Professional Documents
Culture Documents
Convex Analysis PDF
Convex Analysis PDF
Convex Analysis PDF
MORDUKHOVICH • NAM
SYNTHESIS LECTURES ON
COMPUTER MATHEMATICS AND STATISTICS
M
&C Mor gan & Cl aypool Publishers
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
Statistics is Easy!
Dennis Shasha and Manda Wilson
2008
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.
DOI 10.2200/S00554ED1V01Y201312MAS014
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
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.
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
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
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
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.
Acknowledgments
e first author acknowledges continued support by grants from the National Science Foundation
and the second author acknowledges support by the Simons Foundation.
List of Symbols
CHAPTER 1
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:
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
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 :
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.
N r/ are closed in Rn .
It follows that the empty set, the whole space, and any ball IB.xI
lim xk D x:
N
k!1
1.1. PRELIMINARIES 3
is notion allows us to define the following important topological concepts for sets.
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
Definition 1.11 We say that a set ˝ is in Rn if every sequence in ˝ contains a subsequence
converging to some point in ˝ .
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.
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 ˝:
f .x/ f .x/
Proposition 1.20 A subset ˝ of Rn is convex if and only if it contains all convex combinations of its
elements.
6 1. CONVEX SETS AND FUNCTIONS
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
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
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
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 .
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.
e next proposition follows directly from the definition and Proposition 1.25.
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
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
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 ˝ .
10 1. CONVEX SETS AND FUNCTIONS
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.
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.
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
1.3. CONVEX FUNCTIONS 13
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
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 :
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.
1.3. CONVEX FUNCTIONS 15
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 /:
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 /;
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
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 /:
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.
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/ :
Proof. Take x1 ; x2 2 dom , 2 .0; 1/. For any > 0, find yi 2 F .xi / such that
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
and suppose that .f1 ˚ f2 /.x/ > 1 for all x 2 Rn . en f1 ˚ f2 is also convex.
Now we are ready to present the final consequence of Proposition 1.50 in this section that
involves the composition.
Proof. Let F W Rn !
! Rp be a set-valued mapping defined by
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/;
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.
m
X m
X
xD i !i with i D 1; m 2 N;
iD1 iD1
(v) A set ˝ is a .linear/ subspace if and only if ˝ is an affine set containing the origin.
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.
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 ˝ ˝ .
1.4. RELATIVE INTERIORS OF CONVEX SETS 23
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.
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 ˝ .
24 1. CONVEX SETS AND FUNCTIONS
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:
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 1.68 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 ˝ ˝ .
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
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 ˝ .
1.4. RELATIVE INTERIORS OF CONVEX SETS 27
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.
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
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 ˝ .
28 1. CONVEX SETS AND FUNCTIONS
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).
Note that the Lipschitz continuity of f in (1.6) specifies its continuity with a linear rate.
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
d.xI ˝/ kx !k kx yk C ky !k;
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.
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 ˝/.
30 1. CONVEX SETS AND FUNCTIONS
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 ˝/:
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.
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 /;
Since ˝ is closed, this yields !1 C .1 /!2 2 ˝ and so justifies the convexity of ˝ .
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)
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 :
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
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 1.7 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.
34 1. CONVEX SETS AND FUNCTIONS
Exercise 1.10 Verify that the set f.x; t/ 2 Rn R j t kxkg is closed and 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 1.14 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 1.16 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.
(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 1.22 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 .
36 1. CONVEX SETS AND FUNCTIONS
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 1.29 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 .
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.
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)
hv; x wi
N D hxN w;
N x N 0 for any x 2 ˝:
wi
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 :
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 ;
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
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.
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
40 2. SUBDIFFERENTIAL CALCULUS
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
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
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
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 ;:
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.
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
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.
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
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 /:
Proof. To verify (i), fix .v1 ; v2 / 2 N..xN 1 ; xN 2 /I ˝1 ˝2 / and get by the definition that
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
which corresponds to the matrix transposition. e following calculus result describes normals to
solution sets for systems of linear equations.
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;
N A yi D hu;
thu; N vi for all t 2 R; y 2 Rp :
N A .ty/i < hu;
hv; x N D hA y; x
xi xi
N D hy; Ax Axi
N D 0;
Proposition 2.13 Let ˝1 and ˝2 be convex subsets of Rn with xN 2 ˝1 \ ˝2 . en the following
assertions are equivalent:
2.2. NORMALS TO CONVEX SETS 45
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
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
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
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 :
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)
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;
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
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).
hv; x xi
N C 0. f .x//
N D hv; x xi
N 0;
To proceed with the study of Lipschitz continuity of convex functions, we need the follow-
ing two lemmas, which are of their own interest.
Proof. (i) For j j , find t 2 Œ0; 1 with D t . / C .1 t/ . en xN ˙ ei 2 A 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
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.
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
52 2. SUBDIFFERENTIAL CALCULUS
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.
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
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.
Proof. To verify (i)H)(ii), by the continuity of f at xN , for any > 0 find ı > 0 such that
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.
