Convex Analysis PDF

Series ISSN: 1938-1743
MORDUKHOVICH • NAM
SYNTHESIS LECTURES ON
COMPUTER MATHEMATICS AND STATISTICS
M
&C Mor gan & Cl aypool Publishers
Series Editor: Steven G. Krantz, Washington University, St. Louis
An Easy Path to Convex Analysis and

An Easy Path to Convex
Applications Analysis and Applications
Boris S. Mordukhovich, Wayne State University
AN EASY PATH TO CONVEX ANALYSIS AND APPLICATIONS

Nguyen Mau Nam, Portland State University
Convex optimization has an increasing impact on many areas of mathematics, applied sciences, and practical
applications. It is now being taught at many universities and being used by researchers of different fields. As
convex analysis is the mathematical foundation for convex optimization, having deep knowledge of convex
analysis helps students and researchers apply its tools more effectively. The main goal of this book is to provide
an easy access to the most fundamental parts of convex analysis and its applications to optimization. Modern
techniques of variational analysis are employed to clarify and simplify some basic proofs in convex analysis
and build the theory of generalized differentiation for convex functions and sets in finite dimensions. We also
present new applications of convex analysis to location problems in connection with many interesting geometric
problems such as the Fermat-Torricelli problem, the Heron problem, the Sylvester problem, and their
generalizations. Of course, we do not expect to touch every aspect of convex analysis, but the book consists
of sufficient material for a first course on this subject. It can also serve as supplemental reading material for a
Boris S. Mordukhovich
course on convex optimization and applications.
Nguyen Mau Nam
About SYNTHESIs
This volume is a printed version of a work that appears in the Synthesis
MOR GAN & CL AYPOOL
Digital Library of Engineering and Computer Science. Synthesis Lectures
provide concise, original presentations of important research and development
topics, published quickly, in digital and print formats. For more information
visit www.morganclaypool.com
ISBN: 978-1-62705-237-5
SYNTHESIS LECTURES ON
Mor gan &Cl aypool Publishers 90000
MATHEMATICS AND STATISTICS
w w w. m o r g a n c l a y p o o l . c o m
9 781627 052375
Steven G. Krantz, Series Editor
An Easy Path to
Convex Analysis and Applications
Synthesis Lectures on
Mathematics and Statistics
Editor
Steven G. Krantz, Washington University, St. Louis
An Easy Path to Convex Analysis and Applications

Boris S. Mordukhovich and Nguyen Mau Nam
2014
Applications of Affine and Weyl Geometry

Eduardo García-Río, Peter Gilkey, Stana Nikčević, and Ramón Vázquez-Lorenzo
2013
Essentials of Applied Mathematics for Engineers and Scientists, Second Edition

Robert G. Watts
2012
Chaotic Maps: Dynamics, Fractals, and Rapid Fluctuations

Goong Chen and Yu Huang
2011
Matrices in Engineering Problems

Marvin J. Tobias
2011
e Integral: A Crux for Analysis

Steven G. Krantz
2011
Statistics is Easy! Second Edition

Dennis Shasha and Manda Wilson
2010
Lectures on Financial Mathematics: Discrete Asset Pricing

Greg Anderson and Alec N. Kercheval
2010
iii
Jordan Canonical Form: eory and Practice
Steven H. Weintraub
2009
e Geometry of Walker Manifolds

Miguel Brozos-Vázquez, Eduardo García-Río, Peter Gilkey, Stana Nikčević, and Ramón
Vázquez-Lorenzo
2009
An Introduction to Multivariable Mathematics

Leon Simon
2008
Jordan Canonical Form: Application to Differential Equations

Steven H. Weintraub
2008
Statistics is Easy!
Dennis Shasha and Manda Wilson
2008
A Gyrovector Space Approach to Hyperbolic Geometry

Abraham Albert Ungar
2008
Copyright © 2014 by Morgan & Claypool
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means—electronic, mechanical, photocopy, recording, or any other except for brief quotations
in printed reviews, without the prior permission of the publisher.
An Easy Path to Convex Analysis and Applications

www.morganclaypool.com
ISBN: 9781627052375 paperback

ISBN: 9781627052382 ebook
DOI 10.2200/S00554ED1V01Y201312MAS014
A Publication in the Morgan & Claypool Publishers series

SYNTHESIS LECTURES ON MATHEMATICS AND STATISTICS
Lecture #14
Series Editor: Steven G. Krantz, Washington University, St. Louis
Series ISSN
Synthesis Lectures on Mathematics and Statistics
Print 1938-1743 Electronic 1938-1751
An Easy Path to
Convex Analysis and Applications
Boris S. Mordukhovich
Wayne State University
Nguyen Mau Nam

Portland State University
SYNTHESIS LECTURES ON MATHEMATICS AND STATISTICS #14
M
&C Morgan & cLaypool publishers
ABSTRACT
Convex optimization has an increasing impact on many areas of mathematics, applied sciences,
and practical applications. It is now being taught at many universities and being used by re-
searchers of different fields. As convex analysis is the mathematical foundation for convex opti-
mization, having deep knowledge of convex analysis helps students and researchers apply its tools
more effectively. e main goal of this book is to provide an easy access to the most fundamental
parts of convex analysis and its applications to optimization. Modern techniques of variational
analysis are employed to clarify and simplify some basic proofs in convex analysis and build the
theory of generalized differentiation for convex functions and sets in finite dimensions. We also
present new applications of convex analysis to location problems in connection with many in-
teresting geometric problems such as the Fermat-Torricelli problem, the Heron problem, the
Sylvester problem, and their generalizations. Of course, we do not expect to touch every aspect of
convex analysis, but the book consists of sufficient material for a first course on this subject. It can
also serve as supplemental reading material for a course on convex optimization and applications.
KEYWORDS
Affine set, Carathéodory theorem, convex function, convex set, directional deriva-
tive, distance function, Fenchel conjugate, Fermat-Torricelli problem, generalized
differentiation, Helly theorem, minimal time function, Nash equilibrium, normal
cone, Radon theorem, optimal value function, optimization, smallest enclosing ball
problem, set-valued mapping, subdifferential, subgradient, subgradient algorithm,
support function, Weiszfeld algorithm
vii
e first author dedicates this book to the loving memory of his father
S M (1924–1993),
a kind man and a brave warrior.
e second author dedicates this book to the memory of his parents.

ix
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
1 Convex Sets and Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Relative Interiors of Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5 e Distance Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.6 Exercises for Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2 Subdifferential Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.1 Convex Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2 Normals to Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.3 Lipschitz Continuity of Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.4 Subgradients of Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.5 Basic Calculus Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.6 Subgradients of Optimal Value Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.7 Subgradients of Support Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.8 Fenchel Conjugates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.9 Directional Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.10 Subgradients of Supremum Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.11 Exercises for Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3 Remarkable Consequences of Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

3.1 Characterizations of Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.2 Carathéodory eorem and Farkas Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3 Radon eorem and Helly eorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
x
3.4 Tangents to Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.5 Mean Value eorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.6 Horizon Cones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.7 Minimal Time Functions and Minkowski Gauge . . . . . . . . . . . . . . . . . . . . . . . 111
3.8 Subgradients of Minimal Time Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.9 Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.10 Exercises for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4 Applications to Optimization and Location Problems . . . . . . . . . . . . . . . . . . . 129

4.1 Lower Semicontinuity and Existence of Minimizers . . . . . . . . . . . . . . . . . . . . 129
4.2 Optimality Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.3 Subgradient Methods in Convex Optimization . . . . . . . . . . . . . . . . . . . . . . . . 140
4.4 e Fermat-Torricelli Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.5 A Generalized Fermat-Torricelli Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.6 A Generalized Sylvester Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.7 Exercises for Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Solutions and Hints for Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Authors’ Biographies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
xi
Preface
Some geometric properties of convex sets and, to a lesser extent, of convex functions had
been studied before the 1960s by many outstanding mathematicians, first of all by Hermann
Minkowski and Werner Fenchel. At the beginning of the 1960s convex analysis was greatly de-
veloped in the works of R. Tyrrell Rockafellar and Jean-Jacques Moreau who initiated a systematic
study of this new field. As a fundamental part of variational analysis, convex analysis contains a
generalized differentiation theory that can be used to study a large class of mathematical mod-
els with no differentiability assumptions imposed on their initial data. e importance of convex
analysis for many applications in which convex optimization is the first to name has been well
recognized. e presence of convexity makes it possible not only to comprehensively investigate
qualitative properties of optimal solutions and derive necessary and sufficient conditions for opti-
mality but also to develop effective numerical algorithms to solve convex optimization problems,
even with nondifferentiable data. Convex analysis and optimization have an increasing impact
on many areas of mathematics and applications including control systems, estimation and signal
processing, communications and networks, electronic circuit design, data analysis and modeling,
statistics, economics and finance, etc.
ere are several fundamental books devoted to different aspects of convex analysis and
optimization. Among them we mention “Convex Analysis” by Rockafellar [26], “Convex Analysis
and Minimization Algorithms” (in two volumes) by Hiriart-Urruty and Lemaréchal [8] and its
abridge version [9], “Convex Analysis and Nonlinear Optimization” by Borwein and Lewis [4],
“Introductory Lectures on Convex Optimization” by Nesterov [21], and “Convex Optimization”
by Boyd and Vandenberghe [3] as well as other books listed in the bibliography below.
In this big picture of convex analysis and optimization, our book serves as a bridge for
students and researchers who have just started using convex analysis to reach deeper topics in the
field. We give detailed proofs for most of the results presented in the book and also include many
figures and exercises for better understanding the material. e powerful geometric approach
developed in modern variational analysis is adopted and simplified in the convex case in order
to provide the reader with an easy path to access generalized differentiation of convex objects in
finite dimensions. In this way, the book also serves as a starting point for the interested reader to
continue the study of nonconvex variational analysis and applications. It can be of interest from
this viewpoint to experts in convex and variational analysis. Finally, the application part of this
book not only concerns the classical topics of convex optimization related to optimality conditions
and subgradient algorithms but also presents some recent while easily understandable qualitative
and numerical results for important location problems.
xii PREFACE
e book consists of four chapters and is organized as follows. In Chapter 1 we study
fundamental properties of convex sets and functions while paying particular attention to classes
of convex functions important in optimization. Chapter 2 is mainly devoted to developing basic
calculus rules for normals to convex sets and subgradients of convex functions that are in the
mainstream of convex theory. Chapter 3 concerns some additional topics of convex analysis that
are largely used in applications. Chapter 4 is fully devoted to applications of basic results of convex
analysis to problems of convex optimization and selected location problems considered from both
qualitative and numerical viewpoints. Finally, we present at the end of the book Solutions and
Hints to selected exercises.
Exercises are given at the end of each chapter while figures and examples are provided
throughout the whole text. e list of references contains books and selected papers, which are
closely related to the topics considered in the book and may be helpful to the reader for advanced
studies of convex analysis, its applications, and further extensions.
Since only elementary knowledge in linear algebra and basic calculus is required, this book
can be used as a textbook for both undergraduate and graduate level courses in convex analysis and
its applications. In fact, the authors have used these lecture notes for teaching such classes in their
universities as well as while visiting some other schools. We hope that the book will make convex
analysis more accessible to large groups of undergraduate and graduate students, researchers in
different disciplines, and practitioners.

December 2013
xiii
Acknowledgments
e first author acknowledges continued support by grants from the National Science Foundation
and the second author acknowledges support by the Simons Foundation.

December 2013
xv
List of Symbols
R the real numbers

R D . 1; 1 the extended real line
RC the nonnegative real numbers
R> the positive real numbers
N the positive integers
co ˝ convex hull of ˝
aff ˝ affine hull of ˝
int ˝ interior of ˝
ri ˝ relative interior of ˝
span ˝ linear subspace generated by ˝
cone ˝ cone generated by ˝
K˝ convex cone generated by ˝
dim ˝ dimension of ˝
˝ closure of ˝
bd ˝ boundary of ˝
IB closed unit ball
IB.xIN r/ closed ball with center xN and radius r
dom f domain of f
epi f epigraph of f
gph F graph of mapping F
LŒa; b line connecting a and b
d.xI ˝/ distance from x to ˝
˘.xI ˝/ projection of x to ˝
hx; yi inner product of x and y
A adjoint/transpose of linear mapping/matrix A
N.xIN ˝/ normal cone to ˝ at xN
ker A kernel of linear mapping A
D F .x;
N y/
N coderivative to F at .x;
N y/
N
T .xI
N ˝/ tangent cone to ˝ at xN
F1 horizon/asymptotic cone of F
T˝F minimal time function defined by constant dynamic F and target set ˝
L.x; / Lagrangian
F Minkowski gauge of F
xvi LIST OF SYMBOLS
˘F .I ˝/ generalized projection defined by the minimal time function
GF .x;
N t/ generalized ball defined by dynamic F with center xN and radius t
1
CHAPTER 1
Convex Sets and Functions

is chapter presents definitions, examples, and basic properties of convex sets and functions in
the Euclidean space Rn and also contains some related material.
1.1 PRELIMINARIES
We start with reviewing classical notions and properties of the Euclidean space Rn . e proofs
of the results presented in this section can be found in standard books on advanced calculus and
linear algebra.
Let us denote by Rn the set of all n tuples of real numbers x D .x1 ; : : : ; xn /. en Rn is
a linear space with the following operations:
.x1 ; : : : ; xn / C .y1 ; : : : ; yn / D .x1 C y1 ; : : : ; xn C yn /;

.x1 ; : : : ; xn / D .x1 ; : : : ; xn /;
where .x1 ; : : : ; xn /; .y1 ; : : : ; yn / 2 Rn and 2 R. e zero element of Rn and the number zero
of R are often denoted by the same notation 0 if no confusion arises.
For any x D .x1 ; : : : ; xn / 2 Rn , we identify it with the column vector x D Œx1 ; : : : ; xn T ,
where the symbol “T ” stands for vector transposition. Given x D .x1 ; : : : ; xn / 2 Rn and y D
.y1 ; : : : ; yn / 2 Rn , the inner product of x and y is defined by
n
X
hx; yi WD x i yi :
iD1
e following proposition lists some important properties of the inner product in Rn .
Proposition 1.1 For x; y; z 2 Rn and 2 R, we have:

(i) hx; xi 0, and hx; xi D 0 if and only if x D 0.
(ii) hx; yi D hy; xi.
(iii) hx; yi D hx; yi.
(iv) hx; y C zi D hx; yi C hx; zi.
e Euclidean norm of x D .x1 ; : : : ; xn / 2 Rn is defined by

q
kxk WD x12 C : : : C xn2 :
2 1. CONVEX SETS AND FUNCTIONS
p
It follows directly from the definition that kxk D hx; xi.
Proposition 1.2 For any x; y 2 Rn and 2 R, we have:

(i) kxk 0, and kxk D 0 if and only if x D 0.
(ii) kxk D jj kxk.
(iii) kx C yk kxk C kyk .the triangle inequality/.
(iv) jhx; yij kxk kyk .the Cauchy-Schwarz inequality/.
Using the Euclidean norm allows us to introduce the balls in Rn , which can be used to
define other topological notions in Rn .
Definition 1.3 e   centered at xN with radius r 0 and the    of Rn
are defined, respectively, by
˚ ˇ ˚ ˇ
N r/ WD x 2 Rn ˇ kx xk
IB.xI N r and IB WD x 2 Rn ˇ kxk 1 :
It is easy to see that IB D IB.0I 1/ and IB.xI

N r/ D xN C rIB .
Definition 1.4 Let ˝ Rn . en xN is an   of ˝ if there is ı > 0 such that
IB.xI
N ı/ ˝:
e set of all interior points of ˝ is denoted by int ˝ . Moreover, ˝ is said to be  if every point of
˝ is its interior point.
We get that ˝ is open if and only if for every xN 2 ˝ there is ı > 0 such that IB.xI
N ı/ ˝ .
n
It is obvious that the empty set ; and the whole space R are open. Furthermore, any open ball
B.xIN r/ WD fx 2 Rn j kx xkN < rg centered at xN with radius r is open.
Definition 1.5 A set ˝ Rn is  if its complement ˝ c D Rn n˝ is open in Rn .
N r/ are closed in Rn .
It follows that the empty set, the whole space, and any ball IB.xI
Proposition 1.6 (i) e union of any collection of open sets in Rn is open.

(ii) e intersection of any finite collection of open sets in Rn is open.
(iii) e intersection of any collection of closed sets in Rn is closed.
(iv) e union of any finite collection of closed sets in Rn is closed.
Definition 1.7 Let fxk g be a sequence in Rn . We say that fxk g  to xN if

kxk xkN ! 0 as k ! 1. In this case we write
lim xk D x:
N
k!1
1.1. PRELIMINARIES 3
is notion allows us to define the following important topological concepts for sets.
Definition 1.8 Let ˝ be a nonempty subset of Rn . en:

(i) e  of ˝ , denoted by ˝ or cl ˝ , is the collection of limits of all convergent sequences be-
longing to ˝ .
(ii) e  of ˝ , denoted by bd ˝ , is the set ˝ n int ˝ .
We can see that the closure of ˝ is the intersection of all closed sets containing ˝ and that
the interior of ˝ is the union of all open sets contained in ˝ . It follows from the definition that
xN 2 ˝ if and only if for any ı > 0 we have IB.xIN ı/ \ ˝ ¤ ;. Furthermore, xN 2 bd ˝ if and only
N ˝/ intersects both sets ˝ and its complement ˝ c .
if for any ı > 0 the closed ball IB.xI
Definition 1.9 Let fxk g be a sequence in Rn and let fk` g be a strictly increasing sequence of positive
integers. en the new sequence fxk` g is called a  of fxk g.
We say that a set ˝ is bounded if it is contained in a ball centered at the origin with some
radius r > 0, i.e., ˝ IB.0I r/. us a sequence fxk g is bounded if there is r > 0 with
kxk k r for all k 2 N :
e following important result is known as the Bolzano-Weierstrass theorem.
eorem 1.10 Any bounded sequence in Rn contains a convergent subsequence.
e next concept plays a very significant role in analysis and optimization.
Definition 1.11 We say that a set ˝ is  in Rn if every sequence in ˝ contains a subsequence
converging to some point in ˝ .
e following result is a consequence of the Bolzano-Weierstrass theorem.
eorem 1.12 A subset ˝ of Rn is compact if and only if it is closed and bounded.
For subsets ˝ , ˝1 , and ˝2 of Rn and for 2 R, we define the operations:

˚ ˇ ˚ ˇ
˝1 C ˝2 WD x C y ˇ x 2 ˝1 ; y 2 ˝2 ; ˝ WD x ˇ x 2 ˝ :
e next proposition can be proved easily.
Proposition 1.13 Let ˝1 and ˝2 be two subsets of Rn .

(i) If ˝1 is open or ˝2 is open, then ˝1 C ˝2 is open.
(ii) If ˝1 is closed and ˝2 is compact, then ˝2 C ˝2 is closed.
Recall now the notions of bounds for subsets of the real line.
Definition 1.14 Let D be a subset of the real line. A number m 2 R is a   of D if we
have
x m for all x 2 D:
If the set D has a lower bound, then it is  . Similarly, a number M 2 R is an 
 of D if
x M for all x 2 D;
and D is   if it has an upper bound. Furthermore, we say that the set D is 
if it is simultaneously bounded below and above.
Now we are ready to define the concepts of infimum and supremum of sets.
Definition 1.15 Let D R be nonempty and bounded below. e infimum of D , denoted by inf D ,
is the greatest lower bound of D . When D is nonempty and bounded above, its supremum, denoted by
sup D , is the least upper bound of D . If D is not bounded below .resp. above/, then we set inf D WD 1
.resp. sup D WD 1/. We also use the convention that inf ; WD 1 and sup ; WD 1.
e following fundamental axiom ensures that these notions are well-defined.

Completeness Axiom. For every nonempty subset D of R that is bounded above, the least upper bound
of D exists as a real number.
Using the Completeness Axiom, it is easy to see that if a nonempty set is bounded below,
then its greatest lower bound exists as a real number.
roughout the book we consider for convenience extended-real-valued functions, which
take values in R WD . 1; 1. e usual conventions of extended arithmetics are that a C 1 D
1 for any a 2 R, 1 C 1 D 1, and t 1 D 1 for t > 0.
Definition 1.16 Let f W ˝ ! R be an extended-real-valued function and let xN 2 ˝ with f .x/

N <
1. en f is  at xN if for any > 0 there is ı > 0 such that
jf .x/ N < whenever kx

f .x/j xk
N < ı; x 2 ˝:
We say that f is continuous on ˝ if it is continuous at every point of ˝ .
It is obvious from the definition that if f W ˝ ! R is continuous at xN (with f .x/ N < 1),
then it is finite on the intersection of ˝ and a ball centered at xN with some radius r > 0. Further-
more, f W ˝ ! R is continuous at xN (with f .x/N < 1) if and only if for every sequence fxk g in ˝
converging to xN the sequence ff .xk /g converges to f .x/N .
1.2. CONVEX SETS 5
Definition 1.17 Let f W ˝ ! R and let xN 2 ˝ with f .x/
N < 1. We say that f has a  -
 at xN relative to ˝ if there is ı > 0 such that
N for all x 2 IB.xI

f .x/ f .x/ N ı/ \ ˝:
We also say that f has a /  at xN relative to ˝ if
N for all x 2 ˝:
f .x/ f .x/
e notions of local and global maxima can be defined similarly.

Finally in this section, we formulate a fundamental result of mathematical analysis and
optimization known as the Weierstrass existence theorem.
eorem 1.18 Let f W ˝ ! R be a continuous function, where ˝ is a nonempty, compact subset

of Rn . en there exist xN 2 ˝ and uN 2 ˝ such that
˚ ˇ ˚ ˇ
N D inf f .x/ ˇ x 2 ˝ and f .u/
f .x/ N D sup f .x/ ˇ x 2 ˝ :
In Section 4.1 we present some “unilateral” versions of eorem 1.18.
1.2 CONVEX SETS

We start the study of convexity with sets and then proceed to functions. Geometric ideas play
an underlying role in convex analysis, its extensions, and applications. us we implement the
geometric approach in this book.
Given two elements a and b in Rn , define the interval/line segment
˚ ˇ
Œa; b WD a C .1 /b ˇ 2 Œ0; 1 :
Note that if a D b , then this interval reduces to a singleton Œa; b D fag.
Definition 1.19 A subset ˝ of Rn is  if Œa; b ˝ whenever a; b 2 ˝ . Equivalently, ˝ is

convex if a C .1 /b 2 ˝ for all a; b 2 ˝ and 2 Œ0; 1.
P Pm
Given !1 ; : : : ; !m 2 Rn , the element x D m iD1 i !i , where iD1 i D 1 and i 0 for
some m 2 N , is called a convex combination of !1 ; : : : ; !m .
Proposition 1.20 A subset ˝ of Rn is convex if and only if it contains all convex combinations of its
elements.
Figure 1.1: Convex set and nonconvex set.
Proof. e sufficient condition is trivial. To justify the necessity, we show by induction that any
P
convex combination x D m iD1 i !i of elements in ˝ is an element of ˝ . is conclusion follows
directly from the definition for m D 1; 2. Fix now a positive integer m 2 and suppose that every
convex combination of k 2 N elements from ˝ , where k m, belongs to ˝ . Form the convex
combination
mC1
X mC1
X
y WD i !i ; i D 1; i 0
iD1 iD1
and observe that if mC1 D 1, then 1 D 2 D : : : D m D 0, so y D !mC1 2 ˝ . In the case

where mC1 < 1 we get the representations
m
X m
X i
i D 1 mC1 and D 1;
1 mC1
iD1 iD1
which imply in turn the inclusion

m
X i
z WD !i 2 ˝:
1 mC1
iD1
It yields therefore the relationships

m
X i
y D .1 mC1 / !i C mC1 !mC1 D .1 mC1 /z C mC1 !mC1 2 ˝
1 mC1
iD1
and thus completes the proof of the proposition.
Proposition 1.21 Let ˝1 be a convex subset of Rn and let ˝2 be a convex subset of Rp . en the
Cartesian product ˝1 ˝2 is a convex subset of Rn Rp .
1.2. CONVEX SETS 7
Proof. Fix a D .a1 ; a2 /; b D .b1 ; b2 / 2 ˝1 ˝2 , and 2 .0; 1/. en we have a1 ; b1 2 ˝1 and
a2 ; b2 2 ˝2 . e convexity of ˝1 and ˝2 gives us
ai C .1 /bi 2 ˝i for i D 1; 2;
which implies therefore that

a C .1 /b D a1 C .1 /b1 ; a2 C .1 /b2 2 ˝1 ˝2 :
us the Cartesian product ˝1 ˝2 is convex.

Let us continue with the definition of affine mappings.
Definition 1.22 A mapping B W Rn ! Rp is  if there exist a linear mapping A W Rn ! Rp

and an element b 2 Rp such that B.x/ D A.x/ C b for all x 2 Rn .
Every linear mapping is affine. Moreover, B W Rn ! Rp is affine if and only if
B.x C .1 /y/ D B.x/ C .1 /B.y/ for all x; y 2 Rn and 2 R :
Now we show that set convexity is preserved under affine operations.
Proposition 1.23 Let B W Rn ! Rp be an affine mapping. Suppose that ˝ is a convex subset of Rn

and is a convex subset of Rp . en B.˝/ is a convex subset of Rp and B 1 ./ is a convex subset
of Rn .
Proof. Fix any a; b 2 B.˝/ and 2 .0; 1/. en a D B.x/ and b D B.y/ for x; y 2 ˝ . Since ˝
is convex, we have x C .1 /y 2 ˝ . en
a C .1 /b D B.x/ C .1 /B.y/ D B.x C .1 /y/ 2 B.˝/;
which justifies the convexity of the image B.˝/.

Taking now any x; y 2 B 1 ./ and 2 .0; 1/, we get B.x/; B.y/ 2 . is gives us

B.x/ C .1 /B.y/ D B x C .1 /y 2
1
by the convexity of . us we have x C .1 /y 2 B ./, which verifies the convexity of
the inverse image B 1 ./.
Proposition 1.24 Let ˝1 ; ˝2 Rn be convex and let 2 R. en both sets ˝1 C ˝2 and ˝1
are also convex in Rn .
Proof. It follows directly from the definitions.

Next we proceed with intersections of convex sets.
T
Proposition 1.25 Let f˝˛ g˛2I be a collection of convex subsets of Rn . en ˛2I ˝˛ is also a convex
subset of Rn .
T
Proof. Taking any a; b 2 ˛2I ˝˛ , we get that a; b 2 ˝˛ for all ˛ 2 I . e convexity of each
T
˝˛ ensures that a C .1 /b 2 ˝˛ for any 2 .0; 1/. us a C .1 /b 2 ˛2I ˝˛ and
T
the intersection ˛2I ˝˛ is convex.
Definition 1.26 Let ˝ be a subset of Rn . e   of ˝ is defined by

\ n ˇˇ o
co ˝ WD C ˇ C is convex and ˝ C :
e next proposition follows directly from the definition and Proposition 1.25.
Figure 1.2: Nonconvex set and its convex hull.
Proposition 1.27 e convex hull co ˝ is the smallest convex set containing ˝ .
Proof. e convexity of the set co ˝ ˝ follows from Proposition 1.25. On the other hand, for
any convex set C that contains ˝ we clearly have co ˝ C , which verifies the conclusion.
Proposition 1.28 For any subset ˝ of Rn , its convex hull admits the representation
nX
m ˇ X
m o
ˇ
co ˝ D i ai ˇ i D 1; i 0; ai 2 ˝; m 2 N :
iD1 iD1
1.2. CONVEX SETS 9
Proof. Denoting by C the right-hand side of the representation to prove, we obviously have ˝
C . Let us check that the set C is convex. Take any a; b 2 C and get
p
X q
X
a WD ˛i ai ; b WD ˇj bj ;
iD1 j D1
Pp Pq
where ai ; bj 2 ˝ , ˛i ; ˇj 0 with iD1 ˛i D j D1 ˇj D 1, and p; q 2 N . It follows easily that
for every number 2 .0; 1/, we have
p
X q
X
a C .1 /b D ˛i ai C .1 /ˇj bj :
iD1 j D1
en the resulting equality

p
X q
X p
X q
X
˛i C .1 /ˇj D ˛i C .1 / ˇj D 1
iD1 j D1 iD1 j D1
ensures that a C .1 /b 2 C , and thus co ˝ C by the definition of co ˝ . Fix now any a D
Pm
iD1 i ai 2 C with ai 2 ˝ co ˝ for i D 1; : : : ; m. Since the set co ˝ is convex, we conclude
by Proposition 1.20 that a 2 co ˝ and thus co ˝ D C .
Proposition 1.29 e interior int ˝ and closure ˝ of a convex set ˝ Rn are also convex.
Proof. Fix a; b 2 int ˝ and 2 .0; 1/. en find an open set V such that
a 2 V ˝ and so a C .1 /b 2 V C .1 /b ˝:
Since V C .1 /b is open, we get a C .1 /b 2 int ˝ , and thus the set int ˝ is convex.
To verify the convexity of ˝ , we fix a; b 2 ˝ and 2 .0; 1/ and then find sequences fak g
and fbk g in ˝ converging to a and b , respectively. By the convexity of ˝ , the sequence fak C
.1 /bk g lies entirely in ˝ and converges to a C .1 /b . is ensures the inclusion a C
.1 /b 2 ˝ and thus justifies the convexity of the closure ˝ .
To proceed further, for any a; b 2 Rn , define the interval
˚ ˇ
Œa; b/ WD a C .1 /b ˇ 2 .0; 1 :
We can also define the intervals .a; b and .a; b/ in a similar way.
Lemma 1.30 For a convex set ˝ Rn with nonempty interior, take any a 2 int ˝ and b 2 ˝ .
en Œa; b/ int ˝ .
Proof. Since b 2 ˝ , for any > 0, we have b 2 ˝ C IB . Take now 2 .0; 1 and let x WD
2
a C .1 /b . Choosing > 0 such that a C IB ˝ gives us

x C IB D a C .1 /b
C IB
a C .1 / ˝ C IB C IB
D a C .1 /˝ C .1 /IB C IB
h 2 i
aC IB C .1 /˝

˝ C .1 /˝ ˝:
is shows that x 2 int ˝ and thus verifies the inclusion Œa; b/ int ˝ .
Now we establish relationships between taking the interior and closure of convex sets.
Proposition 1.31 Let ˝ Rn be a convex set with nonempty interior. en we have:
(i) int ˝ D ˝ and .ii/ int ˝ D int ˝ .
Proof. (i) Obviously, int ˝ ˝ . For any b 2 ˝ and a 2 int ˝ , define the sequence fxk g by
1 1
xk WD aC 1 b; k 2 N:
k k
Lemma 1.30 ensures that xk 2 int ˝ . Since xk ! b , we have b 2 int ˝ and thus verify (i).
(ii) Since the inclusion int ˝ int ˝ is obvious, it remains to prove the opposite inclusion
int ˝ int ˝ . To proceed, fix any b 2 int ˝ and a 2 int ˝ . If > 0 is sufficiently small, then
1
c WD b C .b a/ 2 ˝ , and hence b D aC c 2 .a; c/ int ˝;
1C 1C
which verifies that int ˝ int ˝ and thus completes the proof.
1.3 CONVEX FUNCTIONS

is section collects basic facts about general (extended-real-valued) convex functions including
their analytic and geometric characterizations, important properties as well as their specifications
for particular subclasses. We also define convex set-valued mappings and use them to study a re-
markable class of optimal value functions employed in what follows.
Definition 1.32 Let f W ˝ ! R be an extended-real-valued function define on a convex set ˝

Rn . en the function f is  on ˝ if

f x C .1 /y f .x/ C .1 /f .y/ for all x; y 2 ˝ and 2 .0; 1/: (1.1)
1.3. CONVEX FUNCTIONS 11
If the inequality in (1.1) is strict for x ¤ y , then f is   on ˝ .
Given a function f W ˝ ! R, the extension of f to Rn is defined by

(
e.x/ WD f .x/ if x 2 ˝;
f
1 otherwise:
Obviously, if f is convex on ˝ , where ˝ is a convex set, then fe is convex on Rn . Furthermore,

if f W Rn ! R is a convex function, then it is also convex on every convex subset of Rn . is
allows to consider without loss of generality extended-real-valued convex functions on the whole
space Rn .
Figure 1.3: Convex function and nonconvex function.
Definition 1.33 e  and  of f W Rn ! R are defined, respectively, by

˚ ˇ
dom f WD x 2 Rn ˇ f .x/ < 1 and
˚ ˇ ˚ ˇ
epi f WD .x; t/ 2 Rn R ˇ x 2 Rn ; t f .x/ D .x; t / 2 Rn R ˇ x 2 dom f; t f .x/ :
Let us illustrate the convexity of functions by examples.
Example 1.34 e following functions are convex:

(i) f .x/ WD ha; xi C b for x 2 Rn , where a 2 Rn and b 2 R.
(ii) g.x/ WD kxk for x 2 Rn .
(iii) h.x/ WD x 2 for x 2 R.
Indeed, the function f in (i) is convex since

f x C .1 /y D ha; x C .1 /yi C b D ha; xi C .1 /ha; yi C b
D .ha; xi C b/ C .1 /.ha; yi C b/
D f .x/ C .1 /f .y/ for all x; y 2 Rn and 2 .0; 1/:
e function g in (ii) is convex since for x; y 2 Rn and 2 .0; 1/, we have

g x C .1 /y D kx C .1 /yk kxk C .1 /kyk D g.x/ C .1 /g.y/;
which follows from the triangle inequality and the fact that k˛uk D j˛j kuk whenever ˛ 2 R
and u 2 Rn . e convexity of the simplest quadratic function h in (iii) follows from a more general
result for the quadratic function on Rn given in the next example.
Example 1.35 Let A be an n n symmetric matrix. It is called positive semidefinite if hAu; ui 0
for all u 2 Rn . Let us check that A is positive semidefinite if and only if the function f W Rn ! R
defined by
1
f .x/ WD hAx; xi; x 2 Rn ;
2
is convex. Indeed, a direct calculation shows that for any x; y 2 Rn and 2 .0; 1/ we have
1
f .x/ C .1 /f .y/ f x C .1 /y D .1 /hA.x y/; x yi: (1.2)
2
If the matrix A is positive semidefinite, then hA.x y/; x yi 0, so the function f is convex
by (1.2). Conversely, assuming the convexity of f and using equality (1.2) for x D u and y D 0
verify that A is positive semidefinite.
e following characterization of convexity is known as the Jensen inequality.
eorem 1.36 A function f W Rn ! R is convex if and only if for any numbers i 0 as i D

P
1; : : : ; m with m n
iD1 i D 1 and for any elements xi 2 R , i D 1; : : : ; m, it holds that
X
m Xm
f i xi i f .xi /: (1.3)
iD1 iD1
Proof. Since (1.3) immediately implies the convexity of f , we only need to prove that any convex
function f satisfies the Jensen inequality (1.3). Arguing by induction and taking into account that
for m D 1 inequality (1.3) holds trivially and for m D 2 inequality (1.3) holds by the definition
of convexity, we suppose that it holds for an integer m WD k with k 2. Fix numbers i 0,
P
i D 1; : : : ; k C 1, with kC1 n
iD1 i D 1 and elements xi 2 R , i D 1; : : : ; k C 1. We obviously have
the relationship
Xk
i D 1 kC1 :
iD1
If kC1 D 1, then i D 0 for all i D 1; : : : ; k and (1.3) obviously holds for m WD k C 1 in this
case. Supposing now that 0 kC1 < 1, we get
k
X i
D1
1 kC1
iD1
and by direct calculations based on convexity arrive at

kC1 h Pk i
X i xi
iD1
f i xi D f .1 kC1 / C kC1 xkC1
1 kC1
iD1
P
k x
iD1 i i
.1 kC1 /f C kC1 f .xkC1 /
1 kC1
Xk
i
D .1 kC1 /f xi C kC1 f .xkC1 /
1 kC1
iD1
k
X i
.1 kC1 / f .xi / C kC1 f .xkC1 /
1 kC1
iD1
kC1
X
D i f .xi /:
iD1
is justifies inequality (1.3) and completes the proof of the theorem.
e next theorem gives a geometric characterization of the function convexity via the con-
vexity of the associated epigraphical set.
eorem 1.37 A function f W Rn ! R is convex if and only if its epigraph epi f is a convex
subset of the product space Rn R.
Proof. Assuming that f is convex, fix pairs .x1 ; t1 /; .x2 ; t2 / 2 epi f and a number 2 .0; 1/.
en we have f .xi / ti for i D 1; 2. us the convexity of f ensures that

f x1 C .1 /x2 f .x1 / C .1 /f .x2 / t1 C .1 /t2 :
is implies therefore that
.x1 ; t1 / C .1 /.x2 ; t2 / D .x1 C .1 /x2 ; t1 C .1 /t2 / 2 epi f;
and thus the epigraph epi f is a convex subset of Rn R.

Conversely, suppose that the set epi f is convex and fix x1 ; x2 2 dom f and a number
2 .0; 1/. en .x1 ; f .x1 //; .x2 ; f .x2 // 2 epi f . is tells us that

x1 C .1 /x2 ; f .x1 / C .1 /f .x2 / D x1 ; f .x1 / C .1 / x2 ; f .x2 / 2 epi f
and thus implies the inequality

f x1 C .1 /x2 f .x1 / C .1 /f .x2 /;
which justifies the convexity of the function f .
Figure 1.4: Epigraphs of convex function and nonconvex function.
Now we show that convexity is preserved under some important operations.
Proposition 1.38 Let fi W Rn ! R be convex functions for all i D 1; : : : ; m. en the following
functions are convex as well:
(i) e multiplication by scalars f for any > 0.
P
(ii) e sum function m iD1 fi .
(iii) e maximum function max1im fi .
Proof. e convexity of f in (i) follows directly from the definition. It is sufficient to prove (ii)
and (iii) for m D 2, since the general cases immediately follow by induction.
(ii) Fix any x; y 2 Rn and 2 .0; 1/. en we have

f1 C f2 x C .1 /y D f1 x C .1 /y C f2 x C .1 /y
f1 .x/ C .1 /f1 .y/ C f2 .x/ C .1 /f2 .y/
D .f1 C f2 /.x/ C .1 /.f1 C f2 /.y/;
which verifies the convexity of the sum function f1 C f2 .
(iii) Denote g WD maxff1 ; f2 g and get for any x; y 2 Rn and 2 .0; 1/ that

fi x C .1 /y fi .x/ C .1 /fi .y/ g.x/ C .1 /g.y/
for i D 1; 2. is directly implies that
˚
g x C .1 /y D max f1 x C .1 /y ; f2 x C .1 /y g.x/ C .1 /g.y/;
which shows that the maximum function g.x/ D maxff1 .x/; f2 .x/g is convex.
e next result concerns the preservation of convexity under function compositions.
Proposition 1.39 Let f W Rn ! R be convex and let ' W R ! R be nondecreasing and convex on
a convex set containing the range of the function f . en the composition ' ı f is convex.
Proof. Take any x1 ; x2 2 Rn and 2 .0; 1/. en we have by the convexity of f that

f x1 C .1 /x2 f .x1 / C .1 /f .x2 /:
Since ' is nondecreasing and it is also convex, it follows that

' ı f x1 C .1 /x2 D ' f x1 C .1 /x2
' f .x1 / C .1 /f .x2 /
' f .x1 / C .1 /' f .x2 /
D .' ı f /.x1 / C .1 /.' ı f /.x2 /;
which verifies the convexity of the composition ' ı f .

Now we consider the composition of a convex function and an affine mapping.
Proposition 1.40 Let B W Rn ! Rp be an affine mapping and let f W Rp ! R be a convex func-

tion. en the composition f ı B is convex.
Proof. Taking any x; y 2 Rn and 2 .0; 1/, we have

f ı B x C .1 /y D f .B.x C .1 /y// D f B.x/
C .1 /B.y/
f B.x/ C .1 /f B.y/ D .f ı B/.x/ C .1 /.f ı B/.y/
and thus justify the convexity of the composition f ı B .

e following simple consequence of Proposition 1.40 is useful in applications.
Corollary 1.41 Let f W Rn ! R be a convex function. For any x; N d 2 Rn , the function 'W R ! R
N d 2 Rn the function '
defined by '.t/ WD f .xN C td / is convex as well. Conversely, if for every x;
defined above is convex, then f is also convex.
Proof. Since B.t/ D xN C td is an affine mapping, the convexity of ' immediately follows from
Proposition 1.40. To prove the converse implication, take any x1 ; x2 2 Rn , 2 .0; 1/ and let
xN WD x2 , d WD x1 x2 . Since '.t/ D f .xN C td / is convex, we have

f x1 C .1 /x2 D f x2 C .x1 x2 / D './ D ' .1/ C .1 /.0/
'.1/ C .1 /'.0/ D f .x1 / C .1 /f .x2 /;
which verifies the convexity of the function f .

e next proposition is trivial while useful in what follows.
Proposition 1.42 Let f W Rn Rp ! R be convex. For .x; N 2 Rn Rp , the functions '.y/ WD

N y/
N y/ and .x/ WD f .x; y/
f .x; N are also convex.
Now we present an important extension of Proposition 1.38(iii).
Proposition 1.43 Let fi W Rn ! R for i 2 I be a collection of convex functions with a nonempty

index set I . en the supremum function f .x/ WD supi2I fi .x/ is convex.
Proof. Fix x1 ; x2 2 Rn and 2 .0; 1/. For every i 2 I , we have

fi x1 C .1 /x2 fi .x1 / C .1 /fi .x2 / f .x1 / C .1 /f .x2 /;
which implies in turn that

f x1 C .1 /x2 D sup fi x1 C .1 /x2 f .x1 / C .1 /f .x2 /:
i2I
is justifies the convexity of the supremum function.

Our next intention is to characterize convexity of smooth functions of one variable. To pro-
ceed, we begin with the following lemma.
Lemma 1.44 Given a convex function f W R ! R, assume that its domain is an open interval I .
For any a; b 2 I and a < x < b , we have the inequalities
f .x/ f .a/ f .b/ f .a/ f .b/ f .x/

:
x a b a b x
a x
Proof. Fix a; b; x as above and form the numbers t WD 2 .0; 1/. en
a b
x a
f .x/ D f a C .x a/ D f a C b a D f a C t.b a/ D f tb C .1 t/a :
b a
is gives us the inequalities f .x/ tf .b/ C .1 t/f .a/ and
x a
f .x/ f .a/ tf .b/ C .1 t/f .a/ f .a/ D t f .b/ f .a/ D f .b/ f .a/ ;
b a
which can be equivalently written as
f .x/ f .a/ f .b/ f .a/

:
x a b a
Similarly, we have the estimate
x b
f .x/ f .b/ tf .b/ C .1 t/f .a/ f .b/ D .t 1/ f .b/ f .a/ D f .b/ f .a/ ;
b a
which finally implies that
f .b/ f .a/ f .b/ f .x/

b a b x
and thus completes the proof of the lemma.
eorem 1.45 Suppose that f W R ! R is differentiable on its domain, which is an open interval
I . en f is convex if and only if the derivative f 0 is nondecreasing on I .
Proof. Suppose that f is convex and fix a < b with a; b 2 I . By Lemma 1.44, we have
f .x/ f .a/ f .b/ f .a/

x a b a
for any x 2 .a; b/. is implies by the derivative definition that
f .b/ f .a/
f 0 .a/ :
b a
Similarly, we arrive at the estimate
f .b/ f .a/
f 0 .b/
b a
and conclude that f 0 .a/ f 0 .b/, i.e., f 0 is a nondecreasing function.
To prove the converse, suppose that f 0 is nondecreasing on I and fix x1 < x2 with x1 ; x2 2
I and t 2 .0; 1/. en
x1 < x t < x2 for x t WD tx1 C .1 t/x2 :
By the mean value theorem, we find c1 ; c2 such that x1 < c1 < x t < c2 < x2 and
f .x t / f .x1 / D f 0 .c1 /.x t x1 / D f 0 .c1 /.1 t /.x2 x1 /;
f .x t / f .x2 / D f 0 .c2 /.x t x2 / D f 0 .c2 /t.x1 x2 /:
is can be equivalently rewritten as

tf .x t / tf .x1 / D f 0 .c1 /t.1 t /.x2 x1 /;
.1 t/f .x t / .1 t/f .x2 / D f 0 .c2 /t.1 t/.x1 x2 /:
Since f 0 .c1 / f 0 .c2 /, adding these equalities yields

f .x t / tf .x1 / C .1 t/f .x2 /;
which justifies the convexity of the function f .

Corollary 1.46 Let f W R ! R be twice differentiable on its domain, which is an open interval I .
en f is convex if and only if f 00 .x/ 0 for all x 2 I .
Proof. Since f 00 .x/ 0 for all x 2 I if and only if the derivative function f 0 is nondecreasing on
this interval. en the conclusion follows directly from eorem 1.45.
Example 1.47 Consider the function

8
<1 if x > 0;
f .x/ WD x
:1 otherwise:
2
To verify its convexity, we get that f 00 .x/ D
> 0 for all x belonging to the domain of f , which
x3
is I D .0; 1/. us this function is convex on R by Corollary 1.46.
Next we define the notion of set-valued mappings (or multifunctions), which plays an im-
portant role in modern convex analysis, its extensions, and applications.
Definition 1.48 We say that F is a -  between Rn and Rp and denote it by
F W Rn !
! Rp if F .x/ is a subset of Rp for every x 2 Rn . e  and  of F are defined,
respectively, by
˚ ˇ ˚ ˇ
dom F WD x 2 Rn ˇ F .x/ ¤ ; and gph F WD .x; y/ 2 Rn Rp ˇ y 2 F .x/ :
Any single-valued mapping F W Rn ! Rp is a particular set-valued mapping where the set

F .x/ is a singleton for every x 2 Rn . It is essential in the following definition that the mapping
F is set-valued.
Definition 1.49 Let F W Rn !

! Rp and let ' W Rn Rp ! R. e   or 
 associated with F and ' is defined by
˚ ˇ
.x/ WD inf '.x; y/ ˇ y 2 F .x/ ; x 2 Rn : (1.4)
roughout this section we assume that .x/ > 1 for every x 2 Rn .
Proposition 1.50 Assume that 'W Rn Rp ! R is a convex function and that F W Rn !

! Rp is of
convex graph. en the optimal value function in (1.4) is convex.
Figure 1.5: Graph of set-valued mapping.
Proof. Take x1 ; x2 2 dom , 2 .0; 1/. For any > 0, find yi 2 F .xi / such that
'.xi ; yi / < .xi / C for i D 1; 2:
It directly implies the inequalities

'.x1 ; y1 / < .x1 / C ; .1 /'.x2 ; y2 / < .1 /.x2 / C .1 /:
Summing up these inequalities and employing the convexity of ' yield

' x1 C .1 /x2 ; y1 C .1 /y2 '.x1 ; y1 / C .1 /'.x2 ; y2 /
< .x1 / C .1 /.x2 / C :
Furthermore, the convexity of gph F gives us

x1 C .1 /x2 ; y1 C .1 /y2 D .x1 ; y1 / C .1 /.x2 ; y2 / 2 gph F;
and therefore y1 C .1 /y2 2 F .x1 C .1 /x2 /. is implies that

x1 C .1 /x2 ' x1 C .1 /x2 ; y1 C .1 /y2 < .x1 / C .1 /.x2 / C :
Letting finally ! 0 ensures the convexity of the optimal value function .

Using Proposition 1.50, we can verify convexity in many situations. For instance, given
two convex functions fi W Rn ! R, i D 1; 2, let '.x; y/ WD f1 .x/ C y and F .x/ WD Œf2 .x/; 1/.
en the function ' is convex and set gph F D epi f2 is convex as well, and hence we justify the
convexity of the sum
.x/ D inf '.x; y/ D f1 .x/ C f2 .x/:
y2F .x/
Another example concerns compositions. Let f W Rp ! R be convex and let B W Rn ! Rp be
affine. Define '.x; y/ WD f .y/ and F .x/ WD fB.x/g. Observe that ' is convex while F is of con-
vex graph. us we have the convex composition

.x/ D inf '.x; y/ D f B.x/ ; x 2 Rn :
y2F .x/
e examples presented above recover the results obtained previously by direct proofs. Now we
establish via Proposition 1.50 the convexity of three new classes of functions.
Proposition 1.51 Let ' W Rp ! R be convex and let B W Rp ! Rn be affine. Consider the set-
valued inverse image mapping B 1 W Rn !! Rp , define
˚ ˇ
f .x/ WD inf '.y/ ˇ y 2 B 1 .x/ ; x 2 Rn ;
and suppose that f .x/ > 1 for all x 2 Rn . en f is a convex function.
1
Proof. Let '.x; y/ '.y/ and F .x/ WD B .x/. en the set
˚ ˇ
gph F D .u; v/ 2 Rn Rp ˇ B.v/ D u
is obviously convex. Since we have the representation
f .x/ D inf '.y/; x 2 Rn ;

y2F .x/
the convexity of f follows directly from Proposition 1.50.
Proposition 1.52 For convex functions f1 ; f2 W Rn ! R, define the  

˚ ˇ
.f1 ˚ f2 /.x/ WD inf f1 .x1 / C f2 .x2 / ˇ x1 C x2 D x
and suppose that .f1 ˚ f2 /.x/ > 1 for all x 2 Rn . en f1 ˚ f2 is also convex.
Proof. Define ' W Rn Rn ! R by '.x1 ; x2 / WD f1 .x1 / C f2 .x2 / and B W Rn Rn ! Rn by

B.x1 ; x2 / WD x1 C x2 . We have
˚ ˇ
inf '.x1 ; x2 / ˇ .x1 ; x2 / 2 B 1 .x/ D .f1 ˚ f2 /.x/ for all x 2 Rn ;
which implies the convexity of .f1 ˚ f2 / by Proposition 1.51.

