2015 Book Set-ValuedOptimization

Vector Optimization
Akhtar A. Khan
Christiane Tammer
Constantin Zălinescu
Set-valued
Optimization
An Introduction with Applications
Vector Optimization
Series Editor:
Johannes Jahn
University of Erlangen-Nürnberg
Department of Mathematics
Cauerstr. 11
81058 Erlangen
Germany
johannes.jahn@fau.de
Vector Optimization
The series in Vector Optimization contains publications in various fields of opti-

mization with vector-valued objective functions, such as multiobjective optimiza-
tion, multi criteria decision making, set optimization, vector-valued game theory
and border areas to financial mathematics, biosystems, semidefinite programming
and multiobjective control theory. Studies of continuous, discrete, combinatorial
and stochastic multiobjective models in interesting fields of operations research
are also included. The series covers mathematical theory, methods and applications
in economics and engineering. These publications being written in English are
primarily monographs and multiple author works containing current advances in
these fields.
More information about this series at

http://www.springer.com/series/8175
Akhtar A. Khan • Christiane Tammer •
Constantin Zălinescu
Set-valued Optimization
An Introduction with Applications
123
Akhtar A. Khan Christiane Tammer
Rochester Institute of Technology School Halle
of Mathematical Sciences Germany
Rochester
New York
USA
Constantin Zălinescu
University “Al. I. Cuza” Iasi
Faculty of Mathematics
Iasi
Romania
ISSN 1867-8971 ISSN 1867-898X (electronic)

ISBN 978-3-642-54264-0 ISBN 978-3-642-54265-7 (eBook)
DOI 10.1007/978-3-642-54265-7
Springer Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014951215
© Springer-Verlag Berlin Heidelberg 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

With our deepest appreciation, we dedicate
this book to the 80th birthday of
Prof. A. Göpfert
Acknowledgments
We take this opportunity to express our most sincere gratitude to Truong Bao, Mihail
Barbosu, Patricia Clark, Marius Durea, Joydeep Dutta, Rosalind Elster, Franco
Giannessi, Mark Gockenbach, Joachim Gwinner, Alfred Göpfert, Frank Heyde,
Jonas Ide, Johannes Jahn, Kathrin Klamroth, Elisabeth Köbis, Andreas Löhne,
Sophia Maggelakis, Juan Enrique Martìnez-Legaz, Antonino Maugeri, Douglas
Meadows, Dumitru Motreanu, Boris Mordukhovich, Zuhair Nashed, Fabio Raciti,
Miguel Sama, Anita Schöbel, Abul Hasan Siddiqi, Radu Strugariu, Doug Ward, and
Petra Weidner for the joint work and for their support.
We would also like to thank Rosalind Elster, Alfred Göpfert, Miguel Sama, and
Doug Ward for carefully reading the manuscript and many valuable suggestions.
Moreover, we are grateful to Andreas Löhne and Brian Winkler for drawing some
of the figures for this book.
I (AAK) take an immense pleasure in expressing my gratitude and deep love to
my mother Shahida Begum who, since my birth, alone bore all the hardships of
life and made uncountable sacrifices for me. It is also a pleasure to acknowledge
the support of my late uncle and his family. I thank my loving wife Basca and my
adorable daughters Sophie and Amelie for all their love and support. In addition, I
would like to thank Mary Hoeft for her boundless motherly love.
We are extremely thankful to Mr. Christian Rauscher of Springer-Verlag for his
limitless patience and unfailing support throughout this process.
Finally, all the authors are grateful to their coauthors for all their support,
understanding, and hard work over this long and fruitful collaboration. We also
acknowledge that Sects. 2.3, 2.4 (partially), 3.1 and 3.2; Chapters 6, 7, 9, and 10
were written by CZ and the remaining parts were prepared by AAK and CT. The
work of A.A. Khan was partially supported by a grant from the Simons Foundation
vii
viii Acknowledgments
(#210443 to Akhtar Khan). The research of C. Zălinescu was supported partially by

the grant PN-II-ID-PCE-2011-3-0084, CNCS-UEFISCDI, Romania.
Rochester, NY, USA Akhtar A. Khan

Halle, Germany Christiane Tammer
Iasi, Romania Constantin Zălinescu
List of Symbols
R The set of real numbers

R R WD R [ f1g [ fC1g
RC Set of non-negative real numbers
.a; b/; a; bŒ; .a; b; a; b; Œa; b/; Œa; bŒ; Œa; b Intervals in R
P Set of positive real numbers
N, N Set of non-negative integers, N D
N n f0g
X; Y; Z; : : : Real linear spaces or topological
linear spaces
jj jj, jj jj Norm in X , norm in X
Y Y WD Y [ f1g [ fC1g
(extended space)
C; K Cones (in Y , Z)
C C , C WD C C Positive and negative dual cones
of C
C# Quasi-interior of the dual cone C C
BV X , BX Open and closed unit balls in the
n.v.s. X
B.k 0 ; "/ Closed ball with the center k 0 and
the radius "
FWX Y Set-valued mapping
f WX ! Y Single-valued map, function
'A;k 0 'A;k 0 .y/ WD infft 2 R j y 2 tk 0
Ag (y 2 Y )
NX .x/ The set of neighborhoods of x 2 X
in the t.v.s. .X; /
NX The set of balanced neighborhoods
of 0 2 X in the t.v.s. .X; /
ix
x List of Symbols
Im f D ff .x/ j x 2 X g Image (range) of the function f W

