ∗ (εn ) → 0, i.e., X ∗ is not uniformly convex.

Theorem 2.8.5 Let X be a Banach space. Then X is uniformly convex if and

only if X ∗ is uniformly smooth.

Proof. Notice
 
tε
ρX ∗ (t) = sup − δX (ε) : 0 < ε ≤ 2 for all t > 0.
2

By interchanging the roles of X and X ∗ , we obtain the result by Theo-

rem 2.8.4.

Theorem 2.8.6 Every uniformly smooth Banach space is reflexive.

100 2. Convexity, Smoothness, Duality Mappings

Proof. Let X be a uniformly smooth Banach space. Then X ∗ is uniformly

convex and hence X ∗ is reflexive. It follows from Theorem 1.9.26 (which states
that the reflexivity of X ∗ implies the reflexivity of X) that X is reflexive.

Fréchet differentiability of norm and uniform smoothness

Uniform smoothness can be characterized by uniform Fréchet differenti-

ability of the norm.

The norm of a Banach space X is said to Fréchet differentiable if for each

x + ty − x
x ∈ SX , lim exists uniformly for y ∈ SX .
t→0 t
In the other words, there exists a function εx (s) with εx (s) → 0 as s → 0
such that

x + ty − x − ty, Jx ≤ |t|εx (|t|) for all y ∈ SX .

In this case, the norm is Gâteaux differentiable and

2 x + ty2 − 12 x2
1
lim sup − y, Jx = 0 for all x ∈ X.
t→0 y∈SX t

On the other hand,

1 1 1
x2 + h, Jx ≤ x + h2 ≤ x2 + h, Jx + b(h)
2 2 2
for all bounded x, h ∈ X, where b is a function defined on R+ such that
b(t)
lim = 0.
t→0 t

We say that the norm of a Banach space X is uniformly Fréchet differentiable

if
x + ty − x
lim exists uniformly for all x, y ∈ SX .
t→0 t
We now establish some results concerning Fréchet differentiability of the
norm of Banach spaces.

Theorem 2.8.7 Let X be a Banach space with a Fréchet differentiable norm.

Then the duality mapping J : X → X ∗ is norm to norm continuous.

Proof. It is sufficient to prove that xn → x ∈ SX ⇒ Jxn → Jx ∈ SX ∗ . Let

{xn } be a sequence in SX such that xn → x. Then x ∈ SX . Because X has a
Fréchet differentiable norm,

x + ty − x
lim = y, Jx uniformly in y ∈ SX ,
t→0 t
2.8. Uniform smoothness 101

i.e., for any ε > 0, there exists δ > 0 such that

x + ty − x
− y, Jx < ε for all y ∈ SX and all t with 0 < |t| ≤ δ.
t
Hence

x + ty − x < t(y, Jx + ε) and x − ty − x < −t(y, Jx − ε),

so that

x + ty − 1 < t(y, Jx + ε) and x − ty − 1 < t(ε − y, Jx).

Note

0 ≤ 1 − x, Jxn  = xn , Jxn  − x, Jxn 

≤ xn − x, Jxn 
≤ xn − xJxn ∗ = xn − x → 0,

i.e., x, Jxn  → 1 as n → ∞. Then there exists n0 ∈ N such that

1 ≤ x, Jxn  + tε for all n ≥ n0 .

Because

1 − tε ≤ x, Jxn  = x, Jx + Jxn  − 1

= x + ty, Jx + x − ty, Jxn  − ty, Jx − Jxn  − 1
≤ x + tyJx∗ + x − tyJxn ∗ − ty, Jx − Jxn  − 1
≤ ty, Jx + tε + 1 + 1 + tε − ty, Jx − ty, Jx − Jxn  − 1
= 2tε − ty, Jx − Jxn  + 1,

this implies that

y, Jx − Jxn  ≤ 3ε for all y ∈ SX .
Similarly, we can show that

y, Jxn − Jx ≤ 3ε for all y ∈ SX .

Thus,
y, Jxn − Jx ≤ 3ε for all n ≥ n0 and y ∈ SX
which gives us
Jxn − Jx∗ < 3ε for all n ≥ n0 .

Therefore, xn → x in X implies Jxn → Jx in X ∗ .

Theorem 2.8.8 Let X be a Banach space. Then the following are equivalent:
(a) X has a uniformly Fréchet differentiable norm.
(b) X ∗ is uniformly convex.
102 2. Convexity, Smoothness, Duality Mappings

Proof. (a) ⇒ (b). Suppose the norm of X is uniformly Fréchet differentiable.

Then for any ε > 0, there exists δ > 0 such that
x + ty − x ε
− y, Jx < for all x, y ∈ SX and all t with 0 < |t| ≤ δ.
t 8
Then for fixed t with 0 < t < δ, we have
tε
x + ty < + ty, Jx + 1
8
and
tε
x − ty < − ty, Jx + 1.
8
As a result
tε
x + ty + x − ty < + 2 for all x, y ∈ SX .
4
Now, let i, j ∈ SX ∗ with i − j∗ ≥ ε > 0, then there exists y0 ∈ SX such that
ε
y0 , i − j > .
2
Note
i + j∗ = sup x, i + j
x∈SX
= sup (x + ty0 , i + x − ty0 , j − ty0 , i − j)
x∈SX
 
tε
< sup x + ty0  + x − ty0  −
x∈SX 2
tε tε
≤ +2−
4 2
tε
≤ 2− .
2
This implies (i + j)/2∗ < 1 − δ(ε). Hence X ∗ is uniformly convex.
(b) ⇒ (a). Let x, y ∈ SX . Then for t > 0,
y, Jx x + ty, Jx − x2
=
x tx
x + tyx − x2
≤
tx
x + ty − x
=
t
x + ty2 − x + tyx
=
tx + ty
x + ty, J(x + ty) − x, J(x + ty)
≤
tx + ty
y, J(x + ty)
=
x + ty
2.8. Uniform smoothness 103

and for t < 0,

y, J(x + ty) x + ty − x y, Jx

≤ ≤ .
x + ty t x

By Theorem 2.4.15, X has a uniformly Fréchet differentiable norm.

Theorem 2.8.9 Let X be a Banach space with uniformly Fréchet differentiable

norm. Then the duality mapping J : X → X ∗ is uniformly continuous on each
bounded set in X.

Proof. Because X ∗ is uniformly convex, the result follows from Theorem 2.4.15

We now study the duality mapping from X ∗ to X. To do so, we define the

conjugate function f ∗ : X ∗ → (−∞, ∞] of any function f : X → (−∞, ∞] by

f ∗ (j) = sup{x, j − f (x) : x ∈ X}, j ∈ X ∗. (2.26)

The conjugate of f ∗ , i.e., the function on X defined by

f ∗∗ (x) = sup{x, j − f ∗ (j) : j ∈ X ∗ }, x∈X

is called the biconjugate of f .

Observation
• f is lower semicontinuous proper convex on X if and only if f ∗∗ = f .

Example 2.8.10 Let C be a nonempty subset of normed space X. Then the

conjugate of the indicator function iC of C is given by

i∗C (j) = sup{x, j : x ∈ C}, j ∈ X ∗.

The function i∗C is called the support function of C.

We now give some basic properties of conjugate functions.

Proposition 2.8.11 Let f ∗ be the conjugate function f . Then

f (x) + f ∗ (j) ≥ x, j for all x ∈ X, j ∈ X ∗ . (2.27)

Proof. It easily follows from (2.26).

The inequality (2.27) is known as the Young inequality. Observe also that if
f is a proper function, then the relation (2.26) can be written as

f ∗ (j) = sup{x, j − f (x) : x ∈ Dom(f )}, j ∈ X ∗ .

104 2. Convexity, Smoothness, Duality Mappings

Proposition 2.8.12 Let f ∗ be the conjugate function of f . Then

(cf )∗ (j) = cf ∗ (c−1 j) for all c > 0 and j ∈ X ∗ .
Proof. For j ∈ X ∗ , we have
(cf )∗ (j) = sup{x, j − (cf )(x) : x ∈ X}
= c sup{c−1 x, j − f (x) : x ∈ X}
= c sup{x, c−1 j − f (x) : x ∈ X}
= cf ∗ (c−1 j).

Proposition 2.8.13 Let X be a normed space and f : X → (−∞, ∞] a proper

convex function. Then the following statements are equivalent:
(a) j ∈ ∂f (x) for x ∈ X.
(b) f (x) + f ∗ (j) ≤ x, j.
(c) f (x) + f ∗ (j) = x, j.
Proof. (b) ⇔ (c). The Young inequality (2.27) shows that (b) and (c) are
equivalent.

(c) ⇔ (a). Suppose condition (c) holds. Then from the Young inequality (2.27),
we find that
f (y) − f (x) ≥ y − x, j for all y ∈ X,
i.e., j ∈ ∂f (x).
Using a similar argument, it follows that (c) ⇒ (a).
Proposition 2.8.14 Let X be a normed space and f : X → (−∞, ∞] a lower
semicontinuous proper convex function. Then j ∈ ∂f (x) ⇔ x ∈ ∂f ∗ (j).
Proof. Because f is a lower semicontinuous convex function, f ∗∗ = f . Observe
that
j ∈ ∂f (x) ⇔ f (x) + f ∗ (j) = x, j
⇔ f ∗∗ (x) + f ∗ (j) = x, j
⇔ x ∈ ∂f ∗ (j).
Proposition 2.8.15 Let X be a Banach space. If f (x) = xp /p, p > 1, then
f ∗ (j) = jq∗ /q, 1/p + 1/q = 1.
Proof. Because Jp (x) = ∂(xp /p) = {j ∈ X ∗ : x, j = xj∗ , j∗ =
xp−1 }, we have
f ∗ (j) = sup {x, j − f (x) = sup {xp − xp /p} = sup {xp /q}.
x∈X x∈X x∈X

Note j∗ = xp−1 so jq∗ = xq(p−1) = xp . Therefore, f ∗ (j) =

jq∗ /q.
2.8. Uniform smoothness 105

Theorem 2.8.16 Let p > 1. Let X be a uniformly smooth Banach space and
let Jp : X → X ∗ and Jq∗ : X ∗ → X be the duality mappings with gauge functions
μp (t) = tp−1 and μq (t) = tq−1 , respectively. Then Jp−1 = Jq∗ .
Proof. The uniform smoothness of X implies that X is reflexive (see
Theorem 2.8.6) and that X ∗ is uniformly convex and reflexive. Note also Jμ is
surjective if and only if X is reflexive. Because Jp is single-valued, it follows
that the inverse Jp−1 : X ∗ = Dom (Jp−1 ) → X = X ∗∗ exists and is given by

Jp−1 (j) = {x ∈ X : j = Jp (x)} for all j ∈ X ∗ .

Now, let Φ(t) = tp /p, t > 0. It is easy to see that Φ(·) = ·p /p is a continuous
convex function and that its conjugate is given by Φ∗ (j∗ ) = jq∗ /q for all
j ∈ X ∗ . Note Jp (x) = ∂Φ(x) and Jq∗ (j) = ∂Φ∗ (j∗ ) for all x ∈ X, j ∈ X ∗ .
Using Proposition 2.8.14, we have
j ∈ ∂Φ(x) if and only if x ∈ ∂Φ∗ (j∗ ).

Therefore, Jp−1 (j) = Jq∗ (j) for all j ∈ X ∗ .

The following inequality is very useful in the existence and approximation

of solutions of nonlinear operator equations.
Theorem 2.8.17 Let X be a Banach space. Then the following are equivalent:
(a) X is uniformly convex.
(b) For any p, 1 < p < ∞ and r > 0, there exists a strictly increasing convex
function gr : R+ → R+ such that gr (0) = 0 and
tx + (1 − t)yp ≤ txp + (1 − t)yp − t(1 − t)gr (x − y) (2.28)
for all x, y ∈ Br [0] and t ∈ [0, 1].
Proof. (a) ⇒ (b). Let X be a uniformly convex Banach space. Assume that
1 < p < ∞. It suffices to prove that (2.28) is true for r = 1. Now we define a
function γ by
γ(ε) = inf{2p−1 (xp + yp ) − x + yp : x, y ∈ BX and x − y ≥ ε}
for all ε ∈ (0, 2].
Because
 p
a+b ap + bp
< for all a, b ≥ 0 and a = b, (2.29)
2 2
we have
γ(ε) ≥ 0 for all 0 < ε ≤ 2.
Suppose that γ(ε) = 0 for some ε > 0. Then there exist sequences {xn } and
{yn } in BX such that xn − yn  ≥ ε for all n ∈ N and
lim 2p−1 (xn p + yn p ) − xn + yn p = 0.
n→∞
106 2. Convexity, Smoothness, Duality Mappings

We may assume a subsequence of {xn } denoted by {xn } such that

a = lim xn , b = lim yn  and c = lim xn + yn 

n→∞ n→∞ n→∞

exist. Thus,  p
a+b ap + bp
= ,
2 2
i.e., equality of inequality (2.29) holds with c = a + b. For a = b > 0, c =
2a = lim xn + yn , it follows from Theorem 2.2.7 that lim xn − yn  = 0, a
n→∞ n→∞
contradiction. Therefore,

γ(ε) > 0 for all ε, 0 < ε ≤ 2.

Now set
 
λxp + (1 − λ)yp − λx + (1 − λ)yp
μ(ε) := inf ,
λ(1 − λ)

where the infimum is taken over all x, y ∈ BX with x − y ≥ ε and λ ∈ (0, 1).
Note μ(ε) ≥ γ(ε)/2p−1 > 0 for all ε, 0 < ε ≤ 2. Thus, it suffices to take as g1
the double dual Young’s function μ∗∗ .
(b) ⇒ (a). Suppose (2.28) is satisfied. For x, y ∈ BX and x − y = ε, we
have
 
x + y  1
 
 2  ≤ 1 − 4 g1 (ε)
≤ 1 − δX (ε),

i.e., δX (ε) ≥ g1 (ε)/4, which shows that X is a uniformly convex Banach

space.

2.9 Banach limit

In this section, we generalize the concept of limit by introducing Banach limits
and we discuss its properties.

Let : c → K be the “limit functional” defined by

(x) = lim xi for x = {xi } ∈ c.

i→∞

Then is a linear functional on c. In order to extend limit on ∞, use the

following notations and results.
Let S be a nonempty set and let B(S) be the Banach space of all bounded
real-valued functions on S with supremum norm.
Example 2.9.1 Let S = N = {1, 2, 3, · · · }. Then B(S) = l∞ .
2.9. Banach limit 107

Let X be a subspace of B(S) and let j be an element of X ∗ . Let e be a

constant function on X defined by e(s) = 1 for all s ∈ S. We will denote j(e) by
j(1). When X contains constants, a linear functional j on X is called a mean
on X if j∗ = j(1) = 1.

The following example shows that there is a subspace of ∞ for which the
mean exists.

Example 2.9.2 Let ∞ = {x = {xi } : sup |xi | < ∞} and X a subset of ∞

i∈N
such that  
1
n
X= x = {xi } ∈ ∞ : lim xi exists .
n→∞ n
i=1

Then X is a linear subspace of ∞. In fact, for x = {xi } and y = {yi } in X,

we have

1 1
n n
lim xi exists and lim yi exists. (2.30)
n→∞ n n→∞ n
i=1 i=1

Hence for scalars α, β, we have

αx + βy = (αx1 + βy1 , · · · , αxi + βyi , · · · ).

Using (2.30), we obtain that

1 1 1
n n n
lim (αxi + βyi ) = α( lim xi ) + β( lim yi )
n→∞ n n→∞ n n→∞ n
i=1 i=1 i=1

exists. It follows that X is a linear subspace of ∞ . We now define j : X → R

by
1
n
j(x) = lim xi for all x ∈ X.
n→∞ n
i=1

Note j(1) = 1 and

1
n
|j(x)| = lim xi
n→∞ n
i=1

1
n
≤ lim sup |xi |
n→∞ n i=1
≤ x∞ ,

and it follows that j∗ = 1. Therefore, j is linear and j∗ = j(1) = 1, i.e., j
is a mean on X.

We now give an equivalent condition for mean.

108 2. Convexity, Smoothness, Duality Mappings

Theorem 2.9.3 Let X be a subspace of B(S) containing constants and j ∈ X ∗ .

Then the following are equivalent:
(a) j is a mean on X, i.e., j∗ = j(1) = 1.
(b) The inequalities

inf x(s) ≤ j(x) ≤ sup x(s)

s∈S s∈S

hold for each x ∈ X.

Proof. (a) ⇒ (b). First, we show that j(x) ≥ 0 for all x ≥ 0. Suppose, for
contraction, that j(x) < 0. Choose a positive number K with x ≤ K. Then

j(K − x) = Kj(1) − j(x) = K − j(x) > K.

Because

j(K − x) ≤ j∗ K − x = K − x = sup |K − x(s)| ≤ K,

s∈S

it follows that
K < j(K − x) ≤ K,
a contradiction. Therefore, j(x) ≥ 0.
Observe that

inf x(s) ≤ x ≤ sup x(s) for each x ∈ X.

s∈S s∈S

Because j(x) ≥ 0 for x ≥ 0, we have

inf x(s) = j( inf x(s)) ≤ j(x) ≤ j(sup x(s)) = sup x(s).

s∈S s∈S s∈S s∈S

(b) ⇒ (a). For x = 1, we have 1 ≤ j(1) ≤ 1 and hence j(1) = 1. Note for each
x ∈ X,
j(x) ≤ sup x(s) ≤ sup |x(s)| = x
s∈S s∈S

and
−j(x) = j(−x) ≤  − x = x,
so |j(x)| ≤ x for each x ∈ X. Thus, j∗ = 1. Therefore, j∗ = j(1) = 1,
i.e., j is a mean on X.

Let f ∈ ∞ . We denote fn (xn+m ) for f (xm+1 , xm+2 , xm+3 , · · · , xm+n , · · · ),

m = 0, 1, 2, · · · . A continuous linear functional j on l∞ is called a Banach limit
if
(L1 ) j∗ = j(1) = 1,
(L2 ) jn (xn ) = jn (xn+1 ) for each x = (x1 , x2 , · · · ) ∈ l∞ .
It is denoted by LIM .
2.9. Banach limit 109

Theorem 2.9.4 (The existence of Banach limits) – There exists a linear

continuous functional j on l∞ such that j∗ = j(1) = 1 and jn (xn ) = jn (xn+1 )
for each x = {xn }n∈N ∈ ∞ .
Proof. Let p : ∞ → R be the functional defined by
x1 + x2 + · · · + xn
p(x) = lim sup .
n→∞ n
Then
x1 + x2 + · · · + xn
−p(−x) = lim inf .
n→∞ n
For x ∈ c, we have
x1 + x2 + · · · + xn
(x) = lim xn = lim = p(x).
n→∞ n→∞ n
Moreover,
p(x + y) ≤ p(x) + p(y) for all x, y ∈ c
and
p(αx) = αp(x) for all x ∈ c and α ≥ 0.
Thus, p is a sublinear functional with (x) = p(x). By the Hahn-Banach
theorem, there is an extension L : ∞ → R of (from c to ∞ ) such that
L(x) ≤ (x) for all x ∈ ∞

and
−p(−x) ≤ L(x) ≤ p(x) for all x ∈ ∞.

Thus, we have
p(1, 1, 1, · · · ) = 1
and
x1 − xn+1
p((x1 , x2 , · · · , xn , · · · ) − (x2 , x3 , · · · , xn+1 , · · · )) = lim sup = 0.
n→∞ n
Hence
L((x1 , x2 , · · · , xn , · · · ) − (x2 , x3 , · · · , xn+1 , · · · )) = 0,
which implies that
L(x1 , x2 , · · · , xn , · · · ) = L(x2 , x3 , · · · , xn+1 · · · )
for all x = (x1 , x2 , · · · xn , · · · ) ∈ ∞.

Therefore, L is a Banach limit.

Observation
• Every Banach limit is a positive functional on ∞ , i.e., LIMn (x) ≥ 0 for all
x ∈ ∞ .
• LIM (1, 1, · · · 1, · · · ) = 1.

We now give elementary properties of Banach limits.

110 2. Convexity, Smoothness, Duality Mappings

Proposition 2.9.5 Let LIM be a Banach limit. Then

lim inf xn ≤ LIM (x) ≤ lim sup xn for each x = (x1 , x2 , · · · ) ∈ l∞ .

n→∞ n→∞

Moreover, if xn → a, then LIM (x) = a.

Proof. For each m ∈ N, we have

LIMn (xn ) = LIMn (xn+1 ) = · · · = LIMn (xn+(m−1) ) ≥ inf xn

n≥m

and hence LIMn (xn ) ≥ sup inf xn = lim inf xn .

m∈N n≥m n→∞

Similarly, since LIMn (xn ) ≤ sup xn , we have LIMn (xn ) ≤ lim sup xn .
n≥m n→∞
Therefore,

lim inf xn ≤ LIM (x) ≤ lim sup xn for each x = (x1 , x2 , · · · ) ∈ l∞ .

n→∞ n→∞

Letting xn → a, we have lim inf xn = lim sup xn = a and hence LIM (x)
n→∞ n→∞
= a.

Proposition 2.9.6 Let a be a real number and let (x1 , x2 , · · · ) ∈ ∞. Then the
following are equivalent:
(a) LIMn (xn ) ≤ a for all Banach limits LIM.
(b) For each ε > 0, there exists m0 ∈ N such that
xn + xn+1 + · · · + xn+m−1
< a + ε f or all m ≥ m0 and n ∈ N. (2.31)
m

Proof. (a) ⇒ (b). Suppose that for {xn } ∈ ∞ , we have LIMn (xn ) ≤ a for all
Banach limits LIM. Define a sublinear functional q : ∞ → R by
 n+m−1 
1 
q(y1 , y2 , · · · ) = lim sup sup yi , {yn } ∈ ∞.
m→∞ n∈N m i=n

By the Hahn-Banach theorem, there exists a linear functional j : ∞ → R such

that
j ≤ q and jn (xn ) = qn (xn ).
It is easy to see that j is a Banach limit. From the assumption, we have
 
1 
n+m−1
qn (xn ) = lim sup sup xi ≤ a.
m→∞ n∈N m i=n

Thus, for ε > 0, there exists m0 ∈ N such that

xn + xn+1 + · · · + xn+m−1
< a + ε for all m ≥ m0 and n ∈ N.
m
2.9. Banach limit 111

(b) ⇒ (a). Suppose for each ε > 0, there exists m0 ∈ N such that (2.31) holds.
Let LIM be a Banach limit. Then
 
xn + xn+1 + · · · + xn+m0 −1
LIMn (xn ) = LIMn ≤ a + ε.
m0

Because ε is an arbitrary positive real number, we have LIMn (xn ) ≤ a.

Proposition 2.9.7 Let a be a real number and let (x1 , x2 , · · · ) ∈ ∞ such that
LIMn (xn ) ≤ a for all Banach limits LIM and lim sup(xn+1 − xn ) ≤ 0. Then
n→∞
lim sup xn ≤ a.
n→∞

Proof. Let ε > 0. By Proposition 2.9.6, there exists m ≥ 2 such that

xn + xn+1 + · · · + xn+m−1 ε
< a + for all n ∈ N.
m 2
Choose n0 ∈ N such that
ε
xn+1 − xn < for all n ≥ n0 .
m−1
Let n ≥ n0 + m. Observe that
xn = xn−i + (xn−i+1 − xn−i ) + · · · + (xn − xn−1 )
iε
≤ xn−i + for each i = 0, 1, · · · , m − 1.
m−1
Thus,
lim sup xn ≤ a + ε.
n→∞

Because ε is arbitrary positive number, we get the conclusion.

We note that if a linear functional j on l∞ satisfying:

lim inf xn ≤ j(x) ≤ lim sup xn for each x = (x1 , x2 , · · · ) ∈ l∞ ,
n→∞ n→∞

then j is a mean on ∞. Thus, every Banach limit on ∞ is a mean on ∞.

Let X be a Banach space, {xn } a bounded sequence in X, and LIM a Banach

limit. Then a point x0 ∈ X is said to be a mean point of {xn } concerning a
Banach limit LIM if
LIMn xn , j = x0 , j for all j ∈ X ∗ .
We establish two preliminary results related to mean points.
Proposition 2.9.8 (Existence of mean points) – Let X be a reflexive
Banach space and {xn } a bounded sequence in X. Then, for a Banach limit
LIM , there exists a point x0 in X such that
LIMn xn , j = x0 , j for all j ∈ X ∗ .
112 2. Convexity, Smoothness, Duality Mappings

Proof. Note the function LIMn xn , j is linear in j. Further, as

|LIMn xn , j| ≤ (sup xn ) · j∗ ,

n∈N

the function LIMn xn , j is also bounded in j. So, we have j0∗ ∈ X ∗∗ such that

LIMn xn , j = j0∗ , j for every j ∈ X ∗ .

Because X is reflexive, there exists x0 ∈ X such that LIMn xn , j = x0 , j for
all j ∈ X ∗ .

Proposition 2.9.9 Let {xn } be a bounded sequence in a Banach space X and

0∞∈ X a mean point of {xn } concerning a Banach limit LIM . Then x0 ∈
x
n=1 co({xk }k≥n ).

Proof. If not, there exists n0 ∈ N such that x0 ∈ / co{xn : n ≥ n0 }. By the

separation theorem, we obtain a point j ∈ X ∗ such that

x0 , j < inf{z, j : z ∈ co{xn : n ≥ n0 }}.

Thus, we have

LIMn xn , j = x0 , j < inf{(xn , j) : n ≥ n0 }

≤ LIMn {xn , j : n ≥ n0 } = LIMn xn , j,

a contradiction.

We now characterize the sequences in ∞ for which all Banach limits coincide.
It is obvious that for any element x ∈ c,

LIM (x) = (x) = lim xn for all Banach limit LIM.

n→∞

However, there exist nonconvergent sequences for which all Banach limits
coincide.

Example 2.9.10 Let x = (1, 0, 1, 0, · · · ) ∈ ∞. Then

(x1 , x2 , · · · xn , · · · ) + (x2 , x3 , · · · , xn+1 , · · · ) = (1, 1, 1, · · · ),

and it follows that

LIMn (xn ) + LIMn (xn+1 ) = LIMn (1) = 1 for all LIM.

Using (L2 ), we have

1
LIMn (xn ) = for all Banach limit LIM.
2
2.9. Banach limit 113

A bounded sequence x = {xi } is said to be almost convergent if all its

Banach limits have the same value at x. Equivalently, x = {xi } ∈ ∞ is almost
convergent if
xn + xn+1 · · · + xn+i−1
lim exists uniformly in n.
i→∞ i
We have seen in Example 2.9.10 that the sequence (1, 0, 1, 0, · · · ) is not conver-
gent, but it is almost convergent.

In optimization theory, the structure of M defined in our next result is of

much interest.

Theorem 2.9.11 Let C be a nonempty closed convex subset of a reflexive

Banach space X, {xn } a bounded sequence in C, LIM a Banach limit, and
ϕ a real-valued function on C defined by ϕ(z) = LIMn xn − z2 , z ∈ C. Then
the set M defined by

M = {u ∈ C : LIMn xn − u2 = inf LIMn xn − z2 } (2.32)

z∈C

is a nonempty closed convex bounded set. Moreover, if X is uniformly convex,

then M has exactly one point.

Proof. First, we show that ϕ is continuous and convex. Let {ym } be a sequence
in C such that ym → y ∈ C. Set L := sup{xn − ym  + xn − y : m, n ∈ N}.
Observe that

xn − ym 2 − xn − y2 ≤ (xn − ym  + xn − y)(xn − ym  − xn − y)

≤ L| xn − ym  − xn − y |
≤ Lym − y for all n, m ∈ N.

Then
LIMn xn − ym 2 ≤ LIMn xn − y2 + Lym − y.
Similarly we have

LIMn xn − y2 ≤ LIMn xn − ym 2 + Lym − y.

Thus, we have
|ϕ(ym ) − ϕ(x)| ≤ Lym − x.
Hence ϕ is continuous on C. Now, let x, y ∈ C and λ ∈ [0, 1]. It is easy to see
that
ϕ((1 − λ)x + λy) ≤ (1 − λ)ϕ(x) + λϕ(y).
Hence ϕ is convex.
Using the fact ((a + b)/2)2 ≤ (a2 + b2 )/2 for all a, b ≥ 0, we have

ym 2 ≤ 2ym − xn 2 + 2xn 2 ,

114 2. Convexity, Smoothness, Duality Mappings

and hence
ym 2 ≤ 2ϕ(ym ) + 2 sup xn 2 ,
n∈N
i.e., ϕ(ym ) → ∞ as ym  → ∞. Thus, ϕ is a continuous convex functional and
ϕ(z) → ∞ as z → ∞. Because X is reflexive, ϕ attains its infimum over C
by Theorem 2.5.8. Then M is a nonempty closed convex set. Moreover, M is
bounded. Indeed, let u ∈ M . Because
u2 ≤ 2u − xn 2 + 2xn 2 for all n ∈ N,
this implies that
u2 ≤ 2ϕ(u) + 2K = 2 inf ϕ(z) + 2K
z∈C

for some K ≥ 0.
Now, suppose X is uniformly convex. Let z1 , z2 ∈ M . Then (z1 +z2 )/2 ∈ M .
Choose r > 0 large enough so that {xn } ∪ M ⊂ Br [0]. Then xn − z1 , xn − z2 ∈
B2r [0] for all n ∈ N. By Theorem 2.8.17, we have
 2
 
xn − z1 + z2  ≤ 1 xn − z1 2 + 1 xn − z2 2 − 1 g2r (z1 − z2 ).
 2  2 2 4
If z1 = z2 , we have
 
z1 + z 2 1 1 1
inf ϕ(z) ≤ ϕ ≤ ϕ(z1 ) + ϕ(z2 ) − g2r (z1 − z2 )
z∈C 2 2 2 4
1
= inf ϕ(z) − g2r (z1 − z2 )
z∈C 4
< inf ϕ(z),
z∈C

a contradiction. Therefore, M has exactly one element.

Let LIM be a Banach limit and let {xn } be a bounded sequence in a Banach
space X. We observe that if ψ : X → R is bounded, Gâteaux differentiable
uniformly on bounded sets, then a function f : X → R defined by f (z) =
LIMn ψ(xn + z) is Gâteaux differentiable with Gâteaux derivative given by
y, f  (z) = LIMn y, ψ  (xn + z) for each y ∈ X.

Using the above facts, we give the following result, which will be used in
convergence of sequences {xn } in Banach spaces with Gâteaux differentiable
norm.
Theorem 2.9.12 Let X be a Banach space with a uniformly Gâteaux differen-
tiable norm and {xn } a bounded sequence in X. Let LIM be a Banach limit
and u ∈ X. Then
LIMn xn − u2 = inf LIMn xn − z2
z∈X

if and only if
LIMn z, J(xn − u) = 0 for all z ∈ X.
2.10. Metric projection and retraction mappings 115

Proof. Let u ∈ X be such that LIMn xn − u2 = inf LIMn xn − z2 .
z∈X
Then u minimizes the continuous convex function φ : X → R+ defined by
φ(z) = LIMn xn − z2 , so we have φ (u) = 0.
Note that the norm of X is Gâteaux differentiable, and Jx is the subdiffer-
ential of the convex function ϕ(x) = x2 /2 at x as the Gâteaux differential of
ϕ. Hence
LIMn z, J(xn − u) = z, φ (u) = 0 for all z ∈ X.
Conversely, suppose that LIMn u − z, J(xn − u) = 0 for all z ∈ X.
If x ∈ X,

xn − x2 − xn − u2 ≥ 2u − x, J(xn − u) for all n ∈ N.

Because LIMn u − x, J(xn − u) = 0 for all x ∈ X, we obtain

LIMn xn − u2 = inf LIMn xn − x2 .

x∈X

Corollary 2.9.13 Let X be a Banach space with a uniformly Gâteaux differ-

entiable norm and C a nonempty closed convex subset of X. Let {xn } be a
bounded sequence in C. Let LIM be a Banach limit and u ∈ C. Then

u ∈ M if and only if LIMn z, J(xn − u) ≤ 0 f or all z ∈ C.

2.10 Metric projection and retraction mappings

Let C be a nonempty subset of a normed space X and let x ∈ X. An element
y0 ∈ C is said to be a best approximation to x if

x − y0  = d(x, C),

where d(x, C) = inf x − y. The number d(x, C) is called the distance from x
y∈C
to C or the error in approximating x by C.

The (possibly empty) set of all best approximations from x to C is denoted

by
PC (x) = {y ∈ C : x − y = d(x, C)}.
This defines a mapping PC from X into 2C and is called the metric projection
onto C. The metric projection mapping is also known as the nearest point
projection mapping, proximity mapping, and best approximation operator.

The set C is said to be a proximinal 2 (respectively, Chebyshev) set if each

x ∈ X has at least (respectively, exactly) one best approximation in C.

2 The term “proximinal” is a combination of the words “proximity” and “minimal” and

was coined by Killgrove.

116 2. Convexity, Smoothness, Duality Mappings

Observation
• C is proximinal if PC (x) = ∅ for all x ∈ X.
• C is Chebyshev if PC (x) is singleton for each x ∈ X.
• The set of best approximations is convex if C is convex.

Some fundamental results on proximinal sets are the following:

First, we observe that every proximinal set must be closed.

Proposition 2.10.1 Let C be a proximinal subset of a Banach space X. Then

C is closed.

Proof. Suppose, for contradiction, that C is not closed. Then there exists a
sequence {xn } in C such that xn → x and x ∈
/ C, but x ∈ X. It follows that

d(x, C) ≤ xn − x → 0,

so that d(x, C) = 0. Because x ∈

/ C, it means that

x − y > 0 for all y ∈ C.

This implies PC (x) = ∅. This contradicts PC (x) = ∅.

Theorem 2.10.2 (The existence of best approximations) – Let C be a

nonempty weakly compact convex subset of a Banach space X and x ∈ X. Then
x has a best approximation in C, i.e., PC (x) = ∅.

Proof. The function f : C → R+ defined by

f (y) = x − y, y∈C

is obviously lower semicontinuous. Because C is weakly compact, we can

apply Theorem 2.5.5, and then there exists y0 ∈ C such that x − y0  =
inf x − y.
y∈C

Corollary 2.10.3 Let C be a nonempty closed convex subset of a reflexive

Banach space X. Then each element x ∈ X has a best approximation in C.

Theorem 2.10.4 (The uniqueness of best approximations) – Let C be

a nonempty convex subset of a strictly convex Banach space X. Then for each
element x ∈ X, C has at most one best approximation.

Proof. Suppose, for contradiction, that y1 , y2 ∈ C are best approximations

to x ∈ X. Because the set of best approximations is convex, it follows that
(y1 + y2 )/2 is also a best approximation to x. Set r := d(x, C). Then

0 ≤ r = x − y1  = x − y2  = x − (y1 + y2 )/2,
2.10. Metric projection and retraction mappings 117

and it follows that

(x − y1 ) + (x − y2 ) = 2r = x − y1  + x − y2 .

By the strict convexity of X we have

x − y1 = t(x − y2 ), t ≥ 0.

Taking the norm in this relation, we obtain r = tr, i.e., t = 1, which gives
us y1 = y2 .

The following example shows that the strict convexity cannot be dropped in
Theorem 2.10.4.
Example 2.10.5 Let X = R2 with norm  · 1 . It is easy to check that X is
not strictly convex. Now, let

C = {(x, y) ∈ R2 : (x, y)1 ≤ 1} = {(x, y) ∈ R2 : |x| + |y| ≤ 1}.

Then C is a closed convex set. The distance from z = (−1, −1) to the set C is
one, and this distance is realized by more than one point of C.
In Theorem 2.10.4, uniqueness of best approximations need not be true for
nonconvex sets.
Example 2.10.6 Let X = R2 with the norm  · 2 and C = SX = {(x, y)} ∈
R2 : x2 +y 2 = 1}. Then X is strictly convex and C is a nonconvex set. However,
all points of C are best approximations to (0, 0) ∈ X.

Theorem 2.10.7 If in a Banach space X, every element possesses at most a

best approximation with respect to every convex set, then X is strictly convex.

Proof. Suppose, for contradiction, that X is not strictly convex. Then there
exist x, y ∈ X, x = y with

x = y = (x + y)/2 = 1.

Furthermore,
tx + (1 − t)y = 1 for all t ∈ [0, 1].
Set C := co({x, y}). Then 0 − z = d(0, C) for all z ∈ C. This means that
every element of C is the best approximation to zero and this clearly contradicts
the uniqueness.

From Corollary 2.10.3 and Theorem 2.10.4 (see also Proposition 2.1.10), we
obtain some important results:
Theorem 2.10.8 Let C be a nonempty weakly compact convex subset of a
strictly convex Banach space X. Then for each x ∈ X, C has the unique best
approximation, i.e., PC (·) is a single-valued metric projection mapping from X
onto C.
118 2. Convexity, Smoothness, Duality Mappings

Corollary 2.10.9 Let C be a nonempty closed convex subset of a strictly convex

reflexive (e.g., uniformly convex) Banach space X and let x ∈ X. Then there
exists a unique element x0 ∈ C such that x − x0  = d(x, C).

Observation
• Every closed convex subset C of a reflexive Banach space is proximinal.
• Every closed convex subset C of a reflexive strictly convex Banach is a Chebyshev
set.
• For every Chebyshev set C, we have
(i) PC (x) is singleton set, i.e., PC is a function from X onto C.
(ii) x − PC (x) = d(x, C) for all x ∈ X.

We now study useful properties of metric projection mappings.

Theorem 2.10.10 Let C be a subset of a normed space X and x ∈ X. Then

PC (x) ⊆ ∂C.

Proof. Let y ∈ PC (x). Suppose y ∈ int(C). Then there exists an ε > 0 such
that Bε (y) ⊂ C. For each n ∈ N, let zn = (1/n)x + (1 − 1/n)y. Then

zn − y = (1/n)x − y.

For sufficiently large N ∈ N, zN − y < ε. Thus, zN ∈ Bε (y) ⊂ C. On the

other hand,

x − zN  = (1 − 1/N )x − y < x − y = d(x, C),

which contradicts the fact that y ∈ PC (x). Therefore, y ∈ ∂C.

Corollary 2.10.11 Let C be a nonempty closed convex subset of a strictly con-
vex reflexive Banach space X and let x ∈ X. Then we have the following:
(a) If x ∈ C, then PC (x) = x.
(b) If x ∈
/ C, then PC (x) ∈ ∂C.

Theorem 2.10.12 Let C be a nonempty closed convex subset of a reflexive

strictly convex Banach space X. If X has the Kadec-Klee property, then the
projection mapping PC of X onto C is continuous.

Proof. Suppose, for contradiction, that PC is not continuous. Then for the
sequence {xn } in X with lim xn = x ∈ X, there exists ε > 0 such that
n→∞

PC (xn ) − PC (x) ≥ ε for all n ∈ N.

Because
|d(xn , C) − d(x, C)| ≤ xn − x,
it follows that

xn − PC (xn ) − x − PC (x) ≤ xn − x.

2.10. Metric projection and retraction mappings 119

This implies that

lim xn − PC (xn ) = x − PC (x). (2.33)

n→∞

Because {PC (xn )} is bounded in C by (2.33), there exists a subsequence

{PC (xni )} of {PC (xn )} such that w − lim PC (xni ) = z ∈ C. Note
i→∞

w − lim (xni − PC (xni )) = x − z. (2.34)

i→∞

By w-lsc of the functional  · , we have

x − z ≤ lim inf xni − PC (xni ) = x − PC (x).

i→∞

This implies z = PC (x) by definition of the function PC . From (2.33) and (2.34)

w − lim (xni − PC (xni )) = x − PC (x) and lim xni − PC (xni ) = x − PC (x).

i→∞ i→∞

Because X has the Kadec-Klee property, we obtain

lim (xni − PC (xni )) = x − PC (x),

i→∞

which implies that lim PC (xni ) = PC (x), which is a contradiction to the

i→∞
assumption that PC (xn ) − PC (x) ≥ ε.
Then following Proposition 2.5.25, we have

Theorem 2.10.13 Let C be a nonempty convex subset of a smooth Banach

space X and let x ∈ X and y ∈ C. Then the following are equivalent:
(a) y is a best approximation to x: x − y = d(x, C).
(b) y is a solution of the variational inequality:

y − z, Jμ (x − y) ≥ 0 f or all z ∈ C,

where Jμ is a duality mapping with gauge function μ.

As an immediate consequence of Theorem 2.10.13, we have

Corollary 2.10.14 Let C be a nonempty convex subset of a Hilbert space H

and PC be the metric projection mapping from H onto C. Let x be an element
in H. Then the following are equivalent:
(a) x − PC (x) = d(x, C).
(b) x − PC (x), PC (x) − z ≥ 0 f or all z ∈ C.

Proposition 2.10.15 Let C be a nonempty closed convex subset of a Hilbert

space X and PC the metric projection from X onto C. Then the following hold:
(a) PC is “idempotent”: PC (PC (x)) = PC (x) f or all x ∈ X.
120 2. Convexity, Smoothness, Duality Mappings

(b) PC is “firmly nonexpansive”:

x − y, PC (x) − PC (y) ≥ PC (x) − PC (y))2 f or all x, y ∈ X.

(c) PC is “nonexpansive”: PC (x) − PC (y) ≤ x − y f or all x, y ∈ X.

(d) PC is “monotone”: PC (x) − PC (y), x − y ≥ 0 f or all x, y ∈ X.
(e) PC is “demiclosed”: xn  x0 and PC (xn ) → y0 ⇒ PC (x0 ) = y0 .

Proof. (a) Observe that PC (x) ∈ C for all x ∈ X and PC (z) = z for all z ∈ C.
Then PC (PC (x)) = PC (x) for all x ∈ X, i.e., PC2 = PC .

(b) Set j := PC (x) − PC (y) for x, y ∈ X. We have

x − y, j = x − PC (x), j + j, j + PC (y) − y, j.

Because from Corollary 2.10.14, we get

x − PC (x), j ≥ 0 and y − PC (y), j ≥ 0,

it follows that
x − y, j ≥ j2 .
(c) This is an immediate consequence of (b).

(d) It follows from (b).

(e) From Corollary 2.10.14, we have

xn − PC (xn ), PC (xn ) − z ≥ 0 for all z ∈ C.

Because xn  x0 and PC (xn ) → y0 , we have

x0 − y0 , y0 − z ≥ 0 for all z ∈ C.

Using Theorem 2.10.13, we obtain x0 − y0  = d(x0 , C). Therefore,

PC (x0 ) = y0 .

Remark 2.10.16 Proposition 2.10.15(c) shows that in a Hilbert space, a metric

projection operator is not only continuous, but also it is Lipschitz continuous and
hence it is uniformly continuous.

The following result is of fundamental importance. It shows that every point

on line segment joining x ∈ X to its best approximation PC (x) ∈ C has PC (x)
as its best approximation.

Proposition 2.10.17 Let C be a Chebyshev set in a Hilbert space H and

x ∈ H. Then PC (x) = PC (y) for all y ∈ co({x, PC (x)}).
2.10. Metric projection and retraction mappings 121

Proof. Suppose, for contradiction, that there exist y ∈ co({x, PC (x)}) and
z ∈ C such that
y − z < y − PC (x).
Set y := λx + (1 − λ)PC (x) for some λ ∈ (0, 1). Then
x − z ≤ x − y + y − z
< x − y + y − PC (x)
= (1 − λ)x − PC (x) + λx − PC (x) = d(x, C),

a contradiction.

If C is a Chebyshev set in a Hilbert space H, then

PC [λx + (1 − λ)PC (x)] = PC (x), x ∈ H, 0 ≤ λ ≤ 1.
Motivated by this fact, we introduce the following:
A Chebyshev subset C of a normed space X is said to be sun if
PC [λx + (1 − λ)PC (x)] = PC (x) for all x ∈ X and λ ≥ 0.
In other words, C is a sun if and only if each point on the ray from PC (x)
through x also has PC (x) as its best approximation in C.

Let C be a nonempty subset of a topological space X and D a nonempty

subset of C. Then a continuous mapping P : C → D is said to be a retraction
if P x = x for all x ∈ D, i.e., P 2 = P . In such case, D is said to be a retract
of C.
Example 2.10.18 Every closed convex subset C of Rn is a retract of Rn .
We have seen in Theorem 2.10.8 that for every weakly compact convex subset
C of a strictly convex Banach space, there exists a metric projection mapping
PC : X → C that may not be continuous. However, every single-valued metric
projection mapping is a retraction if it is continuous.
Theorem 2.10.19 Every closed convex subset C of a uniformly convex Banach
space X is a retract of X.
Proof. By Theorem 2.10.8, there exists a metric projection mapping PC : X →
C such that PC (x) = x for all x ∈ C. By Theorem 2.10.12, PC is continuous.
Therefore, PC is retraction.

We now show that every retraction P with condition (2.35) is sunny non-
expansive (and hence continuous).
Proposition 2.10.20 Let C be a nonempty convex subset of a smooth Banach
space X and D a nonempty subset of C. If P is a retraction of C onto D such
that
x − P x, J(y − P x) ≤ 0 for all x ∈ C and y ∈ D, (2.35)
then P is sunny nonexpansive.
122 2. Convexity, Smoothness, Duality Mappings

Proof. P is sunny: For x ∈ C, set xt := P x + t(x − P x) for all t > 0. Because

C is convex, it follows that xt ∈ C for all t ∈ (0, 1]. Hence

x − P x, J(P x − P xt ) ≥ 0 and xt − P xt , J(P xt − P x) ≥ 0. (2.36)

Because xt − P x = t(x − P x) and t(x − P x), J(P x − P xt ) ≥ 0, we have

xt − P x, J(P x − P xt ) ≥ 0. (2.37)

Combining (2.36) and (2.37), we get

P x − P xt 2 = P x − xt + xt − P xt , J(P x − P xt )
≤ −xt − P x, J(P x − P xt ) + xt − P xt , J(P x − P xt )
≤ 0.

Thus, P x = P xt . Therefore, P is sunny.

P is nonexpansive : For x, z ∈ C, we have from (2.35) that

x − P x, J(P x − P z) ≥ 0 and z − P z, J(P z − P x) ≥ 0.

Hence
x − z − (P x − P z), J(P x − P z) ≥ 0.
This implies that
x − z, J(P x − P z) ≥ P x − P z2

and hence P is nonexpansive.

We now give equivalent formulations of sunny nonexpansive retraction
mappings.
Proposition 2.10.21 Let C be a nonempty convex subset of a smooth Banach
space X, D a nonempty subset of C, and P : C → D a retraction. Then the
following are equivalent:
(a) P is the sunny nonexpansive.
(b) x − P x, J(y − P x) ≤ 0 for all x ∈ C and y ∈ D.
(c) x − y, J(P x − P y) ≥ P x − P y2 for all x, y ∈ C.
Proof. (a)⇒ (b). Let P be the sunny nonexpansive retraction and x ∈ C.
Then P x ∈ D and there exists a point z ∈ D such that P x = z. Set M :=
{z + t(x − z) : t ≥ 0}. Then M is nonempty convex set. Hence for v ∈ M

y − z = P y − P v (as P is sunny, i.e., P v = z)

≤ y − v = y − z + t(z − x) for all y ∈ D.

Hence from Proposition 2.4.7, we have

x − P x, J(y − P x) ≤ 0.
2.10. Metric projection and retraction mappings 123

(b) ⇒ (a). It follows from Proposition 2.10.20.

(b) ⇒ (c). Let x, y ∈ C. Then P x, P y ∈ D and hence from (b), we have

x − P x, J(P y − P x) ≤ 0 and y − P y, J(P x − P y) ≤ 0.

Combining the above inequalities, we get

P x − P y − (x − y), J(P x − P y) ≤ 0.

Hence

P x − P y2 = P x − P y, J(P x − P y)

= P x − P y − (x − y), J(P x − P y) + x − y, J(P x − P y)
≤ x − y, J(P x − P y).

(c) ⇒ (b). Suppose (c) holds. Let x ∈ C and y ∈ D. Replacing y by y = P y in

(c), we have
x − P y, J(P x − P 2 y) ≥ P x − P 2 y2 ,

which implies that

x − y, J(P x − y) ≥ P x − y2 .

Therefore,

x − P x, J(P x − y) = x − y, J(P x − y) + y − P x, J(P x − y)

≥ P x − y2 − P x − y2 = 0.

Finally, we give uniqueness of sunny nonexpansive retraction mappings.

Proposition 2.10.22 Let C be a nonempty convex subset of a smooth Banach

space X and D a nonempty subset of C. If P is a sunny nonexpansive retraction
from C onto D, then P is unique.

Proof. Let Q be another sunny nonexpansive retraction from C onto D. Then,

we have, for each x ∈ C

x − P x, J(y − P x) ≤ 0 and x − Qx, J(y − Qx) ≤ 0 for all y ∈ D.

In particular, because P x and Qx are in D, we have

x − P x, J(Qx − P x) ≤ 0 and x − Qx, J(P x − Qx) ≤ 0,

which imply that P x − Qx2 ≤ 0. Therefore, P x = Qx for all x ∈ C.

124 2. Convexity, Smoothness, Duality Mappings

Bibliographic Notes and Remarks

The results of Sections 2.1∼2.3 are well-known. These results are adapted from
Cioranescu [41], Goebel and Kirk [59], Goebel and Reich [60], Istratescu [73],
Martin [106], and Prus [121].
The results in Sections 2.4∼2.8 are based on Barbu [11], Goebel and Kirk
[59], Prus [121], and Showalter [145]. Theorem 2.4.16 follows from Kaczor [82].
Theorem 2.6.11 is proved in Reich [127]. The interested reader should also
consult the details on duality mappings in Cioranescu [41] and its review by
Reich [131].
Some results in Section 2.9 are based on some results of Ha and Jung [65],
Jung and Park [80], and Shioji and Takahashi [144]. Theorem 2.9.11 is a corner-
stone of the “Optimization Method” expounded in Reich [126, 128]. The unique-
ness of the minimizer is shown in the paper by Reich [130].
The results in Section 2.10 are adapted from several sources (Goebel and
Kirk [59], Goebel and Reich [60], and Takahashi [155]). Theorem 2.10.12 is an
important improvement of Proposition 3.2 of Martin [106]. Proposition 2.10.21
was first proved in Reich [123], where the term “sunny nonexpansive retraction”
was coined. More updated information on (sunny) nonexpansive retractions can
be found in Reich and Kopecká [132].

Exercises

2.1 Let X be a strictly convex Banach space and let x, y ∈ X with x = y.

If x − z = x − w, z − y = w − y and x − y = x − z + z − y,
show that z = w.

2.2 Let X be a uniformly convex Banach space and let δX be the modulus of
convexity of X. Let 0 < ε < r ≤ 2R. Show that δX (ε/R) > 0 and
  
ε
λx + (1 − λ)y ≤ r 1 − 2λ(1 − λ)δX
R

for all x, y ∈ X with x ≤ r, y ≤ r and x − y ≥ ε and λ ∈ [0, 1].

2.3 Let X be a Banach space. Show that X is uniformly convex if and only if
γ(t) > 0 for all t ∈ (0, 2], where

γ(t) = inf{x − y, x∗ − y ∗  : x, y ∈ SX , x − y ≥ t, x∗ ∈ J(x), y ∗ ∈ J(y)}.

2.4 If 1 < p < ∞, and if the Xn s are all strictly convex Banach spaces, show
that

( Xn )p = {x = {xn } : xn ∈ Xn for all n ∈ N and xn pxn < ∞}
n∈N n∈N
2.10. Metric projection and retraction mappings 125

endowed with norm 

x = ( xn pxn )1/p
n∈N

is strictly convex.

2.5 On L2 ([0, 1], dt), we consider the norm

 1/2
1
f  = (f 22 + f 21 ) .
2

Show that this norm is equivalent to  · 2 , but is not smooth.

x = (x1 + x2 ) , x = {xn }n∈N (where

2 2 1/2
2.6 On 1 , we
consider the norm
x1 = |xn |, x2 = ( |xn |2 )1/2 ).
n∈N n∈N
Show that this norm is equivalent to the 1 -norm and that it is strictly
convex.

2.7 Let C be a nonempty closed convex subset of a strictly convex Banach

space X and D a nonempty subset of C. Let x ∈ C and P be a sunny
nonexpansive retraction of C onto D such that P x − y = x − y for
some y ∈ D. Then P x = x.
Chapter 3

Geometric Coefficients of
Banach Spaces

Geometric coefficients play a key role in the existence of fixed points of

Lipschitzian as well as non-Lipschitzian mappings. In this chapter, we discuss
normal structure coefficient, weak normal structure coefficient, Maluta con-
stants, and other related coefficients.

3.1 Asymptotic centers and asymptotic radius

The concept of asymptotic center is introduced, and several useful results are
discussed here.

Let C be a nonempty subset of a Banach space X and {xn } a bounded

sequence in X. Consider the functional ra (·, {xn }) : X → R+ defined by
ra (x, {xn }) = lim sup xn − x, x ∈ X.
n→∞

The infimum of ra (·, {xn }) over C is said to be the asymptotic radius of {xn }
with respect to C and is denoted by ra (C, {xn }). A point z ∈ C is said to be
an asymptotic center of the sequence {xn } with respect to C if
ra (z, {xn }) = inf{ra (x, {xn }) : x ∈ C}.
The set of all asymptotic centers of {xn } with respect to C is denoted by
Za (C, {xn }). This set may be empty, a singleton, or certain infinitely many
points. In fact, if {xn } converges strongly to x ∈ C, then
Za (C, {xn }) = {x}
and if {xn } converges strongly to x and x ∈
/ C, then
ra (C, {xn }) = d(x, C) and Za (C, {xn }) = {y ∈ C : x − y = d(x, C)} = PC (x),
where PC is the metric projection from X into 2C .

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 3,

c Springer Science+Business Media, LLC, 2009
128 3. Geometric Coefficients of Banach Spaces

For any λ ≥ 0, the level set is

Aλ (C, {xn }) = {x ∈ C : ra (x, {xn }) ≤ ra (C, {xn }) + λ}.

It can be easily observed that

(i) A0 (C, {xn }) = Za (C, {xn }), the asymptotic center of {xn } with respect
to C,
(ii) Aλ (C, {xn }) = ∅ for all λ > 0,
(iii) Aλ (C, {xn }) ⊂ Aλ (C, {xn }) if λ < λ,
(iv) Za (C, {xn }) = A0 (C, {xn }) = ∩λ>0 Aλ (C, {xn }) may be empty,

(v) Za (C, {xn }) = ∩λ>0 {∪∞ ∞

n=1 ∩i=n Bra (C,{xn })+λ [xi ]} ∩ C. (3.1)

A bounded sequence {xn } in a Banach space X is said to be regular with

respect to a bounded subset C of X if the asymptotic radii (with respect to C)
of all subsequences of {xn } are the same, i.e.,

ra (C, {xni }) = ra (C, {xn }) for each subsequence {xni } of {xn }.

A regular sequence {xn } in X is said to be asymptotically uniform with respect

to C if Za (C, {xni }) = Za (C, {xn }) for each subsequence {xni } of {xn }.

Observation
• For an arbitrary subsequence {xni } of {xn }, the following fact always holds:

Za (C, {xni }) ⊇ Za (C, {xn }).

• Asymptotic radius: ra (C, {xn }) = inf{ra (x, {xn }) : x ∈ C}.

• Asymptotic center: Za (C, {xn }) = {z ∈ C : ra (z, {xn }) = ra (C, {xn })}.
• For x ∈ X, ra (x, {xn }) = 0 ⇔ lim xn = x.
n→∞

• ra (αx + βy, {xn }) ≤ αra (x, {xn }) + βra (y, {xn }) for all x, y ∈ X and α, β ≥ 0
with α + β = 1.
• |ra (x, {xn }) − ra (y, {xn })| ≤ x − y ≤ ra (x, {xn }) + ra (y, {xn }) for all x, y ∈ X.
• ra (·, {xn }) is convex and nonexpansive (and hence continuous).
• ϕ(·) = ra (·, {xn }) : X → R+ is w-lsc. Indeed, by the continuity of ϕ(·),
ϕ−1 ((−∞, α]) is closed for every α ∈ R. Also convexity of ϕ(·) implies that
ϕ−1 ((−∞, α]) is convex. Thus, ϕ−1 ((−∞, α]) is weakly closed.

We first establish two preliminary results:

Proposition 3.1.1 Let C be a nonempty bounded subset of a Banach space X

and {xn } a bounded sequence in X. Then {xn } has a subsequence that is regular
with respect to C.
3.1. Asymptotic centers and asymptotic radius 129

Proof. Set

r0 := inf{ra (C, {xni }) : {xni } is a subsequence of {xn }}.

Then there is a subsequence {xnk(1) } of {xn } such that

ra (C, {xnk(1) }) ≤ r0 + 1.

Set
r1 := inf{ra (C, {xnk(1)j })},

where the infimum is taken over all subsequences {xnk(1)j } of {xnk(1) }. Let
{xnk(i) } be a subsequence of {xnk(i−1) } and set

ri := inf{ra (C, {xnk(i)j })},

where the infimum is taken over all subsequences {xnk(i)j } of {xnk(i) }. Select a
subsequence {xnk(i+1) } of {xnk(i) } such that

1
ra (C, {xnk(i+1) }) < ri + . (3.2)
i+1
Because r1 ≤ r2 ≤ r3 ≤ · · · and {ri } is bounded above, it follows that lim ri
i→∞
exists (say r). Then from (3.2),

lim ra (C, {xnk(i+1) }) = r.

i→∞

Now consider the diagonal sequence {xnk(k) } and r = ra (C, {xnk(k) }). Because
{xnk(k) } is a subsequence of {xnk(i) }, it follows that ri ≤ r.
Moreover, from (3.2) we have

1
r ≤ ra (C, {xnk(i+1) }) < ri + ,
i+1
which implies that
r ≤ r.
Hence r = r. Note any subsequence {yn } of {xnk(k) } satisfies the following:

{yn } ⊆ {xnk(i) } and {yn } ⊆ {xnk(i+1) }.

Hence ra (C, {yn }) = r, and we conclude that {xnk(k) } is regular with respect
to C.

Proposition 3.1.2 Let C be a separable bounded subset of a Banach space X

and {xn } a bounded sequence in X. Then {xn } has an asymptotically uniform
subsequence.
130 3. Geometric Coefficients of Banach Spaces

Proof. By Proposition 3.1.1, {xn } has a subsequence that is regular with

respect to C. Because C is separable, a routine diagonalization argument can
be used to obtain a subsequence of {xn }, which we again denote by {xn } such
that lim xn − y exists for all y ∈ C. Clearly, such a sequence must be asymp-
n→∞
totically uniform.

We now discuss the existence and uniqueness of asymptotic centers of bounded

sequences.
Theorem 3.1.3 (The existence of of asymptotic centers) – Let {xn } be
a bounded sequence in a Banach space X and C a nonempty subset of X. Then
we have the following:
(a) If C is weakly compact, then Za (C, {xn }) is nonempty.
(b) If C is weakly compact and convex, then Za (C, {xn }) is a nonempty convex
set.
Proof. (a) From (3.1), we obtain that Za (C, {xn }) can be characterized as the
intersection of a decreasing family of weakly closed sets. Thus, Za (C, {xn }) is
nonempty.
(b) Because C is weakly compact convex set and the function ra (·, {xn })
is continuous, it follows from Theorem 2.5.5 that Za (C, {xn }) = {x ∈ C :
ra (x, {xn }) = inf ra (z, {xn })} is nonempty. Also Za (C, {xn }) is convex.
z∈C
Indeed, for x, y ∈ Za (C, {xn }) and t ∈ [0, 1], we have

ra ((1 − t)x + ty, {xn }) = lim sup (1 − t)x + ty − xn 

n→∞
≤ (1 − t) lim sup xn − x + t lim sup xn − y
n→∞ n→∞
= (1 − t)ra (C, {xn }) + tra (C, {xn }) = ra (C, {xn }),

i.e., (1 − t)x + ty ∈ Za (C, {xn }).

Theorem 3.1.4 Let C be a nonempty closed convex subset of a Banach space
X and {xn } a bounded sequence in X. Then

diam(Za (C, {xn })) ≤ 0 (X) ra (C, {xn }).

Proof. Set d = diam(Za (C, {xn })). If Za (C, {xn }) is empty or a singleton,
then we are done. So, we may assume that d > 0. Let 0 < r < d and x, y ∈
Za (C, {xn }) with x − y ≥ d − r. By the convexity of Za (C, {xn }), (x + y)/2 =
z ∈ Za (C, {xn }). Then from the property of modulus of convexity (see Coroll-
ary 2.3.11),

ra (C, {xn }) = ra (z, {xn }) = lim sup xn − (x + y)/2

n→∞
  
d−r
≤ 1 − δX ra (C, {xn }),
ra (C, {xn })
3.1. Asymptotic centers and asymptotic radius 131

so it follows that  
d−r
δX ≤ 0.
ra (C, {xn })
By the definition of 0 (X),
d − r ≤ 0 (X)ra (C, {xn }).

Because r > 0 is arbitrary, it follows that d ≤ 0 (X) ra (C, {xn }).

Using Theorem 3.1.4, we obtain

Theorem 3.1.5 (The uniqueness of asymptotic centers) – Let C be a
nonempty closed convex subset of a uniformly convex Banach space X and {xn }
a bounded sequence in X. Then Za (C, {xn }) is a singleton set.
Proof. Because ra (·, {xn }) is a continuous and convex functional and ra
(z, {xn }) → ∞ as z → ∞, by Theorem 2.5.8, we obtain that Za (C, {xn }) = ∅.
By the uniform convexity of X, 0 (X) = 0, it follows from Theorem 3.1.4 that
diam(Za (C, {xn }) = 0, i.e., Za (C, {xn }) is a singleton set.

The following theorem shows that the asymptotic center enjoys an interesting
inequality.
Theorem 3.1.6 Let C be a nonempty closed convex subset of a uniformly
convex Banach space. Then every bounded sequence {xn } in X has a unique
asymptotic center with respect to C, i.e., Za (C, {xn }) = {z} and
lim sup xn − z < lim sup xn − x f or x = z.
n→∞ n→∞

Proof. The result follows from Theorem 3.1.5.

Observation
• If C is weakly compact, then Za (C, {xn }) is nonempty.
• If C is closed, then Za (C, {xn }) is closed.
• If C is convex, then Za (C, {xn }) is convex.
• Za (C, {xn }) ⊂ ∂C ∪ Za (X, {xn }).

We now give the following result, which is very useful in the study of multi-
valued mappings in Banach spaces.

Proposition 3.1.7 Let C be a nonempty closed convex bounded subset of a

uniformly convex Banach space X and {xn } a sequence in C with asymptotic
center z and asymptotic radius r. For t ∈ (0, 1), let zn = (1 − t)z + txn , n ∈ N.
Then Za (C, {zn }) = z and ra (C, {zn }) = tr.

Proof. Suppose, for contradiction, that Za (C, {zn }) = v = z. Because

zn − z = txn − z for all n ∈ N, (3.3)

132 3. Geometric Coefficients of Banach Spaces

it follows that

ra (C, {zn }) = inf{lim sup zn − w : w ∈ C} ≤ tr.

n→∞

Let ra (C, {zn }) = r . Because the asymptotic center v of {zn } is unique, hence
from Theorem 3.1.6, we have

r = lim sup zn − v < lim sup zn − z = tr.

n→∞ n→∞

For each n ∈ N, we have

xn − v = v − (1 − t)z − txn + (1 − t)z − (1 − t)xn 

≤ v − [(1 − t)z + txn ] + (1 − t)xn − z
= zn − v + (1 − t)xn − z,

which implies that

lim sup xn − v ≤ r + (1 − t)r < r

n→∞

contradicting ra (C, {xn }) = r. Thus, Za (C, {zn }) = z and from (3.3), we have
ra (C, {zn }) = tr.

We present the following result, which has important applications in the

study of fixed point theory of nonlinear mappings:

Theorem 3.1.8 Let C be a nonempty closed convex subset of a uniformly con-

vex Banach space X and {xn } a bounded sequence in C such that Za (C, {xn }) =
{z}. If {ym } is a sequence in C such that lim ra (ym , {xn }) = ra (C, {xn }), then
m→∞
lim ym = z.
m→∞

Proof. Suppose, for contradiction, that {ym } does not converge strongly to z.
Then there exists a subsequence {ymi } of {ym } such that

ymi − z ≥ d > 0 for all i ∈ N.

By the uniform convexity of X, there exists ε > 0 such that

  
d
(ra (C, {xn }) + ε) 1 − δX < ra (C, {xn }).
ra (C, {xn }) + ε

Because ra (z, {xn }) = ra (C, {xn }), there exists n0 ∈ N such that

xn − z ≤ ra (C, {xn }) + ε for all n ≥ n0 .

Because ra (ym , {xn }) → ra (C, {xn }) as m → ∞ and hence ra (ymi , {xn }) →

ra (C, {xn }) as i → ∞, then there exists an integer n0 ∈ N such that

xn − ymi  ≤ ra (C, {xn }) + ε for all n ≥ n0 .

3.1. Asymptotic centers and asymptotic radius 133

Because X is uniformly convex,

    
 
xn − z + ymi  ≤ 1−δX d
(ra (C, {xn })+ε) < ra (C, {xn })
 2  (ra (C, {xn }) + ε)
for all n ≥ max{n0 , n0 }. This implies that
 
z + y mi
ra , {xn } < ra (C, {xn }),
2

which contradicts the uniqueness of the asymptotic center z.

Let C be a nonempty subset of a Banach space X. For x ∈ C, the inward

set of x relative to C is the set
IC (x) = {(1 − t)x + ty : y ∈ C and t ≥ 0}.

Geometrically, it is the union of all rays beginning at x and passing through

other points of C. Let IC (x) denote the closure of IC (x). Then we have
Proposition 3.1.9 Let C be a nonempty closed convex subset of a Banach
space X. If w ∈ IC (x), then (1 − t)x + tw ∈ IC (x) for all t > 0.
Proof. Because w ∈ IC (x), then there exists a sequence {wn } in IC (x) such
that
w = lim wn and wn = (1 − cn )x + cn yn , yn ∈ C, cn ≥ 0.
n→∞
For t > 0, we have
(1 − t)x + twn = (1 − t)x + t[(1 − cn )x + cn yn ]
= (1 − tcn )x + tcn yn ∈ IC (x),
which implies that
(1 − t)x + tw ∈ IC (x).

Proposition 3.1.10 Let C be a nonempty closed convex subset of a Banach

space X. Then the following are equivalent:
(a) w ∈ IC (x).
(b) There exists b > 0 such that (1 − b)x + bw ∈ IC (x).
(c) (1 − b)x + bw ∈ IC (x) for all b > 0.
Proof. (a) ⇒ (c). It follows from Proposition 3.1.9.
(c) ⇒ (b). It is obvious.
(b) ⇒ (a). Suppose that there exists b > 0 such that (1 − b)x + bw ∈ IC (x).
By Proposition 3.1.9,
(1 − a)x + a((1 − b)x + bw) ∈ IC (x) for all a > 0.

Taking a = 1/b, we have w ∈ IC (x).

134 3. Geometric Coefficients of Banach Spaces

Proposition 3.1.11 Let C be a convex subset of a normed linear space X.

Then x − y ∈ IC (x) if and only if

lim d(x − hy, C)/h = 0. (3.4)

h→0+

Proof. Suppose that (3.4) holds. Let x ∈ C. Let ε > 0 be given. Then there
exists b ∈ (0, 1) such that
ε
b−1 d(x − by, C) < .
2
By the definition of distance, there exists u ∈ C such that
bε
x − by − u < d(x − by, C) + .
2
Observe that x + b−1 (u − x) ∈ IC (x). Because

[x + b−1 (u − x)] − (x − y) = b−1 u − (x − by)

 
−1 bε
< b d(x − by, C) +
2
< ε,

it follows that x − y is in the closure of IC (x).

Conversely, suppose that x − y ∈ IC (x). Then there exists a sequence {xn }
in IC (x) such that xn → x − y. Let ε > 0 be given. Then there exists n0 ∈ N
such that
xn − (x − y) < ε for all n ≥ n0 .
Observe that
h−1 d(x − hy, C) ≤ h−1 x − hy − [(1 − h)x + hxn0 ] + h−1 d((1 − h)x + hxn0 , C).
Because xn0 ∈ IC (x) and C is convex, there exists h0 > 0 such that (1 − h0 )x +
h0 xn0 ∈ C. Thus, if 0 < h ≤ h0 , h−1 d((1 − h)x + hxn0 , C) = 0 and

h−1 d(x − hy, C) ≤ h−1 x − hy − [(1 − h)x + hxn0 ]

= x − y − xn0  < ε.

Therefore, lim d(x − hy, C)/h = 0.

h→0+

Proposition 3.1.12 Let C be a nonempty closed convex subset of a uniformly

convex Banach space X and {xn } a bounded sequence in C. If z is the asymp-
totic center of {xn } with respect to C, then it is also the asymptotic center with
respect to IC (z).
Proof. Let v be the asymptotic center of {xn } with respect to IC (z). Suppose
that v = z. Because v = z and C ⊆ IC (z), we have v ∈ IC (z) \ C and
ra (v, {xn }) < ra (z, {xn }) by the uniqueness of the asymptotic center
3.1. Asymptotic centers and asymptotic radius 135

(see Theorem 3.1.6). By the continuity of ra (·, {xn }), there exists w ∈ IC (z) \ C
such that ra (w, {xn }) < ra (z, {xn }). Hence w = (1 − t)z + ty for some y ∈ C
and t > 1. Because ra (·, {xn })) is a convex functional,

ra (y, {xn }) ≤ ra (t−1 w + (1 − t−1 )z, {xn })

≤ t−1 ra (w, {xn }) + (1 − t−1 )ra (z, {xn })
< ra (z, {xn }),

a contradiction. Hence v = z.