1.4. RELATIVE INTERIORS OF CONVEX SETS 21
p
Definition 1.53 A function g W R ! R is called   if

xi yi for all i D 1; : : : ; p H) g.x1 ; : : : ; xp / g.y1 ; : : : ; yp / :
Now we are ready to present the final consequence of Proposition 1.50 in this section that
involves the composition.
Proposition 1.54 Define h W Rn ! Rp by h.x/ WD .f1 .x/; : : : ; fp .x//, where fi W Rn ! R for

i D 1; : : : ; p are convex functions. Suppose that g W Rp ! R is convex and nondecreasing componen-
twise. en the composition g ı h W Rn ! R is a convex function.
Proof. Let F W Rn !
! Rp be a set-valued mapping defined by
F .x/ WD Œf1 .x/; 1/ Œf2 .x/; 1/ : : : Œfp .x/; 1/:
en the graph of F is represented by

˚ ˇ
gph F D .x; t1 ; : : : ; tp / 2 Rn Rp ˇ ti fi .x/ :
Since all fi are convex, the set gph F is convex as well. Define further ' W Rn Rp ! R by
'.x; y/ WD g.y/ and observe, since g is increasing componentwise, that
˚ ˇ
inf '.x; y/ ˇ y 2 F .x/ D g f1 .x/; : : : ; fp .x/ D .g ı h/.x/;
which ensures the convexity of the composition g ı h by Proposition 1.50.
1.4 RELATIVE INTERIORS OF CONVEX SETS

We begin this section with the definition and properties of affine sets. Given two elements a and
b in Rn , the line connecting them is
˚ ˇ
LŒa; b WD a C .1 /b ˇ 2 R :
Note that if a D b , then LŒa; b D fag.
Definition 1.55 A subset ˝ of Rn is  if for any a; b 2 ˝ we have LŒa; b ˝ .
For instance, any point, line, and plane in R3 are affine sets. e empty set and the whole
space are always affine. It follows from the definition that the intersection of any collection of
affine sets is affine. is leads us to the construction of the affine hull of a set.
Definition 1.56 e   of a set ˝ Rn is

\˚ ˇ
aff ˝ WD C ˇ C is affine and ˝ C :
An element x in Rn of the form
m
X m
X
xD i !i with i D 1; m 2 N;
iD1 iD1
is called an affine combination of !1 ; : : : ; !m . e proof of the next proposition is straightforward

and thus is omitted.
Proposition 1.57 e following assertions hold:

(i) A set ˝ is affine if and only if ˝ contains all affine combinations of its elements.
(ii) Let ˝ , ˝1 , and ˝2 be affine subsets of Rn . en the sum ˝1 C ˝2 and the scalar product ˝ for
any 2 R are also affine subsets of Rn .
(iii) Let B W Rn ! Rp be an affine mapping. If ˝ is an affine subset of Rn and is an affine subset
of Rp , then the image B.˝/ is an affine subset of Rp and the inverse image B 1 ./ is an affine subset
of Rn .
(iv) Given ˝ Rn , its affine hull is the smallest affine set containing ˝ . We have
nX
m ˇ X
m o
ˇ
aff ˝ D i !i ˇ i D 1; !i 2 ˝; m 2 N :
iD1 iD1
(v) A set ˝ is a .linear/ subspace if and only if ˝ is an affine set containing the origin.
Next we consider relationships between affine sets and (linear) subspaces.
Lemma 1.58 A nonempty subset ˝ of Rn is affine if and only if ˝ ! is a subspace of Rn for any
! 2 ˝.
Proof. Suppose that a nonempty set ˝ Rn is affine. It follows from Proposition 1.57(v) that
˝ ! is a subspace for any ! 2 ˝ . Conversely, fix ! 2 ˝ and suppose that ˝ ! is a subspace
denoted by L. en the set ˝ D ! C L is obviously affine.
e preceding lemma leads to the following notion.
Definition 1.59 An affine set ˝ is  to a subspace L if ˝ D ! C L for some ! 2 ˝ .
e next proposition justifies the form of the parallel subspace.
Proposition 1.60 Let ˝ be a nonempty, affine subset of Rn . en it is parallel to the unique subspace
L of Rn given by L D ˝ ˝ .
Proof. Given a nonempty, affine set ˝ , fix ! 2 ˝ and come up to the linear subspace L WD ˝ !
parallel to ˝ . To justify the uniqueness of such L, take any !1 ; !2 2 ˝ and any subspaces
L1 ; L2 Rn such that ˝ D !1 C L1 D !2 C L2 . en L1 D !2 !1 C L2 . Since 0 2 L1 , we
have !1 !2 2 L2 . is implies that !2 !1 2 L2 and thus L1 D !2 !1 C L2 L2 . Simi-
larly, we get L2 L1 , which justifies that L1 D L2 .
It remains to verify the representation L D ˝ ˝ . Let ˝ D ! C L with the unique sub-
space L and some ! 2 ˝ . en L D ˝ ! ˝ ˝ . Fix any x D u ! with u; ! 2 ˝ and
observe that ˝ ! is a subspace parallel to ˝ . Hence ˝ ! D L by the uniqueness of L proved
above. is ensures that x 2 ˝ ! D L and thus ˝ ˝ L.
e uniqueness of the parallel subspace shows that the next notion is well defined.
Definition 1.61 e      ; ¤ ˝ Rn is the dimension of the linear

subspace parallel to ˝ . Furthermore, the      ; ¤ ˝ Rn is the dimension
of its affine hull aff ˝ .
To proceed further, we need yet another definition important in what follows.
Definition 1.62 e elements v0 ; : : : ; vm in Rn , m 1, are   if

hX
m m
X i
i vi D 0; i D 0 H) i D 0 for all i D 0; : : : ; m :
iD0 iD0
It is easy to observe the following relationship with the linear independence.
Proposition 1.63 e elements v0 ; : : : ; vm in Rn are affinely independent if and only if the elements
v1 v0 ; : : : ; vm v0 are linearly independent.
Proof. Suppose that v0 ; : : : ; vm are affinely independent and consider the system
m
X m
X
i .vi v0 / D 0; i.e., 0 v0 C i vi D 0;
iD1 iD1
Pm Pm
where 0 WD iD1 i . Since the elements v0 ; : : : ; vm are affinely independent and iD0 i D
0, we have that i D 0 for all i D 1; : : : ; m. us v1 v0 ; : : : ; vm v0 are linearly independent.
e proof of the converse statement is straightforward.
Recall that the span of some set C , span C , is the linear subspace generated by C .
Lemma 1.64 Let ˝ WD afffv0 ; : : : ; vm g, where vi 2 Rn for all i D 0; : : : ; m. en the span of the
set fv1 v0 ; : : : ; vm v0 g is the subspace parallel to ˝ .
Proof. Denote by L the subspace parallel to ˝ . en ˝ v0 D L and therefore vi v0 2 L for
all i D 1; : : : ; m. is gives
˚ ˇ
span vi v0 ˇ i D 1; : : : ; m L:
To prove the converse inclusion, fix any v 2 L and get v C v0 2 ˝ . us we have

m
X m
X
v C v0 D i vi ; i D 1:
iD0 iD0
is implies the relationship

m
X ˚ ˇ
vD i .vi v0 / 2 span vi v0 ˇ i D 1; : : : ; m ;
iD1
which justifies the converse inclusion and hence completes the proof.
e proof of the next proposition is rather straightforward.
Proposition 1.65 e elements v0 ; : : : ; vm are affinely independent in Rn if and only if its affine hull
˝ WD afffv0 ; : : : ; vm g is m-dimensional.
Proof. Suppose that v0 ; : : : ; vm are affinely independent. en Lemma 1.64 tells us that the
subspace L WD spanfvi v0 j i D 1; : : : ; mg is parallel to ˝ . e linear independence of v1
v0 ; : : : ; vm v0 by Proposition 1.63 means that the subspace L is m-dimensional and so is ˝ .
e proof of the converse statement is also straightforward.
Affinely independent systems lead us to the construction of simplices.
Definition 1.66 Let v0 ; : : : ; vm be affinely independent in Rn . en the set

˚ ˇ
m WD co vi ˇ i D 0; : : : ; m
is called an m- in Rn with the vertices vi , i D 0; : : : ; m.
An important role of simplex vertices is revealed by the following proposition.
Proposition 1.67 Consider an m-simplex m with vertices vi for i D 0; : : : ; m. For every v 2 m ,

there is a unique element .0 ; : : : ; m / 2 RmC1
C such that
m
X m
X
vD i vi ; i D 1:
iD0 iD0
Proof. Let .0 ; : : : ; m / 2 RmC1
C and .0 ; : : : ; m / 2 mC1
RC satisfy
m
X m
X m
X m
X
vD i vi D i vi ; i D i D 1:
iD0 iD0 iD0 iD0
is immediately implies the equalities

m
X m
X
.i i /vi D 0; .i i / D 0:
iD0 iD0
Since v0 ; : : : ; vm are affinely independent, we have i D i for i D 0; : : : ; m.

Now we are ready to define a major notion of relative interiors of convex sets.
Definition 1.68 Let ˝ be a convex set. We say that an element v 2 ˝ belongs to the 
 ri ˝ of ˝ if there exists > 0 such that IB.vI / \ aff ˝ ˝ .
We begin the study of relative interiors with the following lemma.
Lemma 1.69 Any linear mapping A W Rn ! Rp is continuous.
Proof. Let fe1 ; : : : ; en g be the standard orthonormal basis of Rn and let vi WD A.ei /, i D 1; : : : ; n.
For any x 2 Rn with x D .x1 ; : : : ; xn /, we have
X
n Xn n
X
A.x/ D A x i ei D xi A.ei / D xi vi :
iD1 iD1 iD1
en the triangle inequality and the Cauchy-Schwarz inequality give us

v v
n u n u n
X uX uX
kA.x/k jxi jkvi k t jxi j t
2 kvi k2 D M kxk;
iD1 iD1 iD1
qP
n
where M WD iD1 kvi k2 . It follows furthermore that
kA.x/ A.y/k D kA.x y/k M kx yk for all x; y 2 Rn ;
which implies the continuity of the mapping A.

e next proposition plays an essential role in what follows.
Proposition 1.70 Let m be an m-simplex in Rn with some m 1. en ri m ¤ ;.

Proof. Consider the vertices v0 ; : : : ; vm of the simplex m and denote
m
1 X
v WD vi :
mC1
iD0
We prove the proposition by showing that v 2 ri m . Define

˚ ˇ
L WD span vi v0 ˇ i D 1; : : : ; m
and observe that L is the m-dimensional subspace of Rn parallel to aff m D afffv0 ; : : : ; vm g. It

is easy to see that for every x 2 L there is a unique collection .0 ; : : : ; m / 2 RmC1 with
m
X m
X
xD i vi ; i D 0:
iD0 iD0
Consider the mapping A W L ! RmC1 , which maps x to the corresponding coefficients

.0 ; : : : ; m / 2 RmC1 as above. en A is linear, and hence it is continuous by Lemma 1.69.
Since A.0/ D 0, we can choose ı > 0 such that
1
kA.u/k < whenever kuk ı:
mC1
Let us now show that .v C ıIB/ \ aff m m , which means that v 2 ri m . To proceed, fix any
x 2 .v C ıIB/ \ aff m and get that x D v C u for some u 2 ıIB . Since v; x 2 aff m and u D
P
x v , we have u 2 L. Denoting A.u/ WD .˛0 ; : : : ; ˛m / gives us the representation u D m
iD0 ˛i vi
Pm
with iD0 ˛i D 0 and the estimate
1
j˛i j kA.u/k < for all i D 0; : : : ; m:
mC1
en implies in turn that
m
X X m
1
vCuD . C ˛i /vi D i vi ;
mC1
iD0 iD0
1 P
where i WD C ˛i 0 for i D 0; : : : ; m. Since m
iD0 i D 1, this ensures that x 2 m .
mC1
us .v C ıIB/ \ aff m m and therefore v 2 ri m .
Lemma 1.71 Let ˝ be a nonempty, convex set in Rn of dimension m 1. en there exist m C 1
affinely independent elements v0 ; : : : ; vm in ˝ .
Proof. Let k WD fv0 ; : : : ; vk g be a k simplex of maximal dimension contained in ˝ . en
v0 ; : : : ; vk are affinely independent. To verify now that k D m, form K WD afffv0 ; : : : ; vk g and
observe that K aff ˝ since fv0 ; : : : ; vk g ˝ . e opposite inclusion also holds since we have
˝ K . Justifying it, we argue by contradiction and suppose that there exists w 2 ˝ such that
w … K . en a direct application of the definition of affine independence shows that v0 ; : : : ; vk ; w
are affinely independent being a subset of ˝ , which is a contradiction. us K D aff ˝ , and hence
we get k D dim K D dim aff ˝ D dim ˝ D m.
e next is one of the most fundamental results of convex finite-dimensional geometry.
eorem 1.72 Let ˝ be a nonempty, convex set in Rn . e following assertions hold:

(i) We always have ri ˝ ¤ ;.
(ii) We have Œa; b/ ri ˝ for any a 2 ri ˝ and b 2 ˝ .
Proof. (i) Let m be the dimension of ˝ . Observe first that the case where m D 0 is trivial since in
this case ˝ is a singleton and ri ˝ D ˝ . Suppose that m 1 and find m C 1 affinely independent
elements v0 ; : : : ; vm in ˝ as in Lemma 1.71. Consider further the m-simplex
˚
m WD co v0 ; : : : ; vm :
We can show that aff m D aff ˝ . To complete the proof, take v 2 ri m , which exists by Propo-
sition 1.70, and get for any small > 0 that
IB.v; / \ aff ˝ D IB.v; / \ aff m m ˝:
is verifies that v 2 ri ˝ by the definition of relative interior.

(ii) Let L be the subspace of Rn parallel to aff ˝ and let m WD dim L. en there is a bijective
linear mapping A W L ! Rm such that both A and A 1 are continuous. Fix x0 2 aff ˝ and define
the mapping f W aff ˝ ! Rm by f .x/ WD A.x x0 /. It is easy to check that f is a bijective
affine mapping and that both f and f 1 are continuous. We also see that a 2 ri ˝ if and only if
f .a/ 2 int f .˝/, and that b 2 ˝ if and only if f .b/ 2 f .˝/. en Œf .a/; f .b// int f .˝/ by
Lemma 1.30. is shows that Œa; b/ ri ˝ .
We conclude this section by the following properties of the relative interior.
Proposition 1.73 Let ˝ be a nonempty, convex subset of Rn . For the convex sets ri ˝ and ˝ , we
have that (i) ri ˝ D ˝ and (ii) ri ˝ D ri ˝ .
Proof. Note that the convexity of ri ˝ follows from eorem 1.72(ii) while the convexity of ˝ was
proved in Proposition 1.29. To justify assertion (i) of this proposition, observe that the inclusion
ri ˝ ˝ is obvious. Fix b 2 ˝ , choose a 2 ri ˝ , and form the sequence
1 1
xk WD aC 1 b; k 2 N;
k k
which converges to b as k ! 1. Since xk 2 ri ˝ by eorem 1.72(ii), we have b 2 ri ˝ . us
˝ ri ˝ , which verifies (i). e proof (ii) is similar to that of Proposition 1.31(ii).
1.5 THE DISTANCE FUNCTION

e last section of this chapter is devoted to the study of the class of distance functions for con-
vex sets, which belongs to the most interesting and important subjects of convex analysis and
its extensions. Functions of this type are intrinsically nondifferentiable while they naturally and
frequently appear in analysis and applications.
Figure 1.6: Distance function.
Given a set ˝ Rn , the distance function associated with ˝ is defined by

˚ ˇ
d.xI ˝/ WD inf kx !k ˇ ! 2 ˝ : (1.5)
Recall further that a mapping f W Rn ! Rp is Lipschitz continuous with constant ` 0 on some

set C Rn if we have the estimate
kf .x/ f .y/k `kx yk for all x; y 2 C: (1.6)
Note that the Lipschitz continuity of f in (1.6) specifies its continuity with a linear rate.
Proposition 1.74 Let ˝ be a nonempty subset of Rn . e following hold:

1.5. THE DISTANCE FUNCTION 29
(i) d.xI ˝/ D 0 if and only if x 2 ˝ .
(ii) e function d.xI ˝/ is Lipschitz continuous with constant ` D 1 on Rn .
Proof. (i) Suppose that d.xI ˝/ D 0. For each k 2 N , find !k 2 ˝ such that
1 1
0 D d.xI ˝/ kx !k k < d.xI ˝/ C D ;
k k
which ensures that the sequence f!k g converges to x , and hence x 2 ˝ .
Conversely, let x 2 ˝ and find a sequence f!k g ˝ converging to x . en
0 d.xI ˝/ kx !k k for all k 2 N;
which implies that d.xI ˝/ D 0 since kx !k k ! 0 as k ! 1.

(ii) For any ! 2 ˝ , we have the estimate
d.xI ˝/ kx !k kx yk C ky !k;

˚ ˇ
d.xI ˝/ kx yk C inf ky !k ˇ ! 2 ˝ D kx yk C d.yI ˝/:
Similarly, we get d.yI ˝/ ky xk C d.xI ˝/ and thus jd.xI ˝/ d.yI ˝/j kx yk, which
justifies by (1.6) the Lipschitz continuity of d.xI ˝/ on Rn with constant ` D 1.
For each x 2 Rn , the Euclidean projection from x to ˝ is defined by
˚ ˇ
˘.xI ˝/ WD ! 2 ˝ ˇ kx !k D d.xI ˝/ : (1.7)
Proposition 1.75 Let ˝ be a nonempty, closed subset of Rn . en for any x 2 Rn the Euclidean
projection ˘.xI ˝/ is nonempty.
Proof. By definition (1.7), for each k 2 N there exists !k 2 ˝ such that
1
d.xI ˝/ kx !k k < d.xI ˝/ C :
k
It is clear that f!k g is a bounded sequence. us it has a subsequence f!k` g that converges to ! .
Since ˝ is closed, ! 2 ˝ . Letting ` ! 1 in the inequality
1
d.xI ˝/ kx !k` k < d.xI ˝/ C ;
k`
we have d.xI ˝/ D kx !k, which ensures that ! 2 ˘.xI ˝/.
An interesting consequence of convexity is the following unique projection property.
Corollary 1.76 If ˝ is a nonempty, closed, convex subset of Rn , then for each x 2 Rn the Euclidean
projection ˘.xI ˝/ is a singleton.
Proof. e nonemptiness of the projection ˘.xI ˝/ follows from Proposition 1.75. To prove the
uniqueness, suppose that !1 ; !2 2 ˘.xI ˝/ with !1 ¤ !2 . en
kx !1 k D kx !2 k D d.xI ˝/:
By the classical parallelogram equality, we have that

!1 C !2 2
2 k!1 !2 k
2kx !1 k2 D kx !1 k2 C kx !2 k2 D 2 x C :
2 2
is directly implies that
!1 C !2 !2 k2 2
2 k!1
x D kx !1 k2 < kx !1 k2 D d.xI ˝/ ;
2 4
!1 C !2
which is a contradiction due to the inclusion 2 ˝.
2
Now we show that the convexity of a nonempty, closed set and its distance function are
equivalent. It is an easy exercise to show that the convexity of an arbitrary set ˝ implies the
convexity of its distance function.
Proposition 1.77 Let ˝ be a nonempty, closed subset of Rn . en the function d.I ˝/ is convex if
and only if the set ˝ is convex.
Proof. Suppose that ˝ is convex. Taking x1 ; x2 2 Rn and !i WD ˘.xi I ˝/, we have
kxi !i k D d.xi I ˝/ for i D 1; 2:
e convexity of ˝ ensures that !1 C .1 /!2 2 ˝ for any 2 .0; 1/. It yields

d x1 C .1 /x2 I ˝ kx1 C .1 /x2 Œ!1 C .1 /!2 k
kx1 !1 k C .1 /kx2 !2 k
D d.x1 I ˝1 / C .1 /d.x2 I ˝2 /;
which implies therefore the convexity of the distance function d.I ˝/ by

d x1 C .1 /x2 I ˝ d.x1 I ˝/ C .1 /d.x2 I ˝/:
1.5. THE DISTANCE FUNCTION 31
To prove the converse implication, suppose that d.I ˝/ is convex and fix any !i 2 ˝ for i D 1; 2
and 2 .0; 1/. en we have

d !1 C .1 /!2 I ˝ d.!1 I ˝/ C .1 /d.!2 I ˝/ D 0:
Since ˝ is closed, this yields !1 C .1 /!2 2 ˝ and so justifies the convexity of ˝ .
Next we characterize the Euclidean projection to convex sets in Rn . In the proposition

below and in what follows we often identify the projection ˘.xI ˝/ with its unique element if ˝
is a nonempty, closed, convex set.
Figure 1.7: Euclidean projection.
Proposition 1.78 Let ˝ be a nonempty, convex subset of Rn and let !N 2 ˝ . en we have !N 2
N ˝/ if and only if
˘.xI
hxN !;
N ! !i N 0 for all ! 2 ˝: (1.8)
Proof. Take !N 2 ˘.xI

N ˝/ and get for any ! 2 ˝ , 2 .0; 1/ that !N C .! N 2 ˝ . us
!/
2
kxN N 2 D d.xI
!k N ˝/ kxN N 2
!N C .! !/k
N 2 2hxN !;
D kxN !k N ! !iN C 2 k! N 2:
!k
is readily implies that

2hxN !;
N ! !i
N k! N 2:
!k
Letting ! 0C , we arrive at (1.8).

To verify the converse, suppose that (1.8) holds. en for any ! 2 ˝ we get
kxN !k2 D kxN !N C !N !k2

D kxN N 2 C k!N !k2 C 2hxN
!k !;
N !N !i
D kxN N 2 C k!N !k2 2hxN
!k !;
N ! !i
N kxN N 2:
!k
us kxN !k
N kxN !k for all ! 2 ˝ , which implies !N 2 ˘.xI
N ˝/ and completes the proof.

We know from the above that for any nonempty, closed, convex set ˝ in Rn the Euclidean
projection ˘.xI ˝/ is a singleton. Now we show that the projection mapping is in fact nonexpan-
sive, i.e., it satisfies the following Lipschitz property.
Proposition 1.79 Let ˝ be a nonempty, closed, convex subset of Rn . en for any elements x1 ; x2 2
Rn we have the estimate
˝ ˛
˘.x1 I ˝/ ˘.x2 I ˝/ 2 ˘.x1 I ˝/ ˘.x2 I ˝/; x1 x2 :
In particular, it implies the Lipschitz continuity of the projection with constant ` D 1:

˘.x1 I ˝/ ˘.x2 I ˝/ kx1 x2 k for all x1 ; x2 2 Rn :
Proof. It follows from the preceding proposition that

˝ ˛
˘.x2 I ˝/ ˘.x1 I ˝/; x1 ˘.x1 I ˝/ 0 for all x1 ; x2 2 Rn :
Changing the roles of x1 ; x2 in the inequality above and summing them up give us
˝ ˛
˘.x1 I ˝/ ˘.x2 I ˝/; x2 x1 C ˘.x1 I ˝/ ˘.x2 I ˝/ 0:
is implies the first estimate in the proposition. Finally, the nonexpansive property of the Eu-
clidean projection follows directly from
˝ ˛
˘.x1 I ˝/ ˘.x2 I ˝/ 2 ˘.x1 I ˝/ ˘.x2 I ˝/; x1 x2
k˘.x1 I ˝/ ˘.x2 I ˝/k kx1 x2 k
for all x1 ; x2 2 Rn , which completes the proof of the proposition.
1.6 EXERCISES FOR CHAPTER 1
Exercise 1.1 Let ˝1 and ˝2 be nonempty, closed, convex subsets of Rn such that ˝1 is bounded
or ˝2 is bounded. Show that ˝1 ˝2 is a nonempty, closed, convex set. Give an example of two
nonempty, closed, convex sets ˝1 ; ˝2 for which ˝1 ˝2 is not closed.
1.6. EXERCISES FOR CHAPTER 1 33
n
Exercise 1.2 Let ˝ be a subset of R . We say that ˝ is a cone if x 2 ˝ whenever 0 and
x 2 ˝ . Show that the following are equivalent:
(i) ˝ is a convex cone.
(ii) x C y 2 ˝ whenever x; y 2 ˝ , and x 2 ˝ whenever x 2 ˝ and 0.
Exercise 1.3 (i) Let ˝ be a nonempty, convex set that contains 0 and let 0 1 2 . Show
that 1 ˝ 2 ˝:
(ii) Let ˝ be a nonempty, convex set and let ˛; ˇ 0. Show that ˛˝ C ˇ˝ .˛ C ˇ/˝ .
Exercise 1.4 (i) Let ˝i for i D 1; : : : ; m be nonempty, convex sets in Rn . Show that
S
x 2 co m iD1 ˝i if and only P if there exist elements !i 2 ˝i and i 0 for i D 1; : : : ; m with
Pm m

iD1 i D 1 such that x D iD1 i !i .
(ii) Let ˝i for i D 1; : : : ; m be nonempty, convex cones in Rn . Show that
m
X m
[
˝i D co f ˝i g:
iD1 iD1
Exercise 1.5 Let ˝ be a nonempty, convex cone in Rn . Show that ˝ is a linear subspace of Rn
if and only if ˝ D ˝ .
Exercise 1.6 Show that the following functions are convex on Rn :

(i) f .x/ D ˛kxk, where ˛ 0.
(ii) f .x/ D kx ak2 , where a 2 Rn .
(iii) f .x/ D kAx bk, where A is an p n matrix and b 2 Rp .
(iv) f .x/ D kxkq , where q 1.
Exercise 1.7 Show that the following functions are convex on the given domains:
(i) f .x/ D e ax ; x 2 R, where a is a constant.
(ii) f .x/ D x q , x 2 Œ0; 1/, where q 1 is a constant.
(iii) f .x/ D ln.x/, x 2 .0; 1/.
(iv) f .x/ D x ln.x/, x 2 .0; 1/.
Exercise 1.8 (i) Give an example of a function f W Rn Rp ! R, which is convex with respect
to each variable but not convex with respect to both.
(ii) Let fi W Rn ! R for i D 1; 2 be convex functions. Can we conclude that the minimum function
minff1 ; f2 g.x/ WD minff1 .x/; f2 .x/g is convex?
Exercise 1.9 Give an example showing that the product of two real-valued convex functions is
not necessarily convex.
Exercise 1.10 Verify that the set f.x; t/ 2 Rn R j t kxkg is closed and convex.
Exercise 1.11 e indicator function associated with a set ˝ is defined by

(
0 if x 2 ˝;
ı.xI ˝/ WD
1 otherwise:
(i) Calculate ı.I ˝/ for ˝ D ;, ˝ D Rn , and ˝ D Œ 1; 1.
(ii) Show that the set ˝ is convex set if and only if its indicator function ı.I ˝/ is convex.
Exercise 1.12 Show that if f W R ! Œ0; 1/ is a convex function, then its q -power f q .x/ WD
.f .x//q is also convex for any q 1.
Exercise 1.13 Let fi W Rn ! R for i D 1; 2 be convex functions. Define

n
˝1 WD f.x;
˚ 1 ; 2 / 2 R n R R j 1 f1 .x/g;
˝2 WD .x; 1 ; 2 / 2 R R R j 2 f2 .x/g:
(i) Show that the sets ˝1 ; ˝2 are convex.
(ii) Define the set-valued mapping F W Rn ! ! R2 by F .x/ WD Œf1 .x/; 1/ Œf2 .x/; 1/ and verify
that the graph of F is ˝1 \ ˝2 .
Exercise 1.14 We say that f W Rn ! R is positively homogeneous if f .˛x/ D ˛f .x/ for all ˛ >
0, and that f is subadditive if f .x C y/ f .x/ C f .y/ for all x; y 2 Rn . Show that a positively
homogeneous function is convex if and only if it is subadditive.
Exercise 1.15 Given a nonempty set ˝ Rn , define

˚ ˇ [
K˝ WD x ˇ 0; x 2 ˝ D ˝:
0
(i) Show that K˝ is a cone.

(ii) Show that K˝ is the smallest cone containing ˝ .
(iii) Show that if ˝ is convex, then the cone K˝ is convex as well.
Exercise 1.16 Let ' W Rn Rp ! R be a convex function and let K be a nonempty, convex
subset of Rp . Suppose that for each x 2 Rn the function '.x; / is bounded below on K . Verify
that the function on Rn defined by f .x/ WD inff'.x; y/ j y 2 Kg is convex.
Exercise 1.17 Let f W Rn ! R be a convex function.

(i) Show that for every ˛ 2 R the level set fx 2 Rn j f .x/ ˛g is convex.
1
(ii) Let ˝ R be a convex set. Is it true in general that the inverse image f .˝/ is a convex
subset of Rn ?
nC1
Exercise 1.18 Let C be a convex subset of R .
(i) For x 2 Rn , define F .x/ WD f 2 R j .x; / 2 C g. Show that F is a set-valued mapping with
convex graph and give an explicit formula for its graph.
(ii) For x 2 Rn , define the function
˚ ˇ
fC .x/ WD inf 2 R ˇ .x; / 2 C : (1.9)
Find an explicit formula for fC when C is the closed unit ball of R2 .

(iii) For the function fC defined in (ii), show that if fC .x/ > 1 for all x 2 Rn , then fC is a
convex function.
Exercise 1.19 Let f W Rn ! R be bounded from below. Define its convexification

nX
m ˇ m
X m
X o
ˇ
.co f /.x/ WD inf i f .xi / ˇ i 0; i D 1; i xi D x; m 2 N : (1.10)
iD1 iD1 iD1
Verify that (1.10) is convex with co f D fC , C WD co .epi f /, and fC defined in (1.9).
Exercise 1.20 We say that f W Rn ! R is quasiconvex if

˚
f x C .1 /y/ max f .x/; f .y/ for all x; y 2 Rn ; 2 .0; 1/:
(i) Show that a function f is quasiconvex if and only if for any ˛ 2 R the level set fx 2
Rn j f .x/ ˛g is a convex set.
(ii) Show that any convex function f W Rn ! R is quasiconvex. Give an example demonstrating
that the converse is not true.
Exercise 1.21 Let a be a nonzero element in Rn and let b 2 R. Show that

˚ ˇ
˝ WD x 2 Rn ˇ ha; xi D b
is an affine set with dim ˝ D 1.
Exercise 1.22 Let the set fv1 ; : : : ; vm g consist of affinely independent elements and let v …
aff fv1 ; : : : ; vm g. Show that v1 ; : : : ; vm ; v are affinely independent.
Exercise 1.23 Suppose that ˝ is a convex subset of Rn with dim ˝ D m, m 1. Let the set
fv1 ; : : : ; vm g ˝ consist of affinely independent elements and let
˚
m WD co v1 ; : : : ; vm :
˚
Show that aff ˝ D aff m D aff v1 ; : : : ; vm .
Exercise 1.24 Let ˝ be a nonempty, convex subset of Rn .
(i) Show that aff ˝ D aff ˝ .
(ii) Show that for any xN 2 ri ˝ and x 2 ˝ there exists t > 0 such that xN C t.xN x/ 2 ˝ .
(iii) Prove Proposition 1.73(ii).
Exercise 1.25 Let ˝1 ; ˝2 Rn be convex sets with ri ˝1 \ ri ˝2 ¤ ;. Show that:

(i) ˝1 \ ˝2 D ˝ 1 \ ˝ 2 .
(ii) ri .˝1 \ ˝2 / D ri ˝1 \ ri˝2 .
Exercise 1.26 Let ˝1 ; ˝2 Rn be convex sets with ˝ 1 D ˝ 2 . Show that ri ˝1 D ri ˝2 .

Exercise 1.27 (i) Let B W Rn ! Rp be an affine mapping and let ˝ be a convex subset of Rn .
Prove the equality
B.ri ˝/ D ri B.˝/:
(ii) Let ˝1 and ˝2 be convex subsets of Rn . Show that ri .˝1 ˝2 / D ri ˝1 ri ˝2 .
Exercise 1.28 Let f W Rn ! R be a convex function. Show that:

(i) aff .epi f / D ˚aff .domˇ f / R :
(ii) ri .epi f / D .x; / ˇ x 2 ri .dom f /; > f .x/ :
Exercise 1.29 Find the explicit formulas for the distance function d.xI ˝/ and the Euclidean
projection ˘.xI ˝/ in the following cases:
n
(i) ˝ is the
˚ closed unit ball
ˇ of R .
2ˇ
(ii) ˝ WD .x1 ; x2 / 2 R jx1 j C jx2 j 1 .
(iii) ˝ WD Œ 1; 1 Œ 1; 1.
Exercise 1.30 Find the formulas for the projection ˘.xI ˝/ in the following cases:
˚ ˇ
(i) ˝ WD ˚x 2 Rn ˇˇha; xi D b , where 0 ¤ a 2 Rn and b 2 R :
(ii) ˝ WD x 2 Rn ˇ Ax D b , where A is an m n matrix with rank A D m and b 2 Rm .
(iii) ˝ is the nonnegative orthant ˝ WD RnC .
Exercise 1.31 For ai bi with i D 1; : : : ; n, define

˚ ˇ
˝ WD x D .x1 ; : : : ; xn / 2 Rn ˇ ai x bi for all i D 1; : : : ; n :
Show that the projection ˘.xI ˝/ has the following representation:

˚ ˇ
˘.xI ˝/ D u D .u1 ; : : : ; un / 2 Rn ˇ ui D maxfai ; minfb; xi gg for i 2 f1; : : : ; ng :
37
CHAPTER 2
Subdifferential Calculus
In this chapter we present basic results of generalized differentiation theory for convex functions
and sets. A geometric approach of variational analysis based on convex separation for extremal
systems of sets allows us to derive general calculus rules for normals to sets and subgradients of
functions. We also include into this chapter some related results on Lipschitz continuity of convex
functions and convex duality. Note that some extended versions of the presented calculus results
can be found in exercises for this chapter, with hints to their proofs given at the end of the book.
2.1 CONVEX SEPARATION

We start this section with convex separation theorems, which play a fundamental role in convex
analysis and particularly in deriving calculus rules by the geometric approach. e first result
concerns the strict separation of a nonempty, closed, convex set and an out-of-set point.
Proposition 2.1 Let ˝ be a nonempty, closed, convex subset of Rn and let xN … ˝ . en there is a
nonzero element v 2 Rn such that
˚ ˇ
sup hv; xi ˇ x 2 ˝ < hv; xi:N (2.1)
Proof. Denote wN WD ˘.xI

N ˝/ and let v WD xN wN ¤ 0. Applying Proposition 1.78 gives us
hv; x wi
N D hxN w;
N x N 0 for any x 2 ˝:
wi
It follows furthermore that
hv; x .xN v/i D hv; x wi

N 0;
which implies in turn that hv; xi hv; xi

N kvk2 and thus verifies (2.1).
In the setting of Proposition 2.1 we choose a number b 2 R such that
˚ ˇ
sup hv; xi ˇ x 2 ˝ < b < hv; xiN
and define the function A.x/ WD hv; xi for which A.x/ N > b and A.x/ < b for all x 2 ˝ . en we
say that xN and ˝ are strictly separated by the hyperplane L WD fx 2 Rn j A.x/ D bg.
e next theorem extends the result of Proposition 2.1 to the case of two nonempty, closed,
convex sets at least one of which is bounded.
38 2. SUBDIFFERENTIAL CALCULUS
eorem 2.2 Let ˝1 and ˝2 be nonempty, closed, convex subsets of Rn with ˝1 \ ˝2 D ;. If
˝1 is bounded or ˝2 is bounded, then there is a nonzero element v 2 Rn such that
˚ ˇ ˚ ˇ
sup hv; xi ˇ x 2 ˝1 < inf hv; yi ˇ y 2 ˝2 :
Proof. Denote ˝ WD ˝1 ˝2 . en ˝ is a nonempty, closed, convex set and 0 … ˝ . Applying

Proposition 2.1 to ˝ and xN D 0, we have
˚ ˇ
WD sup hv; xi ˇ x 2 ˝ < 0 D hv; xi
N with some 0 ¤ v 2 Rn :
For any x 2 ˝1 and y 2 ˝2 , we have hv; x yi , and so hv; xi C hv; yi. erefore,
˚ ˇ ˚ ˇ ˚ ˇ
sup hv; xi ˇ x 2 ˝1 C inf hv; yi ˇ y 2 ˝2 < inf hv; yi ˇ y 2 ˝2 ;
which completes the proof of the theorem.
Figure 2.1: Illustration of convex separation.
Next we show that two nonempty, closed, convex sets with empty intersection in Rn can
be separated, while not strictly in general, if both sets are unbounded.
Corollary 2.3 Let ˝1 and ˝2 be nonempty, closed, convex subsets of Rn such that ˝1 \ ˝2 D ;.
en they are  in the sense that there is a nonzero element v 2 Rn with
˚ ˇ ˚ ˇ
sup hv; xi ˇ x 2 ˝1 inf hv; yi ˇ y 2 ˝2 : (2.2)
2.1. CONVEX SEPARATION 39
Proof. For every fixed k 2 N , we define the set k WD ˝2 \ IB.0I k/ and observe that it is a
closed, bounded, convex set, which is also nonempty if k is sufficiently large. Employing eo-
rem 2.2 allows us to find 0 ¤ uk 2 Rn such that
huk ; xi < huk ; yi for all x 2 ˝1 ; y 2 k :

uk
Denote vk WD and assume without loss of generality that vk ! v ¤ 0 as k ! 1. Fix fur-
kuk k
ther x 2 ˝1 and y 2 ˝2 and take k0 2 N such that y 2 k for all k k0 . en
hvk ; xi < hvk ; yi whenever k k0 :
Letting k ! 1, we have hv; xi hv; yi, which justifies the claim of the corollary.
Now our intention is to establish a general separation theorem for convex sets that may not
be closed. To proceed, we verify first the following useful result.
Figure 2.2: Geometric illustration of Lemma 2.4.
Lemma 2.4 Let ˝ be a nonempty, convex subset of Rn such that 0 2 ˝ n ˝ . en there is a sequence
xk ! 0 as k ! 1 with xk … ˝ for all k 2 N .
Proof. Recall that ri ˝ ¤ ; by eorem 1.72(i). Fix x0 2 ri ˝ and show that tx0 … ˝ for all
t > 0; see Figure 2.2. Indeed, suppose the contrary that tx0 2 ˝ and get by eorem 1.72(ii)
that
t 1
0D x0 C . tx0 / 2 ri ˝ ˝;
1Ct 1Ct
x0
a contradiction. Letting then xk WD gives us xk … ˝ for every k and xk ! 0.
k
Note that Lemma 2.4 may not hold for nonconvex sets. As an example, consider the set
˝ WD Rn nf0g for which 0 2 ˝ n ˝ while the conclusion of the Lemma 2.4 fails.
eorem 2.5 Let ˝1 and ˝2 be nonempty, convex subsets of Rn with ˝1 \ ˝2 D ;. en they
are separated in the sense of (2.2) with some v ¤ 0.
Proof. Consider the set ˝ WD ˝1 ˝2 for which 0 … ˝ . We split the proof of the theorem into
the following two cases:
Case 1: 0 … ˝ . In this case we apply Proposition 2.1 and find v ¤ 0 such that
hv; zi < 0 for all z 2 ˝:
is gives us hv; xi < hv; yi whenever x 2 ˝1 and y 2 ˝2 , which justifies the statement.
Case 2: 0 2 ˝ . In this case we employ Lemma 2.4 and find a sequence xk ! 0 with xk … ˝ .
en Proposition 2.1 produces a sequence fvk g of nonzero elements such that
hvk ; zi < hvk ; xk i for all z 2 ˝:
Without loss of generality, we get (dividing by kvk k and extracting a convergent subsequence)
that vk ! v and kvk k D kvk D 1. is gives us by letting k ! 1 that
hv; zi 0 for all z 2 ˝:
To complete the proof, we just repeat the arguments of Case 1.
Figure 2.3: Extremal system of sets.
eorem 2.5 ensures the separation of convex sets, which do not intersect. To proceed now
with separation of convex sets which may have common points as in Figure 2.3, we introduce
the notion of set extremality borrowed from variational analysis that deals with both convex and
nonconvex sets while emphasizing intrinsic optimality/extremality structures of problems under
consideration; see, e.g., [14] for more details.
2.1. CONVEX SEPARATION 41
n
Definition 2.6 Given two nonempty, convex sets ˝1 ; ˝2 R , we say that the set system f˝1 ; ˝2 g
is  if for any > 0 there is a 2 Rn such that kak < and
.˝1 a/ \ ˝2 D ;:
e next proposition gives us an important example of extremal systems of sets.
Proposition 2.7 Let ˝ be a nonempty, convex subset of Rn and let xN 2 ˝ be a boundary point of
˝ . en the sets ˝ and fxg
N form an extremal system.
N =2/ \ ˝ c ¤ ;. us choosing b 2

Proof. Since xN 2 bd ˝ , for any number > 0 we get IB.xI
N =2/ \ ˝ c and denoting a WD b xN yield kak < and .˝ a/ \ fxg
IB.xI N D ;.
e following result is a separation theorem for extremal systems of convex sets. It admits
a far-going extension to nonconvex settings, which is known as the extremal principle and is at
the heart of variational analysis and its applications [14]. It is worth mentioning that, in con-
trast to general settings of variational analysis, we do not impose here any closedness assumptions
on the sets in question. is will allow us to derive in the subsequent sections of this chapter
various calculus rules for normals to convex sets and subgradients of convex functions by using
a simple variational device without imposing closedness and lower semicontinuity assumptions
conventional in variational analysis.
eorem 2.8 Let ˝1 and ˝2 be two nonempty, convex subsets of Rn . en ˝1 and ˝2 form
an extremal system if and only if there is a nonzero element v 2 Rn which separates the sets ˝1
and ˝2 in the sense of (2.2).
Proof. Suppose that ˝1 and ˝2 form an extremal system of convex sets in Rn . en there is a
sequence fak g in Rn with ak ! 0 and
.˝1 ak / \ ˝2 D ;; k 2N:
For every k 2 N , apply eorem 2.5 to the convex sets ˝1 ak and ˝2 and find a sequence of
nonzero elements uk 2 Rn satisfying
huk ; x ak i huk ; yi for all x 2 ˝1 and y 2 ˝2 ; k 2N:

uk
e normalization vk WD with kvk k D 1 gives us
kuk k
hvk ; x ak i hvk ; yi for all x 2 ˝1 and y 2 ˝2 : (2.3)
us we find v 2 Rn with kvk D 1 such that vk ! v as k ! 1 along a subsequence. Passing to

the limit in (2.3) verifies the validity of (2.2).
For the converse implication, suppose that there is a nonzero element v 2 Rn which sepa-
rates the sets ˝1 and ˝2 in the sense of (2.2). en it is not hard to see that
1
˝1 v \ ˝2 D ; for every k 2 N :
k
us ˝1 and ˝2 form an extremal system.
2.2 NORMALS TO CONVEX SETS

In this section we apply the separation results obtained above (particularly, the extremal version
in eorem 2.8) to the study of generalized normals to convex sets and derive basic rules of the
normal cone calculus.
Definition 2.9 Let ˝ Rn be a convex set with xN 2 ˝ . e   to ˝ at xN is

˚ ˇ
N ˝/ WD v 2 Rn ˇ hv; x xi
N.xI N 0 for all x 2 ˝ : (2.4)
By convention, we put N.xI

N ˝/ WD ; for xN … ˝ .
Figure 2.4: Epigraph of the absolute value function and the normal cone to it at the origin.
We first list some elementary properties of the normal cone defined above.
Proposition 2.10 Let xN 2 ˝ for a convex subset ˝ of Rn . en we have:

(i) N.xI N ˝/ is a closed, convex cone containing the origin.
(ii) If xN 2 int ˝ , then N.xIN ˝/ D f0g.
Proof. To verify (i), fix vi 2 N.xI

N ˝/ and i 0 for i D 1; 2. We get by the definition that
N ˝/ and hvi ; x xi
0 2 N.xI N 0 for all x 2 ˝ implying
h1 v1 C 2 v2 ; x N 0 whenever x 2 ˝:
xi
2.2. NORMALS TO CONVEX SETS 43
us 1 v1 C 2 v2 2 N.xIN ˝/, and hence N.xI
N ˝/ is a convex cone. To check its closedness, fix
a sequence fvk g N.xI
N ˝/ that converges to v . en passing to the limit in
hvk ; x N 0 for all x 2 ˝ as k ! 1

xi
yields hv; x xi
N 0 whenever x 2 ˝ , and so v 2 N.xI N ˝/. is completes the proof of (i).
To check (ii), pick v 2 N.xI
N ˝/ with xN 2 int ˝ and find ı > 0 such that xN C ıIB ˝ .
us for every e 2 IB we have the inequality
hv; xN C ıe xi
N 0;
which shows that ıhv; ei 0, and hence hv; ei 0. Choose t > 0 so small that t v 2 IB . en
hv; tvi 0 and therefore tkvk2 D 0, i.e., v D 0.
We start calculus rules with deriving rather simple while very useful results.
Proposition 2.11 (i) Let ˝1 and ˝2 be nonempty, convex subsets of Rn and Rp , respectively. For
.xN 1 ; xN 2 / 2 ˝1 ˝2 , we have

N .xN 1 ; xN 2 /I ˝1 ˝2 D N.xN 1 I ˝1 / N.xN 2 I ˝2 /:
(ii) Let ˝1 and ˝2 be convex subsets of Rn with xN i 2 ˝i for i D 1; 2. en
N.xN 1 C xN 2 I ˝1 C ˝2 / D N.xN 1 I ˝1 / \ N.xN 2 I ˝2 /:
Proof. To verify (i), fix .v1 ; v2 / 2 N..xN 1 ; xN 2 /I ˝1 ˝2 / and get by the definition that
h.v1 ; v2 /; .x1 ; x2 / .xN 1 ; xN 2 /i D hv1 ; x1 xN 1 i C hv2 ; x2 xN 2 i 0 (2.5)
whenever .x1 ; x2 / 2 ˝1 ˝2 . Putting x2 WD xN 2 in (2.5) gives us
hv1 ; x1 xN 1 i 0 for all x1 2 ˝1 ;
which means that v1 2 N.xN 1 I ˝1 /. Similarly, we obtain v2 2 N.xN 2 I ˝2 / and thus justify the in-
clusion “” in (2.5). e opposite inclusion is obvious.
Let us now verify (ii). Fix v 2 N.xN 1 C xN 2 I ˝1 C ˝2 / and get by the definition that
hv; x1 C x2 .xN 1 C xN 2 /i 0 whenever x1 2 ˝1 ; x2 2 ˝2 :
Putting there x1 WD xN 1 and x2 WD xN 2 gives us v 2 N.xN 1 I ˝1 / \ N.xN 2 I ˝2 /. e opposite inclusion

in (ii) is also straightforward.
Consider further a linear mapping A W Rn ! Rp and associate it with a p n matrix as
usual. e adjoint mapping A W Rp ! Rn is defined by
hAx; yi D hx; A yi for all x 2 Rn ; y 2 Rp ;
which corresponds to the matrix transposition. e following calculus result describes normals to
solution sets for systems of linear equations.
Proposition 2.12 Let BW Rn ! Rp be defined by B.x/ WD Ax C b , where A W Rn ! Rp is a linear

mapping and b 2 Rp . Given c 2 R, consider the solution set
˚ ˇ
˝ WD x 2 Rn ˇ Bx D c :
For any xN 2 ˝ with B.x/

N D c , we have
˚ ˇ
N ˝/ D im A D v 2 Rn ˇ v D A y; y 2 Rp :
N.xI (2.6)
Proof. Take v 2 N.xIN ˝/ and get hv; x xi N 0 for all x such that Ax C b D c . Fixing any
u 2 ker A, we see that A.xN u/ D AxN D c b , i.e., A.xN u/ C b D c , which yields hv; ui 0
whenever u 2 ker A. Since u 2 ker A, it follows that hv; ui D 0 for all u 2 ker A. Arguing by
contradiction, suppose that there is no y 2 Rp with v D A y . us v … W WD A Rp , where the
set W Rn is obviously nonempty, closed, and convex. Employing Proposition 2.1 gives us a
nonzero element uN 2 Rn satisfying
˚ ˇ
sup hu;N wi ˇ w 2 W < hu;N vi;
and implying that 0 < hu;

N vi since 0 2 W . Furthermore
N A yi D hu;
thu; N vi for all t 2 R; y 2 Rp :
N A .ty/i < hu;
N A yi D 0, which shows that hAu;

Hence hu; N yi D 0 whenever y 2 Rp , i.e., AuN D 0. us uN 2
ker A while hv; ui
N > 0. is contradiction justifies the inclusion “” in (2.6).
To verify the opposite inclusion in (2.6), fix any v 2 Rn with v D A y for some y 2 Rp .
For any x 2 ˝ , we have
hv; x N D hA y; x
xi xi
N D hy; Ax Axi
N D 0;
which means that v 2 N.xI

N ˝/ and thus completes the proof of the proposition.
e following proposition allows us to characterize the separation of convex (generally non-
closed) sets in terms of their normal cones.
Proposition 2.13 Let ˝1 and ˝2 be convex subsets of Rn with xN 2 ˝1 \ ˝2 . en the following
assertions are equivalent:
n
(i) f˝1 ; ˝2 g forms an extremal system of two convex sets in R .
(ii) e sets ˝1 and ˝2 are separated
in the sense of (2.2) with some v ¤ 0.
(iii) N.xI
N ˝1 / \ N ˝2 / ¤ f0g.
N.xI
Proof. e equivalence of (i) and (ii) follows from eorem 2.8. Suppose that (ii) is satisfied.
Since xN 2 ˝2 , we have
˚ ˇ
sup hv; xi ˇ x 2 ˝1 hv; xi;
N
and thus hv; x xi N 0 for all x 2 ˝1 , i.e., v 2 N.xI
N ˝1 /. Similarly, we can verify the inclusion
N ˝2 /, and hence arrive at (iii).
v 2 N.xI
Now suppose that (iii) is satisfied and let 0 ¤ v 2 N.xI N ˝1 / \ N ˝2 / . en (2.2)
N.xI
follows directly from the definition of the normal cone. Indeed, for any x 2 ˝1 and y 2 ˝2 one
has
hv; x xiN 0 and h v; y xi N 0:
is implies hv; x xi
N 0 hv; y N , and hence hv; xi hv; yi.
xi
e following corollary provides a normal cone characterization of boundary points.
Corollary 2.14 Let ˝ be a nonempty, convex subset of Rn and let xN 2 ˝ . en xN 2 bd ˝ if and
only if the normal cone to ˝ is nontrivial, i.e., N.xI
N ˝/ ¤ f0g.
Proof. e “only if ” part follows from Proposition 2.10(ii). To justify the “if ” part, take xN 2 bd ˝
and consider the sets ˝1 WD ˝ and ˝2 WD fxg N , which form an extremal system by Proposition 2.7.
Note that N.xIN ˝2 / D Rn . us the claimed result is a direct consequence of Proposition 2.13
since N.xI
N ˝1 / \ Œ N.xI N ˝2 / D N.xIN ˝1 / ¤ f0g.
Now we are ready to establish a major result of the normal cone calculus for convex sets
providing a relationship between normals to set intersections and normals to each set in the in-
tersection. Its proof is also based on convex separation in extremal systems of sets. Note that this
approach allows us to capture general convex sets, which may not be closed.
eorem 2.15 Let ˝1 and ˝2 be convex subsets of Rn with xN 2 ˝1 \ ˝2 . en for any v 2
N ˝1 \ ˝2 / there exist 0 and vi 2 N.xI
N.xI N ˝i /, i D 1; 2, such that
v D v1 C v2 ; .; v1 / ¤ .0; 0/: (2.7)
Proof. Fix v 2 N.xI

N ˝1 \ ˝2 / and get by the definition that
hv; x N 0 for all x 2 ˝1 \ ˝2 :

xi
Denote 1 WD ˝1 Œ0; 1/ and 2 WD f.x; / j x 2 ˝2 ; hv; x xig N . ese convex sets
form an extremal system since we obviously have .x;
N 0/ 2 1 \ 2 and
1 \ .2 .0; a// D ; whenever a > 0:
By eorem 2.8, find 0 ¤ .w; / 2 Rn R such that
hw; xi C 1 hw; yi C 2 for all .x; 1 / 2 1 ; .y; 2 / 2 2 : (2.8)
It is easy to see that 0. Indeed, assuming > 0 and taking into account that .x;
N k/ 2 1
when k > 0 and .x; N 0/ 2 2 , we get
hw; xi
N C k hw; xi;
N
which leads to a contradiction. ere are two possible cases to consider.