X !Y
dom f Domain of the single-valued map
f WX ! Y
dom F Domain of the set-valued map F W
S Y
Dom F Domain of the set-valued map F W
S Y
f 1 .B/ D fx 2 X j f .x/ 2 Bg Inverse image of the set B Y of
the function f W X ! Y
F 1 .B/ WD fx 2 X j F .x/ \ B ¤ ;g Inverse image by F of the set
BF
F C1 .B/ WD fx 2 X j F .x/ Bg Another type of inverse image by
F of the set B F
.r; s/, .r; s, Œr; s/, Œr; s Intervals determined by r; s 2 R,
rs
NO " .xI ˝/ "-Normals
NO .xI ˝/ Fréchet normal cone to ˝ at x 2 ˝
N.xI ˝/ (Basic, limiting, Mordukhovich)
normal cone to ˝ at x
graph F Graph of F
D F .x; y/ (Basic, normal, Mordukhovich)
coderivative of F at .x; y/
epi f Epigraph of f W X ! R
@F .x; y/.y / (Limiting, basic, normal) subdiffer-
ential of F at .x; y/ for y 2 Y
EF W X Y Epigraphical multifunction of F ,
graph EF D epi F
int A, r-int A Interior and relative interior of the
set A
cl A Closure of the set A
bd A, r-bd A Boundary and relative boundary of
the set A
qi A, qri A Quasi-interior and quasi-relative
interior of the set A
span A, spanA Linear and closed linear hulls of the
set A
aff A, affA Affine and closed affine hulls of the
set A
cone A, coneA Conic and closed conic hulls of the
set A
conv A, convA Convex and closed convex hulls of
the set A
core A Algebraic interior (or core) of the
set A
List of Symbols xi
icr A Relative algebraic interior (or

intrinsic core) of the set A
sic A Strong intrinsic core of the set A
A1 Asymptotic cone of the nonempty
set A
RC.A/ Recession cone of the nonempty set
A
P.Y / Power set of Y
PC .Y / Family of all C -proper subsets of Y
Min.A; C / Set of minimal elements of A with
respect to C
Max.A; C / Set of maximal elements of A with
respect to C
.: : :/-PMin.A; C / Set of (. . . ) properly minimal ele-
ments of A with respect to C
WMin.A; C / Set of weakly minimal elements of
A with respect to C
Min.F .S /; / Set of minimal solutions of the set-
valued problem (SP)
QMin.A; C / Q-minimal point of A with respect
to C
T .S; x/ Contingent cone of S at x
C.S; x/ Clarke’s tangent cone of S at x
A.S; x/ Adjacent cone of S at x
DF.x; y/ Contingent derivative of F at
.x; y/ 2 graph .F /
DA F .x; y/ Adjacent derivative of F at .x; y/ 2
graph .F /
DL F .x; y/ Lower Dini derivative of F at
.x; y/ 2 graph .F /
DE F .x; y/ Contingent epiderivative of F at
.x; y/ 2 graph .F /
PC Metric projection mapping
PrX PrX W X Y ! X , PrX .x; y/ D x
f , f Conjugate and biconjugate of f
A Support function of the set A
A Indicator function of the set A
d.x; S /, dS .x/ Distance from x to S
.xn /, fxn g (xn 2 X , n 2 N) Sequences in X
Cl C A WD Cl CC A Upper closure (with respect to C )
of A
inf A Infimum of A Y , .Y; / partially
ordered set
sup A Supremum of A Y , .Y; / par-
tially ordered set
xii List of Symbols
Inf A Infimal set of A Y with respect

to C
F Hyperspace of upper closed sets
I Family of all self-infimal subsets
4F Order relation in F
˚F Addition in F
ˇF Multiplication by non-negative real
numbers in F
4I Order relation in I
˚I Addition in I
ˇI Multiplication by non-negative real
numbers in I
hy ; i, .y ; /, y ./ Linear continuous functional y W
Y ! R
L .X; Y / Set of linear continuous mappings
from X to Y
M .PC .Y /; R/ Set of all functions from PC .Y / to
R
LWY Y Level-set mapping
t.v.s. Real topological vector (linear)
space
n.v.s. Real normed (vector) space
l.c.s. Locally convex topological vector
space
l.s.c. Lower semicontinuous
u.s.c. Upper semicontinuous
w.l.o.g. Without loss of generality
Contents
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1 Motivating Examples . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.2 Book Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
1.3 Useful Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
2 Order Relations and Ordering Cones . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 11
2.1 Order Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 11
2.2 Cone Properties Related to the Topology and the Order .. . . . . . . . . 17
2.3 Convexity Notions for Sets and Set-Valued Maps.. . . . . . . . . . . . . . . . 22
2.4 Solution Concepts in Vector Optimization . . . .. . . . . . . . . . . . . . . . . . . . 28
2.5 Vector Optimization Problems with Variable Ordering
Structure .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
2.6 Solution Concepts in Set-Valued Optimization.. . . . . . . . . . . . . . . . . . . 45
2.6.1 Solution Concepts Based on Vector Approach . . . . . . . . . . 45
2.6.2 Solution Concepts Based on Set Approach . . . . . . . . . . . . . . 48
2.6.3 Solution Concepts Based on Lattice Structure .. . . . . . . . . . 55
2.6.4 The Embedding Approach by Kuroiwa . . . . . . . . . . . . . . . . . . 65
2.6.5 Solution Concepts with Respect to Abstract
Preference Relations . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 67
2.6.6 Set-Valued Optimization Problems
with Variable Ordering Structure .. . . .. . . . . . . . . . . . . . . . . . . . 70
2.6.7 Approximate Solutions of Set-Valued
Optimization Problems . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
2.7 Relationships Between Solution Concepts . . . .. . . . . . . . . . . . . . . . . . . . 74
3 Continuity and Differentiability . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 77
3.1 Continuity Notions for Set-Valued Maps . . . . . .. . . . . . . . . . . . . . . . . . . . 77
3.2 Continuity Properties of Set-Valued Maps Under
Convexity Assumptions . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 90
3.3 Lipschitz Properties for Single-Valued and Set-Valued Maps . . . . 96
3.4 Clarke’s Normal Cone and Subdifferential . . . .. . . . . . . . . . . . . . . . . . . . 102
xiii
xiv Contents
3.5 Limiting Cones and Generalized Differentiability . . . . . . . . . . . . . . . . 103