Proposition 3.1.13 Let C be a nonempty closed convex subset of a reflexive

Banach space X and {xn } a bounded sequence in X. Then there exists a point
v ∈ C such that

(a) ra (v, {xn }) = inf{ra (z, {xn }) : z ∈ IC (v)}.

(b) lim inf x−v, J(xn −v) ≤ 0 for all x ∈ IC (v) if the norm of X is uniformly
n→∞
Gâteaux differentiable.

Proof. (a) Set ϕ(·) = ra (·, {xn }). Observe that ϕ(·) is a continuous and convex
functional and ra (z, {xn }) → ∞ as z → ∞. By Theorem 2.5.8, we obtain
that Za (C, {xn }) = ∅. Let v be an asymptotic center of {xn } with respect to
C. We now show that v is also an asymptotic center of {xn } with respect to
IC (v), i.e.,
ϕ(v) = inf{ϕ(z) : z ∈ IC (v)}.
Set r := inf{ϕ(z) : z ∈ IC (v)}. Suppose that r < ϕ(v). Now for ε > 0,
r+ε < ϕ(v). By the continuity of ϕ, there exists z ∈ IC (v) such that ϕ(z) ≤ r+ε.
Thus,
z = v + t(w − v) for some w ∈ C and t ≥ 1.
By the convexity of ϕ, we obtain

ϕ(v) ≤ ϕ(w) ≤ t−1 ϕ(z) + (1 − t−1 )ϕ(v),

which implies that

ϕ(v) ≤ ϕ(z) ≤ r + ε,

a contradiction. This shows that r = ϕ(v).

(b) For arbitrary y ∈ IC (v) and t > 0, let zt = (1 − t)v + ty. Then zt ∈ IC (v).
By Proposition 2.4.5, we have

2zt − v, J(xn − zt ) ≤ xn − v2 − xn − zt 2 ,

which implies that

2t lim inf y − v, J(xn − zt ) ≤ ϕ2 (v) − ϕ2 (zt ) ≤ 0

n→∞
136 3. Geometric Coefficients of Banach Spaces

and hence
lim inf y − v, J(xn − xt ) ≤ 0.
n→∞
Because X is reflexive with uniformly Gâteaux differential norm, J is uniformly
demicontinuous on bounded subsets of X. Using the above fact and letting
t → 0, we obtain

lim inf y − v, J(xn − v) ≤ 0 for all y ∈ IC (v).

n→∞

Note that when X is separable, by a diagonalization argument, given a

bounded sequence {xn }, there exists a subsequence {xnk } of {xn } such that
lim xnk − z exists for all z ∈ X.
k→∞

In that case, we have the following.

Proposition 3.1.14 Let X be a separable reflexive Banach space with a uni-
formly Gâteaux differentiable norm, C a nonempty closed convex subset of X
and {xn } a bounded sequence in X. Then there exist a point v ∈ C and a
subsequence {xnk } of {xn } such that

lim supx − v, J(xnk − v) ≤ 0 for all x ∈ IC (v).

k→∞

Proof. The result follows from Proposition 3.1.13.

Proposition 3.1.15 Let X be a separable reflexive Banach space with a uni-

formly Gâteaux differentiable norm and {xn } a bounded sequence in X. Then
there exist a point v ∈ X and a subsequence {xnk } of {xn } such that {J(xnk −v)}
converges weakly to zero.

Proof. Now for C = X, IC (v) = X. Hence J(xnk − v)  0 by Proposi-

tion 3.1.14.

3.2 The Opial and uniform Opial conditions

The Opial condition plays an important role in convergence of sequences and
in the study of the demiclosedness principle of nonlinear mappings and the
geometry of Banach spaces.

A Banach space X is said to satisfy the Opial condition if whenever a

sequence {xn } in X converges weakly to x0 ∈ X, then

lim inf xn − x0  < lim inf xn − x for all x ∈ X, x = x0 . (3.5)
n→∞ n→∞

We observe that (3.5) is equivalent to the analogous condition obtained by

replacing lim inf by lim sup. Replacing the strict inequality “ < ” in (3.5)
with “ ≤ , ” we obtain the definition of the so-called non-strict (or weak) Opial
condition.
3.2. The Opial and uniform Opial conditions 137

Example 3.2.1 Every Hilbert space satisfies the Opial condition, i.e., if the
sequence {xn } in a Hilbert space H, converges weakly to x ∈ H, then

lim sup xn − x < lim sup xn − y for all y ∈ H and y = x.
n→∞ n→∞

In fact, because every weakly convergent sequence is necessarily bounded, so we

have that lim sup xn − x and lim sup xn − y are finite. Note
n→∞ n→∞

xn − y2 = xn − x + x − y2 = xn − x2 + x − y2 + 2xn − x, x − y,

so that
lim sup xn − y2 > lim sup xn − x2 .
n→∞ n→∞

The following example shows that Lp [0, 2π], 1 < p < ∞ does not satisfy even
the nonstrict Opial condition for any p = 2.

Example 3.2.2 Let f be a periodic function with period 2π defined by


1 if 0 ≤ t ≤ 4π/3;
f (t) =
−2 if 4π/3 < t ≤ 2π.

Set xn (t) = f (nt), n ∈ N. Then {xn } is a weakly null sequence of Rademacher-

like functions in Lp [0, 2π] for every 1 < p < ∞. Define a function
2π
λp (s) = lim xn − sp = |f (t) − s|p dt,
n→∞ 0

where s ∈ R is treated as the constant function. Note

2π
4π
λ p (0) = −p |f (t)|p−1 sgn(f (t))dt = p(2p−2 − 1),
0 3

λ p (0) = 0 whenever p = 2. It follows that λp (0) is not the minimal value of

λp , except for the case p = 2. Therefore, Lp [0, 2π] does not satisfy even the
nonstrict Opial condition for any p = 2.

We now give necessary and sufficient condition for a space satisfying the
Opial condition.

Proposition 3.2.3 A Banach space X satisfies the Opial condition if and only
if

xn  0 and lim inf xn  = 1 ⇒ lim inf xn − x > 1 for all x = 0. (3.6)
n→∞ n→∞

Proof. Suppose that the condition (3.6) is satisfied. Let un  u and r =

lim inf un − u. If r = 0, then the Opial condition (3.5) follows from the
n→∞
uniqueness of a weak limit.
138 3. Geometric Coefficients of Banach Spaces

If r = lim inf un − u > 0, then xn = r−1 (un − u)  0 and lim inf xn  = 1.
n→∞ n→∞
Hence from (3.6) we have
lim inf xn − x > 1 for x = 0,
n→∞

which implies that

lim inf r−1 (un − u) − x > lim inf r−1 (un − u),
n→∞ n→∞

i.e.,
lim inf un − (u + rx) > lim inf un − u for u = u + rx.
n→∞ n→∞

Hence X satisfies the Opial condition. The inverse implication is obvious.

Proposition 3.2.4 Let X1 , X2 , · · · , Xk be Banach spaces with norm ·1 , ·2 ,
· · · ,  · k , respectively. Let p be a constant in [1, ∞) and put X = X1 × X2 ×
· · · × Xk , where the norm of X is given by
(x1 , x2 , · · · , xk ) = (x1 p2 +x2 p2 +· · ·+xk pk )1/p for all (x1 , x2 , · · · , xk ) ∈ X.
Then the following are equivalent:
(a) X has the Opial condition.
(b) Each Xj has the Opial condition.
Proof. (a) ⇒ (b). Let {xn } be a sequence in Xj such that xn  z. Then a
sequence
 
(0, 0, · · · , 0, xn , 0, · · · , 0)
↑
j th position
in X converges weakly to

(0, 0, · · · , 0, z, 0, · · · , 0).
↑
j th position
Using this fact, one can easily see that (a) implies (b).
Conversely, let {xn } be a sequence in X such that xn  z and let w belong
(1) (2) (k)
to X \ {z}. Set xn := (xn , xn , · · · , xn ) for n ∈ N, z := (z (1) , z (2) , · · · , z (k) )
and w := (w(1) , w(2) , · · · , w(k) ). Because {xn } is a bounded sequence, there
exists a subsequence {xni } of {xn } such that
lim inf xn − wp = lim xni − wp ,
n→∞ i→∞

(j)
and that the limit of {xni − z (j) j } exist for all j = 1, 2, · · · , k. Because Xj
for j, 1 ≤ j ≤ k satisfies the Opial condition,
lim xn(j)
i
− z (j) pj ≤ lim inf xn(j)
i
− w(j) pj
i→∞ n→∞
3.2. The Opial and uniform Opial conditions 139

and
() p () p
lim x()
ni − z  < lim inf x()
ni − w 
i→∞ n→∞

holds for some , 1 ≤ ≤ k because z = w. Therefore, we have

lim inf xn − zp ≤ lim xni − zp

n→∞ i→∞

k
= lim xn(j)
i
− z (j) pj
i→∞
j=1


k
< lim inf xn(j)
i
− w(j) pj
n→∞
j=1
≤ lim xni − wp
i→∞
= lim inf xn − wp .
n→∞

Therefore,
lim inf xn − z < lim inf xn − w.
n→∞ n→∞

The following example shows that Proposition 3.2.4 is not true if the norm
of X is (x1 , x2 , · · · , xk ) = max xi .
1≤i≤k

Example 3.2.5 Let X = R × 2 with the norm

(a, y) = max{|a|, y2 }.

Let en be the nth element of the basis of 2 and {xn = (0, en )} be a sequence in
X. Then we have

xn  = max{0, en 2 } = 1 for all n ∈ N

and xn  0 as n → ∞, but for x = (1, 0), we have

xn − x = (1, en ) = max{1, en 2 } = 1 for all n ∈ N.

Therefore, X does not satisfy the Opial condition, even though R and 2 satisfy
it.

We now consider some classes of Banach spaces that always imply the Opial
condition.

Definition 3.2.6 A Banach space X is said to have a weakly continuous duality

mapping if there exists a gauge function μ such that the duality mapping Jμ (with
gauge function μ) is single-valued and (sequentially) continuous from the weak
topology of X to the weak topology of X ∗ .
140 3. Geometric Coefficients of Banach Spaces

Example 3.2.7 The spaces p (1 < p < ∞) possess duality mappings that are
weakly continuous. To see this, for (x1 , x2 , · · · , xi , · · · ) ∈ p , let

Jμ (x) = (|x1 |p−1 sgn(x1 ), |x2 |p−1 sgn(x2 ), · · · , |xi |p−1 sgn(xi ), · · · )

and μ(t) = tp/q , where 1/p + 1/q = 1. Note

(n) (n) (n)
p  (α1 , α2 , · · · , αi , · · · ) = xn  x = (α1 , α2 , · · · , αi , · · · ) ∈ p

if and only if
(n)
xn p ≤ M for all n ∈ N and αi → αi as n → ∞.

Observe that
(n) (n) (n) (n) (n) (n)
Jμ (xn ) = (|α1 |p−1 sgn(α1 ), |α2 |p−1 sgn(α2 ), · · ·,|αi |p−1 sgn(αi ),· · ·) ∈ q ,

so it follows that

Jμ (xn )q = μ(xn p ) ≤ μ(M ) f or all n ∈ N

and
(n) (n)
|αi |p−1 sgn(αi ) → |αi |p−1 sgn(αi ) as n → ∞.
Thus, we conclude that

q  Jμ (xn )  Jμ (x) ∈ q as n → ∞.

Therefore, Jμ is a weakly continuous mapping from X = p into X ∗ = q.

The following theorem gives an important characterization of Banach spaces

that possess weakly continuous mappings.

Theorem 3.2.8 Let X be a Banach space with a weakly continuous duality

mapping Jμ with function gauge μ. Then we have the following:
(a) If {xn } is a sequence in X such that xn  x, then


lim Φ(xn − y) = 
lim Φ(xn − x) + Φ(x − y) for all y ∈ X, (3.7)
n→∞ n→∞

where 
lim is either lim inf or lim sup.
n→∞ n→∞ n→∞
(b) X has the Opial condition.

Proof. (a) Because Jμ (x) is the Gâteaux derivative of the convex functional
 x
Φ(x) = 0 μ(t)dt, it follows (see Theorem 2.5.23) that
1
Φ(x + y) = Φ(x) + y, Jμ (x + ty)dt for all x, y ∈ X.
0
3.2. The Opial and uniform Opial conditions 141

Let {xn } be a sequence in X such that xn  x and let y be an element in

C. Then Jμ (xn + ty)  Jμ (x + ty), so


lim Φ(xn − y) = 
lim Φ(xn − x + x − y)
n→∞ n→∞

= 
lim Φ(xn − x)
n→∞
1
+
lim x − y, Jμ (xn − x + t(x − y))dt
n→∞ 0
1
= 
lim Φ(xn − x) + x − y, Jμ (t(x − y))dt
n→∞ 0
1
= 
lim Φ(xn − x) + x − y μ(tx − y)dt
n→∞ 0

= 
lim Φ(xn − x) + Φ(x − y).
n→∞

(b) Because Φ is strictly increasing, it follows from (3.7) that X has the
Opial condition.
Observation

• The duality mapping of each Hilbert space (e.g., 2 and Rn ) is the identity
mapping and hence it is weakly continuous. Therefore, every Hilbert space
satisfies the Opial condition (see also Example 3.2.1).

• p (1 < p < ∞) spaces have weakly sequentially continuous duality mappings

(and hence the Opial condition), but the Lp [0, 2π] space (1 < p < ∞, p = 2) fails
to satisfy the Opial condition. It means that the Opial condition is independent
of uniform convexity.

The following Theorem 3.2.9 shows that weak limit of a bounded sequence
is the asymptotic center under some geometric conditions.

Theorem 3.2.9 Let X be a uniformly convex Banach space satisfying the Opial
condition and C a nonempty closed convex subset of X. If {xn } is a sequence
in C such that xn  z, then z is the asymptotic center of {xn } in C.

Proof. From Theorem 3.1.6, Za (C, {xn }) is singleton. Let Za (C, {xn }) =
{u}, u = z. Because xn  z, by the Opial condition,

lim sup xn − z < lim sup xn − u.

n→∞ n→∞

Using again Theorem 3.1.6, we obtain

lim sup xn − u < lim sup xn − z.

n→∞ n→∞

Therefore, z = u.
142 3. Geometric Coefficients of Banach Spaces

Corollary 3.2.10 Let X be a uniformly convex Banach space with a weakly

continuous duality mapping and C a nonempty closed convex subset of X.
If {xn } is a sequence in C such that xn  z, then z is the asymptotic center of
{xn } in C.
Corollary 3.2.11 Let C be a nonempty closed convex subset of a Hilbert space
H. Then the weak limit of a weakly convergent sequence in C coincides with its
asymptotic center with respect to C.
Remark 3.2.12 Corollary 3.2.11 is valid in all sequence spaces p (1 <
p < ∞), but it does not hold in the Lebesgue spaces Lp [0, 2π] (1 < p < ∞, p = 2).
We now introduce notions of uniform Opial condition and locally uniform
Opial condition:
Definition 3.2.13 A Banach space X is said to satisfy the uniform Opial con-
dition if for each t > 0, there exists an r > 0 such that

1 + r ≤ lim inf xn + x

n→∞

for each x ∈ X with x ≥ t and each sequence {xn } in X such that xn  0
and lim inf xn  ≥ 1.
n→∞

It is obvious from Proposition 3.2.3 that the uniform Opial condition implies
the Opial condition.
Definition 3.2.14 A Banach space X is said to satisfy the locally uniform
Opial condition if for any weakly null sequence {xn } in X with lim inf xn  ≥ 1
n→∞
and any t > 0, there is an r > 0 such that

1 + r ≤ lim inf xn + x

n→∞

for every x ∈ X with x ≥ t.

Note that
uniform Opial condition ⇒ locally uniform Opial condition ⇒ Opial condition.

We now define the Opial modulus of X, denoted by rX , as follows:

 
rX (t) = inf lim inf xn + x − 1 ,
n→∞

where t ≥ 0 and the infimum is taken over all x ∈ X with x ≥ t and sequences
{xn } in X such that xn  0 and lim inf xn  ≥ 1.
n→∞

It is obvious that rX is nondecreasing and that X satisfies the uniform Opial

condition if and only if rX (t) > 0 for all t > 0.

Note that in the definition of locally uniform Opial condition, “lim inf ” can
n→∞
be replaced with “lim sup.”
n→∞
3.2. The Opial and uniform Opial conditions 143

Let {xn } be a weakly null sequence in a Banach space X with lim inf
n→∞
xn  ≥ 1. We define the local Opial modulus of X as follows:
 
rX,xn (t) = inf lim inf xn + x − 1 : x ∈ X with x ≥ t .
n→∞

One may easily see that X has the local uniform Opial condition if

rX,xn (t) > 0 for all t > 0

and the uniform Opial condition if

 
rX (t) = inf rX,xn (t) : xn  0 with lim inf xn  ≥ 1 > 0 for all t > 0.
n→∞

We now establish fundamental properties of the Opial modulus.

Proposition 3.2.15 Let X be a Banach space with Opial modulus rX . Then
1 + rX (t1 )
rX (t2 ) − rX (t1 ) ≤ (t2 − t1 ) for all 0 < t1 ≤ t2 ,
t1
i.e., the Opial modulus rX is continuous.
Proof. Let
 
Gt := {xn } in X : xn  0 and lim inf xn  ≥ t .
n→∞

Then for 0 < t1 ≤ t2 , we have

 
1 + rX (t2 ) = inf lim inf xn + x : {xn } ⊂ X, xn  0, lim inf xn  ≥ 1, x ≥ t2
n→∞ n→∞
 
t2
= inf lim inf xn + x : {xn } ∈ G1 , x ≥ t1 .
n→∞ t1

Because t1 /t2 ≤ 1, it follows that

1
1 + rX (t2 ) = inf{lim inf t1 xn + t2 x : {xn } ∈ G1 , x ≥ t1 }
t1 n→∞
t2
= inf{lim inf zn + z : {zn } ∈ G t1 , z ≥ t1 }
t1 n→∞ t2

t2
≤ inf{lim inf zn + z : {zn } ∈ G1 , z ≥ t1 }
t1 n→∞
 
t2
≤ 1 + rX (t1 ) .
t1
The following theorem allows us to estimate the Opial modulus of Banach
spaces.
144 3. Geometric Coefficients of Banach Spaces

Theorem 3.2.16 Let X be a Banach space with a weakly continuous duality

mapping Jμ with gauge function μ. Then rX (t) = Φ−1 (Φ(1) + Φ(t)) − 1 for all
t ≥ 0.
Proof. Let {xn } be a sequence in X such that xn  0. Then we obtain from
(3.7) that
lim inf Φ(xn + y) = lim inf Φ(xn ) + Φ(y) for all y ∈ X. (3.8)
n→∞ n→∞

If lim inf xn  ≥ 1, then

n→∞

Φ(1) + Φ(y) ≤ lim inf Φ(xn ) + Φ(y) = lim inf Φ(xn + y).
n→∞ n→∞

Thus,
Φ−1 (Φ(1) + Φ(y)) ≤ lim inf xn + y for all y ∈ X,
n→∞
and it follows from the definition of rX that
Φ−1 (Φ(1) + Φ(t)) ≤ rX (t) + 1. (3.9)
Now, let xn ∈ SX with xn  0 and x = t. Then from (3.8), we have
1 + rX (t) ≤ lim inf xn + x
n→∞
−1
≤ Φ (Φ(1) + Φ(t)). (3.10)
Combining (3.9) and (3.10), we get

1 + rX (t) = Φ−1 (Φ(1) + Φ(t)) for all t ≥ 0.

Because p (1 < p < ∞) space admits a weakly continuous duality mapping

Jμ with the gauge function μ(t) = tp−1 , we have
Corollary 3.2.17 Let 1 < p < ∞. Then rp (t) = (1 + tp )1/p − 1, t ≥ 0.
Theorem 3.2.18 Let X be a Banach space. Then the following are equivalent:
(a) X has a nonstrict Opial condition.
(b) rX (t) ≥ 0 for all t > 0.
Proof. (a) ⇒(b). Let {xn} be a sequence in X such that xn  0 and lim inf xn  ≥
n→∞
1. Let x ∈ X such that x ≥ t for t > 0. Then
1 ≤ lim inf xn  ≤ lim inf xn + x.
n→∞ n→∞

Hence rX (t) ≥ 0.
(b) ⇒ (a). Let {xn } be a sequence in X such that xn  0 and α :=
lim inf xn  > 0. Let x ∈ X such that x =
 0. If lim inf xn  ≤ x, then
n→∞ n→∞
by w-lsc of the norm, we have
lim inf xn  ≤ lim inf xn + x
n→∞ n→∞

and hence we are done.

3.2. The Opial and uniform Opial conditions 145

If x ≤ lim inf xn , then

n→∞
   
x  xn x
1 + rX 
≤ lim inf  + ,
α n→∞ α α
so
  
x
α 1 + rX ≤ lim inf xn + x. (3.11)
α n→∞

Because x/α > 0, by assumption, we have rX (x/α) ≥ 0. It follows from

(3.11) that
α = lim inf xn  ≤ lim inf xn + x.
n→∞ n→∞

Hence X satisfies a nonstrict Opial condition.

Proposition 3.2.19 A Banach space X satisfies the locally uniform Opial con-
dition if and only if for any sequence {xn } in X that converges weakly to x ∈ X
and for any sequence {ym } in X,

lim sup(lim sup xn − ym ) ≤ lim sup xn − x implies ym → x.

m→∞ n→∞ n→∞

Proof. Assume that X satisfies the locally uniform Opial condition. Let {xn }
be a sequence in X with xn  x ∈ X and {ym } a sequence in X such that

lim sup(lim sup xn − ym ) ≤ lim sup xn − x. (3.12)

m→∞ n→∞ n→∞

Set d := lim sup xn − x. If d = 0, then ym → x. If d > 0, there exists a sub-
n→∞
sequence {xni } of {xn } such that d = lim xni −x. Suppose, for contradiction,
i→∞
that {ym } does not converge to x in norm. Then there exist an ε > 0 and a
subsequence {ymj } of {ym } such that

ymj − x ≥ ε for all j ∈ N.

Set zi := (xni − x)/d. By the local uniform Opial condition of X, we have an

r > 0 such that
ε
1 + r ≤ lim inf zi + z for all z ∈ X with z ≥ .
i→∞ d
In particular, we have

lim inf xni − ymj  ≥ d(1 + r) for all j ∈ N,

i→∞

which gives that

lim sup(lim sup xn − ym ) ≥ d(1 + r) > d = lim sup xn − x,
m→∞ n→∞ n→∞

which contradicts the inequality (3.12).

146 3. Geometric Coefficients of Banach Spaces

Conversely, suppose that X does not satisfy the locally uniform Opial condi-
tion. Then there exist a sequence {xn } in X with xn  0 and lim sup xn  ≥ 1,
n→∞
a constant c > 0, and a sequence {ym } in X with ym  ≥ c for all m ∈ N such
that
1
1+ > lim sup xn − ym  for m ∈ N.
m n→∞

Hence
lim sup(lim sup xn − ym ) ≤ 1 ≤ lim sup xn .
m→∞ n→∞ n→∞

By assumption, we have ym → 0. This contradicts the fact that ym  ≥ c for all
m ∈ N.

3.3 Normal structure

Let C be a nonempty bounded subset of a Banach space X. Then a point
x0 ∈ C is said to be

(i) a diametral point of C if

sup{x0 − x : x ∈ C} = diam(C),

(ii) a nondiametral point of C if

sup{x0 − x : x ∈ C} < diam(C).

A nonempty convex subset C of a Banach space X is said to have normal

structure if each convex bounded subset D of C with at least two points contains
a nondiametral point, i.e., there exists x0 ∈ D such that

sup{x0 − x : x ∈ D} < diam(D).

Geometrically, C is said to have normal structure if for each convex bounded

subset D of C with diam(D) > 0, there exist a point x0 ∈ D and r < diam(D)
such that
D ⊆ Br [x0 ].
The Banach space X is said to have normal structure if every closed convex
bounded subset C of X with diam(C) > 0 has normal structure.

The following theorems state that compact convex subsets of any Banach
space and closed convex bounded subsets of a uniformly convex Banach space
have this geometric property.

Theorem 3.3.1 Every compact convex subset C of a Banach space X has

normal structure.
3.3. Normal structure 147

Proof. Suppose, for contradiction, that C does not have normal structure. Let
D be a convex subset of C that has at least two points. Because C does not have
normal structure, all points of D are diametral. Now we construct a sequence
{xi }∞
i=1 in D such that

xi − xj  = diam(D) for all i, j ∈ N, i = j.

For this, let x1 be an arbitrary point in D. Then there exists a point x2 ∈ D

such that diam(D) = x1 − x2 . Because D is convex, there exists a point
(x1 + x2 )/2 ∈ D. Next we choose a point x3 ∈ D such that
 
 x1 + x2 
diam(D) =  x
 3 − .

2

Proceeding in the same manner, we obtain a sequence {xn } in D such that

 
 x1 + x2 + · · · + xn 

diam(D) = xn+1 − , n ≥ 2.
n 

Because
 
 x1 + x2 + · · · + xn 
diam(D) = x
 n+1 − 

n
 
 (xn+1 − x1 ) + (xn+1 − x2 ) + · · · + (xn+1 − xn ) 
= 



n
1
≤ (xn+1 − x1  + xn+1 − x2  + · · · + xn+1 − xn )
n
≤ diam(D),

it follows that diam(D) = xn+1 − xi , 1 ≤ i ≤ n. This implies that the

sequence {xn } has no convergent subsequences. This contradicts the compact-
ness of C.

Corollary 3.3.2 Every finite-dimensional Banach space has normal structure.

Proof. The result easily follows from Theorem 3.3.1

Theorem 3.3.3 Every closed convex bounded subset C of a uniformly convex

Banach space X has normal structure.

Proof. Let D be a closed convex subset of C with diam(D) = d > 0. Let x1

be an arbitrary point in D. Choose a point x2 ∈ D such that x1 − x2  ≥ d/2.
Because D is convex, (x1 + x2 )/2 ∈ D. Set x0 = (x1 + x2 )/2. By the uniform
convexity,
    
u + v
u ≤ r, v ≤ r and u − v ≥ ε > 0 ⇒   ≤ 1 − δX ε r.
 2  r
148 3. Geometric Coefficients of Banach Spaces

Hence for x ∈ D we have

   
 x1 + x2   
x − x0  =  x −  =  (x − x1 ) + (x − x2 ) 
 2   2 
  
d
≤ d 1 − δX
2d
  
1
= d 1 − δX (3.13)
2
< d. (as δX (1/2) > 0)

Consequently,

sup{x − x0  : x ∈ D} < diam(D).

Theorem 3.3.4 Every uniformly convex Banach space has normal structure.

Proof. It follows from Theorem 3.3.3.

The following class of Banach spaces is more general than the class of uni-
formly convex Banach spaces:
Let X be a Banach space. Given an element z ∈ SX and a constant ε ∈ [0, 2],
we define
   
x + y 

δX (z, ε) = inf 1−  : x, y ∈ BX , x−y ≥ ε, x−y = tz for some t ≥ 0 .
2 

The number δX (z, ε) is called the modulus of convexity in the direction

z ∈ SX . Then Banach space X is said to be uniformly convex in every direction
if δX (z, ε) > 0 for all ε ∈ (0, 2] and z ∈ SX . It is obvious that δX (ε) =
inf{δX (z, ε) : z ∈ SX }.

Observation
• A Banach space X may be uniformly convex in every direction while failing
to be uniformly convex.
• Uniformly convex Banach spaces in every direction are always strictly convex.
• In case of a Hilbert H, δH (z, ε) = 1 − 1 − (ε/2)2 for all z ∈ SH and ε ∈ [0, 2].

The following Theorem 3.3.6 shows that a uniformly convex Banach space
in every direction has normal structure.

Proposition 3.3.5 Let X be a Banach space and C a convex bounded subset

of X with d = diam(C) > 0. Then, for ε > 0, there exists a point x0 ∈ C such
that
  
d−ε
sup{x0 − x : x ∈ C} ≤ d 1 − δX z, for some z ∈ SX .
d
3.3. Normal structure 149

Proof. Given d > ε > 0, we choose x1 , x2 ∈ C such that x1 − x2  ≥ d − ε. Set

x0 := (x1 + x2 )/2 and z := (x1 − x2 )/x1 − x2 . Note x − x1  ≤ d, x − x2  ≤ d
and (x − x1 ) − (x − x2 ) = x1 − x2 = zx1 − x2 . Hence
    
 x1 + x2  d−ε

x − x0  = x − 
2  ≤ d 1 − δX z, d .

Theorem 3.3.6 Let X be a uniformly convex Banach space in every direction.

Then X has normal structure.

Proof. Because X is uniformly convex in every direction, δX (z, ε) > 0 for all
z ∈ X and ε > 0. It follows from Proposition 3.3.5 and the continuity of δX (z, ε)
that X has normal structure.

Example 3.3.7 The space C[0, 1] of continuous real-valued functions with “sup”
norm does not have normal structure. To see it, consider the subset C of
X = C[0, 1] defined by

C = {f ∈ C[0, 1] : 0 = f (0) ≤ f (t) ≤ f (1) = 1, t ∈ [0, 1]}.

Let f1 , f2 ∈ C and λ ∈ [0, 1] and f = λf1 + (1 − λ)f2 . Then f (0) = 0, f (1) = 1

and 0 ≤ f (t) ≤ 1 for all t ∈ [0, 1]. Hence C is convex. Thus, C is a closed
convex bounded subset of X with diam (C) = sup{f − g : f, g ∈ C} = 1. Then
each point of C is a diametral point. In fact, for f0 ∈ C

sup{f0 − f  : f ∈ C} = 1 = diam(C).

Therefore, C does not have normal structure.

The following notion plays an important role in the study of normal struc-
ture.

A bounded sequence {xn } in a Banach space is said to be a diametral

sequence if

lim d(xn+1 , co({x1 , x2 , · · · , xn })) = diam({xn }),

n→∞

where d(x, A) = inf d(x, y).

y∈A

Remark 3.3.8 Any subsequence of a diametral sequence is also diametral.

The following result gives an important fact relating normal structure and
nondiametral sequence.

Proposition 3.3.9 A convex bounded subset C of a Banach space X has

normal structure if and only if it does not contain a diametral sequence.
150 3. Geometric Coefficients of Banach Spaces

Proof. Suppose C contains a diametral sequence {xn }. Then the set

C0 = co({xn }) is a convex subset of C and each point of C0 is a diametral
point. Thus, C fails to have normal structure.
Conversely, suppose that C contains a convex subset D with
d = diam(D) > 0 and each point of D is a diametral point. By induction,
we construct a sequence {xn } in D such that
y0 = x1 ,
n
xi
yn−1 = .
i=1
n
Because yn−1 is a diametral point in D, then for 0 < ε < d, there exists an
xn+1 ∈ D such that
ε
xn+1 − yn−1  > d − 2 .
n
n
Suppose x ∈ co({x1 , x2 , · · · , xn }), say x = i=1 λi xi , where λi ≥ 0 and
n
i=1 λi = 1.
Set 0 < λ := max{λ1 , λ2 , · · · , λn }. Then
  n  
1 λi 1 1 λi
1− ≥ 0 and + 1− = 1.
n λ nλ n i=1 λ
Hence
n  
1 1 1 
n n
1 λi 1
x+ 1− xi = x+ xi − λi xi = yn−1 .
nλ n i=1 λ nλ n i=1 λn i=1
Observe that
ε
d− < xn+1 − yn−1 
n2
n  
1 1 λi
=  (xn+1 − x) + 1− (xn+1 − xi )
nλ n i=1 λ
n  
1 1 λi
≤ xn+1 − x + 1− xn+1 − xi 
nλ n i=1 λ
 
1 1
≤ xn+1 − x + 1 − d.
nλ nλ
Hence
   
ε 1
xn+1 − x ≥ nλ d − 2 − 1 − d
n nλ
 
d ε
= nλ −
nλ n2
ελ
= d−
n
ε
≥ d− ,
n
3.3. Normal structure 151

and it follows that

ε
dist (xn+1 , co({x1 , x2 , · · · , xn })) ≥ d − .
n
Now ε is an arbitrary constant, so therefore, {xn } is a diametral sequence
in D.
Example 3.3.10 In the space 1 , the basis vectors {en } form a diametral
sequence. Hence 1 does not have normal structure.
Theorem 3.3.11 Let X be a reflexive Banach space with the Opial condition.
Then X has normal structure.
Proof. Suppose, for contradiction, that X fails to have normal structure. Then
X contains a diametral sequence {xn } that may converge weakly to 0. Because
{xn } is diametral sequence, by the definition
lim d(xn+1 , co({x1 , x2 , · · · , xn })) = diam({xn }).
n→∞

In particular, lim xn − y = diam({xn }) if y ∈ co({x1 , x2 , · · · }) and lim

n→∞ n→∞
xn − y = diam({xn }) if y ∈ co({x1 , x2 , · · · }). Taking y = 0, we obtain
lim xn  = diam({xn }).
n→∞

Because also
lim xn − x1  = diam({xn }),
n→∞

this contradicts the Opial condition.

Many spaces satisfy a stronger property than normal structure.

Definition 3.3.12 A nonempty convex subset C of a Banach space is said to
have uniformly normal structure if there exists a constant α ∈ (0, 1), indepen-
dent of C, such that each closed convex bounded subset D of C with diam(D) > 0
contains a point x0 ∈ C such that
sup{x0 − x : x ∈ D} ≤ α diam(D).
Theorem 3.3.13 Every uniformly convex Banach space X has uniformly
normal structure.
Proof. For a closed convex bounded subset C of X with d = diam(C) > 0
from (3.13), there exists a point x0 ∈ C such that
  
1
x − x0  ≤ 1 − δX d.
2
This implies that
sup{x − x0  : x ∈ C} ≤ α diam(C),

where α = 1 − δX (1/2) < 1. Therefore, X has uniformly normal structure.

152 3. Geometric Coefficients of Banach Spaces

Let C be a nonempty bounded subset of a Banach space X. We adopt the

following notations:

rx (C) = sup{x − y : y ∈ C}, x ∈ C;

r(C) = inf{rx (C) : x ∈ C} = inf{sup x − y : x ∈ C};
y∈C
Z(C) = {x ∈ C : rx (C) = r(C)};
rX (C) = inf{rx (C) : x ∈ X}.

The number r(C) is called the Chebyshev radius of C and the set Z(C) is called
the Chebyshev center of C. Note that for any x ∈ C

r(C) ≤ rx (C) ≤ diam(C).

Clearly, a point x0 ∈ C is a diametral (nondiametral) point if rx0 (C) =

diam(C) (rx0 (C) < diam(C)). Thus, set C has normal structure if

r(C) < diam(C)

and uniformly normal structure if there exists a constant α ∈ (0, 1), independent
of C, such that
r(C) ≤ α diam(C).

The set C is called diametral if it consists only of diametral points, i.e.,

rx (C) = diam(C) for all x ∈ C,

equivalently

Z(C) = {x ∈ C : rx (C) = diam(C) = r(C)}.

Observation
• The set Z(C) may be empty.
• The set C is diametral if r(C) = diam(C).
• 1
2
diam(C) ≤ rX (C) ≤ r(C) ≤ diam(C).
• co(C) = ∩{Brx (C) (x) : x ∈ C}.
• sup{x0 − x : x ∈ C} = sup{x0 − y : y ∈ co(C)}, x0 ∈ C.
• r(co(C)) ≤ r(C).
• diam(C) = diam(co(C)).

The following result gives an essential condition for the existence of

Chebyshev centers.

Proposition 3.3.14 Let C be a weakly compact convex subset of a Banach

space X. Then Z(C) is a nonempty closed convex subset of C.
3.4. Normal structure coefficient 153

Proof. For x ∈ C, set

 
1
Cn (x) := Br(C)+ n1 [x] = y ∈ C : x − y ≤ r(C) + , n ∈ N.
n

Then Cn (x) is a nonempty closed convex subset of C and hence Cn = x∈C Cn (x)
is a nonempty closed convex subset of C and Cn+1 ⊂ Cn for all n ∈ N. Because

C is weakly compact, it follows that Z(C) = n∈N Cn = ∅.
Proposition 3.3.15 Let X be a Banach space and C a weakly compact convex
subset of X with diam(C) > 0. Suppose C has normal structure. Then

diam(Z(C)) < diam(C).

Proof. Because C has normal structure, there exists at least one nondiametral
point x0 ∈ C, i.e.,

rx0 (C) = sup{x0 − x : x ∈ C} < diam(C).

Let u and v be any two points of Z(C). Then ru (C) = rv (C) = r(C). Because

u − v ≤ sup{u − x : x ∈ Z(C)} ≤ r(C) ≤ rx0 (C) < diam(C),

it follows that
diam(Z(C)) < diam(C).

3.4 Normal structure coefficient

Let X be a Banach space. Then the number N (X) is said to be the normal
structure coefficient if  
diam(C)
N (X) = inf ,
r(C)
where the infimum is taken over all closed convex bounded subsets C of X with
diam(C) > 0.

It is clear that N (X) ≥ 1 and N (X) > 1 if and only if X has uniformly
normal structure.
√
Example 3.4.1 For a Hilbert space H, N (H) = 2. Indeed, let C be the
positive part of the unit ball BH in a Hilbert space H = 2 , i.e.,

C = {x = {xi } : x ≤ 1 and xi ≥ 0, i = 1, 2, · · · }.

√
Then r(C) = inf{sup x − y : x ∈ C} = 1 and diam(C) = 2. Hence
√ y∈C
diam(C) ≥ 2 r(C).

We now give some important properties of the normal structure coefficient.

154 3. Geometric Coefficients of Banach Spaces

Theorem 3.4.2 Let X be a Banach space. Then

1
N (X) ≥ . (3.14)
1 − δX (1)
Proof. Let C be a closed convex bounded subset of X with d = diam (C) > 0
and let d > ε > 0. Choose x and y in C such that x − y ≥ d − ε. Let u be an
element in C and v = (x + y)/2 ∈ C such that

u − v ≥ rv (C) − ε.

By the definition of δX ,
    
u − x + u − y 
u − v =   ≤ d 1 − δX d − ε
 2  d
and by the definition rv (C),

r(C) ≤ rv (C) ≤ u − v + ε.

Thus,   
d−ε
r(C) ≤ 1 − δX diam (C) + ε.
d
Hence by continuity of δX , we have

r(C) ≤ (1 − δX (1))diam(C).

Therefore, we get the desired result.

√ the Hilbert space H, δH (ε) = 1 − (1 − (ε /4)) , which

2 1/2
Remark 3.4.3 For
gives N (H) ≥ 4/ 3, i.e., the estimate (3.14) is not sharp.

Before giving an important example, we observe that the property “uni-

formly normal structure” is stable under small norm perturbations.
Theorem 3.4.4 Let X be a Banach space and let X1 = (X,  ·  ) and X2 =
(X,  ·  ), where  ·  and  ·  are two equivalent norms on X satisfying for
α, β > 0,
αx ≤ x ≤ βx , x ∈ X.
If k = β/α, then
k −1 N (X1 ) ≤ N (X2 ) ≤ kN (X1 ).
Proof. Note for a nonempty bounded convex subset C of X,

α diam· (C) ≤ diam· (C) ≤ β diam· (C).

Hence the result follows from the definition of N (X).

The following example shows that a uniformly convex Banach space (Hilbert
space) has an equivalent norm that fails to have normal structure.
3.4. Normal structure coefficient 155

Example 3.4.5 Let X1 = 2 and let Xλ , λ > 1 denote the space obtained by
renorming the Hilbert space ( 2 ,  · ) as follows:
For x = (x1 , x2 , · · · ) ∈ 2 , set
   
∞ 1/2 
−1 −1
xλ : = max x∞ , λ x = max max |xi |, λ 2
xi .
i∈N
i=1

Because
λ−1 x ≤ xλ ≤ x,
it means that all the norms  · λ are equivalent to norm  · . However the
spaces (Xλ ,  · λ ), λ > 1 are not uniformly convex. A simple calculation shows
that
 √
2(λ2 − 1)1/2 for λ ≤√ 2,
0 (Xλ ) =
2 f or λ ≥ 2.
Then we have the following:
(i) 0 (X√5/2 ) = 1.
√
(ii) For λ = 2, 0 (Xλ ) = 2 and N (Xλ ) = 1, i.e., Xλ fails to have normal
structure, because the set C ⊂ X√2 defined by
 ∞
 
C= x = {xi } : x2i ≤ 1 and xi ≥ 0, i = 1, 2, · · ·
i=1

satisfies r(C) = diam(C) = 1 with respect to the X√2 norm.

Let us now check the validity of Theorem 3.3.13 for uniformly smooth Banach
spaces. To do so, we introduce the notion of super-reflexivity.
Let X and Y be two Banach spaces. We say that Y is finitely representable
in X if for every λ > 0, every finite-dimensional subspace Y0 of Y , there
exist a finite-dimensional subspace X0 of X with dim(X0 ) = dim(Y0 ) and an
isomorphism T : Y0 → X0 such that

T  T −1  ≤ 1 + λ.

This property can be expressed in terms of the Banach-Mazur distance, which

is defined as follows: The Banach-Mazur distance between two Banach spaces
X and Y is denoted by d(X, Y ) and is defined by

d(X, Y ) = inf{T T −1  : T is an isomorphism from X onto Y}.

Thus, Y is finitely representable in X if for any ε ∈ (0, 1) and any finite-

dimensional subspace Y0 of Y , there exists a subspace X0 of X such that
dim(X0 ) = dim(Y0 ) and d(X0 , Y0 ) ≤ 1 + ε.
156 3. Geometric Coefficients of Banach Spaces

Observation
• The Banach space X is finitely representable in itself.
• The relation finite representability is transitive: If a Banach space X is finitely
representable in a Banach space Y and if the Banach space Y is finitely repre-
sentable in another Banach space Z, then X is finitely representable in Z.
We now define the “super-P” property of Banach spaces.
Definition 3.4.6 Let P be a property defined for Banach spaces. Then a
Banach space X is said to have super-P if every Banach space finitely rep-
resentable in X has the property P.
Remark 3.4.7 Every Banach space is finitely representable in itself, so it follows
that if X has super-P for any P, then X has P.
Now we are in a position to define super-reflexivity of Banach spaces.
Definition 3.4.8 A Banach space X is said to be super-reflexive if every
Banach space Y that is finitely representable in X is itself reflexive.
Thus, any super-reflexive Banach space is necessarily reflexive (by
Remark 3.4.7).

For uniform convexity of a Banach space, we have the following:

Theorem 3.4.9 Let X and Y be two Banach spaces with respective moduli of
convexity δX and δY , and suppose Y is finitely representable in X. Then for
each ε ∈ [0, 2), δX (ε) ≤ δY (ε).
Proof. Let ε > 0 and let x, y ∈ Y with x, y ∈ SY and x − yY ≥ ε. Suppose
λ > 0 and let T be an isomorphism of span{x, y} onto some two-dimensional
subspace X0 of X that satisfies
T  ≤ 1 + λ and T −1  = 1.
Set x := T x and y  := T y. Then
x X , y  X ≤ 1 + λ and x − y  X = T x − T yX ≥ x − yY ≥ ε.
By the definition of δX , we have
     
1 x y  ε
 +  ≤ 1 − δX .
2 1 + λ 1 + λ  1+λ
X
By the continuity of δX , we obtain
  
 x + y 
 
 2  ≤ 1 − δX (ε).

Because T −1  = 1, it follows that

        
x + y      
  = T −1 x + y  ≤  x + y  .
 2   2   2 
Y Y X

Therefore, by the definition of δY , we conclude that δY (ε) ≤ δX (ε).

3.4. Normal structure coefficient 157

Corollary 3.4.10 Let X be a Banach space. If Y is finitely representable in

X, then 0 (Y ) ≤ 0 (X).

Corollary 3.4.11 Let X be a uniformly convex Banach space. If a Banach

space Y is finitely representable in X, then Y is uniformly convex.

Corollary 3.4.12 Every uniformly convex Banach space is super-reflexive.

Proof. Because every uniformly convex Banach space is reflexive, it follows

from Corollary 3.4.11 that every uniformly convex space is super-reflexive.

We now discuss the super-property by ultrapower of Banach spaces:

Because the ultrapower {X}U of a Banach space X is finitely representable

in X (see Proposition A.3.10 of Appendix A), we have:

If P is a Banach space property that is inherited by subspaces, then a Banach

space X has super-P if and only if every ultrapower of X has P.

Thus, we have a link between moduli of convexity and smoothness concepts

and ultrapowers:

Let {X}U be an ultrapower of a Banach space X. Then

δX (·) = δ{X}U (·) and ρX (·) = ρ{X}U (·).

Consequently,

0 (X) = 0 ({X}U ) and ρX (0) = ρ{X}U (0).

Theorem 3.4.13 Let X be a Banach space with modulus of smoothness ρX .

If ρX (0) < 1/2, then X is super-reflexive and has normal structure.

Proof. Suppose, for contradiction, that X is not super-reflexive. Then for any
θ < 1, there exist x, y ∈ BX and j1 , j2 ∈ BX ∗ such that

y, j1  = x, j1  = x, j2  = θ and y, j2  = 0.

Hence for all t > 0:

1
ρX (t) ≥ (x + ty + x − ty) − 1
2
1
≥ (x + ty, j1  + x − ty, j2 ) − 1
2
t
= θ(1 + ) − 1.
2
Because θ < 1 is arbitrary, ρX (t) ≥ t/2, a contradiction.
158 3. Geometric Coefficients of Banach Spaces

Again suppose, for contradiction, that X does not have normal structure.
Then there exists a sequence {xn } in BX such that
xn  0, lim xn  = 1 and diam({xn }) ≤ 1.
n→∞

Consider a sequence {jn } in SX ∗ such that xn , jn  = xn , n ∈ N. Because

X ∗ is reflexive, we may assume that {jn } converges weakly to j ∈ X ∗ . Select
i ∈ N such that |xi , j| < 2ε while xn  > 1 − ε for all n ≥ i. Then for k > i
sufficiently large,
ε
xi , jk − j < and |xk , ji | < ε.
2
Consequently, |xi , jk | < ε. For all t ∈ (0, 1), we have
1
ρX (t) ≥ (xi − xk + txi  + xi − xk − txi ) − 1
2
1
≥ (|(1 + t)xi − xk , ji | + |xk − (1 − t)xi , jk |) − 1
2
1
≥ ((1 + t)(1 − ε) − ε + 1 − ε − (1 − t)ε) − 1
2
t
= − 2ε.
2
Because ε > 0 is arbitrary, ρX (t) ≥ t/2, and it follows that ρX (0) ≥ 1/2, a
contradiction to the hypothesis ρX (0) < 1/2.
Theorem 3.4.14 Let X be a Banach space with ρX (0) < 1/2. Then X has
uniformly normal structure.
Proof. Note X has normal structure by Theorem 3.4.13. Suppose, for con-
tradiction, that it is not uniform. Then there exists a sequence {Cn } of closed
convex bounded sets of X, each containing 0 and having diameter 1, for which
lim r(Cn ) = 1, where r(Cn ) = inf{rx (Cn ) : x ∈ Cn }. Define the set C ⊂ {X}U
n→∞
by
C = {Cn }U = {x ∈ {X}U : x = {xn }U , xn ∈ Cn }.
Then C is a closed convex set in {X}U with diam(C) = 1. Moreover, for any
x = {xn }U ∈ C, there exists y = {yn }U ∈ C such that lim xn − yn  = 1, and
n→∞
it means that x − yU = 1 and C is diametral. Because ρX (0) = ρ{X}U (0) < 12 ,
it follows that {X}U has normal structure. This is a contradiction.
Corollary 3.4.15 Let X be a uniformly smooth Banach space. Then X has
uniformly normal structure.
We call a subsequence {yn } a c-subsequence of the sequence {xn } of a Banach
space X if there exists a sequence of integers 1 = p1 ≤ q1 < p2 ≤ q2 < · · · and
scalars αi ≥ 0 such that, for each n ∈ N,

qn 
qn
yn = αi xi , αi = 1.
i=pn i=pn
3.4. Normal structure coefficient 159

Theorem 3.4.16 Every Banach space with a uniformly normal structure is

reflexive.
Proof. Let X be a Banach space with a uniformly normal structure. Let {Cn }
be a decreasing sequence of nonempty closed convex bounded subsets of X. We
need to show that ∩n∈N Cn = ∅. For each n ∈ N, choose xn ∈ Cn . Then for
each ε > 0, there exists a c-subsequence {yn } of {xn } with ym − yn  < ε for
each m, n. Suppose this is not true for some ε > 0. Let Bm = {xn }∞n=m . Then
there exist h, 0 < h < 1 and y1 ∈ co(B1 ) such that

sup{y1 − y : y ∈ co(B1 )} ≤ h diam(B1 ).

Suppose that 0 < h < h1 < 1. Then there exists y1 ∈ co(B1 ) such that

sup{y1 − y : y ∈ co(B1 )} ≤ h1 diam(B1 ).

Because y1 is a finite linear combination of members of B1 , there exists a

c-subsequence {yn } of {xn } such that

sup{yn − y : y ∈ co(Bpn )} ≤ h1 diam(Bpn ) ≤ h1 diam(B1 ),

and it follows that diam({yn }) ≤ h1 diam(B1 ).

By repeating the argument, there exists a successive c-subsequence with
diameter less than or equal to h21 diam(B1 ). We need only repeat the argument
a sufficient number k of times with hk1 diam(B1 ) < ε to obtain a contradiction.
Next by the diagonal method, there exists a c-subsequence of {xn } that is
norm Cauchy, and hence convergent to some y. Therefore, y ∈ ∩n∈N Cn .

We now give the following constants for bounded sequences in Banach spaces
that are very useful to define various geometric coefficients:
Definition 3.4.17 Let {xn } be a bounded sequence in a Banach space X. Then
(1) the real number

diam({xn }) := sup{xm − xn  : m, n ∈ N}

is called the diameter of the sequence {xn },

(2) the real number

diama ({xn }) := lim (sup{xi − xj  : i, j ≥ n})

n→∞

is called the asymptotic diameter of the sequence {xn },

(3) the real number

ra ({xn }) := inf{ra (y, {xn }) : y ∈ co({xn }n∈N )}

is called the asymptotic radius of the sequence {xn } with respect to

co({xn }n∈N ),
160 3. Geometric Coefficients of Banach Spaces

(4) the real number

sep({xn }) := inf{xm − xn  : m = n}

is called the separation of the sequence {xn }.

For a bounded sequence {xn } in a Banach space X, we set

D[{xn }] := lim sup(lim sup xm − xn ).

m→∞ n→∞

It can be easily seen that D[{xn }] ≤ diama ({xn }).

The following example shows that D[{xn }] = diama ({xn }) in general.

Example 3.4.18 Let J be the James quasi-reflexive space, consisting of all real
∞
sequences x := {xn } = n=1 xn en for which
lim xn = 0, en = (0, 0, · · · , 1, 0, · · · ) and xJ < ∞,
n→∞
↑
nth position
where

xJ = sup{[(xp1 − xp2 )2 +(xp2 − xp3 )2 +· · ·+(xpm−1 − xpm )2 +(xpm −xp1 )2 ]1/2}

and the supremum is taken over all choices of m and p1 < p2 < · · · < pm . Then
J is a Banach space with the norm  · J . Consider the sequence {xn } defined
by
xn = en − en+1 , n ∈ N.
√
Then xn ∈ J and xn J = 6 for all n ∈ N. We now show that

D[{xn }] < diama ({xn }).

√
Note xm − xn J = 2 3 for each fixed n ∈ N and for all m ≥ n + 3, and hence
√
D[{xn }] = lim sup(lim sup xm − xn J ) = 2 3.
m→∞ n→∞

On the other hand, for each k ∈ N, if we take n = k, m = n + 1 ≥ k, then

√
xm − xn J = 2 5

and hence
√
sup{xm − xn J : m, n ≥ k} = 2 5 for each k ∈ N,
√
which implies that diama ({xn }) = 2 5. Therefore,
√ √
D[{xn }] = 2 3 < 2 5 = diama ({xn }).

Remark 3.4.19 For each bounded sequence {xn } in a Banach space X, we

have
ra ({xn }) ≤ D[{xn }] ≤ diama ({xn }) ≤ diam({xn }).
3.4. Normal structure coefficient 161

A different form of uniformly normal structure coefficient N (X) is defined

by


diama ({xn })
N (X) = inf : {xn } is a bounded sequence
ra ({xn })

which is not norm convergent .

We now prove an interesting result concerning the uniformly normal struc-

ture coefficient N (X).
Theorem 3.4.20 Let X be a Banach with Ñ (X) = N (X)−1 < 1. Then, for
every bounded sequence {xn } in X, there exists a point z ∈ co({xn }) such that
(a) for every y ∈ X, z − y ≤ ra (y, {xn }),
(b) ra (z, {xn }) ≤ Ñ (X) diama ({xn }).
Proof. (a) For each k ∈ N, set Ak := co({xn }∞ n=k ) and A := ∩k∈N Ak . Observe
that any point z ∈ A satisfies (a). In fact, z ∈ Ak for each k ∈ N, and hence

z − y ≤ sup{x − y : x ∈ Ak } for all y ∈ X,

which implies that

z − y ≤ lim ry (Ak ) = ra (y, {xn }).

k→∞

(b) The reflexivity X implies that each Ak is weakly compact. Hence the sets
A, Za (Ak , {xn }) and Za (A, {xn }) are all nonempty. For each k, choose zk in
Za (Ak , {xn }) and consider a weakly convergent subsequence {zki } of {zk } such
that zki  z. Because z ∈ co({zki }∞ i=j ) ⊂ Akj for any j, by the monotonicity
of the sequence {Ak }, we obtain that z ∈ ∩∞ j=1 Akj = A. Note ra (zk , {xn }) is
a monotonic nondecreasing sequence that has ra (A, {xn }) as an upper bound.
Moreover, because ra (·, {xn }) is weakly lower semicontinuous, we have

lim ra (zk , {xn }) = lim ra (zkj , {xn })

k→∞ j→∞
≥ ra (z, {xn })
≥ ra (A, {xn }).

Hence
lim ra (zk , {xn }) = ra (z, {xn }) = ra (A, {xn }).
k→∞

Note, for any k,

ra (zk , {xn }) = ra (zk , {xn }∞ ∞

n=k ) ≤ Ñ (X) diama ({xn }n=k ) = Ñ (X) diama ({xn }).

Therefore,
ra (z, {xn }) ≤ Ñ (X) diama ({xn }).
162 3. Geometric Coefficients of Banach Spaces

3.5 Weak normal structure coefficient

A Banach space X is said to have weak normal structure if every weakly compact
convex subset C of X with more than one point contains a nondiametral point,
that is, an x0 ∈ C for which

sup{x0 − y : y ∈ C} < diam(C).

Every Banach space that has normal structure also has weak normal struc-
ture, but the converse is not true. For reflexive Banach spaces, these properties
are equivalent.

Example 3.5.1 The space 1 does not have normal structure. Because weak
compactness coincides with compactness in 1 , it follows that 1 does have weak
normal structure.

Let X be a Banach space. Then the number W CS(X) is said to be weak

normal structure coefficient or weakly convergent sequence coefficient if

diam(C)
W CS(X) = inf : C is weakly compact convex subset of X
r(C)

with diam(C) > 0

diama ({xn })
= inf : {xn } is a weakly convergent sequence
ra ({xn })

which is not norm convergent ,

where ra ({xn }) = inf{ra (x, {xn }) : x ∈ co({xn }) is the asymptotic radius of

{xn } relative to co({xn })}.

Observation
• W CS(X) ≥ 1, as r(C) ≤ diam(C).
• A Banach space X has weak uniformly normal structure if W CS(X) > 1.
• If W CS(X) > 1, then X has weak normal structure.

We now give a sharp expression for W CS(X) in terms of D[{xn })].

We begin with the following lemmas:

Proposition 3.5.2 Let {yn } and {zn } be two sequences in a Banach space X
such that α = lim yn  = 0 and zn = yn /yn . Then D[{zn }] = D[{yn }]/α.
n→∞
3.5. Weak normal structure coefficient 163

Proof. For m, n ∈ N, we have

 
 ym yn 
zm − zn  =   ym  −

yn  
 
 ym ym ym yn yn yn 
=   ym  − + − + − 
α α α α yn  
1 1 1 1 1
≤ ym  − + ym − yn  + yn  − .
ym  α α α yn 

Hence

D[{zn }] = lim sup(lim sup zm − zn )

m→∞ n→∞
1 1 1
≤ lim sup(ym  − + lim sup ym − yn )
m→∞ y m  α α n→∞
1 1
= lim sup(lim sup ym − yn ) = D[{yn })].
α m→∞ n→∞ α
Similarly, we can obtain
1
D[{yn }] ≤ D[{zn }]
α
Therefore, D[{zn }] = D[{yn }]/α.
Proposition 3.5.3 Let X be a Banach space and M > 0. Then the following
are equivalent:
(a) M lim sup xn − x ≤ diama ({xn }) for any xn  x (not strongly con-
n→∞
vergent).
(b) M lim sup yn − y ≤ D[{yn }] for any yn  y (not strongly convergent).
n→∞

Proof. (b) ⇒ (a). Because D[{xn }] ≤ diama ({xn }), it is obvious that (b) ⇒
(a).
(a) ⇒ (b). Let xn  x (not strongly convergent) and α := lim sup
n→∞
xn − x =
 0. Then we can choose a subsequence {xm } of {xn } such that
α := lim xm − x. Set zm := (xm − x)/α. Then zm  0 and zm  = 1, by
m→∞
using the diagonal method, we can choose a subsequence {zmi } of {zm } such
that lim zmi − zmj  exists. From (a), we have
i,j→∞,i=j

M = M lim zmi 
i→∞
≤ diama ({zmi }) = D[{zmi }]
1
≤ D[{zm }] = D[{xm }] (by Proposition 3.5.2)
α
1
≤ D[{xn }],
α
164 3. Geometric Coefficients of Banach Spaces

and it follows that

M lim sup xn − x = M α ≤ D[{xn }].

n→∞

We are now able to give an expression for W CS(X) in terms of D[{xn }].

Theorem 3.5.4 Let X be a Banach space. Then

W CS(X) = sup{M > 0 : xn  u ⇒ M lim sup xn − u ≤ D[{xn }]}.

n→∞

Proof. The result follows from Proposition 3.5.3.

We now establish a relation between the Opial modulus and weak normal
structure coefficient of a Banach space.

Theorem 3.5.5 Let X be a Banach space with the Opial modulus rX . Then

W CS(X) ≥ 1 + rX (1).

Proof. Let {xn } be a sequence in X such that xn  x ∈ X. Set b :=

ra (x, {xn }) = lim sup xn − x. Without loss of generality, we may assume that
n→∞
b > 0 and lim xn − x exists. Otherwise, we can consider a sequence {xni }
n→∞
of {xn } such that
lim sup xn − x = lim xni − x.
n→∞ i→∞

Set zn := (xn − x)/b. From the definition of rX , we get

1 + rX (t) ≤ lim inf zn + y for all y ∈ X with y ≥ t.

n→∞

In particular for y = (x − xm )/b and t = y, we have

 
x − xm  xn − xm  xn − xm 
1 + rX ≤ lim inf ≤ lim sup .
b n→∞ b n→∞ b

It follows that
b(1 + rX (1)) ≤ D[{xn }].

Therefore, W CS(X) ≥ 1 + rX (1).

Theorem 3.5.6 Let X be Banach space with the Opial modulus rX .

If rX (1) > 0, then W CS(X) > 1, i.e., X has weak uniformly normal struc-
ture.

Proof. The result follows from the fact that W CS(X) ≥ 1 + rX (1).
3.6. Maluta constant 165

3.6 Maluta constant

Let X be a Banach space. Then the number D(X) is called the Maluta constant
if
 lim sup d(xn+1 , co({xi }ni=1 )) 
n→∞
D(X) = sup ,
diam({xn })
where the supremum is taken over are all nonconstant bounded sequences in X.

The following result shows that D(X) can be defined in several ways.
Proposition 3.6.1 Let X be a Banach space. Then we have the following:
 lim inf d(x 
n+1 , co({xi }i=1 ))
n

(a) D(X) = sup n→∞

diam({xn }) : {xn } ⊂ X .

(b) If X is reflexive, then

 lim sup d(xn+1 , co({xi }ni=1 ))
n→∞
D(X) = sup : {xn } a weakly convergent
diam({xn })

sequence in X .

 lim sup d(xn+1 , co({xi }ni=1 ))

(c) D(X) = sup n→∞
diama ({xn }) : {xn } a nonconvergent

sequence in X .

(d) If X is an infinite-dimensional reflexive space, then

 lim sup d(xn+1 , co({xi }ni=1 ))
n→∞
D(X) = sup : {xn } a weakly, but not strongly
diama ({xn })

convergent sequence in X .

Proof. (a) Let {xn } be a sequence in X and set α := lim sup d(xn , co({xi }n−1
i=1 )).
n→∞
k −1
Then for a subsequence {xnk } of {xn } such that α = lim d(xnk co({xi }ni=1 )),
k→∞
we have
k −1
α = lim d(xnk , co({xni }ni=1 )) ≤ lim inf d(xnk , co({xi }k−1
i=1 )).
k→∞ k→∞

Hence for every {xn } we can find a sequence {yn }(= {xnk }) with diam({yn }) ≤
diam({xn }) such that
lim sup d(xn+1 , co({xi }ni=1 )) lim inf d(yn+1 , co({yi }ni=1 ))
n→∞ n→∞
≤ .
diam({xn }) diam({yn })
Thus, our assertion is proved.
166 3. Geometric Coefficients of Banach Spaces

(b) Let {xn } be a sequence in X with a weakly convergent subsequence

{xnk }. Then
k −1
d(xnk , co({xi }ni=1 )) ≤ d(xnk , co({xi }k−1
i=1 )),
and it follows that
lim inf d(xn+1 , co({xi }ni=1 )) lim sup d(xnk , co({xi }k−1
i=1 ))
n→∞ n→∞
≤ .
diam({xn }) diam({xnk })
From part (a) we can conclude (b).
(c) Let {xn } be a nonconvergent sequence in X. Then for each k ∈ N, we
have
lim sup d(xn+1 , co({xi }ni=1 )) ≤ lim sup d(xn+1 , co({xi }ni=k ))
n→∞ n→∞
≤ D(X) diam({xi }∞
i=k ).

Thus
lim sup d(xn+1 , co({xi }ni=1 )) ≤ lim (lim sup d(xn+1 , co({xi }ni=k )))
n→∞ k→∞ n→∞
≤ D(X) diama ({xn }).
Using the fact that diama ({xn }) ≤ diam({xn }), we obtain
 lim sup d(xn+1 , co({xi }ni=1 )) 
n→∞
D(X) ≤ sup ≤ D(X).
diama ({xn })

(d) It follows easily from (b).

Theorem 3.6.2 Let X be a Banach space. Then we have the following:
(a) If X is a finite-dimensional space, then D(X) = 0.
(b) If X is an infinite-dimensional space with modulus of convexity δX , then
1 1
D(X) ≥ ≥ .
2(1 − δX (1)) 2
Proof. (a) Let X be a finite-dimensional Banach space. Then for every con-
vergent sequence {xn }, we have lim d(xn+1 , co({xi }ni=1 )) = 0. It follows from
n→∞
Proposition 3.6.1(b) that D(X) = 0.
(b) Let X be an infinite-dimensional Banach space and 0 < r < 1. We now
construct a sequence in the following way:

x1 ∈ SX ,
xn+1 ∈ SX such that d(xn+1 , span{xi }ni=1 ) > r for all n ∈ N.
Then for any i, j, i = j we have xi + xj  > r so
1
xi − xj  ≤ 1 − δX (r).
2
3.6. Maluta constant 167

Hence
lim sup d(xn+1 , co({xi }ni=1 ))
n→∞ r
≥ .
diam({xn }) 2(1 − δX (r))
Because δX is continuous at 1, we obtain
1
D(X) ≥ .
2(1 − δX (1))

Theorem 3.6.3 Let X be a Banach space. Then we have the following:

(a) D(X) ≤ Ñ (X).
(b) If X is an infinite-dimensional reflexive Banach space, then
1
D(X) ≤ .
W CS(X)
Proof. (a) For every sequence {xn } in X, we denote co({xn }) by C. Then
diam(C) = diam({xn }) and rx (C) = sup x − y = rx ({xn }) for all x ∈ X. If
y∈C


N 
N
z ∈ C, z = λi xi (λi ≥ 0 and λi = 1),
i=1 i=1

then

N
rz ({xn }) = sup xn − λi xi 
n∈N i=1
≥ lim sup d(xn , co({xi }N
i=1 ))
n→∞
≥ lim sup d(xn , co({xi }n−1
i=1 )).
n→∞

Hence we obtain

r(C) = inf rx (C) = inf rx ({xn }) ≥ lim sup d(xn+1 , co({xi }ni=1 )
x∈C x∈C n→∞

and  
r(C)
Ñ (X) ≥ sup : C = co({xn }) ≥ D(X).
diam(C)
N N
(b) Let x ∈ co({xn }), x = i=1 λi xi (λi ≥ 0, i=1 λi = 1). Then

ra (x, {xn }) = lim sup xn − x ≥ lim sup d(xn+1 , co({xi }N

i=1 )
n→∞ n→∞
≥ lim sup d(xn+1 , co({xi }ni=1 )).
n→∞

Hence

lim sup d(xn+1 , co({xi }ni=1 ) ≤ inf{ra (z, {xn }) : z ∈ co({xn })} for all {xn } in X.
n→∞
168 3. Geometric Coefficients of Banach Spaces

By Proposition 3.6.1(d), we obtain

 lim sup d(x
n+1 , co({xi }i=1 ))
n
n→∞
D(X) = sup : {xn } a weakly, but not strongly
diama ({xn })
!
convergent sequence in X

ra ({xn })
≤ sup : {xn } a weakly, but not strongly convergent
diama ({xn })
!
sequence in X

1
= .
W CS(X)

Theorem 3.6.4 Let X be an infinite-dimensional reflexive Banach space X.

Then
1
W CS(X) = D (X) = ,
D(X)
where D (X) = inf{diam({xn }) : {xn } in X with xn  0 and lim sup
n→∞
xn  = 1}.
Proof. By Theorem 3.6.3(b), it suffices to show that
1
≤ D (X) ≤ W CS(X).
D(X)
First, we show that 1/D(X) ≤ D (X). We take a positive number ε < 1
and choose a sequence {xn } in SX with xn  0 and diam({xn }) < D (X) + ε.
Let {jn } be a sequence in X ∗ such that xn , jn  = jn ∗ = 1 for all n ∈ N.
By Proposition 1.9.25, we can assume that
|xm , jn | < ε whenever m = n.
Suppose that λi , i = 1, 2, · · · n are nonnegative constants such that λ1 + λ2 +
· · · + λn = 1. Then we have

n 
n 
n
xn+1 − λi xi  ≥ |xn+1 − λi xi , jn+1 | ≥ 1 − λi |xi , jn+1 | > 1 − ε.
i=1 i=1 i=1

Thus, lim sup d(xn+1 , co({xi }ni=1 )) ≥ 1 −  and hence

n→∞

1 diam({xn }) D (X) + ε
≤ < .
D(X) 1−ε 1−ε
Because ε ∈ (0, 1) is arbitrary, we have D(X)−1 ≤ D (X).
3.6. Maluta constant 169

We now show that D (X) ≤ W CS(X). Let ε > 0 and choose a weakly
convergent non-norm convergent sequence {xn } in X such that
lim (sup{xi − xj  : i, j ≥ n}) < (W CS(X) + ε)ra (co({xn }), {xn }),
n→∞

where ra (C, {xn }) = inf{ra (z, {xn }) : z ∈ C}.

Let x be the weak limit of {xn } and lim sup xn − x = d. Hence for k a
n→∞
sufficiently large number, we have
diam({xn }∞
n=k ) − ε < d (W CS(X) + ε).

Set yn := d−1 (xn − x). Then lim sup yn  = 1 and yn  0. Therefore,
n→∞

D (X) ≤ diam({yn }∞
n=k ) = d
−1
diam({xn }∞
n=k ) < W CS(X) + (1 + d
−1
)ε,

and it follows that D (X) ≤ W CS(X).

We now establish equivalent expressions for W CS(X).

Theorem 3.6.5 Let X be an infinite-dimensional reflexive Banach space. Then
we have the following:
(a) W CS(X) = inf{diam({xn }) : xn  0 and lim sup xn  = 1} = D (X).
n→∞

(b) W CS(X) = inf{diama ({xn }) : xn  0 and lim sup xn  = 1}.

n→∞

 lim xn − xm 
n,m;n=m
(c) W CS(X) = inf : xn  0 and lim sup xn −xm exists}.
lim sup xn  n,m; n=m
n→∞

Proof. (a) This part follows easily Theorem 3.6.4.