Case 1: D 0. In this case we have w ¤ 0 and
hw; xi hw; yi for all x 2 ˝1 ; y 2 ˝2 :
en following the proof of Proposition 2.13 gives us w 2 N.xI N ˝2 /, and thus
N ˝1 / \ Œ N.xI
(2.7) holds for D 0, v1 D w , and v2 D w .
Case 2: < 0. In this case we let WD > 0 and deduce from (2.8), by using .x; 0/ 2 1 when
x 2 ˝1 , and .x;
N 0/ 2 2 , that
N for all x 2 ˝1 :
hw; xi hw; xi
w
is implies that w 2 N.xI
N ˝1 /, and hence N ˝1 /. To proceed further, we get from (2.8),
2 N.xI

due to .x;
N 0/ 2 1 and .y; ˛/ 2 2 for all y 2 ˝2 with ˛ WD hv; y xi N , that
D E D E
w; xN w; y C hv; y xi N whenever y 2 ˝2 :
Dividing both sides by gives us

Dw E Dw E
; xN ; y C hv; y N whenever y 2 ˝2 :
xi

is clearly implies the inequality

Dw E
C v; y xN 0 for all y 2 ˝2 ;

w w w w
and thus Cv D N ˝2 /. Letting finally v1 WD
C v 2 N.xI N ˝1 / and v2 WD
2 N.xI C

N ˝2 / gives us v D v1 C v2 , which ensures the validity of (2.7) with D 1.
v 2 N.xI
Figure 2.5: Normal cone to set intersection.
Observe that representation (2.7) does not provide a calculus rule for representing normals
to set intersections unless ¤ 0. However, it is easy to derive such an intersection rule directly from
eorem 2.15 while imposing the following basic qualification condition, which plays a crucial role
in the subsequent calculus results; see Figure 2.5.
Corollary 2.16 Let ˝1 and ˝2 be convex subsets of Rn with xN 2 ˝1 \ ˝2 . Assume that the basic
qualification condition (BQC) is satisfied:

N.xIN ˝1 / \ N.xI N ˝2 / D f0g: (2.9)
en we have the normal cone intersection rule
N.xI
N ˝1 \ ˝2 / D N.xI
N ˝1 / C N.xI
N ˝2 /: (2.10)
Proof. For any v 2 N.xI N ˝1 \ ˝2 /, we find by eorem 2.15 a number 0 and elements vi 2
N ˝i / as i D 1; 2 such that the conditions in (2.7) hold. Assuming that D 0 in (2.7) gives us
N.xI
that 0 ¤ v1 D v2 2 N.xI N ˝1 / \ Œ N.xI N ˝2 /, which contradicts BQC (2.9). us > 0 and
v D v1 = C v2 = 2 N.xI N ˝1 / C N.xIN ˝2 /. Hence the inclusion “” in (2.10) is satisfied. e
opposite inclusion in (2.10) can be proved easily using the definition of the normal cone. Indeed,
fix any v 2 N.xIN ˝1 / C N.xI N ˝2 / with v D v1 C v2 , where vi 2 N.xIN ˝i / for i D 1; 2. For any
x 2 ˝1 \ ˝2 , one has
hv; x xi
N D hv1 C v2 ; x xi
N D hv1 ; x xi
N C hv2 ; x xi
N 0;
which implies v 2 N.xI

N ˝1 \ ˝2 / and completes the proof of the corollary.
e following example shows that the intersection rule (2.10) does not hold generally with-
out the validity of the basic qualification condition (2.9).
Example 2.17 Let ˝1 WD f.x; / 2 R2 j x 2 g and let ˝2 WD f.x; / 2 R2 j x 2 g. en

N ˝1 \ ˝2 / D R2 while N.xI
for xN D .0; 0/ we have N.xI N ˝1 / C N.xIN ˝2 / D f0g R.
Next we present an easily verifiable condition ensuring the validity of BQC (2.9). More
general results related to relative interior are presented in Exercise 2.13 and its solution.
Corollary 2.18 Let ˝1 ; ˝2 Rn be such that int ˝1 \ ˝2 ¤ ; or int ˝2 \ ˝1 ¤ ;. en for any
xN 2 ˝1 \ ˝2 both BQC (2.9) and intersection rule (2.10) are satisfied.
Proof. Consider for definiteness the case where int ˝1 \ ˝2 ¤ ;. Let us fix xN 2 ˝1 \ ˝2 and
show that BQC (2.9) holds, which ensures the intersection rule (2.10) by Corollary 2.16. Arguing
by contradiction, suppose the opposite

N.xI
N ˝1 / \ N.xI
N ˝2 / ¤ f0g
and then find by Proposition 2.13 a nonzero element v 2 Rn that separates the sets ˝1 and ˝2 as
in (2.2). Fixing now uN 2 int ˝1 \ ˝2 gives us a number ı > 0 such that uN C ıIB ˝1 . It ensures
by (2.2) that
N for all x 2 uN C ıIB:
hv; xi hv; ui
is implies in turn the inequality
hv; ui N whenever e 2 IB
N C ıhv; ei hv; ui
v
and tells us that hv; ei 0 for all e 2 IB . Choosing finally e WD , we arrive at a contradiction
kvk
and thus complete the proof of this corollary.
Employing induction arguments allows us to derive a counterpart of Corollary 2.16 (and,
similarly, of Corollary 2.18) for finitely many sets.
T
Corollary 2.19 Let ˝i Rn for i D 1; : : : ; m be convex sets and let xN 2 miD1 ˝i . Assume the
validity of the following qualification condition:
hX
m i
N ˝i / H) vi D 0 for all i D 1; : : : ; m :
vi D 0; vi 2 N.xI
iD1
en we have the intersection formula

\ m Xm
N xI
N ˝i D N.xI
N ˝i /:
iD1 iD1
2.3. LIPSCHITZ CONTINUITY OF CONVEX FUNCTIONS 49
2.3 LIPSCHITZ CONTINUITY OF CONVEX FUNCTIONS
In this section we study the fundamental notion of Lipschitz continuity for convex functions. is
notion can be treated as “continuity at a linear rate” being highly important for many aspects of
convex and variational analysis and their applications.
Definition 2.20 A function f W Rn ! R is L  on a set ˝ dom f if there

is a constant ` 0 such that
jf .x/ f .y/j `kx yk for all x; y 2 ˝: (2.11)
We say that f is  L  around xN 2 dom f if there are constants ` 0
and ı > 0 such that (2.11) holds with ˝ D IB.xI
N ı/.
Here we give several verifiable conditions that ensure Lipschitz continuity and characterize
its localized version for general convex functions. One of the most effective characterizations of
local Lipschitz continuity, holding for nonconvex functions as well, is given via the following
notion of the singular (or horizon) subdifferential, which was largely underinvestigated in convex
analysis and was properly recognized only in the framework of variational analysis; see [14, 27].
Note that, in contrast to the (usual) subdifferential of convex analysis defined in the next section,
the singular subdifferential does not reduce to the classical derivative for differentiable functions.
Definition 2.21 Let f W Rn ! R be a convex function and let xN 2 dom f . e  -
 of the function f at xN is defined by
˚ ˇ
N WD v 2 Rn ˇ .v; 0/ 2 N .x;
@1 f .x/ N I epi f :
N f .x/ (2.12)
e following example illustrates the calculation of @1 f .x/
N based on definition (2.12).
Example 2.22 Consider the function

(
0 if jxj 1;
f .x/ WD
1 otherwise:
en epi f D Œ 1; 1 Œ0; 1/ and for xN D 1 we have

N .x; N epi f D Œ0; 1/ . 1; 0;
N f .x//I
which shows that @1 f .x/
N D Œ0; 1/; see Figure 2.6.
e next proposition significantly simplifies the calculation of @1 f .x/

N .
Proposition 2.23 For a convex function f W Rn ! R with xN 2 dom f , we have

@1 f .x/ N dom f /:
N D N.xI
Figure 2.6: Singular subgradients.
Proof. Fix any v 2 @1 f .x/

N with x 2 dom f and observe that .x; f .x// 2 epi f . en using
(2.12) and the normal cone construction (2.4) give us

hv; x xiN D hv; x xi N C 0 f .x/ f .x/
N 0;
which shows that v 2 N.xI N dom f /. Conversely, suppose that v 2 N.xI

N dom f / and fix any
.x; / 2 epi f . en we have f .x/ , and so x 2 dom f . us
hv; x xi
N C 0. f .x//
N D hv; x xi
N 0;
which implies that .v; 0/ 2 N..x; N epi f /, i.e., v 2 @1 f .x/

N f .x//I N .
To proceed with the study of Lipschitz continuity of convex functions, we need the follow-
ing two lemmas, which are of their own interest.
Lemma 2.24 Let fei j i D 1; : : : ; ng be the standard orthonormal basis of Rn . Denote

˚ ˇ
A WD xN ˙ ei ˇ i D 1; : : : ; n ; > 0:
en the following properties hold:

(i) xN C ei 2 co A for j j and i D 1; : : : ; n.
N =n/ co A.
(ii) IB.xI
Proof. (i) For j j , find t 2 Œ0; 1 with D t . / C .1 t/ . en xN ˙ ei 2 A gives us
xN C ei D t.xN ei / C .1 t /.xN C ei / 2 co A:

2.3. LIPSCHITZ CONTINUITY OF CONVEX FUNCTIONS 51
(ii) For x 2 IB.xI
N =n/, we have x D xN C .=n/u with kuk 1. Represent u via fei g by
n
X
uD i ei ;
iD1
qP
n
where ji j iD1 2i D kuk 1 for every i . is gives us

n
X i X 1 n
x D xN C u D xN C ei D xN C i ei :
n n n
iD1 iD1
Denoting i WD i yields j i j . It follows from (i) that xN C i ei D xN C i ei 2 co A, and

thus x 2 co A since it is a convex combination of elements in co A.
Lemma 2.25 If a convex function f W Rn ! R is bounded above on IB.xI

N ı/ for some element xN 2
dom f and number ı > 0, then f is bounded on IB.xI
N ı/.
Proof. Denote m WD f .x/N and take M > 0 with f .x/ M for all x 2 IB.xI
N ı/. Picking any u 2
N ı/, consider the element x WD 2xN u. en x 2 IB.xI
IB.xI N ı/ and
x C u 1 1
m D f .x/
N Df f .x/ C f .u/;
2 2 2
which shows that f .u/ 2m f .x/ 2m M and thus f is bounded on IB.xI
N ı/.
e next major result shows that the local boundedness of a convex function implies its
local Lipschitz continuity.
eorem 2.26 Let f W Rn ! R be convex with xN 2 dom f . If f is bounded above on IB.xI

N ı/
for some ı > 0, then f is Lipschitz continuous on IB.xI
N ı=2/.
Proof. Fix x; y 2 IB.xI

N ı=2/ with x ¤ y and consider the element
ı
u WD x C x y :
2kx yk
x y ı ı ı
Since e WD 2 IB , we have u D x C e 2 xN C IB C IB xN C ıIB . Denoting further
kx yk 2 2 2
ı
˛ WD gives us u D x C ˛.x y/, and thus
2kx yk
1 ˛
xD uC y:
˛C1 ˛C1
It follows from the convexity of f that
1 ˛
f .x/ f .u/ C f .y/;
˛C1 ˛C1
which implies in turn the inequalities
1 1 2kx yk 4M
f .x/ f .y/ .f .u/ f .y// 2M D 2M kx yk
˛C1 ˛C1 ı C 2kx yk ı
with M WD supfjf .x/j j x 2 IB.xIN ı/g < 1 by Lemma 2.25. Interchanging the role of x and y
above, we arrive at the estimate
4M
jf .x/ f .y/j kx yk
ı
and thus verify the Lipschitz continuity of f on IB.xI
N ı=2/.
e following two consequences of eorem 2.26 ensure the automatic local Lipschitz con-
tinuity of a convex function on appropriate sets.
Corollary 2.27 Any extended-real-valued convex function f W Rn ! R is locally Lipschitz contin-

uous on in the interior of its domain int.dom f /.
Proof. Pick xN 2 int.dom f / and choose > 0 such that xN ˙ ei 2 dom f for every i . Consider-
ing the set A from Lemma 2.24, we get IB.xI
N =n/ co A by assertion (ii) of this lemma. Denote
M WD maxff .a/ j a 2 Ag < 1 over the finite set A. Using the representation
m
X m
X
xD i ai with i 0 i D 1; ai 2 A
iD1 iD1
for any x 2 IB.xI

N =n/ shows that
m
X m
X
f .x/ i f .ai / i M D M;
iD1 iD1
and so f is bounded above on IB.xIN =n/. en eorem 2.26 tells us that f is Lipschitz con-
tinuous on IB.xI
N =2n/ and thus it is locally Lipschitz continuous on int .dom f /.
It immediately follows from Corollary 2.27 that any finite convex function on Rn is always
locally Lipschitz continuous on the whole space.
Corollary 2.28 If f W Rn ! R is convex, then it is locally Lipschitz continuous on Rn .

2.4. SUBGRADIENTS OF CONVEX FUNCTIONS 53
e final result of this section provides several characterizations of the local Lipschitz con-
tinuity of extended-real-valued convex functions.
eorem 2.29 Let f W Rn ! R be convex with xN 2 dom f . e following are equivalent:

(i) f is continuous at xN .
(ii) xN 2 int.dom f /.
(iii) f is locally Lipschitz continuous around xN .
(iv) @1 f .x/N D f0g.
Proof. To verify (i)H)(ii), by the continuity of f at xN , for any > 0 find ı > 0 such that
jf .x/ N < whenever x 2 IB.xI

f .x/j N ı/:
Since f .x/
N < 1, this implies that IB.xI
N ı/ dom f , and so xN 2 int.dom f /.
Implication (ii)H)(iii) follows directly from Corollary 2.27 while (iii)H)(i) is obvious. us
conditions (i), (ii), and (iii) are equivalent. To show next that (ii)H)(iv), take xN 2 int.dom f /
for which N.xIN dom f / D f0g. en Proposition 2.23 yields @1 f .x/ N D f0g.
To verify finally (iv)H)(ii), we argue by contradiction and suppose that xN … int.dom f /.
en xN 2 bd.dom f / and it follows from Corollary 2.14 that N.xI N dom f / ¤ f0g, which contra-
dicts the condition @1 f .x/N D f0g by Proposition 2.23.
2.4 SUBGRADIENTS OF CONVEX FUNCTIONS

In this section we define and start studying the concept of subdifferential (or the collection of sub-
gradients) for convex extended-real-valued functions. is notion of generalized derivative is one
of the most fundamental in analysis, with a variety of important applications. e revolutionary
idea behind the subdifferential concept, which distinguishes it from other notions of generalized
derivatives in mathematics, is its set-valuedness as the indication of nonsmoothness of a function
around the reference individual point. It not only provides various advantages and flexibility but
also creates certain difficulties in developing its calculus and applications. e major techniques
of subdifferential analysis revolve around convex separation of sets.
Definition 2.30 Let f W Rn ! R be a convex function and let xN 2 dom f . An element v 2 Rn is

called a  of f at xN if
hv; x xi
N f .x/ N for all x 2 Rn :
f .x/ (2.13)
e collection of all the subgradients of f at xN is called the  of the function at this
point and is denoted by @f .x/
N .
Figure 2.7: Subdifferential of the absolute value function at the origin.
Note that it suffices to take only x 2 dom f in the subgradient definition (2.13). Observe
also that @f .x/
N ¤ ; under more general conditions; see, e.g., Proposition 2.47 as well as other
results in this direction given in [26].
e next proposition shows that the subdifferential @f .x/
N can be equivalently defined ge-
ometrically via the normal cone to the epigraph of f at .x; N .
N f .x//
Proposition 2.31 For any convex function f W Rn ! R and any xN 2 dom f , we have
˚ ˇ
N D v 2 Rn ˇ .v; 1/ 2 N .x;
@f .x/ N epi f :
N f .x//I (2.14)
Proof. Fix v 2 @f .x/

N and .x; / 2 epi f . Since f .x/, we deduce from (2.13) that
˝ ˛
.v; 1/; .x; / .x;
N f .x//
N D hv; x xi
N . f .x//
N
hv; x xi
N .f .x/ f .x//
N 0:
is readily implies that .v; 1/ 2 N..x; N epi f /.

N f .x//I
To verify the opposite inclusion in (2.14), take an arbitrary element v 2 Rn such that
.v; 1/ 2 N..x; N epi f /. For any x 2 dom f , we have .x; f .x// 2 epi f . us
N f .x//I
˝ ˛
hv; x xi
N .f .x/ f .x//
N D .v; 1/; .x; f .x// .x;
N f .x//
N 0;
which shows that v 2 @f .x/

N and so justifies representation (2.14).
Along with representing the subdifferential via the normal cone in (2.14), there is the fol-
lowing simple way to express the normal cone to a set via the subdifferential (as well as the singular
subdifferential (2.12)) of the extended-real-indicator function.
n
Example 2.32 Given a nonempty, convex set ˝ R , define its indicator function by
(
0 if x 2 ˝;
f .x/ D ı.xI ˝/ WD (2.15)
1 otherwise:
en epi f D ˝ Œ0; 1/, and we have by Proposition 2.11(i) that

N .x; N epi f D N.xI
N f .x//I N ˝/ . 1; 0;
which readily implies the relationships
N D @1 f .x/
@f .x/ N D N.xI
N ˝/: (2.16)
A remarkable feature of convex functions is that every local minimizer for a convex function
gives also the global/absolute minimum to the function.
Proposition 2.33 Let f W Rn ! R convex and let xN 2 dom f be a local minimizer of f . en f
attains its global minimum at this point.
Proof. Since xN is a local minimizer of f , there is ı > 0 such that
N for all u 2 IB.xI

f .u/ f .x/ N ı/:
Fix x 2 Rn and construct a sequence of xk WD .1 k 1 /xN C k 1 x as k 2 N . us we have xk 2

N ı/ when k is sufficiently large. It follows from the convexity of f that
IB.xI
1 1
f .x/
N f .xk / .1 k /f .x/
N Ck f .x/;
which readily implies that k 1

f .x/
N k 1
N f .x/ for all x 2 Rn .
f .x/, and hence f .x/
Next we recall the classical notion of derivative for functions on Rn , which we understand
throughout the book in the sense of Fréchet unless otherwise stated; see Section 3.1.
Definition 2.34 We say that f W Rn ! R is .Fréchet/  at xN 2 int.dom f / if there

exists an element v 2 Rn such that
f .x/ f .x/
N hv; x xi
N
lim D 0:
x!xN kx xkN
In this case the element v is uniquely defined and is denoted by rf .x/

N WD v .
e Fermat stationary rule of classical analysis tells us that if f W Rn ! R is differentiable

at xN and attains its local minimum at this point, then rf .x/ N D 0. e following proposition
gives a subdifferential counterpart of this result for general convex functions. Its proof is a direct
consequence of the definitions and Proposition 2.33.
Proposition 2.35 Let f W Rn ! R be convex and let xN 2 dom f . en f attains its local/global
minimum at xN if and only if 0 2 @f .x/
N .
Proof. Suppose that f attains its global minimum at xN . en

N f .x/ for all x 2 Rn ;
f .x/
which can be rewritten as
0 D h0; x xi
f .x/
e definition of the subdifferential shows that this is equivalent to 0 2 @f .x/
N .
Now we show that the subdifferential (2.13) is indeed a singleton for differentiable func-
tions reducing to the classical derivative/gradient at the reference point and clarifying the notion
of differentiability in the case of convex functions.
Proposition 2.36 Let f W Rn ! R be convex and differentiable at xN 2 int.dom f /. en we have
@f .x/ N and
N D frf .x/g
hrf .x/;
N x xi
f .x/ (2.17)
Proof. It follows from the differentiability of f at xN that for any > 0 there is ı > 0 with
kx xk
N f .x/ f .x/
N hrf .x/;
N x xi
N kx N whenever kx
xk xk
N < ı: (2.18)
Consider further the convex function
'.x/ WD f .x/ f .x/
N hrf .x/;
N x xi
N C kx xk;
N x 2 Rn ;
and observe that '.x/ '.x/
N D 0 for all x 2 IB.xI
N ı/. e convexity of ' ensures that '.x/
N for all x 2 Rn . us
'.x/
hrf .x/;
N x xi
N f .x/ f .x/
N C kx N whenever x 2 Rn ;
xk
which yields (2.17) by letting # 0.
It follows from (2.17) that rf .x/ N . Picking now v 2 @f .x/
N 2 @f .x/ N , we get
hv; x xi
N f .x/ f .x/:
N
en the second part of (2.18) gives us that
hv rf .x/;
N x xi
N kx N whenever kx
xk xk
N < ı:
Finally, we observe that kv rf .x/k
N , which yields v D rf .x/
N since > 0 was chosen ar-
bitrarily. us @f .x/ N .
N D frf .x/g
n
Corollary 2.37 Let f W R ! R be a strictly convex function. en the differentiability of f at
xN 2 int.dom f / implies the strict inequality
hrf .x/;
N x xi
N < f .x/ N whenever x ¤ x:
f .x/ N (2.19)
Proof. Since f is convex, we get from Proposition 2.36 that
hrf .x/;
N u xi
N f .u/ N for all u 2 Rn :
f .x/
Fix x ¤ xN and let u WD .x C x/=2

N . It follows from the above that
D x C xN E x C xN 1 1
rf .x/;
N xN f f .x/
N < f .x/ C f .x/
N f .x/:
N
2 2 2 2
us we arrive at the strict inequality (2.19) and complete the proof.
A simple while important example of a nonsmooth convex function is the norm function
on Rn . Here is the calculation of its subdifferential.
Example 2.38 Let p.x/ WD kxk be the Euclidean norm function on Rn . en we have
8
<IB if x D 0;
@p.x/ D n x o
: otherwise:
kxk
To verify this, observe first that the Euclidean norm function p is differentiable at any nonzero
x
point with rp.x/ D as x ¤ 0. It remains to calculate its subdifferential at x D 0. To proceed
kxk
by definition (2.13), we have that v 2 @p.0/ if and only if
hv; xi D hv; x 0i p.x/ p.0/ D kxk for all x 2 Rn :
Letting x D v gives us hv; vi kvk, which implies that kvk 1, i.e., v 2 IB . Now take v 2 IB
and deduce from the Cauchy-Schwarz inequality that
hv; x 0i D hv; xi kvk kxk kxk D p.x/ p.0/ for all x 2 Rn
and thus v 2 @p.0/, which shows that @p.0/ D IB .
e next theorem calculates the subdifferential of the distance function.
eorem 2.39 Let ˝ be a nonempty, closed, convex subset of Rn . en we have

8
<N.xI N ˝/ \ IB if xN 2 ˝;
N ˝/ D n xN ˘.xI
@d.xI N ˝/ o
: otherwise:
d.xI
N ˝/
Proof. Consider the case where xN 2 ˝ and fix any element w 2 @d.xI
N ˝/. It follows from the
definition of the subdifferential that
hw; x xi
N d.xI ˝/ N ˝/ D d.xI ˝/ for all x 2 Rn :
d.xI (2.20)
Since the distance function d.I ˝/ is Lipschitz continuous with constant ` D 1, we have
hw; x xi
N kx N for all x 2 Rn :
xk
is implies that kwk 1, i.e., w 2 IB . It also follows from (2.20) that
hw; x N 0 for all x 2 ˝:
xi
us w 2 N.xI
N ˝/, which verifies that w 2 N.xI N ˝/ \ IB .
To prove the opposite inclusion, fix any w 2 N.xI
N ˝/ \ IB . en kwk 1 and
hw; u N 0 for all u 2 ˝:
xi
us for every x 2 Rn and every u 2 ˝ we have
hw; x xi
N D hw; x u C u xiN D hw; x ui C hw; u xi
N
hw; x ui kwk kx uk kx uk:
Since u is taken arbitrarily in ˝ , this implies hw; x xi
N d.xI ˝/ D d.xI ˝/ d.xI N ˝/, and
hence w 2 @d.xIN ˝/.
Suppose now that xN … ˝ , take zN WD ˘.xI N ˝/ 2 ˝ , and fix any w 2 @d.xI
N ˝/. en
hw; x xi
N d.xI ˝/ d.xI
N ˝/ D d.xI ˝/ kxN zN k
kx zN k kxN zN k for all x 2 Rn :
Denoting p.x/ WD kx zN k, we have
hw; x xi
N p.x/ p.x/ N for all x 2 Rn ;
n xN zN o xN zN
which ensures that w 2 @p.x/ N D . Let us show that w D is a subgradient of
kxN zN k kxN zN k
d.I ˝/ at xN . Indeed, for any x 2 Rn denote px WD ˘.xI ˝/ and get
hw; x xi
N D hw; x zN i C hw; zN xi
N D hw; x zN i kxN zN k
D hw; x px i C hw; px zN i kxN zN k:
Since zN D ˘.xI
N ˝/, it follows from Proposition 1.78 that hxN zN ; px zN i 0, and so we have
hw; px zN i 0. Using the fact that kwk D 1 and the Cauchy-Schwarz inequality gives us
hw; x xi
N D hw; x px i C hw; px zN i kxN zN k
kwk kx px k kxN zN k D kx px k kxN zN k
N ˝/ for all x 2 Rn :
D d.xI ˝/ d.xI
us we arrive at w 2 @d.xI
N ˝/ and complete the proof of the theorem.
We conclude this section by providing first-order and second-order differential charac-
terizations of convexity for differentiable and twice continuously differentiable .C 2 / functions,
respectively, on given convex regions of Rn .
eorem 2.40 Let f W Rn ! R be a differentiable function on its domain D , which is an open

convex set. en f is convex if and only if
hrf .u/; x ui f .x/ f .u/ for all x; u 2 D: (2.21)
Proof. e “only if ” part follows from Proposition 2.36. To justify the converse, suppose that
(2.21) holds and then fix any x1 ; x2 2 D and t 2 .0; 1/. Denoting x t WD tx1 C .1 t/x2 , we have
x t 2 D by the convexity of D . en
hrf .x t /; x1 x t i f .x1 / f .x t /; hrf .x t /; x2 x t i f .x2 / f .x t /:
It follows furthermore that
thrf .x t /; x1 x t i tf .x1 / tf .x t / and

.1 t/hrf .x t /; x2 x t i .1 t/f .x2 / .1 t /f .x t /:
Summing up these inequalities, we arrive at
0 tf .x1 / C .1 t/f .x2 / f .x t /;
which ensures that f .x t / tf .x1 / C .1 t /f .x2 /, and so verifies the convexity of f .
Lemma 2.41 Let f W Rn ! R be convex and twice continuously differentiable on an open subset V
of its domain containing xN . en we have
hr 2 f .x/u;
N ui 0 for all u 2 Rn ;
where r 2 f .x/
N is the Hessian matrix of f at xN .
Proof. Let A WD r 2 f .x/

N , which is symmetric matrix. en
1
f .xN C h/ f .x/
N hrf .x/;
N hi hAh; hi
lim 2 D 0: (2.22)
h!0 khk2
It follows from (2.22) that for any > 0 there is ı > 0 such that
1
khk2 f .xN C h/ f .x/
N hrf .x/;
N hi hAh; hi khk2 for all khk ı:
2
By Proposition 2.36, we readily have
f .xN C h/ f .x/
N hrf .x/;
N hi 0:
Combining the above inequalities ensures that

1
khk2 hAh; hi whenever khk ı: (2.23)
2
u
Observe further that for any 0 ¤ u 2 Rn the element h WD ı satisfies khk ı and, being
kuk
substituted into (2.23), gives us the estimate
1 D u u E
ı 2 ı 2 A ; :
2 kuk kuk
It shows therefore (since the case where u D 0 is trivial) that
2kuk2 hAu; ui whenever u 2 Rn ;
which implies by letting # 0 that 0 hAu; ui for all u 2 Rn .

Now we are ready to justify the aforementioned second-order characterization of convexity
for twice continuously differentiable functions.
eorem 2.42 Let f W Rn ! R be a function twice continuously differentiable on its domain

D , which is an open convex subset of Rn . en f is convex if and only if r 2 f .x/
N is positive
semidefinite for every xN 2 D .
Proof. Taking Lemma 2.41 into account, we only need to verify that if r 2 f .x/ N is positive
semidefinite for every xN 2 D , then f is convex. To proceed, for any x1 ; x2 2 D define x t WD
tx1 C .1 t /x2 and consider the function

'.t/ WD f tx1 C .1 t/x2 tf .x1 / .1 t /f .x2 /; t 2 R :
It is clear that ' is well defined on an open interval I containing .0; 1/. en ' 0 .t/ D
hrf .x t /; x1 x2 i f .x1 / C f .x2 / and ' 00 .t / D hr 2 f .x t /.x1 x2 /; x1 x2 i 0 for every t 2
I since r 2 f .x t / is positive semidefinite. It follows from Corollary 1.46 that ' is convex on I .
Since '.0/ D '.1/ D 0, for any t 2 .0; 1/ we have

'.t/ D ' t.1/ C .1 t /0 t '.1/ C .1 t /'.0/ D 0;
which implies in turn the inequality

f tx1 C .1 t/x2 tf .x1 / C .1 t/f .x2 /;
is justifies that the function f is convex on its domain and thus on Rn .
2.5. BASIC CALCULUS RULES 61
2.5 BASIC CALCULUS RULES
Here we derive basic calculus rules for the convex subdifferential (2.13) and its singular counter-
part (2.12). First observe the obvious subdifferential rule for scalar multiplication.
Proposition 2.43 For a convex function f W Rn ! R with xN 2 dom f , we have
@.˛f /.x/ N and @1 .˛f /.x/

N D ˛@f .x/ N D ˛@1 f .x/
N whenever ˛ > 0:
Let us now establish the fundamental subdifferential sum rules for both constructions (2.13)
and (2.12) that play a crucial role in convex analysis. We give a geometric proof for this result re-
ducing it to the intersection rule for normals to sets, which in turn is based on convex separation of
extremal set systems. Observe that our variational approach does not require any closedness/lower
semicontinuity assumptions on the functions in question.
eorem 2.44 Let fi W Rn ! R for i D 1; 2 be convex functions and let xN 2 dom f1 \ dom f2 .
Impose the singular subdifferential qualification condition

@1 f1 .x/
N \ @1 f2 .x/
N D f0g: (2.24)
en we have the subdifferential sum rules
@.f1 C f2 /.x/
N D @f1 .x/
N C @f2 .x/;
N @1 .f1 C f2 /.x/
N D @1 f1 .x/
N C @1 f2 .x/:
N (2.25)
Proof. For definiteness we verify the first rule in (2.25); the proof of the second result is similar.
Fix an arbitrary subgradient v 2 @.f1 C f2 /.x/
N and define the sets
˝1 WD f.x; 1 ; 2 / 2 Rn R R j 1 f1 .x/g;
˝2 WD f.x; 1 ; 2 / 2 Rn R R j 2 f2 .x/g:
Let us first verify the inclusion

.v; 1; 1/ 2 N .x;
N f1 .x/;
N f2 .x//I
N ˝1 \ ˝2 : (2.26)
For any .x; 1 ; 2 / 2 ˝1 \ ˝2 , we have 1 f1 .x/ and 2 f2 .x/. en
hv; x xiN C . 1/.1 f1 .x//

N C . 1/.2 f2 .x//N
hv; x xi
N f1 .x/ C f2 .x/ f1 .x/
N f2 .x/
N 0;
where the last estimate holds due to v 2 @.f1 C f2 /.x/

N . us (2.26) is satisfied.
To proceed with employing the intersection rule in (2.26), let us first check that the assumed
qualification condition (2.24) yields

N .x;
N f1 .x/;
N f2 .x//I
N ˝1 \ N .x;N f1 .x/;
N f2 .x//I
N ˝2 D f0g; (2.27)
which is the basic qualification condition (2.9) needed for the application of Corollary 2.16 in the
setting of (2.26). Indeed, we have from ˝1 D epi f1 R that

N .x;
N f1 .x/;
N f2 .x//I
N ˝1 D N .x; N epi f1 f0g
N f1 .x//I
and observe similarly the representation

˚ ˇ
N .x;
N f1 .x/; N ˝2 D .v; 0; / 2 Rn R R ˇ .v; / 2 N .x;
N f2 .x//I N epi f2 :
N f2 .x//I
Further, fix any element

.v; 1 ; 2 / 2 N .x;
N f1 .x/;
N f2 .x//I
N ˝1 \ N .x;
N f1 .x/;
N f2 .x//I
N ˝2 :
en 1 D 2 D 0, and hence

.v; 0/ 2 N .x; N epi f1 and . v; 0/ 2 N .x;
N f1 .x//I N epi f2 :
N f2 .x//I
It follows by the definition of the singular subdifferential that

v 2 @1 f1 .x/
N \ @1 f2 .x/
N D f0g:
It yields .v; 1 ; 2 / D .0; 0; 0/ 2 Rn R R. Applying Corollary 2.16 in (2.26) shows that

N .x;N f1 .x/;
N f2 .x//I
N ˝1 \ ˝2 D N .x; N f1 .x/;
N f2 .x//I
N ˝1 C N .x; N f1 .x/;
N f2 .x//I
N ˝2 ;
.v; 1; 1/ D .v1 ; 1 ; 0/ C .v2 ; 0; 2 /
with .v1 ; 1 / 2 N..x; N epi f1 / and .v2 ; 2 / 2 N..x;

N f1 .x//I N epi f2 /. en we get 1 D
N f1 .x//I
2 D 1 and v D v1 C v2 , where vi 2 @fi .x/
N for i D 1; 2. is verifies the inclusion “” in the first
subdifferential sum rule of (2.25).
e opposite inclusion therein follows easily from the definition of the subdifferential. In-
deed, any v 2 @f1 .x/ N can be represented as v D v1 C v2 , where vi 2 @fi .x/
N C @f2 .x/ N for i D 1; 2.
en we have
hv; x xi
N D hv1 ; x xi
N C hv2 ; x xiN
f1 .x/ f1 .x/
N C f2 .x/ f2 .x/
N D .f1 C f1 /.x/ .f1 C f2 /.x/
N
for all x 2 Rn , which readily implies that v 2 @.f1 C f2 /.x/

N and therefore completes the proof
of the theorem.
Note that by Proposition 2.23 the qualification condition (2.24) can be written as

N.xIN dom f1 / \ N.xIN dom f2 / D f0g:
Let us present some useful consequences of eorem 2.44. e sum rule for the (basic) subdif-
ferential (2.13) in the following corollary is known as the Moreau-Rockafellar theorem.
Corollary 2.45 Let fi W Rn ! R for i D 1; 2 be convex functions such that there exists u 2
dom f1 \ dom f2 for which f1 is continuous at u or f2 is continuous at u. en
@.f1 C f2 /.x/ D @f1 .x/ C @f2 .x/ (2.28)
whenever x 2 dom f1 \ dom f2 . Consequently, if both functions fi are finite-valued on Rn , then the
sum rule (2.28) holds for all x 2 Rn .
Proof. Suppose for definiteness that f1 is continuous at u 2 dom f1 \ dom f2 , which implies
that u 2 int.dom f1 / \ dom f2 . en the proof of Corollary 2.18 shows that

N.xI dom f1 / \ N.xI dom f2 / D f0g whenever x 2 dom f1 \ dom f2 :
us the conclusion follows directly from eorem 2.44.

e next corollary for sums of finitely many functions can be derived from eorem 2.44 by
induction, where the validity of the qualification condition in the next step of induction follows
from the singular subdifferential sum rule in (2.25).
Corollary 2.46 Consider finitely many convex functions fi W Rn ! R for i D 1; : : : ; m. e fol-

lowing assertions hold:
T
(i) Let xN 2 miD1 dom fi and let the implication
hX
m i
vi D 0; vi 2 @1 fi .x/
N H) vi D 0 for all i D 1; : : : ; m (2.29)
iD1
hold. en we have the subdifferential sum rules

X
m m
X X
m m
X
1
@ fi .x/
N D @fi .x/;
N @ fi .x/
N D @fi .x/:
N
iD1 iD1 iD1 iD1
Tm
(ii) Suppose that there is u 2 iD1 dom fi such that all (except possibly one) functions fi are continuous
at u. en we have the equality
X
m m
X
@ fi .x/ D @fi .x/
iD1 iD1
Tm
for any common point of the domains x 2 iD1 dom fi .
e next result not only ensures the subdifferentiability of convex functions, but also justifies
the compactness of its subdifferential at interior points of the domain.
Proposition 2.47 For any convex function f W Rn ! R and any xN 2 int.dom f /, we have that
N is nonempty and compact in Rn .
the subgradient set @f .x/
Proof. Let us first verify the boundedness of the subdifferential. It follows from Corollary 2.27
that f is locally Lipschitz continuous around xN with some constant `, i.e., there exists a positive
number such that
jf .x/ f .y/j `kx yk for all x; y 2 IB.xI

N /:
N and any element h 2 Rn with khk ı we have

us for any subgradient v 2 @f .x/
hv; hi f .xN C h/ f .x/

N `khk:
is implies the bound kvk `. e closedness of @f .x/

N follows directly from the definition, and
so we get the compactness of the subgradient set.
It remains to verify that @f .x/
N ¤ ;. Since .x; N 2 bd .epi f /, we deduce from
N f .x//
Corollary 2.14 that N..x; N epi f / ¤ f0g. Take such an element .0; 0/ ¤ .v; / 2
N f .x//I
N..x; N epi f / and get by the definition that
N f .x//I
hv; x xi
N . N 0 for all .x; / 2 epi f:
f .x//
N C 1 yields 0. If D 0, then v 2 @1 f .x/

Using this inequality with x WD xN and D f .x/ N D
f0g by eorem 2.29, which is a contradiction. us > 0 and .v= ; 1/ 2 N..x; N epi f /.
N f .x//I
Hence v= 2 @f .x/N , which completes the proof of the proposition.
Given f W Rn ! R and any ˛ 2 R, define the level set

˚ ˇ
˝˛ WD x 2 Rn ˇ f .x/ ˛ :
Proposition 2.48 Let f W Rn ! R be convex and let xN 2 ˝˛ with f .x/ N D ˛ for some ˛ 2 R. If
N , then the normal cone to the level set ˝˛ at xN is calculated by
0 … @f .x/
N ˝˛ / D @1 f .x/
N.xI N [ R> @f .x/;
N (2.30)
where R> denotes the set of all positive real numbers.

n nC1
Proof. To proceed, consider the set WD R f˛g R and observe that
˝˛ f˛g D epi f \ :
Let us check the relationship

N ˛/I epi f \
N .x; N .x;
N ˛/I D f0g; (2.31)
i.e., the qualification condition (2.9) holds for the sets epi f and . Indeed, take .v; / from
the set on left-hand side of (2.31) and deduce from the structure of that v D 0 and that 0.
If we suppose that > 0, then
1
N ˛/I epi f ;
.v; / D .0; 1/ 2 N .x;

and so 0 2 @f .x/
N , which contradicts the assumption of this proposition. us we have .v; / D
.0; 0/, and then applying the intersection rule of Corollary 2.16 gives us

N.x;N ˝˛ / R D N .x; N ˛/I ˝˛ f˛g D N .x; N ˛/I epi f C .f0g R/:
Fix now any v 2 N.xI N ˝˛ / and get .v; 0/ 2 N.x;N ˝˛ / R. is implies the existence of 0
such that .v; / 2 N..x;N ˛/I epi f / and .v; 0/ D .v; / C .0; /. If D 0, then v 2 @1 f .x/
N .
In the case when > 0 we obtain v 2 @f .x/N and thus verify the inclusion “” in (2.30).
Let us finally prove the opposite inclusion in (2.30). Since ˝˛ dom f , we get
@1 f .x/ N dom f / N.xI

N D N.xI N ˝˛ /:
Take any v 2 R> @f .x/

N and find > 0 such that v 2 @f .x/
N . Taking any x 2 ˝˛ gives us f .x/
N . erefore, we have
˛ D f .x/
1
h v; x xi
N f .x/ f .x/
N 0;
which implies that v 2 N.xI

N ˝˛ / and thus completes the proof.
Corollary 2.49 In the setting of Proposition 2.48 suppose that f is continuous at xN and that 0 …
N . en we have the representation
@f .x/
[
N.xI
N ˝˛ / D RC @f .x/
N D ˛@f .x/;
N ˛ 2 R:
˛0
Proof. Since f is continuous at xN , it follows from eorem 2.29 that @1 f .x/

N D f0g. en the
result of this corollary follows directly from Proposition 2.48.
We now obtain some other calculus rules for subdifferentiation of convex function. For
brevity, let us only give results for the basic subdifferential (2.13) leaving their singular subdiffer-
ential counterparts as exercises; see Section 2.11.
To proceed with deriving chain rules, we present first the following geometric lemma.
Lemma 2.50 Let B W Rn ! Rp be an affine mapping given by B.x/ D A.x/ C b , where A W

Rn ! Rp is a linear mapping and b 2 Rp . For any .x; N 2 gph B , we have
N y/
˚ ˇ
N .x; N gph B D .u; v/ 2 Rn Rp ˇ u D A v :
N y/I
Proof. It is clear that the set gph B is convex and .u; v/ 2 N..x; N gph B/ if and only if
N y/
hu; x xi
N C hv; B.x/ N 0 whenever x 2 Rn :
B.x/i (2.32)
It follows directly from the definitions that
hu; x xi
N C hv; B.x/ B.x/i
N D hu; x xi N C hv; A.x/ A.x/i
N

D hu; x xi
N C hA v; x xi
N
D hu C A v; x xi;
N
which implies the equivalence of (2.32) to hu C A v; x N 0, and so to u D

xi A v .
eorem 2.51 Let f W Rp ! R be a convex function and let B W Rn ! Rp be as in Lemma 2.50

N 2 dom f for some xN 2 Rp . Denote yN WD B.x/
with B.x/ N and assume that
ker A \ @1 f .y/
N D f0g: (2.33)
en we have the subdifferential chain rule

˚ ˇ
@.f ı B/.x/ N D A v ˇ v 2 @f .y/
N D A @f .y/ N : (2.34)
N and form the subsets of Rn Rp R by

Proof. Fix v 2 @.f ı B/.x/
˝1 WD gph B R and ˝2 WD Rn epi f:
It follows from the definition of the subdifferential and the normal cone that .v; 0; 1/ 2
N..x;
N y; N ˝1 \ ˝2 /, where zN WD f .y/
N z/I N . Indeed, fix any .x; y; / 2 ˝1 \ ˝2 . en y D B.x/
and f .y/, and so f .B.x//. us
hv; x xi
N C 0.y y/
N C . 1/. zN / hv; x xi
N Œf .B.x// f .B.x//
N 0:
Employing the intersection rule from Corollary 2.19 under (2.33) gives us

.v; 0; 1/ 2 N .x;
N y;
N z/I
N ˝1 C N .x; N y;
N z/I
N ˝2 ;
which reads that .v; 0; 1/ D .v; w; 0/ C .0; w; 1/ with .v; w/ 2 N..x; N gph B/ and
N y/I
.w; 1/ 2 N..y; N epi f /. en we get
N z/I
v D A w and w 2 @f .y/;
N
which implies in turn that v 2 A .@f .y//

N and thus verifies the inclusion “” in (2.34). e op-
posite inclusion follows directly from the definition of the subdifferential.
In the corollary below we present some conditions that guarantee that the qualification
(2.33) in eorem 2.51 is satisfied.
Corollary 2.52 In the setting of eorem 2.51 suppose that either A is surjective or f is continuous
at the point yN 2 Rn ; the latter assumption is automatic when f is finite-valued. en we have the
subdifferential chain rule (2.34).
Proof. Let us check that the qualification condition (2.33) holds under the assumptions of
this corollary. e surjectivity of A gives us ker A D f0g, and in the second case we have
yN 2 int.dom f /, so @1 f .y/
N D f0g. e validity of (2.33) is obvious.
e last consequence of eorem 2.51 presents a chain rule for normals to sets.
Corollary 2.53 Let ˝ Rp be a convex set and let B W Rn ! Rp be an affine mapping as in

Lemma 2.50 satisfying yN WD B.x/
N 2 ˝ . en we have
1
N.xI
N B .˝// D A .N.yI
N ˝//
under the validity of the qualification condition
ker A \ N.yI
N ˝/ D f0g:
Proof. is follows from eorem 2.51 with f .x/ D ı.xI ˝/.
Given fi W Rn ! R for i D 1; : : : ; m, consider the maximum function
f .x/ WD max fi .x/; x 2 Rn ; (2.35)

iD1;:::;m
and for xN 2 Rn define the set

˚ ˇ
N WD i 2 f1; : : : ; mg ˇ fi .x/
I.x/ N D f .x/
N :
In the proposition below we assume that each function fi for i D 1; : : : ; m is continuous
at the reference point. A more general version follows afterward.
Proposition 2.54 Let fi W Rn ! R, i D 1; : : : ; m, be convex functions. Take any point xN 2

Tm
iD1 dom fi and assume that each fi is continuous at x
N . en we have the maximum rule
[
@ max fi /.x/ N D co @fi .x/:
N
i2I.x/
N
Proof. Let f .x/ be the maximum function defined in (2.35). We obviously have
m
\
epi f D epi fi :
iD1
Since fi .x/ N DW ˛N for any i … I.x/

N < f .x/ N , there exist a neighborhood U of xN and ı > 0 such
that fi .x/ < ˛ whenever .x; ˛/ 2 U .˛N ı; ˛N C ı/. It follows that .x; N 2 int epi fi , and so
N ˛/
N..x; N epi fi / D f.0; 0/g for such indices i .
N ˛/I
Let us show that the qualification condition from Corollary 2.19 is satisfied for the sets
˝i WD epi fi , i D 1; : : : ; m, at .x; N . Fix .vi ; i / 2 N .x;
N ˛/ N epi fi with
N ˛/I
m
X
.vi ; i / D 0:
iD1
en .vi ; i / D .0; 0/ for i … I.x/

N , and hence
X
.vi ; i / D 0:
i2I.x/
N
Note that fi .x/ N D ˛N for i 2 I.x/

N D f .x/ N . e proof of Proposition 2.47 tells us that i 0
P 1
for i 2 I.x/
N . us the equality i2I.x/ N i D 0 yields i D 0, and so vi 2 @ fi .x/ N D f0g for
every i 2 I.x/
N by eorem 2.29. is verifies that .vi ; i / D .0; 0/ for i D 1; : : : ; m. It follows
furthermore that
m
X X
N .x; N epi f D
N f .x//I N .x; N epi fi D
N ˛/I N .x; N epi fi :
N fi .x//I
iD1 i2I.x/
N
For any v 2 @f .x/

N , we have .v; 1/ 2 N..x; N epi f /, which allows us to find .vi ; i / 2
N f .x//I
N..x; N epi fi / for i 2 I.x/
N fi .x//I N such that
X
.v; 1/ D .vi ; i /:
i2I.x/
N
P P 1
is yields i2I.x/N i D 1 and v D N vi . If i D 0, then vi 2 @ fi .x/
i2I.x/ N D f0g. Hence we
assume without loss of generality that i > 0 for every i 2 I.x/N . Since N..x; N epi fi / is a
N fi .x//I
cone, it gives us .vi =i ; 1/ 2 N..x; N epi fi /, and so vi 2 i @fi .x/
N fi .x//I N . We get the represen-
P P
tation v D i2I.x/ N i ui , where ui 2 @fi .x/
N and i2I.x/ N i D 1. us
[
v 2 co @fi .x/:
N
i2I.x/
N
Note that the opposite inclusion follows directly from
@fi .x/ N for all i 2 I.x/;

N @f .x/ N
which in turn follows from the definition. Indeed, for any i 2 I.x/
N and v 2 @fi .x/
N we have
hv; x xi
N fi .x/ fi .x/
N D fi .x/ f .x/
N f .x/ N whenever x 2 Rn ;
f .x/
which implies that v 2 @f .x/

N and thus completes the proof.
To proceed with the general case, define
m
X
˚ ˇ
N WD .1 ; : : : ; m / ˇ i 0;
.x/ i D 1; i fi .x/
N f .x/
N D0
iD1
and observe from the definition of .x/ N hat

˚ ˇ X
N D .1 ; : : : ; m / ˇ i 0;
.x/ i D 1; i D 0 for i … I.x/
N :
i2I.x/
N
eorem 2.55 Let fi W Rn ! R be convex functions for i D 1; : : : ; m. Take any point xN 2
Tm
iD1 dom fi and assume that each fi is continuous at x
N for any i … I.x/N and that the quali-
fication condition
h X i
vi D 0; vi 2 @1 fi .x/
N H) vi D 0 for all i 2 I.x/
N
i2I.x/
N
holds. en we have the maximum rule

[n X ˇ o
ˇ
@ max fi .x/
N D i ı @fi .x/
N ˇ .1 ; : : : ; m / 2 .x/
N ;
i2I.x/
N
where i ı @fi .x/ N WD @1 fi .x/

N if i > 0, and i ı @fi .x/
N WD i @fi .x/ N if i D 0.
Proof. Let f .x/ be the maximum function from (2.35). Repeating the first part of the proof of
Proposition 2.54 tells us that N..x; N epi fi / D f.0; 0/g for i … I.x/
N f .x//I N . e imposed qualifica-
tion condition allows us to employ the intersection rule from Corollary 2.19 to get
m
X X
N .x; N epi f D
N f .x//I N .x; N epi fi D
N f .x//I N .x; N epi fi :
N fi .x//I (2.36)
iD1 i2I.x/
N
For any v 2 @f .x/N , we have .v; 1/ 2 N..x; N epi f /, and thus find by using (2.36) such
N f .x//I
pairs .vi ; i / 2 N..x; N epi fi / for i 2 I.x/
N fi .x//I N that
X
.v; 1/ D .vi ; i /:
i2I.x/
N
P P
is shows that i2I.x/ N i D 1 and v D N vi with vi 2 i ı @fi .x/
i2I.x/ N and i 0 for all i 2
N . erefore, we arrive at
I.x/
[n X ˇ o
ˇ
v2 i ı @fi .x/
N ˇ .1 ; : : : ; m / 2 .x/
N ;
i2I.x/
N
which justifies the inclusion “” in the maximum rule. e opposite inclusion can also be verified
easily by using representations (2.36).
2.6 SUBGRADIENTS OF OPTIMAL VALUE FUNCTIONS

is section is mainly devoted to the study of the class of optimal value/marginal functions defined
in (1.4). To proceed, we first introduce and examine the notion of coderivatives for set-valued
mappings, which is borrowed from variational analysis and plays a crucial role in many aspects of
the theory and applications of set-valued mappings; in particular, those related to subdifferenti-
ation of optimal value functions considered below.
Definition 2.56 Let F W Rn ! ! Rp be a set-valued mapping with convex graph and let .x;
N y/
N 2
p ! n
gph F . e  of F at .x; N is the set-valued mapping D F .x;
N y/ N R ! R with the
N y/W
values ˇ
˚
D F .x;
N y/.v/
N WD u 2 Rn ˇ .u; v/ 2 N .x; N gph F ;
N y/I v 2 Rp : (2.37)
It follows from (2.37) and definition (2.4) of the normal cone to convex sets that the
coderivative values of convex-graph mappings can be calculated by
n ˇ o
ˇ
D F .x;
N y/.v/
N D u 2 Rn ˇ hu; xi
N N D max
hv; yi hu; xi hv; yi :
.x;y/2gph F
Now we present explicit calculations of the coderivative in some particular settings.
Proposition 2.57 Let F W Rn ! Rp be a single-valued mapping given by F .x/ WD Ax C b , where

A W Rn ! Rp is a linear mapping, and where b 2 Rp . For any pair .x; N such that yN D AxN C b ,
N y/
we have the representation
D F .x;
N y/.v/
N D fA vg whenever v 2 Rp :
2.6. SUBGRADIENTS OF OPTIMAL VALUE FUNCTIONS 71
Proof. Lemma 2.50 tells us that
˚ ˇ
N .x; N gph F / D .u; v/ 2 Rn Rp ˇ u D
N y/I A v :
us u 2 D F .x;
N y/.v/
N if and only if u D A v .
Proposition 2.58 Given a convex function f W Rn ! R with xN 2 dom f , define the set-valued
mapping F W Rn !! R by

F .x/ WD f .x/; 1 :
Denoting yN WD f .x/
N , we have
8
ˆ
<@f .x/
N if > 0;
1
D F .x;
N y/./
N D @ f .x/N if D 0;
:̂
; if < 0:
In particular, if f is finite-valued, then D F .x;

N y/./
N N for all 0.
D @f .x/
Proof. It follows from the definition of F that

˚ ˇ ˚ ˇ
gph F D .x; / 2 Rn R ˇ 2 F .x/ D .x; / 2 Rn R ˇ f .x/ D epi f:
us the inclusion v 2 D F .x; N y/./

N is equivalent to .v; / 2 N..x; N epi f /. If > 0, then
N y/I
.v=; 1/ 2 N..x; N epi f /, which implies v= 2 @f .x/
N y/I N , and hence v 2 @f .x/
N . e formula
in the case where D 0 follows from the definition of the singular subdifferential. If < 0, then
D F .x;
N y/./
N D ; because 0 whenever .v; / 2 N..x; N epi f / by the proof of Propo-
N y/I
sition 2.47. To verify the last part of the proposition, observe that in this case D F .x;
N y/.0/
N D
1
@ f .x/N D f0g by Proposition 2.23, so we arrive at the conclusion.
e next simple example is very natural and useful in what follows.
Example 2.59 Define F W Rn ! ! Rp by F .x/ D K for every x 2 Rn , where K Rp

is a nonempty, convex set. en gph F D Rn K , and Proposition 2.11(i) tells us that
N..x; N gph F / D f0g N.yI
N y/I N K/ for any .x; N 2 Rn K . us
N y/
(
f0g if v 2 N.yI
N K/;
D F .x;
N y/.v/
N D
; otherwise:
In particular, for the case where K D Rp we have

(
f0g if v D 0;
D F .x;
N y/.v/
N D
; otherwise:
Now we derive a formula for calculating the subdifferential of the optimal value/marginal function
(1.4) given in the form ˚ ˇ
.x/ WD inf '.x; y/ ˇ y 2 F .x/
via the coderivative of the set-valued mapping F and the subdifferential of the function ' . First
we derive the following useful estimate.
Lemma 2.60 Let ./ be the optimal value function (1.4) generated by a convex-graph mapping
F W Rn ! ! Rp and a convex function ' W Rn Rp ! R. Suppose that .x/ > 1 for all x 2 Rn ,
fix some xN 2 dom , and consider the solution set
˚ ˇ
S.x/
N WD yN 2 F .x/ N ˇ .x/
N D '.x;
N y/
N :
If S.x/
N ¤ ;, then for any yN 2 S.x/
N we have
[
u C D F .x;
N y/.v/
N @.x/:
N (2.38)
.u;v/2@'.x;
N y/
N
Proof. Pick w from the left-hand side of (2.38) and find .u; v/ 2 @'.x; N with
N y/
w u 2 D F .x;
N y/.v/:
N
It gives us .w u; v/ 2 N..x; N gph F / and thus
N y/I
hw u; x xi
N hv; y N 0 for all .x; y/ 2 gph F;
yi
which shows that whenever y 2 F .x/ we have
hw; x xi
N hu; x xi
N C hv; y yi
N '.x; y/ '.x;
N y/
N D '.x; y/ .x/:
N
is implies in turn the estimate
hw; x xi
N inf '.x; y/ .x/
N D .x/ .x/;
N
y2F .x/
which justifies the inclusion w 2 @.x/

N , and hence completes the proof of (2.38).
eorem 2.61 Consider the optimal value function (1.4) under the assumptions of Lemma 2.60.
For any yN 2 S.x/
N , we have the equality
[
@.x/N D u C D F .x;
N y/.v/
N (2.39)
.u;v/2@'.x;
N y/
N
provided the validity of the qualification condition

@1 '.x;
N y/N \ N..x; N gph F / D f.0; 0/g;
N y/I (2.40)
which holds, in particular, when ' is continuous at .x; N .
N y/
2.6. SUBGRADIENTS OF OPTIMAL VALUE FUNCTIONS 73
Proof. Taking Lemma 2.60 into account, it is only required to verify the inclusion “” in (2.39).
To proceed, pick w 2 @.x/ N . For any x 2 Rn , we have
N and yN 2 S.x/
hw; x xi
N .x/ .x/N D .x/ '.x; N y/
N
'.x; y/ '.x; N for all y 2 F .x/:
N y/
is implies that, whenever .x; y/ 2 Rn Rp , the following inequality holds:

hw; x xi N C h0; y yiN '.x; y/ C ı .x; y/I gph F / '.x;
N y/
N C ı .x; N gph F :
N y/I
Denote further f .x; y/ WD '.x; y/ C ı..x; y/I gph F / and deduce from the subdifferential sum
rule in eorem 2.44 under the qualification condition (2.40) that the inclusion

.w; 0/ 2 @f .x;
N y/
N D @'.x; N y/
N C N .x; N gph F
N y/I
is satisfied. is shows that .w; 0/ D .u1 ; v1 / C .u2 ; v2 / with .u1 ; v1 / 2 @'.x; N and .u2 ; v2 / 2
N y/
N..x; N gph F /. is yields v2 D v1 , so .u2 ; v1 / 2 N..x;
N y/I N gph F /. Finally, we get u2 2
N y/I
D F .x; N 1 / and therefore
N y/.v
w D u1 C u2 2 u1 C D F .x;
N y/.v
N 1 /;

Let us present several important consequences of eorem 2.61. e first one concerns the
following chain rule for convex compositions.
Corollary 2.62 Let f W Rn ! R be a convex function and let ' W R ! R be a nondecreasing convex
function. Take xN 2 Rn and denote yN WD f .x/
N assuming that yN 2 dom ' . en
[
@.' ı f /.x/
N D @f .x/N
2@'.y/
N
under the assumption that either @1 '.y/

N D f0g or 0 … @f .x/
N .
Proof. e composition ' ı f is a convex function by Proposition 1.39. Define the set-valued
mapping F .x/ WD Œf .x/; 1/ and observe that
.' ı f /.x/ D inf '.y/

y2F .x/
since ' is nondecreasing. Note each of the assumptions @1 '.y/ N D f0g and 0 … @f .x/
N ensure the
validity of the qualification condition (2.40). It follows from eorem 2.61 that
[
@.' ı f /.x/
N D D F .x;
N y/./:
N
2@'.y/
N
Taking again into account that ' is nondecreasing yields 0 for every 2 @'.y/
N . is tells us
by Proposition 2.58 that
[ [
@.' ı f /.x/
N D D F .x;
N y/./
N D @f .x/;
N
2@'.y/
N 2@'.y/
N
which verifies the claimed chain rule of this corollary.

e next corollary addresses calculating subgradients of the optimal value functions over
parameter-independent constraints.
Corollary 2.63 Let ' W Rn Rp ! R, where the function '.x; / is bounded below on Rp for every
x 2 Rn . Define
˚ ˇ ˚ ˇ
.x/ WD inf '.x; y/ ˇ y 2 Rp ; S.x/ WD y 2 Rp ˇ '.x; y/ D .x/ :
For xN 2 dom and for every yN 2 S.x/

N , we have
˚ ˇ
@.x/N D u 2 Rn ˇ .u; 0/ 2 @'.x;
N y/
N :
Proof. Follows directly from eorem 2.61 for F .x/ Rp and Example 2.59. Note that the
qualification condition (2.40) is satisfied since N..x; N gph F / D f.0; 0/g.
N y/I
e next corollary specifies the general result for problems with affine constraints.
Corollary 2.64 Let B W Rp ! Rn be given by B.y/ WD Ay C b via a linear mapping AW Rp ! Rn

and an element b 2 Rn and let ' W Rp ! R be a convex function bounded below on the inverse image
set B 1 .x/ for all x 2 Rn . Define
˚ ˇ ˚ ˇ
.x/ WD inf '.y/ ˇ B.y/ D xg; S.x/ WD y 2 B 1 .x/ ˇ '.y/ D .x/
and fix xN 2 Rn with .x/

N < 1 and S.x/
N ¤ ;. For any yN 2 S.x/
N , we have

N D .A / 1 @'.y/
@.x/ N :
1
Proof. e setting of this corollary corresponds to eorem 2.61 with F .x/ D B .x/ and
'.x; y/ '.y/. en we have
˚ ˇ
N .x; N gph F D .u; v/ 2 Rn Rp ˇ A u D v ;
N y/I
which therefore implies the coderivative representation

˚ ˇ
D F .x;
N y/.v/
N D u 2 Rn ˇ A u D v D .A / 1 .v/; v 2 Rp :
2.7. SUBGRADIENTS OF SUPPORT FUNCTIONS 75
It is easy to see that the qualification condition (2.40) is automatic in this case. is yields
[
@.x/
N D D F .x;
N y/.v/
N D .A / 1 .@'.y//;
N
v2@'.y/
N
and thus we complete the proof.