3.6 Approximate Cones and Generalized Differentiability . . . . . . . . . . . 107
4 Tangent Cones and Tangent Sets. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
4.1 First-Order Tangent Cones . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 110
4.1.1 The Radial Tangent Cone and the Feasible
Tangent Cone . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 110
4.1.2 The Contingent Cone and the Interiorly
Contingent Cone . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 112
4.1.3 The Adjacent Cone and the Interiorly
Adjacent Cone .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 120
4.2 Modified First-Order Tangent Cones . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 123
4.2.1 The Modified Radial and the Modified
Feasible Tangent Cones . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 124
4.2.2 The Modified Contingent and the Modified
Interiorly Contingent Cones . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 124
4.2.3 The Modified Adjacent and the Modified
Interiorly Adjacent Cones. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 126
4.3 Miscellaneous Properties of First-Order Tangent Cones .. . . . . . . . . 129
4.4 First-Order Tangent Cones on Convex Sets . . .. . . . . . . . . . . . . . . . . . . . 132
4.4.1 Connections Among First-Order Tangent
Cones on Convex Sets . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 132
4.4.2 Properties of First-Order Tangent Cones
on Convex Sets . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 137
4.5 First-Order Local Cone Approximation . . . . . . .. . . . . . . . . . . . . . . . . . . . 143
4.6 Convex Subcones of the Contingent Cone .. . . .. . . . . . . . . . . . . . . . . . . . 147
4.7 First-Order Inversion Theorems and Intersection Formulas . . . . . . 156
4.8 Expressions of the Contingent Cone on Some
Constraint Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 161
4.9 Second-Order Tangent Sets. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 169
4.9.1 Second-Order Radial Tangent Set
and Second-Order Feasible Tangent Set . . . . . . . . . . . . . . . . . 170
4.9.2 Second-Order Contingent Set
and Second-Order Interiorly Contingent Set . . . . . . . . . . . . . 170
4.9.3 Second-Order Adjacent Set and
Second-Order Interiorly Adjacent Set . . . . . . . . . . . . . . . . . . . . 173
4.10 Generalized Second-Order Tangent Sets. . . . . . .. . . . . . . . . . . . . . . . . . . . 175
4.11 Second-Order Asymptotic Tangent Cones .. . . .. . . . . . . . . . . . . . . . . . . . 181
4.11.1 Second-Order Asymptotic Feasible Tangent
Cone and Second-Order Asymptotic Radial
Tangent Cone . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 182
4.11.2 Second-Order Asymptotic Contingent Cone
and Second-Order Asymptotic Interiorly
Contingent Cone . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 183
Contents xv
4.11.3 Second-Order Asymptotic Adjacent Cone

and Second-Order Asymptotic Interiorly
Adjacent Cone .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 185
4.12 Miscellaneous Properties of Second-Order Tangent
Sets and Second-Order Asymptotic Tangent Cones .. . . . . . . . . . . . . . 187
4.13 Second-Order Inversion Theorems .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 192
4.14 Expressions of the Second-Order Contingent Set
on Specific Constraints . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 197
4.15 Miscellaneous Second-Order Tangent Cones. .. . . . . . . . . . . . . . . . . . . . 202
4.15.1 Second-Order Tangent Cones of Ledzewicz
and Schaettler.. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 202
4.15.2 Projective Tangent Cones of Second-Order .. . . . . . . . . . . . . 204
4.15.3 Second-Order Tangent Cone of N. Pavel.. . . . . . . . . . . . . . . . 206
4.15.4 Connections Among the Second-Order
Tangent Cones . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 207
4.16 Second-Order Local Approximation .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 207
4.17 Higher-Order Tangent Cones and Tangent Sets . . . . . . . . . . . . . . . . . . . 210
5 Nonconvex Separation Theorems .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 213
5.1 Separating Functions and Examples . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 213
5.2 Nonlinear Separation . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 217
5.2.1 Construction of Scalarizing Functionals . . . . . . . . . . . . . . . . . 217
5.2.2 Properties of Scalarization Functions . . . . . . . . . . . . . . . . . . . . 219
5.2.3 Continuity Properties . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 224
5.2.4 Lipschitz Properties . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 225
5.2.5 The Formula for the Conjugate
and Subdifferential of 'A for A Convex.. . . . . . . . . . . . . . . . . 231
5.3 Scalarizing Functionals by Hiriart-Urruty and Zaffaroni . . . . . . . . . 232
5.4 Characterization of Solutions of Set-Valued
Optimization Problems by Means of Nonlinear
Scalarizing Functionals .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 236
5.4.1 An Extension of the Functional 'A . . .. . . . . . . . . . . . . . . . . . . . 236
5.4.2 Characterization of Solutions of Set-Valued
Optimization Problems with Lower Set Less
Order Relation lC by Scalarization... . . . . . . . . . . . . . . . . . . . 240
5.5 The Extremal Principle . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 244
6 Hahn-Banach Type Theorems . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 249
6.1 The Hahn–Banach–Kantorovich Theorem . . . .. . . . . . . . . . . . . . . . . . . . 250
6.2 Classical Separation Theorems for Convex Sets . . . . . . . . . . . . . . . . . . 258
6.3 The Core Convex Topology . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 261
6.4 Yang’s Generalization of the Hahn–Banach Theorem . . . . . . . . . . . . 264
6.5 A Sufficient Condition for the Convexity of RC A . . . . . . . . . . . . . . . . 271
xvi Contents
7 Conjugates and Subdifferentials .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 275

7.1 The Strong Conjugate and Subdifferential.. . . .. . . . . . . . . . . . . . . . . . . . 275
7.2 The Weak Subdifferential . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 288
7.3 Subdifferentials Corresponding to Henig Proper Efficiency .. . . . . 296
7.4 Exact Formulas for the Subdifferential of the Sum
and the Composition .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 298
8 Duality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 307
8.1 Duality Assertions for Set-Valued Problems Based
on Vector Approach .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 308
8.1.1 Conjugate Duality for Set-Valued Problems
Based on Vector Approach.. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 308
8.1.2 Lagrange Duality for Set-Valued
Optimization Problems Based on Vector Approach . . . . . 313
on Set Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 317
on Lattice Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 322
8.3.1 Conjugate Duality for F -Valued Problems .. . . . . . . . . . . . . 323
8.3.2 Lagrange Duality for I -Valued Problems .. . . . . . . . . . . . . . 326
8.4 Comparison of Different Approaches to Duality
in Set-Valued Optimization .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 338
8.4.1 Lagrange Duality .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 339
8.4.2 Subdifferentials and Stability . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 341
8.4.3 Duality Statements with Operators as Dual Variables .. . 345
9 Existence Results for Minimal Points . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 349
9.1 Preliminary Notions and Results Concerning
Transitive Relations .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 349
9.2 Existence of Minimal Elements with Respect
to Transitive Relations . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 352
9.3 Existence of Minimal Points with Respect to Cones .. . . . . . . . . . . . . 355
9.4 Types of Convex Cones and Compactness
with Respect to Cones . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 360
9.5 Existence of Optimal Solutions for Vector and Set
Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 362
10 Ekeland Variational Principle . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 369
10.1 Preliminary Notions and Results . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 369
10.2 Minimal Points in Product Spaces . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 373
10.3 Minimal Points in Product Spaces of Isac–Tammer’s Type.. . . . . . 381
10.4 Ekeland’s Variational Principles of Ha’s Type.. . . . . . . . . . . . . . . . . . . . 384
10.5 Ekeland’s Variational Principle for Bi-Set-Valued Maps . . . . . . . . . 390
10.6 EVP Type Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 391
10.7 Error Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 394
Contents xvii
11 Derivatives and Epiderivatives of Set-Valued Maps . . . . . . . . . . . . . . . . . . . . 399