(b) It follows from part (a) that
W CS(X) = inf{diama ({xn }) : {xn } converges weakly to zero and
lim sup xn  = 1}. (3.15)
n→∞

(c) The equality (3.15) allows us to conclude (c) because for every sequence,
we can obtain a subsequence {xn } such that lim xn − xm  exists by
n,m;n=m

a diagonal argument.
Theorem 3.6.6 Let X be a reflexive Banach space. Let
 lim xn − xm  
n,m; n=m
r = inf ,
ra ({xn })
where the infimum is taken over all weakly convergent sequences that are not
convergent and such that lim xn − xm  exists and lim xn − z exists
n,m; n=m n→∞
for every z ∈ co({xn }). Then W CS(X) = r.
170 3. Geometric Coefficients of Banach Spaces

Proof. Because lim xn −xm  = diama ({xn }) if this limit exists, it follows
n,m; n=m
that
r ≥ W CS(X).
We now show that W CS(X) ≤ r. Let {xn } be a sequence in X with xn  0 such
that lim xn − xm  exists and lim xn − z exists for all z ∈ co({xn }).
n,m;n=m n→∞
Set Ak := co({xn }n≥k ). The weak convergence of {xn } to zero implies that
 ∞
k=1 Ak = {0}. Because the function ϕ on X defined by

ϕ(x) = lim xn − x, x ∈ X

n→∞

is weakly lower semicontinuous and Ak is weakly compact, then ϕ attains a

minimum at a point zk in Ak . Because 0 is the unique point that can be
weakly adherent to {zk } we infer that {zk } is weakly null. By the weak lower
semicontinuity of ϕ, we have
ϕ(0) ≤ lim ϕ(zk ).
k→∞

Because {ϕ(zk )} is a nondecreasing sequence that is bounded by ϕ(0), it follows

that
lim ϕ(zk ) ≤ ϕ(0).
k→∞
Thus,
lim ϕ(zk ) = ϕ(0).
k→∞
Observe that
ra ({xn }n≥k ) = ra (co({xn }n≥k ), {xn }n≥k )
= inf{ϕ(y) : y ∈ co({xn }n≥k )}
= ϕ(zk ).
By the definition of r, we have
r lim xn − zk  ≤ lim xn − xm .
n→∞ n,m; n=m

Taking the limit as k → ∞, we obtain

r lim xn  ≤ lim xn − xm .
n→∞ n,m; n=m

Hence
 lim xn − xm  
n,m; n=m
r ≤ inf ,
lim sup xn 
n→∞
where the infimum is taken over all weakly null sequence such that lim
n,m; n=m
xn − xm  exists and lim xn − z exists for every z ∈ co({xn }). Using
n→∞
Theorem 3.6.5 (c), it is clear that this infimum is W CS(X). Hence r ≤ W CS(X).
Therefore, W CS(X) = r.
3.6. Maluta constant 171

We now apply Theorem 3.6.6 to estimate W CS(X) for reflexive Banach

spaces with weakly continuous duality mappings.
Theorem 3.6.7 Let X be a reflexive Banach space with a weakly continuous
duality mapping Jμ with gauge function μ. Then 1 + rX (1) = W CS(X) =
t
Φ−1 (2Φ(1)), where Φ(t) = 0 μ(s)ds, t ≥ 0.
Proof. Using Theorem 3.2.16 and Theorem 3.5.5, we have

W CS(X) ≥ 1 + rX (1) = Φ−1 (2Φ(1)).

To prove equality, we take a weakly null sequence {xn } in SX . We may

assume (through a subsequence if necessary) that lim xn − xm  exists.
n,m;n=m
By Theorem 3.6.6, we have

W CS(X) ≤ lim xn − xm 

n,m;n=m
= lim ( lim xn − xm )
m→∞ n→∞
= lim ( lim Φ−1 (Φ(xn ) + Φ( − xm )))
m→∞ n→∞
= lim Φ−1 (Φ(1) + Φ(xm ))
m→∞
−1
= Φ (2Φ(1)).

This proves the desired result.

Because p (1 < p < ∞), admits a weakly continuous duality mapping Jμ

with the gauge μ(t) = tp−1 , we have
Corollary 3.6.8 Let 1 < p < ∞. Then W CS( p ) = 21/p .
We now establish a relation between W CS(X) and N (X).
Theorem 3.6.9 Let X be a Banach space. Then W CS(X) ≥ N (X).
Proof. Let {xn } be a weakly null sequence in X such that lim xn  = c > 0
n→∞
and lim xn − xm  = d exist. Given ε > 0 with c/2 > ε, we may assume
m,n;m=n
that
xn  ≥ c − ε and xn − xm  ≤ d + ε for all n, m ∈ N.
Let {jn } be a sequence in SX ∗ such that xn , jn  = xn  for all n ∈ N. Using
Proposition 1.9.25, we may assume that |xm , jn | < ε whenever m = n. Note
x ∈ co({xn }) implies that |x, jn | < ε for n sufficiently large. Hence

c − 2ε ≤ xn  − ε = xn − x, jn  ≤ xn − x,

and it gives us r(co({xn })) ≥ c − 2ε. Thus,

diam(co({xn })) d+ε

N (X) ≤ ≤ .
r(co({xn })) c − 2ε
172 3. Geometric Coefficients of Banach Spaces

This shows that

d
N (X) ≤ .
c
Taking the infimum, we obtain that N (X) ≤ W CS(X).
Corollary 3.6.10 Let X be a Banach space. If X has uniformly normal struc-
ture, then X has weak uniformly normal structure.

3.7 GGLD property

Definition 3.7.1 A Banach space X is said to be have generalized Gossez-Lami
Dozo property (GGLD) whenever D[{xn }] > 1 for every weakly null sequence
{xn } such that lim xn  = 1.
n→∞

Example 3.7.2 Consider the space co equivalently renormed by

∞
1
{xn } = {xn }∞ + |x |.
n n
n=1
2

It enjoys the Opial condition and it has a weak normal structure but lacks the
GGLD property.

The following coefficient is very useful for fixed point theory of nonlinear
mappings:
β(X) := inf{D[{xn }] : xn  0, xn  → 1}.
Observation
• A Banach space X has the GGLD property if β(X) > 1.

Proposition 3.7.3 Let X be a Banach space with the GGLD property. If {xn }
is a sequence in X such that xn  x ∈ X with lim xn − x = 0, then
n→∞

lim xn − x < D[{xn }].

n→∞

Proof. Let α := lim xn − x and yn := α−1 (xn − x). Then yn  0 and
n→∞
lim yn  = 1. By the GGLD property, 1 < D[{yn }]. Hence α < D[{xn }].
n→∞

A Banach space X is said to have the semi-Opial condition (SO in short) if for
any nonconstant bounded sequence {xn } in X with xn − xn+1 → 0, there exists
a subsequence {xnk } such that xnk  x ∈ X and lim xnk −x < diam({xn }).
k→∞

The following theorem shows that for reflexive Banach spaces, the SO con-
dition is more general than the GGLD property.
Theorem 3.7.4 Every reflexive Banach space with the GGLD property satisfies
the SO condition.
3.7. GGLD property 173

Proof. Let {xn } be a bounded sequence in X such that xn+1 − xn  → 0.

Because X is reflexive, there exists a subsequence {xnk } of {xn } such that
xnk  x ∈ X. We may assume that r := lim xnk − x > 0, otherwise the
k→∞
result follows immediately.
Now set yk := r−1 (xnk −x). From the GGLD property of space X, we obtain
1 < D[{yk }]. Therefore,

r < lim sup(lim sup xnj − xnk ) ≤ diam({xnk }) ≤ diam({xn }),

k→∞ j→∞

which completes the proof.

Let us give some examples concerning the GGLD property.

Example 3.7.5 The Banach space Xβ := ( 2 , · β ), where

√ xβ = max{x2 ,
βx∞} as has the SO condition for 1 < β < 2, but if 2 < β, Xβ does not
have normal structure and hence Xβ cannot have the GGLD property.

Example 3.7.6 Consider the James space J that consists of sequences x =

{xn } ∈ co such that

xJ = sup{(xp1 − xp2 )2 + (xp2 − xp3 )2 + · · · + (xpn−1 − xpn )2 } < ∞,

where the supremum is taken over all increasing sequences of positive integers
{pi }. The James space J fails to be uniformly convex in every direction (in fact,
does not have normal structure), but J satisfies the GGLD property.

Example 3.7.7 Consider the classic space co of sequences with norm  · 

 1/2
∞
defined by x = x2∞ + i=1 x2i /2i . Then (co ,  · ) is uniformly convex
in every direction, but co fails to have the GGLD property.

Bibliographic Notes and Remarks

Many of the topics and techniques in Section 3.1 follow from Bose and Laskar
[21], Downing and Kirk [49], Goebel and Kirk [59], and Lan and Webb [95].
The results of Section 3.2 are due to Dalby [43], Lami Dozo [50], Lin, Tan,
and Xu [101], and Suzuki [152].
The notion of normal structure was introduced by Brodskii and Milman in
[23]. The uniformly convex and uniformly smooth Banach spaces enjoy uni-
formly normal structure. The results of Sections 3.3∼3.4 are well described in
Goebel and Kirk [59] and Prus [121]. Theorem 3.4.20 is proved in Casini and
Maluta [36].
The normal structure always implies weak normal structure. The results of
Section 3.5 may be found in Kim and Kim [86] and Lin, Tan, and Xu [101].
174 3. Geometric Coefficients of Banach Spaces

The Maluta constant given in Section 3.6 is due to Maluta [103]. The results
described in Section 3.6 can be found in Benavides and Acedo [15], Lin, Tan,
and Xu [101], and Maluta [103]. Theorem 3.6.6 is proved in Benavides [14].

Exercises

3.1 Let H be a Hilbert space and ϕ : H → (−∞, ∞] be a convex, lower semi-

continuous, and proper function. For λ > 0 and x ∈ H, set
 
λ
ϕλ (x) := inf ϕ(y) + x − y2 .
y∈H 2

Let ∂ϕ be the subdifferential of ϕ and Jλ = (I + λ∂ϕ)−1 . Show that

(a) ϕλ is convex and ϕλ (x) = ϕ(Jλ x) + λ2 (∂ϕ)λ (x)2 .
(b) ∂(ϕλ ) = (∂ϕ)λ ; in particular, ϕλ is continuously differentiable and
has Lipschitz continuous derivative.
(c) ϕλ (x) increases to ϕ(x) as λ ↓ 0.

3.2 Let X be a Banach space, and x1 , · · · , xn a finite number of points in X.

Define
1
n
ϕ(z) = xi − z2 for all z ∈ X.
n i=1
Show that ϕ is a convex function and that, if X is reflexive, ϕ attains its
minimum.
3.3 Let H be a Hilbert space. Let C = {x ∈ H : x − a ≤ r} and D = {x ∈
H : x ≤ a + r} be two sets, where a = 0 and r > 0. Let PC and
PD be metric projection mappings, respectively. Show that PC PD is a
nonexpansive retraction of H onto C that is different from PC .
3.4 Let C be a nonempty closed convex subset of a Hilbert space H,
PC : H → C the metric projection mapping onto C, and {xn } a
sequence in H such that xn  x. Show that the asymptotic center of
{xn } with respect to C is PC x.
3.5 Let X be a Banach space. Show that X ∗ has a Fréchet differentiable norm
iff X is reflexive and strictly convex, and has the following property:
if xn  x and xn  → x, then {xn } converges strongly to x.
Chapter 4

Existence Theorems in
Metric Spaces

In this chapter, we study asymptotic fixed point theorems for contraction

mappings and for mappings that are more general than contraction mappings
in metric spaces.

4.1 Contraction mappings and their

generalizations
In this section, we establish a fundamental asymptotic fixed point theorem that
is known as the “Banach contraction principle” and further we give its genera-
lizations in metric spaces.

By an asymptotic fixed point theorem for the mapping T , we mean a theorem

that guarantees the existence of a fixed point of T , if the iterative T n possess
certain properties. Before to establish the Banach contraction principle, we
discuss some basic definitions and results:
Let (X, d) be a metric space and let Lip(X) denote the class of mappings
T : X → X such that
 
d(T n x, T n y)
n
σ(T ) = sup : x, y ∈ X, x = y < ∞
d(x, y)
for all n ∈ N.

Members of Lip(X) are called Lipschitzian mappings and σ(T n ) is the

Lipschitz constant of T n . Note that σ(T ) = 0 if and only if T is constant
on X. For two Lipschitzian mappings T : X → X and S : X → X such that
S(X) ⊆ Dom(T ), we have

σ(T ◦ S) ≤ σ(T )σ(S).

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 4,

c Springer Science+Business Media, LLC, 2009
176 4. Existence Theorems in Metric Spaces

It is clear that the mapping T ∈ Lip(X) if there exists a constant Ln ≥ 0

such that

d(T n x, T n y) ≤ Ln d(x, y) for all x, y ∈ X and n ∈ N. (4.1)

Moreover, the smallest constant Ln for which (4.1) holds is the Lipschitz
constant of T n . A Lipschitzian mapping T : X → X is said to be uniformly
L-Lipschitzian if Ln = L for all n ∈ N. A Lipschitzian mapping is said to be
contraction (nonexpansive) if σ(T ) < 1 (σ(T ) = 1).

The following result plays an important role in proving several existence

theorems in metric spaces.
Proposition 4.1.1 Let (X, d) be a complete metric space and ϕ : X → (−∞, ∞]
a bounded below lower semicontinuous function. Suppose that {xn } is a sequence
in X such that

d(xn , xn+1 ) ≤ ϕ(xn ) − ϕ(xn+1 ) for all n ∈ N0 = N ∪ {0}.

Then {xn } converges to a point v ∈ X and d(xn , v) ≤ ϕ(xn ) − ϕ(v) for all
n ∈ N0 .
Proof. Because

d(xn , xn+1 ) ≤ ϕ(xn ) − ϕ(xn+1 ), n ∈ N0 ,

it follows that {ϕ(xn )} is a decreasing sequence. Moreover, for m ∈ N0


m
d(xn , xn+1 ) ≤ d(x0 , x1 ) + d(x1 , x2 ) + · · · + d(xm , xm+1 )
n=0
≤ ϕ(x0 ) − ϕ(xm+1 )
≤ ϕ(x0 ) − inf ϕ(xn ).
n∈N0

Letting m → ∞, we have
∞

d(xn , xn+1 ) < ∞.
n=0

This implies that {xn } is a Cauchy sequence in X. Because X is complete,

there exists v ∈ X such that lim xn = v. Let m, n ∈ N0 with m > n. Then
n→∞


m−1
d(xn , xm ) ≤ d(xi , xi+1 )
i=n
≤ ϕ(xn ) − ϕ(xm ).

Letting m → ∞, we obtain

d(xn , v) ≤ ϕ(xn ) − lim ϕ(xm ) ≤ ϕ(xn ) − ϕ(v) for all n ∈ N0 .

m→∞
4.1. Contraction mappings and generalizations 177

We now begin with Caristi’s fixed point theorem. To prove it, we need the
following important result.
Theorem 4.1.2 Let X be a complete metric space and ϕ : X → (−∞, ∞] a
proper, bounded below and lower semicontinuous function. Suppose that, for
each u ∈ X with inf ϕ(x) < ϕ(u), there exists a v ∈ X such that
x∈X

u = v and d(u, v) ≤ ϕ(u) − ϕ(v).

Then there exists an x0 ∈ X such that ϕ(x0 ) = inf ϕ(x).

x∈X

Proof. Suppose that inf ϕ(x) < ϕ(y) for every y ∈ X. Let u0 ∈ X with
x∈X
ϕ(u0 ) < ∞. If inf ϕ(x) = ϕ(u0 ), then we are done. Otherwise inf ϕ(x) <
x∈X x∈X
ϕ(u0 ), and there exists a u1 ∈ X such that u0 = u1 and d(u0 , u1 ) ≤ ϕ(u0 ) −
ϕ(u1 ).
Define inductively a sequence {un } in X, starting with u0 . Suppose un−1 ∈
X is known. Then choose un ∈ Sn , where

Sn := {w ∈ X : d(un−1 , w) ≤ ϕ(un−1 ) − ϕ(w)}

such that
1
ϕ(un ) ≤ inf ϕ(w) + {ϕ(un−1 ) − inf ϕ(w)}. (4.2)
w∈Sn 2 w∈Sn

Because un ∈ Sn , we get

d(un−1 , un ) ≤ ϕ(un−1 ) − ϕ(un ), n ∈ N.

Proposition 4.1.1 implies that un → v ∈ X and d(un−1 , v) ≤ ϕ(un−1 ) − ϕ(v).

By hypothesis, there exists a z ∈ X such that z = v and d(v, z) ≤ ϕ(v) − ϕ(z).
Observe that

ϕ(z) ≤ ϕ(v) − d(v, z)

≤ ϕ(v) − d(v, z) + ϕ(un−1 ) − ϕ(v) − d(un−1 , v)
= ϕ(un−1 ) − [d(v, z) + d(un−1 , v)]
≤ ϕ(un−1 ) − d(un−1 , z).

This implies that z ∈ Sn . It follows from (4.2) that

2ϕ(un ) − ϕ(un−1 ) ≤ inf ϕ(w) ≤ ϕ(z).

w∈Sn

Thus,
ϕ(z) < ϕ(v) ≤ lim ϕ(un ) ≤ ϕ(z),
n→∞

a contradiction. Therefore, there exists a point x0 ∈ X such that ϕ(x0 ) =

inf ϕ(x).
x∈X
178 4. Existence Theorems in Metric Spaces

Theorem 4.1.3 (Caristi’s fixed point theorem) – Let X be a complete

metric space and ϕ : X → (−∞, ∞] a proper, bounded below and lower semi-
continuous function. Let T : X → X be a mapping such that

d(x, T x) ≤ ϕ(x) − ϕ(T x) f or all x ∈ X. (4.3)

Then there exists a point v ∈ X such that v = T v and ϕ(v) < ∞.

Proof. Because ϕ is proper, there exists u ∈ X such that ϕ(u) < ∞. Let

C = {x ∈ X : d(u, x) ≤ ϕ(u) − ϕ(x)}.

Then C is a nonempty closed subset of X. We show that C is invariant under

T . For each x ∈ C, we have

d(u, x) ≤ ϕ(u) − ϕ(x)

and hence from (4.3), we have

ϕ(T x) ≤ ϕ(x) − d(x, T x)

≤ ϕ(x) − d(x, T x) + ϕ(u) − ϕ(x) − d(u, x)
= ϕ(u) − [d(x, T x) + d(u, x)]
≤ ϕ(u) − d(u, T x),

and it follows that T x ∈ C.

Suppose, for contradiction, that x = T x for all x ∈ C. Then, for each x ∈ C,
there exists w ∈ C such that

x = w and d(x, w) ≤ ϕ(x) − ϕ(w).

By Theorem 4.1.2, there exists an x0 ∈ C with ϕ(x0 ) = inf ϕ(x). Hence for
x∈C
such an x0 ∈ C, we have

0 < d(x0 , T x0 ) ≤ ϕ(x0 ) − ϕ(T x0 ) ( inf ϕ(x) = ϕ(x0 ) ≤ ϕ(T x0 ))

x∈C
≤ ϕ(T x0 ) − ϕ(T x0 )
= 0,

a contradiction.
Remark 4.1.4 The fixed point of the mapping T in Theorem 4.1.3 need not be
unique.
We now state and prove the Banach contraction principle, which gives a
unique fixed point of the mapping.
Theorem 4.1.5 (Banach’s contraction principle) – Let (X, d) be a com-
plete metric space and T : X → X a contraction mapping with Lipschitz con-
stant k ∈ (0, 1). Then we have the following:
4.1. Contraction mappings and generalizations 179

(a) There exists a unique fixed point v ∈ X.

(b) For arbitrary x0 ∈ X, the Picard iteration process defined by

xn+1 = T xn , n ∈ N0

converges to v.
(c) d(xn , v) ≤ (1 − k)−1 k n d(x0 , x1 ) for all n ∈ N0 .
Proof. (a) Define the function ϕ : X → R+ by ϕ(x) = (1 − k)−1 d(x, T x),
x ∈ X. Hence ϕ is a continuous function. Because T is a contraction mapping,

d(T x, T 2 x) ≤ kd(x, T x), x ∈ X, (4.4)

which implies that

d(x, T x) − kd(x, T x) ≤ d(x, T x) − d(T x, T 2 x).

Hence
1
d(x, T x) ≤ [d(x, T x) − d(T x, T 2 x)]
1−k
= ϕ(x) − ϕ(T x). (4.5)

Let x be an arbitrary element in X and define the sequence {xn } in X by

xn = T n x, n ∈ N0 .

From (4.5), we have

d(xn , xn+1 ) ≤ ϕ(xn ) − ϕ(xn+1 ), n ∈ N0 ,

and it follows from Proposition 4.1.1 that

lim xn = v ∈ X
n→∞

and

d(xn , v) ≤ ϕ(xn ), n ∈ N0 . (4.6)

Because T is continuous and xn+1 = T xn , it follows that v = T v. Suppose z is

another fixed point of T . Then

0 < d(v, z) = d(T v, T z) ≤ kd(v, z) < d(v, z),

a contradiction. Hence T has unique fixed point v ∈ X.

(b) It follows from part (a).
(c) From (4.4) we have that ϕ(xn ) ≤ k n ϕ(x0 ). This implies from (4.6) that
d(xn , v) ≤ k n ϕ(x0 ).

Let us give some examples of contraction mappings.

180 4. Existence Theorems in Metric Spaces

Example 4.1.6 Let X = [a, b] and T : X → X a mapping such that T is


differentiable at every x ∈ (a, b) such that |T (x)| ≤ k < 1. Then, by the mean
value theorem, if x, y ∈ X, there is a point ξ between x and y such that

T x − T y = T (ξ)(x − y).

Thus,

|T x − T y| = |T (ξ)| |x − y| ≤ k|x − y|.
Therefore, T is contraction and it has a unique fixed point.

Example 4.1.7 Let X = R and T : R → R a mapping defined by

1
Tx = x + 1, x ∈ R.
2
Then T is contraction and F (T ) = {2}.

The following example shows that there exists a mapping that is not a con-
traction, but it has a unique fixed point.

Example 4.1.8 Let X = [0, 1] and T : [0, 1] → [0, 1] a mapping defined by

T x = 1 − x, x ∈ [0, 1].

Then T has a unique fixed point 1/2, but T is not a contraction.

Let (X, d) be a metric space. Then a mapping T : X → X is said to be

contractive if
d(T x, T y) < d(x, y) for all x, y ∈ X, x =
 y.
It is clear that the class of contractive mappings falls between the class of
contraction mappings and that of nonexpansive mappings.

Observation
• A contractive mapping can have at most one fixed point.

The contractive mapping may not have a fixed point. It can be seen from
the following example.

Example 4.1.9 Let X be the space c0 consisting of all real sequences x = {xi }
with lim xi = 0 and d(x, y) = x − y = sup |xi − yi |, x = {xi }, y = {yi } ∈ c0 .
i→∞ i∈N
Let BX = {x ∈ c0 : x ≤ 1}. For each x ∈ BX , define

T (x1 , x2 , · · · , xi , · · · ) = (y1 , y2 , · · · , yi , · · · ),

where y1 = (1 + x)/2 and yi = (1 − 1/2i+1 )xi−1 for i = 2, 3, · · · . Note that

|y1 | ≤ 1 and |yi | ≤ |xi−1 | ≤ 1 for all i = 2, 3, · · · . Hence T : BX → BX .
4.1. Contraction mappings and generalizations 181

Suppose x and y are two distinct points in BX . Then

   
x − y 1
T x − T y = sup , 1 − i+1 |xi−1 − yi−1 | : i = 2, 3, · · ·
2 2
   
x − y 1
≤ sup , 1 − i+1 |xi−1 − yi−1 | : i = 2, 3, · · ·
2 2
< x − y.

Suppose that there is a point v ∈ BX such that T v = v. Then v1 = (1+ v)/2 > 0
and for i ≥ 2  
1
|vi | = 1 − i+1 |vi−1 |.
2
This implies for all i ≥ 2
 
1
|vi | = 1− |vi−1 |
2i+1
  
1 1
= 1 − i+1 1 − i |vi−2 |
2 2
...
i−2  
1
= 1− |v1 |
2i+1−k
k=0
 
i−2 
1
≥ 1− |v1 |
2i+1−k
k=0
 
i+1 
1 3
= 1− j
|v1 | > |v1 |.
j=3
2 4

This is not possible, because vi → 0 as i → ∞. Thus, T has no fixed point

in BX .
We note that completeness and boundedness of a metric space do not ensure
the existence of fixed points of contractive mappings. However, contractive
mappings always have fixed points in compact metric spaces.
Theorem 4.1.10 Let X be a compact metric space and T : X → X a con-
tractive mapping. Then T has a unique fixed point v in X. Moreover, for each
x ∈ X, the sequence {T n x} of iterates converges to v.
Proof. For each x ∈ X, define a function ϕ : X → R+ by ϕ(x) = d(x, T x).
Then ϕ is continuous on X and by compactness of X, ϕ attains its minimum,
say ϕ(v), at v ∈ X. Then ϕ(v) = min ϕ(x). If v = T v, then
x∈X

ϕ(T v) = d(T v, T 2 v) < d(v, T v) = ϕ(v),

a contradiction. Hence v = T v. Uniqueness of v follows from the contractive

condition of T .
182 4. Existence Theorems in Metric Spaces

Now, let x0 ∈ X and define a sequence {xn } in X by xn = T n x0 for all

n ∈ N. Set cn := d(T n x0 , v), n ∈ N0 . Because

cn+1 = d(T n+1 x0 , v) < d(T n x0 , v) = cn ,

{cn } is a nonincreasing sequence in R+ . Hence lim cn exists. Suppose lim cn =

n→∞ n→∞
c ≥ 0. Assume that c > 0. Because X is compact, there exists a subsequence
{xni } of {xn } such that xni → z ∈ X. Observe that

0 < c = lim cni = lim d(T ni x0 , v) = d(z, v),

i→∞ i→∞

i.e., z = v. Because T is contractive and continuous,

c = lim d(T ni +1 x0 , v) = d(T z, v) < d(z, v) = c,

i→∞

a contradiction. Thus, c = 0, i.e., z = v. This means that every conver-

gent subsequence of {T n x0 } must converge to v. Therefore, {T n x0 } converges
to v.

The following example shows that in general, even in a Hilbert space for
contractive mappings we cannot have that T n x → x0 for every x ∈ BX and
x0 = T x0 .
Example 4.1.11 Let X = 2 = {(x1 , x2 , · · · , xi , · · · ) : xi real for each i ∈ N
∞ ∞
and i=1 |xi |2 < ∞} and BX = {x ∈ X : x2 = ( i=1 |xi 2 )1/2 ≤ 1}. Define
a mapping T : BX → BX by

T x = (0, α1 x1 , α2 x2 , · · · , αi xi · · · ), x = (x1 , x2 , · · · , xi , · · · ) ∈ BX ,

where α1 = 1; αi = (1−1/i2 ), i = 2, 3, · · · . It is easy to see that T is contractive

with fixed point (0, 0, · · · , 0, · · · ).
Now, let x = (1, 0, · · · , 0, · · · ) ∈ BX , then
n
T n x = (0, 0, · · · , αi , 0, · · · ) for all n ∈ N.
i=1

Thus,
n+2 1
T n x = → as n → ∞,
2(n + 1) 2
and hence T n x  0.
We now consider some important generalizations of the Banach contraction
principle in which the Lipschitz constant k is replaced by some real-valued con-
trol function.
Theorem 4.1.12 (Boyd and Wong’s fixed point theorem) – Let X be a
complete metric space and T : X → X a mapping that satisfies

d(T x, T y) ≤ ψ(d(x, y)) f or all x, y ∈ X, (4.7)

4.1. Contraction mappings and generalizations 183

where ψ : R+ → R+ is upper semicontinuous function from the right (i.e.,

λi ↓ λ ≥ 0 ⇒ lim sup ψ(λi ) ≤ ψ(λ)) such that ψ(t) < t for each t > 0. Then T
i→∞
has a unique fixed point v ∈ X. Moreover, for each x ∈ X, lim T n x = v.
n→∞

Proof. Fix x ∈ X and define a sequence {xn } in X by xn = T n x, n ∈ N0 . Set

dn := d(xn , xn+1 ). We divide the proof into three steps:
Step 1. lim dn = 0.
n→∞
Note

dn+1 = d(xn+1 , xn+2 ) = d(T xn , T xn+1 ) ≤ ψ(dn ), n ∈ N0 .

Hence {dn } is monotonic decreasing and bounded below. Hence lim dn exists.
n→∞
Let lim dn = δ ≥ 0. Assume that δ > 0. By the right continuity of ψ,
n→∞

δ = lim dn+1 ≤ lim ψ(dn ) ≤ ψ(δ) < δ,

n→∞ n→∞

so δ = 0.
Step 2. {xn } is Cauchy sequence.
Assume that {xn } is not Cauchy. Then there exist ε > 0 and integers
mk , nk ∈ N0 such that mk > nk ≥ k and

d(xnk , xmk ) ≥ ε for k = 0, 1, 2, · · · .

Also, choosing mk as small as possible, it may be assumed that

d(xmk −1 , xnk ) < ε.

Hence for each k ∈ N0 , we have

ε ≤ d(xmk , xnk ) ≤ d(xmk , xmk −1 ) + d(xmk −1 , xnk )

≤ d(xmk −1 , xmk ) + ε
= dmk −1 + ε,

and it follows from the fact dmk → 0 that lim d(xmk , xnk ) = ε. Observe that
k→∞

d(xmk , xnk ) ≤ d(xmk , xmk +1 ) + d(xmk +1 , xnk +1 ) + d(xnk +1 , xnk )

≤ dmk + ψ(d(xmk , xnk )) + dnk .

Letting k → ∞ and using the upper semicontinuity of ψ from the right, we

obtain
ε = lim d(xmk , xnk ) ≤ lim ψ(d(xmk , xnk )) ≤ ψ(ε),
k→∞ k→∞

which is a contradiction. Hence {xn } is a Cauchy sequence in X.

Step 3. Existence and uniqueness of fixed points.
184 4. Existence Theorems in Metric Spaces

Because {xn } is Cauchy and X is complete, lim xn = v ∈ X. By continuity

n→∞
of T, we have v = T v. Uniqueness of v easily follows from condition (4.7).

Let Φ denote the class of all mappings ϕ : R+ → R+ satisfying:

(i) ϕ is continuous,
(ii) ϕ(t) < t for all t > 0.

As an immediate consequence of the Boyd-Wong’s fixed point theorem, we

have the following important result, which will be useful in establishing existence
theorems concerning asymptotic contraction mappings.
Corollary 4.1.13 Let X be a complete metric space and T : X → X a mapping
that satisfies
d(T x, T y) ≤ ϕ(d(x, y)) f or all x, y ∈ X,
where ϕ ∈ Φ. Then T has a unique fixed point v ∈ X. Moreover, for each
x ∈ X, lim T n x = v.
n→∞

We now introduce a wider class of mappings that we call “asymptotic con-

tractions.”
Definition 4.1.14 Let (X, d) be a metric space. A mapping T : X → X is said
to be an asymptotic contraction if for each n ∈ N
d(T n x, T n y) ≤ ϕn (d(x, y)) f or all x, y ∈ X, (4.8)
where ϕn : R+ → R+ and ϕn → ϕ ∈ Φ uniformly on the range of d.
The following theorem shows that asymptotic contractions have unique fixed
points.
Theorem 4.1.15 Let X be a complete metric space and T : X → X a con-
tinuous asymptotic contraction for which the mappings ϕn in (4.8) are also
continuous. Assume also that some orbit of T is bounded. Then T has a unique
fixed point v ∈ X and for each x ∈ X, {T n x} converges to v.
Proof. Because the sequence {ϕi } is uniformly convergent, it follows that ϕ is
continuous. For any x, y ∈ X, x = y, we have
lim sup d(T n x, T n y) ≤ lim sup ϕn (d(x, y)) = ϕ(d(x, y)) < d(x, y).
n→∞ n→∞

If there exist x, y ∈ X and ε > 0 such that lim sup d(T n x, T n y) = ε, then
n→∞
there exists k ∈ N such that ϕ(d(T k x, T k y)) < ε because ϕ is continuous, and
ϕ(ε) < ε. It follows that
lim sup d(T n x, T n y) = lim sup d(T n (T k x), T n (T k y)
n→∞ n→∞
≤ lim sup ϕn (d(T k x, T k y))
n→∞
= ϕ(d(T k x, T k y)) < ε,
4.1. Contraction mappings and generalizations 185

a contradiction. Hence

lim d(T n x, T n y) = 0 for any x, y ∈ X. (4.9)

n→∞

Thus, all sequences of the Picard iterates defined by T , are equi-convergent and
bounded.
Now let z0 ∈ X be arbitrary, {zn } be a sequence of Picard iterates of T at the
point z0 , C = {zn } and Fn = {x ∈ C : d(x, T k x) ≤ 1/n, k = 1, · · · , n}. Because
{zn } is bounded, C is bounded. It follows from (4.9) that Fn is nonempty.
Because T is continuous, we have Fn is closed, for any n. Also, we have Fn+1 ⊆
Fn . Let {xn } and {yn } be two arbitrary sequences such that xn , yn ∈ Fn . Let
{nj } be a sequence of integers such that lim d(xnj , ynj ) = lim sup d(xn , yn ).
j→∞ n→∞
Observe that

lim d(xnj , ynj ) ≤ lim (d(xnj , T nj xnj ) + d(T nj xnj , T nj ynj ) + d(ynj , T nj ynj ))
j→∞ j→∞
= lim ϕnj (d(xnj , ynj ))
j→∞
= ϕ( lim d(xnj , ynj )),
j→∞

and hence lim d(xnj , ynj ) = ϕ( lim d(xnj , ynj )), which implies that
j→∞ j→∞
lim d(xnj , ynj ) = 0, because C is bounded. Thus, lim sup d(xn , yn ) = 0 and
j→∞ n→∞
hence lim d(xn , yn ) = 0. This implies that lim diam(Fn ) = 0. By the com-
n→∞ n→∞
pleteness of C, it follows that there exists v ∈ X such that ∩∞
n=1 Fn = {v}.
Because d(v, T v) ≤ 1/n for any n, we have T v = v. From (4.9), we have
lim d(T n x, v) = 0 for any x ∈ X.
n→∞

We now study an important generalization of the Boyd and Wong’s fixed

point theorem in which the control function ϕ is extended in a different direction.
Interestingly, in the following result the continuity condition on ϕ is replaced
by lim ϕn (t) = 0 for all t > 0.
n→∞

Theorem 4.1.16 (Matkowski’s fixed point theorem) – Let X be a com-

plete metric space and T : X → X a mapping that satisfies

d(T x, T y) ≤ ψ(d(x, y)) f or all x, y ∈ X,

where ψ : (0, ∞) → (0, ∞) is nondecreasing and satisfies lim ψ n (t) = 0 for

n→∞
all t > 0. Then T has a unique fixed point v ∈ X and for each x ∈ X,
lim T n x = v.
n→∞

Proof. Fix x0 ∈ X and let xn = T n x0 , n ∈ N. Observe that

0 ≤ lim sup d(xn , xn+1 ) ≤ lim sup ψ n (d(x0 , x1 )) = 0.

n→∞ n→∞
186 4. Existence Theorems in Metric Spaces

Hence lim d(xn , xn+1 ) = 0. Because ψ n (t) → 0 for t > 0, ψ(s) < s for any
n→∞
s > 0. Because lim d(xn , xn+1 ) = 0, given any ε > 0, it is possible to choose n
n→∞
such that
d(xn+1 , xn ) ≤ ε − ψ(ε).
Now for z ∈ Bε [xn ] = {x ∈ X : d(x, xn ) ≤ ε}, we have

d(T z, xn ) ≤ d(T z, T xn ) + d(T xn , xn )

≤ ψ(d(z, xn )) + d(xn+1 , xn )
≤ ψ(ε) + (ε − ψ(ε)) = ε.

Therefore, T : Bε [xn ] → Bε [xn ] and it follows that d(xm , xn ) ≤ ε for all m ≥ n.

Hence {xn } is a Cauchy sequence. The conclusion of the proof follows as in
Theorem 4.1.12.

We now introduce the concept of nearly Lipschitzian mappings:

Let (X, d) be a metric space and fix a sequence {an } in R+ with an → 0.
A mapping T : X → X is said to be nearly Lipschitzian with respect to {an } if
for each n ∈ N, there exists a constant kn ≥ 0 such that

d(T n x, T n y) ≤ kn (d(x, y) + an ) for all x, y ∈ C. (4.10)

The infimum of constants kn for which (4.10) holds is denoted by η(T n ) and
called the nearly Lipschitz constant.
Notice that
 
d(T n x, T n y)
n
η(T ) = sup : x, y ∈ C, x = y .
d(x, y) + an
A nearly Lipschitzian mapping T with sequence {(η(T n ), an )} is said to be
(i) nearly contraction if η(T n ) < 1 for all n ∈ N,
(ii) nearly nonexpansive if η(T n ) = 1 for all n ∈ N,
(iii) nearly asymptotically nonexpansive if η(T n ) ≥ 1 for all n ∈ N and
limn→∞ η(T n ) ≤ 1,
(iv) nearly uniformly k-Lipschitzian if η(T n ) ≤ k for all n ∈ N,
(v) nearly uniformly k-contraction if η(T n ) ≤ k < 1 for all n ∈ N.

Example 4.1.17 Let X = [0, 1] with the usual metric d(x, y) = |x − y| and
T : X → X a mapping defined by

1/2 if x ∈ [0, 1/2],
Tx =
0 if x ∈ (1/2, 1].

Thus, T is discontinuous and non-Lipschitzian. However, it is nearly non-

expansive mapping. Indeed, for a sequence {an } with a1 = 1/2 and an → 0, we
have
d(T x, T y) ≤ d(x, y) + a1 for all x, y ∈ X
4.1. Contraction mappings and generalizations 187

and
d(T n x, T n y) ≤ d(x, y) + an for all x, y ∈ X and n ≥ 2,
because
1
T nx = for all x ∈ [0, 1] and n ≥ 2.
2
We now develop a technique for studying the existence and uniqueness of
fixed points of nearly Lipschitzian mappings.

Theorem 4.1.18 Let X be a complete metric space and T : X → X a contin-

uous nearly Lipschitzian mapping with sequence {(η(T n ), an )}, i.e., for a fixed
sequence {an } in R+ with an → 0 and for each n ∈ N, there exists a constant
η(T n ) > 0 such that

d(T n x, T n y) ≤ η(T n )(d(x, y) + an ) f or all x, y ∈ X.

Suppose η∞ (T ) = lim sup[η(T n )]1/n < 1. Then we have the following:

n→∞
(a) T has a unique fixed point v ∈ X.
(b) For each x ∈ X, the sequence {T n x} converges to v.
∞
(c) d(T n x, v) ≤ i=n η(T i )(d(x, T x)+M ) for all n ∈ N, where M = sup an .
n∈N

Proof. (a) Fix x ∈ X and let xn = T n x, n ∈ N. Set dn := d(xn , xn+1 ). Hence

dn = d(T n x, T n+1 x) ≤ η(T n )(d(x, T x) + an ),

which implies that

∞
 ∞

dn ≤ (d(x, T x) + M ) η(T n )
n=1 n=1

for some M > 0, because lim an = 0. By the Root Test for convergence of
n→∞
∞
series, if η∞ (T ) = lim sup[η(T n )]1/n < 1, then n=1 η(T n ) < ∞. It follows
n→∞
∞
that n=1 dn < ∞ and hence {xn } is a Cauchy sequence. Thus, lim xn exists
n→∞
(say v ∈ X). By the continuity of T , v is fixed point of T . Let w be another
fixed point T . Then
∞
 ∞
 ∞

∞= d(v, w) = d(T n v, T n w) ≤ η(T n )(d(v, w) + an )
n=1 n=1 n=1
∞

≤ (d(u, w) + M ) η(T n ) < ∞,
n=1

a contradiction, hence T has a unique fixed point v ∈ X.

(b) It follows from part (a).
188 4. Existence Theorems in Metric Spaces

(c) If m ∈ N, we have

d(xn , xn+m ) = d(T n x, T n+m x)


n+m−1
≤ d(T i x, T i+1 x)
i=n

n+m−1
≤ η(T i )(d(x, T x) + ai )
i=n

n+m−1
≤ η(T i )(d(x, T x) + M ).
i=n

Letting m → ∞, we obtain
∞
d(xn , v) ≤ i=n η(T i )(d(x, T x) + M ).

Remark 4.1.19 In the case of a nearly uniformly k-Lipschitzian mapping, we

have
lim sup[η(T n )]1/n = lim sup(k)1/n = 1.
n→∞ n→∞

Therefore, the assumptions of Theorem 4.1.18 do not hold for nearly uniformly
k-Lipschitzian mappings.

4.2 Multivalued mappings

Let A be a nonempty subset of a metric space X. For x ∈ X, define

d(x, A) = inf{d(x, y) : y ∈ A}.

Let CB(X) denote the set of nonempty closed bounded subsets of X and K(X)
denote the set of nonempty compact subsets of X. It is clear that K(X) is
included in CB(X).

For A, B ∈ CB(X), define

δ(A, B) = sup{d(x, B) : x ∈ A},

H(A, B) = max{δ(A, B), δ(B, A)} = max{sup d(a, B), sup d(b, A)}.
a∈A b∈B

Example 4.2.1 Let X = R, A = [1, 2] and B = [2, 3]. Then

δ(A, B) = sup d(a, B) = 1 and δ(B, A) = sup d(b, A) = 1.

a∈A b∈B

Hence H(A, B) = max{δ(A, B), δ(B, A)} = 1.

Note that set distance δ is not symmetric. However, δ and H have the
following properties:
4.2. Multivalued mappings 189

Proposition 4.2.2 Let (X, d) be a metric space. Let A, B, C ∈ CB(X). Then

we have the following:
(a) δ(A, B) = 0 ⇔ A ⊂ B.
(b) B ⊂ C ⇒ δ(A, C) ≤ δ(A, B).
(c) δ(A ∪ B, C) = max{δ(A, C), δ(B, C)}.
(d) δ(A, B) ≤ δ(A, C) + δ(C, B).
Proof. (a) By the definition δ, we have

δ(A, B) = 0 ⇔ sup d(x, B) = 0

x∈A
⇔ d(x, B) = 0 for all x ∈ A.

Because B is closed in X,

d(x, B) = 0 ⇔ x ∈ B.

Thus,
δ(A, B) = 0 ⇔ A ⊂ B.
(b) Observe that

B ⊂ C ⇒ d(x, C) ≤ d(x, B) for all x ∈ X.

(c) Observe that

δ(A ∪ B, C) = sup d(x, C) = max{sup d(x, C), sup d(x, C)}.

x∈A∪B x∈A x∈B

(d) Let a ∈ A, b ∈ B and c ∈ C. Then

d(a, b) ≤ d(a, c) + d(c, b),

which implies that

d(a, B) ≤ d(a, c) + d(c, B)
and hence
d(a, B) ≤ d(a, c) + δ(C, B).
Because c ∈ C is arbitrary, we have

d(a, B) ≤ d(a, C) + δ(C, B).

Similarly, because a ∈ A is arbitrary, we have

δ(A, B) ≤ δ(A, C) + δ(C, B).

Proposition 4.2.3 Let (X, d) be a metric space. Then H is a metric on

CB(X).
190 4. Existence Theorems in Metric Spaces

Proof. By the definition of H, we have

H(A, B) ≥ 0 and H(A, B) = H(B, A).

Observe that

H(A, B) = 0 ⇔ max{δ(A, B), δ(B, A)} = 0

⇔ δ(A, B) = 0 and δ(B, A) = 0
⇔ A ⊂ B and B ⊂ A
⇔ A = B.

Using Proposition 4.2.2, we obtain

H(A, B) = max{δ(A, B), δ(B, A)}

≤ max{δ(A, C) + δ(C, B), δ(B, C) + δ(C, A)
≤ max{δ(A, C), δ(C, A)} + max{δ(B, C), δ(C, B)}
= H(A, C) + H(C, B).

The metric H on CB(X) is called the Hausdorff metric. The metric H

depends on the metric d. It is easy to see that the completeness of (X, d)
implies the completeness of (CB(X), H) and (K(X), H).
Remark 4.2.4 Let A, B ∈ CB(X) and a ∈ A. Then for ε > 0, there must
exist a point b ∈ B such that d(a, b) ≤ H(A, B) + ε.
The following proposition gives a characteristic property of the Hausdorff
metric that will be used in Section 8.1.
Proposition 4.2.5 Let X be a metric space. Then

H(A ∪ B, C ∪ D) ≤ max{H(A, C), H(B, D)} for all A, B, C, D ∈ CB(X).

Proof. Observe that

δ(A ∪ B, C ∪ D) = max{δ(A, C ∪ D), δ(B, C ∪ D)}

≤ max{δ(A, C), δ(B, D)}
≤ max{H(A, C), H(B, D)}.

Similarly, we have

δ(C ∪ D, A ∪ B) ≤ max{H(A, C), H(B, D)}

By definition of H, we have

H(A ∪ B, C ∪ D) = max{δ(A ∪ B, C ∪ D), δ(C ∪ D, A ∪ B)}

≤ max{H(A, C), H(B, D)} for all A, B, C, D ∈ CB(X).

Let F (X) denote the family of nonempty closed subsets of a metric space
(X, d). Then we have
4.2. Multivalued mappings 191

Proposition 4.2.6 Let C be a nonempty subset of a metric space (X, d).

Suppose the mapping T : C → F (X) is an upper semicontinuous at x0 ∈ C.
Then the mapping ϕ : C → R+ defined by ϕ(x) = d(x, T x), x ∈ C is lower
semicontinuous at x0 .
Proof. Let ε > 0. By the upper semicontinuity of T at x0 , there exists δ > 0
such that y ∈ Bδ [x0 ] ∩ C implies T y lies in an ε/4-neighborhood of T x0 , and
moreover we may suppose δ ≤ ε/4. Select u ∈ T y such that
ε
d(y, u) ≤ d(y, T y) +
2
and select v ∈ T x0 so that d(u, v) ≤ ε/4. Then
 
ε
d(x0 , T x0 ) − d(y, T y) + ≤ d(x0 , T x0 ) − d(y, u)
2
≤ d(x0 , v) − d(y, u)
≤ d(x0 , y) + d(y, u) + d(u, v) − d(y, u)
≤ d(x0 , y) + d(u, v)
ε ε
≤ δ+ = ,
4 2
and hence
d(x0 , T x0 ) ≤ d(y, T y) + ε.

Therefore, ϕ is lower semicontinuous at x0 .

We now introduce the class of multivalued contraction mappings and obtain

a fixed point theorem for this class of mappings:
Let T be a mapping from a metric space (X, d) into CB(X). Then T is said
to be Lipschitzian if there exists a constant k > 0 such that

H(T x, T y) ≤ kd(x, y) for all x, y ∈ X.

A multivalued Lipschitzian mapping T is said to be contraction (nonexpansive)

if k < 1 (k = 1). Let F (T ) denote the set of fixed points of T , i.e., F (T ) =
{x ∈ X : x ∈ T x}.

Theorem 4.2.7 (Nadler’s fixed point theorem) – Let X be a complete

metric space and T : X → CB(X) a contraction mapping. Then T has a fixed
point in X.

Proof. Let k, 0 < k < 1 be the Lipschitz constant of T . Let x0 ∈ X and

x1 ∈ T x0 . By Remark 4.2.4, there must exist x2 ∈ T x1 such that

d(x1 , x2 ) ≤ H(T x0 , T x1 ) + k.

Similarly, there exists x3 ∈ T x2 such that

d(x2 , x3 ) ≤ H(T x1 , T x2 ) + k 2 .
192 4. Existence Theorems in Metric Spaces

Thus, there exists a sequence {xn } in X such that xn+1 ∈ T xn and

d(xn , xn+1 ) ≤ H(T xn−1 , T xn ) + k n for all n ∈ N.
Notice for each n ∈ N, xn+1 ∈ T xn and so
d(xn , xn+1 ) ≤ H(T xn−1 , T xn ) + k n
≤ kd(xn−1 , xn ) + k n
≤ k[kd(xn−2 , xn−1 ) + k n−1 ] + k n
≤ k 2 d(xn−2 , xn−1 ) + 2k n
···
≤ k n d(x0 , x1 ) + nk n .
∞ ∞
Because n=0 k n < ∞ and n=0 nk n < ∞, we have
∞
 ∞
 ∞

d(xn , xn+1 ) ≤ d(x0 , x1 ) kn + nk n < ∞.
n=0 n=0 n=0

Hence {xn } is a Cauchy sequence. By completeness of X, there exists v ∈ X

such that lim xn = v. Again, by the continuity of T ,
n→∞

lim H(T xn , T v) = 0.
n→∞

Because xn+1 ∈ T xn ,
lim d(xn+1 , T v) = 0,
n→∞

which implies that d(v, T v) = 0. Because T v is closed, it follows that

v ∈ T v.
Example 4.2.8 Let X = [0, 1] and f : [0, 1] → [0, 1] a mapping such that

x/2 + 1/2, 0 ≤ x ≤ 1/2,
f (x) =
−x/2 + 1, 1/2 ≤ x ≤ 1.

Define T : X → 2X by T x = {f (x)} ∪ {0}, x ∈ X. Then T is a multivalued

contraction mapping with F (T ) = {0, 2/3}.
Remark 4.2.9 Example 4.2.8 shows that the fixed point of a multivalued con-
traction mapping is not necessarily unique.
We now discuss a stability result (Theorem 4.2.11) for multivalued contrac-
tion mappings.
Proposition 4.2.10 Let X be a complete metric space and let S, T : X →
CB(X) be two contraction mappings each having Lipschitz constant k < 1, i.e.,
H(Sx, Sy) ≤ kd(x, y) and H(T x, T y) ≤ kd(x, y) for all x, y ∈ X.
Then H(F (S), F (T )) ≤ (1 − k)−1 sup H(Sx, T x).
x∈X
4.2. Multivalued mappings 193

∞
Proof. Let ε > 0 and c > 0 be such that c n=1 nk n < 1. For x0 ∈ F (S),
select x1 ∈ T x0 such that
d(x0 , x1 ) ≤ H(Sx0 , T x0 ) + ε.
Because H(T x0 , T x1 ) ≤ kd(x0 , x1 ), it is possible to select x2 ∈ T x1 such that
cεk
d(x1 , x2 ) ≤ H(T x0 , T x1 ) +
1−k
cεk
≤ kd(x0 , x1 ) + .
1−k
Define {xn } inductively by
cεk n
xn+1 ∈ T xn and d(xn+1 , xn ) ≤ kd(xn , xn−1 ) + .
1−k
Set η := cε/(1 − k). Observe that
d(xn+1 , xn ) ≤ kd(xn , xn−1 ) + ηk n
≤ k(kd(xn−1 , xn−2 ) + ηk n−1 ) + ηk n
≤ k 2 d(xn−1 , xn−2 ) + 2ηk n
···
≤ k n d(x0 , x1 ) + nηk n .
∞ ∞
Because n=1 k n <∞ and n=1 nk n <∞, it follows that {xn } is a Cauchy sequ-
ence in X and it converges to some point v ∈ X. Because lim H(T xn , T v) = 0
n→∞
by continuity of T , it follows from xn+1 ∈ T xn that v ∈ F (T ). Observe that
∞
 ∞
 ∞

d(x0 , v) ≤ d(xn , xn+1 ) ≤ k n d(x0 , x1 ) + η nk n
n=0 n=0 n=0
∞

≤ (1 − k)−1 d(x0 , x1 ) + η nk n
n=0
≤ (1 − k)−1 (d(x0 , x1 ) + ε)
≤ (1 − k)−1 (H(Sx0 , T x0 ) + 2ε).
Interchanging the roles of S and T , we conclude:
For each y0 ∈ F (T ), there exist y1 ∈ Sy0 and u ∈ F (S) such that
d(y0 , u) ≤ (1 − k)−1 (H(Sy0 , T y0 ) + 2ε).

Because ε > 0 is arbitrary, the conclusion follows.

Theorem 4.2.11 Let X be a complete metric space and let Tn : X → CB(X)
(n = 1, 2, · · · ) be contraction mappings each having Lipschitz constant k < 1,
i.e.,
H(Tn x, Tn y) ≤ kd(x, y) f or all x, y ∈ X and n ∈ N.
If lim H(Tn x, T0 x) = 0 uniformly for x ∈ X, then lim H(F (Tn ), F (T0 )) = 0.
n→∞ n→∞
194 4. Existence Theorems in Metric Spaces

Proof. Let ε > 0. Because lim H(Tn x, T0 x) = 0 uniformly for x ∈ X, it is

n→∞
possible to select n0 ∈ N such that
sup H(Tn x, T0 x) ≤ (1 − k)ε for all n ≥ n0 .
x∈X

By Proposition 4.2.10, we have H(F (Tn ), F (T0 )) < ε for all n ≥ n0 .

Next, we extend Nadler’s fixed point theorem for non-self multivalued map-
pings in a metric space. First, we define a metrically convex metric space.
1
Definition 4.2.12 A metric space (X, d) is said to be metrically convex if for
any x, y ∈ X with x = y, there exists z ∈ X, x = y = z such that
d(x, z) + d(z, y) = d(x, y).
We note that in such a space, each two points are the end points of at least
one metric segment. This fact immediately yields a very useful lemma.
Lemma 4.2.13 If C is a nonempty closed subset of a complete and metrically
convex metric space (X, d), then for any x ∈ C, y ∈
/ C, there exists a point
z ∈ ∂C (the boundary of C) such that
d(x, z) + d(z, y) = d(x, y).
Now we are in a position to establish a fundamental result on the existence
of fixed points for non-self multivalued contraction mappings.
Theorem 4.2.14 (Assad and Kirk’s fixed point theorem) – Let (X, d) be
a complete and metrically convex metric space, C a nonempty closed subset of
X, and T : C → CB(X) a contraction mapping, i.e.,
H(T x, T y) ≤ kd(x, y) for all x, y ∈ X,
where k ∈ (0, 1). If T x ⊂ C for each x ∈ ∂C, then T has a fixed point in C.
Proof. We construct a sequence {pn } in C in the following way:
Let p0 ∈ C and p1 ∈ T p0 . If p1 ∈ C, let p1 = p1 . Otherwise, select a point
p1 ∈ ∂C such that
d(p0 , p1 ) + d(p1 , p1 ) = d(p0 , p1 ).
Thus, p1 ∈ C. By Remark 4.2.4, we may choose p2 ∈ T p1 such that
d(p1 , p2 ) ≤ H(T p0 , T p1 ) + k.
Now, if p2 ∈ C, let p2 = p2 , otherwise, let p2 ∈ ∂C such that
d(p1 , p2 ) + d(p2 , p2 ) = d(p1 , p2 ).
Continuing in this manner, we obtain sequences {pn } and {pn } such that for
n ∈ N,
1 The concept of metric convexity was introduced by K. Menger in 1953.
4.2. Multivalued mappings 195

(i) pn+1 ∈ T pn ;
(ii) d(pn+1 , pn ) ≤ H(T pn , T pn−1 ) + k n ,
where pn+1 = pn+1 , if pn+1 ∈ C or
d(pn , pn+1 ) + d(pn+1 , pn+1 ) = d(pn , pn+1 ) if pn+1 ∈
/ C and pn+1 ∈ ∂C. (4.11)
Now, set
P : = {pi ∈ {pn } : pi = pi , i ∈ N};
Q : = {pi ∈ {pn } : pi = pi , i ∈ N}.
Observe that if pi ∈ Q for some i, then pi+1 ∈ P be the boundary condition.
We wish to estimate the distance d(pn , pn+1 ) for n ≥ 2. For this, we consider
three cases:

Case I. pn ∈ P and pn+1 ∈ P .

In this case, we have
d(pn , pn+1 ) = d(pn , pn+1 ) ≤ H(T pn , T pn−1 ) + k n
≤ kd(pn , pn−1 ) + k n .
Case II. pn ∈ P and pn+1 ∈ Q.
By (4.11), we have
d(pn , pn+1 ) ≤ d(pn , pn+1 ) = d(pn , pn+1 )
≤ H(T pn−1 , T pn ) + k n
≤ kd(pn−1 , pn ) + k n .
Case III. pn ∈ Q and pn+1 ∈ P .
By the above observation, two consecutive terms of {pn } cannot be in Q,
hence pn−1 ∈ P and pn−1 = pn−1 . Using this fact, we obtain
d(pn , pn+1 ) ≤ d(pn , pn ) + d(pn , pn+1 )
= d(pn , p n ) + d(pn , pn+1 )
≤ d(pn , pn ) + H(T pn−1 , T pn ) + k n
≤ d(pn , pn ) + αd(pn−1 , pn ) + k n
< d(pn−1 , pn ) + k n
= d(pn−1 , pn ) + k n
≤ H(T pn−2 , T pn−1 ) + k n−1 + k n
≤ kd(pn−2 , pn−1 ) + k n−1 + k n .
The only other possibility, pn ∈ Q, pn+1 ∈ Q cannot occur. Thus, for n ≥ 2,
we have

kd(pn , pn−1 ) + k n , or
d(pn , pn+1 ) = (4.12)
kd(pn−2 , pn−1 ) + k n + k n−1 .
196 4. Existence Theorems in Metric Spaces

Set δ := k −1/2 max{d(p0 , p1 ), d(p1 , p2 )}. We now prove that

d(pn , pn+1 ) ≤ k n/2 (δ + n), n ∈ N. (4.13)

For n = 1
d(p1 , p2 ) ≤ k 1/2 (δ + 1).
For n = 2, we use (4.12) and taking each case separately, we obtain

d(p2 , p3 ) ≤ kd(p1 , p2 ) + k 2
≤ kk 1/2 (δ + 1) + k 2
≤ k(δ + 2);
d(p2 , p3 ) ≤ kd(p0 , p1 ) + k 2 + k
≤ k(k 1/2 δ + k + 1)
≤ k(δ + 2).

Now assume that (4.13) holds for 1 ≤ n ≤ m. Observe that for m ≥ 2

d(pm+1 , pm+2 ) ≤ kd(pm , pm+1 ) + k m+1

≤ k[k m/2 (δ + m)] + k m+1
≤ k (m+1)/2 (δ + m) + k (m+1)/2 k (m+1)/2
≤ k (m+1)/2 [δ + (m + 1)]

or

d(pm+1 , pm+2 ) ≤ kd(pm−1 , pm ) + k m+1 + k m

≤ k[k (m−1)/2 (δ + m − 1)] + k m+1 + k m
≤ k (m+1)/2 (δ + m − 1) + k (m+1)/2 k (m+1)/2 + k (m+1)/2 k (m−1)/2
≤ k (m+1)/2 (δ + m − 1) + k (m+1)/2 + k (m+1)/2
= k (m+1)/2 (δ + m + 1),

and it follows that (4.13) is true for all n ∈ N. Using (4.13) we obtain
∞
 ∞

d(pn , pm ) ≤ δ (k 1/2 )i + i(k 1/2 )i , n > m ≥ 1.
i=m i=m

This means that {pn } is a Cauchy sequence. Because C is closed, {pn } converges
to a point z ∈ C. By our choice of {pn }, there exists a subsequence {pni } of
{pn } such that pni ∈ P , i.e., pni = pni , i = 1, 2, · · · . Note pni ∈ T pni −1 for
i ∈ N by (i) and pni −1 → z imply that T pni −1 → T z as i → ∞ in the Hausdorff
metric H. Because

d(pni , T z) ≤ H(T pni −1 , T z) → 0 as i → ∞,

it follows that d(z, T z) = 0. As T z is closed, z ∈ T z.

4.3. Convexity structure and fixed points 197

4.3 Convexity structure and fixed points

Let C be a nonempty subset of a metric space X and T : C → C a mapping.
Then a sequence {xn } in C is said to be an approximating fixed point sequence
(in short AFPS) of T if lim d(xn , T xn ) = 0.
n→∞

We have seen in the Banach contraction principle that every contraction

mapping has an approximating fixed point sequence in a metric space. In fact,
the Picard iterative sequence (xn+1 = T xn , n ∈ N) is an approximating fixed
point sequence of the contraction mapping T .

The following example shows that the Picard iterative sequence is not
necessarily an approximating fixed point sequence of nonexpansive mappings.

Example 4.3.1 Let X = R and T : R → R a mapping defined by

T x = −x for all x ∈ R.

Note that T is nonexpansive with F (T ) = {0}. However for x0 > 0, the iterative
sequence of the Picard iteration process is

xn+1 = T xn = (−1)n x0 , n ∈ N0 .

Hence d(xn , T xn ) = |(−1)n−1 − (−1)n |x0 = 2x0  0 as n → ∞.

The following Proposition 4.3.9 shows that the convexity structure has an
important role in the existence of AFPS for nonexpansive mappings. We define
convexity structure in a metric space.

Definition 4.3.2 Let (X, d) be a metric space. A continuous mapping W :

X × X × [0, 1] → X is said to be a convex structure2 on X, if for all x, y ∈ X
and λ ∈ [0, 1] the following condition is satisfied:

d(u, W (x, y; λ)) ≤ λd(u, x) + (1 − λ)d(u, y) for all u ∈ X. (4.14)

A metric space X with convex structure is called a convex metric space.

A subset C of a convex metric space X is said to be convex if W (x, y; λ) ∈ C

for all x, y ∈ C and λ ∈ [0, 1]. A convex metric space X is said to have property
(B) if

d(W (u, x; λ), W (u, y; λ)) = (1 − λ)d(x, y) for all u, x, y ∈ X and λ ∈ (0, 1).

Example 4.3.3 A normed space and each of its convex subsets are convex
metric spaces with convexity structure W (x, y; λ) = λx + (1 − λ)y.
2 The convexity structure in a metric space was introduced by W. Takahashi in 1970.
198 4. Existence Theorems in Metric Spaces

Example 4.3.4 Let X be a linear space that is also a metric space with the
following properties:
(i) d(x, y) = d(x − y, 0) for all x, y ∈ X;
(ii) d(λx + (1 − λ)y, 0) ≤ λd(x, 0) + (1 − λ)d(y, 0) for all x, y ∈ X and
λ ∈ [0, 1].
Then X is a convex metric space.
Example 4.3.5 A Fréchet space is not necessarily a convex metric space.
The following propositions are very useful in various applications.
Proposition 4.3.6 Let {Cα : α ∈ Λ} be a family of convex subsets of a convex
metric space X. Then ∩α∈Λ Cα is also a convex subset of X.
Proposition 4.3.7 The open balls Br (x) and the closed balls Br [x] in a convex
metric space X are convex subsets of X.
Proof. For y, z ∈ Br (x) and λ ∈ [0, 1], there exists W (y, z; λ) ∈ X. Because X
is a convex metric space,
d(x, W (y, z; λ)) ≤ λd(x, y) + (1 − λ)d(x, z)
< λr + (1 − λ)r = r.

Therefore, W (y, z; λ) ∈ Br (x). Similarly, Br [x] is a convex subset of X.

Proposition 4.3.8 Let X be a convex metric space. Then
d(x, y) = d(x, W (x, y; λ)) + d(W (x, y; λ), y) for all x, y ∈ X and λ ∈ [0, 1].
Proof. Because X is a convex metric space, we obtain
d(x, y) ≤ d(x, W (x, y; λ)) + d(W (x, y; λ), y)
≤ λd(x, x) + (1 − λ)d(x, y) + λd(x, y) + (1 − λ)d(y, y)
= λd(x, y) + (1 − λ)d(x, y)
= d(x, y)
for all x, y ∈ X and λ ∈ [0, 1]. Therefore,

d(x, y) = d(x, W (x, y; λ) + d(W (x, y; λ), y) for all x, y ∈ X and λ ∈ [0, 1].
We now apply the convexity structure defined in Definition 4.3.2 to obtain
AFPS for nonexpansive mappings in a metric space. Note, similar results are
also discussed in Chapter 5.
Proposition 4.3.9 Let X be a complete convex metric space with property (B),
C a nonempty closed convex subset of X, and T : C → C a nonexpansive
mapping. Then we have the following:
(a) For u ∈ C and t ∈ (0, 1), there exists exactly one point xt ∈ C such that
xt = W (u, T xt ; 1 − t)

(b) If C is bounded, then d(xt , T xt ) → 0 as t → 1, i.e., T has an AFPS.

4.3. Convexity structure and fixed points 199

Proof. (a) For t ∈ (0, 1), consider the mapping Tt : C → C defined by

Tt x = W (u, T x; 1 − t).

By property (B), we have

d(Tt x, Tt y) = td(T x, T y) ≤ td(x, y) for all x, y ∈ C.

By the Banach contraction principle, Tt has exactly one fixed point xt in C.

Therefore,
xt = W (u, T xt ; 1 − t).
(b) By boundedness of C, we get

d(xt , T xt ) = d(T xt , W (u, T xt ; 1 − t))

≤ (1 − t)d(T xt , u) ≤ (1 − t) diam(C) → 0 as t → 1.

Applying Proposition 4.3.9, we have

Theorem 4.3.10 Let X be a complete convex metric space X with property

(B), C a nonempty compact convex subset of X, and T : C → C a nonexpansive
mapping. Then T has a fixed point in C.

Proof. By Proposition 4.3.9, there exists a sequence {xn } in C such that

lim d(xn , T xn ) = 0. (4.15)

n→∞

Because C is compact, there exists a subsequence {xnk } of {xn } such that

xnk → v ∈ C. Hence from (4.15), we have v = T v.

In Theorem 4.3.14, we will see that compactness can be dropped if C has

normal structure. To see this, we extend the notion of normal structure in
metric space X.

For C ⊂ X, we denote the following, which will be used throughout the

remainder of this chapter:

rx (C) = sup{d(x, y) : y ∈ C}, x ∈ C,

r(C) = inf{rx (C) : x ∈ C},
ZC = {x ∈ C : rx (C) = r(C)}.

A point x0 ∈ C is said to be a diametral point of C if

sup{d(x0 , y) : y ∈ C} = diam(C).

A convex metric space X is said to have normal structure if for each closed
convex bounded subset C of X that contains at least two points, there exists
x0 ∈ C that is not a diametral point of C.
200 4. Existence Theorems in Metric Spaces

Example 4.3.11 Every compact convex metric space has normal structure.
A convex metric space X is said to have property (C) if every bounded
decreasing net of nonempty closed convex subsets of X has a nonempty inter-
section. By Smulian’s theorem, every weakly compact convex subset of a Banach
space has property (C).

Using property (C), we have

Proposition 4.3.12 If a convex metric space X has property (C), then ZC is
nonempty, closed, and convex.
Proof. Let Cn (x) = {y ∈ C : d(x, y) ≤ r(C) + 1/n} for n ∈ N and x ∈ X.
It is easily seen that the sets Cn = ∩x∈X Cn (x) form a decreasing sequence of
nonempty closed convex sets, and hence ∩∞ n=1 Cn is nonempty closed convex by
property (C). Because ZC = ∩∞
n=1 Cn , the proof is complete.

Proposition 4.3.13 Let C be a nonempty compact subset of a convex metric

space X and let D be the least closed convex set containing C. If diam(C) > 0,
then there exists an element x0 ∈ D such that sup{d(x, x0 ) : x ∈ C} < diam(C).
Proof. Because C is compact, we may find x1 , x2 ∈ C such that d(x1 , x2 ) =
diam(C). Let C0 ⊂ C be maximal so that C0 ⊃ {x1 , x2 } and d(x, y) = 0
or diam(C) for all x, y ∈ C0 . It is easy to see that C0 is finite. Let C0 =
{x1 , x2 , · · · , xn }. Because X is a convex metric space, we can define
1
y1 = W (x1 , x2 ; ),
2
1
y2 = W (x3 , y1 ; ),
3
···
1
yn−2 = W (xn−1 , yn−3 ; ),
n−1
1
yn−1 = ) = u.
W (xn , yn−2 ;
n
Because C is compact, we can find y0 ∈ C such that
d(y0 , u) = sup{d(x, u) : x ∈ C}.
From (4.14), we obtain
1 n−1
d(y0 , u) ≤ d(y0 , xn ) + d(y0 , yn−2 )
n n  
1 n−1 1 n−2
≤ d(y0 , xn ) + d(y0 , xn−1 ) + d(y0 , yn−3 )
n n n−1 n−1
1 1 n−2
= d(y0 , xn ) + d(y0 , xn−1 ) + d(y0 , yn−3 )
n n n
···
1
n
≤ d(y0 , xk ) ≤ diam(C).
n
k=1
4.4. Normal structure coefficient and fixed points 201

Therefore, if d(y0 , u) = diam(C), then we must have d(y0 , xk ) = diam(C) > 0

for all k = 1, 2, · · · , n. Hence y0 ∈ C0 by definition of C0 . But, then we must
have y0 = xk for some k = 1, 2, · · · , n. This is a contradiction. Therefore,

sup{d(x, u) : x ∈ C} = d(y0 , u) < diam(C).

A closed convex subset C of a convex metric space X is said to have the fixed
point property for nonexpansive mappings if every nonexpansive T : C → C has
a fixed point.
We now prove that every closed convex subset of a convex metric space has
fixed point property for nonexpansive mappings under normal structure.

Theorem 4.3.14 Let X be a convex metric space with property (C). Let C be
a nonempty closed convex bounded subset of X with normal structure and T a
nonexpansive mapping from C into itself. Then T has a fixed point in C.

Proof. Let F be the family of all nonempty closed convex subsets of C, each of
which is mapped into itself by T . By property (C) and Zorn’s lemma, F has a
minimal element C0 . We show that C0 consists of a single point. Let x ∈ ZC0 .
Then
d(T x, T y) ≤ d(x, y) ≤ rx (C0 ) for all y ∈ C0 .
Hence T (C0 ) is contained in the ball B = Br(C0 ) [T x]. Because T (C0 ∩ B) ⊂
C0 ∩ B, the minimality of C0 implies that C0 ⊂ B. Hence rT x (C0 ) ≤ r(C0 ).
Because r(C0 ) ≤ rx (C0 ) for all x ∈ C0 , we have rT x (C0 ) = r(C0 ). Hence
T x ∈ ZC0 and T (ZC0 ) ⊂ ZC0 . By Proposition 4.3.12, ZC0 ∈ F. If z, w ∈ ZC0 ,
then d(z, w) ≤ rz (C0 ) = r(C0 ). Hence, by normal structure,

δ(ZC0 ) ≤ r(C0 ) < δ(C0 ).