Finally in this section, we present a useful application of Corollary 2.64 to deriving a cal-
culus rule for subdifferentiation of infimal convolutions important in convex analysis.
Corollary 2.65 Given convex functions f1 ; f2 W Rn ! R, consider their infimal convolution

˚ ˇ
.f1 ˚ f2 /.x/ WD inf f1 .x1 / C f2 .x2 / ˇ x1 C x2 D x ; x 2 Rn ;
and suppose that .f1 ˚ f2 /.x/ > 1 for all x 2 Rn . Fix xN 2 dom .f1 ˚ f2 / and xN 1 ; xN 2 2 Rn sat-
isfying xN D xN 1 C xN 2 and .f1 ˚ f2 /.x/
N D f1 .xN 1 / C f2 .xN 2 /. en we have
@.f1 ˚ f2 /.x/
N D @f1 .xN 1 / \ @f2 .xN 2 /:
Proof. Define 'W Rn Rn ! R and AW Rn Rn ! Rn by
'.x1 ; x2 / WD f1 .x1 / C f2 .x2 /; A.x1 ; x2 / WD x1 C x2 :
It is easy to see that A v D .v; v/ for any v 2 Rn and that @'.xN 1 ; xN 2 / D .@f1 .xN 1 /; @f2 .xN 2 // for
.xN 1 ; xN 2 / 2 Rn Rn . en v 2 @.f1 ˚ f2 /.x/
N if and only if we have A v D .v; v/ 2 @'.xN 1 ; xN 2 /,
which means that v 2 @f1 .xN 1 / \ @f2 .xN 2 /. us the claimed calculus result follows from Corol-
lary 2.64.
2.7 SUBGRADIENTS OF SUPPORT FUNCTIONS

is short section is devoted to the study of subgradient properties for an important class of
nonsmooth convex functions known as support functions of convex sets.
Definition 2.66 Given a nonempty .not necessarily convex/ subset ˝ of Rn , the  
of ˝ is defined by ˚ ˇ
˝ .x/ WD sup hx; !i ˇ ! 2 ˝ : (2.41)
First we present some properties of (2.41) that can be deduced from the definition.
Proposition 2.67 e following assertions hold:

(i) For any nonempty subset ˝ of Rn , the support function ˝ is subadditive, positively homogeneous,
and hence convex.
(ii) For any nonempty subsets ˝1 ; ˝2 Rn and any x 2 Rn , we have
˚
˝1 C˝2 .x/ D ˝1 .x/ C ˝2 .x/; ˝1 [˝2 .x/ D max ˝1 .x/; ˝2 .x/ :
e next result calculates subgradients of support functions via convex separation.
eorem 2.68 Let ˝ be a nonempty, closed, convex subset of Rn and let

˚ ˇ
S.x/ WD p 2 ˝ ˇ hx; pi D ˝ .x/ :
For any xN 2 dom ˝ , we have the subdifferential formula
@˝ .x/
N D S.x/:
N
Proof. Fix p 2 @˝ .x/

N and get from definition (2.13) that
hp; x xi
N ˝ .x/ N whenever x 2 Rn :
˝ .x/ (2.42)
Taking this into account, it follows for any u 2 Rn that
hp; ui D hp; u C xN xi
N ˝ .u C x/
N ˝ .x/
N ˝ .u/ C ˝ .x/
N N D ˝ .u/: (2.43)
˝ .x/
Assuming that p 62 ˝ , we use the separation result of Proposition 2.1 and find u 2 Rn with
˚ ˇ
˝ .u/ D sup hu; !i ˇ ! 2 ˝ < hp; ui;
which contradicts (2.43) and so verifies p 2 ˝ . From (2.43) we also get hp; xi N . Letting
N ˝ .x/
x D 0 in (2.42) gives us ˝ .x/ N . Hence hp; xi
N hp; xi N and thus p 2 S.x/
N D ˝ .x/ N .
Now fix p 2 S.x/ N and get from p 2 ˝ and hp; xi N that
N D ˝ .x/
hp; x xi
N D hp; xi hp; xi
N ˝ .x/ N for all x 2 Rn :
˝ .x/
is shows that p 2 @˝ .x/

N and thus completes the proof.
Note that eorem 2.68 immediately implies that
˚ ˇ
@˝ .0/ D S.0/ D ! 2 ˝ ˇ h0; !i D ˝ .0/ D 0 D ˝ (2.44)
for any nonempty, closed, convex set ˝ Rn .

e last result of this section justifies a certain monotonicity property of support functions
with respect to set inclusions.
Proposition 2.69 Let ˝1 and ˝2 be a nonempty, closed, convex subsets of Rn . en the inclusion
˝1 ˝2 holds if and only if ˝1 .v/ ˝2 .v/ for all v 2 Rn .
2.8. FENCHEL CONJUGATES 77
n
Proof. Taking ˝1 ˝2 as in the proposition, for any v 2 R we have
˚ ˇ ˚ ˇ
˝1 .v/ D sup hv; xi ˇ x 2 ˝1 sup hv; xi ˇ x 2 ˝2 D ˝2 .v/:
Conversely, suppose that ˝1 .v/ ˝2 .v/ whenever v 2 Rn . Since ˝1 .0/ D ˝2 .0/ D 0, it fol-
lows from definition (2.13) of the subdifferential that
@˝1 .0/ @˝2 .0/;
which yields ˝1 ˝2 by formula (2.44).
2.8 FENCHEL CONJUGATES

Many important issues of convex analysis and its applications (in particular, to optimization) are
based on duality. e following notion plays a crucial role in duality considerations.
Definition 2.70 Given a function f W Rn ! R .not necessarily convex/, its F 
f W Rn ! Œ 1; 1 is
˚ ˇ ˚ ˇ
f .v/ WD sup hv; xi f .x/ ˇ x 2 Rn D sup hv; xi f .x/ ˇ x 2 dom f : (2.45)
Note that f .v/ D 1 is allowed in (2.45) while f .v/ > 1 for all v 2 Rn if dom f ¤
;. It follows directly from the definitions that for any nonempty set ˝ Rn the conjugate to its
indicator function is the support function of ˝ :
˚ ˇ

ı˝ .v/ D sup hv; xi ˇ x 2 ˝ D ˝ .v/; v 2 ˝: (2.46)
e next two propositions can be easily verified.
Proposition 2.71 Let f W Rn ! R be a function, not necessarily convex, with dom f ¤ ;. en
its Fenchel conjugate f is convex on Rn .
Proof. Function (2.45) is convex as the supremum of a family of affine functions.
Proposition 2.72 Let f; g W Rn ! R be such that f .x/ g.x/ for all x 2 Rn . en we have
f .v/ g .v/ for all v 2 Rn .
Proof. For any fixed v 2 Rn , it follows from (2.45) that
hv; xi f .x/ hv; xi g.x/; x 2 Rn :
is readily implies the relationships

˚ ˇ ˚ ˇ
f .v/ D sup hv; xi f .x/ ˇ x 2 Rn sup hv; xi g.x/ ˇ x 2 Rn D g .v/
for all v 2 Rn , and therefore f g on Rn .

Figure 2.8: Fenchel conjugate.
e following two examples illustrate the calculation of conjugate functions.
Example 2.73 (i) Given a 2 Rn and b 2 R, consider the affine function
f .x/ WD ha; xi C b; x 2 Rn :
en it can be seen directly from the definition that

(
b if v D a;
f .v/ D
1 otherwise:
(ii) Given any p > 1, consider the power function

8 p
<x if x 0;
f .x/ WD p
:
1 otherwise:
For any v 2 R, the conjugate of this function is given by

n x p ˇˇ o n xp ˇ
ˇ
o
f .v/ D sup vx ˇx0 D inf vx ˇ x 0 :
p p
It is clear that f .v/ D 0 if v 0 since in this case vx p 1 x p 0 when x 0. Considering

the case of v > 0, we see that function v .x/ WD p 1 x p vx is convex and differentiable on
0 p 1 0 1=.p 1/
.0; 1/ with v .x/
Dx v . us D 0 if and only if x D v
v .x/ , and so v attains
its minimum at x D v 1=.p 1/
. erefore, the conjugate function is calculated by
1 p=.p
f .v/ D 1 v 1/
; v 2 Rn :
p
1 1
Taking q from q D1 p , we express the conjugate function as
8
<0 if v 0;
f .v/ D v q
: otherwise:
q
Note that the calculation in Example 2.73(ii) shows that
xp vq
vx C for any x; v 0:
p q
e first assertion of the next proposition demonstrates that such a relationship in a more general
setting. To formulate the second assertion below, we define the biconjugate of f as the conjugate
of f , i.e., f .x/ WD .f / .x/.
Proposition 2.74 Let f W Rn ! R be a function with dom f ¤ ;. en we have:

(i) hv; xi f .x/ C f .v/ for all x; v 2 Rn .
(ii) f .x/ f .x/ for all x 2 Rn .
Proof. Observe first that (i) is obvious if f .x/ D 1. If x 2 dom f , we get from (2.45) that
f .v/ hv; xi f .x/, which verifies (i). It implies in turn that
˚ ˇ
sup hv; xi f .v/ ˇ v 2 Rn f .x/ for all x; v 2 Rn ;
which thus verifies (ii) and completes the proof.
e following important result reveals a close relationship between subgradients and

Fenchel conjugates of convex functions.
eorem 2.75 For any convex function f W Rn ! R and any xN 2 dom f , we have that v 2
N if and only if
@f .x/
N C f .v/ D hv; xi:
f .x/ N (2.47)
Proof. Taking any v 2 @f .x/
N and using definition (2.13) gives us
f .x/
N C hv; xi N for all x 2 Rn :
f .x/ hv; xi
is readily implies the inequality

˚ ˇ
N C f .v/ D f .x/
f .x/ N C sup hv; xi f .x/ ˇ x 2 Rn hv; xi:
N
Since the opposite inequality holds by Proposition 2.74(i), we arrive at (2.47).

N C f .v/ D hv; xi
Conversely, suppose that f .x/ N . Applying Proposition 2.74(i), we get the
estimate f .v/ hv; xi f .x/ for every x 2 Rn . is shows that v 2 @f .x/

N .
e result obtained allows us to find conditions ensuring that the biconjugate f of a
convex function agrees with the function itself.
Proposition 2.76 Let xN 2 dom f for a convex function f W Rn ! R. Suppose that @f .x/
N ¤ ;. en

we have the equality f .x/ N .
N D f .x/
Proof. By Proposition 2.74(ii) it suffices to verify the opposite inequality therein. Fix v 2 @f .x/
N
and get hv; xi N C f .v/ by the preceding theorem. is shows that
N D f .x/
˚ ˇ
f .x/
N D hv; xi
N f .v/ sup hx;N vi f .v/ ˇ v 2 Rn D f .x/; N
which completes the proof of this proposition.

Taking into account Proposition 2.47, we arrive at the following corollary.
Corollary 2.77 Let f W Rn ! R be a convex function and let xN 2 int.dom f /. en we have the
equality f .x/ N .
N D f .x/
Finally in this section, we prove a necessary and sufficient condition for the validity of the
biconjugacy equality f D f known as the Fenchel-Moreau theorem.
eorem 2.78 Let f W Rn ! R be a function with dom f ¤ ; and let A be the set of all affine
functions of the form '.x/ D ha; xi C b for x 2 Rn , where a 2 Rn and b 2 R. Denote
˚ ˇ
A.f / WD ' 2 A ˇ '.x/ f .x/ for all x 2 Rn :
en A.f / ¤ ; whenever epi f is closed and convex. Moreover, the following are equivalent:
(i) epi f is closed and convex.
(ii) f .x/ D sup'2A.f / '.x/ for all x 2 Rn .
(iii) f .x/ D f .x/ for all x 2 Rn .
Proof. Let us first show that A.f / ¤ ;. Fix any x0 2 dom f and choose 0 < f .x0 /. en
N N / 2 Rn R and > 0 such that
.x0 ; 0 / … epi f . By Proposition 2.1, there exist .v;
hv;
N xi C
N < hv;
N x0 i C
N 0 whenever .x; / 2 epi f: (2.48)
Since .x0 ; f .x0 / C ˛/ 2 epi f for all ˛ 0, we get
.f
N .x/
N C ˛/ <
N 0 whenever ˛ 0:
is implies N < 0 since if not, we can let ˛ ! 1 and arrive at a contradiction. For any x 2
dom f , it follows from (2.48) as .x; f .x// 2 epi f that
hv;
N xi C f
N .x/ < hv;
N x0 i C N 0 for all x 2 dom f:
is allows us to conclude that

vN
f .x/ > h ; x0 xi C 0 if x 2 dom f:
N N
vN
Define now '.x/ WD h ; x0 xi C 0 . en ' 2 A.f /, and so A.f / ¤ ;.
N N
Let us next prove that (i)H)(ii). By definition we need to show that for any 0 < f .x0 /
there is ' 2 A.f / such that 0 < '.x0 /. Since .x0 ; 0 / … epi f , we apply again Proposition 2.1
to obtain (2.48). In the case where x0 2 dom f , it was proved above that ' 2 A.f /. Moreover,

we have '.x0 / D 0 < 0 since N < 0. Consider now the case where x0 … dom f . It follows
N
from (2.48) by taking any x 2 dom f and letting ! 1 that N 0. If N < 0, we can apply the
same procedure and arrive at the conclusion. Hence we only need to consider the case where
N D 0. In this case
N x x0 i C < 0 whenever x 2 dom f:
hv;
Since A.f / ¤ ;, choose '0 2 A.f / and define
'k .x/ WD '0 .x/ C k.hv;

N x x0 i C /; k 2 N :
It is obvious that 'k 2 A.f / and 'k .x0 / D '0 .x0 / C k > 0 for large k . is justifies (ii).
Let us now verify implication (ii)H)(iii). Fix any ' 2 A.f /. en ' f , and hence ' f .
Applying Proposition 2.76 ensures that ' D ' f . It follows therefore that
˚ ˇ
f .x/ D sup '.x/ ˇ ' 2 A.f /g f for every x 2 Rn :
e opposite inequality f f holds by Proposition 2.74(ii), and thus f D f .

e last implication (iii)H)(i) is obvious because the set epi g is always closed and convex
for any function g W Rn ! R.
2.9 DIRECTIONAL DERIVATIVES
Our next topic is directional differentiability of convex functions and its relationships with subdif-
ferentiation. In contrast to classical analysis, directional derivative constructions in convex analysis
are one-sided and related to directions with no classical plus-minus symmetry.
Definition 2.79 Let f W Rn ! R be an extended-real-valued function and let xN 2 dom f . e

  of the function f at the point xN in the direction d 2 Rn is the following
limit—if it exists as either a real number or ˙1:
f .xN C t d / f .x/
N
f 0 .xI
N d / WD lim : (2.49)
t !0C t
Note that construction (2.49) is sometimes called the right directional derivative f at xN in
the direction d . Its left counterpart is defined by
f .xN C t d / f .x/
N
f 0 .xI
N d / WD lim :
t !0 t
It is easy to see from the definitions that
f 0 .xI
N d/ D N d / for all d 2 Rn ;
f 0 .xI
and thus properties of the left directional derivative f 0 .xI

N d / reduce to those of the right one
(2.49), which we study in what follows.
Lemma 2.80 Given a convex function f W Rn ! R with xN 2 dom f and given d 2 Rn , define
f .xN C t d / f .x/
N
'.t / WD ; t > 0:
t
en the function ' is nondecreasing on .0; 1/.
Proof. Fix any numbers 0 < t1 < t2 and get the representation
t1 t1
xN C t1 d D xN C t2 d C 1 x:
N
t2 t2
It follows from the convexity of f that
t1 t1
f .xN C t1 d / f .xN C t2 d / C 1 f .x/;
N
t2 t2
which implies in turn the inequality
f .xN C t1 d / f .x/
N f .xN C t2 d / f .x/
N
'.t1 / D D '.t2 /:
t1 t2
is verifies that ' is nondecreasing on .0; 1/.
2.9. DIRECTIONAL DERIVATIVES 83
e next proposition establishes the directional differentiability of convex functions.
Proposition 2.81 For any convex function f W Rn ! R and any xN 2 dom f , the directional
N d / .and hence its left counterpart/ exists in every direction d 2 Rn . Furthermore,
derivative f 0 .xI
it admits the representation via the function ' is defined in Lemma 2.80:
f 0 .xI
N d / D inf '.t /; d 2 Rn
t>0
Proof. Lemma 2.80 tells us that the function ' is nondecreasing. us we have
f .xN C td / f .x/
N
f 0 .xI
N d / D lim D lim '.t / D inf '.t /;
t !0C t t !0C t >0
which verifies the results claimed in the proposition.
Corollary 2.82 If f W Rn ! R is a convex function, then f 0 .xI

N d / is a real number for any xN 2
n
int.dom f / and d 2 R .
Proof. It follows from eorem 2.29 that f is locally Lipschitz continuous around xN , i.e., there
is ` 0 such that
ˇ f .xN C td / f .x/
N ˇˇ `t kd k
ˇ
ˇ ˇ D `kd k for all small t > 0;
t t
which shows that jf 0 .xI
N d /j `kd k < 1.
To establish relationships between directional derivatives and subgradients of general con-
vex functions, we need the following useful observation.
Lemma 2.83 For any convex function f W Rn ! R and xN 2 dom f , we have
f 0 .xI
N d / f .xN C d / N whenever d 2 Rn :
f .x/
Proof. Using Lemma 2.80, we have for the function ' therein that
'.t/ '.1/ D f .xN C d / N for all t 2 .0; 1/;

f .x/
which justifies the claimed property due to f 0 .xI

N d / D inf t>0 '.t / '.1/.
eorem 2.84 Let f W Rn ! R be convex with xN 2 dom f . e following are equivalent:

(i) v 2 @f .x/N .
N d / for all d 2 Rn .
(ii) hv; d i f 0 .xI
(iii) f 0 .xI N d / for all d 2 Rn .
N d / hv; d i f 0 .xI
Proof. Picking any v 2 @f .x/
N and t > 0, we get
hv; td i f .xN C td / N whenever d 2 Rn ;
f .x/
which verifies the implication (i)H)(ii) by taking the limit as t ! 0C . Assuming now that as-
sertion (ii) holds, we get by Lemma 2.83 that
hv; d i f 0 .xI
N d / f .xN C d / N for all d 2 Rn :
f .x/
It ensures by definition (2.13) that v 2 @f .x/N , and thus assertions (i) and (ii) are equivalent.
It is obvious that (iii) yields (ii). Conversely, if (ii) is satisfied, then for d 2 Rn we have
hv; d i f 0 .xI
N d /, and thus
f 0 .xI
N d/ D N d / hv; d i for any d 2 Rn :
f 0 .xI
is justifies the validity of (iii) and completes the proof of the theorem.
Let us next list some properties of (2.49) as a function of the direction.
Proposition 2.85 For any convex function f W Rn ! R with xN 2 dom f , we define the directional
function .d / WD f 0 .xI
N d /, which satisfies the following properties:
(i) .0/ D 0.
(ii) .d1 C d2 / .d1 / C .d2 / for all d1 ; d2 2 Rn .
(iii) .˛d / D ˛ .d / whenever d 2 Rn and ˛ > 0.
(iv) If furthermore xN 2 int.dom f /, then is finite on Rn .
Proof. It is straightforward to deduce properties (i)–(iii) directly from definition (2.49). For in-
stance, (ii) is satisfied due to the relationships
f .xN C t.d1 C d2 // f .x/ N
.d1 C d2 / D lim
t!0C t
xN C 2td1 C xN C 2t d2
f. / f .x/
N
D lim 2
t!0C t
f .xN C 2t d1 / f .x/
N f .xN C 2t d2 / f .x/
N
lim C lim D .d1 / C .d2 /:
t!0C 2t t!0C 2t
us it remains to check that .d / is finite for every d 2 Rn when xN 2 int.dom f /. To proceed,
choose ˛ > 0 so small that xN C ˛d 2 dom f . It follows from Lemma 2.83 that
.˛d / D f 0 .xI
N ˛d / f .xN C ˛d / f .x/
N < 1:
Employing (iii) gives us .d / < 1. Further, we have from (i) and (ii) that

0D .0/ D d C . d / .d / C . d /; d 2 Rn ;
which implies that .d / . d /. is yields .d / > 1 and so verifies (iv).
2.9. DIRECTIONAL DERIVATIVES 85
To derive the second major result of this section, yet another lemma is needed.
Lemma 2.86 We have the following relationships between a convex function f W Rn ! R and the
directional function defined in Proposition 2.85:
N D @ .0/.
(i) @f .x/
(ii) .v/ D ı˝ .v/ for all v 2 Rn , where ˝ WD @ .0/.

Proof. It follows from eorem 2.84 that v 2 @f .x/

N if and only if
hv; d 0i D hv; d i f 0 .xI

N d/ D .d / D .d / .0/; d 2 Rn :
is is equivalent to v 2 @ .0/, and hence (i) holds.

To justify (ii), let us first show that .v/ D 0 for all v 2 ˝ D @ .0/. Indeed, we have
˚ ˇ

.v/ D sup hv; d i .d / ˇ d 2 Rn hv; 0i .0/ D 0:
Picking now any v 2 @ .0/ gives us
hv; d i D hv; d 0i .d / .0/ D .d /; d 2 Rn ;
which implies therefore that

˚ ˇ

.v/ D sup hv; d i .d / ˇ d 2 Rn 0
and so ensures the validity of .v/ D 0 for any v 2 @ .0/.

It remains to verify that .v/ D 1 if v … @ .0/. For such an element v , find d0 2 Rn
with hv; d0 i > .d0 /. Since is positively homogeneous by Proposition 2.85, it follows that
˚ ˇ

.v/ D sup hv; d i .d / ˇ d 2 Rn sup.hv; t d0 i .t d0 //
t>0
D sup t.hv; d0 i .d0 // D 1;
t>0
which completes the proof of the lemma.
Now we are ready establish a major relationship between the directional derivative and the
subdifferential of an arbitrary convex function.
eorem 2.87 Given a convex function f W Rn ! R and a point xN 2 int.dom f /, we have

˚ ˇ
N d / D max hv; d i ˇ v 2 @f .x/
f 0 .xI N for any d 2 Rn :
Proof. It follows from Proposition 2.76 that
f 0 .xI
N d/ D .d / D
.d /; d 2 Rn ;
by the properties of the function .d / D f 0 .xIN d / from Proposition 2.85. Note that is a finite
convex function in this setting. Employing now Lemma 2.86 tells us that .v/ D ı˝ .v/, where
˝ D @ .0/ D @f .x/ N . Hence we have by (2.46) that
˚ ˇ

.d / D ı˝ .d / D sup hv; d i ˇ v 2 ˝ :
Since the set ˝ D @f .x/
N is compact by Proposition 2.47, we complete the proof.
2.10 SUBGRADIENTS OF SUPREMUM FUNCTIONS

Let T be a nonempty subset of Rp and let g W T Rn ! R. For convenience, we also use the
notation g t .x/ WD g.t; x/. e supremum function f W Rn ! R for g t over T is
f .x/ WD sup g.t; x/ D sup g t .x/; x 2 Rn : (2.50)
t 2T t2T
If the supremum in (2.50) is attained (this happens, in particular, when the index set T is compact
and g.; x/ is continuous), then (2.50) reduces to the maximum function, which can be written in
form (2.35) when T is a finite set.
e main goal of this section is to calculate the subdifferential (2.13) of (2.50) when the
functions g t are convex. Note that in this case the supremum function (2.50) is also convex by
Proposition 1.43. In what follows we assume without mentioning it that the functions g t W Rn ! R
are convex for all t 2 T .
For any point xN 2 Rn , define the active index set
˚ ˇ
N WD t 2 T ˇ g t .x/
S.x/ N D f .x/N ; (2.51)
which may be empty if the supremum in (2.50) is not attained for x D xN . We first present a simple
lower subdifferential estimate for (2.66).
Proposition 2.88 Let dom f ¤ ; for the supremum function (2.50) over an arbitrary index set T .
For any xN 2 dom f , we have the inclusion
[
clco @g t .x/
N @f .x/:
N (2.52)
t 2S.x/
N
Proof. Inclusion (2.52) obviously holds if S.x/

N D ;. Supposing now that S.x/
N ¤ ;, fix t 2 S.x/
N
and v 2 @g t .x/
N . en we get g t .x/ N and therefore
N D f .x/
hv; x xi
N g t .x/ g t .x/
N D g t .x/ f .x/
N f .x/ f .x/;
N
which shows that v 2 @f .x/
N . Since the subgradient set @f .x/
N is a closed and convex, we conclude
that inclusion (2.52) holds in this case.
2.10. SUBGRADIENTS OF SUPREMUM FUNCTIONS 87
Further properties of the supremum/maximum function (2.50) involve the following useful
extensions of continuity. Its lower version will be studied and applied in Chapter 4.
Definition 2.89 Let ˝ be a nonempty subset of Rn . A function f W ˝ ! R is  -

  xN 2 ˝ if for every sequence fxk g in ˝ converging to xN it holds
lim sup f .xk / f .x/:

N
k!1
We say that f is    ˝ if it is upper semicontinuous at every point of this set.
Proposition 2.90 Let T be a nonempty, compact subset of Rp and let g.I x/ be upper semicontinuous
on T for every x 2 Rn . en f in (2.50) is a finite-valued convex function.
Proof. We need to check that f .x/ < 1 for every x 2 Rn . Suppose by contradiction that f .x/
N D
n
1 for some xN 2 R . en there exists a sequence ftk g T such that g.tk ; x/ N ! 1. Since T
is compact, assume without loss of generality that tk ! tN 2 T . By the upper semicontinuity of
N on T we have
g.; x/
N g.tN; x/
1 D lim sup g.tk ; x/ N < 1;
k!1
which is a contradiction that completes the proof of the proposition.
Proposition 2.91 Let the index set ; ¤ T Rp be compact and let g.I x/ be upper semicontinuous
for every x 2 Rn . en the active index set S.x/
N T is nonempty and compact for every xN 2 Rn .
Furthermore, we have the compactness in Rn of the convex hull
[
C WD co @g t .x/:
N
t2S.x/
N
Proof. Fix xN 2 Rn and get from Proposition 2.90 that f .x/ N < 1 for the supremum function
(2.50). Consider an arbitrary sequence ftk g T with g.tk ; x/ N as k ! 1. By the com-
N ! f .x/
N N
pactness of T we find t 2 T such that tk ! t along some subsequence. It follows from the as-
sumed upper semicontinuity of g that
f .x/ N g.tN; x/
N D lim sup g.tk ; x/ N f .x/;
N
k!1
which ensures that tN 2 S.x/

N by the construction of S in (2.51), and thus S.x/
N ¤ ;. Since g.t; x/
N
N , for any t 2 T we get the representation
f .x/
˚ ˇ
N D t 2 T ˇ g.t; x/
S.x/ N f .x/N
implying the closedness of S.x/
N (and hence its compactness since S.x/
N T ), which is an imme-
diate consequence of the upper semicontinuity of g.I x/
N .
It remains to show that the subgradient set
[
Q WD @g t .x/
N
t 2S.x/
N
is compact in Rn ; this immediately implies the claimed compactness of C D co Q, since the

convex hull of a compact set is compact (show it, or see Corollary 3.7 in the next chapter). To
proceed with verifying the compactness of Q, we recall first that Q @f .x/ N by Proposition 2.88.
is implies that the set Q is bounded in Rn , since the subgradient set of the finite-valued convex
function is compact by Proposition 2.47. To check the closedness of Q, take a sequence fvk g Q
converging to some vN and for each k 2 N and find tk 2 S.x/ N such that vk 2 @g tk .x/
N . Since S.x/
N
is compact, we may assume that tk ! tN 2 S.x/ N D g.tN; x/
N . en g.tk ; x/ N D f .x/N and
hvk ; x xi
N g.tk ; x/ g.tk ; x/
N D g.tk ; x/ N for all x 2 Rn
g.tN; x/
by definition (2.13). is gives us the relationships
hv;
N x N D lim suphvk ; x
xi N lim sup g.tk ; x/
xi g.tN; x/
N g.tN; x/ g.tN; x/
N D gtN .x/ gtN .x/;
N
k!1 k!1
which verify that vN 2 @gtN.x/

N Q and thus complete the proof.
e next result is of its own interest while being crucial for the main subdifferential result
of this section.
eorem 2.92 In the setting of Proposition 2.91 we have
N v/ C .v/ for all xN 2 Rn ; v 2 Rn

f 0 .xI
via the support function of the set C defined therein.
Proof. It follows from Proposition 2.81 and Proposition 2.82 that 1 < f 0 .xI N v/ D
inf>0 './ < 1, where the function ' is defined in Lemma 2.80 and is proved to be nonde-
creasing on .0; 1/. us there is a strictly decreasing sequence k # 0 as k ! 1 with
f .xN C k v/ f .x/
N
f 0 .xI
N v/ D lim ; k 2N:
k!1 k
For every k , we select tk 2 T such that f .xN C k v/ D g.tk ; xN C k v/. e compactness of T
allows us to find tN 2 T such that tk ! tN 2 T along some subsequence. Let us show that tN 2 S.x/
N .
Since g.tk ; / is convex and k < 1 for large k , we have

N (2.53)
g.tk ; xN C k v/ D g tk ; k .xN C v/ C .1 k /xN k g.tk ; xN C v/ C .1 k /g.tk ; x/;
2.10. SUBGRADIENTS OF SUPREMUM FUNCTIONS 89
which implies by the continuity of f (any finite-valued convex function is continuous) and the
upper semicontinuity of g.I x/ that
N D lim f .xN C k v/ D lim g.tk ; xN C k v/ g.tN; x/:

f .x/ N
k!1 k!1
N D g.tN; x/
is yields f .x/ N , and so tN 2 S.x/
N . Moreover, for any > 0 and large k , we get
g.tk ; xN C v/ C with WD g.tN; xN C v/
by the upper semicontinuity of g . It follows from (2.53) that

g.tk ; xN C k v/ k . C /
g.tk ; x/
N ! f .x/;
N
1 k
which tells us that limk!1 g.tk ; x/ N due to
N D f .x/
N lim inf g.tk ; x/

f .x/ N g.tN; x/
N lim sup g.tk ; x/ N D f .x/:
N
k!1 k!1
Fix further any > 0 and, taking into account that k < for all k sufficiently large, deduce from
Lemma 2.80 that
f .xN C k v/ f .x/
N g.tk ; xN C k v/ g.tN; x/
N
D
k k
g.tk ; xN C k v/ g.tk ; x/
N

k
g.tk ; xN C v/ g.tk ; x/
N

is ensures in turn the estimates
f .xN C k v/ f .x/
N g.tk ; xN C v/ g.tk ; x/
N
lim sup lim sup
k!1 k k!1
g.tN; xN C v/ g.tN; x/ N

gtN .xN C v/ gtN .x/ N
D ;

which implies the relationships
gtN .xN C v/ gtN .x/
N
f 0 .xI
N v/ inf D gt0N .xI
N v/:
>0
Using finally eorem 2.87 and the inclusion @gtN .x/
N C from Proposition 2.88 give us
f 0 .xI
N v/ gt0N .xI
N v/ D sup hw; vi sup hw; vi D C .v/;
w2@gtN .x/
N w2C
and justify therefore the statement of the theorem.

Now we are ready to establish the main result of this section.
eorem 2.93 Let f be defined in (2.50), where the index set ; ¤ T Rp is compact, and
where the function g.t; / is convex on Rn for every t 2 T . Suppose in addition that the function
g.; x/ is upper semicontinuous on T for every x 2 Rn . en we have
[
N D co
@f .x/ @g t .x/:
N
t 2S.x/
N
Proof. Taking any w 2 @f .x/

N , we deduce from eorem 2.87 and eorem 2.92 that
N v/ C .v/ for all v 2 Rn ;

hw; vi f 0 .x;
where C is defined in Proposition 2.91. us w 2 @C .0/ D C by formula (2.44) for the set
˝ D C , which is a nonempty, closed, convex subset of Rn . e opposite inclusion follows from
Proposition 2.88.
Exercise 2.1 (i) Let ˝ Rn be a nonempty, closed, convex cone and let xN … ˝ . Show that
there exists a nonzero element v 2 Rn such that hv; xi 0 for all x 2 ˝ and hv; xi
N > 0.
(ii) Let ˝ be a subspace of R and let xN … ˝ . Show that there exists a nonzero element v 2 Rn
n
such that hv; xi D 0 for all x 2 ˝ and hv; xi

N > 0.
Exercise 2.2 Let ˝1 and ˝2 be two nonempty, convex subsets of Rn satisfying the condition
ri ˝1 \ ri ˝2 D ;. Show that there is a nonzero element v 2 Rn which separates the sets ˝1 and
˝2 in the sense of (2.2).
Exercise 2.3 Two nonempty, convex subsets of Rn are called properly separated if there exists a
nonzero element v 2 Rn such that
˚ ˇ ˚ ˇ
sup hv; xi ˇ x 2 ˝1 inf hv; yi ˇ y 2 ˝2 ;
˚ ˇ ˚ ˇ
inf hv; xi ˇ x 2 ˝1 < sup hv; yi ˇ y 2 ˝2 :
Prove that ˝1 and ˝2 are properly separated if and only if ri ˝1 \ ri ˝2 D ;.
Exercise 2.4 Let f W Rn ! R be a convex function attaining its minimum at some point xN 2
dom f . Show that the sets f˝1 ; ˝2 g with ˝1 WD epi f and ˝2 WD Rn ff .x/g
N form an extremal
system.
n
Exercise 2.5 Let ˝i for i D 1; 2 be nonempty, convex subsets of R with at least one point in
common and satisfy
int ˝1 ¤ ; and int ˝1 \ ˝2 D ;:
Show that there are elements v1 ; v2 in Rn , not equal to zero simultaneously, and real numbers
˛1 ; ˛2 such that
hvi ; xi ˛i for all x 2 ˝i ; i D 1; 2;
v1 C v2 D 0; and ˛1 C ˛2 D 0:
ˇ
Exercise 2.6 Let ˝ WD fx 2 Rn ˇ x 0g. Show that
˚ ˇ
N ˝/ D v 2 Rn ˇ v 0; hv; xi
N.xI N D 0 for any xN 2 ˝:
Exercise 2.7 Let f W Rn ! R be a convex function differentiable at xN . Show that

˚ ˇ
N .x; N epi f D rf .x/;
N f .x//I N 1 ˇ0 :
Exercise 2.8 Use induction to prove Corollary 2.19.
Exercise 2.9 Calculate the subdifferentials and the singular subdifferentials of the following
convex functions at every point of their domains:
(i) f .x/ D jx 1j C jx 2j; x 2 R.
(ii) f .x/ D e(jxj ; x 2 R.
x 2 1 if jxj 1;
(iii) f .x/ D
1 otherwise on R :
( p
1 x 2 if jxj 1;
(iv) f .x/ D
1 otherwise on R :
2
(v) f .x1 ; x2 / D jx1 j C˚ jx2 j; .x 1 ; x2 / 2 R .
(vi) f .x1 ; x2 / D max
˚ jx1 j; jx2 j ; .x1 ; x2 / 2 R2 .
(vii) f .x/ D max ha; xi b; 0 ; x 2 Rn , where a 2 Rn and b 2 R.
Exercise 2.10 Let f W R ! R be a nondecreasing convex function and let xN 2 R. Show that
v 0 for v 2 @f .x/
N .
Exercise 2.11 Use induction to prove Corollary 2.46.
Exercise 2.12 Let f W Rn ! R be a convex function. Give an alternative proof for the part of
Proposition 2.47 stating that @f .x/
N ¤ ; for any element xN 2 int.dom f /. Does the conclusion
still holds if xN 2 ri.dom f /?
Exercise 2.13 Let ˝1 and ˝2 be nonempty, convex subsets of Rn . Suppose that ri ˝1 \ ri ˝2 ¤
;. Prove that
N.xI
N ˝1 \ ˝2 / D N.xI N ˝2 / for all xN 2 ˝1 \ ˝2 :
N ˝1 / C N.xI
Exercise 2.14 Let f1 ; f2 W Rn ! R be convex functions satisfying the condition
ri .dom f1 / \ ri .dom f2 / ¤ ;:
Prove that for all xN 2 dom f1 \ dom f2 we have
@.f1 C f2 /.x/
N D @f1 .x/
N C @f2 .x/;
N @1 .f1 C f2 /.x/
N D @1 f1 .x/
N C @1 f2 .x/:
N
Exercise 2.15 In the setting of eorem 2.51 show that
N..x;
N y;
N z/I
N ˝1 / \ Œ N..x;
N y;
N zN /I ˝2 /
under the qualification condition (2.33).
Exercise 2.16 Calculate the singular subdifferential of the composition f ı B in the setting of
eorem 2.51.
Exercise 2.17 Consider the composition g ı h in the setting of Proposition 1.54 and take any
xN 2 dom .g ı h/. Prove that
nX
p ˇ o
ˇ
@.g ı h/.x/
N D i vi ˇ .1 ; : : : ; p / 2 @g.y/;
N vi 2 @fi .x/;
N i D 1; : : : ; p
iD1
under the validity of the qualification condition
h p
X i
02 N .1 ; : : : ; p / 2 @1 g.y/
i @fi .x/; N H) i D 0 for i D 1; : : : ; p :
iD1
Exercise 2.18 Calculate the singular subdifferential of the optimal value functions in the setting
of eorem 2.61.
Exercise 2.19 Calculate the support functions for the following sets:
˚ ˇ P
(i) ˝1 WD ˚x D .x1 ; : : : ; xn / 2 Rn ˇˇkxk1 WD niD1 jxi j 1 :
(ii) ˝2 WD x D .x1 ; : : : ; xn / 2 Rn ˇ kxk1 WD maxniD1 jxi j 1 :
Exercise 2.20 Find the Fenchel conjugate for each of the following functions on R:
(i) f .x/ D ex .
ln.x/ x > 0;
(ii) f .x/ D
1 x 0:
2
(iii) f .x/ D ax C bx C c , where a > 0.
(iv) f .x/ D ıIB .x/, where IB D Œ 1; 1.
(v) f .x/ D jxj.
Exercise 2.21 Let ˝ be a nonempty subset of Rn . Find the biconjugate function ı˝

.
Exercise 2.22 Prove the following statements for the listed functions on Rn :
v
(i) .f / .v/ D f , where > 0.

(ii) .f C / .v/ D f .v/ , where 2 R.
(iii) fu .v/ D f .v/ C hu; vi, where u 2 Rn and fu .x/ WD f .x u/.
Exercise 2.23 Let f W Rn ! R be a convex and positively homogeneous function. Show that
f .v/ D ı˝ .v/ for all v 2 Rn , where ˝ WD @f .0/.
Exercise 2.24 Let f W Rn ! R be a function with dom f D ;. Show that f .x/ D f .x/ for
all x 2 Rn .
Exercise 2.25 Let f W R ! R be convex. Prove that @f .x/ D Œf 0 .x/; fC0 .x/, where f 0 .x/
and fC0 .x/ stand for the left and right derivative of f at x , respectively.
Exercise 2.26 Define the function f W R ! R by

( p
1 x2 if jxj 1;
f .x/ WD
1 otherwise:
Show that f is a convex function and calculate its directional derivatives f 0 . 1I d / and f 0 .1I d /
in the direction d D 1.
Exercise 2.27 Let f W Rn ! R be convex and let xN 2 dom f . For an element d 2 Rn , consider
the function ' defined in Lemma 2.80 and show that ' is nondecreasing on . 1; 0/.
Exercise 2.28 Let f W Rn ! R be convex and let xN 2 dom f . Prove the following:
(i) f 0 .xI N d / for all d 2 Rn .
N d / f 0 .xI
(ii) f .x/
N f .xN d / f 0 .xIN d / f 0 .xI
N d / f .xN C d / N , d 2 Rn .
f .x/
Exercise 2.29 Prove the assertions of Proposition 2.67.

Exercise 2.30 Let ˝ Rn be a nonempty, compact set. Using eorem 2.93, calculate the
subdifferential of the support function ˝ .0/.
Exercise 2.31 Let ˝ be a nonempty, compact subset of Rn . Define the function

˚ ˇ
˝ .x/ WD sup kx !k ˇ ! 2 ˝g; x 2 Rn :
Prove that ˝ is a convex function and find its subdifferential @˝ .x/ at any x 2 Rn .
Exercise 2.32 Using the representation

˚ ˇ
kxk D sup hv; xi ˇ kvk 1
and eorem 2.93, calculate the subdifferential @p.x/

N of the norm function p.x/ WD kxk at any
point xN 2 Rn .
Exercise 2.33 Let d.xI ˝/ be the distance function (1.5) associated with some nonempty,
closed, convex subset ˝ of Rn .
(i) Prove the representation
˚ ˇ
d.xI ˝/ D sup hx; vi ˝ .v/ ˇ kvk 1
via the support function of ˝ .

N ˝/ at any point xN 2 Rn .
(ii) Calculate the subdifferential @d.xI
95
CHAPTER 3
Remarkable Consequences of
Convexity
is chapter is devoted to remarkable topics of convex analysis that, together with subdifferential
calculus and related results of Chapter 2, constitute the major achievements of this discipline most
important for many applications—first of all to convex optimization.
3.1 CHARACTERIZATIONS OF DIFFERENTIABILITY

In this section we define another classical notion of differentiability, the Gâteaux derivative, and
show that for convex functions it can be characterized, together with the Fréchet one, via a single-
element subdifferential of convex analysis.
Definition 3.1 Let f W Rn ! R be finite around xN . We say that f is G 

at xN if there is v 2 Rn such that for any d 2 Rn we have
f .xN C td / f .x/
N thv; d i
lim D 0:
t !0 t
en we call v the G  of f at xN and denote by fG0 .x/
N .
It follows from the definition and the equality f 0 .xI N d / that fG0 .x/
N d / D f 0 .xI N D v if
0
and only if the directional derivative f .xIN d / exists and is linear with respect to d with f 0 .xI
N d/ D
hv; d i for all d 2 Rn ; see Exercise 3.1 and its solution. Moreover, if f is Gâteaux differentiable
at xN , then
f .xN C td / f .x/ N
hfG0 .x/;
N d i D lim for all d 2 Rn :
t !0 t
Proposition 3.2 Let f W Rn ! R be a function (not necessarily convex) which is locally Lipschitz
continuous around xN 2 dom f . en the following assertions are equivalent:
(i) f is Fréchet differentiable at xN .
(ii) f is Gâteaux differentiable at xN .
96 3. REMARKABLE CONSEQUENCES OF CONVEXITY
N . For a fixed 0 ¤ d 2 Rn , taking
Proof. Let f be Fréchet differentiable at xN and let v WD rf .x/
into account that kxN C td xk
N D jtj kd k ! 0 as t ! 0, we have
f .xN C td / f .x/
N thv; d i f .xN C t d / f .x/
N hv; t d i
lim D kd k lim D 0;
t!0 t t !0 tkd k
which implies the GaO teaux differentiablity of f at xN with fG0 .x/ N . Note that the local
N D rf .x/
Lipschitz continuity of f around xN is not required in this implication.
Assuming now that f is GaO teaux differentiable at xN with v WD fG0 .x/
N , let us show that
f .xN C h/ f .x/
N hv; hi
lim D 0:
h!0 khk
Suppose by contradiction that it does not hold. en we can find 0 > 0 and a sequence of hk ! 0
with hk ¤ 0 as k 2 N such that
ˇ
ˇf .xN C hk / f .x/N hv; hk ij
0 : (3.1)
khk k
hk
Denote tk WD khk k and dk WD . en tk ! 0, and we can assume without loss of generality
khk k
that dk ! d with kd k D 1. is gives us the estimate
ˇ
ˇf .xN C tk dk / f .x/
N hv; tk dk ij
0 kdk k :
tk
Let ` 0 be a Lipschitz constant of f around xN . e triangle inequality ensures that
ˇ
ˇf .xN C tk dk / f .x/
N hv; tk dk ij
0 kdk k
ˇ tk ˇ
ˇf .xN C tk dk / f .xN C tk d /j C jhv; tk d i hv; tk dk ij ˇf .xN C tk d / f .x/
N hv; tk d ij
C
tk ˇ tk
ˇf .xN C tk d / f .x/
N hv; tk d ij
`kdk d k C kvk kdk d k C ! 0 as k ! 1:
tk
is contradicts (3.1) since 0 kdk k ! 0 , so f is Fréchet differentiable at xN .
Now we are ready to derive a nice subdifferential characterization of differentiability.
eorem 3.3 Let f W Rn ! R be a convex function with xN 2 int(domf /. e following asser-

tions are equivalent:
(i) f is Fréchet differentiable at xN .
(ii) f is Gâteaux differentiable at xN .
N is a singleton.
(iii) @f .x/
3.2. CARATHÉODORY THEOREM AND FARKAS LEMMA 97
Proof. It follows from eorem 2.29 that f is locally Lipschitz continuous around xN , and
hence (i) and (ii) are equivalent by Proposition 3.2. Suppose now that (ii) is satisfied and let
v WD fG0 .x/
N . en f 0 .xI N d / D hv; d i for all d 2 Rn . Observe that @f .x/
N d / D f 0 .xI N ¤ ; since
xN 2 int(domf /. Fix any w 2 @f .x/
N . It follows from eorem 2.84 that
f 0 .xI N d / whenever d 2 Rn ;
N d / hw; d i f 0 .xI
and thus hw; d i D hv; d i, so hw v; d i D 0 for all d . Plugging there d D w v gives us w D v

and thus justifies that @f .x/
N is a singleton.
Suppose finally that (iii) is satisfied with @f .x/
N D fvg and then show that (ii) holds. By
eorem 2.87 we have the representation
f 0 .xI
N d/ D sup hw; d i D hv; d i;
w2@f .x/
N
for all d 2 Rn . is implies that f is GaO teaux and hence Fréchet differentiable at xN .
3.2 CARATHÉODORY THEOREM AND FARKAS LEMMA

We begin with preliminary results on conic and convex hulls of sets that lead us to one of the most
remarkable results of convex geometry established by Constantin Carathéodory in 1911. Recall
that a set ˝ is called a cone if x 2 ˝ for all 0 and x 2 ˝ . It follows from the definition that
0 2 ˝ if ˝ is a nonempty cone. Given a nonempty set ˝ Rn , we say that K˝ is the convex
cone generated by ˝ , or the convex conic hull of ˝ , if it is the intersection of all the convex cones
containing ˝ .
Figure 3.1: Convex cone generated by a set.