11.1 Contingent Derivatives of Set-Valued Maps . . .. . . . . . . . . . . . . . . . . . . . 400
11.1.1 Miscellaneous Graphical Derivatives
of Set-valued Maps . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 407
11.1.2 Convexity Characterization Using
Contingent Derivatives .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 414
11.1.3 Proto-Differentiability,
Semi-Differentiability, and Related
Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 416
11.1.4 Weak Contingent Derivatives of Set-Valued Maps . . . . . . 422
11.1.5 A Lyusternik-Type Theorem Using
Contingent Derivatives .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 426
11.2 Calculus Rules for Derivatives of Set-Valued Maps . . . . . . . . . . . . . . 428
11.2.1 Calculus Rules by a Direct Approach . . . . . . . . . . . . . . . . . . . . 429
11.2.2 Derivative Rules by Using Calculus
of Tangent Cones . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 432
11.3 Contingently C -Absorbing Maps . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 437
11.4 Epiderivatives of Set-Valued Maps.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 445
11.4.1 Contingent Epiderivatives of Set-Valued
Maps with Images in R . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 446
11.4.2 Contingent Epiderivatives in General Spaces .. . . . . . . . . . . 452
11.4.3 Existence Theorems for Contingent Epiderivatives . . . . . 457
11.4.4 Variational Characterization
of the Contingent Epiderivatives . . . . .. . . . . . . . . . . . . . . . . . . . 464
11.5 Generalized Contingent Epiderivatives of Set-Valued Maps . . . . . 470
11.5.1 Existence Theorems for Generalized
Contingent Epiderivatives.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 474
11.5.2 Characterizations of Generalized Contingent
Epiderivatives.. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 478
11.6 Calculus Rules for Contingent Epiderivatives .. . . . . . . . . . . . . . . . . . . . 482
11.7 Second-Order Derivatives of Set-Valued Maps.. . . . . . . . . . . . . . . . . . . 488
11.8 Calculus Rules for Second-Order Contingent Derivatives .. . . . . . . 500
11.9 Second-Order Epiderivatives of Set-Valued Maps . . . . . . . . . . . . . . . . 504
12 Optimality Conditions in Set-Valued Optimization.. . . . . . . . . . . . . . . . . . . . 509
12.1 First-Order Optimality Conditions by the Direct Approach . . . . . . 512
12.2 First-Order Optimality Conditions
by the Dubovitskii-Milyutin Approach . . . . . . . .. . . . . . . . . . . . . . . . . . . . 522
12.2.1 Necessary Optimality Conditions
by the Dubovitskii-Milyutin Approach . . . . . . . . . . . . . . . . . . 523
12.2.2 Inverse Images and Subgradients
of Set-Valued Maps . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 527
12.2.3 Separation Theorems
and the Dubovitskii-Milyutin Lemma .. . . . . . . . . . . . . . . . . . . 534
xviii Contents
12.2.4 Lagrange Multiplier Rules

12.3 Sufficient Optimality Conditions in Set-Valued Optimization .. . . 542
12.3.1 Sufficient Optimality Conditions Under
Convexity and Quasi-Convexity .. . . . .. . . . . . . . . . . . . . . . . . . . 542
Paraconvexity .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 545
Semidifferentiability . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 549
12.4 Second-Order Optimality Conditions in Set-Valued
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 549
12.4.1 Second-Order Optimality Conditions
by the Direct Approach . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 554
12.5 Generalized Dubovitskii-Milyutin Approach
12.5.1 A Separation Theorem for Multiple Closed
and Open Cones . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 559
12.5.2 First-Order Generalized
Dubovitskii-Milyutin Approach . . . . . .. . . . . . . . . . . . . . . . . . . . 562
12.5.3 Second-Order Generalized
Dubovitskii-Milyutin Approach . . . . . .. . . . . . . . . . . . . . . . . . . . 567
12.6 Set-Valued Optimization Problems with a Variable
Order Structure .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 568
12.7 Optimality Conditions for Q-Minimizers
12.7.1 Optimality Conditions for Q-Minimizers
Using Radial Derivatives.. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 572
12.7.2 Optimality Conditions for Q-Minimizers
Using Coderivatives .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 574
12.8 Lagrange Multiplier Rules Based on Limiting Subdifferential . . . 578
12.9 Necessary Conditions for Approximate Solutions
of Set-Valued Optimization Problems . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 591
12.10 Necessary and Sufficient Conditions for Solution
Concepts Based on Set Approach .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 594
12.11 Necessary Conditions for Solution Concepts
with Respect to a General Preference Relation .. . . . . . . . . . . . . . . . . . . 598
12.12 KKT-Points and Corresponding Stability Results . . . . . . . . . . . . . . . . . 600
13 Sensitivity Analysis in Set-Valued Optimization
and Vector Variational Inequalities. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 605
13.1 First Order Sensitivity Analysis in Set-Valued Optimization .. . . . 606
13.2 Second Order Sensitivity Analysis in Set-Valued
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 613
Contents xix
13.3 Sensitivity Analysis in Set-Valued Optimization