Because this contradicts the minimality of C0 , diam(C0 ) = 0 and C0 consists

of a single point.

4.4 Normal structure coefficient and fixed points

In this section, we discuss another convexity structure on metric space and the
existence of fixed points of uniformly L-Lipschitzian mappings in a metric space
with uniformly normal structure.

Let F(X) denote a nonempty family of subsets of a metric space (X, d).
We say that F(X) defines a convexity structure on X if F(X) is stable by
intersection and that F(X) has property (R) if any decreasing sequence {Cn }
of nonempty closed bounded subsets of X with Cn ∈ F(X) has nonvoid inter-
section.

A subset of X is said to be admissible if it is an intersection of closed balls.

We denote by A(X) the family of all admissible subsets of X. It is obvious that
202 4. Existence Theorems in Metric Spaces

A(X) defines a convexity structure on X. In this section, any other convexity

structure F(X) on X is always assumed to contain A(X).

For a bounded subset C of X, we define the admissible hull of C, denoted

by ad(C), as the intersection of all those admissible subsets of X that contain
C, i.e.,
ad(C) = ∩{B : C ⊆ B ⊆ X with B admissible}.
A basic property of admissible hull is given in the following proposition.
Proposition 4.4.1 Let C be a bounded subset of a metric space X and x ∈ X.
Then
rx (ad(C)) = rx (C).
Proof. Suppose r = rx (ad(C)) > rx (C). Then C ⊆ Br [x] for any r with
rx (C) < r < r. It follows that ad(C) ⊆ Br [x]. Hence

rx (ad(C)) = sup{d(x, y) : y ∈ ad(C)} ≤ r < r,

a contradiction.

We introduce normal structure and uniformly normal structure with respect

to convexity structure F(X) in a metric space X, respectively.
Definition 4.4.2 A metric space (X, d) is said to have normal structure if there
exists a convexity structure F(X) such that r(C) < diam(C) for all C ∈ F(X)
that is bounded and consists of more than one point. We say that F(X) is
normal.

Definition 4.4.3 A metric space (X, d) is said to have uniformly normal struc-
ture if there exists a convexity structure F(X) such that r(C) ≤ α · diam(C) for
some constant α ∈ (0, 1) and for all C ∈ F(X) that is bounded and consists of
more than one point. We also say that F(X) is uniformly normal.

We now define the normal structure coefficient of X (with respect to a given

convexity structure F(X)):
The number N (X) is said to be the normal structure coefficient if
 
diam(C)
N (X) = inf ,
R(C)

where the infimum is taken over all bounded C ∈ F(X) with diam(C) > 0.
It is easy to see that X has uniformly normal structure if and only if N (X) > 1.

The following theorem shows that every convexity structure with uniformly
normal structure has property (R).
Theorem 4.4.4 Let X be a complete metric space with a convexity structure
F(X) that is uniformly normal. Let {Cn } be a decreasing sequence of nonempty
bounded subsets of X with Cn ∈ F(X). Then ∩∞ n=1 C n = ∅.
4.4. Normal structure coefficient and fixed points 203

Proof. Without loss of generality, we may assume that diam(Cn ) > 0 for all
n ∈ N. Let η be a real number with Ñ (X) < η < 1, where Ñ (X) = N (X)−1 .
Define a sequence {xn,k } in X as follows:
For arbitrary xn,1 ∈ Cn , n ∈ N, select xn,k ∈ ad({xm,k−1 }m≥n ) such that
sup{d(xn,k , x) : x ∈ ad({xm,k−1 }m≥n ) ≤ η diam(ad{xm,k−1 }m≥n ).
Set An,k := ad({xm,k }m≥n ). Observe that
An,k ⊆ Cn for all n, k ∈ N
and for m ≥ n,
d(xn,k , xm,k ) ≤ sup{d(xn,k , x) : x ∈ An,k−1 }
≤ η diam(An,k−1 )
≤ η diam({xi,k−1 }i≥1 ).
For k ≥ 2, we have
diam({xn,k }) ≤ η diam({xn,k−1 })
≤ η 2 diam({xn,k−2 })
···
≤ η k−1 diam({xn,1 })
≤ η k−1 diam(C1 ).
Now we consider a subsequence {xn,n } of {xn,k }. Then {xn,n } is Cauchy,
because
d(xn,n , xm,m ) ≤ η n−1 diam(C1 ) for m ≥ n.
Therefore, there exists an x ∈ ∩∞
n=1 C n such that {xn,n } converges to x, i.e.,
∩∞
n=1 C n = ∅.

Corollary 4.4.5 Let X be a complete bounded metric space and F(X) a con-
vexity structure of X with uniformly normal structure. Then F(X) has property
(R).
We now introduce the property (P) for metric spaces.
Definition 4.4.6 A metric space (X, d) is said to have property (P) if given any
two bounded sequences {xn } and {zn } in X, one can find some z ∈ ∩∞n=1 ad({zj :
j ≥ n}) such that
lim sup d(z, xn ) ≤ lim sup lim sup d(zj , xn ).
n→∞ j→∞ n→∞

Remark 4.4.7 If X has property (R), then ∩∞ n=1 ad({zj : j ≥ n}) = ∅. Also, if
X is a weakly compact convex subset of a normed space, then admissible hulls
are closed convex and ∩∞
n=1 ad{zj : j ≥ n} = ∅ by weak compactness of X, and
that X possesses property (P) follows directly from the w-lsc of the function
lim sup xn − x.
n→∞
204 4. Existence Theorems in Metric Spaces

We establish the following key result that can be viewed as the metric space
formulation of Theorem 3.4.20.
Theorem 4.4.8 Let (X, d) be a complete bounded metric space with both pro-
perty (P) and uniformly normal structure. Let N (X) be the normal structure
coefficient with respect to the given convexity structure F(X). Then for any
sequence {xn } in X and any constant α > Ñ (X), there exists a point z ∈ X
satisfying the properties:
(a) d(z, y) ≤ lim sup d(xn , y) for all y ∈ X,
n→∞
(b) lim sup d(z, xn ) ≤ α diam({xn }).
n→∞

Proof. (a) For each n ∈ N, let An = ad({xj : j ≥ n}). Then {An } is a

decreasing sequence of admissible subsets of X and hence A := ∩An = ∅ by
Corollary 4.4.5. We observe by Proposition 4.4.1 that

diam(An ) = sup{rx (An ) : x ∈ An }

= sup sup d(x, xj )
x∈An j≥n
= sup sup d(x, xj ) = sup rxj (An )
j≥n x∈An j≥n
= sup sup d(xj , xi )
j≥n i≥n
≤ sup{d(xi , xj ) : i, j ∈ N} = diam({xn }).

For any z ∈ A and any y ∈ X, we have

sup d(y, xj ) = ry (An ) ≥ ry (A) ≥ d(y, z).

j≥n

It follows that
d(y, z) ≤ lim sup d(y, xn ).
n→∞

(b) Without loss of generality, we may assume that diam({xn }) > 0. Then
for α > Ñ (X), we choose ε > 0 so small that

Ñ (X)diam({xn }) + ε ≤ α diam({xn }).

By definition, one can find a zn ∈ An such that

rzn (An ) < r(An ) + ε

≤ Ñ (X) diam(An ) + ε
≤ Ñ (X)diam({xn }) + ε
≤ α diam({xn }).

Hence for each i ≥ 1,

lim sup d(zi , xm ) ≤ α diam({xi }).

m→∞
4.4. Normal structure coefficient and fixed points 205

Now property (P) yields a point z ∈ ∩∞

i=1 ad({zn : n ≥ i}) such that

lim sup d(z, xm ) ≤ lim sup lim sup d(zi , xm ).

m→∞ i→∞ m→∞

Thus, z ∈ A and satisfies

lim sup d(z, xm ) ≤ α diam({xi }).

m→∞

We now present the existence theorem for uniformly L-Lipschitzian map-

pings in a metric space.
Theorem 4.4.9 Let (X, d) be a complete bounded metric space with both pro-
perty (P) and uniformly normal structure and T : X → X a uniformly
L-Lipschitzian mapping with L < Ñ (X)−1/2 . Then T has a fixed point in X.
Proof. Choose a constant α, 1 > α > Ñ (X), such that L < α−1/2 . Let x0 ∈ X.
By Theorem 4.4.8, we can inductively construct a sequence {xm }∞m=1 in X:
for each integer m ≥ 0,
(a) lim sup d(xm+1 , T i xm ) ≤ α diam({T i xm });
i→∞

(b) d(xm+1 , y) ≤ lim sup d(T i xm , y) for all y ∈ X.

i→∞
Set rm := lim sup d(xm+1 , T i xm ) and h := αL2 < 1. Note for each i > j ≥ 0,
i→∞

d(T j xm , T i xm ) ≤ Ld(xm , T i−j xm )

≤ L lim sup d(T n xm−1 , T i−j xm ) (by (b))
n→∞
≤ L2 lim sup d(T n xm−1 , xm )
n→∞
≤ L rm−1 .
2

Observe that
rm = lim sup d(xm+1 , T i xm )
i→∞
≤ α diam({T i xm })
≤ αL2 rm−1 = hrm−1
···
≤ hm r0 .
Hence for each integer i ≥ 0,
d(xm+1 , xm ) ≤ d(xm+1 , T i xm ) + d(xm , T i xm )
≤ d(xm+1 , T i xm ) + lim sup d(T j xm−1 , T i xm )
j→∞

≤ d(xm+1 , T i xm ) + L lim sup d(T j xm−1 , xm )

j→∞

≤ d(xm+1 , T xm ) + Lrm−1 ,
i
206 4. Existence Theorems in Metric Spaces

which implies that

d(xm+1 , xm ) ≤ rm + Lrm−1 ≤ (hm + Lhm−1 )r0 .

This implies that {xm } is Cauchy. Let lim xm = v ∈ X. Observe that

m→∞

d(v, T v) ≤ d(v, xm+1 ) + d(xm+1 , T i xm ) + d(T i xm , T v)

≤ d(z, xm+1 ) + d(xm+1 , T i xm ) + Ld(T i−1 xm , v)
≤ d(v, xm+1 ) + d(xm+1 , T i xm ) + Ld(T i−1 xm , xm+1 ) + Ld(xm+1 , v),

which implies that

d(v, T v) ≤ (1 + L)d(v, xm+1 ) + (1 + L)rm → 0 as m → ∞.

Therefore, v is a fixed point of T .

4.5 Lifschitz’s coefficient and fixed points

In this section, we give an existence theorem concerning uniformly
L-Lipschitzian mappings in a metric space.

First, we define the Lifschitz’s coefficient of a metric space:

Let (X, d) be a metric space. Then the Lifschitz’s coefficient κ(X) is a
number defined by

κ(X) = sup{β > 0 : ∃ α > 1 such that for all x, y ∈ X, for all r > 0,
[d(x, y) > r ⇒ ∃ z ∈ X such thatBαr [x] ∩ Bβr [y] ⊆ Br [z]]}.

It is clear that κ(X) ≥ 1 for any metric space X. For a √strictly convex
Banach space X, κ(X) > 1 and for a Hilbert space H, κ(H) = 2.

We are now in a position to prove a fundamental existence theorem for

uniformly L-Lipschitzian mappings in a metric space with Lifschitz’s coefficient
κ(X).

Theorem 4.5.1 Let (X, d) be a bounded complete metric space and T : X → X

a uniformly L-Lipschitzian mapping with L < κ(X). Then T has a fixed point
in X.

Proof. If κ(X) = 1, then T is contraction and hence T has a unique fixed

point. So, suppose κ(X) > 1. For b ∈ (L, κ(X)), there exists a > 1 such that

∀ u, v ∈ X, r > 0 with d(x, y) > r ⇒ ∃z ∈ X : Bbr [u] ∩ Bar [v] ⊂ Br [z]. (4.16)

For any x ∈ X, let

r(x) = inf{R > 0 : there exists y ∈ X such that lim sup d(T n x, y) ≤ R}.
n→∞
4.5. Lifschitz’s coefficient and fixed points 207

Observe that r is a lower semicontinuous and r(x) = 0 implies x = T x.

Take λ ∈ (0, 1) such that γ = min{λa, λb/L} > 1. We now show that there
exists a sequence {ym } in X that satisfies the following:

r(ym+1 ) ≤ λr(ym ) and d(ym , ym+1 ) ≤ (λ + γ)r(ym ) for all m ∈ N0 . (4.17)

Indeed, consider an arbitrary point y0 ∈ X and assume that y0 , y1 , · · · , ym

are given. We now construct ym+1 . If r(ym ) = 0, then ym+1 = ym . If r(ym ) > 0,
then for a number λr(ym ), there exists n ∈ N such that

d(ym , T n ym ) > λr(ym ).

From the definition of r(ym ), there exists x ∈ X such that

lim sup d(ym , T n x) ≤ r(ym ) < γr(ym ).

n→∞

Hence for i > j

d(T i x, T j ym ) ≤ L d(T i−j x, ym ),
which implies that

lim sup d(T i x, T j ym ) ≤ L lim sup d(T i−j x, ym ) ≤ Lγ r(m).

i→∞ i→∞

Because

Bγr(ym ) [ym ] ∩ BLγr(ym ) [T n ym ] ⊂ Baλr(ym ) [ym ] ∩ Bbλr(ym ) [T n ym ] = C,

the set C is contained in a closed ball centered at w with radius λr(ym ) (Condi-
tion (4.16)). Thus, lim sup d(T n x, w) ≤ λr(ym ). Take w = ym+1 , and it follows
n→∞
from above that {ym } satisfies (4.17).
Note

r(ym+1 ) ≤ λr(ym ) ≤ · · · ≤ λm+1 r(y0 ) → 0 as m → ∞

and
d(ym , ym+1 ) ≤ (λ + γ)r(ym ) → 0 as m → ∞.
Hence {ym } converges to v ∈ M . But because r(v) = 0, v is a fixed point
of T .

Bibliographic Notes and Remarks

Most of the results presented in Section 4.1 may be found in Goebel and Kirk
[59], Khamsi and Kirk [85], Kirk and Sims [91], and Martin [106]. Theorem 4.1.16
first appeared in Matkowski [108] as a generalization of the result of Boyd and
Wong [22], and it was recently extended for non-self mappings in Reich and
208 4. Existence Theorems in Metric Spaces

Zaslavski [133](see also recent results of Agarwal, O’Regan and their coworkers
[1, 117]).
The notion of nearly non-Lipschitzian mappings was introduced by Sahu
[137]. Some existence theorems for demicontinuous nearly contraction and
nearly asymptotically nonexpansive mappings were also established in [137].
The fixed point theory of multivalued contraction self-mappings was first
proved in Nadler [115]. It was extended for multivalued non-self contraction
mappings by Assad and Kirk [4].
The results describes in Sections 4.3∼4.5 can be found in Lifschitz [97], Lim
and Xu [99], and Takahashi [154].

Exercises

4.1 Let X be a complete metric space and T : X → X a mapping such that

T m is contraction for some m ∈ N. Show that T has a unique fixed point.

4.2 Let (X, d) be a metric space and T : X → X a mapping. T is said to be

a Zamfirescu mapping if there exist the real number a, b, and c satisfying
0 < a < 1, 0 < b, c < 1/2 such that for each pair x, y in X, at least one of
the following is true:
(Z1 ) d(T x, T y) ≤ ad(x, y),
(Z2 ) d(T x, T y) ≤ b[d(x, T x) + d(y, T y)],
(Z3 ) d(T x, T y) ≤ c[d(x, T y) + d(y, T x)].

If (X, d) is a complete metric space and T : X → X a Zamfirescu mapping,

show that T has a unique fixed point z ∈ X and for each x ∈ X, {T n x}
converges to z.

4.3 Let T be a mapping from a complete metric space X into itself satisfying
the condition:

d(T x, T y) ≤ ad(x, y) + b[d(x, T x) + d(y, T y)] + c[d(y, T x) + d(x, T y)]

for all x, y ∈ X, where a, b, c are nonnegative real numbers such that

a + 2b + 2c < 1. Show that T has a unique fixed point z ∈ X and for each
x ∈ X, {T n x} converges to z.

4.4 Let T be a mapping from a complete metric space into itself. Assume that
for each ε > 0, there exists δ > 0 such that

d(x, T x) < δ ⇒ T (Bε [x]) ⊂ Bε [x].

If d(T n x, T n+1 x) → 0 for some x ∈ X, show that the sequence {T n x}

converges to z, which is a fixed point of T .
4.5. Lifschitz’s coefficient and fixed points 209

4.5 Let X be a complete metric space and T : X → X an expansion mapping,

i.e., there exists constant k > 1 such that

d(T x, T y) ≥ kd(x, y) for all x, y ∈ X.

Assume that T (X) = X. Show that

(a) T is one to one,
(b) T has a unique fixed point z ∈ X with T n x → z as n → ∞ for some
x ∈ X.

4.6 Let (X, d) be a complete metric space and T : X → CB(X) a mapping.

If α is a function from (0, ∞) to [0, 1) such that lim sup α(r) < t for every
r→t+
t ∈ [0, ∞) and if
H(T x, T y) ≤ α(d(x, y))d(x, y)
for each x, y ∈ X, show that T has a fixed point in X.
Chapter 5

Existence Theorems in
Banach Spaces

This chapter is devoted to a demiclosed principle and existence of fixed points

of Lipschitzian and non-Lipschitzian mappings in Banach spaces.

5.1 Non-self contraction mappings

In Chapter 4, we studied fixed point theorems for single-valued and multivalued
contraction mappings in metric spaces. In this section, we discuss fixed point
theorems concerning non-self contraction mappings in Banach spaces.
Let C be a nonempty subset of a Banach space X. For x ∈ C, the inward
set of x relative to C is the set

IC (x) = {x + t(y − x) : y ∈ C and t ≥ 0}

and the outward set of x relative to C is the set

OC (x) = {x − t(y − x) : y ∈ C and t ≥ 0}.

Let IC (x) and OC (x) denote closures of IC (x) and OC (x), respectively.

Set
d(x + hy, C)
I C (x) := x + {y ∈ X : lim inf = 0}, x ∈ C.
+
h→0 h

Note that for a convex set C, we have

(i) I C (x) = IC (x),
(ii) C ⊆ IC (x).

We now define weakly inward and weakly outward mappings:

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 5,

c Springer Science+Business Media, LLC, 2009
212 5. Existence Theorems in Banach Spaces

Let C be a nonempty subset of a Banach space X and T : C → X a mapping.

Then T is said to be a
(i) inward mapping if T x ∈ IC (x) for all x ∈ C,
(ii) weakly inward mapping if T x ∈ IC (x) for all x ∈ C,
(iii) weakly outward mapping if T x ∈ OC (x) for all x ∈ C.
Let us compare inwardness and weak inwardness conditions with other condi-
tions. Set
(C1) Rothe’s condition: T (∂C) ⊆ C;
(C2) inwardness condition: T x ∈ IC (x) for all x ∈ C;
(C3) weak inwardness condition: T x ∈ IC (x) for all x ∈ C;
(C4) the Leray-Schauder condition (if the interior int(C) of C is nonempty):
there exists a z ∈ int(C) such that

T x − z = μ(x − z) for all x ∈ ∂C and μ > 1.

These boundary conditions hold the implications:

(C1) ⇒ (C2) ⇒ (C3) ⇒ (C4).

The following proposition gives an equivalent formulation of the weakly

inwardness condition.
Proposition 5.1.1 Let C be a nonempty closed convex subset of a Banach
space X. Then T : C → X is weakly inward if and only if
d((1 − h)x + hT x, C)
lim =0 for all x ∈ C. (5.1)
h→0+ h
Proof. Suppose that condition (5.1) holds. Fix x ∈ C. For ε > 0, we may
assume t ∈ (0, 1) and y ∈ C such that

(1 − t)x + tT x − y ≤ d((1 − t)x + tT x, C) + tε.

It follows that

T x − [(1 − t−1 )x + t−1 y] ≤ t−1 d((1 − t)x + tT x, C) + ε.

It is easy to see that T x ∈ IC (x). Hence T is weakly inward.

Conversely, suppose that T is weakly inward, i.e., T x ∈ IC (x) for all x ∈ C.

Hence for ε > 0, there exists y ∈ IC (x) such that

y − T x ≤ ε.

Because C is convex, there exists h0 > 0 such that

(1 − h)x + hy ∈ C for 0 < h ≤ h0 .

5.1. Non-self contraction mappings 213

Hence for these h, we have

d((1 − h)x + hT x, C) (1 − h)x + hT x − {(1 − h)x + hy}]
≤
h h
≤ ε.

Therefore, the condition (5.1) holds.

The following result is an extension of the Banach contraction principle for

non-self contraction mappings.
Theorem 5.1.2 Let C be a nonempty closed convex subset of a Banach space
X and T : C → X a weakly inward contraction mapping. Then T has a unique
fixed point in C.
Prof. Let k, 0 < k < 1 denote Lipschitz constant of T . Choose ε > 0 so small
that k < (1 − ε)/(1 + ε). By Proposition 5.1.1, T satisfies the condition (5.1).
Hence for x ∈ C with x = T x, there exists h ∈ (0, 1) such that
d((1 − h)x + hT x, C) < hεx − T x.
By the definition of distance, there exists y ∈ C such that
(1 − h)x + hT x − y < hεx − T x. (5.2)
By (5.2), we have
hεx − T x > x − y − h(x − T x)
≥ x − y − hx − T x,
which implies that
x − y < (1 + ε)hx − T x. (5.3)
Using (5.2), we have
y − T y ≤ y − [(1 − h)x + hT x] + (1 − h)x + hT x − T x + T x − T y
≤ hεx − T x + (1 − h)x − T x + kx − y
= x − T x + (ε − 1)hx − T x + kx − y
1−ε
= x − T x + (ε − 1)hx − T x + x − y
1+ε
 
1−ε
− − k x − y
1+ε
 
1−ε
< x − T x − − k x − y. (by (5.3))
1+ε
If x = T x, x ∈ C, denote y ∈ C as above by f (x), where f is a self-mapping on
C. Then by putting
 −1
1−ε
ϕ(x) = −k x − T x,
1+ε
214 5. Existence Theorems in Banach Spaces

ϕ : C → R+ is a continuous function and

x − f x < ϕ(x) − ϕ(f x). (5.4)

By Caristi’s theorem, f has a fixed point, which contradicts the strict inequality
(5.4).

We now turn our attention to study fixed points of multivalued mappings in

Banach spaces.
Let C be a nonempty subset of a Banach space X. We say that a mapping
T of C into the family of nonempty subsets of X is weakly inward if T x ⊂ IC (x)
for each x ∈ C. Let F (X) denote the family of nonempty closed subsets of X
and T : C → F (X) a multivalued mapping. Given x ∈ C and α ≥ 1, let Tα (x)
denote the set {z ∈ T x : x − z ≤ α d(x, T x)}.

Theorem 5.1.3 Let C be a nonempty closed convex subset of a Banach space X

and T : C → F (X) an upper semicontinuous mapping satisfying the conditions:
(a) For each x ∈ C, there exists δ = δ(x) > 0 such that

y ∈ Bδ [x] ∩ C ⇒ d(y, T y) ≤ d(y, T x) + kx − y,

where k ∈ (0, 1).

(b) T1 (x) ∩ IC (x) = ∅ for each x ∈ C.
Then T has a fixed point in C.

Proof. Suppose, for contradiction, that T has no fixed point. We may assume
that d(x, T x) > 0 for each x ∈ C. Select ε > 0 such that k < (1 − ε)/(1 + ε).
Given x ∈ C, condition (b) implies the existence of an element z ∈ T1 (x)∩IC (x),
and by Proposition 3.1.11, there exists h ∈ (0, 1) such that

h−1 d((1 − h)x + hz, C) < ε d(x, T x). (5.5)

Set z := (1 − h)x + hz. Observe that z − x = hz − x, and moreover we may
suppose h has been chosen so small that z ∈ Bδ/2 [x] (where δ = δ(x) is from
condition (a)). By (5.5), there exists y ∈ C, y = x such that

z − y < hε d(x, T x). (5.6)

Hence

x − y/z − x ≤ [x − z + z − y]/z − x

= 1 + z − y/z − x
< 1 + z − y/(hd(x, T x))
< 1 + ε,

which implies that

(1 + ε)−1 x − y < z − x. (5.7)

5.1. Non-self contraction mappings 215

Because z ∈ Bδ/2 [x], it follows from (5.7) that y ∈ Bδ [x] and thus
d(y, T y) ≤ d(y, T x) + kx − y
≤ y − z + d(z, T x) + kx − y.
Combining this with (5.6), (5.7), and using the definition of z along with the
fact that z ∈ T1 (x), we obtain
d(y, T y) ≤ y − z + d(x, T x) − x − z + kx − y
< εx − z + d(x, T x) − x − z + kx − y.
= d(x, T x) + kx − y − (1 − ε)x − z
< d(x, T x) + kx − y − (1 − ε)(1 + ε)−1 x − y
= d(x, T x) + [k − (1 − ε)(1 + ε)−1 ]x − y.
Set η = −[k − (1 − ε)(1 + ε)−1 ]. Then η > 0 and
ηx − y ≤ d(x, T x) − d(y, T y).
We now define g : C → C by g(x) = y with y determined as above, and let
ϕ(x) = η −1 d(x, T x). Proposition 4.2.6 implies that ϕ is lower semicontinuous,
so Caristi’s theorem implies the existence of an x0 ∈ C such that x0 = g(x0 ).
But g(x) = y = x for all x ∈ C by definition, and our assumption that T has
no fixed points is contradicted.

The following example shows that condition (b) in Theorem 5.1.3 cannot be
altered to T x ∩ IC (x) = ∅.
Example 5.1.4 Let X = R and C = [0, 1]. Define T : C → K(X) by T x =
{−1, 2} for all x ∈ C. Then T is a constant and
⎧
⎨ [0, ∞) if x = 0,
IC (x) = R if x ∈ (0, 1),
⎩
(−∞, 1] if x = 1.
Therefore, T x ∩ IC (x) = ∅ for all x ∈ C, but T has no fixed point in C.
Theorem 5.1.5 Let C be a nonempty closed convex subset of a Banach space
X and T : C → F (X) an upper semicontinuous mapping that satisfies the
conditions:
(a) For each x ∈ C, there exists δ = δ(x) > 0 such that
y ∈ Bδ [x] ∩ C ⇒ d(y, T y) ≤ d(y, T x) + kx − y,
where k ∈ (0, 1).

(b ) Corresponding with each x ∈ C, there exist constants α = α(x) > 1, μ =
μ(x) ∈ (0, 1) such that
(1 − μ)x + μTα (x) ⊂ C.
Then T has a fixed point in C.
216 5. Existence Theorems in Banach Spaces

Proof. Suppose, for contradiction, that T has no fixed point. Select k  ∈ (k, 1).
Fix x ∈ C, let α = α(x), and choose μ so that μ ≤ min{μ(x), δ(x)} and
αμ d(x, T x) ≤ δ(x). Then

x − y ≥ μ d(x, T x) for all y ∈ (1 − μ)x + μTα (x).

Thus, if ξ ∈ (1, α) is chosen so that ξ − 1 < μ(k  − k), we obtain

(ξ − 1) d(x, T x) ≤ (k  − k)x − y for all y ∈ (1 − μ)x + μTα (x) (5.8)

Now fix z ∈ Tξ (x) and let y = (1 − μ)x + μz. By (5.8), we have

kx − y ≤ k  x − y − (ξ − 1) d(x, T x)

and therefore (using (a) because μx − z ≤ μξ d(x, T x) ≤ δ(x)):

d(y, T y) ≤ d(y, T x) + kx − y

≤ y − z + kx − y
= x − z − x − y + kx − y
≤ ξd(x, T x) − x − y + k  x − y − (ξ − 1) d(x, T x)
= d(x, T x) + (k  − 1)x − y.

Hence
x − y ≤ (1 − k  )−1 [d(x, T x) − d(y, T y)]
and the proof is completed as in Theorem 5.1.3 by taking g(x ) = y and ϕ(x) =
(1 − k  )−1 d(x, T x).
We now derive some existence theorems from Theorems 5.1.3 and 5.1.5.

Theorem 5.1.6 Let C be a nonempty closed convex subset of a Banach space

X and T : C → CB(X) a multivalued contraction mapping that satisfies either
condition (b) or condition (b ). Then T has a fixed point in C.

Proof. It is easy to see that T is automatically upper semicontinuous. Because

d(y, T y) ≤ d(y, T x) + H(T x, T y) for all x, y ∈ C,

it follows that T satisfies condition (a).

Theorem 5.1.7 Let C be a nonempty closed convex subset of a Banach space

X and T : C → K(X) a multivalued contraction mapping for which T x ⊂
IC (x), x ∈ C. Then T has a fixed point in C.

Proof. Under the stated assumptions, condition (b) is automatically satisfied.

Hence the result follows from Theorem 5.1.6.
5.1. Non-self contraction mappings 217

As a slightly different result, we have

Theorem 5.1.8 Let C be a nonempty closed convex subset of a Banach space

X and T : C → CB(X) a contraction mapping. If for each x ∈ C,

d((1 − h)x + hz, C)

lim = 0 uniformly for z ∈ T x,
h→0+ h

then T has a fixed point.

Proof. Let k be the Lipschitz constant of T . Choose real numbers k  and q such
that k < k  < 1, 0 < q < 1 and k  < (1 − q)/(1 + q). Suppose, for contradiction,
that T has no fixed point, i.e., d(x, T x) > 0 for all x ∈ C. For each x ∈ C, take
ε > 0 such that
qd(x, T x) − εd(x, T x) > 0.

By assumption, there exists h ∈ (0, 1) such that

d((1 − h)x + hz, C) < h(q − ε)d(x, T x) for all z ∈ T x.

Choose z ∈ T x such that

x − z < d(x, T x) + hεd(x, T x). (5.9)

For such a z, take y ∈ C such that

(1 − h)x + hz − y < h(q − ε)d(x, T x). (5.10)

Set w := (1 − h)x + hz. From (5.10), we have

w − y < h(q − ε)d(x, T x)

≤ hqx − z − hεd(x, T x)
= qw − x − hεd(x, T x) (since h(z − x) = w − x)
< qw − x

and

x − y ≤ x − w + w − y
< x − w + qw − x
< (1 + q)x − w.

Let ε = [k  x − y − H(T x, T y)]/2. Then, because T is contraction with

Lipschitz constant k and x = y by (5.10), we have ε > 0. Choose u ∈ T x and
v ∈ T y satisfying

w − u < d(w, T x) + ε and u − v ≤ H(T x, T y) + ε .

218 5. Existence Theorems in Banach Spaces

Then we have

d(y, T y) ≤ y − v
≤ y − w + w − u + u − v
< y − w + d(w, T x) + ε + H(T x, T y) + ε
≤ y − w + w − z + k  x − y
= y − w + x − z − hx − z + k  x − y
= y − w + x − z − w − x + k  x − y
< qw − x − hε d(x, T x) + x − z − w − x + k  x − y
= (q − 1)w − x + x − z − hε d(x, T x) + k  x − y
q−1
< x − y + d(x, T x) + k  x − y (from (5.9))
q+1
 
1−q
= d(x, T x) − − k  x − y
q+1
= d(x, T x) − rx − y,

where r = [(1 − q)/(1 + q)] − k  . Now for each x ∈ C, denote y ∈ C as above by

f (x). Then

x − f (x) < ϕ(x) − ϕ(f (x)) for all x ∈ C, (5.11)

where ϕ : C → [0, ∞) is the continuous function defined by ϕ(x) = r−1 d(x, T x)

for all x ∈ C. By Caristi’s fixed point theorem, there exists an x0 ∈ C such
that x0 = f x0 . This contradicts the inequality (5.11). Therefore, T has a fixed
point.

We now consider the existence of fixed points of multivalued contraction

mappings when the domain is not necessarily convex.
Let C be a nonempty subset of a Banach space X. For given x ∈ C and
α ∈ R+ , let IC (x, α) denote the set {z ∈ X : z = x + λ(y − x) for some y ∈ C
and λ ≥ α}. Obviously, IC (x) = IC (x, 0).

Theorem 5.1.9 Let C be a nonempty closed subset of a Banach space X and

T : C → K(X) a contraction mapping satisfying the condition:

T x ⊂ IC (x, 1), x ∈ C.

Then T has a fixed point.

Proof. Suppose, for contradiction, that T has no fixed point. Let k be the
Lipschitz constant of T . Choose q such that 0 < q < 1 and k < (1 − q)/(1 + q).
Let x ∈ C. Then there exists a point z ∈ T x such that

0 < d(x, T x) = x − z.

5.1. Non-self contraction mappings 219

Because z ∈ T x ⊂ IC (x, 1), there exists a yn ∈ C and λn ≥ 1 such that

1
z − (x + λn (yn − x)) < for all n ∈ N.
n
If λn = 1 for all sufficiently large n, then we define y = z. Otherwise, there
exists a subsequence {λni } of {λn } such that λni > 1. We choose N sufficiently
large number such that
1
z − (x + λN (yN − x)) < < qx − z
N
and we define yN = y. Set h := 1/λN and w := (1 − h)x + hz. Observe that

w − x = hx − z

and    
 1 1  q  
   
 λN z − λN x − (y − x) < λN x − z ,
i.e.,
w − y < qh x − z.
It follows that

x − y ≤ x − w + w − y
≤ x − w + qh x − z
= (1 + q)w − x.

Choosing u ∈ T x and v ∈ T y such that

w − u = d(w, T x) and u − v ≤ H(T x, T y),

we have that

d(y, T y) ≤ y − v
≤ y − w + w − u + u − v
≤ y − w + w − z + H(T x, T y)
= y − w + z − x − hx − z + kx − y
< qhx − z + x − z − hx − z + kx − y
= (q − 1)w − x + x − z + kx − y
q−1
< x − y + x − z + kx − y
q+1
= d(x, T x) − rx − y,

where r = [(1 − q)/(1 + q)] − k > 0.

We define the mapping f : C → C by f (x) = y for x ∈ C. Then

x − f x < ϕ(x) − ϕ(f x) for all x ∈ C, (5.12)

220 5. Existence Theorems in Banach Spaces

where ϕ : C → [0, ∞) is the continuous function defined by ϕ(x) = r−1 d(x, T x)

for all x ∈ C. By Caristi’s fixed point theorem, there exists an x0 ∈ C such
that x0 = f x0 . This contradicts the inequality (5.12). Therefore, T has a fixed
point.

In the following theorem, we assume that each point x ∈ C has a nearest

point in the set T x.

Theorem 5.1.10 Let C be a nonempty closed subset of a Banach space X and

T : C → 2X \{∅} a contraction with closed-values and satisfying the condition:

T x ⊂ x + {λ(y − x) : λ ≥ 1, y ∈ C} for all x ∈ C.

Assume that each x ∈ C has a nearest point in T x. Then T has a fixed point.

Proof. Without loss of generality, we may assume that

T x ⊆ x + {λ(y − x) : λ > 1, y ∈ C} for all x ∈ C. (5.13)

Choose q ∈ (0, 1) and ε ∈ (0, 1) such that k < q < (1−ε)/(1+ε). By assumption,
each x ∈ C has a nearest point in T x, then there exists z ∈ T x such that

x − z = d(x, T x).

Set z = f (x). Then f (x) ∈ T x and x − f (x) = d(x, T x). By (5.13), there
exists yn ∈ C and λn > 1 such that

f (x) − (x + λn (yn − x)) → 0 as n → ∞.

Suppose, for contradiction, that T has no fixed point. Then we have a sufficient
large natural number N such that

f (x) − (x + λN (yN − x)) < εd(x, T x). (5.14)

Set
 
1 1 1
h := , w := 1− x+ f (x) = (1 − h)x + hf (x) and g(x) := yN .
λN λN λN

By (5.14), we have

yN − w = yN − ((1 − h)x + hf (x)) < εhd(x, T x) = εhx − f (x).

Also
w − f (x) = (1 − h)x − f (x) = (1 − h)d(x, T x).
5.1. Non-self contraction mappings 221

Observe that

d(yN , T yN ) ≤ yN − w + d(w, T x) + H(T x, T yN )

≤ yN − w + w − f (x) + kx − yN 
≤ yN − w + (1 − h)x − f (x) + kx − yN 
< εhx − f (x) + (1 − h)x − f (x) + kx − yN 
= (ε − 1)hx − f (x) + x − f (x) + kx − yN 
= (ε − 1)w − x + x − f (x) + kx − yN 
= (ε − 1)w − x + qx − yN 
+x − f (x) − (q − k)x − yN . (5.15)

By the choice of the integer N , we see that

εd(x, T x) > f (x) − x − λN (yN − x)

≥ λN yN − x − f (x) − x
= λN yN − x − d(x, T x),

which implies that

1+ε
yN − x < d(x, T x).
λN
By the choice of q, we have
(1 + ε)
yN − x < d(x, T x)
λN
1−ε
< hd(x, T x) (since w − x = hd(x, T x))
q
1−ε
= w − x.
q
It follows from (5.15) that

d(yN , T yN ) < x − f (x) − (q − k)x − yN .

Thus, we have

x − g(x) < ϕ(x) − ϕ(g(x)) for all x ∈ C, (5.16)

where ϕ : C → R+ is the continuous function defined by ϕ(x) = (q−k)−1 d(x, T x),

x ∈ C. By Caristi’s theorem, g has a fixed point that contradicts the strict
inequality (5.16). Therefore, T has a fixed point in C.
Theorem 5.1.11 Let C be a nonempty closed subset of a Banach space X and
T : C → F (X) a contraction. Assume that T is weakly inward on C and that
each x ∈ C has a nearest point in T x. Then T has a fixed point.

Proof. The proof is similar to the proof of Theorem 5.1.10.

222 5. Existence Theorems in Banach Spaces

5.2 Nonexpansive mappings

Let C be a nonempty subset of a normed space X and T : C → X a mapping.
Then T is said to be nonexpansive if

T x − T y ≤ x − y for all x, y ∈ C.

Recall that a sequence {xn } ⊂ C is an approximating fixed point sequence of

T if lim xn − T xn  = 0.
n→∞

The approximating fixed point sequence has a fundamental role in the study
of fixed point theory of nonexpansive mappings. We begin with the existence
and basic properties of approximating fixed point sequences of nonexpansive
mappings.
Proposition 5.2.1 Let C be a nonempty closed convex subset of a Banach
space X and T : C → X a nonexpansive mapping that is weakly inward. Then
for u ∈ C and t ∈ (0, 1), there exists exactly one point xt ∈ C such that

xt = (1 − t)u + T xt .

If C is bounded, then xt − T xt → 0 as t → 1.
Proof. For t ∈ (0, 1), the mapping Tt : C → X defined by

Tt x = (1 − t)u + tT x, x∈C (5.17)

is a contraction with Lipschitz constant t. By Theorem 5.1.2, there exists exactly

one point xt ∈ C such that

xt = Tt xt = (1 − t)u + tT xt .

If C is bounded, then

xt − T xt  = (1 − t)u − T xt  ≤ (1 − t) diam(C) → 0 as t → 1.

As an immediate consequence of Proposition 5.2.1, we have

Corollary 5.2.2 Let C be a nonempty closed convex bounded subset of a Banach
space X and T : C → C a nonexpansive mapping. Then there exists a sequence
{xn } in C such that lim xn − T xn  = 0.
n→∞

Proof. For t ∈ (0, 1), the mapping Tt : C → C defined by (5.17) is a contrac-

tion and it has exactly one fixed point xt in C. Now the result follows from
Proposition 5.2.1

It is clear from the proof of Corollary 5.2.2 that one does not need the
convexity of C. Indeed, this assumption can be replaced by the assumption
that C is star-shaped, i.e., there exists u ∈ C such that (1 − t)u + tx ∈ C for all
x ∈ C and t ∈ [0, 1]. The point u is called star-center of C.
5.2. Nonexpansive mappings 223

There is another way to obtain an approximating fixed point sequence of

nonexpansive mappings defined in a nonconvex domain.

Before proving our next theorem, we need the following lemma:

Lemma 5.2.3 Let {an } and {bn } be two sequences of a normed space X.
If there is a sequence {tn } of real numbers satisfying the conditions:
∞
(i) 0 ≤ tn ≤ t < 1 and n=1 tn = ∞,
(ii) an+1 = (1 − tn )an + tn bn for all n ∈ N,
(iii) lim an  = d,
n→∞
n
(iv) lim sup bn  ≤ d and { i=1 ti bi } is bounded,
n→∞
then d = 0.

Proof. Suppose, for contradiction, that d > 0. It follows from (iv) that
n+m−1 n+m−1
{ i=n ti bi } is bounded for all n and m. Set M = sup{ i=n t i bi  :
n, m ∈ N}. Choose a number N such that N > max{2M/d, 1}. We can choose
a positive ε such that 1 − 2ε exp((N + 1)/(1 − t)) > 1/2. It follows from
k
(i) that there exists a natural k such that N < i=1 ti ≤ N + 1. Because
lim an  = d, lim sup bn  ≤ d and ε is independent of n, without loss of
n→∞ n→∞
generality we may assume that for all n ∈ N,

d(1 − ε) < an  < d(1 + ε) and bn  < d(1 + ε).

k k
Set T = i=1 ti , S = i=1 si and sn = 1 − tn for all n ∈ N. From (ii), we
obtain

ak+1 = s1 s2 · · · sk a1 + t1 s2 s3 · · · sk b1 + · · · + tk−1 sk bk−1

+tk bk , ak+1 ∈ B := co{a1 , b1 , b2 , · · · , bk }.

Let x = T −1
k
i=1 ti bi and

y = S(1 − S)−1 {a1 + t1 (s−1

1 −T
−1
)b1
−1 −1
+t2 (s1 s2 − T )b2 + · · · + tk (S −1 − T −1 )bk }.
−1

Then x, y ∈ B and ak+1 = Sx + (1 − S)y. Hence

d(1 − ε) < ak+1  ≤ Sx + (1 − S)y

≤ Sx + (1 − S)d(1 + ε).
224 5. Existence Theorems in Banach Spaces

It follows that

x > d(1 − S −1 (2 − S)ε)

> d(1 − 2εS −1 )
 k 
= d 1 − 2ε (1 − ti )−1
i=1
 
k  
ti
= d 1 − 2ε exp log 1 +
i=1
1 − ti
 
k 
ti
≥ d 1 − 2ε exp
i=1
1 − ti
≥ d(1 − 2ε exp(T /(1 − t)))
≥ d(1 − 2ε exp((N + 1)/(1 − t)))
> d/2,

because log(1 + u) ≤ u for −1 < u < ∞. Thus, we have

 
 k  d d
x = T −1 
 t b 
i i ≤ T
−1
M≤ M= ,
i=1
2M 2

a contradiction.

Theorem 5.2.4 Let C be a nonempty subset of a normed space and T : C → C

a nonexpansive mapping. For x0 ∈ C, suppose we have a sequence {tn } of real
numbers and a sequence {Ttn } of mappings from C into itself satisfying the
conditions:
∞
(i) 0 ≤ tn ≤ t < 1 and n=0 tn = ∞,
(ii) xn+1 = Ttn xn and Ttn = (1 − tn )I + tn T, n ∈ N0 .
If {xn } is bounded sequence in C, then lim xn − T xn  = 0. Moreover,
n→∞

lim Ttn Ttn−1 · · · Tt0 x0 − Ttn−1 Ttn−2 · · · Tt0 x0  = 0.

n→∞

Proof. Because

xn+1 − T xn+1  ≤ (1 − tn )xn − T xn+1  + tn T xn − T xn+1 

≤ (1 − tn )(xn −xn+1 +xn+1 −T xn+1 )+tn xn+1 − xn 
≤ (1 − tn )xn+1 − T xn+1  + xn+1 − xn 
≤ (1 − tn )xn+1 − T xn+1  + tn xn − T xn ,

this yields
xn+1 − T xn+1  ≤ xn − T xn  for all n ∈ N0 .
5.2. Nonexpansive mappings 225

The sequence {xn − T xn } is nonincreasing; it follows that

lim xn − T xn  exists. (5.18)

n→∞

Suppose lim xn − T xn  = r. Without loss of generality, we may assume that

n→∞
tn > 0 for all n ∈ N0 . Set an := xn − T xn and bn := t−1
n (T xn − T xn+1 ). Then

an+1 = (1 − tn )an + tn bn .

Because

bn  = t−1 −1
n T xn − T xn+1  ≤ tn xn − xn+1  = xn − T xn ,

this implies that

lim sup bn  ≤ r.
n→∞

Observe that

n 
n
 t i bi  =  (T xi − T xi+1 )
i=0 i=0
= T x0 − T xn+1 
≤ x0 − xn+1 ,
n
Hence { i=0 ti bi } is bounded, because {xn } is bounded. By Lemma 5.2.3, we
have lim xn − T xn  = 0.
n→∞
Now, for all n ∈ N

xn+1 − xn  = tn xn − T xn  ≤ txn − T xn ,

which implies that

lim Ttn Ttn−1 · · · Tt0 x0 − Ttn−1 Ttn−2 · · · Tt0 x0  = 0.

n→∞

The notion of asymptotic regularity is of fundamental importance in the

study of fixed point theory of nonlinear mappings.

Let C be a nonempty subset of a normed space X and T : C → C a mapping.

Then T is said to be
(i) asymptotically regular at x0 ∈ C if lim T n x0 − T n+1 x0  = 0;
n→∞
(ii) weakly asymptotically regular at x0 ∈ C if T n x0 − T n+1 x0  0;
(iii) asymptotically regular on C if for any x ∈ C, lim T n x − T n+1 x = 0;
n→∞
(iv) uniformly asymptotically regular if lim (sup T n x − T n+1 x) = 0;
n→∞ x∈C
226 5. Existence Theorems in Banach Spaces

(v) reasonable wanderer in C if starting at any x ∈ C,

∞

T n x − T n+1 x < ∞.
n=0

Note that every uniformly asymptotically regular mapping is asymptotically

regular.

Remark 5.2.5 The asymptotic regularity of a mapping T at a point x0 implies

the existence of an approximating fixed point sequence of that mapping, but the
converse is not true.

It can be easily seen that a contraction mapping enjoys all this properties.
The following example shows that there exists a nonexpansive mapping that is
not necessarily asymptotically regular.

Example 5.2.6 Let X = R and T : X → X defined by T x = −x. Note that T

is nonexpansive, but T is not asymptotically regular.

However, a convex combination of nonexpansive mappings turns out to be

asymptotically regular in a general Banach space. Indeed, we have

Theorem 5.2.7 Let C be a nonempty convex subset of a normed space X and

T : C → C a nonexpansive mapping. For a t ∈ (0, 1), define a mapping
Tt : C → C by
Tt = (1 − t)I + tT.
If for x0 ∈ C, {Ttn x0 } is bounded, then Tt is asymptotically regular at x0 , i.e.,
lim Ttn x0 − Ttn+1 x0  = 0.
n→∞

We now turn to study a demiclosedness principle for nonexpansive mappings

in Banach spaces.

Definition 5.2.8 Let C be a nonempty subset of a Banach space X and T :

C → X a mapping. Then T is said to be demiclosed at v ∈ X if for any sequence
{xn } in C the following implication holds:

xn  u ∈ C and T xn → v imply T u = v.

Our first result concerning the demiclosedness principle of nonexpansive

mappings is in an Opial space.

Theorem 5.2.9 Let X be a Banach space that satisfies the Opial condition,
C a nonempty weakly compact subset of X, and T : C → X a nonexpansive
mapping. Then the mapping I − T is demiclosed.

Proof. Let {xn } be a sequence in C such that xn  x ∈ C and lim

n→∞
(I − T )xn − y = 0 for some y ∈ X. We show that (I − T )x = y.
5.2. Nonexpansive mappings 227

Observe that

xn − T x − y ≤ xn − T xn − y + T xn − T x,

which implies that

lim inf xn − T x − y ≤ lim inf xn − x.

n→∞ n→∞

By the Opial condition, we have

lim inf xn − x < lim inf xn − (T x + y),

n→∞ n→∞

a contradiction. Therefore, (I − T )x = y.

Corollary 5.2.10 Let X be a reflexive Banach space that satisfies the Opial
condition, C a nonempty closed convex subset of X, and T : C → X a non-
expansive mapping. Then I − T is demiclosed.

We now extend the demiclosedness principle of nonexpansive mappings in a

uniformly convex Banach space without Opial’s condition. To do so, we need
the following:

Proposition 5.2.11 Let C be a nonempty convex bounded subset of a uniformly

convex Banach space X and T : C → X a nonexpansive mapping. Then, for
any ε > 0, there exists positive number ζ(ε) > 0 such that x − T x < ε for
all x ∈ co({x0 , x1 }), whenever for x0 , x1 ∈ C with x0 − T x0  ≤ ζ(ε) and
x1 − T x1  ≤ ζ(ε).

Proof. Let x = (1 − λ)x0 + λx1 for some λ ∈ [0, 1]. Suppose x0 − x1  < ε/3,
then
x − x0  = λx1 − x0  < ε/3.
If ζ(ε) < ε/3, then we have

T x − x ≤ T x − T x0  + T x0 − x0  + x0 − x (5.19)
≤ 2x − x0  + ζ(ε)
< ε.

Hence we need only consider pairs of points x0 and x1 with x1 − x0  ≥ ε/3.
Set d := diam(C). Then for any nonnegative number λ with λ < ε/(3d),
ε
x − x0  = λx1 − x0  < .
3
Thus, if ζ(ε) < ε/3 and λ < ε/(3d), from (5.19), we have T x − x < ε.
Now let λ ≥ ε/(3d). If (1 − λ) < ε/(3d), then because x − x1  =
(1 − λ)x1 − x0  < ε/3, we have x − T x < ε.
228 5. Existence Theorems in Banach Spaces

So, without loss of generality we may assume that λ ∈ [ε/(3d), 1 − ε/(3d)]

and x0 − x1  ≥ ε/3. Note

T x − x0  ≤ T x − T x0  + T x0 − x0 
≤ λx1 − x0  + ζ(ε)

and

T x − x1  ≤ T x − T x1  + T x1 − x1 
≤ (1 − λ)x1 − x0  + ζ(ε).

Set
T x − x0 x1 − T x
u := and v := .
λx1 − x0  (1 − λ)x1 − x0 
Then
T x − x0  ζ(ε) 9dζ(ε)
u = ≤1+ ≤1+
λx1 − x0  λx1 − x0  ε2
and
T x − x1  ζ(ε) 9dζ(ε)
v = ≤1+ ≤1+ .
(1 − λ)x1 − x0  (1 − λ)x1 − x0  ε2

Observe that
 
x1 − x0  ε ε
λu + (1 − λ)v = = 1 for all λ ∈ ,1 − .
x1 − x0  3d 3d

By the uniform convexity of X, if ζ(ε) is sufficiently small and positive, it follows

that u − v ≤ ε/d. Because x = λx1 + (1 − λ)x0 , it follows that

T x − x ≤ λ(T x − x1 ) + (1 − λ)(T x − x0 )
 
ε
≤ λ(1 − λ)u − vx1 − x0  < λ(1 − λ) x1 − x0  ≤ ε.
d

Theorem 5.2.12 Let C be a nonempty closed convex bounded subset of a uni-

formly convex Banach space X and T : C → X a nonexpansive mapping. Then
I − T is demiclosed on X.

Proof. Let {xn } be a sequence in C such that xn x and lim xn −T xn −y = 0
n→∞
for some y ∈ X. Set Ty x := T x+y, x ∈ C. Then (I −Ty )xn = (I −T )xn −y → 0.
If (I − Ty )x = 0, then (I − T )x = y. Hence, we may assume without loss of
generality that y = 0. Set εn := xn − T xn . Because εn → 0 as n → ∞, we
may thin out the sequence to make the convergence faster, and we do this in
such a way that for each n

εn ≤ ζ(εn−1 ) < εn−1 ,

5.2. Nonexpansive mappings 229

where ζ(ε) for any ε > 0 is constant as described in the conclusion of Proposi-
tion 5.2.11. Hence for each point z ∈ co({xk : k ≥ n}), we have z −T z ≤ εn−1 .
Because co({xk : k ≥ n}) is weakly compact (and hence weakly closed) and con-
tains the weak limit x of the sequence {xn }, it follows that x−T x ≤ εn−1 → 0
as n → ∞. Therefore, x = T x.

Theorem 5.2.13 Let C be a nonempty closed convex bounded subset of a uni-

formly convex Banach space X and T : C → X a nonexpansive mapping. Then
(I − T )(C) is closed.

Proof. Suppose u ∈ (I − T )(C). Then there is a sequence {xn } in C such that

xn − T xn → u as n → ∞. Because C is a weakly closed and bounded set in
a reflexive Banach space X, it is weakly compact. Hence we may assume that
xn  x ∈ C. By the conclusion of Theorem 5.2.12, we have (I − T )x = u, i.e.,
(I − T )(C) is closed.

We now prove some fundamental existence theorems for nonexpansive

mappings.

Theorem 5.2.14 Let C be a nonempty closed convex bounded subset of a

Banach space X and T : C → X a nonexpansive mapping that is weakly
inward. If I − T is closed, then T has a fixed point in C.

Proof. By Proposition 5.2.1, xt − T xt → 0 as t → 1. Hence 0 lies in the closure

of (I − T )(C). Because I − T is closed, there exists a point v ∈ C such that
(I − T )v = 0.

Theorem 5.2.15 Let X be a reflexive Banach space with the Opial condition.
Let C be a nonempty closed convex bounded subset of X and T : C → C a
nonexpansive mapping. Then T has a fixed point in C.

Proof. By Corollary 5.2.2, there exists a sequence {xn } in C such that lim
n→∞
xn − T xn  = 0. By the reflexivity of X, there is a subsequence {xnk } of {xn }
such that xnk  x ∈ C. By Corollary 5.2.10, I − T is demiclosed at zero, i.e.,
xnk  x ∈ C and xnk − T xnk → 0 imply x − T x = 0. Therefore, x is a fixed
point of T .

Using Theorem 5.2.14, we prove some fundamental existence theorems for

nonexpansive mappings.

Theorem 5.2.16 (Browder’s theorem and Göhde’s theorem) – Let X

be a uniformly convex Banach space and C a nonempty closed convex bounded
subset of X. Then every nonexpansive mapping T : C → C has a fixed point in
C.

Proof. Because (I − T )(C) is closed by Theorem 5.2.13, it follows from

Theorem 5.2.14 that T has fixed point in C.
230 5. Existence Theorems in Banach Spaces

Corollary 5.2.17 (Browder’s theorem) – Let H be a Hilbert space and C a

nonempty closed convex bounded subset of H. Then every nonexpansive mapping
T : C → C has a fixed point in C.
The following result is slightly more general than Theorem 5.2.16.
Theorem 5.2.18 (Kirk’s fixed point theorem) – Let X be a Banach space
and C a nonempty weakly compact convex subset of X with normal structure.
Then every nonexpansive mapping T : C → C has a fixed point.
Proof. Let

F = {Dα ⊂ C : Dα is nonempty closed convex set such that T (Dα ) ⊂ Dα }.

Because C ∈ F, it follows that F is nonempty and it can be partially ordered

by set inclusion. Then, using Zorn’s Lemma (see Appendix A), F has a mini-
mal element, say C0 . We show that C0 has only one element. Suppose, for
contradiction, that C0 contains two elements. Hence diam(C0 ) > 0. Because
C0 is weakly compact convex, Z(C0 ) is nonempty. Let x0 ∈ Z(C0 ). Then for
x ∈ C0
T x0 − T x ≤ x0 − x ≤ rx0 (C0 ) = r(C0 ),
i.e., T x is contained in B = Br(C0 ) [T x0 ]. Thus, T (C0 ) ⊂ B and hence T (B ∩
C0 ) ⊂ B ∩ C0 . The minimality of C0 implies that C0 ⊂ B. Hence rT x0 (C0 ) ≤
r(C0 ). Because r(C0 ) ≤ rT x0 (C0 ), it follows that r(C0 ) = rTx0 (C0 ). Thus,
T x0 ∈ Z(C0 ), i.e., Z(C0 ) is mapped into itself by T . By Proposition 3.3.14,
Z(C0 ) is a nonempty closed convex subset of C0 such that T (Z(C0 )) ⊆ Z(C0 ).
It means that Z(C0 ) ∈ F and also Z(C0 ) is properly contained in C0 by Propo-
sition 3.3.15. This contradicts the minimality of C0 . Therefore, C0 consists of
a single point and hence T has a fixed point in C.

The following examples show that nonexpansive mappings may fail to have
fixed points in general Banach spaces.
Example 5.2.19 Let X be a Banach space and T : X → X a translation
mapping defined by
T x = x + a, a = 0.
Then T is nonexpansive and a fixed point free mapping.

Example 5.2.20 (Sadovski) – Let c0 be the Banach space of null sequences

and C = {x ∈ c0 : x ≤ 1}, the unit closed ball in c0 . Define a mapping
T : C → C by
T (x1 , x2 , · · · , xi , · · · ) = (1, x1 , x2 , x3 , · · · ).
It is obvious that T is nonexpansive on the closed convex bounded set C and
x = (1, 1, 1, · · · ) is a fixed point of T . But (1, 1, 1, · · · ) ∈
/ c0 . In this case, the
Banach space X = c0 is not reflexive and C does not have normal structure.

The following example shows that there exists a non-self nonexpansive map-
ping that has a fixed point.
5.2. Nonexpansive mappings 231

Example 5.2.21 Let X = R, C = [−1, 1] and T : C → X defined by T x =

1 − x, x ∈ C. Then T (−1) = 2 ∈/ C, i.e., T is not a self-mapping. But T does
have a unique fixed point in C.

We now give another example that shows that the boundedness of C in

Browder’s existence theorem may not be essential (even for non-self nonexpan-
sive mappings).

Example 5.2.22 Let X = R, C = R+ and T : C → X defined by T x =

1/(1 + x), x ∈ √
C. Then T is nonexpansive, T (C) = (0, 1], which is bounded,
and T v = v = ( 5 − 1)/2. However, C is unbounded.

Examples 5.2.21 and 5.2.22 indicate that we may be able to extend Browder-
Göhde-Kirk theorem for non-self nonexpansive mappings. We begin with the
following fundamental theorem.

Theorem 5.2.23 Let C be a nonempty closed convex subset of a Hilbert space

H and T : C → H a nonexpansive mapping with T (C) bounded. Then there
exists a z ∈ C such that z − T z = d(T z, C).

Proof. Let P be the metric projection mapping from H onto C. Then P T :

C → C is a nonexpansive mapping. Set D := co(P T (C)). Then P T (C) is
bounded, because T (C) is bounded. Thus, D is a closed convex bounded set
and P T is nonexpansive self-mapping on D. By Browder-Göhde-Kirk’s theorem,
P T has a fixed point z in D. Therefore,

T z − z = T z − P T z = d(T z, C).

In the following corollary, we replace boundedness of T (C) by “boundedness

of C”:

Corollary 5.2.24 Let C be a nonempty closed convex bounded subset of a

Hilbert space H and T : C → H a nonexpansive mapping. Then there exists a
z ∈ C such that z − T z = d(T z, C).

We now apply Theorem 5.2.23 to derive an existence theorem for fixed points
of non-self nonexpansive mappings.

Theorem 5.2.25 Let C be a nonempty closed convex subset of a Hilbert space

H and T : C → H a nonexpansive mapping. Let T (C) be bounded and T (∂C) ⊆
C. Then T has a fixed point.

Proof. By Theorem 5.2.23, there exists a point z ∈ C such that

z − T z = d(T z, C).

If T z ∈ C, then z is a fixed point of T . Otherwise, z ∈ ∂C, and hence T z ∈ C,

since T (∂C) ⊆ C. It follows that z ∈ F (T ).
232 5. Existence Theorems in Banach Spaces

Theorem 5.2.26 Let C be a nonempty closed convex bounded subset of a uni-

formly convex Banach space X and T : C → X a weakly inward nonexpansive
mapping. Then T has a fixed point.

Proof. By Proposition 5.2.1, there exists a path {xt }t∈(0,1) ⊂ C such that
xt − T xt → 0 as t → 1. Set xn := xtn , where tn → 1 as n → ∞. Let z be the
asymptotic center of {xn } with respect to C. Then

ra (T z, {xn }) = lim sup xn − T z

n→∞
≤ lim sup T xn − T z
n→∞
≤ lim sup xn − z = ra (z, {xn }).
n→∞

Because T z ∈ IC (z) and by Proposition 3.1.12, z is the asymptotic center of

{xn } with respect to IC (z), we conclude that T z = z by the uniqueness of the
asymptotic center.

We now discuss the structure of the set of fixed points of nonexpansive

mappings.

Theorem 5.2.27 Let C be a convex subset of a strictly convex Banach space

X and T : C → X a nonexpansive mapping. Then F (T ), the set of fixed points
of T is either empty or convex.

Proof. The example T x = x+a (a = 0) shows that F (T ) = ∅. Next, we assume

that F (T ) = ∅. Set x, y ∈ F (T ) and α ∈ [0, 1]. Then for z = αx + (1 − α)y, we
have
x − T z = T x − T z ≤ x − z = (1 − α)x − y
and
y − T z = T y − T z ≤ y − z = αx − y.
Hence

x − y ≤ x − T z + T z − y ≤ x − z + y − z = x − y.

This implies that

x − y = x − T z + T z − y.
Let a = x − T z and b = T z − y. Then a + b = a + b. Because X is strictly
convex, a = λb for some positive constant λ (see Proposition 2.1.7). This means
that T z is a linear combination of x and y, i.e., T z = βx + (1 − β)y for some
real β. Hence

T z − x = z − x = (1 − α)x − y = (1 − β)x − y,

T z − y = z − y = αx − y = βx − y.

Consequently, α = β, i.e., z = T z.
5.2. Nonexpansive mappings 233

Remark 5.2.28 By the continuity of T, F (T ) is always closed.

Corollary 5.2.29 Let C be a nonempty closed convex subset of a strictly convex

Banach space X and T : C → X a nonexpansive mapping. Then F (T ) is closed
and convex.

Theorem 5.2.27 is not true in a general Banach space. This fact is shown in
the following example:

Example 5.2.30 Let X = R2 be a Banach space with maximum norm defined

by
(a, b) = max{|a|, |b|} f or all x = (a, b) ∈ R2 .
Let T : X → X be a mapping defined by

T (a, b) = (|b|, b) for all (a, b) ∈ R2 .

Then T is nonexpansive and (1, 1) and (1, −1) are fixed points of T . However,
no other point in the segment joining these two points is a fixed point of T .

We now introduce the class of mappings that is properly included in the

class of nonexpansive mappings.
Let Γ denote the set of strictly increasing continuous convex functions
γ : R+ → R+ with γ(0) = 0. Let C be a nonempty convex subset of a Banach
space X. Then a mapping T : C → X is said to be of type (γ) if there exists
γ ∈ Γ such that

γ((1 − t)T x + tT y − T ((1 − t)x + ty)) ≤ x − y − T x − T y

for all x, y ∈ C and t ∈ [0, 1].

It is clear that every mapping of type (γ) is nonexpansive, but the converse is
not true in general. We derive the following interesting result, which shows that
every nonexpansive mapping defined on a convex bounded subset of a uniformly
convex Banach space is a mapping of type (γ).

Theorem 5.2.31 Let C be a nonempty convex bounded subset of a uniformly

convex Banach space X and T : C → X a nonexpansive mapping. Then there
exists a strictly increasing continuous convex function (depending on diam(C))
γ : R+ → R+ with γ(0) = 0 such that

γ((1 − t)T x + tT y − T ((1 − t)x + ty)) ≤ x − y − T x − T y

for all x, y ∈ C and t ∈ [0, 1].

Proof. Because X is uniformly convex, there exists a strictly increasing con-

tinuous convex function g : R+ → R+ with g(0) = 0 such that

2t(1 − t)g(u − v) ≤ 1 − (1 − t)u + tv (5.20)

234 5. Existence Theorems in Banach Spaces

for all u, v ∈ X with u ≤ 1, v ≤ 1 and all t ∈ [0, 1] (see Theorem 2.3.12).
It suffices to show Theorem 5.2.31 when t ∈ (0, 1).
Let x, y ∈ C and z = (1 − t)x + ty for t ∈ (0, 1). Set
Ty − Tz Tz − Tx
d := diam(C), u := and v := .
(1 − t)x − y tx − y

Then we have u ≤ 1, v ≤ 1,

Ty − Tx (1 − t)T x + tT y − T z
(1 − t)u + tv = and u − v = .
x − y t(1 − t)x − y

Take r = (1 − t)T x + tT y − T z and s−1 = t(1 − t)x − y. It follows from
(5.20) that

2t(1 − t)x − yg(rs) ≤ x − y − T x − T y,

which implies that
g(rs)
2 ≤ x − y − T x − T y. (5.21)
s
Observe that rs ≤ 2. Because t(1 − t) ≤ 1/4, it follows that t(1 − t)x − y ≤
d/4 and hence 4/d ≤ s. Note that for rs ≤ 2, the function s → g(rs)/s is
nondecreasing, and then from (5.21) we have
 
d 4r g(4r/d) g(rs)
g =2 ≤2 ≤ x − y − T x − T y.
2 d 4/d s
Therefore,

γ((1 − t)T x + tT y − T ((1 − t)x + ty)) ≤ x − y − T x − T y,

where γ(t) = dg(4t/d)/2, t ≥ 0. It can be easily verified that γ ∈ Γ.

We denote

n
n−1
= {λ = (λ1 , λ2 , · · · , λn ) : λi ≥ 0 (i = 1, 2, · · · , n) and λi = 1}.
i=1

Proposition 5.2.32 Let C be a nonempty convex bounded subset of a uniformly

convex Banach space X and T : C → X a nonexpansive mapping. Then for
each positive integer p, there exists γp ∈ Γ such that for any λ ∈ p−1 and
x1 , x2 , · · · , xp ∈ C,
     
 p p

 λi xi − λ i T xi 
γp T  ≤ max (xi − xj  − T xi − T xj ). (5.22)
1≤i,j≤p
i=1 i=1
5.2. Nonexpansive mappings 235

Proof. Once γp has been defined, we define γp+1 to be any function in Γ

satisfying the condition:
−1
γp+1 (t) ≥ γ2−1 (t) + γp−1 (t + γ2−1 (t)).

We must verify (5.22) for p + 1. Fix λ ∈ p and x1 , x2 , · · · , xp+1 ∈ C. The case

 1. Set
λp+1 = 1 is trivial. We assume that λp+1 =
λj
uj := (1 − λp+1 )xj + λp+1 xp+1 , μj :=
1 − λp+1
and
uj := (1 − λp+1 )T xj + λp+1 T xp+1 for j = 1, 2, · · · , p.
Observe that

p 
p
λj
μj = = 1;
j=1 j=1
1 − λp+1


p+1 
p
λi  p
λi xi = {(1 − λp+1 )xi + λp+1 xp+1 } = μj uj ;
i=1 i=1
1 − λp+1 j=1


p+1 
p
λ i T xi = μj uj ;
i=1 j=1

  p+1   
  p+1

T λ x − λ T x 
 i i i i
i=1 i=1
  p   
  p


= T μj uj − 
μj uj 
j=1 j=1
     
 p p
 p

≤ T μj uj − 
μj T uj  + μj T uj − uj ; (5.23)
j=1 j=1 j=1

     
 p p


γp T μj uj − μj T uj 

j=1 j=1
≤ max (uj − uk  − T uj − T uk ); (5.24)
1≤j,k≤p

uj − uk  − T uj − T uk  ≤ uj − uk  − uj − uk 

+ uk − T uk  + uj − T uj ; (5.25)

γ2 (T uj − uj ) ≤ xj − xp+1  − xj − xp+1 ; (5.26)

uj − uk  − uj − uk  = (1 − λp+1 )(xj − xk  − xj − xk )

≤ xj − xk  − xj − xk . (5.27)
236 5. Existence Theorems in Banach Spaces

Put t := max {xi − xk  − xi − xk }. By (5.26), we have

1≤i,k≤p+1

T uj − uj  ≤ γ2−1 (t).

By (5.25) and (5.27)
uj − uk  − T uj − T uk  ≤ xj − xk  − xj − xk  + 2γ2−1 (t)
≤ t + 2γ2−1 (t). (5.28)
It follows from (5.24) that
  p   
  p

T μj uj − μj T uj  −1 −1
  ≤ γp (t + γ2 (t)). (5.29)
j=1 j=1

From (5.23), (5.28), and (5.29), we have

  p+1  
  
p+1

T λi xi − λ i T xi 
 
i=1 i=1
≤ γp−1 (t + γ2−1 (t)) + γ2−1 (t)
−1
≤ γp+1 (t). (by definition of γp+1 )
Therefore,
     
 p+1 p+1

 λi xi − λi T xi 
γp+1 T  ≤ t= max
1≤i,k≤p+1
(xi −xk −T xi −T xk ).
i=1 i=1

A Banach space X is said to have the convex combination property (CAP) if

for each ε > 0, there exists an integer p(= p(ε)) ≥ 1 such that for all subsets D
in X whose diameters are uniformly bounded,
co(D) ⊂ cop (D) + Br [0],
where  
p
cop (D) = λi xi : λ ∈ p−1
, x1 , x2 , · · · , xp ∈ D .
i=1
We note that every uniformly convex Banach space X has the CAP. The
product of uniformly convex Banach spaces being uniformly convex implies
X × X has the CAP.
Theorem 5.2.33 Let C be a nonempty convex bounded subset of a uniformly
convex Banach space X and T : C → X a nonexpansive mapping. Then there
exists γ ∈ Γ such that for any finite many elements {xi }ni=1 in C and any finite
n
many nonnegative numbers {λi }ni=1 with i=1 λi = 1, the following inequality
holds:
     
 n n

 λi xi − λ i T xi 
γ T  ≤ max (xi − xj  − T xi − T xj ).
1≤i,j≤n
i=1 i=1
5.3. Multivalued nonexpansive mappings 237

Proof. First, determine γp ∈ Γ for p = 2, 3, · · · from Proposition 5.2.32.