Proposition 3.4 Let ˝ be a nonempty subset of Rn . en the convex cone generated by ˝ admits the
following representation:
nX
m ˇ o
ˇ
K˝ D i ai ˇ i 0; ai 2 ˝; m 2 N : (3.2)
iD1
If, in particular, the set ˝ is convex, then K˝ D RC ˝ .
Proof. Denote by C the set on the right-hand side of (3.2) and observe that it is a convex cone
which contains ˝ . It remains to show that C K for any convex cone K containing ˝ . To
P
see it, fix such a cone and form the combination x D m iD1 i ai with any i 0, ai 2 ˝ , and
m 2 N . If i D 0 for all i , then x D 0 2 K . Otherwise, suppose that i > 0 for some i and denote
P
WD m iD1 i > 0. is tells us that
X
m
i
xD ai 2 K;

iD1
which yields C K , so K˝ D C . e last assertion of the proposition is obvious.
Proposition 3.5 For any nonempty set ˝ Rn and any x 2 K˝ n f0g, we have
m
X
xD i ai with i > 0; ai 2 ˝ as i D 1; : : : ; m and m n:
iD1
P
Proof. Let x 2 K˝ n f0g. It follows from Proposition 3.4 that x D m iD1 i ai with some i > 0
and ai 2 ˝ as for i D 1; : : : ; m and m 2 N . If the elements a1 ; : : : ; am are linearly dependent,
then there are numbers i 2 R, not all zeros, such that
m
X
i ai D 0:
iD1
We can assume that I WD fi D 1; : : : ; m j i > 0g ¤ ; and get for any > 0 that
m
X m
X X
m Xm
xD i ai D i a i i ai D i i ai :
iD1 iD1 iD1 iD1
n ˇ o i
i ˇ
Let WD min ˇ i 2 I D 0 , where i0 2 I , and denote ˇi WD i i for i D 1; : : : ; m.
i i0
en we have the relationships
m
X
xD ˇi ai with ˇi0 D 0 and ˇi 0 for any i D 1; : : : ; m:
iD1
Continuing this procedure, after a finite number of steps we represent x as a positive linear com-
bination of linearly independent elements faj j j 2 J g, where J f1; : : : ; mg. us the index
set J cannot contain more than n elements.
e following remarkable result is known as the Carathéodory theorem. It relates convex
hulls of sets with the dimension of the space; see an illustration in Figure 3.2.
Figure 3.2: Carathéodory theorem.
eorem 3.6 Let ˝ be a nonempty subset of Rn . en every x 2 co ˝ can be represented as a

convex combination of no more than n C 1 elements of the set ˝ .
Proof. Denoting B WD f1g ˝ RnC1 , it is easy to observe that co B D f1g co ˝ and that
co B KB . Take now any x 2 co ˝ . en .1; x/ 2 co B and by Proposition 3.5 find i 0 and
.1; ai / 2 B for i D 0; : : : ; m with m n such that
m
X
.1; x/ D i .1; ai /:
iD0
Pm Pm
is implies that x D iD0 i ai with iD0 i D 1, i 0, and m n.
e next useful assertion is a direct consequence of eorem 3.6.
Corollary 3.7 Suppose that a subset ˝ of Rn is compact. en the set co ˝ is also compact.
Now we present two propositions, which lead us to another remarkable result known as the
Farkas lemma; see eorem 3.10.
Proposition 3.8 Let a1 ; : : : ; am be linearly dependent elements in Rn , where ai ¤ 0 for all i D
1; : : : ; m. Define the sets
˚ ˚
˝ WD a1 ; : : : ; am and ˝i WD ˝ n ai ; i D 1; : : : ; m:
en the convex cones generated by these sets are related as

m
[
K˝ D K˝ i : (3.3)
iD1
Proof. Obviously, [m
iD1 K˝i K˝ . To prove the converse inclusion, pick x 2 K˝ and get
m
X
xD i ai with i 0; i D 1; : : : ; m:
iD1
If x D 0, then x 2 [m iD1 K˝i obviously. Otherwise, the linear dependence of ai allows us to find
i 2 R not all zeros such that
m
X
i ai D 0:
iD1
Following the proof of Proposition 3.5, we represent x as

m
X
xD i ai with some i0 2 f1; : : : ; mg:
iD1;i ¤i0
is shows that x 2 [m

iD1 K˝i and thus verifies that equality (3.3) holds.
Proposition 3.9 Let a1 ; : : : ; am be elements of Rn . en the convex cone K˝ generated by the set
˝ WD fa1 ; : : : ; am g is closed in Rn .
Proof. Assume first that the elements a1 ; : : : ; am are linearly independent and take any sequence
fxk g K˝ converging to x . By the construction of K˝ , find numbers ˛ki 0 for each i D
1; : : : ; m and k 2 N such that
Xm
xk D ˛ki ai :
iD1
m
Letting ˛k WD .˛k1 ; : : : ; ˛km / 2 R and arguing by contradiction, it is easy to check that the
sequence f˛k g is bounded. is allows us to suppose without loss of generality that ˛k !
.˛1 ; : : : ; ˛m / as k ! 1 with ˛i 0 for all i D 1; : : : ; m. us we get by passing to the limit
P
that xk ! x D m iD1 ˛i ai 2 K˝ , which verifies the closedness of K˝ .
Considering next arbitrary elements a1 ; : : : ; am , assume without loss of generality that ai ¤
0 for all i D 1; : : : ; m. By the above, it remains to examine the case where a1 ; : : : ; am are linearly
dependent. en by Proposition 3.8 we have the cone representation (3.3), where each set ˝i
contains m 1 elements. If at least one of the sets ˝i consists of linearly dependent elements, we
can represent the corresponding cone K˝i in a similar way via the sets of m 2 elements. is
gives us after a finite number of steps that
p
[
K˝ D KCj ;
j D1
where each set Cj ˝ contains only linear independent elements, and so the cones KCj are
closed. us the cone K˝ under consideration is closed as well.
Now we are ready to derive the fundamental result of convex analysis and optimization
established by Gyula Farkas in 1894.
eorem 3.10 Let ˝ WD fa1 ; : : : ; am g with a1 ; : : : ; am 2 Rn and let b 2 Rn . e following as-

sertions are equivalent:
(i) b 2 K˝ .
(ii) For any x 2 Rn , we have the implication

hai ; xi 0; i D 1; : : : ; m H) hb; xi 0 :
Proof. To verify .i/ H) .ii/, take b 2 K˝ and find i 0 for i D 1; : : : ; m such that
m
X
bD i ai :
iD1
en, given x 2 Rn , the inequalities hai ; xi 0 for all i D 1; : : : ; m imply that

m
X
hb; xi D i hai ; xi 0;
iD1
which is (ii). To prove the converse implication, suppose that b … K˝ and show that (ii) does
not hold in this case. Indeed, since the set K˝ is closed and convex, the strict separation result of
Proposition 2.1 yields the existence of xN 2 Rn satisfying
˚ ˇ
N ˇ u 2 K˝ g < hb; xi:
sup hu; xi N
Since K˝ is a cone, we have 0 D h0; xi N and t u 2 K˝ for all t > 0, u 2 K˝ . us

N < hb; xi
˚ ˇ
N ˇ u 2 K˝ < hb; xi
t sup hu; xi N whenever t > 0:
Dividing both sides of this inequality by t and letting t ! 1 give us
˚ ˇ
sup hu; xiN ˇ u 2 K˝ 0:
It follows that hai ; xi

N 0 for all i D 1; : : : ; m while hb; xi
N > 0. is is a contradiction, which
therefore completes the proof of the theorem.
3.3 RADON THEOREM AND HELLY THEOREM

In this section we present a celebrated theorem of convex geometry discovered by Edward Helly
in 1913. Let us start with the following simple lemma.
Lemma 3.11 Consider arbitrary elements w1 ; : : : ; wm in Rn with m n C 2. en these elements

are affinely dependent.
Proof. Form the elements w2 w1 ; : : : ; wm w1 in Rn and observe that these elements are lin-
early dependent since m 1 > n. us the elements w1 ; : : : ; wm are affinely dependent by Propo-
sition 1.63.
We continue with a theorem by Johann Radon established in 1923 and used by himself to
give a simple proof of the Helly theorem.
eorem 3.12 Consider the elements w1 ; : : : ; wm of Rn with m n C 2 and let I WD

f1; : : : ; mg. en there exist two nonempty, disjoint subsets I1 and I2 of I with I D I1 [ I2
such that for ˝1 WD fwi j i 2 I1 g, ˝2 WD fwi j i 2 I2 g one has
co ˝1 \ co ˝2 ¤ ;:
Proof. Since m n C 2, it follows from Lemma 3.11 that the elements w1 ; : : : ; wm are affinely
dependent. us there are real numbers 1 ; : : : ; m , not all zeros, such that
m
X m
X
i wi D 0; i D 0:
iD1 iD1
Consider the index sets I1 WD fi D 1; : : : ; m j i 0g and I2 WD fi D 1; : : : ; m j i < 0g. Since

Pm P P
D 0, both sets I1 and I2 are nonempty and i2I1 i D i2I2 i . Denoting WD
PiD1 i

i2I1 i , we have the equalities
X X X i X i
i wi D i wi and wi D wi :

i2I1 i2I2 i2I1 i2I2
3.4. TANGENTS TO CONVEX SETS 103
Define ˝1 WD fwi j i 2 I1 g, ˝2 WD fwi j i 2 I2 g and observe that co ˝1 \ co ˝2 ¤ ; since
X i X i
wi D wi 2 co ˝1 \ co ˝2 :

i2I1 i2I2
is completes the proof of the Radon theorem.

Now we are ready to formulate and prove the fundamental Helly theorem.
˚
eorem 3.13 Let O WD ˝1 ; : : : ; ˝m be a collection of convex sets in Rn with m n C 1.
Suppose that the intersection of any subcollection of n C 1 sets from O is nonempty. en we
T
have miD1 ˝i ¤ ;.
Proof. Let us prove the theorem by using induction with respect to the number of elements in
O. e conclusion of the theorem is obvious if m D n C 1. Suppose
˚ that the conclusion
holds for
any collection of m convex sets of Rn with m n C 1. Let ˝1 ; : : : ; ˝m ; ˝mC1 be a collection
of convex sets in Rn such that the intersection of any subcollection of n C 1 sets is nonempty. For
i D 1; : : : ; m C 1, define
mC1
\
i WD ˝j
j D1;j ¤i
and observe that i ˝j whenever j ¤ i and i; j D 1; : : : ; m C 1. It follows from the in-

duction hypothesis that i ¤ ;, and so we can pick wi 2 i for every i D 1; : : : ; m C 1. Ap-
plying the above Radon theorem for the elements w1 ; : : : wmC1 , we find two nonempty, dis-
joint subsets index sets I1 and I2 of I WD f1; : : : ; m C 1g with I D I1 [ I2 such that for W1 WD
fwi j i 2 I1 g and W2 WD fwi j i 2 I2 g one has co W1 \ co W2 ¤ ;. is allows us to pick an ele-
mC1
ment w 2 co W1 \ co W2 . We finally show that w 2 \iD1 ˝i and thus arrive at the conclusion.
Since i ¤ j for i 2 I1 and j 2 I2 , it follows that i ˝j whenever i 2 I1 and j 2 I2 . Fix
any i D 1; : : : ; m C 1 and consider the case where i 2 I1 . en wj 2 j ˝i for every j 2 I2 .
Hence w 2 co W2 D co fwj j j 2 I2 g ˝i due the convexity of ˝i . It shows that w 2 ˝i for
every i 2 I1 . Similarly, we can verify that w 2 ˝i for every i 2 I2 and thus complete the proof
of the theorem.
3.4 TANGENTS TO CONVEX SETS

In Chapter 2 we studied the normal cone to convex sets and learned that the very definition and
major properties of it were related to convex separation. In this section we define the tangent
cone to convex sets independently while observing that it can be constructed via duality from the
normal cone to the set under consideration.
We start with cones. Given a nonempty, convex cone C Rn , the polar of C is
˚ ˇ
C ı WD v 2 Rn ˇ hv; ci 0 for all c 2 C :
Proposition 3.14 For any nonempty, convex cone C Rn , we have
.C ı /ı D cl C:
Proof. e inclusion C .C ı /ı follows directly from the definition, and hence cl C .C ı /ı

since .C ı /ı is a closed set. To verify the converse inclusion, take any c 2 .C ı /ı and suppose by
contradiction that c … cl C . Since C is a cone, by convex separation from Proposition 2.1, find
v ¤ 0 such that
hv; xi 0 for all x 2 C and hv; ci > 0;
which yields v 2 C ı thus contradicts to the fact that c 2 .C ı /ı .
Definition 3.15 Let ˝ be a nonempty, convex set. e   to ˝ at xN 2 ˝ is
N ˝/ WD cl Kf˝
T .xI xg
N D cl RC .˝ x/:
N
Figure 3.3: Tangent cone and normal cone.
Now we get the duality correspondence between normals and tangents to convex sets; see a
demonstration in Figure 3.3.
eorem 3.16 Let ˝ be a convex set and let xN 2 ˝ . en we have the equalities
ı ı
N.xI
N ˝/ D T .xI N ˝/ and T .xI N ˝/ D N.xIN ˝/ : (3.4)
3.4. TANGENTS TO CONVEX SETS 105
ı
Proof. To verify first the inclusion N.xI N ˝/ , pick any v 2 N.xI
N ˝/ ŒT .xI N ˝/ and get by def-
inition that hv; x xi
N 0 for all x 2 ˝ . is implies that
˝ ˛
v; t.x x/N 0 whenever t 0 and x 2 ˝:
Fix w 2 T .xI
N ˝/ and find tk 0 and xk 2 ˝ with tk .xk N ! w as k ! 1. en
x/
ı
hv; wi D lim hv; tk .xk N 0 and thus v 2 T .xI
x/i N ˝/ :
k!1
N ˝/ı and deduce from the relations

To verify the opposite inclusion in (3.4), fix any v 2 ŒT .xI
hv; wi 0 as w 2 T .xI
N ˝/ and x xN 2 T .xI N ˝/ as x 2 ˝ that
hv; x N 0 for all x 2 ˝:

xi
is yields w 2 N.xI N ˝/ and thus completes the proof of the first equality in (3.4). e second
equality in (3.4) follows from the first one due to
ı ıı
N.xI
N ˝/ D T .xI N ˝/ D cl T .xI N ˝/
by Proposition 3.14 and the closedness of T .xI

N ˝/ by Definition 3.15.
e next result, partly based on the Farkas lemma from eorem 3.10, provides effective
representations of the tangent and normal cones to linear inequality systems.
eorem 3.17 Let ai 2 Rn , let bi 2 R for i D 1; : : : ; m, and let

˚ ˇ
˝ WD x 2 Rn ˇ hai ; xi bi :
Suppose that xN 2 ˝ and define the active index set I.x/ N D bi g. en
N WD fi D 1; : : : ; m j hai ; xi
we have the relationships
˚ ˇ
N ˝/ D v 2 Rn ˇ hai ; vi 0 for all i 2 I.x/
T .xI N ; (3.5)
˚ ˇ
N ˝/ D cone ai ˇ i 2 I.x/
N.xI N : (3.6)
Here we use the convention that cone ; WD f0g.
Proof. To justify (3.5), consider the convex cone

˚ ˇ
K WD v 2 Rn ˇ hai ; vi 0 for all i 2 I.x/
N :
It is easy to see that for any x 2 ˝ , any i 2 I.x/

N , and any t 0 we have
˝ ˛
ai ; t.x x/
N D t hai ; xi hai ; xiN t.bi bi / D 0;
and hence RC .˝ N K . us it follows from the closedness of K that
x/
N ˝/ D cl RC .˝
T .xI x/
N K:
To verify the opposite inclusion in (3.5), observe that hai ; xi

N < bi for i … I.x/
N . Using this and
picking an arbitrary element v 2 K give us
hai ; xN C tvi bi for all i … I.x/

N and small t > 0:
Since hai ; vi 0 for all i 2 I.x/

N , it tells us that hai ; xN C t vi bi as i D 1; : : : ; m yielding
1
xN C tv 2 ˝; or v 2 ˝ xN T .xI
N ˝/:
t
is justifies the tangent cone representation (3.5).
Next we prove (3.6) for the normal cone. e first equality in (3.4) ensures that
N ˝/ if and only if hw; vi 0 for all v 2 T .xI

w 2 N.xI N ˝/:
en we see from (3.5) that w 2 N.xIN ˝/ if and only if

hai ; vi 0 for all i 2 I.x/
N H) hw; vi 0
whenever v 2 Rn . e Farkas lemma from eorem 3.10 tells us that w 2 N.xI

N ˝/ if and only if
w 2 conefai j i 2 I.x/g
N , and thus (3.6) holds.
3.5 MEAN VALUE THEOREMS

e mean value theorem for differentiable functions is one of the central results of classical real
analysis. Now we derive its subdifferential counterparts for general convex functions.
eorem 3.18 Let f W R ! R be convex, let a < b , and let the interval Œa; b belong to the
domain of f , which is an open interval in R. en there is c 2 .a; b/ such that
f .b/ f .a/
2 @f .c/: (3.7)
b a
Proof. Define the real-valued function gW Œa; b ! R by

h f .b/ f .a/ i
g.x/ WD f .x/ .x a/ C f .a/
b a
for which g.a/ D g.b/. is implies, by the convexity and hence continuity of g on its open
domain, that g has a local minimum at some c 2 .a; b/. e function
h f .b/ f .a/ i
h.x/ WD .x a/ C f .a/
b a
3.5. MEAN VALUE THEOREMS 107
Figure 3.4: Subdifferential mean value theorem.
is obviously differentiable at c , and hence

˚ n f .b/ f .a/ o
@h.c/ D h0 .c/ D :
b a
e subdifferential Fermat rule and the subdifferential sum rule from Chapter 2 yield
n f .b/ f .a/ o
0 2 @g.c/ D @f .c/ ;
b a
which implies (3.7) and completes the proof of the theorem.
To proceed further with deriving a counterpart of the mean value theorem for convex func-
tions on Rn (eorem 3.20), we first present the following lemma.
Lemma 3.19 Let f W Rn ! R be a convex function and let D be the domain of f , which is an open
subset of Rn . Given a; b 2 D with a ¤ b , define the function 'W R ! R by

'.t / WD f tb C .1 t /a ; t 2 R: (3.8)
en for any number t0 2 .0; 1/ we have

˚ ˇ
@'.t0 / D hv; b ai ˇ v 2 @f .c0 / with c0 WD t0 a C .1 t0 /b:
Proof. Define the vector-valued functions AW R ! Rn and BW R ! Rn by
A.t/ WD .b a/t and B.t/ WD t b C .1 t/a D t .b a/ C a D A.t / C a; t 2 R:

It is clear that '.t/ D .f ı B/.t / and that the adjoint mapping to A is A v D hv; b ai, v 2 Rn .
By the subdifferential chain rule from eorem 2.51, we have
˚ ˇ
@'.t0 / D A @f .c0 / D hv; b ai ˇ v 2 @f .c0 / ;
which verifies the result claimed in the lemma.
eorem 3.20 Let f W Rn ! R be a convex function in the setting of Lemma 3.19. en there
exists an element c 2 .a; b/ such that
˝ ˛ ˚ ˇ
f .b/ f .a/ 2 @f .c/; b a WD hv; b ai ˇ v 2 @f .c/ :
Proof. Consider the function 'W R ! R from (3.8) for which '.0/ D f .a/ and '.1/ D f .b/.
Applying first eorem 3.18 and then Lemma 3.19 to ' , we find t0 2 .0; 1/ such that
˚ ˇ
f .b/ f .a/ D '.1/ '.0/ 2 @'.t0 / D hv; b ai ˇ v 2 @f .c/
with c D t0 a C .1 t0 /b . is justifies our statement.
3.6 HORIZON CONES

is section concerns behavior of unbounded convex sets at infinity.
Definition 3.21 Given a nonempty, closed, convex subset F of Rn and a point x 2 F , the 
 of F at x is defined by the formula
˚ ˇ
F1 .x/ WD d 2 Rn ˇ x C t d 2 F for all t > 0 :
Note that the horizon cone is also known in the literature as the asymptotic cone of F at the point
in question. Another equivalent definition of F1 .x/ is given by
\F x
F1 .x/ D ;
t>0
t
which clearly implies that F1 .x/ is a closed, convex cone in Rn . e next proposition shows that
F1 .x/ is the same for any x 2 F , and so it can be simply denoted by F1 .
Proposition 3.22 Let F be a nonempty, closed, convex subset of Rn . en we have the equality
F1 .x1 / D F1 .x2 / for any x1 ; x2 2 F .
3.6. HORIZON CONES 109
Figure 3.5: Horizon cone.
Proof. It suffices to verify that F1 .x1 / F1 .x2 / whenever x1 ; x2 2 F . Taking any direction
d 2 F1 .x1 / and any number t > 0, we show that x2 C t d 2 F . Consider the sequence
1 1

xk WD x1 C kt d C 1 x2 ; k 2 N :
k k
en xk 2 F for every k because d 2 F1 .x1 / and F is convex. We also have xk ! x2 C td , and
so x2 C td 2 F since F is closed. us d 2 F1 .x2 /.
Corollary 3.23 Let F be a closed, convex subset of Rn that contains the origin. en
\
F1 D tF:
t >0
Proof. It follows from the construction of F1 and Proposition 3.22 that

\F 0 \1 \
F1 D F1 .0/ D D F D tF;
t >0
t t>0
t t>0
which therefore justifies the claim.
Proposition 3.24 Let F be a nonempty, closed, convex subset of Rn . For a given element d 2 Rn ,
the following assertions are equivalent:
(i) d 2 F1 .
(ii) ere are sequences ftk g Œ0; 1/ and ffk g F such that tk ! 0 and tk fk ! d .
Proof. To verify (i)H)(ii), take d 2 F1 and fix xN 2 F . It follows from the definition that
xN C kd 2 F for all k 2 N :
For each k 2 N , this allows us to find fk 2 F such that

1 1
xN C kd D fk ; or equivalently, xN C d D fk :
k k
1
Letting tk WD , we see that tk fk ! d as k ! 1.
k
To prove (ii)H)(i), suppose that there are sequences ftk g Œ0; 1/ and ffk g F with
tk ! 0 and tk fk ! d . Fix x 2 F and verify that d 2 F1 by showing that
x C td 2 F for all t > 0:
Indeed, for any fixed t > 0 we have 0 t tk < 1 when k is sufficiently large. us
.1 t tk /x C t tk fk ! x C td as k ! 1:
It follows from the convexity of F that every element .1 t tk /x C t tk fk belongs to F . Hence

x C td 2 F by the closedness, and thus we get d 2 F1 .
Finally in this section, we present the expected characterization of the set boundedness in
terms of its horizon cone.
eorem 3.25 Let F be a nonempty, closed, convex subset of Rn . en the set F is bounded if
and only if its horizon cone is trivial, i.e., F1 D f0g.
Proof. Suppose F is bounded and take any d 2 F1 . By Proposition 3.24, there exist sequences
ftk g Œ0; 1/ with tk ! 0 and ffk g F such that tk fk ! d as k ! 1. It follows from the
boundedness of F that tk fk ! 0, which shows that d D 0.
To prove the converse implication, suppose by contradiction that F is unbounded while
F1 D f0g. en there is a sequence fxk g F with kxk k ! 1. is allows us to form the se-
quence of (unit) directions
xk
dk WD ; k 2N:
kxk k
It ensures without loss of generality that dk ! d as k ! 1 with kd k D 1. Fix any x 2 F and
observe that for all t > 0 and k 2 N sufficiently large we have

t t
uk WD 1 xC xk 2 F
kxk k kxk k
by the convexity of F . Furthermore, x C td 2 F since uk ! x C td and F is closed. us d 2
F1 , which is a contradiction that completes the proof of the theorem.
3.7. MINIMAL TIME FUNCTIONS AND MINKOWSKI GAUGE 111
3.7 MINIMAL TIME FUNCTIONS AND MINKOWSKI
GAUGE
e main object of this and next sections is a remarkable class of convex functions known as
the minimal time function. ey are relatively new in convex analysis and highly important in
applications. Some of their recent applications to location problems are given in Chapter 4. We
also discuss relationships between minimal time functions and Minkowski gauge functions well-
recognized in convex analysis and optimization.
Definition 3.26 Let F be a nonempty, closed, convex subset of Rn and let ˝ Rn be a nonempty
set, not necessarily convex. e    associated with the sets F and ˝ is defined
by ˚ ˇ
T˝F .x/ WD inf t 0 ˇ .x C tF / \ ˝ ¤ ; : (3.9)
Figure 3.6: Minimal time functions.
e name comes from the following interpretation: (3.9) signifies the minimal time needed for the
point x to reach the target set ˝ along the constant dynamics F ; see Figure 3.6. Note that when F D
IB , the closed unit ball in Rn , the minimal time function (3.9) reduces to the distance function
d.xI ˝/ considered in Section 2.5. In the case where F D f0g, the minimal time function (3.9) is
the indicator function of ˝ . If ˝ D f0g, we get T˝F .x/ D F . x/, where F is the Minkowski
gauge defined later in this section. e minimal time function (3.9) also covers another interesting
class of functions called the directional minimal time function, which corresponds to the case where
F is a nonzero vector.
In this section we study general properties of the minimal time function (3.9) used in what
follows. ese properties are certainly of independent interest.
Observe that T˝F .x/ D 0 if x 2 ˝ , and that T˝F .x/ is finite if there exists t 0 with
.x C tF / \ ˝ ¤ ;;
which holds, in particular, when 0 2 int F . If no such t exists, then T˝F .x/ D 1. us the min-
imal time function (3.9) is an extended-real-valued function in general.
e sets F and ˝ in (3.9) are not necessarily bounded. e next theorem reveals behavior
of T˝F .x/
in the case where ˝ is bounded, while F is not necessarily bounded. It is an exercise
to formulate and prove related results for the case where F is bounded, while ˝ is not necessarily
bounded.
eorem 3.27 Let F be a nonempty, closed, convex subset of Rn and let ˝ Rn be a nonempty,
closed set. Assume that ˝ is bounded. en
T˝F .x/ D 0 implies x 2 ˝ F1 :
e converse holds if we assume additionally that 0 2 F .
Proof. Suppose that T˝F .x/ D 0 and find tk ! 0, tk 0, such that
.x C tk F / \ ˝ ¤ ;; k 2N:
is gives us qk 2 ˝ and fk 2 F with x C tk fk D qk for each k . By the boundedness and

closedness of ˝ , we can select a subsequence of fqk g (no relabeling) that converges to some
q 2 ˝ , and thus tk fk ! q x . It follows from Proposition 3.24 that q x 2 F1 , which justi-
fies x 2 ˝ F1 .
To prove the opposite implication, pick any x 2 ˝ F1 and find q 2 ˝ and d 2 F1 such
that x D q d . Since 0 2 F and d 2 F1 , we get from Corollary 3.23 the inclusions k.q x/ D
kd 2 F for all k 2 N . is shows that
1 1
q x 2 F and so x C F \ ˝ ¤ ;; k 2 N :
k k
1
e latter ensures that 0 T˝F .x/ for all k 2 N , which holds only when T˝F .x/ D 0 and
k
thus completes the proof of the theorem.
e following example illustrates the result of eorem 3.27.
Example 3.28 Let F D R Œ 1; 1 R2 and let ˝ be the disk centered at .1; 0/ with radius
r D 1. Using the representation of F1 in Corollary 3.23, we get F1 D R f0g and ˝ F1 D
R Œ 1; 1. us T˝F .x/ D 0 if and only if x 2 R Œ 1; 1.
e next result characterizes the convexity of the minimal time function (3.9).
Proposition 3.29 Let F be a nonempty, closed, convex subset of Rn . If ˝ Rn is nonempty and

convex, then the minimal time function (3.9) is convex. If furthermore F is bounded and ˝ is closed,
then the converse holds as well.
Proof. Take any x1 ; x2 2 dom T˝F and let 2 .0; 1/. We now show that

T˝F x1 C .1 /x2 T˝F .x1 / C .1 /T˝F .x2 / (3.10)
provided that both sets F and ˝ are convex. Denote i WD T˝F .xi / for i D 1; 2 and, given any
> 0, select numbers ti so that
i ti < i C and .xi C ti F / \ ˝ ¤ ;; i D 1; 2:
It follows from the convexity of F and ˝ that

x1 C .1 /x2 C t1 C .1 /t2 F \ ˝ ¤ ;;
which implies in turn the estimates

T˝F x1 C .1 /x2 t1 C .1 /t2 T˝F .x1 / C .1 /T˝F .x2 / C :
Since is arbitrary, we arrive at (3.10) and thus verifies the convexity of T˝F .
Conversely, suppose that T˝F is convex provided that F is bounded and that ˝ is closed.
en the level set ˚ ˇ
x 2 Rn ˇ T˝F .x/ 0
is definitely convex. We can easily show set reduces to the target set ˝ , which justifies the con-
vexity of the latter.
Definition 3.30 Let F be a nonempty, closed, convex subset of Rn . e M  associated
with the set F is defined by
˚ ˇ
F .x/ WD inf t 0 ˇ x 2 tF ; x 2 Rn : (3.11)
e following theorem summarizes the main properties of Minkowski gauge functions.
eorem 3.31 Let F be a closed, convex set that contains the origin. e Minkowski gauge F
is an extended-real-valued function, which is positively homogeneous and subadditive. Further-
more, F .x/ D 0 if and only if x 2 F1 .
Proof. We first prove that F is subadditive meaning that
F .x1 C x2 / F .x1 / C F .x2 / for all x1 ; x2 2 Rn : (3.12)
is obviously holds if the right-hand side of (3.12) is equal to infinity, i.e., we have that
x1 … dom F or x2 … dom F . Suppose now that x1 ; x2 2 dom F and fix > 0. en there are
numbers t1 ; t2 0 such that
F .x1 / t1 < F .x1 / C ; F .x2 / t2 < F .x2 / C and x1 2 t1 F; x2 2 t2 F:

It follows from the convexity of F that x1 C x2 2 t1 F C t2 F .t1 C t2 /F , and so
F .x1 C x2 / t1 C t2 < F .x1 / C F .x2 / C 2;
which justifies (3.12) since > 0 was chosen arbitrarily. To check the positive homogeneous
property of F , we take any ˛ > 0 and conclude that
˚ ˇ n ˇ t o
ˇ
F .˛x/ D inf t 0 ˇ ˛x 2 tF D inf t 0 ˇ x 2 F
nt ˇ ˛
ˇ t o
D ˛ inf 0 ˇ x 2 F D ˛F .x/:
˛ ˛
It remains to verify the last statement of the theorem. Observe that F .x/ D T˝ F .x/ is a special
case of the minimal time function with ˝ D f0g. Since 0 2 F , eorem 3.27 tells us that F .x/ D
0 if and only if x 2 ˝ . F /1 D F1 , which completes the proof.
We say for simplicity that 'W Rn ! R is ` Lipschitz continuous on Rn if it is Lipschitz con-
tinuous on Rn with Lipschitz constant ` 0. e next result calculates a constant ` of Lipschitz
continuity of the Minkowski gauge on Rn .
Proposition 3.32 Let F be a closed, convex set that contains the origin as an interior point. en the
Minkowski gauge F is `-Lipschitz continuous on Rn with the constant
n1 ˇ o
ˇ
` WD inf ˇ IB.0I r/ F; r > 0 : (3.13)
r
In particular, we have F .x/ `kxk for all x 2 Rn .
Proof. e condition 0 2 int F ensures that ` < 1 for the constant ` from (3.13). Take any r > 0
satisfying IB.0I r/ F and get from the definitions that
˚ ˇ ˚ ˇ ˚ kxk
F .x/ D inf t 0 ˇ x 2 tF inf t 0 ˇ x 2 tIB.0I r/ D inf t 0 j kxk rt D ;
r
which implies that F .x/ `kxk. en the subadditivity of F gives us that
F .x/ F .y/ F .x y/ `kx yk for all x; y 2 Rn ;
and thus F is ` Lipschitz continuous on Rn with constant (3.13).

Next we relate the minimal time function to the Minkowski gauge.
eorem 3.33 Let F be a nonempty, closed, convex set. en for any ˝ ¤ ; we have
˚ ˇ
T˝F .x/ D inf F .! x/ ˇ ! 2 ˝ ; x 2 Rn : (3.14)
Proof. Consider first the case where T˝F .x/
D 1. In this case we have
˚ ˇ
t 0 ˇ .x C tF / \ ˝ ¤ ; D ;;
which implies that ft 0 j ! x 2 tF g D ; for every ! 2 ˝ . us F .! x/ D 1 as ! 2 ˝ ,

and hence the right-hand side of (3.14) is also infinity.
Considering now the case where T˝F .x/ < 1, fix any t 0 such that .x C tF / \ ˝ ¤ ;.
is gives us f 2 F and ! 2 ˝ with x C tf D ! . us ! x 2 tF , and so F .! x/ t ,
which yields inf!2˝ F .! x/ t . It follows from definition (3.9) that
inf F .! x/ T˝F .x/ < 1:

!2˝
To verify the opposite inequality in (3.14), denote WD inf!2˝ F .! x/ and for any > 0 find
! 2 ˝ with F .! x/ < C . By definition of the Minkowski gauge (3.11), we get t 0 such
that t < C and ! x 2 tF . is gives us ! 2 x C tF and therefore
T˝F .x/ t < C :
Since was chosen arbitrarily, we arrive at
T˝F .x/ D inf F .! x/;

!2˝
which implies the remaining inequality in (3.14) and completes the proof.
As a consequence of the obtained result and Proposition 3.32, we establish now the
` Lipschitz continuity of minimal time function with the constant ` calculated in (3.13).
Corollary 3.34 Let F be a closed, convex set that contains the origin as an interior point. en for
any nonempty set ˝ the function T˝F is ` Lipschitz with constant (3.13).
Proof. It follows from the subadditivity of F and the result of Proposition 3.32 that
F .! x/ F .! y/ C F .y x/ F .! y/ C `ky xk for all ! 2 ˝
whenever x; y 2 Rn . en eorem 3.33 tells us that
T˝F .x/ T˝F .y/ C `kx yk;
which implies the ` Lipschitz property of T˝F due to the arbitrary choice of x; y .
Define next the enlargement of the target set ˝ via minimal time function by
˚ ˇ
˝r WD x 2 Rn ˇ T˝F .x/ r ; r > 0: (3.15)
Proposition 3.35 Let F be a nonempty, closed, convex subset of Rn and let ˝ Rn be a nonempty
set. en for any nonempty set ˝ and any x … ˝r with T˝F .x/ < 1 we have
T˝F .x/ D T˝Fr .x/ C r:
Proof. It follows from ˝ ˝r that T˝Fr .x/ T˝F .x/ < 1. Take any t 0 with
.x C tF / \ ˝r ¤ ;
and find f1 2 F and u 2 ˝r such that x C tf1 D u. Since u 2 ˝r , we have T˝F .u/ r . For any
> 0, there is s 0 with s < r C and .u C sF / \ ˝ ¤ ;. is implies the existence of ! 2 ˝
and f2 2 F such that u C sf2 D ! . us
! D u C sf2 D .x C tf1 / C sf2 2 x C .t C s/F
since F is convex. is gives us
T˝F .x/ t C s t C r C and so T˝F .x/ t C r
by the arbitrary choice of > 0. is allows us to conclude that
T˝F .x/ T˝Fr C r:
To verify the opposite inequality, denote WD T˝F .x/ and observe that r < by x … ˝r . For any
> 0, find t 2 Œ0; 1/, f 2 F , and ! 2 ˝ with t < C and x C tf D ! . Hence
! D x C tf D x C .t r/f C rf 2 x C .t r/f C rF and so T˝F .x C .t r/f / r;
which shows that x C .t r/f 2 ˝r . We also see that x C .t r/f 2 x C .t r/F . us
T˝Fr .x/ t r r C , which verifies that
r C T˝Fr D T˝F .x/
since > 0 was chosen arbitrarily. is completes the proof.
Proposition 3.36 Let F be a nonempty, closed, convex subset of Rn . en for any nonempty set ˝ ,
any x 2 dom T˝F , any t 0, and any f 2 F we have
T˝F .x tf / T˝F .x/ C t:
Proof. Fix > 0 and select s 0 so that
T˝F .x/ s < T˝F .x/ C and .x C sF / \ ˝ ¤ ;:
en .x tf C tF C sF / \ ˝ ¤ ;, and hence .x tf C .t C s/F / \ ˝ ¤ ;. It shows that
T˝F .x tf / t C s t C T˝F .x/ C ;
which implies the conclusion of the proposition by letting ! 0C .
3.8. SUBGRADIENTS OF MINIMAL TIME FUNCTIONS 117
3.8 SUBGRADIENTS OF MINIMAL TIME FUNCTIONS
Observing that the minimal time function (3.9) is intrinsically nonsmooth, we study here its gen-
eralized differentiation properties in the case where the target set ˝ is convex. In this case (since
the constant dynamics F is always assumed convex) function (3.9) is convex as well by Proposi-
tion 3.29. e results below present precise formulas for calculating the subdifferential @T˝F .x/ N
depending on the position of xN with respect to the set ˝ F1 .
eorem 3.37 Let both sets 0 2 F Rn and ˝ Rn be nonempty, closed, convex. For any
xN 2 ˝ F1 , the subdifferential of T˝F at xN is calculated by
@T˝F .x/
N D N.xI
N ˝ F1 / \ C ; (3.16)
n
where C WD fv 2 R j F . v/ 1g is defined via (2.41). In particular, we have
@T˝F .x/ N ˝/ \ C for xN 2 ˝:
N D N.xI (3.17)
Proof. Fix any v 2 @T˝F .x/

N . Using (2.13) and the fact that T˝F .x/ D 0 for all x 2 ˝ F1 , we get
v 2 N.xIN ˝ F1 /. Write xN 2 ˝ F1 as xN D !N d with !N 2 ˝ and d 2 F1 . Fix any f 2 F
and define x t WD .xN C d / tf D !N tf as t > 0. en .x t C tF / \ ˝ ¤ ;, and so
D d E
N T˝F .x t / t; or equivalently v;
hv; x t xi f 1 for all t > 0:
t
By letting t ! 1, it implies that hv; f i 1, and hence we get v 2 C . is verifies the inclu-
sion @T˝F .x/
N N.xIN ˝ F1 / \ C .
To prove the opposite inclusion in (3.16), fix any v 2 N.xI N ˝ F1 / \ C and any u 2
dom T˝ . For any > 0, there exist t 2 Œ0; 1/, ! 2 ˝ , and f 2 F such that
F
T˝F .u/ t < T˝F .u/ C and u C tf D !:

Since ˝ ˝ F1 , we have the relationships
hv; u xi
N D hv; ! N C thv; f i t < T˝F .u/ C D T˝F .u/
xi T˝F .x/
N C ;
which show that v 2 @T˝F .x/ N ˝ F1 / \ C @T˝F .x/
N . In this way we get N.xI N and thus justify
the subdifferential formula (3.16) in the general case where xN 2 ˝ F1 .
It remains to verify the simplified representation (3.17) in the case where xN 2 ˝ . Since
˝ ˝ F1 , we obviously have the inclusion
N.xI
N ˝ F1 / \ C N.xI
N ˝/ \ C :
N ˝/ \ C and fix x 2 ˝ F1 . Taking ! 2 ˝
To show the converse inclusion, pick v 2 N.xI
and d 2 F1 with x D ! d gives us t d 2 F since 0 2 F , so hv; t d i 1 for all t > 0. us
hv; d i 0, which readily implies that
hv; x xi
N D hv; ! d xi
N D hv; ! xi
N C hv; d i 0:
In this way we arrive at v 2 N.xI
N ˝ F1 / \ C and thus complete the proof.
e next theorem provides the subdifferential calculation for minimal time functions in the
cases complemented to those in eorem 3.37, i.e., when xN … ˝ F1 and xN … ˝ .
eorem 3.38 Let both sets 0 2 F Rn and ˝ Rn be nonempty, closed, convex and let xN 2
dom T˝F for the minimal time function (3.9). Suppose that ˝ is bounded. en for xN … ˝ F1
we have
@T˝F .x/ N ˝r / \ S ;
N D N.xI (3.18)
where the enlargement ˝r is defined in (3.15), and where
˚ ˇ
S WD v 2 Rn ˇ F . v/ D 1 with r D T˝F .x/
N > 0:
Proof. Pick v 2 @T˝F .x/

N . Observe from the definition that
hv; x N T˝F .x/

xi T˝F .x/
N for all x 2 Rn : (3.19)
en T˝F .x/ r D T˝F .x/ N , and hence hv; x xi

N 0 for any x 2 ˝r . is yields v 2 N.xI
N ˝r /.
Using (3.19) for x WD xN f with f 2 F and Proposition 3.36 ensures that F . v/ 1. Fix
2 .0; r/ and select t 2 R, f 2 F , and ! 2 ˝ so that
r t < r C 2 and ! D xN C tf:
We can write ! D xN C f C .t /f and then see that T˝F .xN C f / t . Applying (3.19)
to x WD xN C f gives us the estimates
hv; f i T˝F .xN C f / T˝F .x/

N t r 2 :
which imply in turn that

1 h v; f i F . v/:
Letting now # 0 tells us that F . v/ 1, i.e., F . v/ D 1 in this case. us v 2 S , which
proves the inclusion @T˝F .x/
N N.xI N ˝r / \ S .
To verify the opposite inclusion in (3.18), take any v 2 N.xIN ˝r / with F . v/ D 1 and
show that (3.19) holds. It follows from eorem 3.37 that v 2 @T˝Fr .x/N , and so
hv; x N T˝Fr .x/ for all x 2 Rn :

xi
Fix x 2 Rn and deduce from Proposition 3.35 that T˝F .x/ r D T˝Fr .x/ in the case t WD
T˝F .x/ > r , which ensures (3.19). Suppose now that t r and for any > 0 choose f 2 F
such that hv; f i > 1 . en Proposition 3.36 tells us that T˝F .x .r t /f / r , and so
x .r t/f 2 ˝r . Since v 2 N.xI N ˝r /, one has hv; x .r t/f xi N 0, which yields

N hv; f i.r t/ .1 /.t r/ D .1 / T˝F .x/ T˝F .x/
hv; x xi N :
Since > 0 was arbitrarily chosen, we get (3.19), and hence verify that v 2 @T˝F .x/
N .
3.8. SUBGRADIENTS OF MINIMAL TIME FUNCTIONS 119
e next result addresses the same setting as in eorem 3.38 while presenting another
formula for calculating the subdifferential of the minimal time function that involves the subdif-
ferential of the Minkowski gauge and generalized projections to the target set.
eorem 3.39 Let both sets 0 2 F Rn and ˝ Rn be nonempty, closed, convex and let
xN 2 dom T˝F . Suppose that ˝ is bounded and that xN … ˝ F1 . en we have

@T˝F .x/
N D @F .!N x/ N \ N.!I
N ˝/ (3.20)
for any !N 2 ˘F .xI

N ˝/ via the generalized projector
˚ ˇ
˘F .xIN ˝/ WD ! 2 ˝ ˇ T˝F .x/N D F .! x/
N : (3.21)
Proof. Fixing any v 2 @T˝F .x/ N ˝/, we get T˝F .x/

N and !N 2 ˘F .xI N D F .!N N and
x/
hv; x N T˝F .x/

xi T˝F .x/
N for all x 2 Rn : (3.22)
It is easy to check the following relationships that hold for any ! 2 ˝ :
hv; ! !i
N D hv; .! !N C x/
N N T˝F .!
xi !N C x/
N T˝F .x/
N
F ! .! !N C x/
N F .!N x/
N D 0:
N ˝/. Denote now e

is shows that v 2 N.!I u WD !N xN and for any t 2 .0; 1/ and u 2 Rn get

hv; t.u u/i T˝F xN t .u e
e u/ T˝F .x/
N
F !N .xN t .u e u// F .!N x/
N
u C t.u e
D F .e u// F .e u/
D F tu C .1 t/e u F .eu/
tF .u/ C .1 t /F .e
u/ F .e u/
D t F .u/ F .e
u/
by applying (3.22) with x WD xN t .u e

u/. It follows that
h v; u e
ui F .u/ u/ for all u 2 Rn :
F .e
us v 2 @F .e u/, and so v 2 @F .!N x/ N , which proves the inclusion “” in (3.20).
To verify the opposite inclusion, write
˚ ˇ ˚ ˇ
F .x/ D inf t 0 ˇ x 2 tF D inf t 0 ˇ . x C tF / \ O ¤ ; D TOF . x/
with O WD f0g. Since xN … ˝ F1 , for any !N 2 ˝ we have xN !N … O F1 D F1 . Further-

more, it follows from v 2 @F .!N x/N by applying eorem 3.38 that v 2 S . Hence it remains
to justify the inclusion
@F .!N x/N \ N.!I N ˝/ N.xI
N ˝r / (3.23)
and then to apply eorem 3.38. To proceed with the proof of (3.23), pick any vector x 2 ˝r
and a positive number arbitrarily small. en there are t < r C , f 2 F , and ! 2 ˝ such that
! D x C tf , which yields
hv; x xi
N D hv; ! tf xi N
D th v; f i C hv; ! !i
N C hv; !N xiN
t C hv; ! !iN C hv; !N xi
N
T˝F .x/
N C C hv; ! !i N C hv; !N xi:
N
Observe that hv; ! N 0 by v 2 N.!I

!i N ˝/. Moreover,
hv; !N xi
N D h v; 0 .!N x/i
N F .0/ F .!N x/
N D T˝F .x/
N
since v 2 @F .!N x/ N . It follows that hv; x xi

N for all x 2 ˝r . Hence v 2 N.xI
N ˝r / be-
cause > 0 was chosen arbitrarily. is completes the proof of the theorem.
3.9 NASH EQUILIBRIUM

is section gives a brief introduction to noncooperative game theory and presents a simple proof
of the existence of Nash equilibrium as a consequence of convexity and Brouwer’s fixed point
theorem. For simplicity, we only consider two-person games while more general situations can be
treated similarly.
Definition 3.40 Let ˝1 and ˝2 be nonempty subsets of Rn and Rp , respectively. A -

  in the case of two players I and II consists of two strategy sets ˝i and two real-valued
functions ui W ˝1 ˝2 ! R for i D 1; 2 called the  . en we refer to this game
as f˝i ; ui g for i D 1; 2.
e key notion of Nash equilibrium was introduced by John Forbes Nash, Jr., in 1950 who
proved the existence of such an equilibrium and was awarded the Nobel Memorial Prize in Eco-
nomics in 1994.
Definition 3.41 Given a noncooperative two-person game f˝i ; ui g, i D 1; 2, a N -

 is an element .xN 1 ; xN 2 / 2 ˝1 ˝2 satisfying the conditions
u1 .x1 ; xN 2 / u1 .xN 1 ; xN 2 / for all x1 2 ˝1 ;

u2 .xN 1 ; x2 / u2 .xN 1 ; xN 2 / for all x2 2 ˝2 :
e conditions above mean that .xN 1 ; xN 2 / is a pair of best strategies that both players agree
with in the sense that xN 1 is a best response of Player I when Player II chooses the strategy xN 2 , and
xN 2 is a best response of Player II when Player I chooses the strategy xN 1 . Let us now consider two
examples to illustrate this concept.
3.9. NASH EQUILIBRIUM 121
Example 3.42 In a two-person game suppose that each player can only choose either strategy
A or B . If both players choose strategy A, they both get 4 points. If Player I chooses strategy A
and Player II chooses strategy B , then Player I gets 1 point and Player II gets 3 points. If Player I
chooses strategy B and Player II chooses strategy A, then Player I gets 3 points and Player II gets
1 point. If both choose strategy B , each player gets 2 points. e payoff function of each player
is represented in the payoff matrix below.
Player II
A B
Player I
A 4; . 4 1; . 3
.
B 3; . 1 2; . 2
In this example, Nash equilibrium occurs when both players choose strategy A, and it also occurs
when both players choose strategy B . Let us consider, for instance, the case where both players
choose strategy B . In this case, given that Player II chooses strategy B , Player I also wants to
keep strategy B because a change of strategy would lead to a reduction of his/her payoff from 2
to 1. Similarly, given that Player I chooses strategy B , Player II wants to keep strategy B because
a change of strategy would also lead to a reduction of his/her payoff from 2 to 1.
Example 3.43 Let us consider another simple two-person game called matching pennies. Sup-
pose that Player I and Player II each has a penny. Each player must secretly turn the penny to
heads or tails and then reveal his/her choices simultaneously. Player I wins the game and gets
Player II’s penny if both coins show the same face (heads-heads or tails-tails). In the other case
where the coins show different faces (heads-tails or tails-heads), Player II wins the game and gets
Player I’s penny. e payoff function of each player is represented in the payoff matrix below.
Player II
Heads Tails
Player I
Heads 1; . 1 1;. 1
.
Tails 1;. 1 1; . 1
In this game it is not hard to see that no Nash equilibrium exists.

Denote by A the matrix that represents the payoff function of Player I, and denote by B
the matrix that represents the payoff function of Player II:
1 1 1 1
AD and B D
1 1 1 1
Now suppose that Player I is playing with a mind reader who knows Player I’s choice of faces. If
Player I decides to turn heads, then Player II knows about it and chooses to turn tails and wins
the game. In the case where Player I decides to turn tails, Player II chooses to turn heads and
again wins the game. us to have a fair game, he/she decides to randomize his/her strategy by,
e.g., tossing the coin instead of putting the coin down. We describe the new game as follows.
Player II
Heads, q1 Tails, q2
Player I
Heads, p1 1; . 1 1;. 1
.
Tails, p2 1;. 1 1; . 1
In this new game, Player I uses a coin randomly with probability of coming up heads p1 and prob-
ability of coming up tails p2 , where p1 C p2 D 1. Similarly, Player II uses another coin randomly
with probability of coming up heads q1 and probability of coming up tails q2 , where q1 C q2 D 1.
e new strategies are called mixed strategies while the original ones are called pure strategies.
Suppose that Player II uses mixed strategy fq1 ; q2 g. en Player I’s expected payoff for
playing the pure strategy heads is
uH .q1 ; q2 / D q1 q2 D 2q1 1:
Similarly, Player I’s expected payoff for playing the pure strategy tails is
uT .q1 ; q2 / D q1 C q2 D 2q1 C 1:
us Player I’s expected payoff for playing mixed strategy fp1 ; p2 g is
u1 .p; q/ D p1 uH .q1 ; q2 / C p2 uT .q1 ; q2 / D p1 .q1 q2 / C p2 . q1 C q2 / D p T Aq;
where p D Œp1 ; p2 T and q D Œq1 ; q2 T .

By the same arguments, if Player I chooses mixed strategy fp1 ; p2 g, then Player II’s ex-
pected payoff for playing mixed strategy fq1 ; q2 g is
u2 .p; q/ D p T Bq D u1 .p; q/:
For this new game, an element .p; N 2 is a Nash equilibrium if

N q/
u1 .p; q/
N u1 .p; N for all p 2 ;
N q/
u2 .p;
N q/ u2 .p; N for all q 2 ;
N q/
˚ ˇ
where WD .p1 ; p2 / ˇ p1 0; p2 0; p1 C p2 D 1 is a nonempty, compact, convex set.
Nash [23, 24] proved the existence of his equilibrium in the class of mixed strategies in a
more general setting. His proof was based on Brouwer’s fixed point theorem with rather involved
3.9. NASH EQUILIBRIUM 123
arguments. Now we present, following Rockafellar [28], a much simpler proof of the Nash equi-
librium theorem that also uses Brouwer’s fixed point theorem while applying in addition just
elementary tools of convex analysis and optimization.
eorem 3.44 Consider a two-person game f˝i ; ui g, where ˝1 Rn and ˝2 Rp are

nonempty, compact, convex sets. Let the payoff functions ui W ˝1 ˝2 ! R be given by
u1 .p; q/ WD p T Aq; u2 .p; q/ WD p T Bq;
where A and B are n p matrices. en this game admits a Nash equilibrium.
Proof. It follows from the definition that an element .p; N 2 ˝1 ˝2 is a Nash equilibrium of
N q/
the game under consideration if and only if
u1 .p; q/
N u1 .p; N for all p 2 ˝1 ;
N q/
u2 .p;
N q/ u2 .p; N for all q 2 ˝2 :
N q/
It is a simple exercise to prove (see also an elementary convex optimization result of Corollary 4.15
in Chapter 4) that this holds if and only if we have the normal cone inclusions
rp u1 .p;
N q/
N 2 N.pI
N ˝1 /; rq u2 .p;
N q/
N 2 N.qI
N ˝2 /: (3.24)
By using the structures of ui and defining ˝ WD ˝1 ˝2 , conditions (3.24) can be expressed in

one of the following two equivalent ways:
N ˝1 / and B T pN 2 N.qI
AqN 2 N.pI N ˝2 /I
N B T p/
.Aq; N 2 N.pI
N ˝1 / N.qI
N ˝2 / D N..p;
N q/I
N ˝1 ˝2 / D N..p;
N q/I
N ˝/:
By Proposition 1.78 on the Euclidean projection, this is equivalent also to
.p;
N q/
N D ˘..p;
N q/ N B T p/I
N C .Aq; N ˝/: (3.25)
Defining now the mapping F W ˝ ! ˝ by
F .p; q/ WD ˘ .p; q/ C .Aq; B T p/I ˝/ with ˝ D ˝1 ˝2
and employing another elementary projection property from Proposition 1.79 allow us to con-
clude that the mapping F is continuous and the set ˝ is nonempty, compact, and convex. By the
classical Brouwer fixed point theorem (see, e.g., [33]), the mapping F has a fixed point .p; N 2 ˝,
N q/
which satisfies (3.25), and thus it is a Nash equilibrium of the game.
Exercise 3.1 Let f W Rn ! R be a function finite around xN . Show that f is Gâteaux dif-
ferentiable at xN with fG0 .x/N D v if and only if the directional derivative f 0 .xI
N d / exists and
0 n
N d / D hv; d i for all d 2 R .
f .xI
Exercise 3.2 Let ˝ be a nonempty, closed, convex subset of Rn .

(i) Prove that the function f .x/ WD Œd.xI ˝/2 is differentiable on Rn and
rf .x/
N D 2ŒxN ˘.xI
N ˝/
for every xN 2 Rn .
(ii) Prove that the function g.x/ WD d.xI ˝/ is differentiable on Rn n˝ .
Exercise 3.3 Let F be a nonempty, compact, convex subset of Rn and let A W Rn ! Rp be a

linear mapping. Define
1
f .x/ WD sup hAx; ui kuk2 :
u2F 2
Prove that f is convex and differentiable.
Exercise 3.4 Give an example of a function f W Rn ! R, which is Gâteuax differentiable but

not Fréchet differentiable. Can such a function be convex?
@f
Exercise 3.5 Let f W Rn ! R be a convex function and let xN 2 int.domf /. Show that if
@xi
at xN exists for all i D 1; : : : ; n, then f is (Fréchet) differentiable at xN .
Exercise 3.6 Let f W Rn ! R be differentiable on Rn . Show that f is strictly convex if and

only if it satisfies the condition
hrf .x/;
N x xi
N < f .x/ N for all x 2 Rn and xN 2 Rn ; x ¤ x:
f .x/ N
Exercise 3.7 Let f W Rn ! R be twice continuously differentiable (C 2 ) on Rn . Prove that if

N is positive definite for all xN 2 Rn , then f is strictly convex. Give an example
the Hessian r 2 f .x/
showing that the converse is not true in general.
Exercise 3.8 A convex function f W Rn ! R is called strongly convex with modulus c > 0 if
for all x1 ; x2 2 Rn and 2 .0; 1/ we have
1
f x1 C .1 /x2 f .x1 / C .1 /f .x2 / c.1 /kx1 x2 k2 :
2
2 n
Prove that the following are equivalent for any C convex function f W R ! R:
(i) f strongly convex.
c
(ii) f k k2 is convex.
2
(iii) hr 2 f .x/d; d i ckd k2 for all x; d 2 Rn .
Exercise 3.9 A set-valued mapping F W Rn !