Using Coderivatives.. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 623
13.4 Sensitivity Analysis for Vector Variational Inequalities . . . . . . . . . . 634
14 Numerical Methods for Solving Set-Valued
Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 645
14.1 A Newton Method for Set-Valued Maps .. . . . . .. . . . . . . . . . . . . . . . . . . . 645
14.2 An Algorithm to Solve Polyhedral Convex Set-Valued
Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 651
14.2.1 Formulation of the Polyhedral Convex
Set-Valued Optimization Problem .. . .. . . . . . . . . . . . . . . . . . . . 653
14.2.2 An Algorithm for Solving Polyhedral
Convex Set-Valued Optimization Problems.. . . . . . . . . . . . . 655
14.2.3 Properties of the Algorithm .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 658
15 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 663
15.1 Set-Valued Approaches to Duality in Vector Optimization . . . . . . . 663
15.1.1 Fenchel Duality for Vector Optimization
Problems Using Corresponding Results
for F -Valued Problems .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 667
15.1.2 Lagrange Duality for Vector Optimization
Problems Based on Results for I -Valued Problems .. . . 670
15.1.3 Duality Assertions for Linear Vector
Optimization Based on Lattice Approach .. . . . . . . . . . . . . . . 677
15.1.4 Further Set-Valued Approaches to Duality
in Linear Vector Optimization . . . . . . . .. . . . . . . . . . . . . . . . . . . . 682
15.2 Applications in Mathematical Finance. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 696
15.3 Set-Valued Optimization in Welfare Economics . . . . . . . . . . . . . . . . . . 701
15.4 Robustness for Vector-Valued Optimization Problems.. . . . . . . . . . . 706
15.4.1 uC -Robustness .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 710
15.4.2 lC -Robustness .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 720
15.4.3 sC -Robustness .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 722
15.4.4 Algorithms for Solving Special Classes
of Set-Valued Optimization Problems .. . . . . . . . . . . . . . . . . . . 724
Appendix . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 727
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 733
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 759
List of Figures
Fig. 2.1 Cone properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 19

Fig. 2.2 Lower set less order relation 4lC . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 49
Fig. 2.3 Upper set less order relation 4uC . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 49
Fig. 2.4 Certainly less order relation 4cC . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 51
p
Fig. 2.5 Possibly less order relation 4C . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 51
Fig. 2.6 The addition and the ordering in I for C D IR2C . . . . . . . . . . . . . . . . . . . 60
Fig. 2.7 The infimum and supremum in I for C D IR2C . . . . . . . . . . . . . . . . . . . . 61
Fig. 4.1 Feasible tangent cone in Example 4.1.6 . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
Fig. 4.2 Feasible tangent cone in Example 4.1.9 . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
Fig. 4.3 N for Example 4.1.16 .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .
T .S; x/ 118
Fig. 4.4 N for Example 4.1.17 .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .
T .S; x/ 118
Fig. 4.5 Failure of the intersection formula
T .S1 \ S2 ; x/
N D T .S1 ; x/ N \ T2 .S; x/ N in Example 4.1.19 .. . . . . . . . . 120
Fig. 4.6 A.S; x/N for Example 4.1.28 .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 122
Fig. 4.7 A.S; x/N for Example 4.1.30 .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 123
Fig. 4.8 The Clarke’s tangent cone in Example 4.2.13 .. .. . . . . . . . . . . . . . . . . . . . 130
Fig. 4.9 Contingent sets in Example 4.10.5 .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 180
Fig. 4.10 Contingent sets in Example 4.10.7 .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 180
Fig. 4.11 Non-connected contingent set of second-order
for Example 4.10.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 182
Fig. 5.1 Level sets of the function 'A;k 0 from (5.5), where
A D K D R2C and k 0 2 int K hold .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 218
Fig. 5.2 y 2 bd A with 'A .y/ D 1 in Example 5.2.5 . . . . . . . . . . . . . . . . . . . . 222
Fig. 11.1 Contingent derivative of the set-valued map F .. . . . . . . . . . . . . . . . . . . . 400
Fig. 11.2 Contingent epiderivative of the set-valued map F . . . . . . . . . . . . . . . . . . 445
Fig. 11.3 A set-valued map and its contingent epiderivative . . . . . . . . . . . . . . . . . . 448
Fig. 11.4 (i) The set-valued map F. (ii) Functions fi ; i D 0; 1; 2 .. . . . . . . . . . . . 451
xxi
xxii List of Figures
Fig. 11.5 (i) A view of epi .F /. (ii) A contradiction

to .1; 0/ 2 GHe-PMin.G.1/; R2C / . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 471
Fig. 11.6 (i) Example 11.5.3: Minimal points.
(ii) Example 11.5.4: Only weakly minimal points . . . . . . . . . . . . . . . . . . 473
Fig. 12.1 The cone C with int C D ; and the corresponding
cone Ck 0 ;" with int Ck 0 ;" ¤ ; . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 582
Fig. 15.1 Two examples for the set Bc . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 679
Fig. 15.2 The primal and dual values in Example 15.1.22 . . . . . . . . . . . . . . . . . . . . 681
Chapter 1
Introduction
Set-valued optimization is a vibrant and expanding branch of applied mathematics

that deals with optimization problems where the objective map and/or the constraint
maps are set-valued maps acting between abstract spaces. Since the notion of set-
valued maps subsumes single-valued maps, set-valued optimization provides an
important generalization and unification of scalar as well as vector optimization
problems. Therefore, this relatively new discipline has justifiably attracted a great
deal of attention in recent years.
Set-valued optimization benefits enormously from the recently-developed tech-
niques of nonsmooth and variational analysis, and it not only provides elegant
proofs of known results in scalar, vector, and nonsmooth optimization, but it also
gives a new spectrum of powerful techniques to these vital branches of applied
mathematics. Even more importantly, there are many research domains that directly
lead to optimization problems which can most satisfactorily be studied in the unified
framework provided by set-valued optimization. For instance, duality principles
in vector optimization, gap functions for vector variational inequalities, inverse
problems for partial differential equations and variational inequalities, fuzzy opti-
mization, image processing, viability theory, and mathematical economics all lead
to optimization problems that can be conveniently cast as set-valued optimization
problems. Moreover, since set-valued maps appear naturally in many branches of
pure and applied mathematics, set-valued optimization has the evident potential
to remain an important and active research topic in both the near and foreseeable
future.
1.1 Motivating Examples
In the following, we collect some examples that depict some of the natural
occurrences of set-valued optimization problems. We begin with the following
example emerging in game theory:
© Springer-Verlag Berlin Heidelberg 2015 1

A.A. Khan et al., Set-valued Optimization, Vector Optimization,
DOI 10.1007/978-3-642-54265-7__1
2 1 Introduction
Example 1.1.1 (Game Theory).