Because X × X has the CAP and hence, given ε > 0, we can determine p
so that
co(D) ⊂ cop (D) + Bε/3 [0] × Bε/3 [0]
for every D ⊂ C × C. Set δ = γp (ε/3). Suppose x1 , x2 , · · · , xn ∈ C satisfy
xi − xj  − T xi − T xj  ≤ δ for all i, j.
Consider D = {(xi , T xi ) ∈ X × X : i = 1, 2, · · · , n}. Thus, for each λ ∈ n−1
,
there exist μ ∈ p−1 and i1 , i2 , · · · , ip ∈ {1, 2, · · · , n} such that
  
 n p
 ε
 λ x − μ 
j ij  <
x
 i i
3
i=1 j=1

and
  
 n p
 ε
 λ T x − μ T x 
ij  < .
 i i j
3
i=1 j=1

In other words, the CAP on X ×X guarantees simultaneous approximability

in X. Observe that
         
 n n
  n p

T λi xi −  
λi T xi  ≤ T λi xi − T μj xi 
 
i=1 i=1 i=1 j=1
  p   
  p


+T μj xi − μj T xij 

j=1 j=1
  
 p n

+
 μj T x ij − λ i T x 
i
j=1 i=1
ε ε ε
≤ + + = ε.
3 3 3
Thus, whenever xi − xj  − T xi − T xj  ≤ δ for all i, j, we have
    
 n n

T λ x − λ T x 
i  ≤ ε.
 i i i
i=1 i=1

Therefore, the construction of γ ∈ Γ such that γ(ε) ≤ δ for this ε − δ

prescription is a simple calculation.

5.3 Multivalued nonexpansive mappings

In this section, we consider the problem of solving the operator equation
x ∈ T x, (5.30)
where T is a multivalued nonexpansive mapping in a Banach space.
238 5. Existence Theorems in Banach Spaces

Definition 5.3.1 Let C be a nonempty subset of a Banach space X. A mapping

T : C → CB(X) is said to be nonexpansive if

H(T x, T y) ≤ x − y for all x, y ∈ C,

where H(·, ·) is the Hausdorff metric on CB(X).

Recall that the graph G(A) of a multivalued mapping A : C → 2Y is

G(A) = {(x, y) ∈ X × Y : x ∈ C, y ∈ Ax},

where Y is another Banach space.

The mapping A is said to be demiclosed at y ∈ Y if

xn (in C)  x and yn ∈ Axn → y ⇒ y ∈ Ax.

First, we show that for every compact-valued nonexpansive mapping T, I −T

is demiclosed in a Banach space with the Opial condition.

Theorem 5.3.2 Let C be a nonempty weakly compact subset of a Banach space

X with the Opial condition and T : C → K(X) a nonexpansive mapping. Then
I − T is demiclosed.

Proof. Because the domain of I − T is weakly compact, we must prove that

the graph of I − T is only sequentially closed. Let (xn , yn ) ∈ G(I − T ) be such
that xn  x and yn → y. Hence x ∈ C.
We now show that y ∈ (I − T )x. Because yn ∈ xn − T xn , yn = xn − zn for
some zn ∈ T xn . By the nonexpansiveness of T , there exists zn ∈ T x such that

zn − zn  ≤ H(T xn , T x) ≤ xn − x. (5.31)

It follows from (5.31) that

lim inf xn − x ≥ lim inf zn − zn 

n→∞ n→∞
= lim inf xn − yn − zn . (5.32)
n→∞

Because T x is compact, zn ∈ T x and yn → y, and then there exists a sub-

sequence {zn i } of {zn } such that

zn i → z ∈ T x and yni → y.

Hence from (5.32) we have

lim inf xni − x ≥ lim inf xni − y − z.

i→∞ i→∞

By the Opial condition

lim inf xni − x < lim inf xni − y − z,

i→∞ i→∞

which implies that y + z = x. Therefore, y = x − z ∈ x − T x.

5.3. Multivalued nonexpansive mappings 239

The fixed point theory of multivalued nonexpansive mappings is however

much more complicated than the corresponding theory of single-valued non-
expansive mappings. We will concentrate on some important existence theorems.

We begin with the existence of fixed points of compact-valued nonexpansive

mappings in Banach spaces.
Theorem 5.3.3 Let X be a Banach space with the Opial condition, C a non-
empty weakly compact convex subset of X, and T : C → K(C) a nonexpansive
mapping. Then T has a fixed point in C.
Proof. Let u be an element in C and let {an } be a sequence in (0, 1) such that
lim an = 1. For each n ∈ N, define Tn : C → K(C) by
n→∞

Tn x = (1 − an )u + an T x, x ∈ C. (5.33)

Then Tn is a contraction mapping. By Theorem 4.2.7 (Nadler’s fixed point

theorem), there exists xn ∈ C such that xn ∈ Tn xn . Because C is weakly
compact, there exists a subsequence {xni } of {xn } such that xni  v ∈ C.
From (5.33) we have
xn = (1 − an )u + an zn ,
where zn ∈ T xn . Observe that

xn − zn  = (1 − an )zn − u.

Hence yn = xn − zn ∈ (I − T )xn and yn → 0. Thus, (xni , yni ) ∈ G(I − T ) and

xni  x and yni → 0, it follows from the demiclosedness of I − T at zero that
0 ∈ (I − T )v, Therefore, v ∈ T v.

Proposition 5.3.4 Let C be a nonempty closed convex subset of a uniformly

convex Banach space X and T : C → K(C) a nonexpansive mapping. Suppose
there exists a bounded sequence {xn } in C such that d(xn , T xn ) → 0 as n → ∞.
Then T has a fixed point in C.

Proof. We may assume that {xn } is regular and thus asymptotically uniform.
Let Za (C, {xn }) = {z} and ra (C, {xn }) = r. Choose yn ∈ T xn such that
xn − yn  → 0 as n → ∞. By the compactness of T z, select zn ∈ T z such that

yn − zn  ≤ H(T xn , T z) ≤ xn − z. (5.34)

Because T z is compact, there exists a subsequence {znk } of {zn } such that

znk → v ∈ T z.
By the regularity of {xn }, we have

ra (C, {xnk }) = ra (C, {xn }) = r and Za (C, {xn }) = Za (C, {xnk }) = {z}.

Because
xnk − v ≤ xnk − ynk  + ynk − znk  + znk − v,
240 5. Existence Theorems in Banach Spaces

it follows from (5.34) that

lim sup xnk − v ≤ r

k→∞

and hence Za (C, {xn }) = z = v.

Theorem 5.3.5 Let X be a uniformly convex Banach space, C a nonempty
closed convex subset of X (with 0 ∈ C), and T : C → K(C) a nonexpansive
mapping. Suppose the set

E = {x ∈ C : λx ∈ T x f or some λ > 1}

is bounded. Then T has a fixed point in C.

Proof. Let {tn } be a sequence in (0,1) with tn → 1. Then for each n ∈ N,
tn T : C → K(C) is a contraction mapping. Then Nadler’s theorem implies that
xn ∈ tn T xn for some xn ∈ C. Now select yn ∈ T xn such that xn = tn yn , which
yields
d(xn , T xn ) ≤ xn − yn  = (t−1
n − 1)xn .

Because {xn } is in E and E is bounded, we obtain that d(xn , T xn ) → 0 as

n → ∞. Therefore, there exists a fixed point of T in C by Proposi-
tion 5.3.4.
Theorem 5.3.6 Let C be a nonempty closed convex bounded subset of a uni-
formly convex Banach space X and T : C → K(X) a nonexpansive mapping.
If T x ⊂ IC (x) for all x ∈ C, then T has a fixed point in C.
Proof. Let u ∈ C be fixed and let {an } be a sequence in (0, 1) such that
lim an = 1. For each n ∈ N, define Tn : C → 2X by
n→∞

Tn x = (1 − an )u + an T x, x ∈ C.

Then Tn is a multivalued contraction with Lipschitz constant an . Because IC (x)

is convex for each x ∈ C, it follows that

Tn x ⊂ IC (x), x ∈ C. (5.35)

Observe that T (C) = ∪x∈C T x is a bounded set. Now let x ∈ C, z ∈ T x and let
K = u − z + diam(∪x∈C T x). If y ∈ Tn x, then y = (1 − an )u + an w for some
w ∈ T x and

d(y, T x) ≤ (1 − an )u − z ≤ (1 − an )K. (5.36)

Also for any x ∈ T x, we have (1 − an )u + an x ∈ Tn x and so

d(x, Tn x) ≤ (1 − an )K. (5.37)

Together (5.36) and (5.37) imply that H(T x, Tn x) ≤ (1 − an )K → 0 uniformly

for x ∈ C as n → ∞.
5.3. Multivalued nonexpansive mappings 241

Because Tn x ⊂ IC (x), x ∈ C by (5.35) and the values of T are compact, Tn

has a fixed point xn ∈ C by Theorem 5.1.7. This means that
xn ∈ Tn xn = (1 − an )u + an T xn .
Thus
d(xn , T xn ) ≤ H(Tn xn , T xn ).
By the uniform convergence of {Tn } we have
d(xn , T xn ) → 0 as n → ∞.
For each n ∈ N, choose yn ∈ T xn such that
xn − yn  = d(xn , T xn ).
By Propositions 3.1.1 and 3.1.2, passing to a subsequence if necessary, we may
assume that {xn } is regular and asymptotically uniform. Let Za (C, {xn }) = {z}
and ra (C, {xn }) = r.
Because T is compact-valued mapping, we can select zn ∈ T z such that
yn − zn  ≤ H(T xn , T z) ≤ xn − z. (5.38)
Let {zni } be a subsequence of {zn } such that zni → v as i → ∞. Hence v ∈ T z.
Because v ∈ T z ⊂ IC (z), there exists t ∈ (0, 1) such that
(1 − t)z + tv ∈ C.
Note Za (C, {xni }) = z, ra (C, {xni }) = r and {xn } is regular and asymptotically
uniform. Set wi := (1 − t)z + txni , i ∈ N, t ∈ (0, 1). Proposition 3.1.7 implies
that Za (C, {wi }) = z and ra (C, {wi }) = tr. Set
pi : = (1 − t)z + tyni
qi : = (1 − t)z + tzni ,
w : = (1 − t)z + tv.
Then for each i ∈ N,
wi − w ≤ wi − pi  + pi − qi  + qi − w
≤ txni − yni  + tyni − zni  + tzni − v
≤ t(xni − yni  + xni − z + zni − v) (from (5.38))

From zni → v and xn − yn → 0, we have

lim sup wi − w ≤ t lim sup xni − z = tr.
i→∞ i→∞

Because Za (C, {wi }) = {z} and ra (C, {wi }) = tr, by the uniqueness of asymp-
totic center we have w = z. Thus, z = v ∈ T z.

Finally, we obtain some existence theorems for multivalued nonexpansive

non-self mappings in which the convexity of domain is not necessary.
242 5. Existence Theorems in Banach Spaces

Theorem 5.3.7 Let C be a nonempty weakly compact star-shaped subset of a

uniformly convex Banach space X and T : C → K(X) a nonexpansive mapping.
If for each x ∈ ∂C, T x ⊂ C and λx + (1 − λ)T x ⊂ C for some λ ∈ (0, 1) or
T x ⊂ int(C), then T has a fixed point in C.
Proof. Let p be the star-center of C and {an } a sequence in (0, 1) with an → 1.
Define
Tn x = (1 − an )p + an T x, x ∈ C, n ∈ N.
Then Tn is a multivalued contraction. By Theorem 4.2.14, each Tn has a fixed
point xn such that

xn = (1 − an )p + an T xn . (5.39)

Because {xn } is bounded, we have ra (C, {xni }) = ra (C, {xn }) and Za (C, {xn }) ⊂
Za (C, {xni }). Let z ∈ Za (C, {xn }). Because T z is compact, there exists zn ∈ T z
such that for yn ∈ T xn

yn − zn  ≤ H(T xn , T z) ≤ xn − z. (5.40)

Let {zni } be a subsequence of {zn } such that zni → v ∈ T z. Because

Za (C, {xn }) ⊂ Za (C, {xni }), it follows that z ∈ Za (C, {xni }). From (5.39),
we get
xni − yni  = (1 − ani )xni − p → 0 as i → ∞.
Hence

lim sup xni − v ≤ lim sup(xni − yni  + yni − zni  + zni − v)
i→∞ i→∞
= lim sup yni − zni 
i→∞
≤ lim sup xni − z (from (5.40))
i→∞
= inf{lim sup xni − x : x ∈ C},
i→∞

and this means that v ∈ Za (C, {xni }).

If z ∈ ∂C, then by hypothesis for v ∈ T z, there exists w ∈ C such that

w = (1 − λ)v + λz, λ ∈ (0, 1).

Suppose that v = z. By uniform convexity of X, we have some δ ∈ (0, 1)

lim sup xni − w ≤ (1 − δ) inf{lim sup xni − y : y ∈ C},

i→∞ i→∞

which is a contradiction of the choice of w. If z ∈ Za (X, {xn }), we have

ra (X, {xni }) ≤ lim sup v − xni 

i→∞
≤ lim sup(v − zni  + zni − yni  + yni − xni )
i→∞
5.4. Asymptotically nonexpansive mappings 243

≤ lim sup zni − yni 

i→∞
≤ lim sup z − xni  = ra (X, {xni }).
i→∞

Hence v ∈ Za (X, {xni }). By uniform convexity of X, we obtain v = z

∈ T z.
Theorem 5.3.8 Let C be a nonempty closed bounded subset of a Banach space
X and T : C → K(X) a mapping that satisfies the following conditions:
(i) H(T x, T y) ≤ x − y for all x, y ∈ C,
(ii) T x ⊂ IC (x, 1) for all x ∈ C,
  
(iii) x∈C (x − T x) = x∈C y∈T x (x − y) is a closed subset of X.
Then T has a fixed point in C.
Proof. Without loss of generality, we may assume that 0 ∈ C. Let {an } be a
sequence in (0, 1) such that an → 1 as n → ∞. For each n ∈ N, define
Tn x = an T x, x ∈ C.
It is easy to see that Tn x ∈ K(X) and Tn x ⊂ IC (x, 1) for each n ∈ N and all
x ∈ C. Furthermore,
H(Tn x, Tn y) ≤ an x − y for all x, y ∈ C, n ∈ N.
By Theorem 5.1.9, there exists xn ∈ C such that
xn = Tn xn = an T xn .
We choose yn ∈ T xn satisfying xn = an yn . By the boundedness of C,
 
1
xn − yn  = − 1 xn  → 0.
an
Because ∪x∈C (x − T x) is closed, we have 0 ∈ ∪x∈C (x − T x). Therefore, there
exists some z ∈ C such that z ∈ T z.

5.4 Asymptotically nonexpansive mappings

Let C be a nonempty subset of a Banach space X. A mapping T : C → C is
said to be asymptotically nonexpansive 1 if for each n ∈ N, there exists a positive
constant kn ≥ 1 with lim kn = 1 such that
n→∞

T x − T n y ≤ kn x − y for all x, y ∈ C.
n

The following example shows that the class of asymptotically nonexpansive

mappings is essentially wider than the class of nonexpansive mappings.
1 The notion of asymptotically nonexpansive mappings was introduced by Goebel and Kirk

in 1972.
244 5. Existence Theorems in Banach Spaces

Example 5.4.1 Let BH be the closed unit ball in the Hilbert space H = 2 and
T : BH → BH a mapping defined by

T (x1 , x2 , x3 , · · · ) = (0, x21 , a2 x2 , a3 x3 , · · · ),

∞
where {ai } is a sequence of real numbers such that 0 < ai < 1 and i=2 ai = 1/2.
Then
T x − T y ≤ 2x − y f or all x, y ∈ BH ,
i.e., T is Lipschitzian, but not nonexpansive. Observe that
n
T n x − T n y ≤ 2 ai x − y f or all x, y ∈ BH and n ≥ 2.
i=2
n
Here kn = 2 i=2 ai → 1 as n → ∞. Therefore, T is asymptotically nonexpan-
sive, but not nonexpansive.
There is also a connection between the demiclosedness principle and the
fixed point theory of asymptotically nonexpansive mappings. Some simple
results concerning the demiclosedness principle of asymptotically nonexpansive
mappings are given in the following theorems:
We first establish Proposition 5.4.2, which shows that the asymptotic center
of every bounded AF P S of an asymptotically nonexpansive mapping is a fixed
point of the mapping in uniformly convex Banach spaces.
Proposition 5.4.2 Let C be a nonempty closed convex subset of a uniformly
convex Banach space X and T : C → C an asymptotically nonexpansive map-
ping. If {yn } is a bounded sequence in C such that lim yn − T yn  = 0 and
n→∞
Za (C, {yn }) = {v}, then v is a fixed point in C.
Proof. We define a sequence {zn } in C by zm = T m v, m ∈ N. For integers
m, n ∈ N, we have

zm − yn  ≤ T m v − T m yn  + T m yn − T m−1 yn  + · · · + T yn − yn 

m−1
≤ km v − yn  + (T yn − yn  + ki yn − T yn ). (5.41)
i=1

Then by (5.41) we have

ra (zm , {yn }) = lim sup yn − zm  ≤ km ra (v, {yn }) = km ra (C, {yn }).
n→∞

Hence

|ra (zm , {yn }) − ra (C, {yn })| ≤ (km − 1)ra (C, {yn }) → 0 as m → ∞.

It follows from Theorem 3.1.8 that T m v → v. By the continuity of T , we have

T z = T ( lim T m z) = lim T m+1 z = z.

m→∞ m→∞
5.4. Asymptotically nonexpansive mappings 245

Theorem 5.4.3 Let X be a uniformly convex Banach space with the Opial
condition, C a nonempty closed convex (not necessarily bounded) subset of X,
and T : C → C an asymptotically nonexpansive mapping. Then I − T is
demiclosed at zero.

Proof. Let {xn } be a sequence in C such that xn  x ∈ C and xn − T xn → 0

as n → ∞. By Theorem 3.2.9 (which states that if a sequence {un } of a
nonempty closed convex subset of a uniformly convex Banach space having the
Opial condition converges weakly to u ∈ C, then u is asymptotic center of
{un } with respect to C), the asymptotic center of {xn } is x. It follows from
Proposition 5.4.2 that x is a fixed point of T .

To prove the next theorem concerning a demiclosedness principle, we need

the following:

Proposition 5.4.4 Let X be a Banach space with the Opial condition, C a

nonempty weakly compact convex subset of X, and T : C → C an asymptoti-
cally nonexpansive mapping. If {xn } is a sequence in C such that xn  x ∈
C and xn − T xn → 0, then {T n x} converges weakly to x.
∞

Proof. Set Am := co({T i x}i≥m ), m ∈ N and A := Am . Because C is weakly
m=1
n
compact, then A is nonempty and A = co(ωw ({T x})) by Theorem 1.9.22.
We show that T n x  x, and this means that A = {x}. Because {xn } is
bounded, we define a functional f : C → R by f (y) = lim sup xn − y, y ∈ C.
n→∞
Suppose, for contradiction, that y0 ∈ A such that y0 = x. Then by the Opial
condition, we have
f (x) < f (y0 ).
Because km − 1 → 0, then for ε := (f (y0 ) − f (x))/(1 + f (x)) > 0, there exists an
integer m0 ∈ N such that km − 1 < ε for all m ≥ m0 . Because y0 ∈ Am0 +1 , there
p
exist an integer p ∈ N and nonnegative numbers t1 , t2 , · · · , tp with i=1 ti = 1
such that
 
  p

y0 − t T m0 +j 
x
 j  < ε. (5.42)
j=1

Note

f (y0 ) = lim sup xn − y0 

n→∞
 
p 
p 
≤ lim sup xn − tj T m0 +j x +  tj T m0 +j x − y0 
n→∞
j=1 j=1

p
< tj lim sup xn − T m0 +j x + ε (by (5.42))
n→∞
j=1
246 5. Existence Theorems in Banach Spaces


p
≤ tj lim sup T m0 +j xn − T m0 +j x + ε (as xn − T xn → 0)
n→∞
j=1
p
≤ tj km0 +j lim sup xn − x + ε
n→∞
j=1
p
≤ tj (ε + 1)f (x) + ε (as km < ε + 1 for m ≥ m0 )
j=1
= (ε + 1)f (x) + ε = f (y0 ),

a contradiction. Therefore, x = y0 , i.e., A = {x}.

Theorem 5.4.5 Let X be a Banach space with the locally uniform Opial con-
dition, C a nonempty weakly compact convex subset of X, and T : C → C an
asymptotically nonexpansive mapping. Then I − T is demiclosed at zero.

Proof. Suppose {xn } is a sequence in C such that xn  x and xn −

T xn → 0. By Proposition 5.4.4, we have T n x  x. Because T is asymptotically
nonexpansive, we have

lim sup(lim sup T n x − T m x) ≤ lim sup(lim sup km T n−m x − x)

m→∞ n→∞ m→∞ n→∞
= lim sup T x − x.
n
n→∞

Proposition 3.2.19 implies that T m x → x. By the continuity of T , we have

x = T x.

We have already shown in Section 5.2 that every nonexpansive mapping is

demiclosed in a uniformly convex Banach space. The following Theorem 5.4.6
shows that Theorem 5.2.12(a) is valid also for asymptotically nonexpansive
mappings.

Theorem 5.4.6 Let C be a nonempty closed convex bounded subset of a uni-

formly convex Banach space X and T : C → C an asymptotically nonexpansive
mapping. Then I − T is demiclosed at zero.

Proof. Let {xn } be a sequence in C such that x  x ∈ C and xn − T xn → 0

as n → ∞. We show that T n x → x. Indeed, because {xn } is weakly convergent
to x, there exists for each integer n ∈ N a convex combination


m(n)
(n) (n)

m(n)
(n)
yn = ti xi+n (ti ≥ 0 and ti = 1)
i=1 i=1

such that yn −x < 1/n. For an arbitrary but fixed j ∈ N, because (I −T )xn →
0, there is an n0 = n0 (ε, j) so large that 1/n0 < ε and (I − T j )xn  < ε for
n ≥ n0 .
5.4. Asymptotically nonexpansive mappings 247

Because X is uniformly convex, by Theorem 5.2.33, there exists a strictly

increasing convex and continuous function g : R+ → R+ , g(0) = 0 such that for
any nonexpansive mapping S : C → X, for any finite many elements {ui }ni=1 in
n
C, and for any finite many nonnegative numbers {ti }ni=1 with i=1 ti = 1, the
following inequality holds:

n 
n
g(S( ti u i ) − ti Sui ) ≤ max (ui − uj  − Sui − Suj ). (5.43)
1≤i, j≤n
i=1 i=1

Suppose Lj is the Lipschitz constant of T j . Then Lj −1 T j is nonexpansive in

C, and it follows from (5.43) that


m(n)
(n)

m(n)
(n)
T yn − yn 
j
≤ T yn −
j
ti T j xi+n  + ti T j xi+n − yn 
i=1 i=1
 m(n)
  
m(n)
(n) (n)
≤ T j ti xi+n − ti T j xi+n 
i=1 i=1


m(n)
(n)
+ ti T j xi+n − xi+n 
i=1

−1
≤ Lj g (xi+n − xk+n 
max
1≤i, k≤m(n)

−1
−Lj T xi+n − T xk+n ) + ε
j j

≤ Lj g −1 (2ε + (1 − L−1
j ) diam(C)) + ε, (5.44)

because

xi+n − xk+n  − L−1

j T xi+n − T xk+n 
j j

≤ xi+n − T j xi+n  + xk+n − T j xk+n 

+ (1 − L−1
j )T xi+n − T xk+n 
j j

≤ 2ε + (1 − L−1
j ) diam(C).

Taking the limit superior as n → ∞ in (5.44), we obtain

lim sup T j yn − yn  ≤ Lj g −1 (2ε + (1 − L−1

j ) diam(C))
n→∞
+ ε for all j ∈ N. (5.45)

For each j ∈ N, we have

T j x − x ≤ T j x − T j yn  + T j yn − yn  + yn − x
≤ (Lj + 1)yn − x + T j yn − yn . (5.46)
248 5. Existence Theorems in Banach Spaces

Because yn → x, it follows from (5.45) and (5.46) that

T j x − x ≤ lim sup((Lj + 1)yn − x + T j yn − yn )

n→∞
≤ Lj g −1 (2ε + (1 − L−1
j ) diam(C)) + ε

= g −1 (0) = 0 as j → ∞ and ε → 0.

This shows that T n x → x and by the continuity of T , we obtain that

x = T x.

We now give a fundamental existence theorem for asymptotically nonexpan-

sive mappings in a uniformly convex Banach space.

Theorem 5.4.7 (Goebel and Kirk’s fixed point theorem) – Let C be a

nonempty closed convex bounded subset of a uniformly convex Banach space X
and T : C → C an asymptotically nonexpansive mapping. Then T has a fixed
point in C.

Proof. For fixed y ∈ C and r > 0, set

Ry := {r : there exists k ∈ N with C ∩ (∩∞

i=k Br [T y]) = ∅} and d := diam(C).
i

Then d ∈ Ry . Hence Ry = ∅. Let r0 = inf{r : r ∈ Ry }. For each ε > 0, we

define
Cε = ∪∞ ∞ i
k=1 (∩i=k Br0 +ε [T y]).

Thus, for each ε > 0, the set Cε ∩ C is nonempty and convex. The reflexivity
of X implies that
∩ε>0 (C ε ∩ C) = ∅.
Note that for x ∈ ∩ε>0 (C ε ∩ C) and η > 0, there exists an integer n0 such that

x − T n y ≤ r0 + η for all n ≥ n0 .

Now, let x ∈ ∩ε>0 (C ε ∩ C) and suppose, for contradiction, that the sequence
{T n x} does not converge strongly to x. Then there exist ε > 0 and a subse-
quence {T ni x} of {T n x} such that

T ni x − x ≥ ε for all i = 1, 2, · · · .

Suppose kn is the Lipschitz constant of T n . Then for m > n, we have

T n x − T m x ≤ kn x − T m−n x.

Assume that r0 > 0 and choose α > 0 such that

  
ε
1 − δX (r0 + α) < r0 .
r0 + α
5.4. Asymptotically nonexpansive mappings 249

Select n such that

 
α
x − T n x ≥ ε and kn r0 + ≤ r0 + α.
2

If n0 ≥ n is sufficiently large, then m > n0 implies

α
x − T m−n y ≤ r0 + .
2
Because
 
α
T n x − T m y ≤ kn x − T m−n y ≤ kn r0 + ≤ r0 + α
2

and
x − T m y ≤ r0 + α,
it follows from the uniform convexity of X that for m > n0 ,
  
1 ε
 (x + T n x) − T m y ≤ 1 − δX (r0 + α) < r0 .
2 r0 + α

This contradicts the definition of r0 . Hence we conclude that r0 = 0 or x = T x.

But r0 = 0 implies that {T n y} is a Cauchy sequence and hence lim T n y =
n→∞
x = T x. Therefore, the set ∩ε>0 (C ε ∩ C) is a singleton that is a fixed point
of T .

In our next existence theorem, the boundedness of C is not necessary.

In particular, we will show that the existence of a fixed point of an asymp-
totically nonexpansive mapping in a uniformly convex Banach space is not only
equivalent to the existence of a bounded orbit at a point, but it is also equivalent
to the existence of a bounded AF P S.

Theorem 5.4.8 Let C be a nonempty closed convex (but not necessarily boun-
ded) subset of a uniformly convex Banach space X and T : C → C an asymp-
totically nonexpansive mapping. Then the following statements are equivalent:
(a) T has a fixed point.
(b) There exists a point x0 ∈ C such that the sequence {T n x0 } is bounded.
(c) There exists a bounded sequence {yn } in C such that lim yn −T yn  = 0.
n→∞

Proof. (a) ⇒ (b) and (a) ⇒ (c) follows easily.

(b) ⇒ (a). Let x0 ∈ C be a point such that the sequence {xn = T n x0 }
is bounded. By Theorem 3.1.5, there exists a unique point z ∈ C such that
Za (C, {xn }) = {z}. Define a sequence {ym } in C by ym = T m z for all m ∈ N.
Let km be the Lipschitz constant of the iterates T m . For n > m ≥ 1, we have

xn − ym  = T m xn−m − T m z ≤ km xn−m − z.

250 5. Existence Theorems in Banach Spaces

This implies that

ra (ym , {xn }) ≤ km ra (z, {xn }) = km ra (C, {xn }).

This shows that lim ra (ym , {xn }) = ra (C, {xn }). By Theorem 3.1.8, lim
m→∞ m→∞
ym = z. By the continuity of T , v is a fixed point of T .
(c) ⇒ (a). Let {yn } be a bounded sequence in C such that lim yn −T yn  =
n→∞
0. Let Za (C, {yn }) = {v}. Therefore, Proposition 5.4.2 implies that v is a fixed
point of T .

We have seen in a Corollary 5.2.29 that F (T ) is closed and convex in strictly

convex Banach space for nonexpansive mappings. However, Corollary 5.2.29 is
not true for asymptotically nonexpansive mappings. In fact, we have
Theorem 5.4.9 Let C be a nonempty closed convex bounded subset of a uni-
formly convex Banach space X and T : C → C an asymptotically nonexpansive
mapping. Then F (T ) is closed and convex.
Proof. The closedness of F (T ) is obvious. To show convexity, it is sufficient to
prove that z = (x + y)/2 ∈ F (T ) for x, y ∈ F (T ). For each n ∈ N, we have
1
x − T n z = T n x − T n z ≤ kn x − z = kn x − y,
2
1
y − T n z = T n y − T n z ≤ kn y − z = kn x − y.
2
By the uniform convexity of X, we have
     
1 2 1 2
z − T z ≤
n
1 − δX kn x − y ≤ 1 − δX kn diam(C)
2 kn 2 kn
and hence T n z → z as n → ∞. It follows from the continuity of T that z is a
fixed point of T .

5.5 Uniformly L-Lipschitzian mappings

Let C be a nonempty subset of a normed space X and T : C → C a mapping.
Then T is said to be uniformly L-Lipschitzian if for each n ∈ N, there exists a
positive constant L such that T n x − T n y ≤ Lx − y for all x, y ∈ C.

Note that every nonexpansive mapping is uniformly L-Lipschitzian with

L = 1 and every asymptotically nonexpansive mapping with sequence {kn }
is also uniformly L-Lipschitzian with L = sup kn .
n∈N

The following proposition shows that the class of uniformly L-Lipschitzian

mappings on C can be characterized as the class of mappings on C that are
nonexpansive relative to some metric on C that is equivalent to the norm.
5.5. Uniformly L-Lipschitzian mappings 251

Proposition 5.5.1 Let C be a nonempty subset of a Banach space X and T :

C → C a uniformly L-Lipschitzian mapping. Then there exists a metric d on C
that is equivalent to the norm metric such that T is nonexpansive with respect
to d.
Proof. Define the metric d on C by

d(x, y) = sup{T n x − T n y : n = 0, 1, 2, · · · }, x, y ∈ C.

Because
x − y ≤ d(x, y) ≤ Lx − y,
this means that the metric d on C is equivalent to the norm metric. Furthermore,
T is nonexpansive with respect to d.
Proposition 5.5.2 Let C be a nonempty subset of a Banach space and ρ a
metric on C satisfying the condition:

αx − y ≤ ρ(x, y) ≤ βx − y for all x, y ∈ C. (5.47)

If T : C → C is a nonexpansive mapping with respect to ρ, then T is uniformly

β/α-Lipschitzian with respect to  · .
Proof. Because

αx − y ≤ ρ(x, y) ≤ βx − y for all x, y ∈ C

and T : C → C is nonexpansive with respect to ρ(x, y), then for n = 1, 2, · · · ,

1 1 β
T n x − T n y ≤ ρ(T n x, T n y) ≤ ρ(x, y) ≤ x − y.
α α α

Therefore, T is uniformly β/α-Lipschitzian.

The following Theorem 5.5.3 shows that the Goebel and Kirk’s fixed point
theorem (see Theorem 5.4.7) for asymptotically nonexpansive mappings remains
valid for a broader class of uniformly L-Lipschitzian mappings with L < γ, where
γ is sufficiently near one.
Theorem 5.5.3 (Goebel and Kirk’s fixed point theorem) – Let C be a
nonempty closed convex bounded subset of a uniformly convex Banach space X.
Then every uniformly L-Lipschitzian mapping T : C → C with L < γ has a
fixed point in C, where γ > 1 is the unique solution of the equation
  
1
t 1 − δX = 1.
t
Proof. We take γ to be the solution of the equation t(1 − δX (1/t)) = 1 and
assume that 1 ≤ L < γ, i.e., L satisfies the inequality:
  
1
L 1 − δX < 1. (5.48)
L
252 5. Existence Theorems in Banach Spaces

For x ∈ C, set d(x) := lim sup x − T n x. Let

n→∞

R = {r > 0 : there exists n ∈ N with C ∩ (∩∞

i=n Br [T x]) = ∅}.
i

Then R is nonempty (because R contains the diameter of C), so we can define

r0 = r0 (x) = inf{r > 0 : r ∈ R}.

For each ε > 0, we define

Cε = ∪∞ ∞ i
n=1 (∩i=n Br0 +ε [T x]).

Hence for each ε > 0, the sets Cε are nonempty and convex. The reflexivity of
X implies that
∩ε>0 (C ε ∩ C) = ∅.
Let z = z(x) ∈ ∩ε>0 (C ε ∩C). Notice that z and r0 have the following properties:

(i) for each ε > 0, Br0 +ε [z] contains almost all terms of the sequence {T i x},

(ii) given u ∈ C and r < r0 , the set {i : u − T i x > r} is infinite.

Now if r0 = 0 or if d(z) = 0, then lim T i x = z yielding z = T z. So we may

i→∞
assume that r0 > 0 and d(z) = lim sup z − T i z > 0.
n→∞
Let ε > 0 with 0 < ε ≤ d(z) and select j ∈ N such that z − T i z ≥ d(z) − ε.
By (i), there exists an integer n0 such that if i ≥ n0 , then

z − T i x ≤ r0 + ε ≤ L(r0 + ε), (as 1 ≤ L)

and it follows for i − j ≥ n0 that

T j z − T i x = T j z − T j (T i−j )x
≤ Lz − T i−j x
≤ L(r0 + ε).

Set w := (z + T j z)/2. By a property of δX , we have

 
 z − T ix + T j z − T ix 
w − T x = 
i  
2 
  
d(z) − ε
≤ 1 − δX L(r0 + ε) for all i ≥ n0 + j.
L(r0 + ε)

This implies (by (ii)) that

  
d(z) − ε
r0 ≤ 1 − δX L(r0 + ε).
L(r0 + ε)
5.5. Uniformly L-Lipschitzian mappings 253

By the continuity of δX , we have

  
d(z)
L 1 − δX ≥ 1.
Lr0
This implies that  
−1 1
d(z) ≤ LδX 1− r0 .
L
From (ii), we have r0 ≤ d(x), and hence
 
−1 1
d(z) ≤ LδX 1− d(x) = αd(x),
L
−1
where α = LδX (1 − 1/L) < 1 because L satisfies (5.48).
To complete the proof, fix x0 ∈ C and define the sequence {xn } by

xm+1 = z(xm ), m = 0, 1, 2, · · · ,

where z(xm ) is selected in the same manner as z(x). Now if for any m we have
r0 (xm ) = 0, then T xm+1 = xm+1 . Otherwise,

xm − xm+1  ≤ 2d(xm ) ≤ 2αm d(x0 ),

which implies that {xm } is a Cauchy sequence. Hence xm → v ∈ C. Note

v − T i v ≤ v − xm  + xm − T i xm  + T i xm − T i v
≤ (1 + L)v − xm  + xm − T i xm 
≤ (1 + L)v − xm  + d(xm ) → 0 as m → ∞.

By the continuity of T , we have v = T v.

Corollary 5.5.4 Let C be a nonempty closed convex bounded subset
√ of Hilbert
space H and T : C → C uniformly L-Lipschitzian with L < γ = 5/2. Then T
has a fixed point.
Let us give an example of a uniformly L-Lipschitzian mapping that is fixed
point free.
Example 5.5.5 Let  · 2 be the usual Euclidean norm on the Hilbert space
H = 2 and let BH be the closed unit ( · 2 -unit) ball and let S : 2 → 2 be the
right shift operator defined by

S(x1 , x2 , · · · ) = (0, x1 , x2 , · · · ).

Then the mapping T : BH → BH defined by

(1 − x2 )e + Sx
Tx = , x ∈ BH , e = (1, 0, 0, · · · )
(1 − x2 )e + Sx2
is uniformly L-Lipschitzian with L = 2, but it has no fixed point in BH .
254 5. Existence Theorems in Banach Spaces

Applying Theorem 5.5.3, we have

Theorem 5.5.6 Let C be a nonempty closed convex bounded subset of a uni-
formly convex Banach space X. If T : C → C is nonexpansive with respect to a
metric ρ(x, y) on C satisfying (5.47), where β/α < γ for γ as in Theorem 5.5.3,
then T has a fixed point in C.
The following theorem has a sharper estimate for L than Goebel and Kirk’s
fixed point theorem even in a more general Banach space.
Theorem 5.5.7 (Casini and Maluta’s fixed point theorem) - Every
Banach space X with uniformly normal structure has the fixed point property
for uniformly L-Lipschitzian mappings with L < N (X).
Proof. Let C be a nonempty closed convex bounded subset of X and T : C → C
a uniformly L-Lipschitzian mapping. For any x0 ∈ C, consider the sequence
{T n x0 } in C. By Theorem 3.4.20(b) for {T n x0 }, there exists x1 ∈ co({T n x0 })
such that

ra (x1 , {T n x0 }) ≤ Ñ (X) diama ({T n x0 }), (5.49)

where Ñ (X) = 1/N (X). Observe that

diama ({T n x0 }) = lim (sup{T i x0 − T j x0  : i, j ≥ k})

k→∞
≤ sup T i x0 − T j x0 
i≥j≥0

≤ L sup x0 − T i−j x0 

i≥j≥0
≤ L sup x0 − T n x0 .
n∈N

From (5.49), we have

ra (x1 , {T n x0 }) ≤ LÑ (X) sup x0 − T n x0 . (5.50)

n∈N

Moreover, for ∈ N, we have

ra (T  x1 , {T n x0 }) = lim sup T  x1 − T n x0 
n→∞
≤ L lim sup x1 − T n− x0 
n→∞
= Lra (x1 , {T n x0 }). (5.51)

From (5.50) and (5.51), we have

ra (T  x1 , {T n x0 }) ≤ L2 Ñ (X) sup x0 − T n x0 .

n∈N

By Theorem 3.4.20(a), we have

x1 − T  x1  ≤ ra (T  x1 , {T n x0 }) ≤ L2 Ñ (X) sup x0 − T n x0 ,

n∈N
5.5. Uniformly L-Lipschitzian mappings 255

which implies that

sup x1 − T  x1  ≤ L2 Ñ (X) sup x0 − T n x0 .

∈N n∈N

Thus, for any x0 ∈ C, we can inductively define a sequence {xm }m≥0 in the
following manner:

sup xm+1 − T n xm+1  ≤ L2 Ñ (X) sup xm − T n xm  for all m ∈ N0

n∈N n∈N

and

ra (xm+1 , {T n xm }) ≤ LÑ (X) sup xm − T n xm  for all m ∈ N0 .

n∈N

Set Dm := sup xm − T n xm  for m ≥ 0 and η := L2 Ñ (X) < 1. Then

n∈N

Dm+1 ≤ ηDm ≤ η 2 Dm−1 ≤ · · · ≤ η m+1 D0 → 0 as m → ∞.

Observe that

xm − xm+1  ≤ xm − T n xm  + xm+1 − T n xm  ≤ Dm + xm+1 − T n xm .

Taking the limit superior as n → ∞, we have

xm − xm+1  ≤ (1 + LÑ (X))Dm ,

and it follows that {xm } is a Cauchy sequence in C. Let lim xm = v ∈ C.

m→∞
Hence

v − T v ≤ v − xm  + xm − T xm  + T xm − T v
≤ (1 + L)xm − v + Dm → 0 as m → ∞.

The following result is a slight generalization of Theorem 5.5.7.

Theorem 5.5.8 Let X be a Banach space with uniformly normal structure, C
a nonempty bounded subset of X, and T : C → C a uniformly L-Lipschitzian
mapping with L < N (X). Suppose that there exists a nonempty closed convex
bounded subset M of C with the following property (P):

x ∈ M implies ωw ({T n x}) ⊂ M. (5.52)

Then T has a fixed point in M .

Proof. For any x0 ∈ M and each n ∈ N, consider a sequence {T j x0 }j≥n in C.
By Theorem 3.4.20(b), we have yn ∈ co({T j x0 }j≥n ) such that

lim sup T j x0 − yn  ≤ Ñ (X) diama ({T j x0 }j≥n ). (5.53)

j→∞
256 5. Existence Theorems in Banach Spaces

Theorem 3.4.16 implies that X is reflexive, and there exists a subsequence {yni }
of {yn } such that yni  x1 ∈ X. From (5.53) and w-lsc of the functional
ra (·, {T n x0 }), we have
ra (x1 , {T j x0 }) ≤ lim inf ra (yni , {T j x0 })
i→∞
≤ lim sup ra (yn , {T j x0 })
n→∞

≤ Ñ (X) diama ({T n x0 }).

It can be easily seen that x1 ∈ ∩∞
n=1 co({T x0 }j≥n ) and that
j

x1 − y ≤ lim sup T n x0 − y for all y ∈ X.

n→∞

Using Theorem 1.9.22, we obtain that co(ωw ({T n x0 })) = ∩∞

n=1 co({T x0 }j≥n ).
j

It follows that x1 ∈ co(ωw ({T x0 })). Using property (P) we obtain that
n

x1 ∈ M . By repeating the above process, we can obtain a sequence {xm }

in M with the properties:

for all integers m ≥ 0,

lim sup xm+1 − T n xm  ≤ Ñ (X) diama ({T n xm }) (5.54)
n→∞

and
xm+1 − y ≤ lim sup T n xm − y for all y ∈ X. (5.55)
n→∞

Set Dm := sup xm − T n xm  for all m = 0, 1, 2, · · · . Note

n∈N

diama ({T n xm }) = lim (sup{T i xm − T j xm  : i, j ≥ k})

k→∞
≤ sup T i xm − T j xm 
i≥j≥0

≤ L sup xm − T i−j xm 

i≥j≥0
≤ LDm .
Moreover, from (5.54) we have, for ∈N
lim sup T  xm+1 − T n xm  ≤ L lim sup xm+1 − T n− xm 
n→∞ n→∞
≤ L lim sup xm+1 − T n xm 
n→∞

≤ LÑ (X) diama ({T n xm })

≤ L2 Ñ (X)Dm .
From (5.55), we have
xm+1 − T  xm+1  ≤ lim sup T n xm − T  xm+1 
n→∞

≤ L2 Ñ (X)Dm ,
5.5. Uniformly L-Lipschitzian mappings 257

which implies that

Dm+1 ≤ ηDm for all m = 0, 1, 2, · · · ,

where η = L2 Ñ (X) < 1. Now proceeding with the same argument as in the
proof of Theorem 5.5.7, we conclude that {xm } converges strongly to a fixed
point of T in M .

Using weak uniformly normal structure coefficient W CS(X), we now

establish an existence theorem for asymptotically regular Lipschitzian mappings.
Before proving Theorem 5.5.10, we first establish a preliminary result.

Proposition 5.5.9 Let C be a nonempty closed subset of a Banach space X

and T : C → C an asymptotically regular mapping such that for some m ∈ N,
T m is continuous. If lim sup T ni x − z = 0 for some x ∈ C and z ∈ C, then
i→∞
z ∈ F (T ).

Proof. Note T ni x → z as i → ∞. So

z − T ni +m x ≤ z − T ni x + T ni x − T ni +m x

m−1
≤ z − T ni x + T ni +ν x − T ni +ν+1 x.
ν=0

By the asymptotic regularity of T , T ni +m x → z as i → ∞. Because T m is

continuous, it follows that

T m z = T m ( lim T ni x) = lim (T ni +m x) = z.
i→∞ i→∞

m
Because z = T z = T 2m
z = ··· = T ms
z for all s ∈ N,

z − T z = T ms z − T ms+1  → 0 as s → ∞.

Therefore, z ∈ F (T ).
Theorem 5.5.10 Let X be a Banach space with W CS(X) > 1, C a nonempty
weakly compact convex subset of X, T : C → C a Lipschitzian mapping such
that lim inf σ(T n ) < W CS(X). If T is asymptotically regular on C, then T
n→∞
has a fixed point.
Proof. Because one can easily construct a nonempty closed convex separable
subset C0 of C such that C0 is invariant under T , i.e., T (C0 ) ⊂ C0 , we may
assume that C is itself separable. The separability of C makes it possible to
select a subsequence {ni } of natural numbers such that

lim inf σ(T n ) = lim σ(T ni ) < W CS(X)

n→∞ i→∞

and
{T ni x} converges weakly for every x ∈ C.
258 5. Existence Theorems in Banach Spaces

Now we can construct a sequence {xm } in C in the following way:


x0 ∈ C arbitrary,
xm+1 = w − lim T ni xm , m ≥ 0.
i→∞

By the asymptotic regularity of T on C, we have

xm+1 = w − lim T ni +k xm for all k ≥ 0.

i→∞

Set
rm := lim sup xm+1 − T ni xm  and L := lim sup σ(T ni ).
i→∞ i→∞

By the definition of W CS(X), we obtain

W CS(X) = sup{M > 0 : M · lim sup xn − u ≤ D[{xn }]},

n→∞

where the supremum is taken over all weakly (not strongly) convergent sequences
{xn } in X with xn  u. Then we have
1
rm = lim sup xm+1 − T ni xm  ≤ D[{T ni xm }].
i→∞ W CS(X)

However, from the w-lsc of the norm of X, we obtain that

D[{T ni xm }] = lim sup(lim sup T ni xm − T nj xm )

i→∞ j→∞

≤ lim sup(lim sup(T ni xm − T ni +nj xm 

i→∞ j→∞

+T xm − T nj xm ))
ni +nj

≤ lim sup(lim sup(σ(T ni )xm − T nj xm 

i→∞ j→∞
i −1
n
+ T nj +ν xm − T nj +ν+1 xm ))
ν=0
= lim sup σ(T ni ) × lim sup xm − T nj xm 
i→∞ j→∞
≤ L lim sup(lim sup T nk xm−1 − T nj xm )
j→∞ k→∞

= L lim sup(lim sup(T nj xm − T nj +nk xm−1 

j→∞ k→∞

+T xm−1 − T nk xm−1 )

nj +nk

≤ L lim sup(lim sup(σ(T nj )xm − T nk xm−1 

j→∞ k→∞
nj −1

+ T nk +ν xm−1 − T nk +ν+1 xm−1 ))
ν=0
= L2 rm−1 .
5.6. Non-Lipschitzian mappings 259

Hence
L2
rm ≤ rm−1 = ηrm−1 for all m ∈ N,
W CS(X)
where η = L2 /W CS(X) < 1.
Again by w-lsc of the norm of X, we have

xm − xm+1  ≤ lim sup(xm − T ni xm  + T ni xm − xm+1 )

i→∞
≤ lim sup xm − T ni xm  + lim sup xm+1 − T ni xm 
i→∞ i→∞
≤ lim sup(lim sup T nj xm−1 − T ni xm ) + rm
i→∞ j→∞

≤ lim sup(lim sup(T ni xm − T ni +nj xm−1 

i→∞ j→∞

+T ni +nj
xm−1 − T nj xm−1 )) + rm
≤ lim sup(lim sup(σ(T ni )xm − T nj xm−1 
i→∞ j→∞
i −1
n
+ T nj +ν xm−1 − T nj +ν+1 xm−1 )) + rm
ν=0
= Lrm−1 + rm
≤ (L + η)rm−1 ,

and it follows that {xm } is a Cauchy sequence in C. Let lim xm = p ∈ C.

m→∞
Note

p − T ni p ≤ p − xm+1  + xm+1 − T ni xm  + T ni xm − T ni p
≤ p − xm+1  + xm+1 − T ni xm  + σ(T ni )xm − p,

which implies that

lim sup p − T ni p ≤ xm+1 − p + rm + Lxm − p → 0 as m → ∞.

i→∞

Hence T ni p → p as i → ∞. Therefore, T p = p by Proposition 5.5.9.

5.6 Non-Lipschitzian mappings

Let C be a nonempty subset of a Banach space X, T : C → C a mapping, and
fix a sequence {an } in R+ with an → 0. Recall that
 n 
T x − T n y
n
η(T ) = sup : x, y ∈ C, x = y
x − y + an
is nearly Lipschitz constant of T n . Then T is nearly Lipschitzian with sequence
{(η(T n ), an )} if

T n x − T n y ≤ η(T n )(x − y + an ) for all x, y ∈ C and n ∈ N.

260 5. Existence Theorems in Banach Spaces

T is nearly asymptotically nonexpansive with sequence {(η(T n ), an )} if for each

n ∈ N, η(T n ) ≥ 1 with lim η(T n ) = 1 and
n→∞

T n x − T n y ≤ η(T n )(x − y + an ) for all x, y ∈ C.

In this section, we study fixed point theorems for non-Lipschitzian mappings

in Banach spaces. We begin with the following preliminary result.

Proposition 5.6.1 Let C be a nonempty closed subset of a Banach space and

T : C → C a demicontinuous mapping. Suppose that T n u → x∗ as n → ∞ for
some u, x∗ ∈ C. Then x∗ is a fixed point of T .

Proof. Let un = T n u for all n ∈ N. Then {un } and {T un } converge strongly to

x∗ . By the demicontinuity of T, {T un } converges weakly to T x∗ . By uniqueness
of weak limits of {T un }, we have x∗ = T x∗ .
In view of Proposition 5.6.1, we remark that Theorem 4.1.18 is valid for
demicontinuous nearly Lipschitzian mappings. In fact,

Theorem 5.6.2 Let C be a nonempty closed subset of a Banach space and T :

C → C a demicontinuous nearly Lipschitzian mapping with sequence {(η(T n ),
an )}. Suppose η∞ (T ) = lim sup[η(T n )]1/n < 1. Then we have the following:
n→∞
(a) T has a unique fixed point v ∈ C.
(b) For each x ∈ C, the sequence {T n x} converges to v.
∞
(c) T n x − v ≤ i=n η(T i )(x − T x + M ) for all x ∈ C and n ∈ N, where
M = sup an .
n∈N

Proof. Let x ∈ C. By Theorem 4.1.18, {T n x} is a Cauchy sequence in C.

Let lim T n x = v ∈ C. It follows from Proposition 5.6.1 that v is a fixed point
n→∞
of T .
We now give demiclosedness principle for nearly Lipschitzian mappings in a
Banach space.

Theorem 5.6.3 Let X be a Banach space with a weakly continuous duality

mapping Jμ : X → X ∗ with gauge function μ. Let C be a nonempty closed
convex subset of X and T : C → C a uniformly continuous nearly Lipschitzian
mapping with sequence {(η(T n ), an )} such that lim η(T n ) = 1. Then I − T is
n→∞
demiclosed at zero.

Proof. Let {xn } be a sequence in C such that xn  x and lim xn −T xn  = 0.

n→∞
Then x ∈ C because C is weakly closed. The uniform continuity of T implies
that
lim xn − T m xn  = 0 for all m ∈ N.
n→∞
5.6. Non-Lipschitzian mappings 261

It follows that T m xn  x for all m ∈ N. Set rm := lim sup T m xn − x, m ∈ N.

n→∞
Let m, s ∈ N. Because T m+s xn  x as n → ∞, by the Opial condition, we
have
rm+s = lim sup T m+s xn − x < lim sup T m+s xn − T s x
n→∞ n→∞
≤ lim sup η(T s )(T m xn − x + as )
n→∞
= η(T s )(rm + as ).
It follows that
lim sup rs ≤ rm ,
s→∞

which implies that

lim sup rs ≤ lim inf rm .
s→∞ m→∞

Thus, lim rm exists. Suppose lim rm = r for some r > 0. Noting by

m→∞ m→∞
Theorem 2.5.23 that
1
Φ(x + y) = Φ(x) + y, Jμ (x + ty)dt for all x, y ∈ X.
0

For m, s ∈ N, we have
Φ(T m+s xn − x) = Φ(T m+s xn − T m x + T m x − x)
= Φ(T m+s xn − T m x)
1
+ T m x − x, Jμ (T m+s xn − T m x + t(T m x − x))dt
0
≤ Φ(η(T m )(T s xn − x + am ))
1
+ T m x − x, Jμ (T m+s xn − T m x + t(T m x − x))dt.
0

Because T m+s xn  x as n → ∞, we obtain

Φ(rm+s ) = Φ(lim sup T m+s xn − x)
n→∞
≤ Φ(η(T m )(rs + am ))
1
− T m x − x, Jμ ((1 − t)(T m x − x))dt
0
= Φ(η(T m )(rs + am ))
1
− T m x − xμ(tT m x − x)dt
0
= Φ(η(T m )(rs + am )) − Φ(T m x − x),
which implies that
Φ(T m x − x) ≤ Φ(η(T m )(rs + am )) − Φ(rm+s ).
262 5. Existence Theorems in Banach Spaces

Because lim rs exists, we have

s→∞

Φ(T m x − x) ≤ Φ(η(T m )(r + am )) − Φ(r) → 0 as m → ∞.

Thus, T m x → x. Therefore, by the continuity of T , we have x = T x.

Corollary 5.6.4 Let X be a Banach space with a weakly continuous duality
mapping Jμ : X → X ∗ with gauge function μ. Let C be a nonempty closed
convex subset of X and T : C → C an asymptotically nonexpansive mapping.
Then I − T is demiclosed at zero.

The following theorem is an extension of Theorem 5.4.8 for demicontinuous

nearly asymptotically nonexpansive mappings.
Theorem 5.6.5 Let C be a nonempty closed convex (but not necessarily
bounded) subset of a uniformly convex Banach space X and T : C → C a
demicontinuous nearly asymptotically nonexpansive mapping. Then the follow-
ing statements are equivalent:
(a) T has a fixed point in C.
(b) There is a point x0 ∈ C such that the sequence {T n x0 } is bounded.
Proof. (a) ⇒ (b) follows easily.
(b) ⇒ (a). Assume that x0 ∈ C is such that the sequence {xn } defined by
xn = T n x0 is bounded. By Theorem 3.1.5, let Za (C, {xn }) = {z}, and let {ym }
be a sequence in C defined by ym = T m z for m = 1, 2, · · · .
For two integers n > m ≥ 1, we have

xn − ym  = T n x0 − T m z = T m (T n−m x0 ) − T m z

and hence
ra (ym , {xn }) ≤ η(T m )(ra (z, {xn }) + am ).
This shows that ra (ym , {xn }) → ra (C, {xn }) as m → ∞. By Theorem 3.1.8,
this would imply that T m z → z as m → ∞. Because T is demicontinuous,
hence by Proposition 5.6.1, z ∈ F (T ).

In the following results, W CS(X) plays an important role in the existence

of fixed points of nearly Lipschitzian mappings.
Theorem 5.6.6 Let X be a Banach space with W CS(X) > 1, C a nonempty
weakly compact convex subset of X, and T : C → C a demicontinuous nearly
Lipschitzian mapping with sequence {(an , η(T n ))} such that η(T n ) → 1 as n →
∞. If T is weakly asymptotically regular on C, i.e., T n x − T n+1 x  0 for all
x ∈ C, then T has a fixed point in C.
Proof. Let U be a free ultrafilter on N. We then define a mapping S on C by

Sx = w − lim T n x, x ∈ C.
U
5.6. Non-Lipschitzian mappings 263

Because C is weakly compact, Sx is well defined for all x ∈ C. The asymptotic

nonexpansiveness of T clearly implies that S is a nonexpansive on C. Hence, S
has a fixed point v ∈ C, i.e.,
w − lim T n v = v.
U

This yields a subsequence {T ni v} of {T n v} converging weakly to v. By the

property of W CS(X)
1
lim sup T ni v − v ≤ D[{T ni v}]. (5.56)
i→∞ W CS(X)
By the weak asymptotic regularity of T , we have
T nt +p v  v as t → ∞ for any p ≥ 0.
On the other hand, for each i, j ∈ N with i > j, by the w-lsc of the norm  · 
we have
T ni v − T nj v ≤ η(T nj )(v − T ni −nj v + anj )
≤ η(T nj )(lim inf T nt +p v − T ni −nj v + anj )
t→∞
(with p = ni − nj )
≤ η(T nj )[η(T ni −nj )(lim inf v − T nt v + ani −nj ) + anj ].
t→∞

Taking the limit superior as i → ∞, we get

lim sup T ni v − T nj v ≤ η(T nj ) lim sup v − T nt v + anj
i→∞ t→∞

and hence from (5.56), we have

1
lim sup T ni v − v ≤ D[{T ni v}]
i→∞ W CS(X)
1
≤ lim sup( lim T ni v − T nj v)
W CS(X) j→∞ i→∞
1
≤ lim sup v − T nt v,
W CS(X) t→∞
which implies that
(W CS(X) − 1) lim sup T ni v − v ≤ 0,
i→∞

i.e., lim T ni v = v. By the demicontinuity of T , lim T ni v = v implies w −

i→∞ i→∞
lim T ni +1 v = T v. By weak asymptotic regularity of T , we have T ni v−T ni +1 v →
i→∞
0 implies w − lim T ni +1 v = v. Hence by the uniqueness of weak limit of
i→∞
{T ni +1
v}, we conclude that v = T v.
Corollary 5.6.7 Let X be a Banach space with W CS(X) > 1, C a nonempty
weakly compact convex subset of X, and T : C → C an asymptotically non-
expansive mapping. If T is weakly asymptotically regular on C, then T has a
fixed point in C.
264 5. Existence Theorems in Banach Spaces

5.7 Pseudocontractive mappings

In this section, our aim is to study a class of continuous pseudocontractive
mappings in Banach spaces. Strongly pseudocontractive mappings will play an
important role in many of the existence theorems for pseudocontractive map-
pings.

Let X be a Banach space with dual X ∗ . Then a mapping T with domain

Dom(T ) and range R(T ) in X is said to be strongly pseudocontractive if there
exists a positive constant k and such that

x − y ≤ (1 + t)(x − y) − kt(T x − T y)

for all x, y ∈ Dom(T ) and all t > 0.

For k = 1, such mappings are called pseudocontractive.

Following Proposition 2.4.7, we are able to formulate an equivalent definition

of strongly pseudocontractive mapping as follows:

The mapping T : Dom(T ) ⊆ X → X is strongly pseudocontractive if for

each x, y ∈ Dom(T ), there exist a positive constant k and j(x − y) ∈ J(x − y)
such that

T x − T y, j(x − y) ≤ kx − y2 .

It is easy to see that the mapping T is pseudocontractive if for each x, y ∈ D(T ),

there exists j(x − y) ∈ J(x − y) such that

T x − T y, j(x − y) ≤ x − y2 . (5.57)

We note that every nonexpansive (contraction) mapping is pseudocontractive

(strongly pseudocontractive), but the converse is not true. In fact, if T is non-
expansive with domain Dom(T ), for each x, y ∈ Dom(T ) and j(x−y) ∈ J(x−y),
we have

T x − T y, j(x − y) ≤ T x − T yj(x − y)∗ ≤ x − y2 .

We now give examples of pseudocontractive mappings that are not non-

expansive.
Example 5.7.1 Let H = R2 be the Hilbert space under the usual Euclidean
inner product. If x = (a, b) ∈ H, we define x⊥ = (b, −a) ∈ H. Trivially, we
have

x, x⊥  = 0, x⊥  = x;

x⊥ , y ⊥  = x, y, x⊥ − y ⊥  = x − y;

and
x, y ⊥  + x⊥ , y = 0 for all x, y ∈ H.
5.7. Pseudocontractive mappings 265

Let C be the closed unit ball in H, C1 = {x ∈ H : x ≤ 1/2} and C2 =

{x ∈ H : 1/2 ≤ x ≤ 1}.
We define the mapping T : C → C by

x + x⊥ if x ∈ C1 ,
Tx =
x/x − x + x⊥ if x ∈ C2 .

We now show that T is Lipschitz continuous. One easily shows that

√
T x − T y = 2x − y for all x, y ∈ C1 .

For x, y ∈ C2 , we have
 2
 x y  2
 
 x − y  =
x · y
(x · y − x, y)

1
= (x − y2 − (x − y)2 )
x · y
2
≤ x − y2
x · y
≤ 8x − y2 ,

which implies that

 
 x y 
T x − T y ≤   ⊥ ⊥
 x y  + x − y + x − y 
−

≤ 5x − y.

Now let x and y be in the interiors of C1 and C2 , respectively. Then there

exist λ ∈ (0, 1) and z ∈ C1 ∩ C2 for which z = λx + (1 − λ)y. Hence

T x − T y ≤ T x − T z + T z − T y
√
≤ 2x − z + 5z − y
≤ 5x − z + 5z − y
= 5x − y.

Thus, T x − T y ≤ 5x − y for all x, y ∈ C, i.e., T is Lipschitzian on C.

We now show that T is pseudocontractive. For x, y ∈ C, set Γ(x, y) :=
x − y2 − T x − T y, x − y. Hence to show T is a pseudocontractive, we need to
prove that Γ(x, y) ≥ 0 for all x, y ∈ C. We consider the following three cases:
Case 1. x, y ∈ C1 :
Obviously, Γ(x, y) ≥ 0 for all x, y ∈ C1 .
Case 2. x ∈ C1 and y ∈ C2 :
266 5. Existence Theorems in Banach Spaces

We have
 
x, y
Γ(x, y) = x − y2 − x2 − y2 + y −
y
x, y
= 2y2 − y + (1 − 2y) .
y

Because 1 − 2y ≤ 0 for y ∈ C2 , x, y/(xy) has its minimum, for fixed
x and y when x, y/(xy) = 1. We conclude that

Γ(x, y) ≥ 2y2 − y + x − 2xy

= (y − x)(2y − 1)
≥ 0 for all x ∈ C1 , y ∈ C2 .

Case 3. x, y ∈ C2 :
Observe that
 
1 1
T x − T y, x − y = x − x2 + y − y2 + 2 − − x, y
x y
x, y
= x − x2 + y − y2 + (2xy − x − y) .
xy

Hence Γ(x, y) = 2x2 + 2y2 − x − y − (4xy − x − y)x, y/
(xy). It is easy to see that 4xy − x − y ≥ 0 for all x, y ∈ C2 . Hence
for fixed x and y, Γ(x, y) has a minimum when x, y/(xy) = 1. This
minimum is 2x2 + 2y2 − 4xy = 2(x − y)2 . Thus, Γ(x, y) ≥ 0 for
all x, y ∈ C2 . Therefore, T is Lipschitz continuous pseudocontractive, but it is
not nonexpansive.

Example 5.7.2 Let X = R and T : Dom(T ) = [0, 1] → R be defined by

T x = (1 − x2/3 )3/2 , x ∈ [0, 1].

Because T is monotonically decreasing, T is pseudocontractive. Observe that

     3/2  3/2
1 1 15 3
T 3 −T 3 = −
4 2 16 4
|(15)3/2 − (12)3/2 | 7 1 1
= > = 3− 3 .
64 64 4 2

Hence T is not nonexpansive. Thus, continuous pseudocontractive is not neces-

sarily nonexpansive.

We now establish an equivalent definition of pseudocontractive mapping in

a Hilbert space.
5.7. Pseudocontractive mappings 267

Proposition 5.7.3 Let H be a Hilbert space. Then following are equivalent:

(a) T is a pseudocontractive mapping with domain Dom(T ).
(b) T x − T y2 ≤ x − y2 + (I − T )x − (I − T )y2 for all x, y ∈ Dom(T ).
Proof. (a) ⇒ (b). Let T be a pseudocontractive with Dom(T ). Then from
(5.57), we have

T x − T y, x − y ≤ x − y2 for all x, y ∈ Dom(T ).

Observe that

T x − T y2 = (I − T )x − (I − T )y − (x − y)2

= x − y2 + (I − T )x − (I − T )y2
−2(I − T )x − (I − T )y, x − y
= x − y2 + (I − T )x − (I − T )2
−2{x − y2 − T x − T y, x − y}
≤ x − y2 + (I − T )x − (I − T )y2 .

(b) ⇒ (a). Suppose for all x, y ∈ Dom(T ),

T x − T y2 ≤ x − y2 + (I − T )x − (I − T )y2

holds. Then we have

T x − T y2 ≤ x − y2 + x − y − (T x − T y)2

≤ x − y2 + x − y2 + T x − T y2 − 2T x − T y, x − y,

and it follows that T x − T y, x − y ≤ x − y2 .

Proposition 5.7.4 Let H be a Hilbert space and T a nonlinear mapping on
H with domain Dom(T ). Then T is strongly pseudocontractive if the following
inequality is satisfied:

T x − T y2 ≤ x − y2 + k(I − T )x − (I − T )y2 (5.58)

for all x, y ∈ Dom(T ), where k ∈ (0, 1).

Proof. From (5.58), we have for all x, y ∈ Dom(T )

T x − T y2 ≤ x − y2 + k{x − y2 + T x − T y2 − 2T x − T y, x − y},

which implies that

1−k
T x − T y, x − y ≤ t(k)x − y2 − T x − T y2 ,
2k

where t(k) = (1 + k)/(2k).

268 5. Existence Theorems in Banach Spaces

The pseudocontractive mappings are easily seen to be more general than the
nonexpansive mappings. They derive their importance in nonlinear functional
analysis via their connection with an important class of nonlinear operators
defined as follows:

Let X be a Banach space. An operator A : Dom(A) ⊆ X → X is said to be

accretive if for each x, y ∈ Dom(T ) and t > 0, the following inequality holds:

x − y ≤ x − y + t(Ax − Ay). (5.59)

An operator A is said to be dissipative if −A is accretive and A is expansive if

Ax − Ay ≥ x − y for all x, y ∈ Dom(A).

Example 5.7.5 Let X = R and A : Dom(A) ⊆ X → R a real-valued increasing

(nonincreasing) function. Then A is accretive (dissipative).
By Proposition 2.4.7, we obtain
Proposition 5.7.6 Let X be a Banach space and T a nonlinear operator on X
with domain Dom(T ). Then the following are equivalent:
(a) T is an accretive operator.
(b) For each x, y ∈ Dom(T ), there exists j(x − y) ∈ J(x − y) such that

T x − T y, j(x − y) ≥ 0.

Proposition 5.7.7 Let X be a Banach space and T : Dom(T ) ⊆ X → X a

mapping. Then T is pseudocontractive if and only if I − T is accretive.

Proposition 5.7.8 Let X be a Banach space and A : Dom(T ) ⊆ X → X an

operator. If (I + tA) is expansive for all t > 0, then the following hold:
(i) A is accretive.
(ii) (I +tA)−1 exists and (I +tA)−1 is a nonexpansive mapping from R(I +tA)
into Dom(A).

Proof. Proposition 5.7.8 follows from (5.59).

We now introduce a semigroup of type ω:

Let C be a nonempty closed subset of a Banach space X and ω a real number.
A semigroup of type ω on C is a function S : R+ ×C → C satisfying the following
conditions:
(i) S(t1 + t2 )x = S(t1 )S(t2 )x for t1 , t2 ≥ 0 and x ∈ C;
(ii) S(t)x − S(t)y ≤ eωt x − y for t ≥ 0 and x, y ∈ C;
5.7. Pseudocontractive mappings 269

(iii) S(0)x = x for x ∈ C;

(iv) S(t)x is continuous in t ≥ 0 for each x ∈ C.
For each t > 0 and x ∈ C, let At x = (S(t)x − x)/t, Dom(A) = {x ∈ C :
lim At x exists} and Ax = lim At x for all x ∈ Dom(A). Then A is called the
t→0+ t→0+
(strong) generator of a semigroup S.
If ω = 0, S is said to be semigroup of nonexpansive mappings.
Remark 5.7.9 If C is a nonempty closed convex subset of a Banach space X
and A : C → X a continuous mapping, the following are equivalent:
(i) A is the generator of a semigroup S of type ω on C.
(ii) For x, y ∈ C, there exists j(x − y) ∈ J(x − y) such that

Ax − Ay, j(x − y) ≤ ωx − y2

d(x + hAx, C)
and lim+ = 0 for all x ∈ C.
h→0 h
The following proposition plays a key role in the existence of solutions of
nonlinear operator equations.
Proposition 5.7.10 Let C be a nonempty closed convex subset of a Banach
space X and T : C → X a continuous strongly pseudocontractive mapping with
constant k ∈ (0, 1) such that

d(x + hT x, C)
lim+ = 0.
h→0 h
Then for each ε > 0 with εk < 1, the range of (I − εT ) contains C.
Proof. Let u be an element in C and ε a positive number such that εk < 1.
For each x ∈ C, define Bx = εT x + u − x. Then B is continuous on C and for
x, y ∈ C, there exists j(x − y) ∈ J(x − y) such that

Bx − By, j(x − y) = ε(T x − T y) − (x − y), j(x − y)

≤ (εk − 1)x − y2 .