! Rn is said to be monotone if
hv2 v1 ; x2 x1 i 0 whenever vi 2 F .xi /; i D 1; 2:

F is called strictly monotone if
hv2 v1 ; x2 x1 i > 0 whenever vi 2 F .xi /; x1 ¤ x2 ; i D 1; 2:
It is called strongly monotone with modulus ` > 0 if
hv2 v1 ; x2 x1 i `kx1 x2 k2 whenever vi 2 F .xi /; i D 1; 2:
(i) Prove that for any convex function f W Rn ! R the mapping F .x/ WD @f .x/ is monotone.
(ii) Furthermore, for any convex function f W Rn ! R the subdifferential mapping @f .x/ is
strictly monotone if and only if f is strictly convex, and it is strongly monotone if and only if
f is strongly convex.
Exercise 3.10 Prove Corollary 3.7.
Exercise 3.11 Let A be an p n matrix and let b 2 Rp . Use the Farkas lemma to show that the
system Ax D b , x 0 has a feasible solution x 2 Rn if and only if the system AT y 0, b T y > 0
has no feasible solution y 2 Rp .
Exercise 3.12 Deduce Carathéodory’s theorem from Radon’s theorem and its proof.
Exercise 3.13 Use Carathéodory’s theorem to prove Helly’s theorem.
Exercise 3.14 Let C be a subspace of Rn . Show that the polar of C is represented by

˚ ˇ
C ı D C ? WD v 2 Rn ˇ hv; ci D 0 for all c 2 C :
Exercise 3.15 Let C be a nonempty, convex cone of Rn . Show that

˚ ˇ
C ı D v 2 Rn ˇ hv; ci 1 for all c 2 C :
Exercise 3.16 Let ˝1 and ˝2 be convex sets with int ˝1 \ ˝2 ¤ ;. Show that
T .xI
N ˝1 \ ˝2 / D T .xI N ˝2 / for any xN 2 ˝1 \ ˝2 :
N ˝1 / \ T .xI
Exercise 3.17 Let A be p n matrix, let
˚ ˇ
˝ WD x 2 Rn ˇ Ax D 0; x 0 ;
and let xN D .xN 1 ; : : : ; xN n / 2 ˝ . Verify the tangent and normal cone formulas:
ˇ
N ˝/ D fu D .u1 ; : : : ; un / 2 Rn ˇ Au D 0; ui 0 for any i with xN i D 0 ;
T .xI
˚ ˇ
N ˝/ D v 2 Rn ˇ v D A y
N.xI z; y 2 Rp ; z 2 Rn ; zi 0; hz; xi
N D0 :
Exercise 3.18 Give an example of a closed, convex set ˝ R2 with a point xN 2 ˝ such that
the cone RC .˝ N is not closed.
x/
Exercise 3.19 Let f W Rn ! R be a convex function such that there is ` 0 satisfying @f .x/
ÌB . Show that f is ` Lipschitz continuous on Rn .
Exercise 3.20 Let f W R ! R be convex. Prove that f is nondecreasing if and only if @f .x/
Œ0; 1/ for all x 2 R.
Exercise 3.21 Find the horizon cones of the following sets:

(i) F WD f.x; y/ 2 R2 j y x 2 g.
(ii) F WD f.x; y/ 2 R2 j y jxjg.
(iii) F WD ˚f.x; y/ 2 R2 j x > 0; y p
1=xg.
2
(iv) F WD .x; y/ 2 R j x > 0; y x 2 C 1g.
Exercise 3.22 Let vi 2 Rn and bi > 0 for i D 1; : : : ; m. Consider the set

˚ ˇ
F WD x 2 Rn ˇ hvi ; xi bi for all i D 1; : : : ; mg: (3.26)
Prove the following representation of the horizon cone and Minkowski gauge for F :
˚ ˇ
F1 D x 2 Rn ˇ hvi ; xi 0 for all i D 1; : : : ; m ;
n hvi ; ui o
F .u/ D max 0; max :
i2f1;:::;mg bi
Exercise 3.23 Let A; B be nonempty, closed, convex subsets of Rn .

(i) Show that .A B/1 D A1 B1 .
(ii) Find conditions ensuring that .A C B/1 D A1 C B1 .
Exercise 3.24 Let F ¤ ; be a closed, bounded, convex set and let ˝ ¤ ; be a closed set in
Rn . Prove that T˝F .x/ D 0 if and only if x 2 ˝ . is means that the condition 0 2 F in eo-
rem 3.27(i) is not required in this case.
Exercise 3.25 Let F ¤ ; be a closed, bounded, convex set and let ˝ ¤ ; be a closed, convex
set. Consider the sets C and S defined in eorem 3.37 and eorem 3.38, respectively. Prove
the following:
(i) If xN 2 ˝ , then
@T˝F .x/ N ˝/ \ C :
N D N.xI
(ii) If xN … ˝ , then
@T˝F .x/ N ˝r / \ S ; r WD T˝F .x/
N D N.xI N > 0:
(iii) If xN … ˝ , then
@T˝F .x/
N D @F .!N x/
N \ N.!I
N ˝/
for any !N 2 ˘F .xI
N ˝/
Exercise 3.26 Give a direct proof of characterizations (3.24) of Nash equilibrium.
Exercise 3.27 In Example 3.43 with mixed strategies, find .p; N that yields a Nash equilibrium
N q/
in this game.
129
CHAPTER 4
Applications to Optimization
and Location Problems
In this chapter we give some applications of the results of convex analysis from Chapters 1–3
to selected problems of convex optimization and facility location. Applications to optimization
problems are mainly classical, being presented however in a simplified and unified way. On the
other hand, applications to location problems are mostly based on recent journal publications and
have not been systematically discussed from the viewpoint of convex analysis in monographs and
textbooks.
4.1 LOWER SEMICONTINUITY AND EXISTENCE OF

MINIMIZERS
We first discuss the notion of lower semicontinuity for extended-real-valued functions, which
plays a crucial role in justifying the existence of absolute minimizers and related topics. Note that
most of the results presented below do not require the convexity of f .
Definition 4.1 An extended-real-valued function f W Rn ! R is   .l.s.c./

 xN 2 Rn if for every ˛ 2 R with f .x/
N > ˛ there is ı > 0 such that
f .x/ > ˛ for all x 2 IB.xI

N ı/: (4.1)
We simply say that f is   if it is l.s.c. at every point of Rn .
e following examples illustrate the validity or violation of lower semicontinuity.
Example 4.2 (i) Define the function f W R ! R by

(
x 2 if x ¤ 0;
f .x/ WD
1 if x D 0:
It is easy to see that it is l.s.c. at xN D 0 and in fact on the whole real line. On the other hand, the
replacement of f .0/ D 1 by f .0/ D 1 destroys the lower semicontinuity at xN D 0.
130 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
Figure 4.1: Which function is l.s.c?
(ii) Consider the indicator function of the set Œ 1; 1 R given by

(
0 if jxj 1;
f .x/ WD
1 otherwise:
en f is a lower semicontinuous function on R. However, the small modification

(
0 if jxj < 1;
g.x/ WD
1 otherwise
violates the lower semicontinuity on R. In fact, g is not l.s.c. at x D ˙1 while it is lower semi-
continuous at any other point of the real line.
e next four propositions present useful characterizations of lower semicontinuity.
Proposition 4.3 Let f W Rn ! R and let xN 2 dom f . en f is l.s.c. at xN if and only if for any
> 0 there is ı > 0 such that
f .x/
N < f .x/ whenever x 2 IB.xI
N ı/: (4.2)
Proof. Suppose that f is l.s.c. at xN . Since WD f .x/

N N , there is ı > 0 with
< f .x/
f .x/
N D < f .x/ for all x 2 IB.xI
N ı/:
Conversely, suppose that for any > 0 there is ı > 0 such that (4.2) holds. Fix any < f .x/
N and
choose > 0 for which < f .x/N . en
< f .x/
N < f .x/ whenever x 2 IB.xI
N ı/
for some ı > 0, which completes the proof.

4.1. LOWER SEMICONTINUITY AND EXISTENCE OF MINIMIZERS 131
n
Proposition 4.4 e function f W R ! R is l.s.c. if and only if for any number 2 R the level set
fx 2 Rn j f .x/ g is closed.
Proof. Suppose that f is l.s.c. For any fix 2 R, denote

˚ ˇ
˝ WD x 2 Rn ˇ f .x/ :
If xN … ˝ , we have f .x/
N > . By definition, find ı > 0 such that
f .x/ > for all x 2 IB.xI

N ı/;
which shows that IB.xIN ı/ ˝ c , the complement of ˝ . us ˝ c is open, and so ˝ is closed.
To verify the converse implication, suppose that the set fx 2 Rn j f .x/ g is closed for
any 2 R. Fix xN 2 Rn and 2 R with f .x/N > and then let
˚ ˇ
WD x 2 Rn ˇ f .x/ :
N ı/ c . is implies
Since is closed and xN … , there is ı > 0 such that IB.xI
f .x/ > for all x 2 IB.xI

N ı/
Proposition 4.5 e function f W Rn ! R is l.s.c. if and only if its epigraph is closed.
Proof. To verify the “only if ” part, fix .x;

N / … epi f meaning < f .x/
N . For any > 0 with
N > C > , there is ı > 0 such that f .x/ > C whenever x 2 IB.xI
f .x/ N ı/. us
IB.xI
N ı/ . ; C / .epi f /c ;
and hence the set epi f is closed.

Conversely, suppose that epi f is closed. Employing Proposition 4.4, it suffices to verify
that for any 2 R the set ˝ WD fx 2 Rn j f .x/ g is closed. To see this, fix any sequence
fxk g ˝ that converges to xN . Since f .xk / , we have .xk ; / 2 epi f for all k . It follows from
N / that .x;
.xk ; / ! .x; N / 2 epi f , and so f .x/
N . us xN 2 ˝ , which justifies the closedness
of the set ˝ .
Proposition 4.6 e function f W Rn ! R is l.s.c. at xN if and only if for any sequence xk ! xN we

have lim infk!1 f .xk / f .x/
N .
Proof. If f is l.s.c. at xN , take a sequence of xk ! xN . For any < f .x/ N , there exists ı > 0
such that (4.1) holds. en xk 2 IB.xI N ı/, and so < f .xk / for large k . It follows that
lim infk!1 f .xk /, and hence f .x/ N lim infk!1 f .xk / by the arbitrary choice of < f .x/ N .
To verify the converse, suppose by contradiction that f is not l.s.c. at xN . en there is <
1
N such that for every ı > 0 we have xı 2 IB.xI
f .x/ N ı/ with f .xı /. Applying this to ık WD
k
gives us a sequence fxk g Rn converging to xN with f .xk /. It yields
N > lim inf f .xk /;

f .x/
k!1
which is a contradiction that completes the proof.

Now we are ready to obtain a significant result on the lower semicontinuity of the minimal
value function (3.9), which is useful in what follows.
eorem 4.7 In the setting of eorem 3.27 with 0 2 F we have that the minimal time function
T˝F is lower semicontinuous.
Proof. Pick xN 2 Rn and fix an arbitrary sequence fxk g converging to xN as k ! 1. Based on

Proposition 4.6, we check that
lim inf T˝F .xk / T˝F .x/:

N
k!1
e above inequality obviously holds if lim infk!1 T˝F .xk / D 1; thus it remains to consider the
case of lim infk!1 T˝F .xk / D 2 Œ0; 1/. Since it suffices to show that T˝F .x/N , we assume
without loss of generality that T˝F .xk / ! . By the definition of the minimal time function, we
find a sequence ftk g Œ0; 1/ such that
1
T˝F .xk / tk < T˝F .xk / C and .xk C tk F / \ ˝ ¤ ; as k 2 N :
k
For each k 2 N , fix fk 2 F and !k 2 ˝ with xk C tk fk D !k and suppose without loss of gen-
erality that !k ! ! 2 ˝ . Consider the two cases: > 0 and D 0. If > 0, then
! xN
fk ! 2 F as k ! 1;

which implies that .xN C F / \ ˝ ¤ ;, and hence T˝F .x/ N . Consider the other case where
D 0. In this case the sequence ftk g converges to 0, and so tk fk ! ! xN . It follows from Propo-
sition 3.24 that ! xN 2 F1 , which yields xN 2 ˝ F1 . Employing eorem 3.27 tells us that
T˝F .x/
N D 0 and also T˝F .x/
N . is verifies the lower semicontinuity of T˝F in setting of e-
orem 3.27.
4.1. LOWER SEMICONTINUITY AND EXISTENCE OF MINIMIZERS 133
e following two propositions indicate two cases of preserving lower semicontinuity under
important operations on functions.
Pm
Proposition 4.8 Let fi W Rn ! R, i D 1; : : : ; m, be l.s.c. at some point xN . en the sum iD1 fi
is l.s.c. at this point.
Proof. It suffices to verify it for m D 2. Taking xk ! xN , we get
lim inf fi .xk / f .x/

N for i D 1; 2;
k!1
which immediately implies that

lim inf f1 .xk / C f2 .xk / lim inf f1 .xk / C lim inf f2 .xk / f1 .x/
N C f2 .x/;
N
k!1 k!1 k!1
and thus f1 C f2 is l.s.c. at xN by Proposition 4.6.
Proposition 4.9 Let I be any nonempty index set and let fi W Rn ! R be l.s.c. for all i 2 I . en
the supremum function f .x/ WD supi2I fi .x/ is also lower semicontinuous.
Proof. For any number 2 R, we have directly from the definition that
˚ ˇ ˚ ˇ \˚ ˇ
x 2 Rn ˇ f .x/ D x 2 Rn ˇ sup fi .x/ D x 2 Rn ˇ fi .x/ :
i2I i2I
us all the level sets of f are closed as intersections of closed sets. is insures the lower semi-
continuity of f by Proposition 4.4.
e next result provides two interrelated unilateral versions of the classical Weierstrass ex-
istence theorem presented in eorem 1.18. e crucial difference is that now we assume the
lower semicontinuity versus continuity while ensuring the existence of only minima, not together
with maxima as in eorem 1.18.
eorem 4.10 Let f W Rn ! R be a l.s.c. function. e following hold:

(i) e function f attains its absolute minimum on any nonempty, compact subset ˝ Rn that
intersects its domain.
(ii) Assume that infff .x/ j x 2 Rn g < 1 and that there is a real number > infff .x/ j x 2 Rn g
for which the level set fx 2 Rn j f .x/ < g of f is bounded. en f attains its absolute minimum
on Rn at some point xN 2 dom f .
Proof. To verify (i), show first that f is bounded below on ˝ . Suppose on the contrary that for
every k 2 N there is xk 2 ˝ such that f .xk / k . Since ˝ is compact, we get without loss of
generality that xk ! xN 2 ˝ . It follows from the lower semicontinuity that
N lim inf f .xk / D

f .x/ 1;
k!1
which is a contradiction showing that f is bounded below on ˝ . To justify now assertion (i),
observe that WD infx2˝ f .x/ < 1. Take fxk g ˝ such that f .xk / ! . By the compactness
of ˝ , suppose that xk ! xN 2 ˝ as k ! 1. en
N lim inf f .xk / D ;

f .x/
k!1
which shows that f .x/N D , and so xN realizes the absolute minimum of f on ˝ .

It remains to prove assertion (ii). It is not hard to show that f is bounded below, so WD
infx2Rn f .x/ is a real number. Let fxk g Rn for which f .xk / ! as k ! 1 and let k0 2 N
be such that f .xk / < for all k k0 . Since the level set fx 2 Rn j f .x/ < g is bounded, we
select a subsequence of fxk g (no relabeling) that converges to some xN . en
N lim inf f .xk / ;

f .x/
k!1
which shows that xN 2 dom f is a minimum point of f on Rn .

In contrast to all the results given above in this section, the next theorem requires convexity.
It provides effective characterizations of the level set boundedness imposed in eorem 4.10(ii).
Note that condition (iv) of the theorem below is known as coercivity.
eorem 4.11 Let f W Rn ! R be convex and let infff .x/ j x 2 Rn g < 1. e following as-
sertions are equivalent:
(i) ere exists a real number ˇ > infff .x/ j x 2 Rn g for which the level set Lˇ WD fx 2
Rn j f .x/ < ˇg is bounded.
(ii) All the level sets fx 2 Rn j f .x/ < g of f are bounded.
(iii) limkxk!1 f .x/ D 1.
f .x/
(iv) lim infkxk!1 > 0.
kxk
Proof. Let WD infff .x/ j x 2 Rn g, which is a real number or 1. We first verify (i)H)(ii).
Suppose by contradiction that there is 2 R such that the level set L is not bounded. en
there exists a sequence fxk g L with kxk k ! 1. Fix ˛ 2 . ; ˇ/, xN 2 L˛ and define the convex
combination
1 1 1
yk WD xN C p xk xN D p xk C 1 p x:
N
kxk k kxk k kxk k
4.2. OPTIMALITY CONDITIONS 135
By the convexity of f , we have the relationships
1 1 1
f .yk / p f .xk / C .1 p N <p
/f .x/ C .1 p /˛ ! ˛ < ˇ;
kxk k kxk k kxk k kxk k
and hence yk 2 Lˇ for k sufficiently large, which ensures the boundedness of fyk g. It follows
from the triangle inequality that
x 1 p 1
k
kyk k p 1 p kxk
N kxk k 1 p kxk
N ! 1;
kxk k kxk k kxk k
which is a contradiction. us (ii) holds.
Next we show that (ii)H)(iii). Indeed, the violation of (iii) yields the existence of a constant
M > 0 and a sequence fxk g such that kxk k ! 1 as k ! 1 and f .xk / < M for all k . But this
obviously contradicts the boundedness of the level set fx 2 Rn j f .x/ < M g.
Finally, we justify (iii)H)(i) leaving the proof of (iii)”(iv) as an exercise. Suppose that
(iii) is satisfied and for M > , find ˛ such that f .x/ M whenever kxk > ˛ . en
˚ ˇ
LM WD x 2 Rn ˇ f .x/ < M g IB.0I ˛/;
which shows that this set is bounded and thus verifies (i).
4.2 OPTIMALITY CONDITIONS

In this section we first derive necessary and sufficient optimality conditions for optimal solutions
to the following set-constrained convex optimization problem:
minimize f .x/ subject to x 2 ˝; (4.3)
where ˝ is a nonempty, convex set and f W Rn ! R is a convex function.
Definition 4.12 We say that xN 2 ˝ is a    to (4.3) if f .x/

N < 1 and there
is > 0 such that
f .x/ f .x/N for all x 2 IB.xI
N / \ ˝: (4.4)
If f .x/ f .x/
N for all x 2 ˝ , then xN 2 ˝ is a    to (4.3).
e next result shows, in particular, that there is no difference between local and global
solutions for problems of convex optimization.
Proposition 4.13 e following are equivalent in the convex setting of (4.3):

(i) xN is a local optimal solution to the constrained problem (4.3).
(ii) xN is a local optimal solution to the unconstrained problem
minimize g.x/ WD f .x/ C ı.xI ˝/ on Rn : (4.5)

(iii) xN is a global optimal solution to the unconstrained problem (4.5).
(iv) xN is a global optimal solution to the constrained problem (4.3).
Proof. To verify (i)H)(ii), select > 0 such that (4.4) is satisfied. Since g.x/ D f .x/ for x 2
N / and since g.x/ D 1 for x 2 IB.xI
˝ \ IB.xI N / and x … ˝ , we have
N for all x 2 IB.xI

g.x/ g.x/ N /;
which means (ii). Implication (ii)H)(iii) is a consequence of Proposition 2.33. e verification

of (iii)H)(iv) is similar to that of (i)H)(ii), and (vi)H)(i) is obvious.
In what follows we use the term “optimal solutions” for problems of convex optimization,
without mentioning global or local. Observe that the objective function g of the unconstrained
problem (4.5) associated with the constrained one (4.3) is always extended-real-valued when ˝ ¤
Rn . is reflects the “infinite penalty” for the constraint violation.
e next theorem presents basic necessary and sufficient optimality conditions for optimal
solutions to the constrained problem (4.3) with a general (convex) cost function.
eorem 4.14 Consider the convex problem (4.3) with f W Rn ! R such that f .x/
N < 1 for
some point xN 2 ˝ under the qualification condition

@1 f .x/
N \ N.xIN ˝/ D f0g; (4.6)
which automatically holds if f is continuous at xN . en the following are equivalent:

(i) xN is an optimal solution to (4.3).
(ii) 0 2 @f .x/
N C N.xI N ˝/, i.e., @f .x/ N ˝/ ¤ ;.
N \ Œ N.xI
(iii) f 0 .xI
N d / 0 for all d 2 T .xI
N ˝/.
Proof. Observe first that @1 f .x/

N D f0g and thus (4.6) holds if f is continuous at xN . is follows
from eorem 2.29. To verify now (i)H)(ii), we use the unconstrained form (4.5) of problem
(4.3) due to Proposition 4.13. en Proposition 2.35 provides the optimal solution characteriza-
tion 0 2 @g.x/
N . Applying the subdifferential sum rule from eorem 2.44 under the qualification
condition (4.6) gives us the optimality condition in (ii).
To show that (ii)H)(iii), we have by the tangent-normal duality (3.4) the equivalence

N ˝/ ” hv; d i 0 for all v 2 N.xI
d 2 T .xI N ˝/ :
Fix an element w 2 @f .x/ N with w 2 N.xI N ˝/. en hw; d i 0 for all d 2 T .xI
N ˝/ and we
arrive at f 0 .xI
N d / hw; d i 0 by eorem 2.84, achieving (iii) in this way.
It remains to justify implication (iii)H)(i). It follows from Lemma 2.83 that
f .xN C d / N f 0 .xI
f .x/ N d / 0 for all d 2 T .xI
N ˝/:

By definition of the tangent cone, we have T .xI
N ˝/ D cl cone.˝ N , which tells us that d WD
x/
N ˝/ if w 2 ˝ . is yields
w xN 2 T .xI
f .w/ f .x/
N D f .w xN C x/
N f .x/
N D f .xN C d / f .x/
N 0;
which verifies (i) and completes the proof of the theorem.
Corollary 4.15 Assume that in the framework of eorem 4.14 the cost function f W Rn ! R is
differentiable at xN 2 dom f . en the following are equivalent:
(i) xN is an optimal solution to problem (4.3).
(ii) rf .x/ N 2 N.xI N ˝/, which reduces to rf .x/
N D 0 if xN 2 int ˝ .
Proof. Since the differentiability of the convex function f at xN implies its continuity and so Lip-
schitz continuity around this point by eorem 2.29. Since @f .x/ N D frf .x/g N in this case by
eorem 3.3, we deduce this corollary directly from eorem 4.14.
Next we consider the following convex optimization problem containing functional in-
equality constraints given by finite convex functions, along with the set/geometric constraint as in
(4.3). For simplicity, suppose that f is also finite on Rn . Problem .P / is:
minimize f .x/
subject to gi .x/ 0 for i D 1; : : : ; m;
x 2 ˝:
Define the Lagrange function for .P / involving only the cost and functional constraints by
m
X
L.x; / WD f .x/ C i gi .x/
iD1
with i 2 R and the active index set given by I.x/ WD fi 2 f1; : : : ; mg j gi .x/ D 0g.
eorem 4.16 Let xN be an optimal solution to .P /. en there are multipliers 0 0 and
D .1 ; : : : ; m / 2 Rm , not equal to zero simultaneously, such that i 0 as i D 0; : : : ; m,
m
X
0 2 0 @f .x/
N C i @gi .x/
N C N.xI
N ˝/;
iD1
and i gi .x/
N D 0 for all i D 1; : : : ; m.
Proof. Consider the finite convex function
˚ ˇ
N gi .x/ ˇ i D 1; : : : ; m
'.x/ WD max f .x/ f .x/;
and observe that xN solves the following problem with no functional constraints:
minimize '.x/ subject to x 2 ˝:
Since ' is finite and hence locally Lipschitz continuous, we get from eorem 4.14 that 0 2
@'.x/
N C N.xIN ˝/. e subdifferential maximum rule from Proposition 2.54 yields
[
0 2 co @f .x/
N [Œ @gi .x/
N C N.xI N ˝/:
i2I.x/
N
P
is gives us 0 0 and i 0 for i 2 I.x/
N such that 0 C i2I.x/
N i D 1 and
X
0 2 0 @f .x/
N C i @gi .x/
N C N.xI
N ˝/: (4.7)
i2I.x/
N
Letting i WD 0 for i … I.x/

N , we get i gi .x/
N D 0 for all i D 1; : : : ; m with .0 ; / ¤ 0.
Many applications including numerical algorithms to solve problem .P / require effective
constraint qualification conditions that ensure 0 ¤ 0 in eorem 4.16. e following one is clas-
sical in convex optimization.
Definition 4.17 S’   .CQ/ holds in .P / if there exists u 2 ˝

such that gi .u/ < 0 for all i D 1; : : : ; m.
e next theorem provides necessary and sufficient conditions for convex optimization in
the qualified, normal, or Karush-Kuhn-Tucker .KKT/ form.
eorem 4.18 Suppose that Slater’s CQ holds in .P /. en xN is an optimal solution to .P / if

and only if there exist nonnegative Lagrange multipliers .1 ; : : : ; m / 2 Rm such that
m
X
0 2 @f .x/
N C i @gi .x/
N C N.xI
N ˝/ (4.8)
iD1
and i gi .x/
N D 0 for all i D 1; : : : ; m.
Proof. To verify the necessity part, we only need to show that 0 ¤ 0 in (4.7). Suppose on the
contrary that 0 D 0 and find .1 ; : : : ; m / ¤ 0, vi 2 @gi .x/
N , and v 2 N.xI
N ˝/ satisfying
m
X
0D i vi C v:
iD1
is readily implies that
m
X m
X m
X
0D i hvi ; x xi
N C hv; x xi
N i gi .x/ gi .x/
N D i gi .x/ for all x 2 ˝;
iD1 iD1 iD1
P
which contradicts since the Slater condition implies m iD1 i gi .u/ < 0.
To check the sufficiency of (4.8), take v0 2 @f .x/N , vi 2 @gi .x/
N , and v 2 N.xI N ˝/ such that
P
0 D v0 C m v
iD1 i i C v with g
i i .x/
N D 0 and i 0 for all i D 1; : : : ; m . en the optimality
of xN in .P / follows directly from the definitions of the subdifferential and normal cone. Indeed,
for any x 2 ˝ with gi .x/ 0 for i D 1; : : : ; m one has
m
X
0 D hi vi C v; x xi
N D i hvi ; x xi
N C hv; x xi
N
iD1
m
X
i Œgi .x/ gi .x/
N C f .x/ f .x/
N
iD1
Xm
D i gi .x/ C f .x/ f .x/
N f .x/ f .x/;
N
iD1
which completes the proof.

Finally in this section, we consider the convex optimization problem .Q/ with inequality
and linear equality constraints
minimize f .x/
subject to gi .x/ 0 for i D 1; : : : ; m;
Ax D b;
where A is an p n matrix and b 2 Rp .
Corollary 4.19 Suppose that Slater’s CQ holds for problem .Q/ with the set ˝ WD fx 2 Rn j Ax
b D 0g in Definition 4.17. en xN is an optimal solution to .Q/ if and only if there exist nonnegative
multipliers 1 ; : : : ; m such that
m
X
0 2 @f .x/
N C N C im A
i @gi .x/
iD1
and i gi .x/
N D 0 for all i D 1; : : : ; m.
Proof. e result follows from eorem 4.16 and Proposition 2.12, where the normal cone to the
set ˝ is calculated via the image of the adjoint operator A .
e concluding result presents the classical version of the Lagrange multiplier rule for con-
vex problems with differentiable data.
Corollary 4.20 Let f and gi for i D 1; : : : ; m be differentiable at xN and let frgi .x/ N be
N j i 2 I.x/g
n
linearly independent. en xN is an optimal solution to problem .P / with ˝ D R if and only if there
are nonnegative multipliers 1 ; : : : ; m such that
m
X
0 2 rf .x/
N C i rgi .x/
N D Lx .x;
N /
iD1
and i gi .x/
N D 0 for all i D 1; : : : ; m.
Proof. In this case N.xI

N ˝/ D f0g and it follows from (4.7) that 0 D 0 contradicts the linear
independence of the gradient elements frgi .x/ N .
N j i 2 I.x/g
4.3 SUBGRADIENT METHODS IN CONVEX

OPTIMIZATION
is section is devoted to algorithmic aspects of convex optimization and presents two versions of
the classical subgradient method: one for unconstrained and the other for constrained optimiza-
tion problems. First we study the unconstrained problem:
minimize f .x/ subject to x 2 Rn ; (4.9)
where f W Rn ! R is a convex function. Let f˛k g as k 2 N be a sequence of positive numbers. e

subgradient algorithm generated by f˛k g is defined as follows. Given a starting point x1 2 Rn ,
consider the iterative procedure:
xkC1 WD xk ˛k vk with some vk 2 @f .xk /; k 2N: (4.10)
In this section we assume that problem (4.9) admits an optimal solution and that f is convex and
Lipschitz continuous on Rn . Our main goal is to find conditions that ensure the convergence of
the algorithm. We split the proof of the main result into several propositions, which are of their
own interest. e first one gives us an important subgradient property of convex Lipschitzian
functions.
Proposition 4.21 Let ` 0 be a Lipschitz constant of f on Rn . For any x 2 Rn , we have the

subgradient estimate
kvk ` whenever v 2 @f .x/:
4.3. SUBGRADIENT METHODS IN CONVEX OPTIMIZATION 141
n
Proof. Fix x 2 R and take any subgradient v 2 @f .x/. It follows from the definition that
hv; u xi f .u/ f .x/ for all u 2 Rn :
Combining this with the Lipschitz property of f gives us
hv; u xi ` ku xk ; u 2 Rn ;
which yields in turn that
hv; y C x xi ` ky C x xk ; i.e. hv; yi ` kyk ; y 2 Rn :
Using y WD v , we arrive at the conclusion kvk `.
Proposition 4.22 Let ` 0 be a Lipschitz constant of f on Rn . For the sequence fxk g generated by
algorithm (4.10), we have

kxkC1 xk2 kxk xk2 2˛k f .xk / f .x/ C ˛k2 `2 for all x 2 Rn ; k 2 N : (4.11)
Proof. Fix x 2 Rn , k 2 N and get from (4.10) that
kxkC1 xk2 D kxk ˛k vk xk2 D kxk x ˛k vk k2

D kxk xk2 2˛k hvk ; xk xi C ˛k2 kvk k2 :
Since vk 2 @f .xk / and kvk k ` by Proposition 4.21, it follows that
hvk ; x xk i f .x/ f .xk /:
is gives us (4.11) and completes the proof of the proposition.

e next proposition provides an estimate of closeness of iterations to the optimal value of
(4.9) for an arbitrary sequence f˛k g in the algorithm (4.10). Denote
˚ ˇ ˚
V WD min f .x/ ˇ x 2 Rn and Vk WD min f .x1 /; : : : ; f .xk / :
Proposition 4.23 Let ` 0 be a Lipschitz constant of f on Rn and let S be the set of optimal
solutions to (4.9). For all k 2 N , we have
P
k
d.x1 I S/2 C `2 ˛i2
iD1
0 Vk V : (4.12)
P
k
2 ˛i
iD1
Proof. Substituting any xN 2 S into estimate (4.11) gives us

kxkC1 N 2 kxk
xk N 2
xk 2˛k f .xk / N C ˛k2 `2 :
f .x/
N , we get
Since Vk f .xi / for all i D 1; : : : ; k and V D f .x/
k
X k
X

kxkC1 N 2 kx1
xk N 2
xk 2 ˛i f .xi / N C `2
f .x/ ˛i2
iD1 iD1
k
X k
X
kx1 N 2
xk 2 ˛i .Vk V / C `2 ˛i2 :
iD1 iD1
It follows from kxkC1 N 2 0 that

xk
k
X k
X

2 ˛i Vk V kx1 N 2 C `2
xk ˛i2 :
iD1 iD1
Since Vk V for every k , we deduce the estimate

P
kx1 N 2 C `2 kiD1 ˛i2
xk
0 Vk V P
2 kiD1 ˛i
and hence arrive at (4.12) since xN 2 S was chosen arbitrarily.
e corollary below shows the number of steps in order to achieve an appropriate error
estimate for the optimal value in the case of constant step sizes.
Corollary 4.24 Suppose that ˛k D for all k 2 N in the setting of Proposition 4.23. en there
exists a positive constant C such that
0 Vk V < `2 whenever k > C = 2 :
Proof. From (4.12) we readily get

a C k`2 2 a `2
0 Vk V D C with a WD d.x1 I S /2 :
2k 2k 2
Since a=.2k/ C `2 =2 < `2 as k > a=.`2 2 /, the conclusion holds with C WD a=`2 .
As a direct consequence of Proposition 4.23, we get our first convergence result.
Proposition 4.25 In the setting of Proposition 4.23 suppose in addition that
1
X
˛k D 1 and lim ˛k D 0: (4.13)
k!1
kD1
en we have the convergence Vk ! V as k ! 1.

Proof. Given any > 0, choose K 2 N such that ˛i ` < for all i K . For any k > K , using
(4.12) ensures the estimates
P
K P
k
d.x1 I S /2 C `2 ˛i2 `2 ˛i2
iD1 iDK
0 Vk V C
P
k P
k
2 ˛i 2 ˛i
iD1 iD1
P
K P
k
d.x1 I S/2 C `2 ˛i2 ˛i
iD1 iDK
C
P
k P
k
2 ˛i 2 ˛i
iD1 iD1
P
K
d.x1 I S/2 C `2 ˛i2
iD1
C =2:
P
k
2 ˛i
iD1
P1 Pk
e assumption kD1 ˛k D 1 yields limk!1 iD1 ˛i D 1, and hence
0 lim inf.Vk V / lim sup.Vk V / =2:

k!1 k!1
Since > 0 is chosen arbitrarily, one has limk!1 .Vk V / D 0 or Vk ! V as k ! 1.

e next result gives us important information about the value convergence of the subgra-
dient algorithm (4.10).
Proposition 4.26 In the setting of Proposition 4.25 we have
lim inf f .xk / D V :

k!1
Proof. Since V is the optimal value in (4.9), we get f .xk / V for any k 2 N . is yields
lim inf f .xk / V :

k!1
To verify the opposite inequality, suppose on the contrary that lim infk!1 f .xk / > V and then
find a small number > 0 such that
lim inf f .xk / 2 > V D minn f .x/ D f .x/;

N
k!1 x2R
where xN is an optimal solution to problem (4.9). us
lim inf f .xk / > f .x/

N C :
k!1
In the case where lim infk!1 f .xk / < 1, by the definition of lim inf, there is k0 2 N such that
for every k k0 we have that
f .xk / > lim inf f .xk / ;

k!1
and hence f .xk / f .x/N > for every k k0 . Note that this conclusion also holds if
lim infk!1 f .xk / D 1. It follows from (4.11) with x D xN that

kxkC1 xkN 2 kxk xk N 2 2˛k f .xk / f .x/ N C ˛k2 `2 for every k k0 : (4.14)
Let k1 be such that ˛k `2 < for all k k1 and let kN1 WD maxfk1 ; k0 g. For any k > kN1 , it easily
follows from estimate (4.14) that
2 k
X

kxkC1 N 2 kxk
xk N 2
xk ˛k : : : xkN1 xN ˛j ;
j DkN1
P1
which contradicts the assumption kD1 ˛k D 1 in (4.13) and so completes the proof.
Now we are ready to justify the convergence of the subgradient algorithm (4.10).
eorem 4.27 Suppose that the generating sequence f˛k g in (4.10) satisfies
1
X 1
X
˛k D 1 and ˛k2 < 1: (4.15)
kD1 kD1
en the sequence fVk g converges to the optimal value V and the sequence of iterations fxk g in
(4.10) converges to some optimal solution xN of (4.9).
Proof. Since (4.15) implies (4.13), the sequence fVk g converges to V by Proposition 4.25. It
remains to prove that fxk g converges to some optimal solution xN of problem (4.9). Let ` 0 be
a Lipschitz constant of f on Rn . Since the set S of optimal solutions to (4.9) is assumed to be
nonempty, we pick any ex 2 S and substitute x D ex in estimate (4.11) of Proposition 4.22. It gives
us
kxkC1 e x k2 kxk e x k2 2˛k f .xk / V C ˛k2 `2 for all k 2 N : (4.16)
Since f .xk / V 0, inequality (4.16) yields
kxkC1 x k2 kxk
e x k2 C ˛k2 `2 :
e
en we get by induction that

k
X
kxkC1 x k2 kx1
e x k2 C `2
e ˛i2 : (4.17)
iD1
P1 2
e right-hand side of (4.17) is finite by our assumption that kD1 ˛k < 1, and hence
sup kxkC1 e
x k < 1:
k2N
is implies that the sequence fxk g is bounded. Furthermore, Proposition 4.26 tells us that
lim inf f .xk / D V :

k!1
Let fxkj g be a subsequence of fxk g such that
lim f .xkj / D V : (4.18)

j !1
en fxkj g is also bounded and contains a limiting point xN . Without loss of generality assume
that xkj ! xN as j ! 1. We directly deduce from (4.18) and the continuity of f that xN is an
optimal solution to (4.9). Fix any j 2 N and any k kj . us we can rewrite (4.16) with ex D xN
and see that
k
X
2
kxkC1 xk 2
N kxk xk
N C2
˛k2 `2 : : : xk
j

xN C ` 2
˛i2 :
iDkj
Taking first the limit as k ! 1 and then the limit as j ! 1 in the resulting inequality above
gives us the estimate
1
X
2
lim sup kxkC1 N 2 lim xkj
xk xN C `2 lim ˛i2 :
k!1 j !1 j !1
iDkj
P1
Since xkj ! xN and kD1 ˛k2 < 1, this implies that
lim sup kxkC1 N 2 D 0;

xk
k!1
and thus justifies the claimed convergence of xk ! xN as k ! 1.

Finally in this section, we present a version of the subgradient algorithm for convex prob-
lems of constrained optimization given by
minimize f .x/ subject to x 2 ˝; (4.19)
where f W Rn ! R is a convex function that is Lipschitz continuous on Rn with some constant

` 0, while—in contrast to (4.9)—we have now the constraint on x given by a nonempty, closed,
convex set ˝ Rn . e following counterpart of the subgradient algorithm (4.10) for constrained
problems is known as the projected subgradient method.
Given a sequence of positive numbers f˛k g and given a starting point x1 2 ˝ , the sequence
of iterations fxk g is constructed by
xkC1 WD ˘.xk ˛k vk I ˝/ with some vk 2 @f .xk /; k 2 N; (4.20)
where ˘.xI ˝/ stands for the (unique) Euclidean projection of x onto ˝ .

e next proposition, which is a counterpart of Proposition 4.22 for algorithm (4.20), plays
a major role in justifying convergence results for the projected subgradient method.
Proposition 4.28 In the general setting of (4.20) we have the estimate

kxkC1 xk2 kxk xk2 2˛k f .xk / f .x/ C ˛k2 `2 for all x 2 ˝; k 2 N :
Proof. Define zkC1 WD xk ˛k vk and fix any x 2 ˝ . It follows from Proposition 4.22 that

kzkC1 xk2 kxk xk2 2˛k f .xk / f .x/ C ˛k2 `2 for all k 2 N : (4.21)
Now by projecting zkC1 onto ˝ and using Proposition 1.79, we get
kxkC1 xk D k˘.zkC1 I ˝/ ˘.xI ˝/k kzkC1 xk :
Combining this with estimate (4.21) gives us

kxkC1 xk2 kxk xk2 2˛k f .xk / f .x/ C ˛k2 `2 for all k 2 N;
which thus completes the proof of the proposition.

Based on this proposition and proceeding in the same way as for the subgradient algorithm
(4.10) above, we can derive similar convergence results for the projected subgradient algorithm
(4.20) under the assumptions (4.13) on the generating sequence f˛k g.
4.4 THE FERMAT-TORRICELLI PROBLEM

In 1643 Pierre de Fermat proposed to Evangelista Torricelli the following optimization problem:
Given three points in the plane, find another point such that the sum of its distances to the given
points is minimal. is problem was solved by Torricelli and was named the Fermat-Torricelli
problem. A more general version of the Fermat-Torricelli problem asks for a point that minimizes
the sum of the distances to a finite number of given points in Rn . is is one of the main problems
in location science, which has been studied by many researchers. e first numerical algorithm for
solving the general Fermat-Torricelli problem was introduced by Endre Weiszfeld in 1937. As-
sumptions that guarantee the convergence of the Weiszfeld algorithm were given by Harold Kuhn
in 1972. Kuhn also pointed out an example in which the Weiszfeld algorithm fails to converge.
Several new algorithms have been introduced recently to improve the Weiszfeld algorithm. e
4.4. THE FERMAT-TORRICELLI PROBLEM 147
goal of this section is to revisit the Fermat-Torricelli problem from both theoretical and numerical
viewpoints by using techniques of convex analysis and optimization.
Given points ai 2 Rn as i D 1; : : : ; m, define the (nonsmooth) convex function
m
X
'.x/ WD kx ai k: (4.22)
iD1
en the mathematical description of the Fermat-Torricelli problem is
minimize '.x/ over all x 2 Rn : (4.23)
Proposition 4.29 e Fermat-Torricelli problem (4.23) has at least one solution.
Proof. e conclusion follows directly from eorem 4.10.
Proposition 4.30 Suppose that ai as i D 1; : : : ; m are not collinear, i.e., they do not belong to the
same line. en the function ' in (4.22) is strictly convex, and hence the Fermat-Torricelli problem
(4.23) has a unique solution.
Proof. Since each function 'i .x/ WD kx ai k as i D 1; : : : ; m is obviously convex, their sum
P
'D m n
iD1 'i is convex as well, i.e., for any x; y 2 R and 2 .0; 1/ we have

' x C .1 /y '.x/ C .1 /'.y/: (4.24)
N yN 2 Rn with xN ¤ yN and 2 .0; 1/

Supposing by contradiction that ' is not strictly convex, find x;
such that (4.24) holds as equality. It follows that

'i xN C .1 /yN D 'i .x/ N C .1 /'i .y/ N for all i D 1; : : : ; m;
which ensures therefore that
k.xN ai / C .1 /.yN ai /k D k.xN ai /k C k.1 /.yN ai /k; i D 1; : : : ; m:
If xN ¤ ai and yN ¤ ai , then there exists ti > 0 such that ti .xN ai / D .1 /.yN ai /, and hence
1
xN ai D i .yN ai / with i WD :
ti
Since xN ¤ yN , we have i ¤ 1. us
1 i
ai D xN yN 2 L.x;
N y/;
N
1 i 1 i
where L.x; N signifies the line connecting xN and yN . Both cases where xN D ai and yN D ai give us
N y/
ai 2 L.x; N . Hence ai 2 L.x;
N y/ N for i D 1; : : : ; m, a contradiction.
N y/
We proceed further with solving the Fermat-Torricelli problem for three points in Rn . Let
us first present two elementary lemmas.
Lemma 4.31 Let a1 and a2 be any two unit vectors in Rn . en we have a1 a2 2 IB if and
1
only if ha1 ; a2 i .
2
Proof. If a1 a2 2 IB , then a1 C a2 2 IB and
ka1 C a2 k 1:
By squaring both sides, we get the inequality
ka1 C a2 k2 1;
which subsequently implies that
ka1 k2 C 2ha1 ; a2 i C ka2 k2 D 2 C 2ha1 ; a2 i 1:
1
erefore, ha1 ; a2 i . e proof of the converse implication is similar.
2
Lemma 4.32 Let a1 ; a2 ; a3 2 Rn with ka1 k D ka2 k D ka3 k D 1. en a1 C a2 C a3 D 0 if and

only if we have the equalities
1
ha1 ; a2 i D ha2 ; a3 i D ha3 ; a1 i D : (4.25)
2
Proof. It follows from the condition a1 C a2 C a3 D 0 that
ka1 k2 C ha1 ; a2 i C ha1 ; a3 i D 0; (4.26)
ha2 ; a1 i C ka2 k2 C ha2 ; a3 i D 0; (4.27)

2
ha3 ; a1 i C ha3 ; a2 i C ka3 k D 0: (4.28)
Subtracting (4.28) from (4.26) gives us ha1 ; a2 i D ha2 ; a3 i and similarly ha2 ; a3 i D ha3 ; a1 i. Sub-
stituting these equalities into (4.26), (4.27), and (4.28) ensures the validity of (4.25).
Conversely, assume that (4.25) holds. en we have
3
X 3
X
ka1 C a2 C a3 k2 D kai k2 C hai ; aj i D 0;
iD1 i;j D1;i ¤j
which readily yields the equality a1 C a2 C a3 D 0 and thus completes the proof.
Now we use subdifferentiation of convex functions to solve the Fermat-Torricelli problem
for three points in Rn . Given two nonzero vectors u; v 2 Rn , define
hu; vi
cos.u; v/ WD
kuk kvk
and also consider, when xN ¤ ai , the three vectors
xN ai
vi D ; i D 1; 2; 3:
kxN ai k
Geometrically, each vector vi is a unit one pointing in the direction from the vertex ai to xN .
Observe that the classical (three-point) Fermat-Torricelli problem always has a unique solution
even if the given points belong to the same line. Indeed, in the latter case the middle point solves
the problem.
e next theorem fully describes optimal solutions of the Fermat-Torricelli problem for
three points in two distinct cases.
eorem 4.33 Consider the Fermat-Torricelli problem (4.23) for the three points a1 ; a2 ; a3 2
Rn . e following assertions hold:
(i) Let xN … fa1 ; a2 ; a3 g. en xN is the solution of the problem if and only if
cos.v1 ; v2 / D cos.v2 ; v3 / D cos.v3 ; v1 / D 1=2:
(ii) Let xN 2 fa1 ; a2 ; a3 g, say xN D a1 . en xN is the solution of the problem if and only if
cos.v2 ; v3 / 1=2:
Proof. In case (i) function (4.22) is differentiable at xN . en xN solves (4.23) if and only if
r'.x/
N D v1 C v2 C v3 D 0:
By Lemma 4.32 this holds if and only if

˚
hvi ; vj i D cos.vi ; vj / D 1=2 for any i; j 2 1; 2; 3 with i ¤ j:
To verify (ii), employ the subdifferential Fermat rule (2.35) and sum rule from Corollary 2.45,
which tell us that xN D a1 solves the Fermat-Torricelli problem if and only if
0 2 @'.a1 / D v2 C v3 C IB:
Since v2 and v3 are unit vectors, it follows from Lemma 4.31 that
hv2 ; v3 i D cos.v2 ; v3 / 1=2;

Figure 4.2: Construction of the Torricelli point.
Example 4.34 is example illustrates geometrically the solution construction for the Fermat-
Torricelli problem in the plane. Consider (4.23) with the three given points A, B , and C on
R2 as shown in Figure 4.2. If one of the angles of the triangle ABC is greater than or equal to
120ı , then the corresponding vertex is the solution to (4.23) by eorem 4.33(ii). Consider the
case where none of the angles of the triangle is greater than or equal to 120ı . Construct the two
equilateral triangles ABD and ACE and let S be the intersection of DC and BE as in the figure.
Two quadrilaterals ADBC and ABCE are convex, and hence S lies inside the triangle ABC . It is
clear that two triangles DAC and BAE are congruent. A rotation of 60ı about A maps the triangle
DAC to the triangle BAE . e rotation maps CD to BE , so †DSB D 60ı . Let T be the image of
S though this rotation. en T belongs to BE . It follows that †AS T D †ASE D 60ı . Moreover,
†DSA D 60ı , and hence †BSA D 120ı . It is now clear that †AS C D 120ı and †BS C D 120ı .
By eorem 4.33(i) the point S is the solution of the problem under consideration.
Next we present the Weiszfeld algorithm [32] to solve numerically the Fermat-Torricelli
problem in the general case of (4.23) for m points in Rn that are not collinear. is is our standing
assumption in the rest of this section. To proceed, observe that the gradient of the function ' from
(4.22) is calculated by
m
X x ai ˚
r'.x/ D for x … a1 ; : : : ; am :
kx ai k
iD1
Solving the gradient equation r'.x/ D 0 gives us

Pm ai
iD1
kx ai k
xD DW f .x/: (4.29)
Pm 1
iD1
kx ai k
n
We define f .x/ on the whole space R by letting f .x/ WD x for x 2 fa1 ; : : : ; am g.
Weiszfeld algorithm consists of choosing an arbitrary starting point x1 2 Rn , taking f is
from (4.29), and defining recurrently
xkC1 WD f .xk / for k 2 N; (4.30)
where f .x/ is called the algorithm mapping. In the rest of this section we present the precise
conditions and the proof of convergence for Weiszfeld algorithm in the way suggested by Kuhn
[10] who corrected some errors from [32]. In contrast to [10], we employ below subdifferential
rules of convex analysis, which allow us to simplify the proof of convergence.
e next proposition ensures decreasing the function value of ' after each iteration, i.e.,
the descent property of the algorithm.
Proposition 4.35 If f .x/ ¤ x , then '.f .x// < '.x/.

Proof. e assumption of f .x/ ¤ x tells us that x is not a vertex from fa1 ; : : : ; am g. Observe
also that the point f .x/ is a unique minimizer of the strictly convex function
m
X kz a i k2
g.z/ WD
kx ai k
iD1
and that g.f .x// < g.x/ D '.x/ since f .x/ ¤ x . We have furthermore that
m
X kf .x/ ai k2
g f .x/ D
kx ai k
iD1
m
X .kx ai k C kf .x/ ai k kx ai k/2
D
kx ai k
iD1
m
X
.kf .x/ ai k kx ai k/2
D '.x/ C 2 '.f .x// '.x/ C :
kx ai k
iD1
is clearly implies the inequality

m
X
.kf .x/ ai k kx ai k/2
2' f .x/ C < 2'.x/
kx ai k
iD1

e following two propositions show how the algorithm mapping f behaves near a vertex
that solves the Fermat-Torricelli problem. First we present a necessary and sufficient condition
ensuring the optimality of a vertex in (4.23). Define
m
X ai aj
Rj WD ; j D 1; : : : ; m:
kai aj k
iD1;i ¤j
Proposition 4.36 e vertex aj solves problem (4.23) if and only if kRj k 1.
Proof. Employing the Fermat rule from Proposition 2.35 and the subdifferential sum rule from
Corollary 2.45 shows that the vertex aj solves problem (4.23) if and only if
0 2 @'.aj / D Rj C IB.0; 1/;
which is clearly equivalent to the condition kRj k 1.