In this book we deal with set order relations like those defined below. Let C Y
be a proper convex and pointed cone in a linear space Y . Furthermore, let A; B Y
be arbitrarily chosen sets. Then the lower set less order relation lC is defined by
(see Definition 2.6.9)
A lC B ” A C C B ” 8b 2 B 9a 2 A W a C b:
Moreover, we will also use the upper set less order relation uC defined by (see
Definition 2.6.9)
A uC B ” A B C ” 8a 2 A 9b 2 B W a C b:
These set order relations can be applied to problems in game theory. We consider
a vector-valued two-person game. Let M and N be nonempty sets, Y a linear
space and let f W M N ! Y be a single-valued map. Suppose that Player 1
chooses a 2 M and Player 2 next chooses b 2 N . Then f .a; b/ describes the loss
for Player 1.
Consider a set-valued map F W M Y defined by
F .a/ WD ff .a; b/ j b 2 N g:
In the case that Player 2 is cooperative toward Player 1, Player 1 may choose the
lower set less order relation lC and look for lC -minimal solutions a 2 M in the
sense of Definition 2.6.19, i.e., Player 1 is looking for a 2 M with
F .a/ lC F .a/ for some a 2 M H) F .a/ lC F .a/:
This means that the vector-valued game can be formulated as a set-valued optimiza-
tion problem, and Player 1 has to solve the following:
lC minimize F .a/; subject to a 2 M: (SP lC )
When Player 2 is non-cooperative (Player 2 wants Player 1 to lose), Player 1

should choose the upper set less order relation uC and look for uC -minimal
solutions a 2 M in the sense of Definition 2.6.19. Thus, Player 1 is looking for
a 2 M with
F .a/ uC F .a/ for some a 2 M H) F .a/ uC F .a/:
Then the vector-valued game can again be formulated as a set-valued optimization

where Player 1 is now looking to solve the problem:
uC minimize F .a/; subject to a 2 M: (SP uC )

1.1 Motivating Examples 3
Further interesting applications of similar set-valued approaches are given in the

paper by Fernandez, Hinojosa and Puerto [183].
To further illustrate the wide range of set-valued optimization problems, we now
examine their connection with duality principles in vector optimization and gap
functions for vector variational inequalities.
Example 1.1.2 (Duality Principles). Consider the following vector optimization
problem:
minimizeC f .x/ subject to x 2 S D fx 2 M j g.x/ 2 Kg (P)
where M Rn ; C Rp and K Rm are all pointed closed convex cones, and

where f W Rn ! Rp and g W Rn ! Rm are single-valued mappings. Let L be
the family of p m matrices such that K C and let L W M L ! Rp be
the vector-valued Lagrangian given by L.x; / D f .x/ C g.x/: We define a
set-valued map ˚ W L Rp as follows:
˚./ D Min.fL.x; /j x 2 M g; C /:
Then the dual problem associated with (P) is the following set optimization
problem:
maximizeC ˚./ subject to 2 L : (D)
We derive duality statements for vector optimization problems using a set-valued

approach in detail in Sect. 15.1. Specifically, we discuss set-valued approaches to
duality for linear vector optimization and give some applications in mathematical
finance.
Example 1.1.3 (Gap Functions). Let X and Y be normed spaces and let L .X; Y /
be the set of all linear continuous operators from X to Y: Let Q X be a nonempty
closed convex set, let Q0 X be a pointed closed convex cone and let W X ! Y
be a given map.
Consider the following vector variational inequality (VVI, for short) : Find
xN 2 Q such that
h .x/;
N x xi
N … Q0 nf0g for every x 2 Q:
We define a set-valued mapping ˚ W X Y by
˚.x/ D Max.fh .x/;

N x Qi; x 2 Qg; Q0/:
Then the above VVI is equivalent to the following set optimization problem:
minimizeQ0 ˚.x/ subject to x 2 Q:

4 1 Introduction
Robust optimization has become very popular in the last few years, possibly
because of its many applications and its close approximation to real-world scenarios.
Usually when solving an optimization problem, all parameters are given. Instead, it
might happen that only an estimated value or a set of possible parameter values
are available. This is especially true in problems of traffic optimization, scheduling,
portfolio optimization, network flow and network design.
On the other hand, many real-world optimization problems instead require the
minimization of multiple conflicting objectives, e.g. the maximization of expected
return versus the minimization of risk in portfolio optimization, the minimization
of production time versus the minimization of manufacturing equipment cost,
or the maximization of tumor control versus the minimization of normal tissue
complication in radiotherapy treatment design.
Example 1.1.4 (Robust Optimization). Robust optimization can be applied to prob-
lems where a solution is required that hedges against the worst case that may happen.
For example, consider the problem of placing a landing pad for rescue helicopters
at a ski resort. The location of the landing pad should be chosen in such a way that
the flight time to all the resort’s protected slopes is minimized in the worst case,
even though flight times are uncertain due to unknown future weather conditions.
Similarly, in determining the flight schedule of an airline, one would want to be able
to serve as many passengers as possible in a cost-effective manner, even though the
exact number of passengers will not be known when the schedule is fixed.
A concept of robustness for vector-valued optimization problems was introduced
and studied in a paper by Ehrgott, Ide and Schöbel [159]. Using set order relations,
we derive and explore new concepts of robustness in Sect. 15.4.
Example 1.1.5 (Stochastic/Fuzzy Programming).
Another rich source for set optimization problems is the situation in mathematical
programming when the objective and the constraint functions cannot be assigned
exact values, but are allowed to vary in a specified range. This leads to set-valued
data, for example, those found in stochastic and fuzzy programming. In fuzzy
programming, this fact is usually characterized by a membership function, but in the
general setting of set optimization, the use of membership functions can be avoided
as whole sets can be taken into account.
We also note that set optimization makes several other interesting appearances in
sensitivity analysis for vector optimization problems.
1.2 Book Structure
The organization of this book is as follows:

Chapter 2 begins with an introduction to order relations and cone properties
and then gives a detailed overview of solution concepts in vector-valued as well as
1.2 Book Structure 5
set-valued optimization. We introduce and discuss the following solution concepts

for set-valued optimization problems:
• solution concepts based on a vector approach,
• solution concepts based on a set approach,
• solution concepts based on lattice structure.
Furthermore, we present the embedding approach of Kuroiwa and show how it is
possible to transform a set-valued optimization problem into a vector optimization
problem. Solution concepts for set-valued optimization problems with respect
to abstract preference relations and for set-valued problems with variable order
structure are also studied. Subsequently, we introduce approximate solutions of set-
valued optimization problems, and finally, the relationships between the different
solution concepts are studied.
In Chap. 3, we present continuity notions for set-valued mappings and corre-
sponding properties under certain convexity assumptions. Additionally, we intro-
duce Lipschitz properties for both single-valued and set-valued maps as well as
recall some concepts of generalized differentiability and corresponding calculus
rules.
Chapter 4 is a thorough introduction to tangent cones of first-order and tangent
sets and tangent cones of second-order. This chapter includes definitions, properties,
and examples of some of the most commonly used first-order tangent cones
and second-order tangent sets and cones. Specifically, we study six of the most
commonly used first-order tangent cones and six of their modified analogues as
well as give numerous equivalent characterizations for a selection of them. We
also study the notion of the local cone approximation from an abstract point of
view, illustrating not only the key features of tangent cones, but also suggesting
new ways for their generation. Furthermore, we investigate convex cones which lie
between the contingent cone and the Clarke’s tangent cone: two of the most popular
tangent cones of first-order. In this discussion, we also study recession tangent
cones, the Treiman tangent cone, the prototangent cone, and the quasi-strict tangent
cone. We go on to study six of the most commonly used second-order tangent
sets and a related notion of second-order generalized tangent sets. We examine, in
sufficient detail, second-order asymptotic tangent cones and second-order projective
tangent cones as well as collect their intrinsic features. Useful connections among
various tangent cones and tangent sets are also thoroughly investigated. A brief
description of some higher-order tangent cones and higher-order tangent sets is
likewise provided.
In Chap. 5 we introduce nonlinear scalarization methods, which are techniques
that are important from both a theoretical and computational point of view.
We introduce several different scalarizing functionals and discuss their properties;
particularly we examine their monotonicity, continuity, Lipschitz continuity, sub-
linearity, and convexity. Using these nonlinear functionals, we provide nonconvex
separation theorems and also derive necessary optimality conditions for solutions
of set-valued optimization problems (Sect. 12.8). Generally, these functionals prove
useful in a variety of proofs, but especially in the proof of minimal point theorems
6 1 Introduction
in Chap. 10. Continuing, we study characterizations of the solutions of set-valued