Let x ∈ C and for each h > 0, let xh be an element of C such that

d(x + hT x, C) ≥ x + hT x − xh  − h2 .

Note that
x + hT x − xh 
lim = 0.
h→0+ h
Define
εxh + hu + (1 − h)x
yh = .
1+ε
270 5. Existence Theorems in Banach Spaces

Because C is convex and xh , u and x are in C, it follows that yh ∈ C, whenever

h ∈ (0, 1). Hence if h ∈ (0, 1) and λ = (1 + ε)−1 h, we have
d(x + λBx, C) ≤ x + λBx − yh 
= [(1 + ε)x + hεT x + hu − hx]
−[εxh + hu + (1 − h)x](1 + ε)−1
= ε(1 + ε)−1 x + hT x − xh .
Hence lim+ d(x + λBx, C)/λ = 0. So B is the generator of a semigroup V of
λ→0
type (kε − 1) on C. Because
V (t)x − V (t)y ≤ e(kε−1)t x − y for t > 0,
V (t) is a contraction from C into itself. Hence V (t) has a unique fixed point
xt ∈ C. Because
V (s)xt = V (s)V (t)xt = V (t)V (s)xt ,
there is a unique point z ∈ C such that V (t)z = z for all t ≥ 0. Thus Bz = 0
and z − εT z = u. Therefore, the range of (I − εT ) contains C.
The following example shows that convexity of C cannot be removed from
Proposition 5.7.10.
Example 5.7.11 Let X = R2 be the Euclidean space, C = {(x, y) ∈ R2 :
x2 + y 2 = 1} and A(x, y) = (y, −x) for all (x, y) ∈ C. Then A is genera-
tor of a semigroup U of type 0 on C (in particular, U (t)(x, y) = (x cos t +
y sin t, −x sin t + y cos t)), but the image of C under I − εA does not intersect
C for any ε > 0.

We now introduce more general classes of nonlinear operators:

Let α : R+ → R+ be a function such that α(0) = 0 and lim inf α(r) > 0 for
r→r0
all r0 > 0.
A mapping A : C ⊆ X → X is said to be α-strongly accretive if for each
x, y ∈ C, there exists j(x − y) ∈ J(x − y) such that
Ax − Ay, j(x − y) ≥ α(x − y)x − y. (5.60)
A mapping T is said to be α-strongly pseudocontractive if I − T is α-strongly
accretive.

If α(r) = kr for some k > 0, then A is strongly accretive (with strongly

accretive constant k) as in this case (5.60) reduces to
Ax − Ay, j(x − y) ≥ kx − y2
and T = I − A (k ∈ (0, 1)) is strongly pseudocontractive (with strongly pseu-
docontractive constant 1 − k) as in this case (5.60) takes the form of
T x − T y, j(x − y) ≤ (1 − k)x − y2 .
5.7. Pseudocontractive mappings 271

One may easily see that the pseudocontractivity of T implies the strong
pseudocontractivity of tT for t ∈ (0, 1). Also, the accretiveness of A implies the
strong accretiveness of tI + A for t > 0.

Note that every continuous strongly accretive self-mapping on X is surjec-

tive.

The following proposition shows that the sum of accretive and strongly
accretive mappings is strongly accretive.
Proposition 5.7.12 Let X be a smooth Banach space, and T : X → X a
strongly accretive mapping with the strongly accretive constant k ∈ (0, 1). Let
S : X → X be an accretive mapping. Then T + S : X → X is also a strongly
accretive mapping with the strongly accretive constant k.
Proof. Because S is accretive and T is strongly accretive with the strongly
accretive constant k, then for any x, y ∈ X, we have
Sx − Sy, J(x − y) ≥ 0 and T x − T y, J(x − y) ≥ kx − y2 .
Hence
(T + S)x − (T + S)y, J(x − y) = T x − T y, J(x − y) + Sx − Sy, J(x − y)
≥ kx − y2 .
Applying Proposition 5.7.10, we obtain the existence of zeros for continuous
and α-strongly accretive mappings.
Theorem 5.7.13 Let C be a nonempty closed convex subset of a Banach space
X and A : C → X a continuous and α-strongly accretive mapping such that
(i) lim inf α(r) > Ax0  for some x0 ∈ C,
r→∞

d(x − hAx, C)
(ii) lim+ = 0 for all x ∈ C.
h→0 h
Then A has a unique zero in C.
Proof. We may assume that x0 = 0. Proposition 5.7.10 implies that C ⊂
(I + A)(C). Because (I + A) is invertible, the mapping g = (I + A)−1 is
a nonexpansive self-mapping on C and the fixed points of g are zeros of A.
It suffices to show that g has a fixed point.
Let D = {x ∈ C : Ax = tx for some t < 0}. Then D is bounded. Indeed, for
x ∈ D, we have
Ax = tx for some t < 0 and Ax − A0, j ≥ α(x)x for some j ∈ J(x),
which imply that
α(x)x ≤ Axx + A0x
= tx2 + A0x.
Because t < 0, α(x) < A0. It follows that D is bounded.
272 5. Existence Theorems in Banach Spaces

Similarly, one can show that E = {y ∈ C : g(y) = λy for some λ > 1} is also
bounded.
Next, we show that (I − g)(C) is a closed set of X. For this, let {yn } be a
sequence in C such that yn − g(yn ) → u for some u ∈ X. Set xn := g(yn ). Then
yn − g(yn ) = Axn → u as n → ∞.
Because for m, n ∈ N, there exists j(xn − xm ) ∈ J(xn − xm ) such that
Axn − Axm , j(xn − xm ) ≥ α(xn − xm )xn − xm ,
this yields
α(xn − xm ) ≤ Axn − Axm .
Hence {xn } is a Cauchy sequence. Let xn → x for some x ∈ C. By the continuity
of I + A, yn → y for some y ∈ C. Thus, (I − g)(y) = u, i.e., (I − g)(C) is closed.
Now, let {tn } be a sequence in (0, 1) with tn → 1. Then tn g(yn ) = yn for
some yn ∈ C, and it follows that
yn − g(yn ) = (1 − t−1
n )yn .

Because {yn } is in E and E is bounded, yn − g(yn ) → 0 ∈ (I − g)(C). Therefore,

g has a fixed point in C.

Note that the mapping A satisfies condition (ii) of Theorem 5.7.13 if and
only if I − A is weakly inward on C. Therefore, Theorem 5.7.13 yields the
following useful existence results for strongly pseudocontractive mappings in
Banach spaces.
Corollary 5.7.14 Let C be a nonempty closed convex subset of a Banach space
X and T : C → X a weakly inward continuous α-strongly pseudocontractive
mapping. Then T has a unique fixed point in C.
Corollary 5.7.15 Let C be a nonempty closed convex subset of a Banach space
X and T : C → C a continuous α-strongly pseudocontractive mapping. Then T
has a unique fixed point in C.
We now give fundamental properties and existence results for pseudocon-
tractive mappings in Banach spaces.
Proposition 5.7.16 Let C be a nonempty subset of a Banach space X and
T : C → X a continuous pseudocontractive mapping. Let AT : C → X be
a mapping defined by AT := I + r(I − T ) for any r > 0. Then we have the
following:
(a) AT is one-one and A−1 T is nonexpansive.
−1
(b) F (T ) = F (AT ).
(c) If C is closed, then AT (C) is closed.
(d) If C is closed and convex and T is weakly inward, then the range of AT
contains C, i.e., C ⊂ AT (C).
5.7. Pseudocontractive mappings 273

Proof. (a) By pseudocontractivity of T,

x − y ≤ [I + r(I − T )]x − [I + r(I − T )]y = AT x − AT y for all x, y ∈ C,

and it follows that AT is one-one. Therefore, A−1T is nonexpansive.

(b) and (c) are obvious.
(d) Let z be a point in C. Then it suffices to show that there exists x ∈ C
such that z = AT x. Define g : C → X by g(x) = (1 + r)−1 (rT x + z). Then g is
weakly inward and continuous. Let x, y ∈ C, there exists j(x − y) ∈ J(x − y)
such that
T x − T y, j(x − y) ≤ x − y2 ,
which implies that
r
g(x) − g(y), j(x − y) = T x − T y, j(x − y)
1+r
r
≤ x − y2 .
1+r
Then g is continuous and a r/(1 + r)-strongly pseudocontractive mapping.
By Corollary 5.7.14, there exists x ∈ C with g(x) = x, i.e., z = AT (x).

Theorem 5.7.17 Let X be a uniformly convex Banach space satisfying the

Opial condition. Let C be a nonempty closed convex subset of X and T : C → X
a weakly inward continuous pseudocontractive mapping. Then I − T is demi-
closed at zero.

Proof. Let {xn } be a sequence in C with xn  z and lim xn − T xn  = 0.

n→∞
By Theorem 3.2.9, we have Za (C, {xn }) = {z}.
Let AT : C → X be a mapping defined by AT := I + r(I − T ) for any
r > 0. Then Proposition 5.7.16 (d) implies that C ⊂ AT (C) and because AT
is one-one, we conclude that g : C → C defined by g = A−1 T is nonexpansive.
Because AT (xn ) = xn + r(xn − T xn ), it follows that xn = g(xn + r(xn − T xn )).
Now

xn − g(xn ) = g(xn + r(xn − T xn )) − g(xn )

≤ rxn − T xn  → 0 as n → ∞.

Because

ra (g(z), {xn }) = lim sup xn − g(z)

n→∞
≤ lim sup g(xn ) − g(z) ≤ lim sup xn − z,
n→∞ n→∞

it follows that g(z) = z.

We now give existence results for pseudocontractive mappings.

274 5. Existence Theorems in Banach Spaces

Proposition 5.7.18 Let {xn } be a bounded sequence in a Hilbert space H and

{rn } a strictly decreasing sequence in R+ such that

rn xn − rm xn , xn − xm  ≤ 0 for all m, n ∈ N.

Then there exists x ∈ H such that xn → x.

Proof. Observe that

2rn xn − rm xn , xn − xm  = (rn + rm )xn − xm 2 + (rn − rm )(xn 2 − xm 2 )

for all m, n ∈ N. This implies that {xn } is a Cauchy sequence and hence there
exists x ∈ H such that xn → x.

Theorem 5.7.19 Let C be a nonempty closed convex bounded subset of a Hilbert

space H and T : C → H a weakly inward continuous pseudocontractive mapping.
Then T has a fixed point.

Proof. Let u be an element in C and let {tn } be a strictly increasing sequence

in (0, 1) with tn → 1. Define a mapping Tn : C → X by

Tn x = (1 − tn )u + tn T x, x ∈ C, n ∈ N.

Then for each n ∈ N, Tn is a continuous strongly pseudocontractive mapping

and also Tn is weakly inward because C is convex. By Corollary 5.7.14, there
exists exactly one point xn ∈ C such that xn = (1 − tn )u + tn T xn .
Set rn = t−1
n − 1. Then

rn xn − rm xm , xn − xm  = (rn − rm )u, xn − xm  + T xn − T xm

−(xn − xm ), xn − xm 
≤ (rn − xm )u, xn − xm .

Without loss of generality, we may assume that u = 0. By Proposition 5.7.18,

xn → x. It follows from the fact xn − T xn → 0 and continuity of T that
x = T x.
Theorem 5.7.20 Let X be a uniformly convex Banach space and C a nonempty
closed convex subset of X (with 0 ∈ C). Let T : C → X be a weakly inward
continuous pseudocontractive mapping. Then T has fixed point in C if and only
if the set E = {x ∈ C : T x = λx f or some λ > 1} is bounded.
Proof. Suppose that E is bounded. Define a mapping AT : C → X by
AT := I + r(I − T ) for any r > 0. Proposition 5.7.16 implies that C ⊂ AT (C)
and because AT is one-one, we conclude that g : C → C defined by g = A−1 T is
nonexpansive.
Now, by Theorem 5.3.5, it suffices to show that the set

D = {y ∈ C : g(y) = μy for some μ > 1}

5.7. Pseudocontractive mappings 275

is bounded. Suppose that g(y) = μy for μ > 1. Select x ∈ C such that y = AT x.

Then T x = (1 + (μ − 1)/(μr))x, i.e., x ∈ E. Because x = μy, it follows that D
is bounded.
Conversely, suppose that v is a fixed point of T and x ∈ E. Then T x = λx
for some λ > 1. By the pseudocontractivity of T ,

x − v ≤ (1 + r)(x − v) − r(λx − v) = (1 + r − rλ)x − v.

By choosing r = (λ − 1)−1 , we have x − v ≤ v. Therefore, E is

bounded.

Corollary 5.7.21 Let C be a nonempty closed convex subset of a Banach space

X and T : C → C a Lipschitzian pseudocontractive mapping. Suppose C has
the fixed point property for nonexpansive mappings. Then T has a fixed point
in C.

Theorem 5.7.22 Let C be a nonempty closed convex subset of a Banach space

X and T : C → X a weakly inward continuous pseudocontractive mapping.
Suppose C has the fixed point property for nonexpansive self-mappings. Then T
has a fixed point in C.

Proof. Note the mapping g : C → C defined by g := A−1 T is nonexpansive

by Proposition 5.7.16. Therefore, there exists v ∈ C such that v = AT v =
v + r(v − T v) from which v ∈ F (T ).

Corollary 5.7.23 Let C be a nonempty closed convex subset of a Banach space

and T : C → X a weakly inward nonexpansive mapping. Suppose C has the fixed
point property for nonexpansive self-mappings. Then T has a fixed point in C.

Finally, we discuss the structure of the set of fixed points of pseudocontrac-

tive mappings.
Theorem 5.7.24 Let C be a nonempty closed convex subset of a strictly convex
Banach space X and T : C → X a weakly inward continuous pseudocontractive
mapping. Then F (T ) is closed and convex.
Proof. Proposition 5.7.16 implies that the mapping g := A−1
T is nonexpansive
and F (T ) = F (g). Hence F (T ) is closed and convex by Corollary 5.2.29.

Bibliographic Notes and Remarks

The existence theorem for single-valued non-self contraction mappings in the
form presented here is due to Caristi [35], but the techniques for existence results
for multivalued non-self contraction mappings are based on those of Downing
and Kirk [49], Mizoguchi and Takahashi [110], Xu [166], and Yi and Zhao [170].
The properties of nonexpansive mappings presented in Section 5.2 are based
on Bruck [32] and Goebel and Kirk [59]. An extension of Theorem 5.2.7 can
276 5. Existence Theorems in Banach Spaces

be found in Baillon, Reich, and Bruck [8]. The first fixed point theorem in the
noncompact setting for nonexpansive mappings was proved by Browder [26] in
Hilbert space in 1965. Theorem 5.2.16 was proved independently by Browder
[27] and Göhde [61] in 1965. Theorem 5.2.18 was proved in Kirk [88] in the same
year in a slightly more general situation. Theorem 5.2.23 and 5.2.25 are due to
Singh and Watson [149]. Theorem 5.2.26 is based on the interesting result of
Caristi [35].
Different techniques for solving operator equation (5.30) presented in Sec-
tion 5.3 are based on the results of Downing and Kirk [49], Lami Dozo [50],
Morales [112], and Yi and Zhao [170]. Theorem 5.3.6 can also be found in Reich
[124]. Theorems 5.4.7 and 5.5.3 were proved by Goebel and Kirk in [57] and
[58], respectively. Other important results presented in Sections 5.4∼5.5 can be
found in Benavides and Xu [17], Casini and Maluta [36], Gornicki [62], Lim and
Xu [98], and Xu [164].
Section 5.6 contains some new results for non-Lipschitzian mappings. Corol-
lary 5.6.7 is the main result of Benavides, Acedo, and Xu [16].
Example 5.7.1 was constructed in Chidume and Mutangadura [39].
An important result Proposition 5.7.10 is due to Martin [105]. Other results pre-
sented in Section 5.7 are discussed in Goebel and Reich [60], Khamsi and Kirk
[85], Kirk and Schoneberg [90], Molales [113], and Reinermann and Schoneberg
[134].
The approximate fixed point property plays an important role in the fixed
point theory of nonexpansive, asymptotically nonexpansive, and pseudocontrac-
tive mappings. This property is also described in Matoušková and Reich [109],
Reich [129], and Shafrir [143].

Exercises

5.1 Let C be a nonempty closed subset of a Banach space X and T : C → X

a continuous mapping with the property that for each x ∈ C, there is
an αx , 0 < αx ≤ 1 such that (1 − αx )x + αx T x ∈ C. Let x1 ∈ C and
inductively for n ∈ N define xn+1 ∈ C by xn+1 = (1 − αn )xn + αn T xn ,
where αn ∈ (0, 1] is chosen so that xn+1 ∈ C. Show that
∞
(a) if z = lim xn exists and n=1 αn = ∞, then z is a fixed point of T ,
n→∞
(b) if T is a contraction mapping, then z = lim xn exists.
n→∞

5.2 Let X be a Banach space and T : X → X a mapping satisfies the condition:

T x − T y ≤ ϕ(x − y) for all x, y ∈ X,

where ϕ : [0, ∞) → [0, ∞) is a continuous function and ϕ(t) < t for t > 0.
Show that I − T is bijective and (I − T )−1 is continuous on X.
5.7. Pseudocontractive mappings 277

5.3 Let C be a nonempty closed convex subset of a Banach space X and T :

C → X a weakly inward contraction mapping. Show that T has a unique
fixed point in C.
5.4 Let C be a nonempty closed convex subset of a Banach space X and T :
C → C a nonexpansive mapping. If A = I − T , show that R(I + λA) ⊃ C
for every λ > 0.
5.5 Let C be a nonempty subset of a normed space X. A function α of C ×C →
[0, ∞) is symmetric if α(x, y) = α(y, x) for all x, y ∈ C. Let T : C → C
be a mapping. T is said to be generalized nonexpansive if there exist
symmetric functions αi , i = 1, 2, 3, · · · , 5 of C × C into [0, ∞) such that
 5 
sup αi (x, y) : x, y ∈ C ≤ 1 and for all x, y in X,
i=1

T x−T y ≤ a1 x−y+a2 x−T y+a3 y−T x+a4 x−T x+a5 y−T y,

where ai = αi (x, y).

If C is a nonempty convex subset of a uniformly convex Banach space
X, and T : C → C is a generalized nonexpansive mapping with F (T ) = ∅,
show that for each t ∈ (0, 1), the mapping defined by

Tt x = (1 − t)x + tT x, x ∈ C

is asymptotically regular.
5.6 Let X be a strictly convex Banach space and C a weakly compact convex
subset of X that has normal structure. Let S = {T1 , T2 , · · · , Tn } be a
finite commuting family of nonexpansive mappings of C into itself. Show
that ∩ni=1 F (Ti ) = ∅.
5.7 Let X be a Banach space, C a nonempty closed bounded subset of X
that is star-shaped with respect to 0, and T : C → C an asymptotically
nonexpansive mapping with sequence {kn } and uniformly asymptotically
regular (i.e., for each ε > 0, there exists n0 ∈ N such that T n x−T n+1 x ≤
ε for all n ≥ n0 and all x ∈ C). Let {λn } be a sequence in (0, 1) such that
lim λn = 1. Show that
n→∞

(a) for each n ∈ N, there exists exactly one xn ∈ C such that xn =

(λn /kn )T n xn ,
(b) xn − T xn → 0 as n → ∞.
5.8 Let C be a nonempty subset of a normed space X and T : C → C a
mapping. T is said to be weakly asymptotically semicontractive if there
exist a mapping S : C × C → C and a sequence {kn } ⊂ [1, ∞) such
that T x = S(x, x) for all x ∈ C while for each fixed x ∈ C, S(·, x) is
asymptotically nonexpansive with sequence {kn } and for fixed x ∈ C and
fixed n ∈ N, the mapping y → S(·, y)n x is compact on C.
278 5. Existence Theorems in Banach Spaces

If X is a reflexive Banach space possessing a weakly continuous duality

mapping, C is a nonempty closed convex bounded subset of X, and T :
C → C is weakly asymptotically semicontractive with data (S, {kn }) and
satisfies the condition:
for each ε > 0, there exists an n0 ∈ N such that for all n ≥ n0 and all
z ∈ C,
S(·, z)n+1 (z) − S(·, z)n (z) < ε,
show that

(a) inf {x − T x} = 0,

x∈C

(b) if (I − T )(C) is closed, it follows that F (T ) = ∅.

5.9 Let X be a Banach space with GGLD and the Opial condition, C a nonempty
weakly compact convex subset of X, and T : C → C an asymptotically
nonexpansive mapping. Show that I − T is demiclosed at 0.

5.10 Let X = R2 with the 1 norm, i.e., (x, y) = |x| + |y|. Let A and B be
operators in X defined by

D(A) = D(B) = {(0, 0), (0, 1)}

and

A(0, 0) = (0, 0) = B(0, 1), A(0, 1) = (1, 1/2), B(0, 0) = (1, 1).

Show that both A and B are accretive, but A + B is not.

Chapter 6

Approximation of Fixed
Points

The purpose of this chapter is to develop iterative techniques for approximation

of fixed points of nonlinear mappings by using the Picard, Mann, and Ishikawa
iteration processes.

6.1 Basic properties and lemmas

In this section, we develop preliminary results for approximation of fixed points
of nonlinear mappings.

Proposition 6.1.1 Let X be a Banach space satisfying the Opial condition, C

a nonempty weakly compact subset of X, and T : C → C a mapping such that
(i) F (T ) = ∅,
(ii) I − T is demiclosed at zero.
Let {xn } be a sequence in C satisfying the following properties:
(D1 ) lim xn − p exists for all p ∈ F (T );
n→∞
(D2 ) {xn } is an AFPS, i.e., lim xn − T xn  = 0.
n→∞
Then {xn } converges weakly to a fixed point of T .

Proof. Because C is weakly compact, it follows that {xn } has a weakly conver-
gent subsequence {xnj }. Suppose {xnj } converges weakly to p. Because {xnj }
⊂ C and C is weakly closed, then p ∈ C. From (D2 ), lim ||xn − T xn || = 0 and
n→∞
because I − T is demiclosed at zero, we have (I − T )p = 0, so that p ∈ F (T ).
To complete the proof, we show that {xn } converges weakly to a fixed point of
T ; it suffices to show that ωw ({xn }) consists of exactly one point, namely, p.
Suppose there exists another subsequence {xnk } of {xn } that converges weakly
to some q = p. As in the case of p, we must have q ∈ C and q ∈ F (T ). It follows

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 6,

c Springer Science+Business Media, LLC, 2009
280 6. Approximation of Fixed Points

from (D1 ) that lim ||xn − p|| and lim ||xn − q|| exist. Because X satisfies the
n→∞ n→∞
Opial condition, we have

lim ||xn − p|| = lim ||xnj − p|| < lim ||xnj − q|| = lim ||xn − q||,
n→∞ j→∞ j→∞ n→∞

lim ||xn − q|| = lim ||xnk − q|| < lim ||xnk − p|| = lim ||xn − p||,
n→∞ k→∞ k→∞ n→∞

a contradiction. Hence p = q and {xn } converges weakly to p.

Proposition 6.1.2 Let C be a nonempty closed convex subset of a Banach

space X and T : C → C a mapping such that
(i) F (T ) = ∅,
(ii) I − T is demiclosed at zero.
Let {xn } be a sequence in C that satisfies properties (D1 ) and (D2 ). Suppose
{xn } holds one of the following conditions:
(a) X is uniformly convex with Fŕechet differentiable norm and

lim xn , J(p − q) exists for all p, q ∈ F (T ). (6.1)

n→∞

(b) X is reflexive, X ∗ has the Kadec-Klee property, and lim txn +(1−t)p−q
n→∞
exists for all t ∈ [0, 1] and for some p, q ∈ ωw ({xn }).
Then {xn } converges weakly to a fixed point of T .

Proof. We show that ωw ({xn }) has exactly one point. Let u, v ∈ ωw ({xn })
with u = v. Then for some subsequences {xni } and {xnj } of {xn }, we have
xni  u and xnj  v. By (D2 ), lim xn − T xn  = 0, which implies by the
n→∞
demiclosedness of I − T at zero that u, v ∈ ωw ({xn }) ⊂ F (T ).
(a) From (6.1), we have

u, J(p − q) = d (say), and v, J(p − q) = d;

so

u − v, J(p − q) = 0 for all p, q ∈ F (T ). (6.2)

From (6.2) we obtain that

u − v2 = u − v, J(u − v) = 0,

a contradiction. Hence ωw ({xn }) is singleton. Therefore, {xn } converges weakly

to a fixed point of T .
(b) By assumption, lim txn + (1 − t)u − v exists. Corollary 2.4.17 guaran-
n→∞
tees that u = v. Hence ωw ({xn }) is singleton. Therefore, {xn } converges weakly
to a fixed point of T .
6.1. Basic properties and lemmas 281

Proposition 6.1.3 Let X be a reflexive Banach space, C a nonempty closed

convex subset of X, {xn } a bounded sequence in C, and T : C → C a nonex-
pansive mapping. Suppose {xn } satisfies one of the following conditions:
(D2 ) lim xn − T xn  = 0.
n→∞
(D3 ) lim xn+1 − T xn  = 0.
n→∞
Then the set M defined by (2.32) is a nonempty closed convex bounded and
T -invariant subset of C.

Proof. Define a real-valued function ϕ on C by ϕ(z) = LIMn xn −z2 for each
z ∈ C. Theorem 2.9.11 implies that M is a nonempty closed convex bounded
set. If lim xn − T xn  = 0, then for y ∈ M
n→∞

ϕ(T y) = LIMn xn − T y2 ≤ LIMn T xn − T y2

≤ LIMn xn − y2 = ϕ(y).

Hence T y ∈ M , i.e., M is T -invariant.

Suppose now that lim xn+1 − T xn  = 0. Observe that for y ∈ M
n→∞

ϕ(T y) = LIMn xn − T y2

= LIMn xn+1 − T y2 (as LIMn (an ) = LIMn (an+1 ))
≤ LIMn T xn − T y2
≤ LIMn xn − y2 = ϕ(y).

Hence M is T -invariant.

We now give some useful lemmas:

Lemma 6.1.4 Let {αn } be a sequence of nonnegative numbers such that

∞ ∞
n=1 αn =∞. Suppose that βn > 0 for all n ∈ N and n=1 αn βn < ∞.
Then lim inf βn = 0.
n→∞

Proof. Suppose, for contradiction, that lim inf βn = δ for some δ > 0. Then
n→∞
there exists n0 ∈ such that βn > δ/2 for all n ≥ n0 . Then
 δ 
αn βn > αn = ∞,
2
n≥n0 n≥n0

a contradiction. Therefore, δ = 0.

Lemma 6.1.5 Let {δn } be a sequence of nonnegative numbers satisfying:

δn+1 ≤ βn δn + γn for all n ∈ N,

282 6. Approximation of Fixed Points

where {βn } and {γn } are sequences of nonnegative numbers such that
∞

{βn } ⊆ [1, ∞), (βn − 1) < ∞, (6.3)
n=1
∞

γn < ∞. (6.4)
n=1

Then lim δn exists. If lim inf δn = 0, then lim δn = 0.

n→∞ n→∞ n→∞

Proof. For m, n ∈ N, we have

δn+m+1 ≤ βn+m δn+m + γn+m

≤ βn+m (δn+m + γn+m )
≤ βn+m (βn+m−1 (δn+m−1 + γn+m−1 ) + γn+m )
···
 n+m  
n+m 
≤ βi δn + γi .
i=n i=n

Hence
 ∞  ∞
 
lim sup δm ≤ βi δn + γi . (6.5)
m→∞
i=n i=n
 ∞  ∞

By the conditions (6.3) and (6.4), we have lim βi = 1 and lim γi =
n→∞ n→∞
i=n i=n
0. It follows from (6.5) that lim sup δn ≤ lim inf δn . Therefore, lim δn exists.
n→∞ n→∞ n→∞

Suppose lim inf δn = 0. Then lim δn = lim inf δn = 0.

n→∞ n→∞ n→∞

Lemma 6.1.6 Let {an }, {bn }, and {tn } be three sequences of nonnegative
numbers such that

an+1 ≤ (1 − tn )an + bn tn , n ∈ N, (6.6)

∞
where tn ∈ [0, 1], n=1 tn = ∞ and lim bn = 0. Then lim an = 0.
n→∞ n→∞

Proof. By (6.6),
n n 
 n 
0 ≤ an+1 ≤ (1 − ti )ak + ti (1 − tj ) bi . (6.7)
i=k i=k j=i+1

Observe that

n n n  n 
ti (1 − tj ) ≤ 1 for all n, k ∈ N and (1 − ti ) ≤ exp − ti → 0.
i=k j=i+1 i=k i=k
6.1. Basic properties and lemmas 283

Given ε > 0, pick k such that bi ≤ ε for all i ≥ k, from (6.7), we have

0 ≤ lim inf an ≤ lim sup an ≤ ε.

n→∞ n→∞

Letting ε → 0, we obtain lim an = 0.

n→∞

Lemma 6.1.7 Let {xn } and {yn } be two sequences in a uniformly convex
Banach space X such that

xn+1 = (1 − αn )xn + αn yn and yn  ≤ xn , n ∈ N,

∞
where {αn } is a sequence of nonnegative numbers in [0, 1] with n=1 min{αn ,
1 − αn } = ∞. Then 0 ∈ {xn − yn }.

Proof. Suppose, for contradiction, that xn − yn  ≥ ε > 0 for all n ∈ N.

Observe that
xn+1  ≤ xn  ≤ · · · ≤ x1  for all n ∈ N.
Then

xn+1  = (1 − αn )xn + αn yn 

  
ε
≤ xn  1 − 2 min{αn , 1 − αn }δX .
x1 

Inductively, we have
n−1   
ε
xn  ≤ 1 − 2 min{αi , 1 − αi }δX x1  for all n > 1.
i=1
x1 
∞
Because i=1 min{αi , 1 − αi } = ∞, it follows that
   n−1
  
ε
xn  ≤ x1 exp − 2δX min αi , 1 − αi → 0 as n → ∞
x1  i=1

and hence lim xn  = lim yn  = 0. This is a contradiction.

n→∞ n→∞

Lemma 6.1.8 Let C be a nonempty closed convex bounded subset of a Banach

space X and {Tn } a sequence of Lipschitzian self-mappings of C such that
∞
(i) Ln (≥ 1) is the Lipschitz constant of Tn with n=1 (Ln − 1) < ∞,
(ii) F := ∩∞n=1 F (Tn ), the set of common fixed points of {Tn } is nonempty.

For a given x1 ∈ C, define the sequence {xn } by

xn+1 = Tn xn , n ∈ N. (6.8)
284 6. Approximation of Fixed Points

Then the following hold:

(a) lim xn − p exists for all p ∈ F.
n→∞

(b) If X is uniformly convex, then lim txn + (1 − t)f1 − f2  exists for all
n→∞
f1 , f2 ∈ F, and t ∈ [0, 1].
(c) If X is uniformly convex with Fréchet differentiable norm, then lim
n→∞
xn − p, J(p − q) exists for all p, q ∈ F.
Proof. (a) Let p ∈ F. Then from (6.8), we have
xn+1 − p = Tn xn − p ≤ Ln xn − p, n ∈ N,
∞
and it follows from Lemma 6.1.5 that lim xn − p exists, because n=1
n→∞
(Ln − 1) < ∞.
(b) Let p, q ∈ F. Set
an (t) : = txn + (1 − t)p − q,
Sn,m : = Tn+m−1 Tn+m−2 · · · Tn ,
bn,m = Sn,m (txn + (1 − t)p) − (txn+m + (1 − t)p).
We show lim an (t) exists for all t ∈ [0, 1]. Note that lim an (0) and lim an (1)
n→∞ n→∞ n→∞
exist, so it remains to show that lim an (t) exists for all t ∈ (0, 1).
n→∞
Observe that for x, y ∈ C
xn+m = Sn,m xn ;
 n+m−1   ∞ 
Sn,m x − Sn,m y ≤ Li x − y ≤ Li x − y;
i=n i=n

and
an+m (t) = txn+m + (1 − t)p − q
≤ txn+m + (1 − t)p − Sn,m (txn + (1 − t)p)
+Sn,m (txn + (1 − t)p) − q
 n+m−1 
≤ bn,m + Li txn + (1 − t)p − q
i=n
 ∞ 
≤ bn,m + Li an (t) = bn,m + Hn an (t), (6.9)
i=n
∞
where Hn = i=n Li .
By Theorem 5.2.31, there exists a strictly increasing continuous function
γ : R+ → R+ with γ(0) = 0 such that
γ(S(tx + (1 − t)y) − (tSx + (1 − t)Sy)) ≤ x − y − Sx − Sy
6.1. Basic properties and lemmas 285

for all nonexpansive S : C → X and t ∈ [0, 1]. It then follows that

γ(Hn−1 bn,m ) = γ(Hn−1 Sn,m (txn + (1 − t)p) − (tSn,m xn + (1 − t)p))

≤ xn − p − Hn−1 xn+m − p,

which implies that

bn,m ≤ Hn γ −1 (xn − p − Hn−1 xn+m − p). (6.10)

Because lim xn − p exists by part (a) and lim Hn = 1, then from (6.10)
n→∞ n→∞
we obtain that lim bn,m = 0. Hence from (6.9), we have
n,m→∞

lim sup am (t) ≤ lim bn,m + lim inf Hn an (t) = lim inf an (t).
m→∞ n,m→∞ n→∞ n→∞

Therefore, lim an (t) exists for all t ∈ [0, 1].

n→∞
(c) Because the norm of X is Fréchet differentiable,

1 1 1
x2 + h, Jx ≤ x + h2 ≤ x2 + h, Jx + b(h) (6.11)
2 2 2
for all bounded x, h ∈ X, where J is normalized duality mapping and b is the
b(t)
function defined on R+ such that lim = 0.
t→0 t
Taking x = p − q and h = t(xn − p) in (6.11), we get

1 1 2
p − q2 + txn − p, J(p − q) ≤ a (t)
2 2 n
1
≤ p − q2 + txn − p, J(p − q) + b(txn − p).
2

Because for each t ∈ (0, 1), lim an (t) exists, it follows that
n→∞

1
p − q2 + t lim supxn − p, J(p − q)
2 n→∞
1
≤ lim a2 (t)
2 n→∞ n
1
≤ p − q2 + t lim inf xn − p, J(p − q) + o(t).
2 n→∞

This yields

o(t)
lim supxn − p, J(p − q) ≤ lim inf xn − p, J(p − q) + .
n→∞ n→∞ t

On letting t → 0+ , we obtain that lim xn − p, J(p − q) exists.

n→∞
286 6. Approximation of Fixed Points

6.2 Convergence of successive iterates

In this section, we develop techniques for convergence of {T n x} to fixed points
of nonlinear operators in Banach spaces. Recall for contraction mapping T , the
sequence of iterates {T n x} converges strongly in Banach spaces. The following
example shows that even if a nonexpansive mapping T has a unique fixed point,
then {T n x} need not converge to it.

Example 6.2.1 Let C = BH be the unit ball of theHilbert space H = 2 and

∞
{an } a sequence of real numbers in [0, 1] such that n=1 an > 0. Consider the
linear mapping T : C → C defined by

T (x1 , x2 , · · · ) = (0, a1 x1 , a2 x2 , · · · ).

The origin is the only fixed point of T . It is easy to see that the sequence
of iterates {T n e} with e = (1, 0, 0, · · · ) converges weakly to 0, but it does not
converge strongly to 0.

We first study strong convergence of {T n x} for nonexpansive mappings in a

Banach space.

Theorem 6.2.2 (Browder and Petryshyn’s theorem) – Let X be a

Banach space and T an asymptotically regular nonexpansive self-mapping of
X. Suppose that T has a fixed point and that I − T maps closed bounded subsets
of X into closed subsets of X. Then for each x ∈ X, {T n x} converges strongly
to an element of F (T ).

Proof. Let p ∈ F (T ). Then {T n x − p} is a nonincreasing sequence.

It suffices therefore to show that there exists a subsequence of {T n x} that con-
verges strongly to a fixed point of T . Let S be the strong closure of sequence
{T n x}. By the asymptotic regularity of T ,

(I − T )T n x → 0 as n → ∞.

Hence 0 lies in the strong closure of (I −T )(S) and because the latter is closed by
hypothesis (as S is closed and bounded), 0 lies in (I − T )(S). Hence there exists
a subsequence {T ni x} such that T ni x → v ∈ S such that (I − T )v = 0. Hence
T n x → v.

We now turn our attention to the study of weak convergence of the iterates
of nonlinear mappings.

Theorem 6.2.3 Let X be a Banach space satisfying the Opial condition, C a

nonempty weakly compact convex subset of X, and T : C → C a nonexpansive
mapping with F (T ) = ∅. If for x ∈ C, T n x − T n+1 x → 0 as n → ∞, then
{T n x} converges weakly to an element of F (T ).
6.2. Convergence of successive iterates 287

Proof. For x ∈ C, define a sequence {xn } in C by xn = T n x, n ∈ N0 . Then

for v ∈ F (T ),
xn+1 − v = T n+1 x − v ≤ T n x − v ≤ xn − v for all n ∈ N0 ,
and it follows that lim xn − v exists. By assumption
n→∞

T n x − T n+1 x = xn − T xn → 0 as n → ∞.
Thus, {T n x} is an AFPS of T . By Theorem 5.2.9, I − T is demiclosed at zero.
Applying Proposition 6.1.1, we conclude that {T n x} converges weakly to some
z ∈ F (T ).

We now study weak convergence of iterates of mappings that are more gen-
eral than nonexpansive mappings. We begin with the following proposition:

Proposition 6.2.4 Let C be a nonempty closed convex subset of a uniformly

convex Banach space X satisfying the Opial condition and T : C → C a nearly
asymptotically nonexpansive mapping. Suppose that x0 is the asymptotic cen-
ter of the bounded sequence {T n x} for some x ∈ C. If the weak limit z of a
subsequence {T ni x} of {T n x} is a fixed point of T , then x0 coincides with z.
Proof. It is obvious that ra (C, {T n x}) ≥ ra (C, {T ni x}). Because T ni x  z, it
follows from Theorem 3.2.9 that Za (C, {T ni x}) = {z} and so, for any ε > 0, we
can choose an integer io such that
ε
z − T ni0 x ≤ ra (C, {T ni x}) + .
2
Because z is a fixed point of T and T is nearly asymptotically nonexpansive, we
can choose an integer j0 such that for all j ≥ j0
z − T ni0 +j x ≤ η(T j )(z − T ni0 x + aj )
ε
≤ η(T j )(ra (C, {T ni x}) + + aj )
2
≤ η(T j )(ra (C, {T n x}) + ε + aj ),

and it follows that

lim sup z − T n x = ra (C, {T n x}).
n→∞

By the uniqueness of asymptotic center, we have z = x0 .

Theorem 6.2.5 Let X be a uniformly convex Banach space satisfying the Opial
condition and C a nonempty closed convex (but not necessarily bounded) subset
of X. Let T : C → C be a demicontinuous nearly asymptotically nonexpansive
mapping and x ∈ C. Then {T n x} converges weakly to a fixed point of T if and
only if T is weakly asymptotically regular at x.
288 6. Approximation of Fixed Points

Proof. Assume that T n x  z as n → ∞. We show that z ∈ F (T ). By Theo-

rem 3.2.9, Za (C, {T n x}) = {z}. As in proof of Theorem 5.6.5, we have z ∈ F (T ).
Because T n x  z as n → ∞, it follows that T n+1 x − T n x  0 as n → ∞.
Conversely, suppose that T n+1 x − T n x  0 as n → ∞. First, we show that
ωw ({T n x}) ⊆ F (T ). Let y ∈ ωw ({T n x}). Then we have a subsequence {T ni x}
of {T n x} such that T ni x  y. By the weak asymptotic regularity of T , we have

T ni +m x  y as i → ∞ for m = 0, 1, · · · .

By Theorem 3.2.9, we have

Za (C, {T ni +m x}) = {y} for m = 0, 1, 2, · · · .

Let {ys } be a sequence in C defined by ys = T s y for s ∈ N. For m, s ∈ N with

m > s, we have

ys − T ni +m x = T s y − T s (T ni +m−s x)

≤ η(T s )(y − T ni +m−s x + as ),

which implies that

ra (ys , {T m x}) ≤ η(T s )(ra (y, {T m x}) + as ).

By Theorem 3.1.8, T s y → y as s → ∞. By the demicontinuity of T , we

obtain from Proposition 5.6.1 that T y = y. Thus, ωw ({T n x}) ⊆ F (T ) is veri-
fied. To complete the proof, we show that ωw ({T n x}) is singleton. Let u, v ∈
ωw ({T n x}). Then we have two subsequences {T nj x} and {T nk x} of {T n x} such
that T nj x  u and T nk x  v. Then u, v ∈ F (T ). Let Za (C, {T n x}) = {z}.
It follows from Proposition 6.2.4 that u = v = z. This proves that ωw ({T n x}) =
{z}.

Corollary 6.2.6 Let X be a uniformly convex Banach space satisfying the

Opial condition and C a nonempty closed convex (but not necessarily bounded)
subset of X. If T : C → C is an asymptotically nonexpansive mapping and
x ∈ C, then {T n x} converges weakly to a fixed point of T if and only if T is
weakly asymptotically regular at x.

Proof. Because every asymptotically nonexpansive mapping is nearly asymp-

totically nonexpansive mapping, the result follows from Theorem 6.2.5.

6.3 Mann iteration process

We have already seen in Section 6.2 that asymptotic regularity of nonlinear map-
pings T is required even for weak convergence of {T n x}. We drop asymptotic
regularity of nonlinear mappings when using the Mann iteration process.
6.3. Mann iteration process 289

Definition 6.3.1 Let C be a nonempty convex subset of a linear space X and

T : C → C a mapping. Let A = [ai,j ] be an infinite real matrix satisfying:
(A1 ) A is a lower matrix with nonnegative entries (an,i ≥ 0 for all n, i ∈ N
and an,i = 0 for all i > n), i.e.,
⎛ ⎞
a1,1 0 0 ··· 0 0
⎜ a2,1 a2,2 0 ··· 0 0 ⎟
⎜ ⎟
⎜ a3,1 a3,2 a3,3 · · · 0 0 ⎟
A=⎜ ⎜ ···
⎟,
⎜ ··· ··· ··· ··· ··· ⎟ ⎟
⎝ an,1 an,2 an,3 · · · 0 0 ⎠
··· ··· ··· ··· ··· ···
n
(A2 ) the sum of each row is 1, i.e., i=1 an,i = 1 for all n ∈ N,
(A3 ) lim an,i = 0 for all i ∈ N.
n→∞

Define a sequence {xn } in C by x1 ∈ C and

n
xn+1 = T ( an,i xi ), n ∈ N. (6.12)
i=1

Then the sequence {xn } defined by (6.12) is called the Mann iteration.

Such an iteration process is called the Mann iteration process.1

Example 6.3.2 Let Λ define the Cesaro matrix, i.e.,
⎛ ⎞
1 0 0 0 ··· ··· ···
⎜ 1 1
0 0 ··· ··· ··· ⎟
⎜ 21 2 ⎟
⎜ 1 1
0 ··· ··· ··· ⎟
Λ=⎜ 3
⎜ ··· ···
3 3 ⎟.
⎟
⎜ 1 ··· ··· ··· ··· ··· ⎟
⎝ 1 1
··· 1
0 ··· ⎠
n n n n
··· ··· ··· ··· ··· ··· ···

It is easy to see that Λ satisfies all the hypothesis related to the matrix A. Then
the sequence {xn } in C defined by (6.12) reduces to

1
n
xn+1 = T ( xi ), n ∈ N.
n i=1

In Definition 6.3.1, the matrix A is very general. The most useful Mann
iteration process can be obtained by choosing the matrix A as follows:

an,i = (1 − an,n )an−1,i , i = 1, 2, · · · , n and n = 2, 3, · · · , (6.13)

and
either an,n = 1 or an,n < 1 for all n ∈ N. (6.14)
1 It was introduced by W.R. Mann in 1953.
290 6. Approximation of Fixed Points

The entries of matrix A satisfying conditions (6.13) and (6.14) can be con-
structed by choosing a sequence of nonnegative numbers in [0, 1] as below:
Choose a sequence {αn } of nonnegative numbers satisfying the conditions:
∞

0 ≤ αn < 1 for all n ∈ N and αn = ∞ (6.15)
n=1

and then we can define the entries of A by

⎧
⎪
⎪ a1,1 = 1, a1,i = 0 for i > 1;
⎨
an+1,n+1 = αnfor n ∈ N;
n
⎪ an+1,i = ai,i k=i (1 − αk ) for i = 1, 2, · · · , n;
⎪
⎩
an+1,i = 0 for i > n + 1, n ∈ N.
More precisely, we now define the Mann iteration process, which will be used
to approximate fixed points of nonlinear mappings.
Definition 6.3.3 Let C be a nonempty convex subset of a linear space X and
T : C → C a mapping. Let {αn } be a sequence of nonnegative numbers satisfying
(6.15). Define a sequence {xn } in C by

x1 ∈ C;
(6.16)
xn+1 = M (xn , αn , T ), n ∈ N;

where M (xn , αn , T ) = (1 − αn )xn + αn T xn . Then sequence {xn } is called the

(normal) Mann iteration.
Using convexity structure defined in Section 4.3, we now define the Mann
iteration in a metric space.

Definition 6.3.4 Let C be a nonempty convex subset of a convex metric space

X and T : C → C a mapping. Let {αn } be a sequence satisfying (6.15). Define
a sequence {xn } in C by

x1 ∈ C;
xn+1 = W (T xn , xn ; αn ), n ∈ N.

Then {xn } is called the Mann iteration.

First, we study a convergent Mann iteration for arbitrary continuous map-

pings.
Theorem 6.3.5 Let C be a nonempty closed convex subset of a Banach space
X and T : C → C a continuous mapping. If the Mann iteration {xn } defined
by (6.16) converges strongly to a point p ∈ C, then p is a fixed point of T .
Proof. Let lim xn = p. Suppose, for contradiction, that p = T p. Set εn :=
n→∞
xn − T xn − (p − T p). Because lim xn = p and T is continuous, it follows that
n→∞

lim εn = lim [(xn − T xn ) − (p − T p)] = 0.

n→∞ n→∞
6.3. Mann iteration process 291

Because p − T p > 0, there exists n0 ∈ N such that εn  < p − T p/3 and
xn − xm  < p − T p/3 for all n, m ≥ n0 . Let N be any positive integer such
n0 +N
that i=n 0
αi ≥ 1. Because xi+1 − xi = αi (T xi − xi ), it follows that
 n0 +N   n0 +N 
     
xn0 − xn0 +N +1  = 
 (xi − x )
i+1  = 
 αi (p − T p + ε )
i 
i=n0 i=n0
 n   n 
 0 +N   0 +N 
≥ 
 αi (p − T p) −
 
 α ε 
i i
i=n0 i=n0
n
0 +N  
p − T p
≥ αi p − T p −
i=n0
3
2p − T p
≥ .
3

The contradiction proves the result.

We have shown that if the Mann iteration is convergent to v for a continuous

mapping T , then v is a fixed point of T . But if T is not continuous, then there
is no guarantee that, even if the Mann iteration converges strongly to z, then z
will be a fixed point of T . Let us give an example of a discontinuous mapping.
Example 6.3.6 Let X = C = [0, 1] and T : C → C a mapping defined by
⎧
⎨ 0, if x = 0,
Tx = 1, if 0 < x < 1,
⎩
0, if x = 1.

Then T 0 = 0 and the Mann iteration {xn } defined by (6.16) with x1 ∈ (0, 1)
and αn = 1/n, n ∈ N converges to 1, which is not a fixed point of T .
The following results are very useful for approximation of fixed points of
nonexpansive type mappings.
Proposition 6.3.7 Let C be a nonempty convex subset of a normed space X
and T : C → C a mapping with a fixed point p in C such that

T x − p ≤ x − p for all x ∈ C.

Then for the Mann iteration {xn } defined by (6.16) with {αn } in [0, 1],
lim xn − p exists.
n→∞

Proof. Because

xn+1 − p ≤ (1 − αn )xn − p + αn T xn − p ≤ xn − p for all n ∈ N,

it follows that lim xn − p exists.

n→∞
292 6. Approximation of Fixed Points

Proposition 6.3.8 Let C be a nonempty convex subset of a uniformly convex

Banach space X and T : C → C a mapping with F (T ) = ∅ satisfying the
condition:
T x − p ≤ x − p for all x ∈ C and p ∈ F (T ).
∞
Define a sequence {xn } in C by (6.16) with the restriction that n=1 min{αn ,
1 − αn } = ∞. Then lim inf xn − T xn  = 0.
n→∞

Proof. Proposition 6.3.7 implies that lim xn − p exists for p ∈ F (T ).

n→∞
Observe that
T xn − p ≤ xn − p for all n ∈ N
and
xn+1 − p = (1 − αn )xn + αn (T xn − p) for all n ∈ N.

Applying Lemma 6.1.7, we obtain that lim inf xn − T xn  = 0.

n→∞

6.4 Nonexpansive and quasi-nonexpansive

mappings
We begin with a basic result on approximation of fixed points of nonexpansive
mappings in a uniformly convex Banach space with compact setting.

Theorem 6.4.1 (Krasnoselski) – Let C be a nonempty closed convex bounded

subset of a uniformly convex Banach space X and T a nonexpansive mapping of
C into a compact subset of C. Let x1 ∈ C be an arbitrary point in C. Then the
sequence {xn } defined by

1 1
(K) xn+1 = (xn + T xn ) = M (xn , , T ), n∈N
2 2
converges strongly to a fixed point of T in C.

Proof. Note F (T ) = ∅ by Schauder’s theorem. Let p ∈ F (T ). Then lim

n→∞
xn − p exists by Proposition 6.3.7. Proposition 6.3.8 implies that

lim inf xn − T xn  = 0. (6.17)

n→∞

Note
1
xn+1 − T xn+1  ≤ (xn − T xn+1  + T xn − T xn+1 )
2
1
≤ (xn − xn+1  + xn+1 − T xn+1  + xn − xn+1 ),
2
which gives
xn+1 − T xn+1  ≤ xn − T xn  for all n ∈ N.
6.4. Nonexpansive and quasi-nonexpansive mappings 293

This means that lim xn − T xn  exists. Using (6.17), we obtain

n→∞

xn − T xn  → 0 as n → ∞. (6.18)

Because {T xn } is in a compact set, there exists a subsequence {T xnk } of

{T xn } such that T xnk → v ∈ C. Hence from (6.18), we have xnk → v. Because
T is continuous, v is a fixed point of T . Note lim xn − v = lim xnk − v
n→∞ k→∞
exists. Therefore, {xn } converges strongly to a fixed point of T in C.

The following example shows that without the asymptotic regularity condi-
tion, the Picard iteration cannot be used to approximate fixed points of non-
expansive mapping, but the iteration procedure given by (K) can be used to
locate fixed point of the same mapping.
Example 6.4.2 Let X = C, which has the usual absolute value metric for
complex numbers, C = {z ∈ C : |z| ≤ 1}, and T : C → C a mapping defined by
√
T z = iz, z ∈ C, where i = −1.

It is easy to see that T is nonexpansive mapping with a fixed point 0 ∈ C.

Now, let z0 = 0 be an arbitrary point in C. Then the Picard iteration of T
is given by
zn+1 = T zn = in+1 z0 , n = 0, 1, 2, · · · .
Hence

|T n z0 − T n+1 z0 | = |in z0 − in+1 z0 |

√
= |in | · |1 − i| · |z0 | = 2|z0 |  0 as n → ∞.

Note T is not asymptotically regular at z0 , and {T n z0 = in z0 } does not converge

to zero.
However, from (K) we have
 n+1
1 1+i 1+i
zn+1 = (zn + T zn ) = zn = z0 .
2 2 2
Because
∞
 ∞ n
1+i 1+i
|zn − zn+1 | = 1− |z0 |
n=0 n=0
2 2
∞   n+1
1
= √ |z0 |
n=0
2
√
= ( 2 + 1)|z0 | < ∞,

it follows that {zn } is Cauchy sequence. Let lim zn = p. Because zn − T zn →

n→∞
0 as n → ∞, it follows that p = 0. Therefore, {zn } converges to the fixed
point 0.
294 6. Approximation of Fixed Points

The following result shows that the sequence {xn } defined by

xn+1 = M (xn , αn , T ), n∈N

converges strongly to a fixed point of nonexpansive mapping T without the

assumption of convexity of domain.

Theorem 6.4.3 (Ishikawa) – Let C be a nonempty closed subset of a Banach

space X and let T be a nonexpansive mapping from C into a compact subset of
X. Suppose there exist x1 ∈ C and a sequence {αn } of real numbers satisfying
the conditions:
∞
(i) 0 ≤ αn ≤ α < 1 and n=1 αn = ∞,
(ii) xn ∈ C for all n ∈ N, where xn+1 = M (xn , αn , T ).
Then {xn } converges strongly to an element of F (T ).

Proof. Let D denote co(T (C)∪{x1 }). Then D is compact by Mazur’s theorem.
The sequence {xn } is clearly in D. It follows from assumptions (i) ∼ (ii) that
{xn } is a compact sequence in C. Hence there exists a subsequence {xnk } of
{xn } such that xnk → v ∈ C, as C is closed. As in proof of Theorem 5.2.4, {xn }
is an AFPS for T , i.e., xn − T xn → 0. Thus, v ∈ F (T ). Note lim xn − v
n→∞
exists by Proposition 6.3.7. Therefore, lim xn − v = lim xnk − v = 0.
n→∞ k→∞

The next result is similar to Theorem 6.2.2 (Browder and Petryshyn’s

theorem), but the asymptotic regularity condition of T is not necessary.

Theorem 6.4.4 (Groetsch) – Let C be a nonempty closed convex subset of

a uniformly convex Banach space X, T : C → C a nonexpansive mapping that
has at least one fixed point, and {αn } a sequence of nonnegative numbers such
∞
that 0 ≤ αn < 1 and n=1 αn (1 − αn ) = ∞. Suppose that I − T maps closed
bounded subsets of C into closed subsets of C. Then the Mann iteration {xn }
defined by (6.16) converges strongly to a fixed point of T .

Proof. By Proposition 6.3.8, we have lim inf xn − T xn  = 0. It is easy to show

n→∞
that lim xn − T xn  exists (see Theorem 5.2.4). It follows that
n→∞

lim xn − T xn  = 0. (6.19)

n→∞

Let S be the strong closure of {xn }. By (6.19) and the fact that (I − T )(S) is
closed, 0 ∈ (I − T )(S). Hence there exists a subsequence {xni } converging to v,
where (I − T )v = 0. Therefore, {xn } converges strongly to v, as lim xn − v
n→∞
exists.

We now consider a class of mappings that properly includes the class of

nonexpansive mappings with fixed points.
6.4. Nonexpansive and quasi-nonexpansive mappings 295

Definition 6.4.5 Let C be a nonempty subset of a normed space X and T :

C → C a mapping that has at least one fixed point p in C. Then T is said to
be quasi-nonexpansive if
T x − p ≤ x − p for all x ∈ C.
Observation
• A nonexpansive mapping with at least one fixed point is quasi-nonexpansive.
• A linear quasi-nonexpansive is nonexpansive.

The following example shows that there exists a nonlinear continuous quasi-
nonexpansive mapping that is not nonexpansive.
Example 6.4.6 Let X = l∞ , C = BX = {x ∈ l∞ : x∞ ≤ 1} and T : C → C
a mapping defined by
T x = (0, x21 , x22 , x23 , · · · ) for x = (x1 , x2 , x3 , · · · ) ∈ C.
It is clear that T is a nonlinear continuous self-mapping on C with unique fixed
point 0. Moreover,
T x − 0∞ = (0, x21 , x22 , · · · )∞ ≤ (0, x1 , x2 , · · · )∞ = x − p for all x ∈ C,
i.e., T is quasi-nonexpansive mapping. However, T is not nonexpansive. Indeed,
for x = (1/2, 1/2, · · · ) and y = (3/4, 3/4, · · · ), we have
 
 5 5  5 1
T x − T y∞ =  0, , , · · · 

 = > = x − y∞ .
16 16 ∞ 16 4
We now show that Theorem 5.2.4 is also true for quasi-nonexpansive map-
pings in a uniformly convex Banach space.
Theorem 6.4.7 Let C be a nonempty closed convex subset of a uniformly con-
vex Banach space X and T : C → C a quasi-nonexpansive mapping that has at
least one fixed point p. Let {xn } be the Mann iteration defined by
xn+1 = Tαn xn = M (xn , αn , T ), n ∈ N,
where {αn } is a sequence of nonnegative numbers that is bounded away from 0
and 1. Then {xn } has the following properties:
(D1 ) lim xn − p exists.
n→∞
(D2 ) lim xn − T xn  = 0.
n→∞
Moreover, lim Tαn Tαn−1 · · · Tα1 x1 − Tαn−1 Tαn−2 · · · Tα1 x1  = 0.
n→∞

Proof. (a) It follows that from Proposition 6.3.7.

(b) Suppose lim xn − p = r. Because
n→∞

xn+1 − p = (1 − αn )(xn − p) + αn (T xn − p) and T xn − p ≤ xn − p,

it follows from Theorem 2.3.13 that lim xn − T xn  = 0.

n→∞
296 6. Approximation of Fixed Points

Corollary 6.4.8 Let C be a nonempty closed convex subset of a uniformly con-

vex Banach space X and T : C → C a quasi-nonexpansive mapping that has at
least one fixed point p. Let {xn } be the Mann iterative sequence defined by

xn+1 = Tα xn = M (xn , α, T ), n ∈ N.

Then Tα is asymptotically regular for each x1 ∈ C, i.e., lim Tαn x1 −Tαn+1 x1 = 0.
n→∞

Recall that a self-mapping T on a nonempty subset C of a Banach space X

is demicompact if every bounded {xn } in C such that {xn − T xn } converges
strongly contains a convergent subsequence.

The following example demonstrates that there is no connection between

continuity and demicompactness of mappings.
Example 6.4.9 Let X = C = [0, 1] with the usual metric and T : C → C a
mapping defined by

x/2, if x = 0,
Tx =
1, if x = 0.

Then T is not continuous. However, T is demicompact. In fact, if {xn } is a

bounded sequence in C such that xn − T xn → 0 as n → ∞, then by the Bolzano-
Weierstrass theorem, it follows that {xn } has a convergent subsequence.
The following example shows that there is a demicompact mapping that is
not compact.
Example 6.4.10 Let X = 2 and C = {e1 , e2 , · · · , en , · · · } be the usual
orthonormal basis for 2 . Define T : C → C by

T (ei ) = ei+1 , i ∈ N.

Then T is continuous (in fact, an isometry), but not compact. However, T is

demicompact. Indeed, if {ei }i∈N is a bounded sequence in C such that ei − T ei
converges, {ei }i∈N must be finite.

We now introduce a condition that ensures strong convergence of iterative

sequences to fixed points of nonexpansive type mappings.
Condition I. Let C be a nonempty subset of a Banach space X and T :
C → C a mapping with F (T ) = ∅. Then T is said to satisfy Condition I if there
exists a nondecreasing function f : [0, ∞) → [0, ∞) with f (0) = 0, f (t) > 0 for
t ∈ (0, ∞) such that

x − T x ≥ f (d(x, F (T )) for all x ∈ C.

T is said to satisfy Condition II if there exists a constant c > 0 such that

x − T x ≥ c d(x, F (T )) for all x ∈ C.

6.4. Nonexpansive and quasi-nonexpansive mappings 297

It is easy to see that mappings that satisfy Condition II also satisfy

Condition I.

The following example shows that there exists a mapping that is quasi-
nonexpansive mapping and satisfies Condition II.
Example 6.4.11 Let C be a nonempty subset of a Banach space X and T :
C → C a mapping such that

T x − T y ≤ ax − y + bx − T x + cy − T y for all x, y ∈ C,

where a, b, c ≥ 0 with a + b + c ≤ 1. If F (T ) = ∅, then T satisfies condition II.

Indeed, if p ∈ F (T ), then for x ∈ C

T x − p ≤ ax − p + bx − T x ≤ ax − p + b(x − p + p − T x),

which implies that

a+b
T x − p ≤ x − p.
1−b
Hence T is quasi-nonexpansive. Observe that

T x − p ≥ T x − x − x − p ≥ x − p − x − T x

and
T x − p ≤ ax − p + bx − T x.
Hence
ax − p + bx − T x ≥ x − p − x − T x,
which gives
1−a
x − T x ≥ x − p.
a+b
The constant (1 − a)/(1 + b) is positive because 0 < a, b < 1. Thus, Condition II
holds.
We now establish a relationship between mappings that satisfy Condition I
and those that are demicompact.
Proposition 6.4.12 Let C be a nonempty closed bounded subset of a Banach
space X and T : C → C a mapping with F (T ) =  ∅. If I − T maps closed
bounded subsets of C onto closed subsets of X, then T satisfies Condition I on
C.
Proof. Let M = sup{d(x, F (T )) : x ∈ C}. If M = 0, then F (T ) = C and
Condition I follows trivially. If M > 0, then for 0 < r < M , define

Cr = {x ∈ C : d(x, F (T )) ≥ r}

and
f (r) = inf{x − T x : x ∈ Cr }.
298 6. Approximation of Fixed Points

Note that Cr is nonempty closed bounded. We show that f (r) > 0 for arbitrary
r, 0 < r < M .
By hypothesis, (I − T )(Cr ) = {x − T x : x ∈ Cr } is closed. If 0 ∈ (I − T )(Cr )
then 0 = z − T z for some z ∈ Cr and hence z ∈ F (T ), but d(z, F (T )) ≥ r > 0
a contradiction. Therefore, 0 = (I − T )(Cr ).
Suppose now that f (r) = inf{x − T x : x ∈ Cr } = 0. Then there exists a
sequence {xn } in Cr such that xn −T xn  → 0. Note {xn −T xn } ⊆ (I −T )(Cr )
is a closed set. Thus, we obtain 0 ∈ (I − T )(Cr ), contradicting our statement
above that 0 ∈/ (I − T )(Cr ). Therefore, f (r) > 0 for r < M .
We extend the domain of f to R+ by defining f (0) = 0 and f (r) = sup{f (s) :
s < M } for r ≥ M . It is easy to verify that f so defined fulfills the hypotheses of
Condition I; in particular, x − T x ≥ f (d(x, F (T ))) for all x ∈ C.

We now give strong convergence of the Mann iteration for quasi-nonexpansive

mappings satisfying Condition I.

Theorem 6.4.13 Let C be a nonempty closed convex subset of a uniformly

convex Banach space X and T : C → C a continuous quasi-nonexpansive map-
ping. If T satisfies Condition I, then for arbitrary x1 ∈ C, the Mann iteration
{xn } defined by (6.16), where {αn } is a sequence of nonnegative numbers in
[0, 1] that is bounded away from 0 and 1, converges strongly to a fixed point
of T .

Proof. Because for p ∈ F (T ), xn+1 − p ≤ xn − p, it follows that

d(xn+1 , F (T )) ≤ d(xn , F (T )).

Thus, lim d(xn , F (T )) exists. By Theorem 6.4.7, lim xn − T xn  = 0. Note

n→∞ n→∞

xn − T xn  ≥ f (d(xn , F (T ))), n ∈ N,

which gives that lim d(xn , F (T )) = 0. Then for given ε > 0, there exists n0 ∈ N
n→∞
such that
ε
d(xn , F (T )) < for all n ≥ n0 .
2
Note for all n, m ≥ n0 and p ∈ F (T )

xn − xm  ≤ xn − p + p − xm 
≤ 2xn0 − p,

which implies that

xn − xm  ≤ ε.
Thus, {xn } is a Cauchy sequence and lim xn = z ∈ C. Therefore, xn −T xn → 0
n→∞
implies by the continuity of T that z ∈ F (T ).

A consequence of Proposition 6.4.12 and Theorem 6.4.13 is the following:

6.4. Nonexpansive and quasi-nonexpansive mappings 299

Corollary 6.4.14 (Browder and Petryshyn) – Let C be a nonempty closed

convex subset of a uniformly convex Banach space X and T : C → C a non-
expansive mapping. For t ∈ (0, 1), let Tt be given by Tt = tI + (1 − t)T . If I − T
maps closed bounded sets of C onto closed subsets of X and F (T ) = ∅, then for
each x ∈ C, {Ttn x} converges strongly to a fixed point of T .
In many applications, compactness is a strong condition. We now study the
problem of approximation of fixed points of nonexpansive and quasi-nonexpansive
mappings in the noncompact setting.
Theorem 6.4.15 Let X be a Banach space satisfying the Opial condition, C a
weakly compact subset of X, and T : C → X a nonexpansive mapping. Given a
sequence {xn } in C defined by (6.16), where {αn } is a sequence of nonnegative
∞
numbers such that 0 ≤ αn ≤ α < 1 and n=1 αn = ∞. Then {xn } converges
weakly to a fixed point of T .
Proof. Theorem 5.2.4 implies that lim xn − T xn  = 0. Because C is weakly
n→∞
compact, there exists a subsequence {xnk } of {xn } that converges weakly to
p ∈ C. By Theorem 5.2.9, I − T is demiclosed at zero, p = T p. Thus, all
the assumptions of Proposition 6.1.1 are satisfied. Therefore, {xn } converges
weakly to a fixed point of T by Proposition 6.1.1.

Theorem 6.4.16 Let X be a uniformly convex Banach space satisfying the

Opial condition, C a nonempty closed convex subset of X, and T : C → C
a quasi-nonexpansive mapping that has at least one fixed point. If I − T is
demiclosed at zero, then the Mann iteration {xn } defined in Theorem 6.4.7
converges weakly to a fixed point of T .

Proof. Theorem 6.4.7 implies that {xn } has properties (D1 ) ∼ (D2 ). Therefore,
the conclusion follows from Proposition 6.1.1.

We have seen in Section 3.2 that there exists a class of uniformly convex
Banach spaces without the Opial condition (e.g., Lp spaces, p = 2). Therefore,
Theorem 6.4.15 is not true for such Banach spaces. The following theorem deals
with the problem of approximation of fixed points of nonexpansive mappings in
a uniformly convex Banach space without the Opial condition.
Theorem 6.4.17 Let X be a uniformly convex Banach space with a Fŕechet
differentiable norm, C a nonempty closed convex bounded subset of X, and
T : C → C a nonexpansive mapping. Then for each x1 ∈ C, the Mann iteration
∞
{xn } defined by (6.16) with the restriction that n=1 min{αn , 1 − αn } = ∞
converges weakly to a fixed point of T .
Proof. Set
Tn := (1 − αn )I + αn T, n ∈ N. (6.20)
It is easy to see that F (T ) ⊆ F (Tn ) and Tn is nonexpansive. It follows from
Lemma 6.1.8 that lim xn , J(p − q) exists for all p, q ∈ F (T ).
n→∞
300 6. Approximation of Fixed Points

Observe that

(i) I − T is demiclosed at zero by Theorem 5.2.12;

(ii) lim xn − p exists for all p ∈ F (T ) by Proposition 6.3.7,

n→∞

(iii) lim inf xn − T xn  = 0 by Proposition 6.3.8 and lim xn − T xn  exists
n→∞ n→∞
by (5.18) imply that xn − T xn  → 0.

Thus, all the assumptions of Proposition 6.1.2 are satisfied. Therefore, {xn }
converges weakly to a fixed point of T .

The following example shows that there exists a Banach space that does not
satisfy the Opial condition and its norm is not Fŕechet differentiable. However,
its dual does have the Kadec-Klee property.

Example 6.4.18 Let X = R2 with the norm given by x = x21 + x22
and Y = Lp [0, 1] with 1 < p < ∞ and p = 2. Then the Cartesian product
X × Y equipped with the 2 -norm is uniformly convex, it does not satisfy the
Opial condition, and its norm is not Fŕechet differentiable. However, its dual
does have the Kadec-Klee property.

The following theorem is more general than Theorem 6.4.17.

Theorem 6.4.19 Let X be a uniformly convex Banach space such that its dual
has the Kadec-Klee property, C a nonempty closed convex bounded subset of X,
and T : C → C a nonexpansive mapping. Then the Mann iteration defined by
∞
(6.16) with the restriction that n=1 min{αn , 1 − αn } = ∞ converges weakly to
a fixed point of T .

Proof. It follows from the proof of Theorem 6.4.17 that Tn defined by (6.20)
is a nonexpansive mapping. Observe that

(i) I − T is demiclosed at zero;

(ii) lim txn + (1 − t)p − q exists for all p, q ∈ F (T ) and t ∈ [0, 1] by

n→∞
Lemma 6.1.8(b);

(iii) xn − T xn → 0.

Therefore, {xn } converges weakly to a fixed point of T by Proposi-

tion 6.1.2.

6.5 The modified Mann iteration process

In this section, we study weak convergence of the modified Mann iteration
process to fixed points of mappings that are more general than nonexpansive
mappings in Banach spaces.
6.5. The modified Mann iteration process 301

First, we modify the Mann iteration process and prove a useful lemma:

Let C be a nonempty convex subset of a linear space X, T : C → C a

mapping, and {αn } a real sequence such that 0 < a ≤ αn ≤ b < 1 for all n ∈ N.
Then the sequence {xn } in C defined by

x1 ∈ C;
(6.21)
xn+1 = M (xn , αn , T n ), n ∈ N

is called the modified Mann iteration.