Proposition 4.36 from convex analysis allows us to essentially simplify the proof of the
following result established by Kuhn. We use the notation:

f s .x/ WD f f s 1 .x/ for s 2 N with f 0 .x/ WD x; x 2 Rn :
Proposition 4.37 Suppose that the vertex aj does not solve (4.23). en there is ı > 0 such that the
condition 0 < kx aj k ı implies the existence of s 2 N with
kf s .x/ aj k > ı and kf s 1

.x/ aj k ı: (4.31)
Proof. For any x … fa1 ; : : : ; am g, we have from (4.29) that

Pm ai aj
iD1;i¤j
kx ai k
f .x/ aj D and so
Pm 1
iD1
kx ai k
Pm
ai aj
iD1;i ¤j
f .x/ aj kx ai k
lim D lim D Rj ; j D 1; : : : ; m:
x!aj kx aj k x!aj Pm kx aj k
1 C iD1;i ¤j
kx ai k
It follows from Proposition 4.36 that
kf .x/ aj k
lim D kRj k > 1; j D 1; : : : ; m; (4.32)
x!aj kx aj k
which allows us to select > 0; ı > 0 such that
kf .x/ aj k
> 1 C whenever 0 < kx aj k ı:
kx aj k
q 1
us for every s 2 N the validity of 0 < kf .x/ aj k ı with any q D 1; : : : ; s yields
kf s .x/ aj k .1 C /kf s 1
.x/ aj k : : : .1 C /s kx aj k:
Since .1 C /s kx aj k ! 1 as s ! 1, it gives us (4.31) and completes the proof.
We finally present the main result on the convergence of the Weiszfeld algorithm.
eorem 4.38 Let fxk g be the sequence generated by Weiszfeld algorithm (4.30) and let xk …
fa1 ; : : : ; am g for k 2 N . en fxk g converges to the unique solution xN of (4.23).
Proof. If xk D xkC1 for some k D k0 , then fxk g is a constant sequence for k k0 , and hence
it converges to xk0 . Since f .xk0 / D xk0 and xk0 is not a vertex, xk0 solves the problem. us
we can assume that xkC1 ¤ xk for every k . It follows from Proposition 4.35 that the sequence
f'.xk /g of nonnegative numbers is decreasing, so it converges. is gives us

lim '.xk / '.xkC1 / D 0: (4.33)
k!1
By the construction of fxk g, we have that each xk belongs to the compact set co fa1 ; : : : ; am g.
is allows us to select from fxk g a subsequence fxk g converging to some point zN as ! 1.
We have to show that zN D xN , i.e., it is the unique solution to (4.23). Indeed, we get from (4.33)
that

lim '.xk / ' f .xk / D 0;
!1
which implies by the continuity of the functions ' and f that '.z/ N , i.e., f .z/
N D '.f .z// N D zN and
thus zN solves (4.23) if it is not a vertex.
It remains to consider the case where zN is a vertex, say a1 . Suppose by contradiction that zN D
a1 ¤ xN and select ı > 0 sufficiently small so that (4.31) holds for some s 2 N and that IB.a1 I ı/
does not contain xN and ai for i D 2; : : : ; m. Since xk ! zN D a1 , we can assume without loss of
generality that this sequence is entirely contained in IB.a1 I ı/. For x D xk1 select l1 so that xl1 2
IB.a1 I ı/ and f .xl1 / … IB.a1 I ı/. Choosing now k > l1 and applying Proposition 4.37 give us
l2 > l1 with xl2 2 IB.a1 I ı/ and f .xl2 / … IB.a1 I ı/. Repeating this procedure, find a subsequence
fxl g such that xl 2 IB.a1 I ı/ while f .xl / is not in this ball. Extracting another subsequence
if needed, we can assume that fxl g converges to some point qN satisfying f .q/ N D qN . If qN is not a
vertex, then it solves (4.23) by the above, which is a contradiction because the solution xN is not
in IB.a1 I ı/. us qN is a vertex, which must be a1 since the other vertexes are also not in the ball.
en we have
kf .xl / a1 k
lim D 1;
!1 kxl a1 k

which contradicts (4.32) in Proposition 4.37 and completes the proof of the theorem.
4.5 A GENERALIZED FERMAT-TORRICELLI PROBLEM
ere are several generalized versions of the Fermat-Torricelli problem known in the literature
under different names; see, e.g., our papers [15, 16, 17] and the references therein. Note that
the constrained version of the problem below is also called a “generalized Heron problem” after
Heron from Alexandria who formulated its simplest version in the plane in the 1st century AD.
Following [15, 16], we consider here a generalized Fermat-Torricelli problem defined via the
minimal time functions (3.9) for finitely many closed, convex target sets in Rn with a constraint
set of the same structure. In [6, 11, 29] and their references, the reader can find some particular
versions of this problem and related material on location problems. We also refer the reader to
[15, 17] to nonconvex problems of this type.
Let the target sets ˝i for i D 1; : : : ; m and the constraint set ˝0 be nonempty, closed, con-
vex subsets of Rn . Consider the minimal time functions TF .xI ˝i / D T˝Fi .x/ from (3.9) for ˝i ,
i D 1; : : : ; m, with the constant dynamics defined by a closed, bounded, convex set F Rn that
contains the origin as an interior point. en the generalized Fermat-Torricelli problem is formu-
lated as follows:
m
X
minimize H.x/ WD TF .xI ˝i / subject to x 2 ˝0 : (4.34)
iD1
Our first goal in this section is to establish the existence and uniqueness of solutions to the opti-
mization problem (4.34). We split the proof into several propositions of their own interest. Let
us start with the following property of the Minkowski gauge from (3.11).
Proposition 4.39 Assume that F is a closed, bounded, convex subset of Rn such that 0 2 int F . en
F .x/ D 1 if and only if x 2 bd F .
Proof. Suppose that F .x/ D 1. en there exists a sequence of real numbers ftk g that converges
to t D 1 and such that x 2 tk F for every k . Since F is closed, it yields x 2 F . To complete the
proof of the implication, we show that x … int F . By contradiction, assume that x 2 int F and
choose t > 0 so small that x C tx 2 F . en x 2 .1 C t/ 1 F . is tells us F .x/ .1 C t/ 1 <
1, a contradiction.
Now let x 2 bd F . Since F is closed, we have x 2 F , and hence F .x/ 1. Suppose on the
contrary that WD F .x/ < 1. en there exists a sequence of real numbers ftk g that converges
to with x 2 tk F for every k . en x 2 F D F C .1 /0 F . By Proposition 3.32, the
Minkowski gauge F is continuous, so the set fu 2 Rn j F .u/ < 1g is an open subset of F
containing x . us x 2 int F , which completes the proof.
Recall further that a set Q Rn is said to be strictly convex if for any x; y 2 Q with x ¤ y
and for any t 2 .0; 1/ we have tx C .1 t/y 2 int Q.
4.5. A GENERALIZED FERMAT-TORRICELLI PROBLEM 155
n
Proposition 4.40 Assume that F is a closed, bounded, strictly convex subset of R and such that
0 2 int F . en
F .x C y/ D F .x/ C F .y/ with x; y ¤ 0 (4.35)
if and only if x D y for some > 0.
Proof. Denote ˛ WD F .x/, ˇ WD F .y/ and observe that ˛; ˇ > 0 since F is bounded. If the
linearity property (4.35) holds, then
x C y
F D 1;
˛Cˇ
x ˛ y ˇ
F C D 1:
˛˛Cˇ ˇ˛Cˇ
is allows us to conclude that
x ˛ y ˇ
C 2 bd F:
˛˛Cˇ ˇ˛Cˇ
x y ˛
Since 2 F , 2 F , and 2 .0; 1/, it follows from the strict convexity of F that
˛ ˇ ˛Cˇ
x y F .x/
D ; i.e., x D y with D :
˛ ˇ F .y/
e opposite implication in the proposition is obvious.
Proposition 4.41 Let F be as in Proposition 4.40 and let ˝ be a nonempty, closed, convex subset of
Rn . For any x 2 Rn , the set ˘F .xI ˝/ is a singleton.
Proof. We only need to consider the case where x … ˝ . It is clear that ˘F .xI ˝/ ¤ ;. Suppose
by contradiction that there are u1 ; u2 2 ˘F .xI ˝/ with u1 ¤ u2 . en
TF .xI ˝/ D F .u1 x/ D F .u2 x/ D r > 0;

u1 x u2 x
which implies that 2 bd F and 2 bd F . Since F is strictly convex, we have
r r
1 u1 x 1 u2 x 1 u1 C u2
C D x 2 int F;
2 r 2 r r 2
u1 C u2
and so we get F .u x/ < r D TF .xI ˝/ with u WD 2 ˝ . is contradiction completes
2
the proof of the proposition.
In what follows we identify the generalized projection ˘F .xI ˝/ with its unique element
when both sets F and ˝ are strictly convex.
Proposition 4.42 Let F be a closed, bounded, convex subset of Rn such that 0 2 int F and let ˝ be
a nonempty, closed set ˝ Rn with xN … ˝ . If !N 2 ˘F .xIN ˝/, then !N 2 bd ˝ .
Proof. Suppose that !N 2 int ˝ . en there is > 0 such that IB.!I
N / ˝ . Define
xN !N
z WD !N C 2 IB.!I
N /:
kxN !k
N
and observe from the construction of the Minkowski gauge (3.11) that
xN !N
F .z x/
N D F !N C xN D 1 F .!N x/
N < F .!N N D TF .xI
x/ N ˝/:
kxN !k
N kxN !k
N
is shows that !N must be a boundary point of ˝ .
Figure 4.3: ree sets that satisfy assumptions of Proposition 4.43.
Proposition 4.43 In the setting of problem (4.34) suppose that the sets ˝i for i D 1; : : : ; m are
strictly convex. If for any x; y 2 ˝0 with x ¤ y there is i0 2 f1; : : : ; mg such that
L.x; y/ \ ˝i0 D ;
for the line L.x; y/ connecting x and y , then the function

m
X
H.x/ D TF .xI ˝i /
iD1
defined in (4.34) is strictly convex on the constraint set ˝0 .

Proof. Suppose by contradiction that there are x; y 2 ˝0 , x ¤ y , and t 2 .0; 1/ with

H tx C .1 t/y D t H.x/ C .1 t/H.y/:
Using the convexity of each function TF .xI ˝i / for i D 1; : : : ; m, we have

TF tx C .1 t/yI ˝i D t TF .xI ˝i / C .1 t/TF .yI ˝i /; i D 1; : : : ; m: (4.36)
Suppose further that L.x; y/ \ ˝i0 D ; for some i0 2 f1; : : : ; mg and define
u WD ˘F .xI ˝i0 / and v WD ˘F .yI ˝i0 /:
en it follows from (4.36), eorem 3.33, and the convexity of the Minkowski gauge that
tF .u x/ C .1 t/F .v y/ D t TF .xI ˝i0 / C .1 t/TF .yI ˝i0 /

D TF tx C .1 t/yI ˝i0
F t u C .1 t /v .tx C .1 t/y/
tF .u x/ C .1 t /F .v y/;
which shows that tu C .1 t/v D ˘F .tx C .1 t/yI ˝i0 /. us u D v , since otherwise we have
tu C .1 t/v 2 int ˝i0 , a contradiction. is gives us u D v D ˘F .tx C .1 t /yI ˝i0 /. Apply-
ing (4.36) again tells us that

F u .tx C .1 t/y D F t.u x/ C F .1 t /.u y/ ;
so u x; u y ¤ 0 due to x; y … ˝i0 . By Proposition 4.40, we find > 0 such that t.u x/ D

.1 t /.u y/, which yields
.1 t/
u x D .u y/ with D ¤ 1:
t
is justifies the inclusion
1
uD x y 2 L.x; y/;
1 1
which is a contradiction that completes the proof of the proposition.
Now we are ready to establish the existence and uniqueness theorem for (4.34).
eorem 4.44 In the setting of Proposition 4.43 suppose further that at least one of the sets
among ˝i for i D 0; : : : ; m is bounded. en the generalized Fermat-Torricelli problem (4.34)
has a unique solution.
Proof. By the strict convexity of H from Proposition 4.43, it suffices to show that problem (4.34)
admits an optimal solution. If ˝0 is bounded, then it is compact and the conclusion follows
from the (Weierstrass) eorem 4.10 since the function H in (4.34) Lipschitz continuous by
Corollary 3.34. Consider now the case where one of the sets ˝i for i D 1; : : : ; m is bounded; say
it is ˝1 . In this case we get that for any ˛ 2 R the set
˚ ˇ ˚ ˇ
L˛ D x 2 ˝0 ˇ H.x/ ˛ x 2 ˝0 ˇ TF .xI ˝1 / ˛ ˝0 \ .˝1 ˛F /
is compact. is ensures the existence of a minimizer for (4.34) by eorem 4.10.
Next we completely solve by means of convex analysis and optimization the following sim-
plified version of problem (4.34):
3
X
minimize H.x/ D d.xI ˝i /; x 2 R2 ; (4.37)
iD1
where the target sets ˝i WD IB.ci I ri / for i D 1; 2; 3 are the closed balls of R2 whose centers ci are
supposed to be distinct from each other to avoid triviality.
We first study properties of this problem that enable us to derive verifiable conditions in
terms of the initial data .ci ; ri / ensuring the uniqueness of solutions to (4.37).
Proposition 4.45 e solution set S of (4.37) is a nonempty, compact, convex set in R2 .
Proof. e existence of solutions (4.37) follows from the proof of eorem 4.44. It is also easy to
see that S is closed and bounded. e convexity of S can be deduced from the convexity of the
distance functions d.xI ˝i /.
To proceed, for any u 2 R2 define the index subset
˚ ˇ
I .u/ WD i 2 f1; 2; 3g ˇ u 2 ˝i ; (4.38)
which is widely used in what follows.
Proposition 4.46 Suppose that xN does not belong to any of the sets ˝i , i D 1; 2; 3, i.e., I .x/N D ;.
en xN is an optimal solution to (4.37) if and only if xN solves the classical Fermat-Torricelli problem
(4.23) generated by the centers of the balls ai WD ci for i D 1; 2; 3 in (4.22).
Proof. Assume that xN is a solution to (4.37) with xN … ˝i for i D 1; 2; 3. Choose ı > 0 such that
N ı/ \ ˝i D ; as i D 1; 2; 3 and get following for every x 2 IB.xI
IB.xI N ı/:
3
X 3
X 3
X 3
X
kxN ci k ri D inf H.x/ D H.x/
N H.x/ D kx ci k ri :
iD1 iD1 x2R2 iD1 iD1
is shows that xN is a local minimum of the problem
3
X
minimize F .x/ WD kx ci k; x 2 R2 ; (4.39)
iD1
so it is a global absolute minimum of this problem due to the convexity of F . us xN solves the
classical Fermat-Torricelli problem (4.23) generated by c1 ; c2 ; c3 . e converse implication is also
straightforward.
e next result gives a condition for the uniqueness of solutions to (4.37) in terms of the
index subset defined in (4.38).
Proposition 4.47 Let xN be a solution to (4.37) such that I .x/

N D ; in (4.38). en xN is the unique
solution to (4.37), i.e., S D fxgN .
Proof. Suppose by contradiction that there is u 2 S such that u ¤ xN . Since I .x/ N D ;, we have
xN … ˝i for i D 1; 2; 3. en Proposition 4.46 tells us that xN solves the classical Fermat-Torricelli
problem generated by c1 ; c2 ; c3 . It follows from the convexity of S in Proposition 4.45 that
N u S . us we find v ¤ xN such that v 2 S and v … ˝i for i D 1; 2; 3. It shows that v also
Œx;
solves the classical Fermat-Torricelli problem (4.39). is contradicts the uniqueness result for
(4.39) and hence justifies the claim of the proposition.
e following proposition verifies a visible geometric property of ellipsoids.
Proposition 4.48 Given ˛ > 0 and a; b 2 R2 with a ¤ b , consider the set

˚ ˇ
E WD x 2 R2 ˇ kx ak C kx bk D ˛
and suppose that ka bk < ˛ . For x; y 2 E with x ¤ y , we have

x C y x C y

a C b < ˛;
2 2
xCy
which ensures, in particular, that … E.
2
Proof. It follows from the triangle inequality that

x C y x C y 1

a C b D kx aCy ak C kx bCy bk
2 2 2
1
kx ak C ky ak C kx bk C ky bk D ˛:
2
Suppose by contradiction that
x C y x C y

a C b D ˛; (4.40)
2 2
xCy
which means that 2 E . Since our norm is (always) Euclidean, it gives
2
x a D ˇ.y a/ and x b D .y b/
for some numbers ˇ; 2 .0; 1/ n f1g since x ¤ y . is implies in turn that
1 ˇ 1
aD x y 2 L.x; y/ and b D x y 2 L.x; y/:
1 ˇ 1 ˇ 1 1
xCy
Since ˛ > ka bk, we see that x; y 2 L.a; b/ n Œa; b. To check now that 2 Œa; b, impose
2
xCy
without loss of generality the following order of points: x , , a; b , and y . en
2
x C y x C y

a C b < kx ak C kx bk D ˛;
2 2
xCy
which contradicts (4.40). Since 2 Œa; b, it yields
2
x C y x C y

a C b D ka bk < ˛;
2 2
xCy
which is a contradiction. us … E and the proof is complete.
2
e next result ensures the uniqueness of solutions to (4.37) via the index set (4.38) differ-
ently from Proposition 4.47.
Proposition 4.49 Let S be the solution set for problem (4.37) and let xN 2 S satisfy the conditions:
I .x/
N D fig and xN … Œcj ; ck , where i; j; k are distinct indices in f1; 2; 3g. en S is a singleton, namely
N .
S D fxg
Proof. Suppose for definiteness that I .x/

N D f1g. en we have xN 2 ˝1 , xN … ˝2 , xN … ˝3 , and
xN … Œc2 ; c3 . Denote
˛ WD kxN c2 k C kxN c3 k D inf H.x/ C r2 C r3

x2R2
and observe that ˛ > kc2 c3 k. Consider the ellipse

˚ ˇ
E WD x 2 R2 ˇ kx c2 k C kx c3 k D ˛
for which we get xN 2 E \ ˝1 . Arguing by contradiction, suppose that S is not a singleton, i.e.,
there is u 2 S with u ¤ xN . e convexity of S implies that Œx;
N u S . We can choose v 2 Œx;
N u
sufficiently close to but different from xN so that Œx;
N v \ ˝2 D ; and Œx;
N v \ ˝3 D ;. Let us first
show that v … ˝1 . Assuming the contrary gives us d.vI ˝1 / D 0, and hence
H.v/ D inf H.x/ D d.vI ˝2 / C d.vI ˝3 / D kv c2 k r2 C kv c3 k r3 ;

x2R2
xN C v
which implies that v 2 E . It follows from the convexity of S and ˝1 that 2 S \ ˝1 . We
2
xN C v xN C v xN C v
have furthermore that … ˝2 and … ˝3 , which tell us that 2 E . But this
2 2 2
cannot happen due to Proposition 4.48. erefore, v 2 S and I .v/ D ;, which contradict the
result from Proposition 4.47 and thus justify that S is a singleton.
Figure 4.4: A Fermat-Torricelli problem with jI.x/j

N D 2.
Now we verify the expected boundary property of optimal solutions to (4.37).
Proposition 4.50 Let xN 2 S and let I .x/

N D fi; j g f1; 2; 3g. en xN 2 bd .˝i \ ˝j /.
Proof. Assume for definiteness I .x/ N D f1; 2g, i.e., xN 2 ˝1 \ ˝2 and xN … ˝3 . Arguing by con-
tradiction, suppose that xN 2 int .˝1 \ ˝2 /, and therefore there exists ı > 0 such that IB.xI
N ı/
˝1 \ ˝2 . Denote p WD ˘.xI N ˝3 /, WD kxN pk > 0 and take q 2 Œx; N ı/ satisfying the
N p \ IB.x;
condition kq pk < kxN pk D d.xI N ˝3 /. en we get
3
X 3
X
H.q/ D d.qI ˝i / D d.qI ˝3 / kq pk < kxN pk D d.xI
N ˝3 / D N ˝i / D H.x/;
d.xI N
iD1 iD1
which contradicts the fact that xN 2 S and thus completes the proof.
e following result gives us verifiable necessary and sufficient conditions for the solution
uniqueness in (4.37).
Figure 4.5: A generalized Fermat-Torricelli problem with infinite solutions.
eorem 4.51 Let \3iD1 ˝i D ;. en problem (4.37) has more than one solutions if and only
if there is a set Œci ; cj \ ˝k for distinct indices i; j; k 2 f1; 2; 3g that contains a point u 2 R2
belonging to the interior of ˝k and such that the index set I .u/ is a singleton.
Proof. Let u 2 Œc1 ; c2 \ ˝3 , u 2 int ˝3 satisfy the conditions in the “if ” part of the theo-
rem. en u … ˝1 , u … ˝2 and we can choose v ¤ u with v 2 Œc1 ; c2 ], v … ˝1 , v … ˝2 , and
v 2 int ˝3 . is shows that the set I .v/ is a singleton. Let us verify that in this case u is a solu-
tion to (4.37). To proceed, we observe first that
H.u/ D d.uI ˝1 / C d.uI ˝2 / C d.uI ˝3 / D kc1 c2 k r1 r2 :
Choose ı > 0 such that IB.uI ı/ \ ˝1 D ;, IB.uI ı/ \ ˝2 D ;, and IB.uI ı/ ˝3 . For every
x 2 IB.uI ı/, we get
H.x/ D d.xI ˝2 / C d.xI ˝3 / D kx c1 k C kx c2 k r1 r2 kc1 c2 k r1 r2 D H.u/;
which implies that u is a local (and hence global) minimizer of H. Similar arguments show that
v 2 S , so the solution set S is not a singleton.
To justify the “only if ” part, suppose that S is not a singleton and pick two distinct elements
N yN 2 S , which yields Œx;
x; N S by Proposition 4.45. If there is a solution to (4.37) not belonging
N y
to ˝i for every i 2 f1; 2; 3g, then S reduces to a singleton due to Proposition 4.47, a contradiction.
us we can assume by the above that ˝1 contains infinitely many solutions. en its interior
int ˝1 also contains infinitely many solutions by the obvious strict convexity of the ball ˝1 . If
there is such a solution u with I .u/ D f1g, then u 2 int ˝1 and u … ˝2 as well as u … ˝3 . is
implies that u 2 Œc2 ; c3 , since the opposite ensures the solution uniqueness by Proposition 4.49.
Consider the case where the set I .u/ contains two elements for every solution belonging to int ˝1 .
en there are infinitely many solutions belonging to the intersection of these two sets, which is
strictly convex in this case. Hence there is a solution to (4.37) that belongs to the interior of this
intersection, which contradicts Proposition 4.50 and thus completes the proof of the theorem.
Let us now consider yet another particular case of problem (4.34), where the constant dy-
namics F is given by the closed unit ball. It is formulated as
m
X
minimize D.x/ WD d.xI ˝i / subject to x 2 ˝0 : (4.41)
iD1
For x 2 Rn , define the index subsets of f1; : : : ; mg by

˚ ˇ ˚ ˇ
I.x/ WD i 2 f1; : : : ; mg ˇ x 2 ˝i and J.x/ WD i 2 f1; : : : ; mg ˇ x … ˝i : (4.42)
It is obvious that I.x/ [ J.x/ D f1; : : : ; mg and I.x/ \ J.x/ D ; for all x 2 Rn . e next theo-
rem fully characterizes optimal solutions to (4.41) via projections to ˝i .
eorem 4.52 e following assertions hold for problem (4.41):

(i) Let xN 2 ˝0 be a solution to (4.41). For ai .x/ N as i 2 J.x/
N Ai .x/ N , we have
X X
ai .x/
N 2 Ai .x/
N C N.xI N ˝0 /; (4.43)
i2J.x/
N i2I.x/
N
where each set Ai .x/

N , i D 1; : : : ; m, is calculated by
8
ˆ xN ˘.xI N ˝i /
ˆ
< for xN … ˝i ;
d.xI
N ˝i /
Ai .x/
N D (4.44)
ˆ
:̂
N.xIN ˝i / \ IB for xN 2 ˝i :
(ii) Conversely, if xN 2 ˝0 satisfies (4.43), then xN is a solution to (4.41).
Proof. To verify (i), fix an optimal solution xN to problem (4.41) and equivalently describe it as an
optimal solution to the unconstrained optimization problem:
minimize D.x/ C ı.xI ˝/; x 2 Rn : (4.45)
Applying the subdifferential Fermat rule to (4.45), we characterize xN by
X
n
02@ d.I ˝i / C ı.I ˝/ .x/:
N (4.46)
iD1
Since all of the functions d.I ˝i / for i D 1; : : : ; n are convex and continuous, we employ the
subdifferential sum rule of eorem 2.15 to (4.46) and arrive at
X X
02 @d.xI N ˝i / C d.xI
N ˝i / C N.xI
N ˝0 /
i2J.x/
N
X Xi2I.x/
N
D Ai .x/
N C Ai .x/
N C N.xI
N ˝0 /
i2J.x/
N i2I.x/
N
X X
D ai .x/
N C Ai .x/
N C N.xI
N ˝0 /;
i2J.x/
N i2I.x/
N
which justifies the claimed inclusion

X X
ai .x/
N 2 Ai .x/
N C N.xI
N ˝0 /:
i2J.x/
N i2I.x/
N
e converse assertion (ii) is verified similarly by taking into account the “if and only if ” property
of the proof of (i).
Let us specifically examine problem (4.41) in the case where the intersections of the con-
straint set ˝0 with the target sets ˝i are empty.
Corollary 4.53 Consider problem (4.41) with ˝0 \ ˝i D ; for i D 1; : : : ; m. Given xN 2 ˝0 , de-

fine the elements
xN ˘.xIN ˝i /
ai .x/
N WD ¤ 0; i D 1; : : : ; m: (4.47)
d.xI
N ˝i /
en xN 2 ˝0 is an optimal solution to (4.41) if and only if
m
X
ai .x/
N 2 N.xI
N ˝0 /: (4.48)
iD1
Proof. In this case we have that xN … ˝i for all i D 1; : : : ; m, which means that I.x/
N D ; and
N D f1; : : : ; mg. us our statement is a direct consequence of eorem 4.52.
J.x/
For more detailed solution of (4.41), consider its special case of m D 3.
eorem 4.54 Let m D 3 in the framework of eorem 4.52, where the target sets ˝1 ; ˝2 , and
˝3 are pairwise disjoint, and where ˝0 D Rn . e following hold for the optimal solution xN to
(4.41) with the sets ai .x/
N defined by (4.47):
(i) If xN belongs to one of the sets ˝i , say ˝1 , then we have
˝
a2 ; a3 i 1=2 and a2 a3 2 N.xI
N ˝1 /:
(ii) If xN does not belong to any of the three sets ˝1 , ˝2 , and ˝3 , then
˚
hai ; aj i D 1=2 for i ¤ j as i; j 2 1; 2; 3 :
Conversely, if xN satisfies either (i) or (ii), then it is an optimal solution to (4.41).
Proof. To verify (i), we have from eorem 4.52 that xN 2 ˝1 solves (4.41) if and only if
0 2 @d.xI
N ˝1 / C @d.xI
N ˝2 / C @d.xI
N ˝3 /
D @d.xI
N ˝1 / C a2 C a3 D N.xI
N ˝1 / \ IB C a2 C a3 :
us a2 a3 2 N.xI N ˝1 / \ IB , which is equivalent to the validity of a2 a3 2 IB and a2

N ˝1 /. By Lemma 4.31, this is also equivalent to
a3 2 N.xI
˝
a2 ; a3 i 1=2 and a2 a3 2 N.xI N ˝1 /;
which gives (i). To justify (ii), we similarly conclude that in the case where xN … ˝i for i D 1; 2; 3,
this element solves (4.41) if and only if
0 2 @d.xI
N ˝1 / C @d.xI
N ˝2 / C @d.xI
N ˝3 / D a1 C a2 C a3 :
en Lemma 4.32 tells us that the latter is equivalent to

˚
hai ; aj i D 1=2 for i ¤ j as i; j 2 1; 2; 3 :
Conversely, the validity of either (i) or (ii) and the convexity of ˝i ensure that
0 2 @d.xI
N ˝1 / C @d.xI N ˝3 / D @D.x/;
N ˝2 / C @d.xI N
and thus xN is an optimal solution to the problem under consideration.
Figure 4.6: Generalized Fermat-Torricelli problem for three disks.
Example 4.55 Let the sets ˝1 , ˝2 , and ˝3 in problem (4.37) be closed balls in R2 of radius
r D 1 centered at the points .0; 2/, . 2; 0/, and .2; 0/, respectively. We can easily see that the point
.0; 1/ 2 ˝1 satisfies all the conditions in eorem 4.54(i), and hence it is an optimal solution (in
fact the unique one) to this problem.
More generally, consider problem (4.37) in R2 generated by three pairwise disjoint disks
denoted by ˝i , i D 1; 2; 3. Let c1 , c2 , and c3 be the centers of the disks. Assume first that either
the line segment Œc2 ; c3 2 R2 intersects ˝1 , or Œc1 ; c3 intersects ˝2 , or Œc1 ; c2 intersects ˝3 . It
is not hard to check that any point of the intersections (say, of the sets ˝1 and Œc2 ; c3 for defi-
niteness) is an optimal solution to the problem under consideration, since it satisfies the necessary
and sufficient optimality conditions of eorem 4.54(i). Indeed, if xN is such a point, then a2 and
a3 are unit vectors with ha2 ; a3 i D 1 and a2 a3 D 0 2 N.xI N ˝1 /. If the above intersection
assumptions are violated, we define the three points q1 ; q2 , and q3 as follows. Let u and v be in-
tersection points of Œc1 ; c2 and Œc1 ; c3 with the boundary of the disk centered at c1 , respectively.
en we see that there is a unique point q1 on the minor curve generated by u and v such that
the measures of angle c1 q1 c2 and c1 q1 c3 are equal. e points q2 and q3 are defined similarly.
eorem 4.54 yields that whenever the angle c2 q1 c3 , or c1 q2 c3 , or c2 q3 c1 equals or exceeds 120ı
(say, the angle c2 q1 c3 does), then the point xN WD q1 is an optimal solution to the problem under
consideration. Indeed, in this case a2 and a3 from (4.47) are unit vectors with ha2 ; a3 i 1=2
and a2 a3 2 N.xI N ˝1 / because the vector a2 a3 is orthogonal to ˝1 .
If none of these angles equals or exceeds 120ı , there is a point q not belonging to ˝i
as i D 1; 2; 3 such that the angles c1 qc2 D c2 qc3 D c3 qc1 are of 120ı and that q is an optimal
solution to the problem under consideration. Observe that in this case the point q is also a unique
optimal solution to the classical Fermat-Torricelli problem determined by the points c1 ; c2 , and
c3 ; see Figure 4.6.
Finally in this section, we discuss the application of the subgradient algorithm from Sec-
tion 4.3 to the numerical solution of the generalized Fermat-Torricelli problem (4.34).
eorem 4.56 Consider the generalized Fermat-Torricelli problem (4.34) and assume that S ¤
; for its solution set. Picking a sequence f˛k g of positive numbers and a starting point x1 2 ˝0 ,
form the iterative sequence
m
X
xkC1 WD ˘ xk ˛k vi k I ˝0 ; k D 1; 2; : : : ; (4.49)
iD1
with an arbitrary choice of subgradient elements
vik 2 @F .!ik xk / \ N.!i k I ˝i / for some !ik 2 ˘F .xk I ˝i / if xk … ˝i (4.50)
and vik WD 0 otherwise, where ˘F .xI ˝/ stands for the generalized projection operator (3.21),
and where F is the Minkowski gauge (3.11). Define the value sequence
˚ ˇ
Vk WD min H.xj / ˇ j D 1; : : : ; k :
(i) If the sequence f˛k g satisfies conditions (4.13), then fVk g converges to the optimal value V of
problem (4.34). Furthermore, we have the estimate
P
d.x1 I S/2 C `2 1 2
kD1 ˛k
0 Vk V Pk ;
2 iD1 ˛k
where ` 0 is a Lipschitz constant of cost function H./ on Rn .

(ii) If the sequence f˛k g satisfies conditions (4.15), then fVk g converges to the optimal value V
and fxk g in (4.49) converges to an optimal solution xN 2 S of problem (4.34).
Proof. is is a direct application of the projected subgradient algorithm described in Proposi-
tion 4.28 to the constrained optimization problem (4.34) by taking into account the results for
the subgradient method in convex optimization presented in Section 4.3.
Let us specify the above algorithm for the case of the unit ball F D IB in (4.34).
Corollary 4.57 Let F D IB Rn in (4.34), let f˛k g be a sequence of positive numbers, and let
x1 2 ˝0 be a starting point. Consider algorithm (4.49) with @F ./ calculated by
8
ˆ ˘.xk I ˝i / xk
ˆ
< if xk … ˝i ;
d.xk I ˝i /
@F .!ki xk / D
ˆ
:̂
0; if xk 2 ˝i :
en all the conclusions of eorem 4.56 hold for this specification.
Proof. Immediately follows from eorem 4.56 with F .x/ D kxk.
Next we illustrate applications of the subgradient algorithm of eorem 4.56 to solving

the generalized Fermat-Torricelli problem (4.34) formulated via the minimal time function with
non-ball dynamics. Note that to implement the subgradient algorithm (4.49), it is sufficient to
calculate just one subgradient from the set on the right-hand side of (4.50) for each target set. In
the following example we consider for definiteness the dynamics F given by the square Œ 1; 1
Œ 1; 1 in the plane. In this case the corresponding Minkowski gauge (3.11) is given by the formula
˚
F .x1 ; x2 / D max jx1 j; jx2 j ; .x1 ; x2 / 2 R2 : (4.51)
Example 4.58 Consider the unconstrained generalized Fermat-Torricelli problem (4.34) with
the dynamics F D Œ 1; 1 Œ 1; 1 and the square targets ˝i of right position (with sides parallel
to the axes) centered at .ai ; bi / with radii ri (half of the side) as i D 1; : : : ; m. Given a sequence
Figure 4.7: Minimal time function with square dynamic
of positive numbers f˛k g and a starting point x1 , construct the iterations xk D .x1k ; x2k / by
algorithm (4.49). By eorem 3.39, subgradients vi k can be chosen by
8
ˆ
ˆ .1; 0/ if jx2k bi j x1k ai and x1k > ai C ri ;
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ . 1; 0/ if jx2k bi j ai x1k and x1k < ai ri ;
ˆ
ˆ
ˆ
<
vik D .0; 1/ if jx1k ai j x2k bi and x2k > bi C ri ;
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ .0; 1/ if jx1k ai j bi x2k and x2k < bi ri ;
ˆ
ˆ
ˆ
ˆ
:̂
.0; 0/ otherwise.
4.6 A GENERALIZED SYLVESTER PROBLEM

In [31] James Joseph Sylvester proposed the smallest enclosing circle problem: Given a finite number
of points in the plane, find the smallest circle that encloses all of the points. After more than a
century, location problems of this type remain active as they are mathematically beautiful and
have meaningful real-world applications.
is section is devoted to the study by means of convex analysis of the following generalized
version of the Sylvester problem formulated in [18]; see also [19, 20] for further developments.
Let F Rn be a closed, bounded, convex set containing the origin as an interior point. Define
the generalized ball centered at x 2 Rn with radius r > 0 by
GF .xI r/ WD x C rF:
4.6. A GENERALIZED SYLVESTER PROBLEM 169
It is obvious that for F D IB the generalized ball GF .xI r/ reduces to the closed ball of radius
r centered at x . Given now finitely many nonempty, closed, convex sets ˝i for i D 0; : : : ; m,
the generalized Sylvester problem is formulated as follows: Find a point x 2 ˝0 and the smallest
number r 0 such that
GF .xI r/ \ ˝i ¤ ; for all i D 1; : : : ; m:
Observe that when ˝0 D R2 and all the sets ˝i for i D 1; : : : ; m are singletons in R2 , the gen-
eralized Sylvester problem becomes the classical Sylvester enclosing circle problem. e sets ˝i
for i D 1; : : : ; m are often called the target sets.
To study the generalized Sylvester problem, we introduce the cost function
˚ ˇ
T .x/ WD max TF .xI ˝i / ˇ i D 1; : : : ; m
via the minimal time function (3.9) and consider the following optimization problem:
minimize T .x/ subject to x 2 ˝0 : (4.52)
We can show that the generalized ball GF .xI r/ with xN 2 ˝0 is an optimal solution of the gener-
alized Sylvester problem if and only if xN is an optimal solution of the optimization problem (4.52)
with r D T .x/N .
Our first result gives sufficient conditions for the existence of solutions to (4.52).
Tm 4.59 e generalized Sylvester problem (4.52) admits an optimal solution if the set
Proposition
˝0 \ iD1 .˝i ˛F / is bounded for all ˛ 0. is holds, in particular, if there exists an index
i0 2 f0; : : : ; mg such that the set ˝i0 is compact.
Proof. For any ˛ 0, consider the level set

˚ ˇ
L˛ WD x 2 ˝0 ˇ T .x/ ˛
and then show that in fact we have

m
\
L˛ D ˝0 \ Œ .˝i ˛F /: (4.53)
iD1
Indeed, it holds for any nonempty, closed, convex set Q that TF .xI Q/ ˛ if and only if .x C
˛F / \ Q ¤ ;, which is equivalent to x 2 Q ˛F . If x 2 L˛ , then x 2 ˝0 and TF .xI ˝i / ˛
for i D 1; : : : ; m. us x 2 ˝i ˛F for i D 1; : : : ; m, which verifies the inclusion “” in (4.53).
e proof of the opposite inclusion is also straightforward.
It is clear furthermore that the bound-
Tm
edness of the set ˝0 \ iD1 .˝i ˛F / for all ˛ 0 ensures that the level sets for the cost
function in (4.52) are bounded. us it follows from eorem 4.11 that problem (4.52) has an
optimal solution. e last statement is obvious.
Now we continue with the study of the uniqueness of optimal solutions to (4.52) splitting
the proof into several propositions.
Proposition 4.60 Consider a convex function g.x/ 0 on a nonempty, convex set ˝ Rn . en
the function defined by h.x/ WD .g.x//2 is strictly convex on ˝ if and only if g is not constant on any
line segment Œa; b ˝ with a ¤ b .
Proof. Suppose that h is strictly convex on ˝ and suppose on the contrary that g is constant on
a line segment Œa; b with a ¤ b , then h is also constant on this line segment, so it cannot be
strictly convex. Conversely, suppose g is not constant on any segment Œa; b ˝ with a ¤ b .
Observe that h is a convex function on ˝ since it is a composition of a quadratic function, which
is a nondecreasing convex function on Œ0; 1/, and a convex function whose range is a subset
of Œ0; 1/. Let us show that it is strictly convex on ˝ . Suppose by contradiction that there are
t 2 .0; 1/ and x; y 2 ˝ , x ¤ y , with
h.z/ D th.x/ C .1 t /h.y/ with z WD tx C .1 t/y:
en .g.z//2 D t.g.x//2 C .1 t /.g.y//2 . Since g.z/ tg.x/ C .1 t/g.y/, one has
2 2 2
g.z/ t 2 g.x/ C .1 t/2 g.y/ C 2t .1 t /g.x/g.y/;
which readily implies that

2 2 2 2
t g.x/ C .1 t/ g.y/ t 2 g.x/ C .1 t /2 g.y/ C 2t .1 t/g.x/g.y/:
us .g.x/ g.y//2 0, so g.x/ D g.y/. is shows that h.x/ D h.z/ D h.y/ for the point
z 2 .x; y/ defined above. We arrive at a contradiction by showing that h is constant on the line
segment Œz; y. Indeed, fix any u 2 .z; y/ and observe that
h.u/ h.z/ C .1 /h.y/ D h.z/ for some 2 .0; 1/:
On the other hand, since z lies between x and u, we have
h.z/ h.x/ C .1 /h.u/ h.z/ C .1 /h.z/ D h.z/ for some 2 .0; 1/:
It gives us h.z/ D h.x/ C .1 /h.u/ p D h.z/ C .1 /h.u/, and hence h.u/ D h.z/. is
contradicts the assumption that g.u/ D h.u/ is not constant on any line segment Œa; b with
a ¤ b and thus completes the proof of the proposition.
Proposition 4.61 Let both sets F and ˝ in Rn be nonempty, closed, strictly convex, where F contains
the origin as an interior point. en the convex function TF .I ˝/ is not constant on any line segment
Œa; b Rn such that a ¤ b and Œa; b \ ˝ D ;.
Proof. To prove the statement of the proposition, suppose on the contrary that there is Œa; b
Rn with a ¤ b such that Œa; b \ ˝ D ; and TF .xI ˝/ D r for all x 2 Œa; b. Choosing u 2
˘F .aI ˝/ and v 2 ˘F .bI ˝/, we get F .u a/ D F .v b/ D r . Let us verify that u ¤ v . In-
deed, the opposite gives us F .u a/ D F .u b/ D r , which implies by r > 0 that
u a u b
F D F D 1;
r r
u a u b
and thus 2 bd F and 2 bd F . Since F is strictly convex, it yields
r r
1u a 1u b 1 a C b
C D u 2 int F:
2 r 2 r r 2
Hence we arrive at a contradiction by having
a C b a C b
TF I ˝ F u < r:
2 2
To finish the proof of the proposition, fix t 2 .0; 1/ and get

r D TF ta C .1 t/bI ˝ F t u C .1 t/v .ta C .1 t /b/
tF .u a/ C .1 t/F .v b/ D r;
which implies that tu C .1 t/v 2 ˘F .ta C .1 t /bI ˝/. us t u C .1 t /v 2 bd ˝ , which

contradicts the strict convexity of ˝ and completes the proof.
Figure 4.8: Minimal time function for strictly convex sets.
Proposition 4.62 Let ˝ Rn be a nonempty, closed, convex set and let hi W ˝ ! R be nonnegative,
continuous, convex functions for i D 1; : : : ; m. Define
˚
.x/ WD max h1 .x/; : : : ; hm .x/
and suppose that .x/ D r > 0 on Œa; b for some line segment Œa; b ˝ with a ¤ b . en there
exists a line segment Œ˛; ˇ Œa; b with ˛ ¤ ˇ and an index i0 2 f1; : : : ; mg such that we have
hi0 .x/ D r for all x 2 Œ˛; ˇ.
Proof. ere is nothing to prove for m D 1. When m D 2, consider the function

˚
.x/ WD max h1 .x/; h2 .x/ :
en the conclusion is obvious if h1 .x/ D r for all x 2 Œa; b. Otherwise we find x0 2 Œa; b with
h1 .x0 / < r and a subinterval Œ˛; ˇ Œa; b, ˛ ¤ ˇ , such that
h1 .x/ < r for all x 2 Œ˛; ˇ:
us h2 .x/ D r on this subinterval. To proceed further by induction, we assume that the conclu-
sion holds for a positive integer m 2 and show that the conclusion holds for m C 1 functions.
Denote ˚
.x/ WD max h1 .x/; : : : ; hm .x/; hmC1 .x/ :
en we have .x/ D maxfh1 .x/; k1 .x/g with k1 .x/ WD maxfh2 .x/; : : : ; hmC1 .x/g, and the con-
clusion follows from the case of two functions and the induction hypothesis.
Now we are ready to study the uniqueness of an optimal solution to the generalized Sylvester
problem (4.52).
eorem 4.63 In the setting of problem (4.52) suppose further that the sets F and ˝i for
i D 1; : : : ; m are strictly convex. en problem (4.52) has at most one optimal solution if and
only if \m iD0 ˝i contains at most one element.
Proof. Define K.x/ WD .T .x//2 and consider the optimization problem

minimize K.x/ subject to x 2 ˝0 : (4.54)
It is obvious that xN 2 ˝0 is an optimal solution to (4.52) if and only if it is an optimal solution
to problem (4.54). We are going to prove that if \m iD0 ˝i contains at most one element, then K is
strictly convex on ˝0 , and hence problem (4.52) has at most one optimal solution. Indeed, suppose
that K is not strictly convex on ˝0 . By Proposition 4.60, we find a line segment Œa; b ˝0 with
a ¤ b and a number r 0 so that
˚ ˇ
T .x/ D max TF .xI ˝i / ˇ i D 1; : : : ; m D r for all x 2 Œa; b:
T
It is clear that r > 0, since otherwise Œa; b m iD0 ˝i , which contradicts the assumption that
this set contains at most one element. Proposition 4.62 tells us that there exist a line segment
Œ˛; ˇ Œa; b with ˛ ¤ ˇ and an index i0 2 f1; : : : ; mg such that
TF .xI ˝i0 / D r for all x 2 Œ˛; ˇ:

Since r > 0, we have x … ˝i0 for all x 2 Œ˛; ˇ, which contradicts Proposition 4.61.
Conversely, let problem (4.52) have at most one solution. Suppose now that the set \m iD0 ˝i
T
contains more than one element. It is clear that miD0 ˝ i is the solution set of (4.52) with the op-
timal value equal to zero. us this problem has more than one solution, which is a contradiction
that completes the proof of the theorem.
e following corollary gives a sufficient condition for the uniqueness of optimal solutions
to the generalized Sylvester problem (4.52).
Corollary 4.64 In the setting of (4.52) suppose that the sets F and ˝i for i D 1; : : : ; m are strictly
Tm
convex. If iD0 ˝i contains at most one element and if one of the sets among ˝i for i D 0; : : : ; m is
bounded, then problem (4.52) has a unique optimal solution.
T
Proof. Since m iD0 ˝i contains at most one element, problem (4.52) has at most one optimal
solution by eorem 4.63. us it has exactly one solution because the solution set is nonempty
by Proposition 4.59.
Next we consider a special unconstrained case of (4.52) given by
˚ ˇ
minimize M.x/ WD max d.xI ˝i / ˇ i D 1; : : : ; m subject to x 2 Rn ; (4.55)
for nonempty, closed, convex sets ˝i under the assumption that \niD1 ˝i D ;. Problem (4.55)
corresponds to (4.52) with F D IB Rn . Denote the set of active indices for (4.55) by
˚ ˇ
I.x/ WD i 2 f1; : : : ; mg ˇ M.x/ D d.xI ˝i / ; x 2 Rn :
Further, for each x … ˝i , define the singleton
x ˘.xI ˝i /
ai .x/ WD (4.56)
d.xI ˝i /
where the projection ˘.xI ˝i / is unique. We have M.x/ > 0 as i 2 I.x/ due to the assumption
Tm
iD1 ˝i D ;. is ensures that x … ˝i automatically for any i 2 I.x/.
Let us now present necessary and sufficient optimality conditions for (4.55) based on the gen-
eral results of convex analysis and optimization developed above.
eorem 4.65 e element xN 2 Rn is an optimal solution to problem (4.55) under the assump-
tions made if and only if we have the inclusion
˚ ˇ
xN 2 co ˘.xIN ˝i / ˇ i 2 I.x/
N (4.57)
In particular, for the case where ˝i D fai g, i D 1; : : : ; m, problem (4.55) has a unique solution
and xN is the problem generated by ai if and only if
˚ ˇ
xN 2 co ai ˇ i 2 I.x/ N : (4.58)
Proof. It follows from Proposition 2.35 and Proposition 2.54 that the condition
[
0 2 co @d.xIN ˝i /
i2I.x/
N
is necessary and sufficient for optimality of xN in (4.55). Employing now the distance function
subdifferentiation from eorem 2.39 gives us the minimum characterization
˚ ˇ
N ˇ i 2 I.x/
0 2 co ai .x/ N ; (4.59)
via the singletons ai .x/
N calculated in (4.56). It is easy to observe that (4.59) holds if and only if
P
there are i 0 for i 2 I.x/
N such that i2I.x/ N i D 1 and
X xN !N i
0D i with !i WD ˘.xI
N ˝i /:
M.x/
N
i2I.x/
N
is equation is clearly equivalent to

X X
0D i .xN !N i /; or xN D i !N i ;
i2I.x/
N i2I.x/
N
which in turn is equivalent to inclusion (4.57). Finally, for ˝i D fai g as i D 1; : : : ; m, the latter
inclusion obviously reduces to condition (4.58). In this case the problem has a unique optimal
solution by Corollary 4.64.
To formulate the following result, we say that the ball IB.xI
N r/ touches the set ˝i if the in-
tersection set ˝i \ IB.xI
N r/ is a singleton.
Corollary 4.66 Consider problem (4.55) under the assumptions above. en any smallest intersecting
ball touches at least two target sets among ˝i , i D 1; : : : ; m.
Proof. Let us first verify that there are at least two indices in I.x/
N provided that xN is a solution to
(4.55). Suppose the contrary, i.e., I.x/
N D fi0 g. en it follows from (4.57) that
xN 2 ˘.xI N ˝i0 / D M.x/
N ˝i0 / and d.xI N > 0;
a contradiction. Next we show that IB.xIN r/ with r WD M.x/
N touches ˝i when i 2 I.x/
N . Indeed,
if on the contrary there are u; v 2 ˝i such that
N r/ \ ˝i with u ¤ v;
u; v 2 IB.xI
en, by taking into account that d.xI
N ˝i / D r , we get
ku xk N ˝i / and kv
N r D d.xI xk
N r D d.xI
N ˝i /;
so u; v 2 ˘.xI
N ˝i /. is contradicts the fact that the projector ˘.xI
N ˝i / is a singleton and thus
completes the proof of the corollary.
Now we illustrate the results obtained by solving the following two examples.
Example 4.67 Let xN solve the smallest enclosing ball problem generated by ai 2 R2 , i D 1; 2; 3.
Corollary 4.66 tells us that there are either two or three active indices. If I.x/ N consists of
two indices, say I.x/N D f2; 3g, then xN 2 co fa2 ; a3 g and kxN a2 k D kxN a3 k > kxN a1 k by
a2 C a3
eorem 4.65. In this case we have xN D and ha2 a1 ; a3 a1 i < 0. If conversely
2
ha2 a1 ; a3 a1 i < 0, then the angle of the triangle formed by a1 , a2 , and a3 at the vertex
a1 is obtuse; we include here the case where all the three points ai are on a straight line. It is
a C a a2 C a3
2 3
easy to see that I D f2; 3g and that the point xN D satisfies the assumption
2 2
of eorem 4.65 thus solving the problem. Observe that in this case the solution of (4.55) is the
midpoint of the side opposite to the obtuse vertex.
If none of the angles of the triangle formed by a1 ; a2 ; a3 is obtuse, then the active index set
N consists of three elements. In this case xN is the unique point satisfying
I.x/
˚
xN 2 co a1 ; a2 ; a3 and kxN a1 k D kxN a2 k D kxN a3 k:
In the other words, xN is the center of the circumscribing circle of the triangle.
Figure 4.9: Smallest intersecting ball problem for three disks.
Example 4.68 Let ˝i D IB.ci ; ri / with ri > 0 and i D 1; 2; 3 be disjoint disks in R2 , and let xN
be the unique solution of the corresponding problem (4.55). Consider first the case where one of
the line segments connecting two of the centers intersects the other disks, e.g., the line segment
connecting c2 and c3 intersects ˝1 . Denote u2 WD c2 c3 \ bd ˝2 and u3 WD c2 c3 \ bd ˝3 , and let
xN be the midpoint of u2 u3 . en I.x/ N D f2; 3g and we can apply eorem 4.65 to see that xN is
the solution of the problem.
It remains to consider the case where any line segment connecting two centers of the disks
does not intersect the remaining disk. In this case denote
u1 WD c1 c2 \ bd ˝1 ; v1 WD c1 c3 \ bd ˝1 ;
u2 WD c2 c3 \ bd ˝2 ; v2 WD c2 c1 \ bd ˝2 ;
u3 WD c3 c1 \ bd ˝3 ; v3 WD c3 c2 \ bd ˝3 ;
a1 WD c1 m1 \ bd ˝1 ; a2 D c2 m2 \ bd ˝2 ; a3 WD c3 m3 \ bd ˝3 ;
where m1 is the midpoint of u2 v3 , m2 is the midpoint of u3 v1 , and m3 is the midpoint of u1 v2 .

Suppose that one of the angles u2 a1 v3 , u3 a2 v1 , u1 a3 v2 is greater than or equal to 90ı ; e.g.,
2 2 2
u2 a1 v3 is greater than 90ı . en I.m1 / D f2; 3g, and thus xN D m1 is the unique solution of this
2
problem by eorem 4.65. If now we suppose that all of the aforementioned angles are less than
or equal to 90ı , then I.x/
N D 3 and the smallest disk we are looking for is the unique disk that
touches three other disks. e construction of this disk is the celebrated problem of Apollonius.
Finally in this section, we present and illustrate the subgradient algorithm to solve numeri-
cally the generalized Sylvester problem (4.52).
eorem 4.69 In the setting of problem (4.52) suppose that the solution set S of the problem
is nonempty. Fix x1 2 ˝0 and define the sequences of iterates by

xkC1 WD ˘ xk ˛k vk I ˝0 ; k 2 N;
where f˛k g is a sequence of positive numbers, and where

f0g if xk 2 ˝i ;
vk 2
@F .!k xk / \ N.!k I ˝i / if xk … ˝i ;
where !k 2 ˘F .xk I ˝i / for some i 2 I.xk /. Define the value sequence

˚ ˇ
Vk WD min T .xj / ˇ j D 1; : : : ; k :
(i) If the sequence f˛k g satisfies conditions (4.13), then fVk g converges to the optimal value V of
the problem. Furthermore, we have the estimate
P
d.x1 I S/2 C `2 1 2
kD1 ˛k
0 Vk V P ;
2 kiD1 ˛k
where ` 0 is a Lipschitz constant of cost function T ./ on Rn .