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 .
54 2. SUBDIFFERENTIAL CALCULUS
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)
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.
2.4. SUBGRADIENTS OF CONVEX FUNCTIONS 55
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:
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. 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
2.4. SUBGRADIENTS OF CONVEX FUNCTIONS 57
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)
hrf .x/;
N u xi
N f .u/ N for all u 2 Rn :
f .x/
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
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
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
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 .
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
60 2. SUBDIFFERENTIAL CALCULUS
By Proposition 2.36, we readily have
f .xN C h/ f .x/
N hrf .x/;
N hi 0:
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;
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.
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)
@.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:
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
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
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 :
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
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
hv; x xi
N
. N 0 for all .x; / 2 epi f:
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)
˝˛ f˛g D epi f \ :
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
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. 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)
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
ker A \ @1 f .y/
N D f0g: (2.33)
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:
2.5. BASIC CALCULUS RULES 67
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
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.
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
Proof. Let f .x/ be the maximum function defined in (2.35). We obviously have
m
\
epi f D epi fi :
iD1
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/
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
70 2. SUBDIFFERENTIAL CALCULUS
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).
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
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:
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/
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
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/
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 /;
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
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
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
74 2. SUBDIFFERENTIAL CALCULUS
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
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/ :
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.
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
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 /:
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.
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.
@˝ .x/
N D S.x/:
N
hp; x xi
N ˝ .x/ N whenever x 2 Rn :
˝ .x/ (2.42)
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/
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
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 .
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 .
f .x/ WD ha; xi C b; x 2 Rn :
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/.
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 ;
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)
80 2. SUBDIFFERENTIAL CALCULUS
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
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
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 .
2.8. FENCHEL CONJUGATES 81
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)
.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:
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 :
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
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.
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
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/.
Now we are ready establish a major relationship between the directional derivative and the
subdifferential of an arbitrary convex function.
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
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
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
hvk ; x xi
N g.tk ; x/ g.tk ; x/
N D g.tk ; x/ N for all x 2 Rn
g.tN; x/
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
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 g.tN; x/
is yields f .x/ N , and so tN 2 S.x/
N . Moreover, for any > 0 and large k , we get
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
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
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
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.
2.11. EXERCISES FOR CHAPTER 2 91
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 2.9 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 2.12 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 /?
92 2. SUBDIFFERENTIAL CALCULUS
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
ri .dom f1 / \ ri .dom f2 / ¤ ;:
@.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
N..x;
N y;
N z/I
N ˝1 / \ Œ N..x;
N y;
N zN /I ˝2 /
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
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 :
2.11. EXERCISES FOR CHAPTER 2 93
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 2.22 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.
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/
Prove that ˝ is a convex function and find its subdifferential @˝ .x/ at any x 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
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.
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.
f 0 .xI N d / whenever d 2 Rn ;
N d / hw; d i f 0 .xI
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 .
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
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
3.2. CARATHÉODORY THEOREM AND FARKAS LEMMA 99
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.
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.
100 3. REMARKABLE CONSEQUENCES OF CONVEXITY
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:
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
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˝ .
3.2. CARATHÉODORY THEOREM AND FARKAS LEMMA 101
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.
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
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
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.
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
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
.C ı /ı D cl C:
N ˝/ WD cl Kf˝
T .xI xg
N D cl RC .˝ x/:
N
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
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 ˝/
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.
N ˝/ D cl RC .˝
T .xI x/
N K:
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
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)
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/
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
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
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 .
110 3. REMARKABLE CONSEQUENCES OF CONVEXITY
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 :
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:
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)
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.
112 3. REMARKABLE CONSEQUENCES OF CONVEXITY
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
.x C tk F / \ ˝ ¤ ;; k 2N:
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).
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
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)
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 .
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
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
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)
3.7. MINIMAL TIME FUNCTIONS AND MINKOWSKI GAUGE 115
Proof. Consider first the case where T˝F .x/
D 1. In this case we have
˚ ˇ
t 0 ˇ .x C tF / \ ˝ ¤ ; D ;;
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
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
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)
116 3. REMARKABLE CONSEQUENCES OF CONVEXITY
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)
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:
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
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)
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:
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/
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
hv; !N xi
N D h v; 0 .!N x/i
N F .0/ F .!N x/
N D T˝F .x/
N
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.
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
1 1 1 1
AD and B D
1 1 1 1
122 3. REMARKABLE CONSEQUENCES OF CONVEXITY
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/
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.
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)
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 ˝/:
.p;
N q/
N D ˘..p;
N q/ N B T p/I
N C .Aq; N ˝/: (3.25)
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.
124 3. REMARKABLE CONSEQUENCES OF CONVEXITY
3.10 EXERCISES FOR CHAPTER 3
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
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˝ .
hrf .x/;
N x xi
N < f .x/ N for all x 2 Rn and xN 2 Rn ; x ¤ x:
f .x/ N
Exercise 3.8 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
3.10. EXERCISES FOR CHAPTER 3 125
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 3.11 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 3.16 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
126 3. REMARKABLE CONSEQUENCES OF CONVEXITY
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/
`IB . 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.