Lemma 6.5.1 Let C be a nonempty convex subset of a Banach space X and
T : C → C an asymptotically nonexpansive mapping with sequence {kn } such
∞
that n=1 (kn − 1) < ∞. Define the modified Mann iteration {xn } by (6.21).
Then we have the following:

(a) If p is a fixed point of T , it follows that lim xn − p exists.

n→∞

(b) If lim xn − T n xn  = 0, then lim xn − T xn  = 0.

n→∞ n→∞

Proof. (a) Let p be the fixed point of T . From (6.21), we have

xn+1 − p ≤ (1 − αn )xn − p + αn T n xn − p
≤ (1 − αn )xn − p + αn kn xn − p
≤ kn xn − p for all n ∈ N.
∞
Because n=1 (kn − 1) is convergent, it follows from Lemma 6.1.5 that lim
n→∞
xn − p exists.
(b) For each n ∈ N, set dn := xn − T n xn  and L = sup kn . Then we have
n∈N

xn+1 − T xn+1  ≤ xn+1 − T n+1 xn+1  + T n+1 xn+1 − T xn+1 

≤ dn+1 + Lxn+1 − T n xn+1 
≤ dn+1 + L(xn+1 − xn  + xn − T n xn 
+T n xn − T n xn+1 
≤ dn+1 + L(αn dn + dn + Lxn − xn+1 )
≤ dn+1 + L(2 + L)dn → 0 as n → ∞.

Using Lemma 6.5.1, we prove weak convergence of the modified Mann itera-
tion {xn } defined by (6.21) in uniformly convex Banach spaces.
Theorem 6.5.2 Let X be a uniformly convex Banach space satisfying the Opial
condition, C a nonempty closed convex bounded subset of X, and T : C → C
∞
an asymptotically nonexpansive mapping with sequence {kn } such that n=1
(kn − 1) < ∞. Let {αn } be a sequence of nonnegative numbers in (0, 1) bounded
away from 0 and 1. Then the modified Mann iteration {xn } defined by (6.21)
converges weakly to a fixed point of T .
302 6. Approximation of Fixed Points

Proof. Let p be a fixed point of T . By Lemma 6.5.1(a), lim xn − p exists.

n→∞
Because

lim sup T n xn − p ≤ lim sup(kn xn − p) ≤ lim xn − p

n→∞ n→∞ n→∞

and
lim (1 − αn )(xn − p) + αn (T n xn − p) = lim xn+1 − p,
n→∞ n→∞

it follows from Theorem 2.3.13 that lim xn − T xn 

n
= 0, which in turn
n→∞
implies by Lemma 6.5.1(b) that lim xn − T xn  = 0. By Theorem 5.4.3, I − T
n→∞
is demiclosed at zero. Therefore, {xn } converges weakly to a fixed point of T
by Proposition 6.1.1.

The following convergence theorems extend Theorems 6.4.17 and 6.4.19 for
asymptotically nonexpansive mappings, respectively.
Theorem 6.5.3 Let X be a uniformly convex Banach space with Fréchet differ-
entiable norm, C a nonempty closed convex bounded subset of X, and T : C → C
∞
an asymptotically nonexpansive mapping such that n=1 (kn − 1) < ∞. Then
for each x1 ∈ C, the Mann iteration {xn } defined by (6.21), where {αn } is a
sequence of nonnegative numbers bounded away from 0 and 1, converges weakly
to a fixed point of T .
Proof. Set Tn := (1 − αn )I + αn T n , n ∈ N. It is easily seen that F (T ) ⊆ F (Tn )
and Tn is Lipschitzian with Lipschitz constant Ln = (1 − αn ) + αn kn ≥ 1.
Because
∞ ∞ ∞
(Ln − 1) = αn (kn − 1) ≤ (kn − 1) < ∞,
n=1 n=1 n=1

it follows from Lemma 6.1.8 that lim xn , J(p − q) exists for all p, q ∈ F (T ).
n→∞
Observe that
(i) lim xn − p exists for all p ∈ F (T ),
n→∞

(ii) xn − T xn → 0,
(iii) I − T is demiclosed at zero,
(iv) lim xn , J(p − q) exists for all p, q ∈ F (T ).
n→∞

Hence result follows from Proposition 6.1.2.

Theorem 6.5.4 Let X be a uniformly convex Banach space such that X ∗ has
the Kadec-Klee property, C a nonempty closed convex bounded subset of X, and
T : C → C an asymptotically nonexpansive mapping with sequence {kn } such
∞
that n=1 (kn −1) < ∞. Then for each x1 ∈ C, the Mann iteration {xn } defined
by (6.21), where {αn } is a sequence of nonnegative numbers bounded away from
0 and 1, converges weakly to a fixed point of T .
6.6. The Ishikawa iteration process 303

Proof. As in proof of Theorem 6.5.3, we have the following:

(i) lim xn − p exists for all p ∈ F (T ),
n→∞

(ii) xn − T xn → 0,
(iii) I − T is demiclosed at zero,
(iv) lim txn + (1 − t)p − q exists for all p, q ∈ F (T ) and t ∈ [0, 1] by
n→∞
Lemma 6.1.8.
Hence Theorem 6.5.4 follows from Proposition 6.1.2.

6.6 The Ishikawa iteration process

In this section, we discuss the problem of approximation of fixed points of
pseudocontractive mappings and develop iterative methods to deal with such
problems in Hilbert spaces. We have seen in Section 6.4 that the Mann itera-
tion converges (strongly) to fixed points of nonexpansive mapping in finite-
dimensional Banach spaces. The following example shows that there exists a
Lipschitz pseudocontractive mapping with a unique fixed point for which the
Mann iteration fails to converge.
Example 6.6.1 Let H, C1 , C2 , C, and T be as in Example 5.7.1. Observe that

T x2 = 2x2 for all x ∈ C1

and
T x2 = 1 + 2x2 − 2x for all x ∈ C2 .
It is clear that the origin is the only fixed point of T .
We now show that no Mann iteration for T is convergent to the origin for
any nonzero starting point. Let x ∈ C be such that x = 0. If x ∈ C1 , then

(1 − λ)x + λT x2 = (1 + λ2 )x2 > x2 for λ ∈ (0, 1),

i.e., the Mann iterate of x is actually further away form zero than x is.
If x ∈ C2 , then
  2
 λ 
(1 − λ)x + λT x = 
2  ⊥
+ 1 − 2λ x + λx 
x
 2 
λ
= 1+ − 2λ + λ2 x2
x
> 0 for λ ∈ (0, 1).

Thus, the Mann iteration {xn } defined by (6.16) has the following properties:
(i) if x1 ∈ C1 , then xn+1  > xn  for all n ∈ N,
√
(ii) if x1 ∈ C2 , then xn+1  ≥ xn / 2 for all n ∈ N.
304 6. Approximation of Fixed Points

For convergence of such a sequence to origin, xn would have to lie in neigh-

borhood C1 of the origin for all n > n0 for some n0 ∈ N. But this is not possible
because we already established for C1 that xn+1  > xn  for all n > n0 .

We now introduce an iteration process for approximation of fixed points of

pseudocontractive mappings:

Definition 6.6.2 Let C be a nonempty convex subset of a linear space X and

T : C → C a mapping. Let {αn } and {βn } be two sequences of real numbers in
[0,1] satisfying the following conditions:

(i) 0 ≤ αn ≤ βn ≤ 1 and lim βn = 0,

n→∞

∞
(ii) n=1 αn βn = ∞.

For arbitrary x1 ∈ C, define a sequence {xn } in C by


xn+1 = (1 − αn )xn + αn T yn ,
(6.22)
yn = (1 − βn )xn + βn T xn , n ∈ N.

Then {xn } is called the Ishikawa iteration.2

Before proving a theorem, we first establish two preliminary results:

Proposition 6.6.3 Let C be a nonempty convex subset of a Hilbert space

H and T : C → C a pseudocontractive mapping. Then

(1 − α)(x − y) + α(T x − T y)2 ≤ x − y2 + α2 x − y − (T x − T y)2

for all x, y ∈ C and α ∈ [0, 1].

Proof. Let x, y ∈ C. Then from the identity

(1 − λ)x + λy2 = (1 − λ)x2 + λy2 − λ(1 − λ)x − y2 , λ ∈ [0, 1];

we have

(1 − α)(x − y) + α(T x − T y)2

= (1 − α)x − y2 + αT x − T y2 − α(1 − α)x − y − (T x − T y)2
≤ (1 − α)x − y2 + α(x − y2 + x − y − (T x − T y)2 )
−α(1 − α)x − y − (T x − T y)2
≤ x − y + α2 x − y − (T x − T y)2 .

2 This iteration process was introduced by Ishikawa in 1974.

6.6. The Ishikawa iteration process 305

Proposition 6.6.4 Let C be a nonempty convex subset of a Hilbert space H

and T : C → C a pseudocontractive mapping. For 0 ≤ α ≤ β ≤ 1, define a
mapping Tα,β : C → C by Tα,β x = (1 − α)x + αT [(1 − β)x + βT x], x ∈ C. Then

Tα,β x − Tα,β y ≤ x − y2 − αβ(1 − 2β)x − y − (T x − T y)2

−α(β − α)x − y − (T ux − T uy )2
+αβT x − T y − (T ux − T uy )2 .

for all x, y ∈ C, where ux = (1 − β)x + βT x and uy = (1 − β)y + βT y.

Proof. Let x, y ∈ C. By Proposition 6.6.3, we have

ux − uy 2 = (1 − β)(x − y) + β(T x − T y)2

≤ x − y2 + β 2 x − y − (T x − T y)2 ,
ux − T ux 2 = (1 − β)(x − T ux ) + β(T x − T ux )2
= (1 − β)x − T ux 2 + βT x − T ux 2 − β(1 − β)x − T x2 .

Because T is pseudocontractive,

T ux − T uy 2 ≤ ux − uy 2 + ux − uy − (T ux − T uy )2

≤ ux − uy 2
+(1 − β)(x − y) + β(T x − T y) − (T ux − T uy )2
≤ x − y2 + β 2 x − y − (T x − T y)2
+(1 − β)x − y − (T ux − T uy )2
+βT x − T y − (T ux − T uy )2
−β(1 − β)x − y − (T x − T y)2
≤ x − y2 − β(1 − 2β)x − y − (T x − T y)2
+(1 − β)x − y − (T ux − T uy )2
+βT x − T y − (T ux − T uy )2 .

Hence

Tα,β x − Tα,β y2

= (1 − α)(x − y) + α(T ux − T uy )2
= (1 − α)x − y2 + αT ux − T uy 2
−α(1 − α)x − y − (T ux − T uy )2
≤ (1 − α)x − y2 + α{x − y2 − β(1 − 2β)x − y − (T x − T y)2
+(1 − β)x − y − (T ux − T uy )2 + βT x − T y − (T ux − T uy )2 }
−α(1 − α)x − y − (T ux − T uy )2
≤ x − y2 − αβ(1 − 2β)x − y − (T x − T y)2
−α(β − α)x − y − (T ux − T uy )2
+αβT x − T y − (T ux − T uy )2 .
306 6. Approximation of Fixed Points

Theorem 6.6.5 Let C be a nonempty compact convex subset of a Hilbert space

H and T : C → C a Lipschitzian pseudocontractive mapping with F (T ) = ∅. Let
{xn } be the Ishikawa iteration defined by (6.22). Then {xn } converges strongly
to a fixed point of T .
Proof. (a) Let p ∈ F (T ). Set Tn xn := Tαn ,βn xn . Then xn+1 = Tn xn , n ∈ N.
From Proposition 6.6.4, we have
xn+1 − p2 ≤ xn − p2 − αn βn (1 − 2βn )xn − T xn 2
−αn (βn − αn )xn − T yn 2 + αn βn T xn − T yn 2 .
Because αn ≤ βn , it follows that
xn+1 − p2 ≤ xn − p2 − αn βn (1 − 2βn )xn − T xn 2
+ αn βn T xn − T yn 2 . (6.23)
Suppose T is L-Lipschitzian mapping. Then
T xn − T yn  ≤ Lxn − yn  ≤ Lβn xn − T xn 
Hence from (6.23)
xn+1 − p2 ≤ xn − p2 − αn βn (1 − 2βn − L2 βn2 )xn − T xn 2 .
Because lim βn = 0, there exists a number n0 ∈ N such that 2βn + L2 βn2 ≤ 1/2
n→∞
for all n ≥ n0 . Hence
1
xn+1 − p2 ≤ xn − p2 − αn βn xn − T xn 2 for all n ≥ n0 , (6.24)
2
which gives that
1 
n
αi βi xi − T xi 2 ≤ xn0 − p2 − xn+1 − p2 .
2 i=n
0

Because C is bounded, {xn+1 − p} is bounded. Therefore, the series on the

left-hand side is bounded. From condition (ii), this implies that
lim inf xn − T xn  = 0,
n→∞

which in turn implies from the compactness of C that there exists subsequence
{xnj } such that lim xnj = v, where v ∈ F (T ).
j→∞
Because v ∈ F (T ), it follows from (6.24) that
xn+1 − v ≤ xn − v for all n ≥ n0 . (6.25)
Let ε > 0. Then there exists an Ni,0 such that
xNi,0 − v ≤ ε for all Ni,0 ≥ n0 .
Hence from (6.25), we get
xn − v ≤ ε for all n ≥ Ni,0 .

This completes the proof of Theorem 6.6.5.

6.7. The S-iteration process 307

6.7 The S-iteration process

For C a convex subset of a linear space X and T a mapping of C into itself, the
iterative sequence {xn } of the S-iteration process is generated from x1 ∈ C and
is defined by 
xn+1 = (1 − αn )T xn + αn T yn ,
(6.26)
yn = (1 − βn )xn + βn T xn , n ∈ N,
where {αn } and {βn } are real sequences in (0,1) satisfying the condition:
∞

αn βn (1 − βn ) = ∞. (6.27)
n=1

Let us compare the rate of convergence of the Picard, Mann, and S-iteration
processes for contraction mappings.
Proposition 6.7.1 Let C be a nonempty closed convex subset of a Banach
space X and T : C → C a contraction mapping with Lipschitz constant k and
a unique fixed point p. For u1 = v1 = w1 ∈ C, define sequences {un }, {vn } and
{wn } in C as follows:
Picard iteration: un+1 = T un , n ∈ N
Mann iteration: vn+1 = (1 − αn )vn + αn T vn , n ∈ N
S-iteration: wn+1 = (1 − αn )T wn + αn T yn ,
yn = (1 − βn )wn + βn T wn , n ∈ N,
where {αn } and {βn } are real sequences in (0, 1). Then we have the following:
(a) un+1 − p ≤ kun − p for all n ∈ N.
(b) vn+1 − p ≤ vn − p for all n ∈ N.
(c) wn+1 − p ≤ k[1 − (1 − k)αn βn ]wn − p for all n ∈ N.
Proof. Part (a) is obvious.
(b) Now part (b) follows from
vn+1 − p = (1 − αn )(vn − p) + αn (T vn − p)
≤ (1 − αn )vn − p + kαn vn − p
≤ [1 − (1 − k)αn ]vn − p for all n ∈ N.
(c) For all n ∈ N, we have
wn+1 − p ≤ (1 − αn )kwn − p + αn kyn − p
≤ k[(1 − αn )wn − p + αn ((1 − βn )wn − p + kβn wn − p)]
= k[1 − (1 − k)αn βn ]wn − p.

It is obvious that the rate convergence of the S-iteration process is faster

than the Picard iteration process and the Picard iteration process is faster than
the Mann iteration process for contraction mappings.
We now discuss the S-iteration process for nonexpansive mappings.
308 6. Approximation of Fixed Points

Lemma 6.7.2 Let X be a normed space, C a nonempty convex subset of X,

and T : C → C a nonexpansive mapping. If {xn } is the iterative process defined
by (6.26), then lim xn − T xn  exists.
n→∞

Proof. Set an := xn − T xn for all n ∈ N. Then, we have

an+1  ≤ (1 − αn )T xn − T xn+1  + αn T yn − T xn+1 

≤ (1 − αn )xn − xn+1  + αn yn − xn+1 . (6.28)

Because

yn − T yn  ≤ (1 − βn )xn − T yn  + βn T xn − T yn 
≤ (1 − βn )xn − T yn  + βn2 an ,
xn+1 − yn  ≤ (1 − αn )yn − T xn  + αn yn − T yn 
= (1 − αn )(1 − βn )an  + αn ((1 − βn )xn − T yn  + βn2 an )
≤ (1 − αn )(1 − βn )an  + αn ((1 − βn )(xn − T xn  (6.29)
+T xn − T yn ) + βn2 an )
≤ [(1 − αn )(1 − βn ) + αn (1 − βn )(1 + βn ) + αn βn2 ]an 
= (1 − βn + αn βn )an ,
xn+1 − xn  ≤ (1 − αn )an  + αn xn − T yn 
≤ (1 − αn )an  + αn (xn − T xn  + T xn − T yn ) (6.30)
≤ (1 + αn βn )an .

From (6.28), (6.29), and (6.30), we have

an+1  ≤ [(1 − αn )(1 + αn βn ) + αn (1 − βn + αn βn )]an 

= an ,

so that {an } is nonincreasing and hence lim an  exists.

n→∞

Theorem 6.7.3 Let C be a nonempty closed convex (not necessary bounded)

subset of a uniformly convex Banach space X and T : C → C a nonexpansive
mapping. Let {xn } be the sequence defined by (6.26) with the restriction:

lim αn βn (1 − αn ) exists and lim αn βn (1 − βn ) = 0. (6.31)

n→∞ n→∞

Then, for arbitrary initial value x1 ∈ C, {xn − T xn } converges to some

constant rC (T ), which is independent of the choice of the initial value x1 ∈ C.

Proof. Lemma 6.7.2 implies that lim xn − T xn  exists. We denote this
n→∞
limit by r(x1 ). Let {x∗n } be another iterative sequence generated by (6.26) with
the same restriction on iteration parameters {αn } and {βn } as the sequence
{xn }, but with the initial value x∗1 ∈ C. It follows from Lemma 6.7.2 that
lim x∗n − T x∗n  = r(x∗1 ).
n→∞
6.7. The S-iteration process 309

Observe that

T yn − T yn∗  ≤ yn − yn∗ 

≤ (1 − βn )xn − x∗n  + βn T xn − T x∗n 
≤ xn − x∗n 

and

xn+1 − x∗n+1  = (1 − αn )(T xn − T x∗n ) + αn (T yn − T yn∗ )

≤ (1 − αn )xn − x∗n  + αn T yn − T yn∗ 
≤ (1 − αn )xn − x∗n  + αn yn − yn∗  (6.32)
≤ xn − x∗n .

This shows that lim xn − x∗n  exists. Let lim xn − x∗n  = d for some d > 0.
n→∞ n→∞

By using Corollary 2.3.10, we obtain that

yn − yn∗  = (1 − βn )(xn − x∗n ) + βn (T xn − T x∗n )

 
xn − x∗n − (T xn − T x∗n )
≤ [1 − 2βn (1 − βn )δX xn − x∗n ,
xn − x∗n 
it follows from (6.32) that

xn+1 − x∗n+1  ≤ xn − x∗n 

 
xn − x∗n − (T xn − T x∗n )
−2αn βn (1 − βn )xn − x∗n δX .
xn − x∗n 
This gives us
∞
  
xn − x∗n − (T xn − T x∗n )
αn βn (1 − βn )xn − x∗n δX ≤ x1 − x∗1  < ∞.
n=1
xn − x∗n 

Because lim αn βn (1 − βn ) = 0 and lim xn − x∗n  = d, we

n→∞ n→∞
 
xn − x∗n − (T xn − T x∗n )
lim δX = 0.
n→∞ xn − x∗n 
Because δX is strictly increasing and continuous and lim xn − x∗n  = d > 0,
n→∞
we have
lim xn − x∗n − (T xn − T x∗n ) = 0.
n→∞

Observe that

xn − T xn  − x∗n − T x∗n  ≤ xn − T xn − (x∗n − T x∗n ),

which implies that

lim xn − T xn  − x∗n − T x∗n  = 0.

n→∞
310 6. Approximation of Fixed Points

Thus, r(x1 ) = r(x∗1 ). Because

xn+1 − T xn+1  ≤ xn − T xn  ≤ x1 − T x1  for all n ∈ N, and x1 ∈ C,

it follows that
rC (T ) = inf{x − T x : x ∈ C}.

Theorem 6.7.4 Let X be a real uniformly convex Banach space with a Fréchet
differentiable norm or that satisfies Opials condition. Let C be a nonempty
closed convex (not necessary bounded) subset of and T : C → C a nonexpansive
mapping with F (T ) = ∅. Let {xn } be the sequence defined by (6.26) with the
restriction (6.31). Then {xn } converges weakly to a fixed point of T .

Proof. Set Tn x := (1 − αn )T x + αn T ((1 − βn )x + βn T x) for all x ∈ C and

n ∈ N. Then for each n ∈ N, the mapping Tn : C → C is also nonexpansive and
the S-iterative sequence {xn } defined by (6.26) can be written as

xn+1 = Tn xn for all n ∈ N.

Furthermore, we have F (T ) ⊆ F (Tn ) for all n ∈ N. Because F (T ) = ∅, by

Theorem 6.7.3, we see that lim xn − T xn  = 0. The remainder of the proof
n→∞
is followed by Theorems 6.4.16 and 6.4.17.

Bibliographic Notes and Remarks

The important tools presented in Section 6.1 were presented in Deng [46],
Groetsch [64], Osilike and Aniagbosor [118], Reich [125], and Tan and Xu
[159, 160].
Theorem 6.2.2 is established by Browder and Petryshyn [30]. Other results
of Section 6.2 belong to Gornicki [62]. Theorem 6.2.5 is an improvement of the
main result of Bose [20]. Theorem 6.3.5 is proved in Chidume and Mutangadura
[39].
Theorems 6.4.3 and 6.4.4 are established by Ishikawa [71] and Groetsch [63],
respectively. Theorem 6.4.17 is similar to the result established by Reich in
[125]. Other results presented in Section 6.4 are discussed in Krasnoselskii [93]
Senter and Dotson [142], and Singh and Whitfield [147]. A survey of iteration
processes can be found in an article [135] by Rhoades. The convergence of
the modified Mann iteration process for asymptotically nonexpansive mappings
presented here is discussed in Chidume, Shahzad, and Zegeye [40] and Tan and
Xu [160]. Theorem 6.6.5 is proved in Ishikawa [70]. Notice that S-iteration
process is faster than the Picard iteration process. The S-iteration process was
introduced in Agarwal, O’Regan, and Sahu [2].
6.7. The S-iteration process 311

Exercises

6.1 Let C be a compact convex subset of a Hilbert space H and let P(C) denote
family of all bounded proximinal subsets of C. Let T : C → P(C) be a
nonexpansive mapping with fixed point p. Define the sequence {xn } of
Mann iterates by x1 ∈ C,

xn+1 = αn xn + (1 − αn )yn , n ∈ N,

where yn ∈ T xn is such that yn − p = d(p, T xn ) and {αn } is a real

∞
sequence such that 0 ≤ αn < 1 and n=1 αn = ∞. Show that there
exists a subsequence {xnk } of {xn } that converges to a fixed point of T.
6.2 Let X be a uniformly convex Banach space that satisfies the Opial con-
dition or has a Fréchet differentiable norm, C a closed convex bounded
subset of X, and T : C → C an asymptotically nonexpansive mapping
∞
with sequence {kn } such that n=1 (kn − 1) < ∞. Suppose that x1 is a
given point in C and {αn } and {βn } are real sequences such that {αn } is
bounded away from 0 and 1 and {βn } is bounded away from 1. Show that
the sequence {xn } defined by the modified Ishikawa iteration process:
(I) xn+1 = (1 − αn )xn + αn T n ((1 − βn )xn + βn T n x), n ∈ N
converges weakly to a fixed point of T .
6.3 Let C be a nonempty subset of a Banach space X and T : C → C a
mapping. T is said to be asymptotically quasi-nonexpansive if F (T ) = ∅
and there exists a sequence {kn } ⊂ [1, ∞) with lim kn = 1 such that
n→∞

T x − p ≤ kn x − p for all x ∈ C, p ∈ F (T ) and n ∈ N.

n

If C is a nonempty compact convex subset of a uniformly convex Banach

space X and T : C → C is a uniformly continuous asymptotically quasi-
∞
nonexpansive mapping with sequence {kn } such that n=1 (kn − 1) < ∞,
show that sequence {xn } defined by (I) converges strongly to a fixed point
of T .
6.4 Let C be a nonempty subset of a Banach space X and T : C → C a
mapping. T is said to be asymptotically nonexpansive in the intermediate
sense provided T is uniformly continuous and

lim sup sup (T n x − T n y − x − y) ≤ 0.

n→∞ x,y∈C

If C is a nonempty closed convex subset of a uniformly convex Banach

space X and T : C → C is completely continuous and asymptotically
nonexpansive in the intermediate sense with
∞

cn = sup (T n x − T n y − x − y) ∨ 0 such that cn < ∞,
x,y∈C n=1
312 6. Approximation of Fixed Points

show that sequence {xn } defined by (I) converges strongly to a fixed point
of T .
6.5 Let C be a nonempty closed convex subset of a Banach space X and T :
C → C a contraction mapping with Lipschitz constant k and a unique
fixed point p. For u1 = v1 = w1 ∈ C, define sequences {un }, {vn } and
{wn } in C as follows:
Picard iteration: un+1 = T un , n ∈ N;
Mann iteration: vn+1 = (1 − αn )vn + αn T vn , n ∈ N;
S-iteration: wn+1 = (1 − αn )T wn + αn T yn ,
yn = (1 − βn )wn + βn T wn , n ∈ N;
where {αn } and {βn } are real sequences in (0, 1). Show that
(a) un+1 − p ≤ kun − p for all n ∈ N.
(b) vn+1 − p ≤ vn − p for all n ∈ N.
(c) wn+1 − p ≤ kwn − p for all n ∈ N.
6.6 Let C be a nonempty closed convex subset of a Hilbert space H and T :
C → C a nonexpansive mapping. Show that the following are equivalent:
(a) F (T ) = ∅.
n−1
(b) For any x ∈ C, { n1 i=0 T i x} converges weakly to a fixed point of
T.
6.7 Let C be a nonempty subset of a Hilbert space and T : C → C a mapping.
The mapping T is said to satisfy condition (A) if F (T ) = ∅ and there
exists a real positive number λ such that
x − T x, x − p ≥ λx − T x2 for all x ∈ C and p ∈ F (T ).

If C is a nonempty closed convex subset of a Hilbert space H, T : C → C

is a mapping that satisfies condition (A), I − T is demiclosed at zero, and
{xn } is a sequence in C generated by xn+1 = (1 − αn )xn + αn T xn , n ∈ N
with 0 < a < αn ≤ b < 1, show that {xn } converges weakly to an element
of F (T ).
6.8 Let C be a nonempty convex subset of a Banach space X and Ti : C →
C (i = 1, 2, · · · , k) nonexpansive mappings. Let
S = α0 + α1 T1 + α2 T2 + · · · + αk Tk ,
k
where αi ≥ 0, α0 > 0 and i=1 αi = 1. If {xn } is a bounded sequence in
C defined by
xn+1 = Sxn , n ∈ N,
show that xn − Sxn → 0 as n → ∞.
6.7. The S-iteration process 313

6.9 Let X be a Banach space that satisfies the Opial condition and C a weakly
compact convex subset of X. Let Ti (i = 1, 2, · · · , k) and {xn } be as in
Exercise 6.8. Show that {xn } converges weakly to a fixed point of S.

6.10 Let C be a nonempty closed convex subset of a Banach space. Let {Ti : i =
1, 2, · · · , k} be k asymptotically quasi-nonexpansive self-mappings of C,
i.e., Tin x − qi  ≤ kin x − qi  for all x ∈ C, qi ∈ F (Ti ), i ∈ {1, 2, 3, · · · , k}.
k
Suppose that F = i=1 F (Ti ) = ∅, x0 ∈ C, {αn } ⊂ (s, 1 − s) for some
∞
s ∈ (0, 1), and n=1 (kin − 1) < ∞ for all i ∈ {1, 2, 3, · · · , k}. Show that
the implicity iterative sequence {xn } generated by

xn = αn xn−1 + (1 − αn )Tim xn , n ∈ N,

where n = (m−1)k +i, i ∈ {1, 2, 3, · · · , k} converges strongly to a common

fixed point if and only if lim inf d(xn , F ) = 0.
n→∞
Chapter 7

Strong Convergence
Theorems

In this chapter, we prove convergence theorems for approximants of self-

mappings and non-self mappings in Banach spaces. We also study a Halpern’s
type iteration process for approximation of fixed points of nonexpansive
mappings in a Banach space with a uniformly Gâteaux differentiable norm.

7.1 Convergence of approximants of

self-mappings
In this section, we study strong convergence of approximants of nonexpansive
and asymptotically nonexpansive type self-mappings in Banach spaces.
First, we establish a fundamental strong convergence theorem for nonexpan-
sive mappings in a Hilbert space.

Theorem 7.1.1 (Browder’s convergence theorem) – Let C be a nonempty

closed convex bounded subset of a Hilbert space H. Let u be an element in C
and Gt : C → C, t ∈ (0, 1) the family of mappings defined by

Gt x = (1 − t)u + tT x, x ∈ C.

Then the following hold:

(a) There is exactly one fixed point xt of Gt , i.e.,

xt = (1 − t)u + tT xt . (7.1)

(b) The path {xt } converges strongly to P u as t → 1, where P is the metric

projection mapping from C onto F (T ).

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 7,

c Springer Science+Business Media, LLC, 2009
316 7. Strong Convergence Theorems

Proof. (a) Note for each t ∈ (0, 1), Gt is a contraction mapping of C into itself.
Hence Gt has a unique fixed point xt in C.
(b) Because F (T ) is a nonempty closed convex subset of C, there exists an
element u0 ∈ F (T ) that is the nearest point of u. By boundedness of {xt }, there
exists a subsequence {xtn } of {xt } such that xtn  z ∈ C. Write xtn = xn .
Because xn − T xn → 0, it follows that z = T z. Indeed, for z = T z

lim sup xn − z < lim sup xn − T z

n→∞ n→∞
≤ lim sup(xn − T xn  + T xn − T z)
n→∞
≤ lim sup xn − z,
n→∞

a contradiction, because H has the Opial condition. Observe that

(1 − tn )xn + tn (xn − T xn ) = (1 − tn )u

and
(1 − tn )u0 + tn (u0 − T u0 ) = (1 − tn )u0 .
Subtracting and taking the inner product of the difference with xn − u0 , we get

(1 − tn )xn − u0 , xn − u0  + tn U xn − U u0 , xn − u0 
= (1 − tn )u − u0 , xn − u0 ,

where U = I − T . Because U = I − T is monotone, U xn − U u0 , xn − u0  ≥ 0,

it follows that

xn − u0 2 ≤ u − u0 , xn − u0  for all n ∈ N.

Because u0 ∈ F (T ) is the nearest point to u,

u − u0 , z − u0  ≤ 0,

which gives

xn − u0 2 ≤ u − u0 , xn − u0 
= u − u0 , xn − z + u − u0 , z − u0 
≤ u − u0 , xn − z.

Thus, from xn  z, we obtain xn → u0 as n → ∞. We show that xt → u0 as

t → 1, i.e., u0 is the only strong cluster point of {xt }. Suppose, for contradiction,
that {xtn } is another subsequence of {xt } such that xtn → v = u0 as n → ∞.
Set xn := xtn . Because xn − T xn → 0, it follows that v ∈ F (T ). From (7.1),
we have
xt − T xt = (1 − t)(u − T xt ). (7.2)
7.1. Convergence of approximants of self-mappings 317

Because for y ∈ F (T )

xt − T xt , xt − y = xt − T y + T y − T xt , xt − y
= xt − y2 − T xt − T y, xt − y
≥ 0,

this gives from (7.2) that u − T xt , xt − y ≥ 0. Thus, xt − u, xt − y ≤ 0 for

all t ∈ (0, 1) and y ∈ F (T ). It follows that

u0 − u, u0 − v ≤ 0 and v − u, v − u0  ≤ 0,

which imply that u0 = v, a contradiction. Therefore, {xt } converges strongly

to P u, where P is metric projection mapping from C onto F (T ).
We now prove strong convergence of path {xt } in a more general situation.
Proposition 7.1.2 Let C be a nonempty subset of a Banach space X and T :
C → X a pseudocontractive mapping such that for some u ∈ C, the equation

x = (1 − t)u + tT x (7.3)

has a unique solution xt in C for each t ∈ (0, 1). If F (T ) = ∅, there exists

j(xt − v) ∈ J(xt − v) such that

xt − u, j(xt − v) ≤ 0 for all v ∈ F (T ) and t ∈ (0, 1).

Proof. From (7.3) we have

xt − T xt = (1 − t)(u − T xt ) for all t ∈ (0, 1).

For y ∈ F (T ), there exists j(xt − y) ∈ J(xt − y) such that

xt − T xt , j(xt − y) = xt − T y + T y − T xt , j(xt − y)

= xt − y2 − T xt − T y, j(xt − y)
≥ 0,

which implies that

u − T xt , j(xt − y) ≥ 0.
It follows from (7.3) that

xt − u, j(xt − y) ≤ 0 for all y ∈ F (T ) and t ∈ (0, 1).

Theorem 7.1.3 Let X be a reflexive Banach space with a weakly continuous

duality mapping J : X → X ∗ . Let C be a nonempty closed subset of X and
T : C → X a demicontinuous pseudocontractive mapping such that for some
u ∈ C, the equation defined by (7.3) has a unique solution xt in C for each
t ∈ (0, 1). If the path {xt } is bounded, then it converges strongly to a fixed point
of T as t → 1.
318 7. Strong Convergence Theorems

Proof. Because {xt } is bounded, {T xt } is bounded by (7.3) and

xt − T xt  = (1 − t)u − T xt  ≤ (1 − t)diam({u − T xt }) → 0.

Because X is reflexive and {xt } is bounded, there exists a subsequence {xtn }

of {xt } such that xtn  v as tn → 1. Write xtn := xn . Because (t−1 − 1)xt =
(t−1 − 1)u + T xt − xt , it follows that

(t−1
n − 1)xn − (t−1
m − 1)xm , J(xn − xm )
= (t−1 −1
n − tm )u, J(xn − xm )
+T xn − T xm − (xn − xm ), J(xn − xm )
≤ (t−1 −1
n − tm )u, J(xn − xm ) for all n, m ∈ N.

Taking the limit as m → ∞, we obtain

(t−1 −1
n − 1)xn , J(xn − v) ≤ (tn − 1)u, J(xn − v),

and thus,
xn − u, J(xn − v) ≤ 0.
Hence

xn − v2 = xn − v, J(xn − v) = xn − u, J(xn − v) + u − v, J(xn − v).

Therefore, xn → v as n → ∞. Because T xn → v by xn − T xn → 0, it follows

from the demicontinuity of T that v ∈ F (T ).
We show that v is the only strong cluster point of {xt }. Suppose, for con-
tradiction, that {xtn } is another subsequence of {xt } such that xtn → w (= v)
as tn → 1. It can be easily seen that w = T w. Thus, from Proposition 7.1.2,
we have

xtn − u, J(xn − w) ≤ 0 and xtn − u, J(xtn − v) ≤ 0

which imply that

v − u, J(v − w) ≤ 0 and w − u, J(w − v) ≤ 0.

Hence

u − w2 = v − w, J(v − w) = v − u, J(v − w) + u − w, J(v − w) ≤ 0,

a contradiction. Therefore, {xt } converges strongly to a fixed point of T as

t → 1.
Corollary 7.1.4 Let X be a reflexive Banach space with a weakly continuous
duality mapping J : X → X ∗ . Let C be a nonempty closed subset of X and
T : C → X a nonexpansive mapping such that for some u ∈ C, the equation
(7.3) has a unique solution xt in C for each t ∈ (0, 1). If the path {xt } is
bounded, then it converges strongly to a fixed point of T as t → 1.
7.1. Convergence of approximants of self-mappings 319

Applying Theorem 7.1.3, we obtain

Theorem 7.1.5 Let X be a reflexive Banach space with a weakly continuous
duality mapping, C a nonempty closed convex bounded subset of X, u an element
in C, and T : C → C a continuous pseudocontractive mapping. Then the
following hold:
(a) For each t ∈ (0, 1), there exists exactly one xt ∈ C such that

xt = (1 − t)u + tT xt . (7.4)

(b) {xt } converges strongly to a fixed point of T as t → 1.

Proof. (a) For each t ∈ (0, 1), define Gt : C → C by

Gt x = (1 − t)u + tT x, x ∈ C.

Then Gt is well defined because u ∈ C and T (C) ⊂ C. Because for each

t ∈ (0, 1), Gt is strongly pseudocontractive, it follows from Corollary 5.7.15 that
Gt has exactly one fixed point xt ∈ C.
(b) It follows from Theorem 7.1.3.
Corollary 7.1.6 Let X be a reflexive Banach space with a weakly continuous
duality mapping J : X → X ∗ , C a nonempty closed convex bounded subset of
X, and T : C → C a continuous pseudocontractive mapping. Then F (T ) is a
sunny nonexpansive retract of C.
Proof. For each u ∈ C, by Theorem 7.1.5, there is a unique path {xt } defined
by (7.4) such that lim xt = v ∈ F (T ). Then there exists a mapping P from C
t→1
onto F (T ) defined by P u = lim xt , as u is an arbitrary element of C.
t→1
Because

xt − u, J(xt − y) ≤ 0 for all y ∈ F (T ) and t ∈ (0, 1),

this implies that

P u − u, J(P u − y) ≤ 0 for all u ∈ C, y ∈ F (T ).

Therefore, by Proposition 2.10.21, P is the sunny nonexpansive retraction from

C onto F (T ).

Next, we study a strong convergence theorem for the following more general
class of mappings:

Definition 7.1.7 Let C be a nonempty subset of a Banach space X and T :

C → C a mapping. Then T is said to be asymptotically pseudocontractive if for
each n ∈ N and x, y ∈ C, there exist a sequence {kn } in [1, ∞) with lim kn = 1
n→∞
and j(x − y) ∈ J(x − y) such that T n x − T n y, j(x − y) ≤ kn x − y2 .
320 7. Strong Convergence Theorems

We note that every asymptotically nonexpansive mapping is asymptotically

pseudocontractive, but the converse is not true. In fact, if T is asymptotically
nonexpansive with domain Dom(T ) and sequence {kn }, then for each n ∈ N
and x, y ∈ Dom(T ), there exists j(x − y) ∈ J(x − y) such that

T n x − T n y, j(x − y) ≤ T n x − T n yx − y ≤ kn x − y2 .

Theorem 7.1.8 Let X be a reflexive Banach space with a weakly continuous

duality mapping J : X → X ∗ . Let C be a nonempty closed subset of X and
T : C → C a demicontinuous asymptotically pseudocontractive mapping with
sequence {kn }. Let u be an element in C and {tn } a sequence of nonnegative
numbers in (0, 1) such that tn → 1 and lim (kn − 1)/(1 − tn ) = 0. Let {xn } be
n→∞
a bounded sequence in C with xn − T xn → 0 such that

xn = (1 − tn )u + tn T n xn for all n ∈ N. (7.5)

If I − T is demiclosed at zero, then {xn } converges strongly to a fixed point

of T .

Proof. From (7.5), we have

xn − T n xn = (1 − tn )(u − T n xn ) and tn (u − T n xn ) = u − xn .

Thus, whenever y ∈ F (T ), we have

(1 − tn )u − T n xn , J(xn − y) = xn − T n xn , J(xn − y)

= xn − y + y − T n xn , J(xn − y)
= xn − y2 − T n xn − T n y, J(xn − y)
≥ −(kn − 1)xn − y2 ,

which yields

kn − 1 kn − 1
xn − u, J(xn − y) ≤ xn − y2 ≤ K (7.6)
1 − tn 1 − tn

for some K ≥ 0.
Because X is reflexive, there exists a subsequence {xni } of {xn } such that
xni  v ∈ C. Because I − T is demiclosed at zero, v = T v. Hence

xni − v2 = xni − v, J(xni − v)

= xni − u, J(xni − v) + u − v, J(xni − v)
kni − 1
≤ K + u − v, J(xni − v).
1 − tni

From J(xni − v) →∗ 0 and (kni − 1)/(1 − tni ) → 0, we get xni → v.

7.1. Convergence of approximants of self-mappings 321

We now show that v is only strong cluster point of {xn }. Suppose, for
contradiction, that {xnj } is another subsequence of {xn } such that xnj →
w ∈ C. Because xnj − T xnj → 0, it follows that T xnj → w. By demicon-
tinuity of T , we have that T xnk  T w. Hence T w = w. From (7.6), we have

v − u, J(v − w) ≤ 0 and w − u, J(w − v) ≤ 0,

which imply that

v − w2 = v − w, J(u − w) = v − u, J(v − w) + u − w, J(v − w) ≤ 0,

a contradiction. Therefore, {xn } converges strongly to a fixed point of T .

Corollary 7.1.9 Let X be a reflexive Banach space with a weakly continuous

duality mapping J : X → X ∗ . Let C be a nonempty closed subset of X and
T : C → C an asymptotically nonexpansive mapping with sequence {kn }. Let
u be an element in C and {tn } a sequence of real numbers in (0, 1) such that
tn → 1 and lim (kn − 1)/(1 − tn ) = 0. Let {xn } be a bounded sequence in C
n→∞
with xn − T xn → 0 such that xn = (1 − tn )u + tn T n xn for all n ∈ N. Then {xn }
converges strongly to a fixed point of T .

The following result is very useful for strong convergence of AFPS of self-
mappings as well as non-self mappings.

Theorem 7.1.10 Let X be a reflexive Banach space whose norm is uniformly

Gâteaux differentiable, C a nonempty closed convex subset of X, T : C → X a
demicontinuous mapping with F (T ) = ∅, and A : C → C a continuous strongly
pseudocontractive mapping with constant k ∈ [0, 1). Let {αn } be a sequence in
R+ with lim αn = 0 and {xn } a bounded sequence in C such that xn −T xn → 0
n→∞
as n → ∞ and

xn − Axn , J(xn − p) ≤ αn xn − p2 for all n ∈ N and p ∈ F (T ). (7.7)

Suppose the set M = {x ∈ C : LIMn xn − x2 = inf LIM xn − y2 } contains
y∈C
a fixed point of T , where LIM is a Banach limit. Then {xn } converges strongly
to an element of M ∩ F (T ).

Proof. By Theorem 2.9.11, M is a nonempty closed convex and bounded set.

By assumption, T has a fixed point in M . Denote such a fixed point by v.
It follows from Corollary 2.9.13 that

LIMn z, J(xn − v) ≤ 0 for all x ∈ C.

In particular,

LIMn Av − v, J(xn − v) ≤ 0. (7.8)

322 7. Strong Convergence Theorems

From (7.7), we obtain

LIMn xn − Axn , J(xn − v ≤ 0. (7.9)

Combining (7.8) and (7.9), we have

LIMn xn − v2 = LIMn [xn − Axn , J(xn − v) + Axn − Av, J(xn − v)
+Av − v, J(xn − v)]
≤ kLIMn xn − v2 ,

i.e., (1 − k)LIMn xn − v2 ≤ 0. Therefore, there is a subsequence {xni } of

{xn } that converges strongly to v. To complete the proof, let {xnj } be another
subsequence of {xn } such that xnj → z as j → ∞. Because xnj − T xnj → 0,
it follows that T xnj → z. By demicontinuity of T , we have that T z = z. From
(7.7), we have

v − Av, J(v − z) ≤ 0 and z − Az, J(z − v) ≤ 0.

Hence z = v. This proves that {xn } converges strongly to v.

Corollary 7.1.11 Let X be a reflexive Banach space whose norm is uniformly
Gâteaux differentiable, C a nonempty closed convex subset of X, and T : C → X
a demicontinuous mapping with F (T ) = ∅. Let u be an element in C, {αn } a
sequence in R+ with lim αn = 0, and {xn } a bounded sequence in C such that
n→∞
xn − T xn → 0 as n → ∞ and

xn − u, J(xn − p) ≤ αn xn − p2 for all n ∈ N and all p ∈ F (T ).

Suppose the set M = {x ∈ C : LIMn xn − x2 = inf LIM xn − y2 } contains
y∈C
a fixed point of T , where LIM is a Banach limit. Then {xn } converges strongly
to an element of F (T ).

We now prove a notable strong convergence theorem for nonexpansive map-

pings in a uniformly smooth Banach space.
Theorem 7.1.12 (Reich’s convergence theorem) – Let C be a nonempty
closed convex subset of a uniformly smooth Banach space X, x an element in
C, T : C → C a nonexpansive mapping, and Gt : C → C, t ∈ (0, 1), the family
of mappings defined by Gt (x) = (1 − t)x + tT Gt (x). If T has a fixed point, then
for each x ∈ C, lim Gt (x) exists and is a fixed point of T .
t→1

Proof. Let {tn } be a sequence of real numbers in (0, 1) such that tn → 1.

Set xn := Gtn (x). Because F (T ) = ∅, it follows that {xn } is bounded and
xn − T xn → 0 as n → ∞. Then the set M defined by (2.32) is a nonempty
closed convex bounded T -invariant subset of C (see Proposition 6.1.3). Note
every uniformly smooth Banach space is reflexive and has normal structure.
Hence every closed convex bounded set of X has fixed point property. Thus,
7.1. Convergence of approximants of self-mappings 323

T has a fixed point in M . Observe that {xn } satisfies (7.7) with αn = 0 for
all n ∈ N (see Proposition 7.1.2). It follows from Corollary 7.1.11 that {xn }
converges strongly to an element of F (T ).

Applying Corollary 7.1.11, we obtain

Theorem 7.1.13 Let X be a reflexive Banach space with a uniformly Gâteaux

differentiable norm, C a nonempty closed convex subset of X, and T : C →
C an asymptotically nonexpansive mapping with sequence {kn }. Let u be an
element in C and {tn } a sequence of real numbers in (0, 1) such that tn → 1
and (kn − 1)/(1 − tn ) → 0. Then the following hold:

(a) There exists exactly one point xn ∈ C such that

xn = (1 − tn )u + tn T n xn , n ∈ N.

(b) If {xn } is a bounded AF P S of T and M = {x ∈ C : LIMn xn − x2 =

inf LIMn xn − y2 } contains a fixed point of T , then {xn } converges
y∈C
strongly to an element of F (T ).

Proof. (a) Because lim (kn − 1)/(1 − tn ) = 0, then there exists a sufficiently
n→∞
large natural number n0 such that kn tn < 1 for all n ≥ n0 . For each n ∈ N,
define Tn : C → C by

Tn x = (1 − tn )u + tn T n x, x ∈ C.

Because for each n ≥ n0 , Tn is contraction, there exists exactly one fixed point
xn ∈ C of Tn . We may assume that xn = u for all n = 1, 2, · · · , n0 − 1. Then

xn = (1 − tn )u + tn T n xn for all n ∈ N.

(b) As in the proof of Theorem 7.1.8, it can be easily seen that {xn } satisfies
the inequality (7.6). Note that M is a nonempty closed convex bounded set.
Moreover, T has a fixed point in M by assumption.
Observe that
(i) (7.7) is satisfied with αn = (kn − 1)/(1 − tn ) → 0 as n → ∞,
(ii) T has a fixed point in M ,
(iii) xn − T xn  → 0 as n → ∞.
Hence this part follows from Corollary 7.1.11.

The following proposition shows that for a bounded AFPS, the set M satisfies
the property (P ) defined by (5.52).

Proposition 7.1.14 Let C be a nonempty closed convex bounded subset of a

reflexive Banach space X and T : C → C asymptotically nonexpansive. Let
{xn } be an AFPS. Then the set M satisfies property (P ).
324 7. Strong Convergence Theorems

Proof. By Theorem 2.9.11, M is a nonempty closed convex bounded subset of

C. Let x ∈ M . Because {T m x} is bounded in C, there exists a subsequence
{T mj x} of {T m x} such that T mj x  u ∈ C. Let kn be the Lipschitz constant
of T n . By w-lsc of the function ϕ(z) = LIMn xn − z2 , we have

ϕ(u) = lim inf ϕ(T mj x)

j→∞
≤ lim sup ϕ(T m x)
m→∞
= lim sup(LIMn xn − T m x2 )
m→∞
≤ lim sup(LIMn (xn − T xn +T xn − T 2 xn  + · · · + T m−1 xn − T m xn 
m→∞
+T m xn − T m x)2 )
≤ lim sup(LIMn (km xn − x))2
m→∞
= ϕ(x) = inf ϕ(z).
z∈M

Thus, u ∈ M . Therefore, M has property (P ).

Applying Theorem 5.5.8 and Proposition 7.1.14, we obtain

Theorem 7.1.15 Let C be a nonempty closed convex bounded subset of a uni-

formly smooth Banach space X and T : C → C an asymptotically nonexpansive
mapping with sequence {kn }. Let u ∈ C and {tn } a sequence in (0, 1) such that
tn → 1 and (kn − 1)/(1 − tn ) → 0. Suppose the sequence {xn } defined by (7.5)
is an AFPS of T . Then {xn } converges strongly to a fixed point of T .

Proof. By Proposition 7.1.14, the set M has property (P). It follows from
Theorem 5.5.8 that T has a fixed point in M . Therefore, by Theorem 7.1.13,
{xn } converges strongly to an element of F (T ).

7.2 Convergence of approximants of non-self

mappings
In this section, we discuss strong convergence of approximants of non-self non-
expansive mappings.
The following theorem is an extension of Browder’s strong convergence theorem
for non-self nonexpansive mappings with unbounded domain.

Theorem 7.2.1 (Singh and Watson’s convergence theorem) – Let C

be a nonempty closed convex subset of a Hilbert space H and T : C → H a
nonexpansive mapping such that T (∂C) ⊆ C and T (C) is bounded. Let u be an
element in C and define Gt : C → H by

Gt x = (1 − t)u + tT x, x ∈ C and t ∈ (0, 1).

7.2. Approximants of non-self mappings 325

Let xt = Gt xt . Then {xt } converges strongly to v as t → 1, where v is the fixed

point of T closest to u.
Proof. Note F (T ) is nonempty by Theorem 5.2.25. Then for any y ∈ F (T ),
we have
xt − y ≤ u − y for all t ∈ (0, 1),
so {xt } is bounded. By boundedness of {T xt }, we obtain that

xt − T xt  ≤ (1 − t) sup u − T xt  → 0 as t → 1.
t∈(0,1)

Because H is reflexive, {xt } has a weakly convergent subsequence. Let {xtn }

be subsequence of {xt } such that xtn  z as tn → 1. Write xn = xtn . Because
I − T is demiclosed at zero, z ∈ F (T ). Because F (T ) is a nonempty closed
convex set in C by Corollary 5.2.29, there exists a unique point v ∈ F (T ) that
is closest to u, i.e., v ∈ F (T ) is the nearest point projection of u. Now, for
y ∈ F (T ), we have

xt − u + t(u − y)2 = t2 T xt − y2

≤ t2 xt − y2 = t2 xt − u + u − y2

and hence

xt − u2 + t2 u − y2 + 2txt − u, u − y = xt − u + t(u − y)2

≤ t2 (xt − u2 + u − y2 + 2xt − u, u − y).

It follows that
2t
xt − u2 ≤ xt − u, y − u ≤ xt − u, y − u ≤ xt − u · y − u.
1+t
Hence xt − u ≤ y − u. By w-lsc of the norm of H,

z − u ≤ lim inf xn − u ≤ y − u for all y ∈ F (T ).

n→∞

But v is the nearest point projection of u. Therefore, z = v is the unique element

in F (T ) that is the nearest point projection of u. This shows that xn  v as
n → ∞. It remains to show that the convergence is strong. Because

xn − u2 = xn − v + v − u2 = xn − v2 + u − v2 + 2xn − v, v − u,

this implies that

xn − v2 = xn − u2 − u − v2 − 2xn − v, v − u

≤ −2xn − v, v − u → 0 as n → ∞.

Therefore, {xt } converges strongly to v.

We now establish a strong convergence theorem for non-self mappings in a

Banach space.
326 7. Strong Convergence Theorems

Theorem 7.2.2 Let X be a uniformly convex Banach space with a uniformly

Gâteaux differentiable norm, C a nonempty closed convex subset of X, u an
element in C, and T : C → X a weakly inward nonexpansive mapping with
F (T ) = ∅. Suppose for t ∈ (0, 1), the contraction Gt : C → X defined by

Gt x = (1 − t)u + tT x, x ∈ C (7.10)

has a unique fixed point xt ∈ C. Then {xt } converges strongly to a fixed point
of T as t → 1.

Proof. Because F (T ) is nonempty, then {xt } is bounded. In fact, we have

xt − v ≤ u − v for all v ∈ F (T ) and t ∈ (0, 1).

We now show that {xt } converges strongly to a fixed point of T as t → 1 . To this

end, let {tn } be a sequence of real numbers in (0,1) such that tn → 1 as n → ∞.
Set xn := xtn . Then we can define ϕ : C → [0, ∞) by ϕ(x) = LIMn xn − x2 .
Then the set M defined by (2.32) is a nonempty closed convex bounded subset
of C. Because

xn − T xn  = (1 − tn )T xn − u → 0 as n → ∞, (7.11)

it follows that for x ∈ M

ϕ(T x) = LIMn xn − T x2

≤ LIMn T xn − T x2
≤ LIMn xn − x2 = ϕ(x). (7.12)

By Theorem 2.9.11, M consists of one point, say z. We now show that this z
is a fixed point of T . Because T is weakly inward, there are some vn ∈ C and
λn ≥ 0 such that

wn := z + λn (vn − z) → T z strongly.

If λn ≤ 1 for infinitely many n and for these n, then we have wn ∈ C and hence
T z ∈ C. Thus, we have T z = z by (7.12). So, we may assume λn > 1 for all
sufficiently large n. We then write vn = rn wn + (1 − rn )z, where rn = λ−1 n .
Suppose rn → 1. Then vn → T z and hence T z ∈ C. By (7.12), we have T z = z.
So, without loss of generality, we may assume rn ≤ a < 1. By Theorem 2.8.17,
there exists a continuous increasing function g = gr : [0, ∞) → [0, ∞) with
g(0) = 0 such that

λx + (1 − λ)y2 ≤ λx2 + (1 − λ)y2 − λ(1 − λ)g(x − y),

for all x, y ∈ Br [0] and λ ∈ [0, 1], where Br [0] (the closed ball centered at 0 and
with radius r) is big enough so that Br [0] contains z and {wn }. It follows that

ϕ(λx + (1 − λ)y) ≤ λϕ(x) + (1 − λ)ϕ(y) − λ(1 − λ)g(x − y)

7.3. Convergence of Halpern iteration process 327

for all x, y ∈ Br [0] and λ ∈ [0, 1]. Because vn ∈ C, we obtain that

ϕ(z) ≤ ϕ(vn )
≤ rn ϕ(wn ) + (1 − rn )ϕ(z) − rn (1 − rn )g(wn − z)

and hence

(1 − a)g(wn − z) ≤ (1 − rn )g(wn − z)

≤ ϕ(wn ) − ϕ(z).

Taking the limit as n → ∞, we obtain

(1 − a)g(T z − z) ≤ ϕ(T z) − ϕ(z)

≤ 0. (by (7.12))

Therefore, T z = z, i.e., z is a fixed point of T . Observe that

(i) xn − T xn → 0 by (7.11),
(ii) (7.7) is satisfied with αn = 0,
(iii) the set M contains a fixed point z of T.
By Corollary 7.1.11, we conclude that {xt } converges strongly to z as
t → 1.

7.3 Convergence of Halpern iteration process

In Chapter 6, we have seen that the Mann and S-iteration processes are weakly
convergent for nonexpansive mappings even in uniformly convex Banach spaces.
The purpose of this section is to develop an iteration process so that it can
generate a strongly convergent sequence in a Banach space.

Definition 7.3.1 Let C be a nonempty convex subset of a linear space X and

T : C → C a mapping. Let u ∈ C and {αn } a sequence in [0, 1]. Then a
sequence {xn } in C defined by

x0 ∈ C
(7.13)
xn+1 = αn u + (1 − αn )T xn , n ≥ 0

is called the Halpern iteration.

We now prove the main convergence theorem of this section.

Theorem 7.3.2 Let X be a Banach space with a uniformly Gâteaux differ-
entiable norm, C a nonempty closed convex subset of X, and T : C → C a
nonexpansive mapping with F (T ) = ∅. Let u ∈ C and {αn } be a sequence of
real numbers in [0, 1] that satisfies
∞
 ∞

lim αn = 0, αn = ∞ and |αn+1 − αn | < ∞. (7.14)
n→∞
n=0 n=0
328 7. Strong Convergence Theorems

Suppose that {zt } converges strongly to z ∈ F (T ) as t → 1, where for t ∈

(0, 1), zt is a unique element of C that satisfies zt = (1 − t)u + tT zt . Then the
sequence {xn } defined by (7.13) converges strongly to z.
Proof. Because F (T ) = ∅, it follows that {xn } and {T xn } are bounded. Set
K := sup{u + T xn  : n ∈ N}. From (7.13), we have

xn+1 − xn  ≤ |αn − αn−1 |(u + T xn−1 ) + (1 − αn )xn − xn−1 

≤ |αn − αn−1 |K + (1 − αn )xn − xn−1 .

Hence for m, n ∈ N, we have

xn+m+1 − xn+m 
 n+m−1
   n+m−1 
≤ |αk+1 − αk | K + (1 − αk+1 ) xm+1 − xm 
k=m k=m
 n+m−1
   n+m−1
 
≤ |αk+1 − αk | K + exp − αk+1 xm+1 − xm .
k=m k=m

∞
So the boundedness of {xn } and k=0 αk = ∞ yield

∞ 
lim sup xn+1 − xn  = lim sup xn+m+1 − xn+m  ≤ |αk+1 − αk | K
n→∞ n→∞
k=m

∞
for all m ∈ N. Because k=0 |αk+1 − αk | < ∞, we get
lim xn+1 − xn  = 0. Notice
n→∞

xn − T xn  ≤ xn − xn+1  + xn+1 − T xn 

≤ xn − xn+1  + αn u − T xn  → 0 as n → ∞.

Let LIM be a Banach limit. Then, we get

LIMn xn − T zt 2 ≤ LIMn xn − zt 2 .

Because t(xn − T zt ) = (xn − zt ) − (1 − t)(xn − u), we have

t2 xn − T zt 2 ≥ xn − zt 2 − 2(1 − t)xn − u, J(xn − zt )

= (2t − 1)xn − zt 2 + 2(1 − t)u − zt , J(xn − zt )

for all n ∈ N. These inequalities yield

1−t
LIMn xn − zt 2 ≥ LIMn u − zt , J(xn − zt ).
2
Letting t go to 1, we get

0 ≥ LIMn u − z, J(xn − z),

7.3. Convergence of Halpern iteration process 329

because X has uniformly Gâteaux differentiable norm. Because xn − xn+1  →

0 as n → ∞, we obtain

lim |u − z, J(xn+1 − z) − u − z, J(xn − z)| = 0.

n→∞

Hence by Proposition 2.9.7, we obtain

lim supu − z, J(xn − z) ≤ 0. (7.15)

n→∞

Because (1 − αn )(T xn − z) = (xn+1 − z) − αn (u − z), we have

(1 − αn )(T xn − z)2 ≥ xn+1 − z2 − 2αn u − z, J(xn+1 − z),

it follows that

xn+1 − z2 ≤ (1 − αn )xn − z2 + 2(1 − (1 − αn ))u − z, J(xn+1 − z)

for each n ∈ N. Let ε > 0. From (7.15), there exists n0 ∈ N such that

u − z, J(xn − z) ≤ ε/2 for all n ≥ n0 .

Then we have
 n+n0 −1   n+n0 −1 
xn+n0 − z ≤ 2
(1 − αk ) xn0 − z + 1 −
2
(1 − αk ) ε
k=n0 k=n0

∞
for all n ∈ N. By the condition k=0 αk = ∞, we have

lim sup xn − z2 = lim sup xn+n0 − z2 ≤ ε.

n→∞ n→∞

Therefore, {xn } converges strongly to z, because ε is an arbitrary positive real

number.
Corollary 7.3.3 Let C be a nonempty closed convex subset of a uniformly
smooth Banach space and T : C → C a nonexpansive mapping with F (T ) = ∅.
Let u ∈ C and {αn } a sequence of real numbers in [0, 1] satisfying (7.14).
Then the sequence {xn } defined by (7.13) converges strongly to a fixed point
of T .

Bibliographic Notes and Remarks

The main results presented in Section 7.1 are proved in Browder [29], Lim and
Xu [98], Morales and Jung [114], Reich [126], and Takahashi and Ueda [158].
Theorem 7.2.1 is due to Singh and Watson [149]. The strong convergence of
approximants of nonexpansive non-self mappings can be found in Jung and Kim
[78], Xu [165], and Xu and Yin [168]. Theorem 7.3.2 follows from Shioji and
Takahashi [144]. Such strong convergence results have been recently generalized
by viscosity approximation method (see Moudafi [111], Xu [167]).
330 7. Strong Convergence Theorems

Exercises

7.1 Let C be a nonempty closed convex subset of a Hilbert space H. Let

T : C → C be a nonexpansive mapping and f : C → C a contraction
mapping. Let {xn } be the sequence defined by the scheme

1 εn
xn = T xn + f xn ,
1 + εn 1 + εn

where εn is a sequence (0, 1) with εn → 0. Show that {xn } converges

strongly to the unique solution of the variational inequality:

(I − f )x̃, x̃ − x ≤ 0 for all x ∈ F (T ).

7.2 Let H be a Hilbert space, C a closed convex subset of H, T : C → C a

nonexpansive mapping with F (T ) = ∅, and f : C → C a contraction. Let
{xn } be a sequence in C defined by

x0 ∈ C,
xn+1 = (1 − αn )T xn + αn f (xn ), n ≥ 0,

where {αn } is a sequence in (0, 1) satisfies

(H1) αn → 0;
∞
(H2) n=0 αn = ∞;
∞
(H3) either n=0 |αn+1 − αn | < ∞ or lim (αn+1 /αn ) = 1.
n→∞

Show that under the hypotheses (H1) ∼ (H3), xn → x̃, where x̃ is the
unique solution of the variational inequality:

(I − f )x̃, x̃ − x ≤ 0 for all x ∈ F (T ).

7.3 Let C be a nonempty closed convex subset of a uniformly smooth Banach

space X and T : C → C a nonexpansive mapping with F (T ) = ∅. If ΠC
is the set of all contractions on C, show that the path {xt } defined by

xt = tf xt + (1 − t)T xt , t ∈ (0, 1), f ∈ ΠC ,

converges strongly to a point in F (T ). If we define Q : ΠC → F (T ) by

Q(f ) = lim xt , f ∈ ΠC ,
t→0+

show that Q(f ) solves the variational inequality:

(I − f )Q(f ), J(Q(f ) − v) ≤ 0, f ∈ ΠC and v ∈ F (T ).

7.3. Convergence of Halpern iteration process 331

7.4 Let C be a nonempty closed convex subset of a Banach space X. Let

A : C → C be a continuous strongly pseudocontractive with constant
k ∈ [0, 1) and T : C → C a continuous pseudocontractive mapping. Show
that

(a) for each t ∈ (0, 1), there exists unique solution xt ∈ C of equation

x = tAx + (1 − t)T x.

(b) Moreover, if v is a fixed point of T , then for each t ∈ (0, 1), there
exists j(xt − v) ∈ J(xt − v) such that

xt − Axt , j(xt − v) ≤ 0;

(c) {xt } is bounded.

7.5 Let C be a nonempty closed convex subset of a Banach space X that has
a uniformly Gâteaux differentiable norm and T : C → C a nonexpansive
mapping with F (T ) = ∅. For a fixed δ ∈ (0, 1), define S : C → C by

Sx := (1 − δ)x + δT x

for all x ∈ C. Assume that {zt } converges strongly to a fixed point z of

T as t → 0, where zt is the unique element of C that satisfies

zt = tu + (1 − t)T zt

for arbitrary u ∈ C. Let {αn } be a real sequence in (0, 1) that satisfies

the following conditions:
(i) lim αn = 0;
n→∞
∞
(ii) n=1 αn = ∞.

For arbitrary x0 ∈ C, let the sequence {xn } be defined iteratively by

xn+1 = αn u + (1 − αn )Sxn .

Show that {xn } converges strongly to a fixed point of T .

Chapter 8

Applications of Fixed Point

Theorems

The aim of this chapter is to sketch applications of fixed point theorems in

attractors of hyperbolic IFS, best approximation theory, operator equations,
variational inequalities, and variational inclusions.

8.1 Attractors of the IFS

Definition 8.1.1 A (hyperbolic) iterated function system consists of a com-
plete metric space (X, d) together with a finite set of contraction mappings Tn :
X → X, with respective Lipschitz constants kn , for n = 1, 2, · · · , N .

The abbreviation “IFS” is used for “iterated function system”. The notation
for the IFS is {X; Tn , n = 1, 2, · · · , N } and its Lipschitz constant is k =
max{kn : n = 1, 2, · · · , N }.

Let {X; Tn , n = 1, 2, · · · , N } be a hyperbolic iterated function system with

Lipschitz constant k. Then the fixed point of the mapping W : K(X) → K(X)
defined by
W (B) = ∪N n=1 Tn (B) for all B ∈ K(X),

is called the attractor of the IFS .

The following propositions tell us how to make a contraction mapping on

(K(X), H) out of a contraction mapping on a metric space (X, d).

Proposition 8.1.2 Let (X, d) be a metric space and T : X → X a continuous

mapping. Then T maps K(X) into itself.

Proof. Let C be a nonempty compact subset of X. Then clearly T (C) =

{T x : x ∈ C} is nonempty. We show that T (C) is compact. Let {yn = T xn }
be a sequence in C. Because C is compact, there is a subsequence {xni } that

R.P. Agarwal et al., Fixed Point Theory for Lipschitzian-type Mappings with Applications,
Topological Fixed Point Theory and Its Applications 6, DOI 10.1007/978-0-387-75818-3 8,

c Springer Science+Business Media, LLC, 2009
334 8. Applications of Fixed Point Theorems

converges to a point in x̂ ∈ C. The continuity of T implies that {yni = T xni }

is a subsequence of {yn } that converges to ŷ = T x̂ ∈ T (C).
Proposition 8.1.3 Let T : X → X be a contraction mapping on the metric
space (X, d) with Lipschitz constant k. Then T : K(X) → K(X) defined by

T (B) = {T x : x ∈ B} for all B ∈ K(X)

is a contraction mapping on (K(X), H) with Lipschitz constant k.

Proof. Because T : X → X is continuous, it follows from Proposition 8.1.2
that T maps K(X) into itself. Now let B, C ∈ K(X). Then

δ(T (B), T (C)) = sup{inf{d(T x, T y) : y ∈ C} : x ∈ B}

≤ sup{inf{kd(x, y) : y ∈ C} : x ∈ B} = k · δ(B, C).

Similarly, δ(T (C), T (B)) ≤ k · δ(C, B). Hence

H(T (B), T (C) = max{δ(T (B), T (C), δ(T (C), T (B))}

≤ k max{δ(B, C), δ(C, B)}
≤ k · H(B, C).

The next proposition provides an important method for combining contrac-

tion mappings on (K(X), H) to produce new contraction mappings on (K(X), H).

Proposition 8.1.4 Let (X, d) be a metric space. Let {Tn : n = 1, 2, · · · , N }

be contraction mappings on (K(X), H). Let the Lipschitz constant for Tn be
denoted by kn for each n. Define W : K(X) → K(X) by

W (B) = T1 (B) ∪ T2 (B) ∪ · · · ∪ TN (B) = ∪N

n=1 Tn (B) for all B ∈ K(X).

Then W is a contraction mapping with Lipschitz constant k = max{kn : n =

1, 2, · · · , N }.
Proof. We show the claim for N = 2. An inductive argument then completes
the proof. Let B, C ∈ K(X). We have

H(W (B), W (C)) = H(T1 (B) ∪ T2 (B), T1 (C) ∪ T2 (C))

≤ max{H(T1 (B), T1 (C)), H(T2 (B), T2 (C))}
(by Proposition 4.2.5)
≤ max{k1 H(B, C), k2 H(B, C)} ≤ k H(B, C).

The following theorem gives the main result concerning a hyperbolic IFS.
Theorem 8.1.5 Let {X; Tn , n = 1, 2, · · · , N } be a hyperbolic iterated function
system with Lipschitz constant k. Then the mapping W : K(X) → K(X) defined
by
W (B) = ∪N n=1 Tn (B) for all B ∈ K(X),
8.2. Best approximation theory 335

is a contraction mapping on the complete metric space (K(X), H) with Lipschitz

constant k, i.e.,

H(W (B), W (C)) ≤ kH(B, C) for all B, C ∈ K(X).

Its unique fixed point A ∈ K(X) satisfies

A = W (A) = ∪N
n=1 Tn (A)

and is given by A = lim W n (B) for any B ∈ K(X).

n→∞

Proof. By Proposition 8.1.4, W : K(X) → K(X) is a contraction mapping.

Hence the result follows from the Banach contraction principle.

8.2 Best approximation theory

Recall that when C is a nonempty subset of a normed space X, the set of best
approximation to x ∈ X from C is

PC (x) = {y ∈ C : x − y = d(x, C)},

where PC is the metric projection from X into 2C .

We begin with the fundamental result concerning invariance best approxi-

mation.

Theorem 8.2.1 (Brosowski’s theorem) – Let X be a Banach space and T :

X → X a nonexpansive mapping with a fixed point x ∈ X. Let C be a nonempty
subset of X such that T (C) ⊆ C. Suppose PC (x) is a nonempty compact convex
subset of C. Then T has a fixed point in PC (x).