(ii) If the sequence f˛k g satisfies conditions (4.15), then fVk g converges to the optimal value V
and fxk g in (4.49) converges to an optimal solution xN 2 S of the problem.
Proof. Follows from the results of Section 4.3 applied to problem (4.52).
e next examples illustrate the application of the subgradient algorithm of eorem 4.69
in particular situations.
Example 4.70 Consider problem (4.55) in R2 with m target sets given by the squares

˝i D S.!i I ri / WD !1i ri ; !1i C ri !2i ri ; !2i C ri ; i D 1; : : : ; m:
Denote the vertices of the i th square by v1i WD .!1i C ri ; !2i C ri /; v2i WD .!1i ri ; !2i C
ri /; v3i WD .!1i ri ; !2i ri /; v4i WD .!1i C ri ; !2i ri /, and let xk WD .x1k ; x2k /. Fix an in-
dex i 2 I.xk / and then choose the elements vk from eorem 4.69 by
8 xk v1i
ˆ
ˆ if x1k !1i > ri and x2k !2i > ri ;
ˆ
ˆ kxk v1i k
ˆ
ˆ xk v2i
ˆ
ˆ if x1k !1i < ri and x2k !2i > ri ;
ˆ
ˆ
ˆ
ˆ kxk v2i k
ˆ
ˆ xk v3i
ˆ
ˆ if x1k !1i < ri and x2k !2i < ri ;
< kxk v3i k
vk D xk v4i
ˆ if x1k !1i > ri and x2k !2i < ri ;
ˆ
ˆ kxk v4i k
ˆ
ˆ
ˆ
ˆ .0; 1/ if jx1k !1i j ri and x2k !2i > ri ;
ˆ
ˆ
ˆ
ˆ .0; 1/ if jx1k !1i j ri and x2k !2i < ri ;
ˆ
ˆ
ˆ
ˆ .1; 0/ if x1k !1i > ri and jx2k !2i j ri ;
:̂
. 1; 0/ if x1k !1i < ri and jx2k !2i j ri :
Figure 4.10: Euclidean projection to a square.
Now we consider the case of a non-Euclidean norm in the minimal time function.
˚ ˇ
Example 4.71 Consider problem (4.52) with F WD .x1 ; x2 / 2 R2 ˇ jx1 j C jx2 j 1 and a
nonempty, closed, convex target set ˝ . e generalized ball GF .xIN t/ centered at xN D .xN 1 ; xN 2 /
with radius t > 0 has the following diamond shape:
˚ ˇ
N t/ D .x1 ; x2 / 2 R2 ˇ jx1 xN 1 j C jx2 xN 2 j t :
GF .xI
e distance from xN to ˝ and the corresponding projection are given by

˚ ˇ
N ˝/ D min t 0 ˇ GF .xI
TF .xI N t/ \ ˝ ¤ ; ; (4.60)
˘F .xI N t / \ ˝ with t D TF .xI

N ˝/ D GF .xI N ˝/:
Let ˝i be the squares S.!i I ri /, i D 1; : : : ; m, given in Example 4.70. en problem (4.55) can be
interpreted as follows: Find a set of the diamond shape in R2 that intersects all m given squares.
Using the same notation as in Example 4.70, we can see that the elements vk in eorem 4.69
are chosen by
8
ˆ
ˆ .1; 1/ if x1k !1i > ri and x2k !2i > ri ;
ˆ
ˆ
ˆ
ˆ . 1; 1/ if x1k !1i < ri and x2k !2i > ri ;
ˆ
ˆ
ˆ . 1; 1/ if x1k !1i < ri and x2k !2i < ri ;
ˆ
ˆ
<
.1; 1/ if x1k !1i > ri and x2k !2i < ri ;
vk D
ˆ
ˆ .0; 1/ if jx1k !1i j ri and x2k !2i > ri ;
ˆ
ˆ
ˆ
ˆ .0; 1/ if jx1k !1i j ri and x2k !2i < ri ;
ˆ
ˆ
ˆ
ˆ .1; 0/ if x1k !1i > ri and jx2k !2i j ri ;
:̂
. 1; 0/ if x1k !1i < ri and jx2k !2i j ri :
Exercise 4.1 Given a lower semicontinuous function f W Rn ! R, show that its multiplication
.˛f / 7! ˛f .x/ by any scalar ˛ > 0 is lower semicontinuous as well.
Exercise 4.2 Given a function f W Rn ! R, show that its lower semicontinuity is equivalent
to openess, for every 2 R, of the set
˚ ˇ
U WD x 2 Rn ˇ f .x/ > :
Exercise 4.3 Give an example of two lower semicontinuous real-valued functions whose prod-
uct is not lower semicontinuous.
Figure 4.11: Minimal time function with diamond-shape dynamic.
Exercise 4.4 Let f W Rn ! R. Prove the following assertions:

(i) f is upper semicontinuous (u.s.c.) at xN in the sense of Definition 2.89 if and only if f is
lower semicontinuous at this point.
(ii) f is u.s.c. at xN if and only if lim supk!1 f .xk / f .x/
N for every xk ! xN .
(iii) f is u.s.c. at xN if and only if we have lim supx!xN f .x/ f .x/
N .
n n
(iv) f is u.s.c. on R if and only if the upper level set fx 2 R j f .x/ g is closed for any real
number .
(v) f is u.s.c. on Rn if and only if its hypergraphical set (or simply ) hypo f WD
f.x; / 2 Rn R j f .x/g is closed.
Exercise 4.5 Show that a function f W Rn ! R is continuous at xN if and only if it is both lower
semicontinuous and upper semicontinuous at this point.
Exercise 4.6 Prove the equivalence (iii)”(iv) in eorem 4.11.
Exercise 4.7 Consider the real-valued function

m
X
g.x/ WD kx ai k2 ; x 2 Rn ;
iD1
where ai 2 Rn for i D 1; : : : ; m. Show that g is a convex function and has an absolute minimum
a1 C a2 C : : : C am
at the mean point xN D .
m
Exercise 4.8 Consider the problem:
minimize kxk2 subject to Ax D b;
where A is an p n matrix with rank A D p , and where b 2 Rp . Find the unique optimal solution
of this problem.
Exercise 4.9 Give a direct proof of Corollary 4.15 without using eorem 4.14.
Exercise 4.10 Given ai 2 Rn for i D 1; : : : ; m, write MATLAB programs to implement the

subgradient algorithm for minimizing the function f given by:
P
(i) f .x/ WD m kx ai k; x 2 Rn .
PiD1
m
(ii) f .x/ WD iD1 kx ai k1 ; x 2 Rn .
P
(iii) f .x/ WD miD1 kx ai k1 ; x 2 Rn .
Exercise 4.11 Write a MATLAB program to implement the projected subgradient algorithm
for solving the problem:
minimize kxk1 subject to Ax D b;
where A is an p n matrix with rank A D p , and where b 2 Rp .
Exercise 4.12 Consider the optimization problem (4.9) and the iterative procedure (4.10) under
the standing assumptions imposed in Section 4.3. Suppose that the optimal value V is known and
that the sequence f˛k g is given by
8
ˆ
<0 if 0 2 @f .xk /;
˛k WD f .xk / V (4.61)
:̂ 2
otherwise;
kvk k
where vk 2 @f .xk /. Prove that:
(i) f .xk / ! V as k ! 1.
`d.x1 I S/
(ii) 0 Vk V p for k 2 N .
k
Exercise 4.13 Let ˝i for i D 1; : : : ; m be nonempty, closed, convex subsets of Rn having a

nonempty intersection. Use the iterative procedure (4.10) for the function
˚ ˇ
f .x/ WD max d.xI ˝i / ˇ i D 1; : : : ; m
Tm
with the sequence f˛k g given in (4.61) to derive an algorithm for finding xN 2 iD1 ˝i .
Exercise 4.14 Derive the result of Corollary 4.20 from condition (ii) of eorem 4.14.
Exercise 4.15 Formulate and prove a counterpart of eorem 4.27 for the projected subgradient
algorithm (4.20) to solve the constrained optimization problem (4.19).
Exercise 4.16 Prove the converse implication in Lemma 4.31.
Exercise 4.17 Write a MATLAB program to implement the Weiszfeld algorithm.
Exercise 4.18 Let ai 2 R for i D 1; : : : ; m. Solve the problem

m
X
minimize jx ai j; x 2 R:
iD1
Exercise 4.19 Give a simplified proof of eorem 4.51.
Exercise 4.20 Prove assertion (ii) of eorem 4.52.

Exercise 4.21 Give a complete proof of eorem 4.56.
Exercise 4.22 Let ˝i WD Œai ; bi 2 R for i D 1; : : : ; m be m disjoint intervals. Solve the fol-
lowing problem:
m
X
minimize f .x/ WD d.xI ˝i /; x 2 R:
iD1
Exercise 4.23 Write a MATLAB program to implement the subgradient algorithm for solving
the problem:
minimize f .x/; x 2 Rn ;
Pm
where f .x/ WD iD1 d.xI ˝i / with ˝i WD IB.ci I ri / for i D 1; : : : ; m.
Exercise 4.24 Complete the solution of the problem in Example 4.58. Write a MATLAB
program to implement the algorithm.
Exercise 4.25 Give a simplified proof of eorem 4.63 in the case where F is the closed unit
ball of Rn and the target sets are closed Euclidean balls of Rn .
Exercise 4.26 Give a simplified proof of eorem 4.63 in the case where F is the closed unit
ball of Rn .
Exercise 4.27 Write a MATLAB program to implement the subgradient algorithm for solving
the problem:
minimize f .x/; x 2 Rn ;
where f .x/ WD maxfd.xI ˝i / j i D 1; : : : ; mg with ˝i WD IB.ci I ri / for i D 1; : : : ; m. Note that
if xN is an optimal solution to this problem with the optimal value r , then IB.xIN r/ is the smallest
ball that intersects ˝i for i D 1; : : : ; m.
Exercise 4.28 Give a complete proof of eorem 4.69.
183
Solutions and Hints for

Selected Exercises
is final part of the book contains hints for selected exercises. Rather detailed solutions
are given for several problems.
EXERCISES FOR CHAPTER 1
Exercise 1.1. Suppose that ˝1 is bounded. Fix a sequence fxk g ˝1 ˝2 converging to xN .

For every k 2 N , xk is represented as xk D yk zk with yk 2 ˝1 , zk 2 ˝2 . Since ˝1 is closed
and bounded, fyk g has a subsequence fyk` g converging to yN 2 ˝1 . en fzk` g converges to
yN xN 2 ˝2 . us xN 2 ˝1 ˝2 and ˝ is closed. Let ˝1 WD f.x; y/ 2 R2 j y 1=x; x > 0g
and ˝2 WD f.x; y/ 2 R2 j y 1=x; x > 0g . en ˝1 and ˝2 are nonempty, closed, convex
sets while ˝1 ˝2 D f.x; y/ 2 R2 j y > 0 is not closed.
Exercise 1.4. (i) Use Exercise 1.3(ii).

(ii) We only consider the case where m D 2 since the general case is similar. From the
definition of cones, both sets ˝1 and ˝2 contain 0, and so ˝1 [ ˝2 ˝1 C ˝2 . en
co f˝1 [ ˝2 g ˝1 C ˝2 . To verify the opposite inclusion, fix any x D w1 C w2 2 ˝1 C ˝2
with w1 2 ˝1 and w2 2 ˝2 and then write x D Œ.2w1 C 2w2 /=2 2 co f˝1 [ ˝2 g.
Exercise 1.7. Compute the first or second derivative of each function. en apply either
eorem 1.45 or Corollary 1.46.
Exercise 1.8. (i) Use f .x; y/ WD xy; .x; y/ 2 R2 . (ii) Take f1 .x/ WD x and f2 .x/ WD x .
Exercise 1.9. Let f1 .x/ WD x and f2 .x/ WD x 2 , x 2 R.
Exercise 1.12. Use Proposition 1.39. Note that the function f .x/ WD x q is convex on Œ0; 1/.
Exercise 1.16. Define F W Rn !! Rp by F .x/ WD K , x 2 Rn . en gph F is a convex subset of

Rn Rp . e conclusion follows from Proposition 1.50.
184 SOLUTIONS AND HINTS FOR EXERCISES
Exercise 1.18. (i) gph F D C . (ii) Consider the function
( p
1 x2 if jxj 1;
f .x/ WD
1 otherwise:
(iii) Apply Proposition 1.50.

p
Exercise 1.20. (i) Use the definition. (ii) Take f .x/ WD jxj.
Pm Pm
Exercise 1.22. Suppose that v C iD1 i vi D 0 with C iD1 i D 0. If ¤ 0, then
Pm i
iD1 D 1, and so

m
X i ˚
vD vi 2 aff v1 ; : : : ; vm ;

iD1
which is a contradiction. us D 0, and hence i D 0 for i D 1; : : : ; m because the elements

v1 ; : : : ; vm are affinely independent.
Exercise 1.23. Since m ˝ aff ˝ , we have aff m aff ˝ . To prove the opposite inclusion,
it remains to show that ˝ aff m . By contradiction, suppose that there are v 2 ˝ such that
v … aff m . It follows from Exercise 1.22 that the vectors v; v1 ; : : : ; vm are affinely independent
and all of them belong to ˝ . is contradicts the condition dim ˝ D m. e second equality
can also be proved similarly.
Exercise 1.24. (i) We obviously have aff ˝ aff ˝ . Observe that the set aff ˝ is closed and
contains ˝ since aff ˝ D w0 C L, where w0 2 ˝ and L is the linear subspace parallel to aff ˝ ,
which is a closed set. us ˝ aff ˝ and the conclusion follows.
(ii) Choose ı > 0 so that IB.xI
N ı/ \ aff ˝ ˝ . Since xN C t.xN x/ D .1 C t/xN C . t/x is an
affine combination of some elements in aff ˝ D aff ˝ , we have xN C t .xN x/ 2 aff ˝ . us
u WD xN C t.xN x/ 2 ˝ when t is small.
(iii) It follows from (i) that ri ˝ ri ˝ . Fix any x 2 ri ˝ and xN 2 ri ˝ . By (ii) we have
z WD x C t.x x/ N 2 ˝ when t > 0 is small, and so x D z=.1 C t / C t x=.1
N N ri ˝ .
C t / 2 .z; x/
Exercise 1.26. If ˝ 1 D ˝ 2 , then ri ˝ 1 D ri ˝ 2 , and so ri ˝1 D ri ˝2 by Exercise 1.24(iii).
Exercise 1.27. (i) Fix y 2 B.ri ˝/ and find x 2 ri ˝ such that y D B.x/. en pick a vector
yN 2 ri B.˝/ B.˝/ and take xN 2 ˝ with yN D B.x/ N . If x D xN , then y D yN 2 ri B.˝/. Consider
the case where x ¤ xN . Following the solution of Exercise 1.24(iii), find e x 2 ˝ such that x 2 .e
x ; x/
N
and let e x / 2 B.˝/. us y D B.x/ 2 .B.e
y WD B.e x /; B.x//
N D .e N , and so y 2 ri B.˝/. To
y ; y/
complete the proof, it remains to show that B.˝/ D B.ri ˝/ and then use Exercise 1.26 to obtain
SOLUTIONS AND HINTS FOR EXERCISES 185
the inclusion
ri B.˝/ D ri B.ri ˝/ B.ri ˝/:
Since B.ri ˝/ B.˝/, the continuity of B and Proposition 1.73 imply that
B.˝/ B.˝/ D B.ri ˝/ B.ri ˝/:
us we have B.˝/ B.ri ˝/ and complete the proof.

(ii) Consider B.x; y/ WD x y defined on Rn Rn .
Exercise 1.28. (i) Let us show that
aff .epi f / D aff .dom f / R : (4.62)
Fix any .x; / 2 aff .epi f /. en there exist i 2 R and .xi ; i / 2 epi f for i D 1; : : : ; m with
Pm
iD1 i D 1 such that
Xm
.x; / D i .xi ; i /:
iD1
Pm
Since xi 2 dom f for i D 1; : : : ; m, we have x D iD1 i xi 2 aff .dom f /, which yields
aff .epi f / aff .dom f / R :
To verify the opposite inclusion, fix any .x; / 2 aff .dom f / R and find xi 2 dom f and i 2
P Pm
R for i D 1; : : : ; m with m
iD1 i D 1 and x D iD1 i xi . Define further ˛i WD f .xi / 2 R and
Pm
˛ WD iD1 i ˛i for which .xi ; ˛i / 2 epi f and .xi ; ˛i C 1/ 2 epi f . us
m
X
i .xi ; ˛i / D .x; ˛/ 2 aff .epi f /;
iD1
Xm
i .xi ; ˛i C 1/ D .x; ˛ C 1/ 2 aff .epi f /:
iD1
Letting WD ˛ C 1 gives us
.x; / D .x; ˛/ C .1 /.x; ˛ C 1/ 2 aff .epi f /;
which justifies the opposite inclusion in (4.62).

(ii) See the proof of [9, Proposition 1.1.9].
b
ha; xi
Exercise 1.30. (i) ˘.xI ˝/ D x C a.
kak2
(ii) ˘.xI ˝/ D x C AT .AAT / 1 .b Ax/.
(iii) ˘.xI ˝/ D maxfx; 0g (componentwise maximum).
Exercise 2.1. (i) Apply Proposition 2.1 and the fact that tx 2 ˝ for all t 0.
(ii) Use (i) and observe that x 2 ˝ whenever x 2 ˝ .
Exercise 2.2. Observe that the sets ri ˝i for i D 1; 2 are nonempty and convex. By eorem 2.5,
there is 0 ¤ v 2 Rn such that
hv; xi hv; yi for all x 2 ri ˝1 ; y 2 ri ˝2 :
Fix x0 2 ri ˝1 and y0 2 ri ˝2 . It follows from eorem 1.72(ii) that for any x 2 ˝1 , y 2 ˝2 ,

and t 2 .0; 1/ we have that x C t.x0 x/ 2 ri ˝1 and y C t.y0 y/ 2 ri ˝2 . en apply the
inequality above for these elements and let t ! 0C .
Exercise 2.3. We split the solution into four steps.

Step 1: If L is a subspace of Rn that contains a nonempty, convex set ˝ with xN … ˝ and xN 2 L, then
there is 0 ¤ v 2 L such that ˚ ˇ
sup hv; xi ˇ x 2 ˝ < hv; xi:N (4.63)
Proof. By Proposition 2.1, find w 2 Rn such that
˚ ˇ
sup hw; xi ˇ x 2 ˝ < hw; xi:
N (4.64)
Represent Rn as L ˚ L? , where L? WD fu 2 Rn j hu; xi D 0 for all x 2 Lg. en w D u C v

with u 2 L? and v 2 L. Substituting w into (4.64) and taking into account that hu; xi D 0 for
all x 2 ˝ and that hu; xi
N D 0, we get (4.63). Note that (4.63) yields v ¤ 0.
Step 2: Let ˝ Rn be a nonempty, convex set such that 0 … ri ˝ . en the sets ˝ and f0g can be
separated properly.
Proof. In the case where 0 … ˝ , we can separate ˝ and f0g strictly, so the conclusion is obvi-
ous. Suppose now that 0 2 ˝ n ri ˝ . Let L WD aff ˝ D span ˝ , where the latter holds due to
0 2 aff ˝ , which follows from the fact that ˝ L since L is closed. Repeating the proof of
Lemma 2.4, find a sequence fxk g L with xk ! 0 and xk … ˝ for every k . It follows from
Step 1 that there is fvk g L with vk ¤ 0 and
˚ ˇ
sup hvk ; xi ˇ x 2 ˝ < 0:
Dividing both sides by kvk k, we suppose without loss of generality that kvk k D 1 and that vk !
v 2 L with kvk D 1. en ˚ ˇ
sup hv; xi ˇ x 2 ˝ 0:
To finish this proof, it remains to show that there exists x 2 ˝ with hv; xi < 0. By contradiction,
suppose that hv; xi 0 for all x 2 ˝ . en hv; xi D 0 on ˝ . Since v 2 L D aff ˝ , we can rep-
P Pm
resent v as v D m iD1 i wi , where iD1 i D 1 and wi 2 ˝ for i D 1; : : : ; m. is implies that
2
Pm
kvk D hv; vi D iD1 i hv; wi i D 0, which is a contradiction.
Step 3: Let ˝ Rn be a nonempty, convex set. If the sets ˝ and f0g can be separated properly, then
we have 0 … ri ˝ .
Proof. By definition, find 0 ¤ v 2 Rn such that
˚ ˇ
sup hv; xi ˇ x 2 ˝ 0
and then take xN 2 ˝ with hv; xiN < 0. Arguing by contradiction, suppose that 0 2 ri ˝ . en
Exercise 1.24(ii) gives us t > 0 such that 0 C t.0 x/
N D t xN 2 ˝ . us hv; t xi N 0, which
yields hv; xi
N 0, a contradiction.
Step 4: Justifying the statement of Exercise 2.3.
Proof. Define ˝ WD ˝1 ˝2 . By Exercise 1.27(ii) we have ri ˝1 \ ri ˝2 D ; if and only if
0 … ri .˝1 ˝2 / D ri ˝1 ri ˝2 :
en the conclusion follows from Step 2 and Step 3.
Exercise 2.4. Observe that ˝1 \ Œ˝2 .0; / D ; for any > 0.
Exercise 2.5. Employing eorem 2.5 and arguments similar to those in Exercise 2.2, we can
show that there is 0 ¤ v 2 Rn such that
hv; xi hv; yi for all x 2 ˝1 ; y 2 ˝2 :
By Proposition 2.13, choose 0 ¤ v 2 N.xI N ˝2 / and, using the normal cone

N ˝1 / \ Œ N.xI
definition, arrive at the conclusion with v1 WD v , v2 WD v , ˛1 WD hv1 ; xi
N , and ˛2 WD hv2 ; xi:
N
Exercise 2.6. Fix xN 0 and v 2 N.xI

N ˝/. en
hv; x N 0 for all x 2 ˝:

xi
Using this inequality for 0 2 ˝ and 2xN 2 ˝ , we get hv; xi

N D 0. Since ˝ D . 1; 0 : : :
. 1; 0, it follows from Proposition 2.11 that v 0. e opposite inclusion is a direct conse-
quence of the definition.
Exercise 2.8. Suppose that this holds for any k convex sets with k m and m 2. Consider the
sets ˝i for i D 1; : : : ; m C 1 satisfying the qualification condition
h mC1
X i
N ˝i / H) vi D 0 for all i D 1; : : : ; m C 1 :
vi D 0; vi 2 N.xI
iD1
Since 0 2 N.xI
N ˝mC1 /, this condition also holds for the first m sets. us
\ m Xm
N xI
N ˝i D N.xI
N ˝i /:
iD1 iD1
en apply the induction hypothesis to 1 WD \m

iD1 ˝i and 2 WD ˝mC1 .
Exercise 2.10. Since v 2 @f .x/

N , we have
v.x x/
N f .x/ N for all x 2 R :
f .x/
en pick any x1 < xN and use the inequality above.
Exercise 2.12. Define the function

(
f .x/
N if x D x;
N
g.x/ WD f .x/ C ıfxg
N .x/ D
1 otherwise
via the indicator function of fxg

N . en epi g D fxg N 1/, and hence N..x;
N Œf .x/; N epi g/ D
N g.x//I
n n n
R . 1; 0. It follows from Proposition 2.31 that @g.x/ N D R and @ıfxg N D R . Employing
N .x/
the subdifferential sum rule from Corollary 2.45 gives us
Rn D @g.x/ N C Rn ;
N D @f .x/
which ensures that @f .x/

N ¤ ;.
Exercise 2.13. We proceed step by step as in the proof of eorem 2.15. It follows from Exer-
cise 1.28 that ri 1 D ri ˝1 .0; 1/ and
˚ ˇ
ri 2 D .x; / ˇ x 2 ri ˝1 ; < hv; x xi
N :
It is easy to check, arguing by contradiction, that ri 1 \ ri 2 D ;. Applying the separation result

from Exercise 2.3, we find 0 ¤ .w; / 2 Rn R such that
hw; xi C 1 hw; yi C 2 for all .x; 1 / 2 1 ; .y; 2 / 2 2 :
x; e
Moreover, there are .e y; e
1 / 2 1 and .e 2 / 2 2 satisfying
xi C e
hw;e yi C e
1 < hw; e 2 :
As in the proof of eorem 2.15, it remains to verify that < 0. By contradiction, assume that
D 0 and then get
hw; xi hw; yi for all x 2 ˝1 ; y 2 ˝2 ;

hw;ex i < hw; e
y i with e
x 2 ˝1 ; e
y 2 ˝2 :
us the sets ˝1 and ˝2 are properly separated, and hence it follows from Exercise 2.3 that
ri ˝1 \ ri ˝2 D ;. is is a contradiction, which shows that < 0. To complete the proof, we
only need to repeat the proof of Case 2 in eorem 2.15.
Exercise 2.14. Use Exercise 2.13 and proceed as in the proof of eorem 2.44.
Exercise 2.16. Use the same procedure as in the proof of eorem 2.51 and the construction of
the singular subdifferential to obtain
˚ ˇ
@1 .f ı B/.x/ N D A v ˇ v 2 @1 f .y/
N D A @1 f .y/ N : (4.65)
1
Alternatively, observe the equality dom .f ı B/ D B .dom f /. en apply Proposition 2.23
and Corollary 2.53 to complete the proof.
Exercise 2.17. Use the subdifferential formula from eorem 2.61 and then the representation
of the composition from Proposition 1.54.
Exercise 2.19. For any v 2 Rn , we have ˝1 .v/ D kvk1 and ˝2 .v/ D kvk1 .
Exercise 2.21. Use first the representations

˚ ˇ ˚ ˇ
.ı˝ / .v/ D sup hv; xi ı.xI ˝/ ˇ x 2 Rn D sup hv; xi ˇ x 2 ˝ D ˝ .v/; v 2 Rn :
Applying then the Fenchel conjugate one more time gives us

˚ ˇ
.˝ / .x/ D sup hx; vi ˝ .v/ ˇ v 2 Rn D ıco ˝ .x/:
Exercise 2.23. Since f W Rn ! R is convex and positively homogeneous, we have @f .0/ ¤ ; and
f .0/ D 0. is implies in the case where v 2 ˝ D @f .0/ that hv; xi f .x/ for all x 2 Rn , and
so the equality holds when x D 0. us we have
˚ ˇ
f .v/ D sup hv; xi f .x/ ˇ x 2 Rn D 0:
If v … ˝ D f .0/, there is xN 2 Rn satisfying hv; xi N . It gives us by homogeneity that

N > f .x/
˚ ˇ ˚ ˇ
f .v/ D sup hv; xi f .x/ ˇ x 2 Rn sup hv; t xi N N ˇ t > 0 D 1:
f .t x/
Exercise 2.26. f 0 . 1I d / D 1 and f 0 .1I d / D 1.
Exercise 2.27. Write

t2 t2
xN C t2 d D .xN C t1 d / C .1 /xN for fixed t1 < t2 < 0:
t1 t1
Exercise 2.28. (i) For d 2 Rn , consider the function ' defined in Lemma 2.80. en show that
the inequalities < 0 < t imply that './ '.t /.
(ii) It follows from (i) that '.t / '.1/ for all t 2 .0; 1/. Similarly, '. 1/ '.t / for all
t 2 . 1; 0/. is easily yields the conclusion.
Exercise 2.31. Fix xN 2 Rn , define P .xI N , and consider the

N ˝/ WD fw 2 ˝ j kxN wk D ˝ .x/g
following function
fw .x/ WD kx wk for w 2 ˝ and x 2 Rn :
en we get from eorem 2.93 that
˚ ˇ
@˝ .x/ N ˇ w 2 P .xI
N D co @fw .x/ N ˝/ :
xN w
In the case where ˝ is not a singleton we have @fw .x/
N D , and so
˝ .x/
N
xN co P .xI
N ˝/
@˝ .x/
N D :
˝ .x/
N
e case where ˝ is a singleton is trivial.
Exercise 2.33. (i) Fix any v 2 Rn with kvk 1 and any w 2 ˝ . en
hx; vi ˝ .v/ hx; vi hw; vi D hx w; vi kx wk kvk kx wk:
It follows therefore that ˚

sup hx; vi ˝ .v/ d.xI ˝/:
kvk1
To verify the opposite inequality, observe first that it holds trivially if x 2 ˝ . In the case where
x … ˝ , let w WD ˘.xI ˝/ and u WD x w ; thus kuk D d.xI ˝/. Following the proof of Propo-
sition 2.1, we obtain the estimate
hu; zi hu; xi kuk2 for all z 2 ˝:
Divide both sides by kuk and let v WD u=kuk. en ˝ .v/ hv; xi d.xI ˝/, and so
˚
d.xI ˝/ sup hx; vi ˝ .v/ :
kvk1

Exercise 3.1. Suppose that f is Gâteaux differentiable at xN with fG0 .x/
N D v . en
f .xN C td / f .x/
N f .xN C td / f .x/
N
hfG0 .x/;
N vi D lim D lim for all d 2 Rn :
t!0 t t!0C t
0 n
is implies by the definition of the directional derivative that f .xI
N d / D hv; d i for all d 2 R .
0 n
Conversely, suppose that f .xI
N d / D hv; d i for all d 2 R . en
f .xN C t d / f .x/
N
f 0 .xI
N d / D hv; d i D lim for all d 2 Rn :
t !0C t
N d / D hv; d i D hv; d i for all d 2 Rn . us

N d / D f 0 .xI
Moreover, f 0 .xI
f .xN C t d / f .x/
N
f 0 .xI
N d / D hv; d i D lim for all d 2 Rn :
t !0 t
is implies that f is Gâteaux differentiable at xN with fG0 .x/
N D v.
Exercise 3.2. (i) Since the function '.t / WD t 2 is nondecreasing on Œ0; 1/, following the proof
of Corollary 2.62, we can show that if g W Rn ! Œ0; 1/ is a convex function, then @.g 2 /.x/ N D
2g.x/@g.
N N for every xN 2 Rn , where g 2 .x/ WD Œg.x/2 . It follows from eorem 2.39 that
x/
(
f0g if xN 2 ˝;
@f .x/
N D 2d.xIN ˝/@d.xI N ˝/ D ˚
2ŒxN ˘.xI N ˝/ otherwise:
˚
us in any case @f .x/
N D 2ŒxN ˘.xI N ˝/ is a singleton, which implies the differentiability of
f at xN by eorem 3.3 and rf .x/ N D 2ŒxN ˘.xI N ˝/.
(ii) Use eorem 2.39 in the out-of-set case and eorem 3.3.
1
Exercise 3.3. For every u 2 F , the function fu .x/ WD hAx; ui kuk2 is convex, and hence f
2
is a convex function by Proposition 1.43. It follows from the equality hAx; ui D hx; A ui that
rfu .x/ D A u:
1
Fix x 2 Rn and observe that the function gx .u/ WD hAx; ui kuk2 has a unique maximizer on
2
F denoted by u.x/. By eorem 2.93 we have @f .x/ D fA u.x/g, and thus f is differentiable
at x by eorem 3.3. We refer the reader to [22] for more general results and their applications
to optimization.
Exercise 3.5. Consider the set Y WD fd 2 Rn j f 0 .xI N d / D f 0 .xI

N d /g and verify that Y is a
subspace containing the standard orthogonal basis fei j i D 1; : : : ; ng. us V D Rn , which
implies the differentiability of f at xN . Indeed, fix any v 2 @f .x/
N and get by eorem 2.84(iii)
0 n
that f .xI
N d / D hv; d i for all d 2 R . Hence @f .x/ N is a singleton, which ensures that f is
differentiable at xN by eorem 3.3.
Exercise 3.6. Use Corollary 2.37 and follow the proof of eorem 2.40.
Exercise 3.7. An appropriate example is given by the function f .x/ WD x 4 on R.
Exercise 3.10. Fix any sequence fxk g co ˝ . By the Carathéodory theorem, find k;i 0 and
wk;i 2 ˝ for i D 0; : : : ; n with
n
X n
X
xk D k;i wk;i ; k;i D 1 for every k 2 N :
iD0 iD0
en use the boundedness of fk;i gk for i D 0; : : : ; n and the compactness of ˝ to extract a
subsequence of fxk g converging to some element of co ˝ .
Exercise 3.16. By Corollary 2.18 we have N.xI N ˝1 \ ˝2 / D N.xI N ˝2 /. Applying

N ˝1 / C N.xI
eorem 3.16 tells us that
ı
T .xI
N ˝1 \ ˝2 / D N.xI N ˝1 \ ˝2 /
ı ı ı
D N.xIN ˝1 / C N.xI N ˝2 / D N.xI N ˝1 / \ N.xIN ˝2 /
D T .xI
N ˝1 / \ T .xI
N ˝2 /:
Exercise 3.18. Consider the set ˝ WD f.x; y/ 2 R2 j y x 2 g.
Exercise 3.19. Fix any a; b 2 Rn with a ¤ b . It follows from eorem 3.20 that there exists
c 2 .a; b/ such that
f .b/ f .a/ D hv; b ai for some c 2 @f .c/:
en jf .b/ f .a/j kvk kb ak `kb ak.
Exercise 3.21. (i) F1 D f0g RC . (ii) F1 D F . (iii) F1 D RC RC .

(iv) F1 D f.x; y/ j y jxjg.
Exercise 3.23. (i) Apply the definition. (ii) A condition is A1 \ . B1 / D f0g.
Exercise 3.24. Suppose that T˝F .x/ D 0. en there exists a sequence of tk ! 0, tk 0, with
.x C tk F / \ ˝ ¤ ; for every k 2 N . Choose fk 2 F and wk 2 ˝ such that x C tk fk D wk .
Since F is bounded, the sequence fwk g converges to x , and so x 2 ˝ since ˝ is closed. e
converse implication is trivial.
Exercise 3.25. Follow the proof of eorem 3.37 and eorem 3.38.

Exercise 4.2. e set U D Lc is closed by Proposition 4.4.
Exercise 4.3. Consider the l.s.c. function f .x/ WD 0 for x ¤ 0 and f .0/ WD 1. en the product
of f f is not l.s.c.
Pm
Exercise 4.7. Solve the equation rg.x/ D 0, where rg.x/ D iD1 2.x ai /.
Exercise 4.8. Let ˝ WD fx 2 Rn j Ax D bg. en N.xI ˝/ D fAT j 2 Rp g for x 2 ˝ . Corol-

lary 4.15 tells us that x 2 ˝ is an optimal solution to this problem if and only if there exists 2 Rp
such that
2x D AT and Ax D b:
Solving this system of equations yields x D AT .AAT / 1 b .
Exercise 4.11. e MATLAB function sign x gives us a subgradient of the function

f .x/ WD kxk1 for all x 2 R. Considering the set ˝ WD fx 2 Rn j Ax D b , we have
˘.xI ˝/ D x C AT .AAT / 1 .b Ax/, which can be found by minimizing the quadratic
function kx uk2 over u 2 ˝ ; see Exercise 4.8.
Exercise 4.12. If 0 2 @f .xk0 / for some k0 2 N , then ˛k0 D 0, and hence xk D xk0 2 S and
f .xk / D V for all k k0 . us we assume without loss of generality that 0 … @f .xk / as k 2 N .
It follows from the proof of Proposition 4.23 that
k
X k
X
2 2
0 kxkC1 xk
N kx1 xk
N 2 ˛i Œf .xi / VC ˛i2 kvi k2
iD1 iD1
for any xN 2 S . Substituting here the formula for ˛k gives us

k
X k
X
Œf .xi / V 2 Œf .xi / V 2
0 kxkC1 N 2 kx1
xk N 2
xk 2 C
kvi k2 kvi k2
iD1 iD1
k
X Œf .xi / V 2
D kx1 N 2
xk and
kvi k2
iD1
k
X Œf .xi / V 2
kx1 N 2:
xk
kvi k2
iD1
Since kvi k ` for every i , we get
k
X Œf .xi / V 2
kx1 N 2;
xk
`2
iD1

k
X
Œf .xi / V 2 `2 kx1 N 2:
xk (4.66)
iD1
P
en the series 1 kD1 Œf .xk / V 2 is convergent. Hence f .xk / ! V as k ! 1, which justifies
(i). From (4.66) and the fact that Vk f .xj / for i 2 f1; : : : ; kg it follows that
k.Vk V /2 `2 kx1 N 2:
xk
`kx1 xk
N
us 0 Vk V p . Since xN was taken arbitrarily in S , we arrive at (ii).
k
Exercise 4.17. Define the MATLAB function
Pm ai
iD1
kx ai k
f .x/ WD ; x … fa1 ; : : : ; am g
Pm 1
iD1
kx ai k
and f .x/ WD x for x 2 fa1 ; : : : ; am g. Use the FOR loop with a stopping criteria to find
xkC1 WD f .xk /, where x0 is a starting point.
Exercise 4.18. Use the subdifferential Fermat rule. Consider the cases where m is odd and where
m is even. In the first case the problem has a unique optimal solution. In the second case the
problem has infinitely many solutions.
Exercise 4.22. For the interval ˝ WD Œa; b R, the subdifferential formula for the distance func-
tion 2.39 gives us 8
ˆ
ˆ f0g if xN 2 .a; b/;
ˆ
ˆ
ˆ
ˆ
ˆ
<Œ0; 1 if xN D b;
@d.xI
N ˝/ D f1g if xN > b;
ˆ
ˆ
ˆ
ˆ Œ 1; 0 if xN D a;
ˆ
ˆ
:̂f 1g if xN < b:
Exercise 4.23. For ˝ WD IB.cI r/ Rn , a subgradient of the function d.xI ˝/ at xN is
8
<0 if xN 2 ˝;
v.x/
N WD xN c
: if xN … ˝:
kxN ck
en a subgradient of the function f can be found by the subdifferential sum rule of Corol-
lary 2.46. Note that the qualification condition (2.29) holds automatically in this case.
Exercise 4.27. Use the subgradient formula in the hint of Exercise 4.23 together with Propo-
sition 2.54. Note that we can easily find an element i 2 I.x/
N , and then any subgradient of the
function fi .x/ WD d.xI ˝i / at xN belongs to @f .x/
N .
195
Bibliography
[1] H. H. Bauschke and P. L. Combettes, Convex Analysis and Monotone Operator eory in
Hilbert Spaces, Springer, 2011. DOI: 10.1007/978-1-4419-9467-7.
[2] D. P. Bertsekas, A. Nedić and A. E. Ozdaglar, Convex Analysis and Optimization, Athena
Scientific, 2003.
[3] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
DOI: 10.1017/CBO9780511804441. xi
[4] J. M. Borwein and A. S. Lewis, Convex Analysis and Nonlinear Optimization, Springer,
2000. DOI: 10.1007/978-1-4757-9859-3. xi
[5] J. M. Borwein and Q. J. Zhu, Techniques of Variational Analysis, Springer, 2005.
[6] Z. Drezner and H. W. Hamacher (Eds.), Facility Location: Applications and eory, Springer,
2004. 154
[7] G. Colombo and P. R. Wolenski, “Variational analysis for a class of minimal time functions
in Hilbert spaces,” J. Convex Anal., vol. 11, pp. 335–361, 2004.
[8] J.-B. Hiriart-Urruty and C. Lemaréchal, Convex Analysis and Minimization Algorithms,
Vols. I and II, Springer, 1993. DOI: 10.1007/978-3-662-06409-2. xi
[9] J.-B. Hiriart-Urruty and C. Lemaréchal, Fundamentals of Convex Analysis, Springer, 2001.
DOI: 10.1007/978-3-642-56468-0. xi, 185
[10] H. W. Kuhn, “A note on Fermat-Torricelli problem,” Math. Program., vol. 54, pp. 98–107,
1973. DOI: 10.1007/BF01584648. 151
[11] Y. S. Kupitz, H. Martini and M. Spirova, “e Fermat-Torricelli problem, Part I: A dis-
crete gradient-method approach,” Optim. eory Appl., vol. 158, pp. 305–327, 2013. DOI:
10.1007/s10957-013-0266-z. 154
[12] S. R. Lay, Convex Sets and eir Applications, Wiley & Sons, 1982.
[13] G. G. Magaril-Il’yaev and V. M. Tikhomirov, Convex Analysis: eory and Applications,

American Mathematical Society, 2003.
196 BIBLIOGRAPHY
[14] B. S. Mordukhovich, Variational Analayis and Generalized Differentiation, Vols. I and II,
Springer, 2006. DOI: 10.1007/3-540-31246-3. 40, 41, 49
[15] B. S. Mordukhovich and N. M. Nam, “Applications of variational analysis to a general-

ized Fermat-Torricelli problem,” Optim. eory Appl., vol. 148, pp. 431–454, 2011. DOI:
10.1007/s10957-010-9761-7. 154
[16] B. S. Mordukhovich, N. M. Nam and J. Salinas, “Solving a generalized Heron problem

by means of convex analysis,” Amer. Math. Monthly, vol. 119, pp. 87–99, 2012. DOI:
10.4169/amer.math.monthly.119.02.087. 154
[17] B. S. Mordukhovich, N. M. Nam and J. Salinas, “Applications of variational analysis

to a generalized Heron problem,” Applic. Anal., vol. 91, pp. 1915–1942, 2012. DOI:
10.1080/00036811.2011.604849. 154
[18] B. S. Mordukhovich, N. M. Nam and C. Villalobos, “e smallest enclosing ball problem
and the smallest intersecting ball problem: Existence and uniqueness of solutions,” Optim.
Letters, vol. 154, pp. 768–791, 2012. DOI: 10.1007/s11590-012-0483-7. 168
[19] N. M. Nam and N. Hoang, “A generalized Sylvester problem and a generalized Fermat-
Torricelli problem,” J. Convex Anal., vol. 20, pp. 669–687, 2013. 168
[20] N. M. Nam, N. Hoang and N. T. An, “Constructions of solutions to generalized Sylvester

and Fermat-Torricelli problems for Euclidean balls,” Optim. eory Appl., in press. DOI:
10.1007/s10957-013-0366-9. 168
[21] Y. Nesterov, Introductory Lectures on Convex Optimization: A Basic Course, Kluwer Academic
Publishers, 2004. DOI: 10.1007/978-1-4419-8853-9. xi
[22] Y. Nesterov, “Smooth minimization of nonsmooth functions,” Math Program., vol. 103,
pp. 127–152, 2005. DOI: 10.1007/s10107-004-0552-5. 191
[23] J. F. Nash, Noncooperative Games, Doctoral thesis, Princeton University, 1950. 122
[24] J. F. Nash, “Equilibrium points in N-person games,” Proc. Nat. Acad. Sci. vol. 36, pp. 48–49,
1950. DOI: 10.1073/pnas.36.1.48. 122
[25] B. T. Polyak, An Introduction to Optimization, Optimization Solftware, 1987.
[26] R. T. Rockafellar, Convex Analysis, Princeton University Press, 1970. xi, 54
[27] R. T. Rockafellar and R. J-B. Wets, Variational Analysis, Springer, 1998. DOI: 10.1007/978-
3-642-02431-3. 49
BIBLIOGRAPHY 197
[28] R. T. Rockafellar, “Applications of convex variational analysis to Nash equilibrium,” Pro-
ceedings of 7th International Conference on Nonlinear Analysis and Convex Analysis (Busan,
Korea), pp. 173–183, 2011. 123
[29] A. Ruszczynski, Nonlinear Optimization, Princeton University Press, 2006. 154
[30] N. Z. Shor, Minimization Methods for Nondifferentiable Functions, Springer, 1985. DOI:
10.1007/978-3-642-82118-9.
[31] J. J. Sylvester, “A question in the geometry of situation,” Quart. J. Math., vol. 1, p. 79, 1857.
168
[32] E. Weiszfeld, “Sur le point pour lequel la somme des distances de n points donnés est min-
imum,” ToO hoku Mathematics Journal, vol. 43, pp. 355–386, 1937. 150, 151
[33] E. Zeidler, Applied Functional Analysis. Applications to Mathematical Physics, Springer, 1997.
123
199
Authors’ Biographies
BORIS MORDUKHOVICH
Boris Mordukhovich is Distinguished University Professor of Mathematics at Wayne State Uni-
versity holding also Chair Professorship of Mathematics and Statistics at the King Fahd Univer-
sity of Petroleum and Minerals. He has more than 350 publications including several mono-
graphs. Among his best known achievements are the introduction and development of powerful
constructions of generalized differentiation and their applications to broad classes of problems in
variational analysis, optimization, equilibrium, control, economics, engineering, and other fields.
Mordukhovich is a SIAM Fellow, an AMS Fellow, and a recipient of many international awards
and honors including Doctor Honoris Causa degrees from six universities over the world. He is
named a Highly Cited Researcher in Mathematics by the Institute of Scientific Information (ISI).
His research has been supported by continued grants from the National Science Foundations.
Department of Mathematics, Wayne State University, Detroit, MI 48202
boris@math.wayne.edu. Homepage: math.wayne.edu/boris
NGUYEN MAU NAM

Nguyen Mau Nam received his B.S. from Hue University, Vietnam, in 1998 and a Ph.D. from
Wayne State University in 2007. He is currently Assistant Professor of Mathematics at Portland
State University. He has around 30 papers on convex and variational analysis, optimization, and
their applications published in or accepted by high-level mathematical journals. His research is
supported by a collaboration grant from the Simons Foundation.
Fariborz Maseeh Department of Mathematics and Statistics,
Portland State University, Portland, OR 97201
mau.nam.nguyen@pdx.edu. Homepage: web.pdx.edu/mnn3
201
Index
affine combination, 22 descent property, 151

affine hull, 21 dimension of convex set, 23
affine mapping, 7 directional derivative, 82
affine set, 21 distance function, 28
affinely independent, 23 domain, 11
asymptotic cone, 108 domain of multifunction, 18
biconjugacy equality, 80 epigraph, 11, 131

Bolzano-Weierstrass theorem, 3 Euclidean projection, 146
boundary of set, 3 Euclidean projection to set, 29
Brouwer’s fixed point theorem, 120 extended-real-valued function, 4
Carathéodory theorem, 99 Farkas lemma, 105
closed ball, 2 Farkas’ lemma, 99
closure of set, 3 Fenchel conjugate, 77
coderivative, 70 Fenchel-Moreau theorem, 80
coercivity, 134 Fermat rule, 56
compact set, 3 Fermat-Torricelli problem, 146
complement of set, 2 Fréchet differentiability, 55
cone, 33, 97
constrained optimization, 135 GaO teaux differentiability, 95
convex combination, 5 generalized ball, 168
convex combinations, 99 generalized Fermat-Torricelli problem, 154
convex cone, 33 generalized Sylvester problem, 169
convex cone generated by a set, 97 global/absolute minimum, 4
convex conic hull, 97 graph of multifunction, 18
convex function, 11
convex hull, 8, 99 Helly’s theorem, 102
convex optimization, 135 horizon cone, 108, 117
convex set, 5 hypograph, 179
derivative, 55 image, 44
202 INDEX
indicator function, 54 positively homogeneous functions, 34
infimal convolution, 20, 75 projected subgradient method, 145
infimum, 4
interior of set, 2 Radon’s theorem, 102
relative interior, 25
Jensen inequality, 12
set separation, 39, 40
Karush-Kuhn-Tucker condition, 138 set separation, strict, 37
set-valued mapping/multifunction, 18
Lagrange function, 137 simplex, 24
level set, 34, 113, 131 singular subdifferential, 49
Lipschitz continuity, 114 Slater constraint qualification, 138
Lipschitz continuous functions, 49 smallest enclosing circle problem, 168
local minimum, 4 span, 23
locally Lipschitz functions, 49 strictly convex, 147, 155
lower semicontinuous, 178 strictly convex functions, 170
lower semicontinuous function, 129 strictly convex sets, 154
strictly monotone mappings, 125
marginal function, 72 strongly convex, 124
maximum function, 14, 86 strongly monotone mappings, 125
mean value theorem, 17, 106 subadditive functions, 34
minimal time function, 111 subdifferential, 53, 117
Minkowski gauge, 111, 113, 156 subgradient, 53
mixed strategies, 120 subgradient algorithm, 141
monotone mappings, 125 subgradients, 53, 117
Moreau-Rockafellar theorem, 63 subspace, 22
support function, 117
Nash equilibrium, 120
support functions, 75
noncooperative games, 120
supremum, 4
normal cone, 42, 117
supremum function, 16
open ball, 2 tangent cone, 104
optimal value function, 70 two-person games, 120
optimal value/marginal function, 18
upper semicontinuous function, 87
parallel subspace, 22
polar of a set, 103 Weierstrass existence theorem, 5
positive semidefinite matrix, 12 Weiszfeld algorithm, 150, 181

Convex Analysis PDF

Uploaded by

Copyright:

Available Formats

You might also like

Convex Analysis PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Convex Analysis PDF

Uploaded by

Copyright:

Available Formats

Series ISSN: 1938-1743

Series Editor: Steven G. Krantz, Washington University, St. Louis

An Easy Path to Convex Analysis and

AN EASY PATH TO CONVEX ANALYSIS AND APPLICATIONS

An Easy Path to Convex Analysis and Applications

Applications of Aﬃne and Weyl Geometry

Essentials of Applied Mathematics for Engineers and Scientists, Second Edition

Chaotic Maps: Dynamics, Fractals, and Rapid Fluctuations

Matrices in Engineering Problems

e Integral: A Crux for Analysis

Statistics is Easy! Second Edition

Lectures on Financial Mathematics: Discrete Asset Pricing

e Geometry of Walker Manifolds

An Introduction to Multivariable Mathematics

Jordan Canonical Form: Application to Diﬀerential Equations

A Gyrovector Space Approach to Hyperbolic Geometry

An Easy Path to Convex Analysis and Applications

ISBN: 9781627052375 paperback

A Publication in the Morgan & Claypool Publishers series

Nguyen Mau Nam

SYNTHESIS LECTURES ON MATHEMATICS AND STATISTICS #14

e second author dedicates this book to the memory of his parents.

1 Convex Sets and Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

3 Remarkable Consequences of Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

4 Applications to Optimization and Location Problems . . . . . . . . . . . . . . . . . . . 129

Solutions and Hints for Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

Authors’ Biographies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

Boris S. Mordukhovich and Nguyen Mau Nam

Boris S. Mordukhovich and Nguyen Mau Nam

R the real numbers

Convex Sets and Functions

.x1 ; : : : ; xn / C .y1 ; : : : ; yn / D .x1 C y1 ; : : : ; xn C yn /;

e following proposition lists some important properties of the inner product in Rn .

Proposition 1.1 For x; y; z 2 Rn and  2 R, we have:

e Euclidean norm of x D .x1 ; : : : ; xn / 2 Rn is deﬁned by

Proposition 1.2 For any x; y 2 Rn and  2 R, we have:

It is easy to see that IB D IB.0I 1/ and IB.xI

Deﬁnition 1.5 A set ˝  Rn is  if its complement ˝ c D Rn n˝ is open in Rn .

Proposition 1.6 (i) e union of any collection of open sets in Rn is open.

Deﬁnition 1.7 Let fxk g be a sequence in Rn . We say that fxk g  to xN if

Deﬁnition 1.8 Let ˝ be a nonempty subset of Rn . en:

kxk k  r for all k 2 N :

e following important result is known as the Bolzano-Weierstrass theorem.

eorem 1.10 Any bounded sequence in Rn contains a convergent subsequence.

e next concept plays a very signiﬁcant role in analysis and optimization.

e following result is a consequence of the Bolzano-Weierstrass theorem.

eorem 1.12 A subset ˝ of Rn is compact if and only if it is closed and bounded.

For subsets ˝ , ˝1 , and ˝2 of Rn and for  2 R, we deﬁne the operations:

e next proposition can be proved easily.

Proposition 1.13 Let ˝1 and ˝2 be two subsets of Rn .

e following fundamental axiom ensures that these notions are well-deﬁned.

Deﬁnition 1.16 Let f W ˝ ! R be an extended-real-valued function and let xN 2 ˝ with f .x/

jf .x/ N <  whenever kx

We say that f is continuous on ˝ if it is continuous at every point of ˝ .

N for all x 2 IB.xI

We also say that f has a /  at xN relative to ˝ if

e notions of local and global maxima can be deﬁned similarly.

eorem 1.18 Let f W ˝ ! R be a continuous function, where ˝ is a nonempty, compact subset

Proposition 1.1 For x; y; z 2 Rn and 2 R, we have:

Proposition 1.2 For any x; y 2 Rn and 2 R, we have:

Deﬁnition 1.5 A set ˝ Rn is  if its complement ˝ c D Rn n˝ is open in Rn .

kxk k r for all k 2 N :

For subsets ˝ , ˝1 , and ˝2 of Rn and for 2 R, we deﬁne the operations:

jf .x/ N < whenever kx

Note that if a D b , then this interval reduces to a singleton Œa; b D fag.

Deﬁnition 1.19 A subset ˝ of Rn is  if Œa; b ˝ whenever a; b 2 ˝ . Equivalently, ˝ is

and observe that if mC1 D 1, then 1 D 2 D : : : D m D 0, so y D !mC1 2 ˝ . In the case

ai C .1 /bi 2 ˝i for i D 1; 2;

B.x C .1 /y/ D B.x/ C .1 /B.y/ for all x; y 2 Rn and 2 R :

a C .1 /b D B.x/ C .1 /B.y/ D B.x C .1 /y/ 2 B.˝/;

a 2 V ˝ and so a C .1 /b 2 V C .1 /b ˝:

Deﬁnition 1.32 Let f W ˝ ! R be an extended-real-valued function deﬁne on a convex set ˝

e function g in (ii) is convex since for x; y 2 Rn and 2 .0; 1/, we have

eorem 1.36 A function f W Rn ! R is convex if and only if for any numbers i 0 as i D

.x1 ; t1 / C .1 /.x2 ; t2 / D .x1 C .1 /x2 ; t1 C .1 /t2 / 2 epi f;

Proof. Taking any x; y 2 Rn and 2 .0; 1/, we have

Proof. Fix x1 ; x2 2 Rn and 2 .0; 1/. For every i 2 I , we have

Since f 0 .c1 / f 0 .c2 /, adding these equalities yields

roughout this section we assume that .x/ > 1 for every x 2 Rn .

'.xi ; yi / < .xi / C for i D 1; 2:

and therefore y1 C .1 /y2 2 F .x1 C .1 /x2 /. is implies that