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 3.25 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 3.27 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.
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
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.
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)
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/
If xN … ˝ , we have f .x/
N > . By definition, find ı > 0 such that
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
IB.xI
N ı/ . ; C / .epi f /c ;
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.
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.
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.
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
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).
e next result shows, in particular, that there is no difference between local and global
solutions for problems of convex optimization.
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
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)
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 ˝/:
4.2. OPTIMALITY CONDITIONS 137
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;
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.
138 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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:
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
e next theorem provides necessary and sufficient conditions for convex optimization in
the qualified, normal, or Karush-Kuhn-Tucker .KKT/ form.
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
4.2. OPTIMALITY CONDITIONS 139
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
minimize f .x/
subject to gi .x/ 0 for i D 1; : : : ; m;
Ax D b;
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 .
140 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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.
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.
hv; u xi ` ku xk ; u 2 Rn ;
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)
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
142 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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
Proof. Since V is the optimal value in (4.9), we get f .xk / V for any k 2 N . is yields
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
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
sup kxkC1 e
x k < 1:
k2N
is implies that the sequence fxk g is bounded. Furthermore, Proposition 4.26 tells us that
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
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)
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)
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/
148 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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
ka1 C a2 k 1:
ka1 C a2 k2 1;
1
erefore, ha1 ; a2 i . e proof of the converse implication is similar.
2
which readily yields the equality a1 C a2 C a3 D 0 and thus completes the proof.
4.4. THE FERMAT-TORRICELLI PROBLEM 149
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
(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:
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
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
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
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
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 .1 C /kf s 1
.x/ aj k : : : .1 C /s kx aj k:
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.
154 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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
which implies in turn that
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
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
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 ;
Suppose further that L.x; y/ \ ˝i0 D ; for some i0 2 f1; : : : ; mg and define
en it follows from (4.36), eorem 3.33, and the convexity of the Minkowski gauge that
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/ ;
.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.
158 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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).
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)
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
4.5. A GENERALIZED FERMAT-TORRICELLI PROBLEM 159
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).
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.
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
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.
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).
162 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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
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
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
4.5. A GENERALIZED FERMAT-TORRICELLI PROBLEM 163
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
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 .
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:
X
n
02@ d.I ˝i / C ı.I ˝/ .x/:
N (4.46)
iD1
164 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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
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.
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 /:
4.5. A GENERALIZED FERMAT-TORRICELLI PROBLEM 165
(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 :
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 :
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 :
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
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
166 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
.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
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 :
4.5. A GENERALIZED FERMAT-TORRICELLI PROBLEM 167
(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
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.
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
168 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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.
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:
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.
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.
170 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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
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/;
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.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 ;.
4.6. A GENERALIZED SYLVESTER PROBLEM 171
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;
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/
172 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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.
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
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.
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)
174 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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
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.
4.6. A GENERALIZED SYLVESTER PROBLEM 175
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.
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.
176 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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 ;
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;
(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
Proof. Follows from the results of Section 4.3 applied to problem (4.52).
4.6. A GENERALIZED SYLVESTER PROBLEM 177
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 :
Now we consider the case of a non-Euclidean norm in the minimal time function.
178 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
˚ 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
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.
4.7. EXERCISES FOR CHAPTER 4 179
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.
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
180 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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 4.11 Write a MATLAB program to implement the projected subgradient algorithm
for solving the problem:
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 4.14 Derive the result of Corollary 4.20 from condition (ii) of eorem 4.14.
4.7. EXERCISES FOR CHAPTER 4 181
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 4.22 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 ;
182 4. APPLICATIONS TO OPTIMIZATION AND LOCATION PROBLEMS
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 4.29 Complete the solution of the problem in Example 4.70. Write a MATLAB
program to implement the algorithm.
Exercise 4.30 Complete the solution of the problem in Example 4.71. Write a MATLAB
program to implement the algorithm.
183
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 1.12. Use Proposition 1.39. Note that the function f .x/ WD x q is convex on Œ0; 1/.
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 1.27. (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
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
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).
186 SOLUTIONS AND HINTS FOR EXERCISES
EXERCISES FOR CHAPTER 2
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
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
SOLUTIONS AND HINTS FOR EXERCISES 187
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 :
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
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
188 SOLUTIONS AND HINTS FOR EXERCISES
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
v.x x/
N f .x/ N for all x 2 R :
f .x/
Rn D @g.x/ N C Rn ;
N D @f .x/
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 :
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
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 2.23. 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:
Exercise 2.33. (i) Fix any v 2 Rn with kvk 1 and any w 2 ˝ . en
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
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
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
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 3.6. Use Corollary 2.37 and follow the proof of eorem 2.40.
192 SOLUTIONS AND HINTS FOR EXERCISES
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 3.19. 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 3.24. 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 4.12. 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
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
[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.
[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
[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
[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
Index
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