Proof. First, we show that T : PC (x) → PC (x). Let y ∈ PC (x). Then

T y − x ≤ y − x = inf x − z
z∈C

implying that T y ∈ PC (x).

Let {an } be a sequence in (0, 1) such that an → 1. Define Tn : PC (x) →
PC (x) by
Tn x = (1 − an )u + an T x for all x ∈ PC (x).
By the Banach contraction principle, each Tn has a unique fixed point, say
xn . Because PC (x) is compact, there exists a subsequence {xni } of {xn } such
that xni → z ∈ PC (x). Because xn − T xn  → as n → ∞, it follows that
z ∈ F (T ).

The following result is an improvement of Brosowski’s theorem.

336 8. Applications of Fixed Point Theorems

Theorem 8.2.2 Let X be a Banach space and T : X → X a nonexpansive

mapping with a fixed point x ∈ X. Let C be a nonempty subset of X such
that T (∂C) ⊆ C. Suppose PC (x) is weakly compact and star-shaped. Assume
that
(i) T is nonexpansive on PC (x),
(ii) T x − T x ≤ x − x for all x ∈ PC (x),
(iii) I − T is demiclosed on PC (x).
Then PC (x) ∩ F (T ) = ∅.
Proof. First, we show that T is a self-mapping on PC (x). Let y ∈ PC (x).
Theorem 2.10.10 implies that y ∈ ∂C. As T (∂C) ⊆ C, so T y ∈ C. Because
T x = x and T is nonexpansive, we have

T y − x = T y − T x ≤ y − x = inf x − z.
z∈C

Because T y ∈ C, it follows that T y ∈ PC (x). Hence T is a self-mapping on

PC (x). Now, let p be the star-center of PC (x) and {tn } a sequence of real
numbers in (0, 1) with tn → 1. Define Tn : PC (x) → PC (x) by

Tn x = (1 − tn )p + tn T x, x ∈ PC (x).

For each n ∈ N, Tn is a contraction, so there exists exactly one fixed point xn

of Tn .
Now

xn − T xn  = (1 − tn )T xn − p ≤ (1 − tn )K → 0 as n → ∞

for some K ≥ 0. Because PC (x) is weakly compact, there exists a subsequence

{xni } of {xn } such that xni  z ∈ PC (x). Because I − T is demiclosed on
PC (x) and xni − T xni → 0 as i → ∞, it follows that z ∈ F (T ). Therefore,
PC (x) ∩ F (T ) = ∅.

8.3 Solutions of operator equations

In this section, we study applications of fixed point theorems to solutions of
operator equations in Banach spaces.
We give an application of the Browder-Göhde-Kirk’s fixed point theorem for
the existence of solutions of the operator equation x − T x = f .
Theorem 8.3.1 Let X be a uniformly convex Banach space X, f an element
in X, and T : X → X a nonexpansive mapping. Then the operator equation

x − Tx = f (8.1)

has a solution x if and only if for any x0 ∈ X, the sequence of Picard iterates
{xn } in X defined by xn+1 = T xn + f, n ∈ N0 is bounded.
8.3. Solutions of operator equations 337

Proof. Let Tf be the mapping from X into X given by

Tf (u) = T u + f.

Then u is a solution of (8.1) if and only if u is a fixed point of Tf . It is easy to

see that Tf is nonexpansive. Suppose Tf has a fixed point u ∈ X. Then

xn+1 − u ≤ xn − u for all n ∈ N.

Hence {xn } is bounded.

Conversely, suppose that {xn } is bounded. Let d = diam({xn }) and

Bd [x] = {y ∈ X : x − y ≤ d} for each x ∈ X.

Set Cn := ∩i≥n Bd [xi ]. Then Cn is a nonempty convex set for each n, and
Tf (Cn ) ⊂ Cn+1 . Let C be the closure of the union of Cn for n ∈ N, i.e.,
C = ∪n∈N Cn . Because Cn increases with n, C is closed convex bounded subset
of X. Because Tf maps C into itself, Tf has a fixed point in C by the Browder-
Göhde-Kirk theorem.

Next we show that if T is nonexpansive, then the Mann iteration converges

strongly to the solution of the operator equation x + T x = f .

Theorem 8.3.2 Let H be a Hilbert space and T : H → H a monotonic non-

expansive operator. For f ∈ H, define S : H → H by

Sx = −T x + f, x ∈ H. (8.2)

Then we have the following:

(a) The Mann iteration {xn } defined by
∞

xn+1 = M (xn , αn , S) with αn ∈ [0, 1], αn (1 − αn ) = ∞,
n=1

converges strongly to the unique solution x = v of the operator equation

x + T x = f. (8.3)

(b) If αn = (n + 1)−1 , then xn+1 − v ≤ (n + 1)−1/2 x1 − v for all n ∈ N.

Proof. From (8.2), we have

Sx − Sy = T x − T y ≤ x − y for all x, y ∈ H.

The solution v of operator equation (8.3) is a fixed point of the nonexpansive

operator S. By the monotonicity of T ,

Sx − Sy, x − y = −T x − T y, x − y ≤ 0 for all x, y ∈ H.

338 8. Applications of Fixed Point Theorems

Hence

xn+1 − v2 = (1 − αn )(xn − v) + αn (Sxn − Sv)2

= (1 − αn )2 xn − v2 + 2αn (1 − αn )Sxn − Sv, xn − v
+αn2 Sxn − Sv2
≤ (1 − αn )2 xn − v2 + αn2 Sxn − Sv2 (as αn (1 − αn ) ≥ 0)
≤ (1 − αn )2 xn − v2 + αn2 xn − v2 , (8.4)

which implies that

xn+1 − v2 ≤ [1 − 2αn (1 − αn )]xn − v2

n
≤ [1 − 2αi (1 − αi )]x1 − v2 . (8.5)
i=1

∞
Because i=1 αi (1 − αi ) = ∞, it follows that
n  
n 
[1 − 2αi (1 − αi )] ≤ exp − 2 αi (1 − αi ) → 0 as n → ∞.
i=1 i=1

Hence from (8.5), {xn } converges strongly to the unique solution of the
operator equation (8.3).
(b) Note

xn+1 − v = (1 − αn )(xn − v) + αn (Sxn − v)

≤ (1 − αn )xn − v + αn Sxn − v
≤ xn − v ≤ · · · ≤ x1 − v for all n ∈ N.

Hence from (8.4), we have

 2  2
n 1
xn+1 − v2 ≤ xn − v2 + xn − v2 ,
n+1 n+1

which gives that

(n + 1)2 xn+1 − v2 − n2 xn − v2 ≤ x1 − v2 .

Summing from n = 1 to m, we get

(m + 1)2 xm+1 − v2 − x1 − v2 ≤ mx1 − v2 .

Hence
1
xm+1 − v ≤ √ x1 − v for all m ∈ N.
m+1
8.4. Differential and integral equations 339

8.4 Differential and integral equations

Let f (x, y) be a continuous real-valued function on [a, b] × [c, d]. The Cauchy
initial value problem is to find a continuous differentiable function y on [a, b]
satisfying the differential equation
dy
= f (x, y), y(x0 ) = y0 . (8.6)
dx
Consider the Banach space C[a, b] of continuous real-valued functions with
supremum norm defined by y = sup{|y(x)| : x ∈ [a, b]}.
Integrating (8.6), we obtain an integral equation
x
y(x) = y0 + f (t, y(t))dt. (8.7)
x0

The problem (8.6) is equivalent the problem solving the integral equation
(8.7).
We define an integral operator T : C[a, b] → C[a, b] by
x
T y(x) = y0 + f (t, y(t))dt.
x0

Thus, a solution of Cauchy initial value problem (8.6) corresponds with a

fixed point of T . One may easily check that if T is contraction, then the problem
(8.6) has a unique solution.

Now our purpose is to impose certain conditions on f under which the inte-
gral operator T is Lipschitzian with σ(T ) < 1.
Theorem 8.4.1 Let f (x, y) be a continuous function of Dom(f ) = [a, b] × [c, d]
such that f is Lipschitzian with respect to y, i.e., there exists L > 0 such that
|f (x, u) − f (x, v)| ≤ L|u − v| f or all u, v ∈ [c, d] and f or x ∈ [a, b].
Suppose (x0 , y0 ) ∈ int(Dom(f )). Then for sufficiently small h > 0, there exists
a unique solution of the problem (8.6).
Proof. Let M = sup{|f (x, y)| : x, y ∈ Dom(f )} and choose h > 0 such that
Lh < 1 and [x0 − h, x0 + h] ⊆ [a, b]. Set
C := {y ∈ C[x0 − h, x0 + h] : |y(x) − y0 | ≤ M h}.
Then C is a closed subset of the complete metric space C[x0 − h, x0 + h] and
hence C is complete. Note T : C → C is a contraction mapping. Indeed, for
x ∈ [x0 − h, x0 + h] and two continuous functions y1 , y2 ∈ C, we have
 x 
 
T y1 − T y2  =   f (x, y1 ) − f (x, y2 )dt 

x0
≤ |x − x0 | sup L|y1 (s) − y2 (s)|
s∈[x0 −h,x0 +h]

≤ Lhy1 − y2 .
340 8. Applications of Fixed Point Theorems

Therefore, T has a unique fixed point implying that the problem (8.6) has a
unique solution.

Now, consider the Fredholm integral equation for an unknown function y :

[a, b] → R (−∞ < a < b < ∞):

b
y(x) = f (x) + λ k(x, t)y(t)dt, (8.8)
a

where
k(x, t) is continuous on [a, b] × [a, b]

and
f (x) is continuous on [a, b].

Consider the Banach space X = C[a, b] of continuous real-valued functions

with supremum norm y = sup{|y(x)| : x ∈ [a, b]} and define an operator
T : C[a, b] → C[a, b] by

b
T y(x) = f (x) + λ k(x, t)y(t)dt. (8.9)
a

Thus, a solution of Fredholm integral equation (8.8) is a fixed point of T .

We now impose a restriction on the real number λ such that T becomes a

contraction.

Theorem 8.4.2 Let K(x, t) be a continuous function on [a, b] × [a, b] with M =

sup{|k(x, t)| : x, t ∈ [a, b]}, f a continuous function on [a, b], and λ a real number
such that M (b − a)|λ| < 1. Then the Fredholm integral equation (8.8) has a
unique solution.

Proof. It is sufficient to show that the mapping T defined by (8.9) is a con-

traction. For two continuous functions y1 , y2 ∈ C[a, b], we have

b
T y1 − T y2  = sup |λ|| k(x, t)[y1 (t) − y2 (t)]dt|
x∈[a,b] a
b
≤ |λ| sup |k(x, t)||y1 (t) − y2 (t)|dt
x∈[a,b] a
b
≤ |λ|M sup |y1 (t) − y2 (t)|dt
a t∈[a,b]
b
= |λ|M y1 − y2  dt.
a
8.5. Variational inequality 341

8.5 Variational inequality

Let C be a convex subset of a smooth Banach space X, D a nonempty subset of
C, and A : C → C a mapping. We consider the following variational inequality
V ID (C, I − A):

to find a z ∈ D such that (I − A)z, J(z − y) ≤ 0 for all y ∈ D,

where J is the duality mapping from X into X ∗ .

The set of solutions of the variational inequality V ID (C, I − A) is denoted
by ΩD (I − A), i.e.,

ΩD (I − A) = {z ∈ C : (I − A)z, J(z − y) ≤ 0 for all y ∈ D}.

Proposition 8.5.1 Let C be a nonempty convex subset of a smooth Banach

space X. Let A : C → C be strongly pseudocontractive with constant k ∈ [0, 1).
Then variational inequality V ID (C, I − A) has at most one solution.

Proof. Let x∗ and y ∗ be two distinct solutions of V ID (C, I − A). Then

x∗ − Ax∗ , J(x∗ − y ∗ ) ≤ 0

and
y ∗ − Ay ∗ , J(y ∗ − x∗ ) ≤ 0.
Adding these inequalities, we get

x∗ − y ∗ − (Ax∗ − Ay ∗ ), J(x∗ − y ∗ ) ≤ 0,

which implies that

x∗ − y ∗ 2 ≤ Ax∗ − Ay ∗ , J(x∗ − y ∗ )

≤ kx∗ − y ∗ 2 ,

a contradiction. Therefore, x∗ = y ∗ .

Proposition 8.5.2 Let C be a nonempty convex subset of a smooth Banach

space X and D a nonempty subset of C. Let A : C → C be a mapping and
let P be the sunny nonexpansive retraction from C onto D. Then following are
equivalent:
(a) z is a fixed point of P A.
(b) z is a solution of variational inequality V ID (C, I − A).

Proof. Let x ∈ C and x0 ∈ D. Then from Proposition 2.10.21, we have

x0 = P x if and only if x0 − x, J(x0 − y) ≤ 0 for all y ∈ D. (8.10)

For z ∈ C, we obtain from (8.10) that

z = P f z if and only if z − f z, J(z − y) ≤ 0 for all y ∈ D.

342 8. Applications of Fixed Point Theorems

Theorem 8.5.3 Let X be a reflexive Banach space with a uniformly Gâteaux

differentiable norm, C a nonempty closed convex bounded subset of X, A : C →
C a continuous strongly pseudocontractive mapping with constant k ∈ [0, 1), and
T : C → X a weakly inward continuous pseudocontractive mapping. Suppose
every nonempty closed convex bounded subset of C has fixed point property for
nonexpansive self-mappings. Then we have the following:
(a) for each t ∈ (0, 1), there exists a path {xt } in C defined by
xt = (1 − t)Axt + tT xt , (8.11)

(b) {xt } converges strongly to z ∈ F (T ) as t → 1,

(c) z is the unique solution of the variational inequality V IF (T ) (C, I − A).
Proof. (a) For each t ∈ (0, 1), define the mapping Gt : C → X by

t x = (1 − t)Ax + tT x,
GA x ∈ C.
Note for each t ∈ (0, 1), GAt is weakly inward continuous strongly pseudocon-
tractive. By Corollary 5.7.14, GAt has exactly one fixed point xt in C.
(b) Because the mapping (2I − T ) has a nonexpansive inverse g, then g maps
C into itself. Note xt − T xt → 0 as t → 1 implies that xt − g(xt ) → 0 as t → 1.
Write xn = xtn . Then, we have
xn − T xn → 0 as n → ∞. (8.12)
Because {xn } is bounded, we define ϕ : C → R by ϕ(z) = LIMn +

xn − z2 , z ∈ C. As in proof of Theorem 7.1.12, the set M defined by (2.32)

is a nonempty closed convex bounded g-invariant subset of C and hence the
nonexpansive mapping g has a fixed point in M by assumption. Denote such a
fixed point by v.
On the other hand, by monotonicity of I − T , we have
xt − T xt , J(xt − y) ≥ 0 for all y ∈ F (T ).
From (8.11), we have xt − T xt = 1−t
t (Axt − xt ). Thus,
xt − Axt , J(xt − y) ≤ 0 for all y ∈ F (T ). (8.13)
Clearly, the sequence {xn } satisfies (7.7). Therefore, by Theorem 7.1.10, {xn }
converges strongly to the fixed point v.
We finally prove that the path {xt } converges strongly. Toward this end,
we assume that {tn } is another subsequence in (0, 1) such that xtn → v  as
tn → 1. By (8.12), we obtain v  ∈ F (T ). From (8.13), we have that
v − Av, J(v − v  ) ≤ 0 and v  − Av  , J(v  − v) ≤ 0.
We must have v = v  . Therefore, {xt } converges strongly to v ∈ F (T ).
(c) Because xt → v ∈ F (T ) as t → 1, it follows from (8.13) that

v − Av, J(v − y) ≤ 0 for all y ∈ F (T ).

8.6. Variational inclusion problem 343

Corollary 8.5.4 Let X be a reflexive Banach space with a uniformly Gâteaux

differentiable norm, C a nonempty closed convex subset of X, A : C → C a
contraction mapping with Lipschitz constant k ∈ [0, 1), and T : C → C a non-
expansive with F (T ) = ∅. Suppose every nonempty closed convex bounded subset
of C has fixed point property for nonexpansive self-mappings. Then conclusions
of Theorem 8.5.3 hold.

Proof. It suffices to show that the path {xt } defined by (8.11) is bounded. Let
y ∈ F (T ). Then

xt − y ≤ (1 − t)Axt − y + tT xt − y

≤ (1 − t)(Axt − Ay + Ay − y) + tT xt − y
≤ (1 − t)(kxt − y + Ay − y) + txt − y

which implies that

1
xt − y ≤ Ay − y.
1−k

Hence {xt } is bounded.

Theorem 8.5.5 Let X be a reflexive Banach space with a uniformly Gâteaux

differentiable norm, C a nonempty closed convex subset of X, and T : C → C
a nonexpansive mapping such that F (T ) = ∅. Suppose that every closed convex
bounded subset of C has fixed point property for nonexpansive self-mappings.
Then F (T ) is the sunny nonexpansive retract of C.

Proof. The proof is followed by Corollary 8.5.4.

8.6 Variational inclusion problem

In this section, we study the existence and uniqueness of solutions and the
convergence of the Mann iteration for a variational inclusion problem in a
Banach space.
Let A, T be two self-mappings defined on a Banach space X, g : X → X ∗
another mapping, and ϕ : X ∗ → (−∞, ∞] a proper convex lower semicontinuous
function. Let us consider the following variational inclusion problem:


to find an u ∈ X such that g(u) ∈ Dom(∂ϕ),
(8.14)
T u − Au − f, v − g(u) ≥ ϕ(g(u)) − ϕ(v) for all v ∈ X ∗ .

We begin with the following basic result:

Proposition 8.6.1 Let X be a reflexive Banach space. Then the following are
equivalent:
344 8. Applications of Fixed Point Theorems

(a) x∗ ∈ X is a solution of variational inclusion problem (8.14).

(b) x∗ ∈ X is a fixed point of the mapping S : X → 2X :

S(x) = f − (T x − Ax + ∂ϕ(g(x)) + x, x ∈ X.

(c) x∗ ∈ X is a solution of equation f ∈ T x − Ax + ∂ϕ(g(x)), x ∈ X.

Proof. (a) ⇒ (c). If x∗ is a solution of the variational inclusion problem (8.14),
then g(x∗ ) ∈ Dom(∂ϕ) and

T x∗ − Ax∗ − f, v − g(x∗ ) ≥ ϕ(g(x∗ )) − ϕ(v) for all v ∈ X ∗ .

By the definition of subdifferential of ϕ, it follows from the above expression

that

f + Ax∗ − T x∗ ∈ ∂ϕ(g(x∗ )). (8.15)

This implies that x∗ is a solution of equation f ∈ T x − Ax + ∂ϕ(g(x)).

(c) ⇒ (b). Adding x∗ to both sides of (8.15), we have

x∗ ∈ f − (T x∗ − Ax∗ + ∂ϕ(g(x∗ ))) + x∗ = Sx∗ . (8.16)

This implies that x∗ is a fixed point of S in X.

(b) ⇒ (a). From (8.16), we have f − (T x∗ − Ax∗ ) ∈ ∂ϕ(g(x∗ )), hence from
the definition of ∂ϕ, it follows that

ϕ(v) − ϕ(g(x∗ )) ≥ f − (T x∗ − Ax∗ ), v − g(x∗ ) for all v ∈ X ∗ ,

i.e.,
T x∗ − Ax∗ − f, v − g(x∗ ) ≥ ϕ(g(x∗ )) − ϕ(v) for all v ∈ X ∗ .

Thus, x∗ is a solution of the variational inclusion problem (8.14).

Theorem 8.6.2 Let X be a uniformly smooth Banach space and let T, A :
X → X, g : X → X ∗ be three continuous mappings. Let ϕ : X ∗ → (−∞, ∞] be
a function with a continuous Gâteaux differential ∂ϕ and satisfying the following
conditions:
(i) T − A : X → X is a strongly accretive mapping with constant k ∈ (0, 1),
(ii) ϕog : X → X is accretive.
For any given f ∈ X, define a mapping S : X → X by

Sx = f − (T x − Ax + ∂ϕ(g(x))) + x, x ∈ X. (8.17)
If S(X) is bounded, then for any given x1 ∈ X, the Mann iterative sequence
{xn } defined by

xn+1 = M (xn , αn , S), n ∈ N, (8.18)

8.6. Variational inclusion problem 345

where {αn } is the sequence in [0, 1] with the restriction

∞

αn → 0, αn = ∞, (8.19)
n=0

converges strongly to the unique solution of variational inclusion (8.14).

Proof. First we show that the variational inclusion problem (8.14) has a unique
solution.
From conditions (i) ∼ (ii) and Proposition 5.7.12, the mapping T −A+∂ϕog :
X → X is a strongly accretive continuous mapping with a strongly accretive
constant k ∈ (0, 1). Because T − A + ∂og is a continuous strongly accretive
mapping, T − A + ∂ϕog is surjective. Therefore, for any given f ∈ X, the
equation f = (T − A + ∂ϕog)(x) has a solution x∗ . Because X is reflexive, by
Proposition 8.6.1, x∗ is a solution of variational inclusion (8.14), and it is also
a fixed point of the self-mapping S defined by (8.17).
We now break the proof into the following three steps:

Step 1: x∗ is the unique solution of (8.14) in X.

Suppose, for contradiction, u∗ ∈ X is another solution of (8.14). Then, u∗
is also a fixed point of S. Hence, we have
x∗ − u∗ 2 = x∗ − u∗ , J(x∗ − u∗ )
= Sx∗ − Su∗ , J(x∗ − u∗ )
= f − (T − A + ∂ϕog)(x∗ ) + x∗
−(f − (T − A + ∂ϕog)(u∗ ) + u∗ ), J(x∗ − u∗ )
= x∗ − u∗ 2 − (T − A + ∂ϕog)(x∗ )
−(T − A + ∂ϕog)(u∗ ), J(x∗ − u∗ )
≤ x∗ − u∗ 2 − kx∗ − u∗ 2 ,
a contradiction. Hence x∗ = u∗ .

Step 2: {xn } is bounded.

Because S(X) is bounded, let
K = sup{Sx − x∗  + x1 − x∗  : x ∈ X}. (8.20)
Now we show that
xn − x∗  ≤ K for all n ∈ N. (8.21)
In fact, for n = 1 it follows from (8.20) that x1 − x∗  ≤ K. Suppose (8.21) is
true for n = k ≥ 1, then for n = k + 1, we have
xk+1 − x∗  = (1 − αk )(xk − x∗ ) + αk (Sxk − x∗ )
≤ (1 − αk )xk − x∗  + αk Sxk − x∗ 
≤ K,
346 8. Applications of Fixed Point Theorems

Hence {xn } is bounded.

Step 3: xn → x∗ .
From (8.18) and Proposition 2.4.6(b), we obtain that
xn+1 − x∗ 2 = (1 − αn )(xn − x∗ ) + αn (Sxn − x∗ )2
≤ (1 − αn )2 xn − x∗ 2 + 2αn Sxn − x∗ , J(xn+1 − x∗ )
= (1 − αn )2 xn − x∗ 2 + 2αn Sxn − x∗ , J(xn − x∗ )
+2αn Sxn − x∗ , J(xn+1 − x∗ ) − J(xn − x∗ ). (8.22)
Observe that
Sxn − x∗ , J(xn − x∗ ) = f − (T − A + ∂ϕog)(xn ) + xn
−(f − (T − A + ∂ϕog)(x∗ ) − x∗ ), J(xn − x∗ )
= xn − x∗ 2 − (T − A + ∂ϕog)(xn )
−(T − A + ∂ϕog))(x∗ ), J(xn − x∗ )
≤ (1 − k)xn − x∗ 2 .
Set βn := |Sxn − x∗ , J(xn+1 − x∗ ) − J(xn − x∗ )|. Then
βn ≤ KJ(xn+1 − x∗ ) − J(xn − x∗ )∗ .
Observe that
xn+1 − x∗ − (xn − x∗ ) = xn+1 − xn
= αn (Sxn − xn ). (8.23)
Again because {xn }, {Sxn } are bounded, and αn → 0, hence, from (8.23), we
have
xn+1 − x∗ − (xn − x∗ ) → 0 as n → ∞.
By the uniform continuity of J, J(xn+1 − x∗ ) − J(xn − x∗ )∗ → 0. Thus, we
have
βn → 0. (8.24)
So from (8.22), (8.23), and (8.24), we have
xn+1 − x∗ 2 ≤ [(1 − αn )2 + 2αn (1 − k)]xn − x∗ 2 + 2αn βn
= (1 + αn2 − 2αn k)xn − x∗ 2 + 2αn βn
= [1 − αn k + αn (αn − k)]xn − x∗ 2 + 2αn βn . (8.25)
Because αn → 0, there exists an n0 ∈ N such that αn < k for n ≥ n0 . Hence
for any n ≥ n0 , from (8.25) we have
xn+1 − x∗ 2 ≤ (1 − αn k)xn − x∗ 2 + 2αn βn . (8.26)
Let an = xn − x∗ 2 , tn = αn k, bn = 2αn βn . Therefore, by Lemma 6.1.6,
xn → x∗ as n → ∞.
8.6. Variational inclusion problem 347

Bibliographic Notes and Remarks

The results presented in Section 8.1 may be found in Barnsly [12]. Theorem 8.2.1
is due to Brosowski [24]. Theorem 8.2.2 follows from Carbone [34]. General best
approximation theory is well described in Singh, Watson, and Srivastava [150].
The results described in Section 8.3 are stated in Browder and Petryshyn
[30] and Dotson [48]. Theorems 8.4.1 and 8.4.2 can be found in standard books
(Goebel and Kirk [59], Kreyszig [94], and Zeidler [171]).
The existence and approximation of solutions of variational inclusion
problem (8.14) were studied in Chang [37].

Exercises
8.1 Let C be a closed convex cone in a Hilbert space H and T : C → H a
mapping. Show that the following are equivalent:
(a) Find x ∈ C such that

T x ∈ C ∗ (dual cone) and x, T x = 0. (C.P.)

(b) Find x ∈ C such that gx = x, where g : C → C is defined by
gx = PC (x − ρT x), and ρ > 0 is a real number.
8.2 Let C be a closed convex cone of a Hilbert space H. Show that the comple-
mentarity problem (C.P.) has a solution if and only if T (x) = PC x−T PC x
for x ∈ H has a fixed point in H. If x0 = T x0 , show that x = PC x0 is a
solution of the complementarity problem.
8.3 Let C be a nonempty compact convex subset of a normed pace X and
T : C → X a continuous mapping. Show that there exists a point u ∈ C
such that u − T u = d(T u, C).
8.4 Let K(s, t) and w(s, t) be continuous real functions on the unit square
[0, 1]2 , and let v(s) be a continuous real function on [0, 1]. Suppose that
|w(s, t1 ) − w(s, t2 )| ≤ N |t1 − t2 | for all 0 ≤ t1 , t2 , s ≤ 1.
Show that there is a unique continuous real function y(s) on [0, 1] such
that s
y(s) = v(s) + K(s, t)w(t, y(t))dt.
0

8.5 Let K(s, t, u) be a continuous function on 0 ≤ s, t ≤ 1, u ≥ 0 such that

|K(x, t, u1 ) − K(s, t, u2 )| ≤ N (s, t)|u1 − u2 |,
where N (s, t) is a continuous function satisfying
1
N (s, t)dt ≤ k < 1,
0
348 8. Applications of Fixed Point Theorems

for every 0 ≤ s ≤ 1. Show that for every y ∈ C[0, 1], there exists a unique
function y ∈ C[0, 1] such that
1
y(s) = v(s) + K(s, t, y(t))dt.
0
Appendix A

A.1 Basic inequalities

Lemma A.1.1 Let a, b ∈ R+ and 2 ≤ p < ∞. Then we have the following:
(a) ap + bp ≤ (a2 + b2 )p/2 ,
(b) (a2 + b2 )p/2 ≤ 2(p−2)/2 (ap + bp ).

Proof. We note that both the inequalities hold if either a or b is zero. So we

prove the Lemma for a = 0 and b = 0.

(a) Because
a2 b2
≤ 1 and 2 ≤ 1,
a2 +b 2 a + b2
we have
 p/2  p/2
ap bp a2 b2
+ = +
(a2 + b2 )p/2 (a2 + b2 )p/2 a2 + b2 a2 + b2
2 2
a b
≤ + 2 (since p/2 ≤ 1)
a2 + b2 a + b2
= 1

(b) It is obvious for p = 2. So, assume that p > 2. Set p := p/2 > 1 and
q  = p /(p − 1) = p/(p − 2). Then 1/p + 1/q  = 1. By Holder’s inequality, we
have
     
a2 + b2 ≤ ((a2 )p + (b2 )p )1/p (1q + 1q )1/q
= 2(p−2)/p (ap + bp )2/p ,

which implies that

(a2 + b2 )p/2 ≤ 2(p−2)/2 (ap + bp ).

350 A. Appendix

A.2 Partially ordered set

Let ≤ be a relation on a nonempty set X. Then the relation ≤ is said to be
partially ordered if it is
(i) reflexive: a ≤ a for all a ∈ X;
(ii) antisymmetric: a ≤ b and b ≤ a ⇒ a = b for some a, b ∈ X;
(iii) transitive: a ≤ b and b ≤ c ⇒ a ≤ c for some a, b, c ∈ X.
The ordered pair (X, ≤) is called a partially ordered set.

Totally ordered (or linearly ordered) set - Let (X, ≤) be a partially ordered
set. Then a subset S of X is said to be totally ordered or linearly ordered if for
all a, b ∈ S either a ≤ b or b ≤ a, i.e., all the elements of S are comparable.

Infimum and supremum of a set - Let X be a partially ordered set with

relation ≤ and let S be a nonempty subset of X. Then an element u ∈ X
(v ∈ S) is said to be an upper bound of S (a lower bound of S) if
x ≤ u for all x ∈ S (v ≤ x for all x ∈ S).
The least upper bound of S is called the supremum and it is denoted sup S.
The greatest lower bound of S is called the infimum and it is denoted by inf S.

Minimum and maximum of a set - Let S be a nonempty subset of a partially

ordered set (X, ≤). If inf S exists and belongs to S, then it is called a minimum
of S. Similarly, if sup S exists and belongs to S, then it is called a maximum of S.

Minimal and maximal elements of a set - Let (X, ≤) be a partially ordered

set. An element m ∈ X is said to be minimal if x ≤ m for x ∈ X ⇒ x = m.
Similarly, an element m ∈ X is said to maximal if m ≤ x for x ∈ X ⇒ x = m.

We now state a very useful lemma that is known as Zorn’s lemma.

Lemma A.2.1 (Zorn’s Lemma) - Let (X, ≤) be a partially ordered set in
which every chain has an upper bound. Then X has a maximal element.

A.3 Ultrapowers of Banach spaces

Let Λ denote an index set.
Definition A.3.1 Let F be a nonempty family of subsets of Λ. Then the family
F is said to be a filter on Λ if
(F1 ) F is closed under super set:
A ∈ F and A ⊆ B ⊆ Λ ⇒ B ∈ F,
(F2 ) F is closed under intersection:
A, B ∈ F ⇒ A ∩ B ∈ F.
A.3. Ultrapowers of Banach spaces 351

Observation
• 2Λ , the power set of index set Λ, defines a filter.
• A filter F is proper if F = 2Λ .

For i0 ∈ Λ, let Fi0 = {A ⊆ Λ : i0 ∈ A}. Then a filter of the form Fi0 for
some i0 is called a non-free filter.
Definition A.3.2 Let U be a filter on Λ. Then U is said to be an ultrafilter on
Λ if it is maximal with respect to ordering of filters on Λ by inclusions, i.e., if
U ⊆ F and F is a filter on Λ, then F = U.
Observation
• If U is an ultrafilter on Λ, then it is not properly contained in any other filter
on Λ.

Definition A.3.3 Let (X, τ ) be a topological space, U an ultrafilter on Λ and

{xi }i∈Λ a subset in X. We say

lim xi (≡ τ − lim xi ) = x
U U

if for every neighborhood U of x we have {i ∈ Λ : xi ∈ U } ∈ U.

Observation
• If X is a Hausdorff topological space, then the limit along U of the set {xi }i∈Λ
in X is always unique.
• If {xn } is a bounded sequence in R and U is an ultrafilter on N, then lim xn
U
exists and
lim inf xn ≤ lim xn ≤ lim sup xn .
n→∞ U n→∞

We now give basic properties of ultrafilter.

Proposition A.3.4 Let X be a Hausdorff topological vector space and U an
ultrafilter over an index set Λ. If {xi }i∈Λ and {yi }i∈Λ are two subsets of X
such that lim xi and lim yi exist,
U U

lim(xi + yi ) = lim xi + lim yi

U U U

and
lim αxi = α lim xi for any scalar α ∈ K.
U U

Proposition A.3.5 Let {xn } be a sequence in a metric space X and U an

ultrafilter (over an index set N) such that lim xn = x. Then there exists a
U
subsequence of {xn } that converges to x.

Proposition A.3.6 Let X be a Hausdorff topological space. Then X is compact

if and only if lim xi exists for all {xi }i∈I ⊂ X and any ultrafilter over Λ.
U
352 A. Appendix

Now we are in position to define an ultrapower of a Banach space:

Let X be a Banach space and U an ultrafilter over an index set Λ. Let

∞ (X) denote the space

{{yn } : yn ∈ X and {yn } ∈ ∞}

with the norm {yn }∞ (X) := sup yn  and let N be the closed subspace of
n∈N
∞ (X)
{{yn } : yn ∈ X and lim yn  = 0}.
U

Definition A.3.7 Let X be a Banach space. Then the Banach space ultrapower
of X over U is defined to be the Banach space quotient

{X}U := ∞ (X)/N .

The norm  · U in {X}U is the usual quotient norm, i.e., {xi }U = lim xi .
U

We now give some useful properties of an ultrapower of Banach spaces.

Proposition A.3.8 Let X be a Banach space. Then {X}U is a Banach space.

Proposition A.3.9 The ultrapower {X}U of a Banach space X contains a

subspace isometrically isomorphic to X.

Observation
• X is a subspace of {X}U .
• {H}U is a Hilbert space, i.e.,

{xi } + {yi }2U + {xi } − {yi }2U = 2{xi }2U + 2{yi }2U for all {xi }, {yi } ∈ {H}U .

Proposition A.3.10 The ultrapower {X}U of a Banach space X is finitely

representable in X.
Bibliography

[1] R. P. Agarwal, M. El-Gebeily and D. O’Regan, Generalized contractions in

partially ordered metric spaces, Applicable Analysis, 87 (2008), 109–116.

[2] R. P. Agarwal, D. O’Regan and D. R. Sahu, Iterative construction of

fixed points of nearly asymptotically nonexpansive mappings, J. Nonlinear
Convex Anal., 8 (1) (2007), 61–79.

[3] I. D. Arandelovic, On a fixed point theorem of Kirk, J. Math. Anal. Appl.,

301(2) (2005), 384–385.

[4] N. A. Assad and W. A. Kirk, Fixed point theorems for set-valued

mappings of contractive type, Pacific J. Math., 43 (1972), 553–562.

[5] G. Bachman and L. Narici, Functional analysis, Academic Press,

New York, 1966.

[6] J. S. Bae, Reflexivity of a Banach space with a uniformly normal structure,

Proc. Amer. Math. Soc., 90(2) (1984), 269–270.

[7] J. S. Bae and M. S. Park, Fixed point theorems in metric spaces with
uniformly normal structure, J. Korean Math. Soc., 30(1) (1993), 51–62.

[8] J. B. Baillon, S. Reich and R. E. Bruck, On the asymptotic behavior of

nonexpansive mappings and semigroups, Houston J. Math. 4 (1978), 1–9.

[9] S. Banach, Sur les operations dans les ensembles abstrait et leur applica-
tion aux equations, integrals, Fund. Math., 3(1922), 133–181.

[10] V. Barbu, Nonlinear semigroups and differential equations in Banach

spaces, Noordhoff International Publishing, Leiden, 1976.

[11] V. Barbu and Th. Precupanu, Convexity and optimization in Banach

spaces, Revised edition, Translated from the Romanian, Editura
Academiei, Bucharest; Sijthoff and Noordhoff International Publishers,
Alphen aan den Rijn, 1978.

[12] M. F. Barnsley, Fractals everywhere, Academic Press, San Diego, 1988.

354 Bibliography

[13] B. Beauzamy, Introduction to Banach spaces and their geometry, Vol. 68

of North-Holland Mathematics Studies, North-Holland, Amsterdam, 2nd
edition, 1985.

[14] T. D. Benavides, Fixed point theorems for uniformly Lipschitzian map-

pings and asymptotically regular mappings, Nonlinear Anal., Theory
Methods Appl., 32(1) (1998), 15–27.

[15] T. D. Benavides and G. L. Acedo, Lower bounds for normal structure

coefficients, Proc. R. Soc. Edinb., 121(3–4) (1992), 245–252.

[16] T. D. Benavides G. L. Acedo and H. K. Xu, Weak uniform normal struc-

ture and iterative fixed points of nonexpansive mappings, Colloq. Math.
68(1) (1995), 17–23.

[17] T. D. Benavides and H. K. Xu, A new geometrical coefficient for Banach

spaces and its applications in fixed point theory, Nonlinear Anal., Theory
Methods Appl., 25(3) (1995), 311–325.

[18] V. Berinde, Iterative approximation of fixed points, Efemeride, Baia Mare,

Romania 2002.
[19] B. Bollobas, Linear analysis an introductory course, Cambridge University
Press, Cambridge, 1990.

[20] S. C. Bose, Weak convergence to the fixed point of an asymptotically

nonexpansive map, Proc. Amer. Math. Soc., 68 (1978), 305–308.

[21] S. C. Bose and S. K. Laskar, Fixed point theorems for certain class of
mappings, J. Math. Phys. Sci., 19 (1985), 503–509.

[22] D. W. Boyd and J. S. W. Wong, On nonlinear contractions, Proc. Amer.

Math. Soc., 20(2) (1969), 458–464.

[23] M. S. Brodskii and D. P. Milman, On the center of a convex set (Russian),

Dokl. Akad. Nauk SSSR 59 (1948), 837–840.

[24] B. Brosowski, Fixpunktsa’tze in der Approximationstheorie (German)

Mathematica (Cluj), 11 (34) (1969) 195–220.

[25] L. E. J. Brouwer, Uber Abbildung von Mannigfaltigkeiten, Math. Ann. 71

(1912), 97–115.

[26] F. E. Browder, Fixed-point theorems for noncompact mappings in Hilbert

space, Proc. Natl. Acad. Sci. U.S.A. 53 (1965), 1272–1276.

[27] F. E. Browder, Nonexpansive nonlinear operators in a Banach space, Proc.

Natl. Acad. Sci., 54 (1965), 1041–1044.

[28] F. E. Browder, Nonlinear operators and nonlinear equations of evolution

in Banach spaces, Proc. Symposia Pure Math., Vol XVIII-Part 2, 1967.
Bibliography 355

[29] F. E. Browder, Convergence of approximants to fixed points of nonexpan-

sive non-linear mappings in Banach spaces, Arch. Rational Mech. Anal.,
24 (1967), 82–90.

[30] F. E. Browder and W. V. Petryshyn, The solution by iteration of nonlinear

functional equations in Banach spaces, Bull. Amer. Math. Soc., 72 (1966),
571–575.

[31] R. E. Bruck, On the almost-convergence of iterates of a nonexpansive

mapping in Hilbert space and the structure of the weak w-limit set, Israel
J. Math., 29 (1978), 1–16.

[32] R. E. Bruck, On the convex approximation property and the asymptotic

behavior of nonlinear contractions in Banach spaces, Israel J. Math., 38
(1981), 304–314.

[33] W. L. Bynum, Normal structure coefficients for Banach spaces, Pacific J.

Math., 86 (2) (1980), 427–436.

[34] A. Carbone, Some results on invariant approximation, Inter. J. Math.

Math. Sci., 17 (3) (1994), 483–487.

[35] J. Caristi, Fixed point theorems for mappings satisfying inwardness con-
ditions, Trans. Amer. Math. Soc., 215 (1976), 241–251.

[36] E. Casini and E. Maluta, Fixed points of uniformly Lipschitzian map-

pings in spaces with uniformly normal structure, Nonlinear Anal., Theory
Methods Appl., 9(1) (1985), 103–108.

[37] S. S. Chang, The Mann and Ishikawa iterative ipproximation of solutions

to variational inclusions with accretive type mappings, Comput. Math.
Appl. 37 (1999) 17–24.

[38] S. S. Chang, Some results for asymptotically pseudocontractive mappings

and asymptotically nonexpansive mappings, Proc. Amer. Math. Soc.,
129(3) (2000), 845–853.

[39] C. E. Chidume and S. A. Mutangadura, An example on the Mann iteration

method for Lipschitz pseudocontractions, Proc. Amer. Math. Soc., 129
(2001) 2359–2363.

[40] C. E. Chidume, N. Shahzad and H. Zegeye, Convergence theorems for

mappings which are asymptotically nonexpansive in the intermediate
sense, Numer. Funct. Anal. Optim., 25(3–4) (2003), 239–257.

[41] L. Cioranescu, Geometry of Banach spaces, duality mappings and non-

linear problems, Kluwer Academic Publishers, Dordrecht, 1990.

[42] James A. Clarkson, Uniformly convex spaces, Trans. Amer. Math. Soc.,
40 (3) (1936), 396–414.
356 Bibliography

[43] T. Dalby, Relationships between properties that imply the weak fixed
point property, J. Math. Anal. Appl., 253(2) (2001), 578–589.

[44] K. Deimling, Ordinary differential equations in Banach spaces (Lecture

notes in mathematics), Springer-Verlag, Berlin, 1977.

[45] K. Deimling, Nonlinear functional analysis, Springer-Verlag, Berlin, 1985.

[46] L. Deng, Convergence of the Ishikawa iteration process for nonexpansive

mappings, J. Math. Anal. Appl., 199(3) (1996), 769–775.

[47] F. Deutsch, Best approximation in inner product spaces, Springer-Verlag,

Berlin, 2001.

[48] W. G. Dotson, Jr., An iterative process for nonlinear monotonic non-

expansive operators in Hilbert space, Math. Computation, 32(141) (1978),
223–225.

[49] D. Downing and W. A. Kirk, Fixed point theorems for set-valued map-
pings in metric and Banach spaces, Math. Japonica, 22 (1977), 99–112.

[50] E. Lami Dozo, Multivalued nonexpansive mappings and Opial’s condition,

Proc. Amer. Math. Soc., 38(2) (1973), 286–292.

[51] M. Edelstein, A remark on a theorem of Krasnoselskii, Amer. Math. Mon.,

73 (1966), 509–510.

[52] M. Edelstein, The construction of an asymptotic center with a fixed-point

property, Bull. Amer. Math. Soc., 78 (2), 1972, 206–208.

[53] J. G. Falset and E. L. Fuster, Normal structure and fixed point property,
Glasg. Math. J., 38(1) (1996), 29–37.

[54] J. G. Falset, B. Sims and M. A. Smyth, The demiclosedness principle for

mappings of asymptotically nonexpansive type, Houston J. Math. 22(1)
(1996), 101–108.

[55] J. G. Falset, A. J. Melado and E. L. Fuster, Isomorphically expansive

mappings in 2 , Proc. Amer. Math. Soc., 125(9) (1997), 2633–2636.

[56] P. J. S. G. Ferreira, The existence and uniqueness of the minimum norm

solution to certain linear and nonlinear problems, Signal Processing, 55(1)
(1996), 137–139.

[57] K. Goebel and W. A. Kirk, A fixed point theorem for asymptotically

nonexpansive mappings, Proc. Amer. Math. Soc., 35(1) (1972), 171–174.

[58] K. Goebel and W. A. Kirk, A fixed point theorem for transformations

whose iterates have uniform Lipschitz constant, Stud. Math., 47 (1973),
135–140.
Bibliography 357

[59] K. Goebel and W. A. Kirk, Topics in metric fixed point theory, Cambridge
University Press, Cambridge, 1990.

[60] K. Goebel and S. Reich, Uniform convexity, hyperbolic geometry, and

nonexpansive mappings, Marcel Dekker, Inc., New York, 1984.

[61] D. Göhde, Zum Prinzip der kontraktiven Abbildung, (German) Math.

Nachr., 30 (1965), 251–258.

[62] J. Gornicki, Weak convergence theorems for asymptotically nonexpan-

sive mappings in uniformly convex Banach spaces, Comment Math. Univ.
Carolinae, 30(2) (1989), 249–252.

[63] C. W. Groetsch, A note on segmenting Mann iterates, J. Math. Anal.

Appl., 40 (1972), 369–372.

[64] C. W. Groetsch, A nonstationary iterative process for nonexpansive map-

pings, Proc. Amer. Math. Soc., 43(1) (1974), 155–158.

[65] K. S. Ha and J. S. Jung, Strong convergence theorems for accretive oper-

ators in Banach spaces, J. Math. Anal. Appl., 147(2) (1990), 330–339.

[66] B. Halpern, Fixed points of nonexpanding maps, Bull. Amer. Math. Soc.
73 (1967), 957–961.

[67] A. M. Harder and T. L. Hicks, Fixed point theory and iteration proce-
dures, Indian J. Pure Applied Math., 19(1) (1988), 17–26.

[68] T. L. Hicks and M. D. Humphries, A note on fixed point theorems,

J. Approx. Theorey, 34 (1982), 221–225.

[69] H. Ishihara and W. Takahashi, Modules of convexity, characteristic of

convexity and fixed point theorems, Kodai Math. J., 10 (1987), 197–208.

[70] S. Ishikawa, Fixed points by a new iteration method, Proc. Amer. Math.
Soc., 44 (1974), 147–150.

[71] S. Ishikawa, Fixed points and iteration of a nonexpansive mapping in a

Banach space, Proc. Amer. Math. Soc., 59 (1976), 65–71.

[72] V. I. Istratescu, Fixed point theory an introduction, Mathematics and its

Applications, 7, D. Reidel Publishing Co., Dordrecht-Boston, Mass., 1981.

[73] V. I. Istratescu, Strict convexity and complex strict convexity, Marcel

Dekker, Inc., New York, 1984.

[74] K. D. Joshi, Introduction to general topology, John Wiley & Sons, Inc.,
New York, 1983.

[75] M. C. Joshi and R. K. Bose, Some topics in nonlinear functional analysis,

John Wiley & Sons, Inc., New York, 1985.
358 Bibliography

[76] J. S. Jung, Convergences of nonexpansive iteration processes in Banach

spaces, J. Math. Anal. Appl., 273(1) (2002), 153–159.

[77] J. S. Jung, Y. J. Cho and B. S. Lee, Asymptotic behavior of nonexpan-

sive iterations in Banach spaces, Comm. Appl. Nonlinear Anal., 7 (2000),
63–67.

[78] J. S. Jung and T. H. Kim, Strong convergence of approximating fixed

points for nonexpansive nonself-mappings in Banach spaces, Kodai Math.
J., 21 (1998), 259–272.

[79] J. S. Jung and J. S. Park, Convergence of nonlinear semigroups and resol-

vents of their generators, Comm. Korean Math. Soc., 8(4) (1993), 629–641.

[80] J. S. Jung and J. S. Park, Asymptotic behavior of nonexpansive sequences

and mean points, Proc. Amer. Math. Soc., 124(2) (1996), 475–480.

[81] J. S. Jung and D. R. Sahu, Dual convergence of iteration processes for non-
expansive mappings in Banach spaces, Czech. Math. J., 53 (128) (2003),
397–404.

[82] W. Kaczor, Weak convergence of almost orbits of asymptotically non-

expansive commutative semigroups, J. Math. Anal. Appl., 272(2) (2002),
565–574.

[83] M. A. Khamsi, Uniform smoothness implies super-normal structure pro-

perty, Nonlinear Anal. Theory Methods Appl., 19(11) (1992), 1063–1069.

[84] M. A. Khamsi, On uniform Opial condition and uniform Kadec-Klee pro-

perty in Banach and metric spaces, Nonlinear Anal., Theory Methods
Appl., 26(10) (1996), 1733–1748.

[85] M. A. Khamsi and W. A. Kirk, An introduction to metric spaces and fixed

point theory, John Wiley & Sons, Inc., Hoboken, N.J, 2001.

[86] T. H. Kim and E. S. Kim, Iterative fixed points of non-Lipschitzian self-

mappings, Kodai Math. J., 18(2) (1995), 275–283.

[87] T. H. Kim and E. S. Kim, Fixed point theorems for non-Lipschitzian

mappings in Banach spaces, Math. Japonica, 45(1) (1997), 61–67.

[88] W. A. Kirk, A fixed point theorem for mappings which do not increase
distances, Amer. Math. Monthly, 72 (1965), 1004–1006.

[89] W. A. Kirk, Fixed points of asymptotic contractions, J. Math. Anal. Appl.,

277(2) (2003), 645–650.

[90] W. A. Kirk and R. Schoneberg, Some results on pseudocontractive map-

pings, Pacific J. Math., 71 (1977), 89–100.
Bibliography 359

[91] W. A. Kirk and B. Sims, Handbook of metric fixed point theory, Kluwer
Academic Publishers, Dordrecht, 2001.

[92] W. A. Kirk, C. M. Yanez and S. S. Shin, Asymptotically nonexpansive

mappings, Nonlinear Anal., Theory Methods Appl., 33(1) (1998), 1–12.

[93] M. A. Krasnoselskii, Two remarks on the method of successive approxi-

mations, Uspehi Mat. Nauk 10 (1955), 123–127.

[94] E. Kreyszig, Introductory functional analysis with applications, John

Wiley and Sons, New York, 1989.

[95] K. Q. Lan and J. R. L. Webb, A-properness and fixed point theorems for
dissipative type maps, Abstract Appl. Anal., 4(2) (1999), 83–100.

[96] A. T. M. Lau, N. Shioji and W. Takahashi, Existence of nonexpansive

retractions for amenable semigroups of nonexpansive mappings and non-
linear ergodic theorems in Banach spaces, J. Funct. Anal., 161(1) (1999),
62–75.

[97] E. A. Lifschitz, Fixed point theorems for operators in strongly convex

spaces (Russian) Voronez. Gos. Univ. Trudy Mat. Fak. 16 (1975), 23–28.

[98] T. C. Lim and H. K. Xu, Fixed point theorems for asymptotically non-
expansaive mappings, Nonlinear Anal., Theory Methods Appl., 22(11)
(1994), 1345–1355.

[99] T. C. Lim and H. K. Xu, Uniformly Lipschitzian mappings in metric

spaces with uniform normal structure, Nonlinear Anal., Theory Methods
Appl., 25(11) (1995), 1231–1235.

[100] B. V. Limaye, Functional analysis, New Age International (P) Limited

Publishers, New Delhi, India 1997.

[101] P. K. Lin, K. K. Tan and H. K. Xu, Demiclosedness principle and

asymptotic behavior for asymptotically nonexpansive mappings, Non-
linear Anal., Theory Methods Appl., 24(6) (1995), 929–946.

[102] D. G. Luenberger, Optimization by vector space methods, John Wiley

and Sons. Inc., New York, 1969.

[103] E. Maluta, Uniformly normal structure and related coefficients, Pacific J.

Math., 111 (1984), 357–369.

[104] W. R. Mann, Mean value methods in iteration, Proc. Amer. Math. Soc.,
4 (1953), 506–510.

[105] R. H. Martin, Differential equations on closed subsets of a Banach space,

Trans. Amer. Math. Soc., 179 (1973), 399–414.
360 Bibliography

[106] R. H. Martin, Nonlinear operators and differential equations in Banach

spaces, John Wiley and Sons, New York, 1976.
[107] St. Maruster, The solution by iteration of nonlinear equations in Hilbert
spaces, Proc. Amer. Math. Soc., 63 (1977), 69–73.
[108] J. Matkowski, Integrable solutions of functional equations, Dissertationes
Math., 127 (1975), 1–68.
[109] E. Matoušková and S. Reich, Reflexivity and approximate fixed points,
Studia Math., 159 (2003), 403–415.
[110] N. Mizoguchi and W. Takahashi, Fixed point theorems for multivalued
mappings on complete metric spaces, J. Math. Anal. Appl., 141 (1989),
177–188.
[111] A. Moudafi, Viscosity approximation methods for fixed points problems,
J. Math. Anal. Appl., 241(1) (2000), 46–55.
[112] C. H. Morales, Set-valued mappings in Banach spaces, Houston J. Math.,
9(2) (1983), 245–253.
[113] C. H. Morales, Surjectivity theorems for multivalued mappings of accretive
type, Comment. Math. Univ. Caroline, 26(2) (1985), 397–413.
[114] C. H. Morales and J. S. Jung, Convergence of paths for pseudocontrac-
tive mappings in Banach spaces, Proc. Amer. Math. Soc. 128 (2000),
3411–3419.
[115] S. B. Nadler, Multi-valued contraction mappings, Pacific J. Math., 30
(1969), 475–488.
[116] Z. Opial, Weak convergence of the sequence of successive approximations
for non-expansive mappings, Bull. Amer. Math. Soc., 73 (1967), 591–597.
[117] D. O’Regan and N. S. Papageorgiou, Fixed point theorems of Matkowski
type on complete gauge spaces, J. Nonlinear Functional Anal. Diff. Equa-
tions, to appear.
[118] M.O. Osilike and S. C. Aniagbosor, Weak and strong convergence
theorems for fixed points of asymptotically nonexpansive mappings, Math.
Comput. Model., 32 (2000), 1181–1191.
[119] D. Pascali and K. Silviu Sburlan, Nonlinear mappings of monotone type,
Sijthoff & Noordhoff International Publishers, Alphen aan den Rijn, 1978.
[120] S. Prus, On Bynum’s fixed point theorem, Atti Semin. Mat. Fis. Univ.
Modena, 38(2) (1990), 535–545.
[121] S. Prus, Geometrical background of metric fixed point theory, Handbook
of metric fixed point theory, pp. 93–132, Kluwer Acad. Publ., Dordrecht,
2001.
Bibliography 361

[122] L. Qihou, Iteration sequences for asymptotically quasi-nonexpansive

mapping with an error member of uniform convex Banach space, J. Math.
Anal. Appl., 266 (2002), 468–471.

[123] S. Reich, Asymptotic behavior of contractions in Banach spaces, J. Math.

Anal. Appl. 44 (1973), 57–70

[124] S. Reich, Approximate selections, best approximations, fixed points, and

invariant sets, J. Math. Anal. Appl. 62 (1978), 104–113

[125] S. Reich, Weak convergence theorems for nonexpansive mappings in

Banach spaces, J. Math. Anal. Appl., 67(2) (1979), 274–276.

[126] S. Reich, Strong convergence theorems for resolvents of accretive operators

in Banach spaces, J. Math. Anal. Appl., 75 (1980), 287–292.

[127] S. Reich, On the asymptotic behavior of nonlinear semigroups and the

range of accretive operators, J. Math. Anal. Appl., 79 (1981), 113–126.

[128] S. Reich, Convergence, resolvent consistency, and the fixed point property
for nonexpansive mappings, Contemporary Math., 18 (1983), 167–174.

[129] S. Reich, The almost fixed point property for nonexpansive mappings,
Proc. Amer. Math. Soc., 88 (1983), 44–46.

[130] S. Reich, Nonlinear semigroups, holomorphic mappings, and integral equa-

tions, Proc. Symp. Pure Math., Vol. 45, Part 2, Amer. Math. Soc., Provi-
dence, RI, (1986) 307–324.

[131] S. Reich, Book Review: Geometry of Banach spaces, duality mappings

and nonlinear problems, Bull. Amer. Math. Soc., (N.S.) 26 (2) (1992),
367–370.

[132] S. Reich and E. Kopecká, Nonexpansive retracts in Banach spaces, Banach

Center Publications, 77 (2007), 215–225.

[133] S. Reich and A. J. Zaslavski, A fixed point theorem for Matkowski con-
tractions, Fixed Point Theory 8 (2007), 303–307.

[134] J. Reinermann and R. Schoneberg, Some results and problems in the fixed
point theory for nonexpansive and pseudocontractive mappings in Hilbert
space, Fixed Point Theory Appl., (Srinivasa Swaminathan, ed.) (Proc.
Semin. Halifax, 1975), Academic Press, New York, 1976, pp. 187–196.

[135] B. E. Rhoades, Some fixed point iteration procedures, Int. J. Math. Math.
Sci., 14(1) (1991), 1–16.

[136] H. L. Royden, Real analysis, Prentice-Hall of India, New Delhi, India,

1988.
362 Bibliography

[137] D. R. Sahu, Fixed points of demicontinuous nearly Lipschitzian mappings

in Banach spaces, Comment. Math. Univ. Caroline, 46(4) (2005), 653–666.

[138] J. Schauder, Der Fixpunktsatzin Funktionalraiimen, Studia Mathematika,

2 (1930), 171–180.

[139] J. Schu, A fixed point theorem for nonexpansive mappings on star-shaped

domains, Z. Anal. Anwend., 10(4) (1991), 417–431.

[140] J. Schu, Approximation of fixed points of asymptotically nonexpansive

mappings, Proc. Amer. Math. Soc., 112(1) (1991), 143–151.

[141] R. Schoneberg, On the structure of fixed point sets of pseudocontractive

mappings-II, Comment Math. Univ. Carol., 18(2) (1977), 299–310.

[142] H. F. Senter and W. G. Dotson, Jr., Approximating fixed points of non-

expansive mappings, Proc. Amer. Math. Soc., 44(2) (1974), 375–380.

[143] I. Shafrir, The approximate fixed point property in Banach and hyperbolic
spaces, Israel J. Math., 71 (1990), 211–223.

[144] N. Shioji and W. Takahashi, Strong convergence of approximated

sequences for nonexpansive mappings in Banach spaces, Proc. Amer.
Math. Soc., 125(12) (1997), 3641–3645.

[145] R. E. Showalter, Monotone operators in Banach space and nonlinear par-

tial differential equations, Mathematical Surveys and Monographs, 49
Amer. Math. Soc., Providence, RI, 1997.

[146] A. H. Siddiqi, Functional analysis with applications, Tata McGraw-Hill

Publishing Company Limited, New Delhi, 1992.

[147] K. L. Singh and J. H. M. Whitfield, Sequence of iterates for nonexpan-

sive type mappings in Banach spaces, Nonlinear Anal. Appl., 80 (1982),
243–260.

[148] S. P. Singh and B. Watson, Proximity maps and fixed points, J. Approx.
Theorey, 39 (1983), 72–76.

[149] S. P. Singh and B. Watson, On approximating fixed points, Proc. Sym-

posia in Pure Math., 45(2) (1986) 393–395.

[150] S. P. Singh, B. Watson and P. Srivastava, Fixed point theory and best
approximation: The KKM-map principle, Series: Mathematics and Its
Applications, Springer-Verlag, Berlin, 1997.

[151] Z. H. Sun, Strong convergence of an implicit iteration process for a finite

family of asymptotically quasi nonexpansive mappings, J. Math. Anal.
Appl., 286 (2003), 351–358.
Bibliography 363

[152] T. Suzuki, Common fixed points of two nonexpansive mappings in Banach

spaces, Bull. Austral. Math. Soc., 69(1) (2004), 1–18.

[153] S. Swaminathan, Normal structure in Banach spaces and its generalisa-

tions, Contemporary Mathematics, 18 (1983), 201–215.

[154] W. Takahashi, A convexity in metric space and nonexpansive mappings-I,

Kodai Math. Sem. Rep., 22 (1970), 142–149.

[155] W. Takahashi, Nonlinear functional analysis, Fixed point theory and its
applications, Yokohama Publishers, Yokohama, 2000.

[156] W. Takahashi and D. H. Jeong, Fixed point theorem for nonexpansive

semigroups on Banach space, Proc. Amer. Math. Soc., 122(4) (1994),
1175–1179.

[157] W. Takahashi and G. E. Kim, Strong convergence of approximants to

fixed points of nonexpansive nonself-mappings in Banach spaces, Non-
linear Anal., Theory Methods Appl., 32(3) (1998), 447–454.

[158] W. Takahashi and Y. Ueda, On Reich’s strong convergence theorems

for resolvents of accretive operators, J. Math. Anal. Appl., 104 (1984),
546–553.

[159] K. K. Tan and H. K. Xu, Approximating fixed points of nonexpansive

mappings by the Ishikawa iteration process, J. Math. Anal. Appl., 178(2)
(1993), 301–308.

[160] K. K. Tan and H. K. Xu, Fixed point iteration processes for asymptotically
nonexpansive mappings, Proc. Amer. Math. Soc., 122(3) (1994), 733–739.

[161] J. R. L. Webb, Approximation solvability of nonlinear equations, Non-

linear Anal. Funct. Appl., Proc. Spring. Sch., 49 (1982), 234–257.

[162] C. S. Wong, Approximation to fixed points of generalized nonexpansive

mappings, Proc. Amer. Math. Soc., 54 (1976), 93–97.

[163] Z. B. Xu and G. F. Roach, Characteristic inequalities of uniformly convex

and uniformly smooth Banach spaces, J. Math. Anal. Appl., 157(1) (1991),
189–210.

[164] H. K. Xu, Existence and convergence for fixed points of mappings of

asymptotically nonexpansive type, Nonlinear Anal., Theory Methods
Appl., 16(12) (1991), 1139–1146.

[165] H. K. Xu, Approximating curves of nonexpansive nonself-mappings in

Banach spaces, C.R. Acad. Sci., 325(2) (1997), 151–156.

[166] H. K. Xu, Multivalued nonexpansive mappings in Banach spaces, Non-

linear Anal., Theory Methods Appl., 43(6) (2001), 693–706.
364 Bibliography

[167] H. K. Xu, Viscosity approximation methods for nonexpansive mappings,

J. Math. Anal. Appl., 298 (2004), 279–291.

[168] H. K. Xu and X.M. Yin, Strong convergence theorems for nonexpansive

nonself mappings, Nonlinear Anal., Theory Methods Appl., 24(2) (1995),
223–228.

[169] C. M. Yanez, A remark on weakly inward contractions, Nonlinear Anal.,

Theory Methods Appl., 16(10) (1991), 847–848.

[170] H. W. Yi and Y. C. Zhao, Fixed point theorems for weakly inward mul-
tivalued mappings and their randomizations, J. Math. Anal. Appl., 183
(1994), 613–619.

[171] E. Zeidler, Nonlinear functional analysis and its applications, Springer-

Verlag, Berlin, 1986.
Index

A Cauchy initial value problem 339

Absolutely convergent series 20 Cauchy sequence 6
Accretive 268 Cauchy-Schwarz inequality 16
Admissible hull of set 202 Characteristic of convexity 96
Admissible set 201 Chebyshev set 115
Almost convergent 113 Coercive 69
α-monotone mapping 68 Closed ball 1
α-strongly accretive 270 Closed convex hull 10
α-strongly pseudocontractive 270 Closed half-space 30
Approximating fixed point sequence 197 Closed mapping 45
Assad and Kirk’s fixed point theorem 194 Closed set 3,18
Asymptotic contraction mapping 184 Closure of a set 3
Asymptotically nonexpansive 243 Closure of convex hull 10
Asymptotically regular mapping 225 Cluster point 6
Attractor 333 Coefficient of convexity 96
B Compact set 5,32
Banach-Alaoglu’s theorem 43 Compact topological space 5
Banach contraction principle 178 Complete inner product space 18
Banach limit 108 Complete metric space 6
Banach-Mazur distance 155 Completely continuous mapping 46
Banach space 18 Condition I 296
Base for a topology 4 Conjugate function 103
Basis of linear space 20 Continuous mapping 7
Best approximation 115 Contraction mapping 44
Biconjugate function 103 Contractive mapping 44
Bilinear functional 34 Contractivity factor 334
Boundary of set 3 Convergent sequence 6
Bounded below function 80 Convergent series 19
Bounded linear operator 23 Convex function 79
Bounded mapping 45 Convex metric space 197
Bounded set 1 Convex structure 197
Brouwer’s fixed point theorem 47 Convex functional 28
Browder’s convergence theorem 315 Convex hull 9
Browder’s theorem 229 Convex metric space 197
Browder and Petryshyn’s theorem 286 Convex set 9
Brosowski’s theorem 335
D
C Demiclosed mapping 45
Cantor’s intersection theorem 20 Demicompact mapping 46
Caristi’s fixed point theorem 178 Demicontinuous mapping 45
366 Index

Dense set 21 I
Diameter of a set 1 IFS 333
Diametral point 199 Index set 7
Distance between points 1 Indicator function 80
Directed set 5 Induced topology 11
Directional derivative 83 Infinite-dimensional space 9
Dissipative 268 Inner product 15
Domain 80 Inner product space 15
Dual space 26 Interior of set 2
Duality mapping 67,74 Interior point of set 2
Duality mapping with gauge 74 Inward mapping 212
Duality pairing 34 Inward set 211
Ishikawa iteration process 304
E
Eberlein-Smulian theorem 41 J
Epigraph of function 80 James theorem 36
Equivalent norm 14 K
Euclidean space 15 Kadec-Klee property 57
Existence of Banach limit 109 Kirk’s fixed point theorem 230
Expansive 268
L
F Lifschitz’s coefficient 206
Finite ε-net 32 Limit functional 106
Finite-dimensional space 9 Limit point 18
Finite representable space 155 Linear combination 9
Firmly nonexpansive 120 Linear functional 22
Fixed point 46 Linear operator 22
Fixed point property 201 Line segment 49
Fréchet differential norm 100 Linear space 8
Fréchet metric 12 Linear span 9
Fredholem integral equation 340 Linear topology 15
Lipschitz condition 44
G Lipschitz constant 175
Gâteuax differential norm 91 Lipschitzian mapping 175
Gauge function 74 Locally bounded mapping 45
Generalized duality mapping 74 Locally convex topology 15
Generalized sequence 5 Locally convex topological space 15
Goebel and Kirk’s fixed point Locally uniformly convex space 66
theorem 248, 251 Lower semicontinuous function 80
Göhde’s fixed point theorem 229 Lower semicontinuous multivalued
H mapping 46
Hahn-Banach theorem 29 M
Halpern type iteration 327 m-linear functional 34
Hausdorff metric 29 Maluta’s fixed point theorem 254
Hausdorff topological space 5 Mann iteration 289
Heine-Borel theorem 33 Mann iteration process 288
Hemicontinuous mapping 45 Matkowski’s fixed point theorem 185
Hilbert space 18 Maximal monotone 69
Hyperbolic iterated function system 333 Mazur’s theorem 33
Hyperplane 30 Mean 107
Index 367

Mean point 111 Q

Metric 1 Quasi-nonexpansive mapping 295
Metrically convex 194
R
Metric projection mapping 115
Reasonable wanderer 226
Metric topology 2
Reflexive Banach space 35
Metric space 1
Reich’s convergence theorem 322
Metrizable space 2
Reisz representation theorem 28
Minimal norm 52
Retract 121
Modulus of convexity 58
Retraction mapping 121
Modulus of local convexity 66
Modulus of smoothness 94 S
Monotone 68 Scalar 9
Multivalued mapping 46 Schauder basis 20
Schauder’s fixed point theorem 47
N
Schur property 37
Nadler’s fixed point theorem 191
Seminorm 14
Natural embedding mapping 35
Seminormed space 14
Nearly asymptotically nonexpansive 186
Signum function 68
Nearly contraction mapping 186
Singh and Watson’s convergence
Nearly Lipschitzian mapping 186
theorem 324
Nearly Lipschitz constant 186
Smooth Banach space 91
Nearly nonexpansive mapping 186
Solution of variational inequality
Nearly uniformly k-contraction 186
Star-shaped 222
Nearly uniformly k-Lipschitzian 186
Strictly convex Banach space 50
Neighborhood of a point 4
Strictly convex function 79
Net 5
Strictly monotone mapping 68
Nonexpansive mapping 44
Strongly monotone mapping 68
Nonlinear operator 22
Strongly pseudocontractive 264
Non-Lipschitzian mapping 44, 259
Subbase 4
Norm 10
Subdifferential 85
Norm topology 11
Subgradient 85
Normal structure 146
Sublinear functional 28
Normed space 10
Subnet 5
O Subspace 9
Open ball 1 Sun 121
Open half-space 30 Super reflexive 156
Open cover 5 Support function of set 103
Open set 2 Supporting hyperplane 31
Outward set 211
T
P T0 -space 4
Parallelogram law 16 T1 -space 4
Partially ordered set 350 T2 -space 5
Point of support 31 Topology 1
Precompact 45 Topological linear space 15
Pre-Hilbert space 15 Topological space 2
Product space 8 Totally bounded set 32
Proper function 80 Translation invariant 11
Proximinal set 115 Tychonoff’s fixed point theorem 47
Pseudocontractive mapping 264 Tychonoff’s theorem 8
368 Index

U Vector 9
Ultrafilter 351 Vector space 8
Ultrapower 157
W
Unbounded set 2
Weak Cauchy sequence 36
Uniform boundedness principle 25
Weakly asymptotically regular 225
Uniformly asymptotically regular 225
Weakly closed mapping 45
Uniformly continuous mapping 43
Weakly closed set 36
Uniformly convex Banach space 53,96
Weakly compact set 37
Uniformly Fréchet differential norm 100
Weakly complete 36
Uniformly Gâteuax differential norm 91
Weakly continuous mapping 45
Uniformly L-Lipschitzian mapping 176
Weakly convergent sequence 36
Uniformly nonsquare 96
Weakly inward mapping 212
Uniformly normal structure 151
Weakly outward mapping 212
Uniformly smooth Banach space 98
Weak topology 36
Upper semicontinuous multivalued
Weak* topology 38
mapping 46
Weierstrass theorem 33
V Z
Variational inclusion problem 343 Zorn’s lemma 350
Variational inequality 119 Zero vector 8