optimization problems using nonlinear scalarizing functionals. Finally, we present
the extremal principle of Kruger and Mordukhovich and discuss its relationship to
the separation properties of nonconvex sets. This extremal principle is later applied
in Sect. 12.9 to derive a subdifferential variational principle for set-valued mappings,
and again in Sect. 12.11 to prove a first order necessary condition for fully-localized
minimizers of set-valued optimization problems.
In Chap. 6, we present a generalization of the Hahn–Banach–Kantorovich exten-
sion theorem to K-convex set-valued maps as well as Yang’s extension theorem. We
also give classical separation theorems for convex sets, the core convex topology on
a linear space, and a criterion for the convexity of the cone generated by a set.
To each type of efficiency for optimization problems, it is possible to associate
notions of a conjugate and subdifferential for either vector valued functions or set-
valued maps. Thus Chap. 7 studies the conjugate and subdifferential corresponding
to the notion of strong efficiency as well as the subdifferentials corresponding to
weak and Henig-type efficiencies. For the strong conjugate and subdifferential,
we establish results similar to the convex scalar case. For the other types of
subdifferential, we establish formulas for the subdifferentials of the sum and
composition of both functions and set-valued maps.
Chapter 8 contains presentations of duality assertions for set-valued optimization
problems in infinite dimensional spaces where the solution concept is based on a
vector, set, or lattice approach. We give conjugate duality statements for problems
where the solution concept is based on a vector approach. The notions of conjugate
maps and subdifferentials as well as a perturbation approach used for deriving these
duality assertions are also all provided. A Lagrange duality for set-valued problems
based on the vector approach is also given. Furthermore, we consider set-valued
optimization problems where the solution concept is given by a set order relation
(introduced by Kuroiwa) and derive several corresponding saddle point assertions.
For set-valued problems where the solution concept is based on a lattice structure,
we present duality theorems that are based on the consequent use of the infimum
and supremum (in the sense of greatest lower and least upper bounds with respect
to a partial ordering). We derive conjugate duality assertions as well as Lagrange
duality statements. A comparison of different approaches to duality in set-valued
optimization is then given at the end of the chapter.
In Chap. 9 we establish several existence results for minimal points with respect
to transitive relations and then apply them in topological vector spaces for quasi-
orders generated by convex cones. We continue with the presentation of several
types of convex cones along with related notions of compactness. We end the chapter
with some existence results for vector and set optimization problems.
Chapter 10 presents existence results for the minimal points of subsets of the
Cartesian product of a complete metric space and a topological vector space with
respect to order relations determined by generalized set-valued metrics. Such results
are in turn useful for deriving EVP type results for vector and scalar functions. We
then derive EVP results of Ha’s type as well as an EVP result for bi-set-valued maps.
1.2 Book Structure 7
We end the chapter with an application to error bounds for set-valued optimization
problems.
Chapter 11 is devoted to the derivatives of set-valued maps, which are motivated
by the geometric interpretation of the classical derivative for single-valued maps
as a local approximation of its graph. We study the contingent derivative, the
contingent epiderivative, and examine several analogues of these notions such as
the adjacent derivative, the Clarke derivative, the lower Dini derivative, and the
radial derivative. We also investigate their second-order analogues: the second-
order contingent derivative, the second-order adjacent derivative, the second-order
lower Dini derivative, and their epiderivative counterparts. We give several calculus
rules for the first-order derivatives and epiderivatives along with their second-order
extensions. The existence of contingent epiderivatives is a vital issue in set-valued
optimization and we subsequently present a comprehensive study pertaining to
the existence of contingent epiderivatives, their generalized analogues, and their
various characterizations. We also study other important issues such as proto-
differentiability and semi-differentiability along with several other related notions.
Numerous additional concepts which play a critical role in the differentiability of
set-valued maps are also explored in depth.
In Chap. 12, we give a wide variety of optimality conditions in set-valued
optimization. To be specific, we give necessary first-order and second-order opti-
mality conditions in set-valued optimization by a direct approach using graphical
derivatives and epiderivatives. We also necessary first-order and second-order opti-
mality conditions in set-valued optimization by the Dubovitskii-Milyutin approach
using graphical derivatives and epiderivatives. Sufficient optimality conditions using
graphical derivatives and epiderivatives are also given. A brief discussion of opti-
mality conditions for Q-minimizers using graphical derivatives and coderivatives
is also made. We describe the necessary optimality conditions for set-valued
optimization problems in Asplund spaces using the limiting subdifferential of
Mordukhovich, where the solution concept is based on a vector approach. The
given results do not assume convexity concerning the objective set-valued map or
feasible set, nor do they assume that the ordering cone of the image space has a
nonempty interior. In order to prove the optimality conditions, we use the nonlinear
scalarization method presented in Sect. 5.2. Furthermore, we derive necessary
conditions for Q-minimizers of the set-valued optimization problem in Banach
spaces using Clarke’s coderivative and Ioffe’s approximate coderivative. In the proof
of the necessary optimality condition, we use the scalarizing functional by Hiriart-
Urruty (introduced in Sect. 5.3). Moreover, we also detail necessary conditions for
approximate solutions of a set-valued optimization problem in Asplund spaces using
the extremal principle of variational analysis given in Sect. 5.5. Using a directional
derivative of the set-valued map, we develop necessary and sufficient conditions for
solutions described by the set approach. We also present necessary conditions for
those solutions introduced in Sect. 2.6.5 that are related to applications in welfare
economics (see Sect. 15.3). The main tool for the proof of these necessary conditions
is the extended extremal principle given in Theorem 5.5.10. Due to its importance
for the derivation of algorithms, we also consider set-valued optimization problems
8 1 Introduction
from the point of view of the stability of KKT points. Corresponding results for a
vector approach in finite dimensional spaces are likewise given.
Chapter 13 addresses sensitivity analysis for a variety of problems. Here we
study first-order and second-order sensitivity analysis of set-valued optimization
problems. We introduce several set-valued perturbation maps and investigate their
differentiability properties and give related results. Sensitivity analysis is done by
using both the graphical derivatives such as the contingent derivatives and the lower
Dini derivatives as well as by using the coderivatives. In this chapter, we will also
study differentiability properties of set-valued gap functions associated to vector
variational inequalities and sensitivity analysis for vector variational inequalities.
For vector variational inequalities, both the first-order as well as the second-order
results are given.
In Chap. 14, we introduce several solution procedures for set-valued optimization
problems. Section 14.1 gives a Newton method for solving general set-valued
optimization problems. We also present an algorithm for a special class of set-
valued optimization problems where the objective map is polyhedral, convex, and
set-valued (Sect. 14.2). We further discuss the noteworthy relationship between mul-
tiobjective optimization problems with uncertainties and set-valued optimization
and then go on to present solution methods for these special set-valued optimization
problems (see Sect. 15.4 and particularly Sect. 15.4.4).
Chapter 15 develops several applications of set-valued optimization. We use
the solution concept based on a lattice approach to derive a useful dual problem
to a given vector optimization problem and subsequently to obtain corresponding
duality assertions. This approach is characterized by an embedding of the image
space of the vector optimization problem into a complete lattice. The results are
discussed for the linear case, and then some applications of the duality results to
mathematical finance are given. Specifically, we study a Markowitz-type bicriterial
portfolio optimization problem where the expected return of the portfolio should be
maximized and the risk of the portfolio, measured by the Conditional Value at Risk,
should be minimized. Moreover, we give an interpretation of the corresponding dual
problem in the financial sense. In addition, we apply a variational approach and
necessary conditions for set-valued optimization problems to derive extensions of
the second fundamental theorem of welfare economics to nonconvex economies
with general preference relations.
Finally, the last chapter deals with uncertainty in vector optimization, a topic
which is of obvious importance in many applications. Most real-world optimization
problems are contaminated with uncertain data and some parameters are either not
known or not known exactly. In this chapter we apply solution concepts for set-
valued optimization problems based on a set approach, particularly making use
of the set order relations introduced by Kuroiwa in order to study robustness for
vector-valued optimization problems. The presented algorithms can be considered
as algorithms for solving special classes of set-valued optimization problems where
the solution concept is given by different set order relations.
1.3 Useful Notation 9
1.3 Useful Notation
We conclude this chapter by introducing some notation:

Let X; Y be arbitrary nonempty sets. A function F W X ! 2Y is called a set-
valued map, and is denoted by F W X Y . So, if F W X Y is a set-valued
map, the image of F at x 2 X is a (possibly empty) subset F .x/ of Y . The domain
of F is
dom F WD fx 2 X j F .x/ ¤ ;g;
while its image is
Im F WD fy 2 Y j 9 x 2 X W y 2 F .x/g:
The set-valued map F W X Y is usually identified with its graph,
graph F WD f.x; y/ 2 X Y j y 2 F .x/g:
In this way, with each set-valued map one associates a relation and vice versa. It is
obvious that
dom F D PrX .graph F /
and
Im F D PrY .graph F /;
where PrX .graph F / denotes the projection of graph F onto X and PrY .graph F /
the projection of graph F onto Y . The image by F of the set A X is
[
F .A/ WD F .x/I
x2A
so, Im F D F .X /. The inverse image by F of the set B F is
F 1 .B/ WD fx 2 X j F .x/ \ B ¤ ;g:
In fact, F 1 .B/ is the image of B given by the inverse set-valued map F 1 W

Y X defined by
F 1 .y/ WD fx 2 X j y 2 F .x/gI
so F 1 .y/ D F 1 .fyg/. Of course, dom F 1 D Im F and Im F 1 D dom F . We

shall also use another type of inverse image: F C1 .B/ WD fx 2 X j F .x/ Bg.
10 1 Introduction
When W X W is another set-valued map, the composition of and F is the

set-valued map ı F W X W defined by . ı F /.x/ WD fz 2 W j 9 y 2 F .x/
with z 2 .y/g; note that

graph. ı F / D PrX W graph F W \ X graph :
When Y is a linear space and F; F1 ; F2 W X Y , we define the sum F1 C F2

and multiplication by a scalar F as the set-valued maps F1 C F2 ; F W X Y
defined by
.F1 C F2 /.x/ WD F1 .x/ C F2 .x/;
. F /.x/ WD F .x/
with the usual conventions that A C ; WD ; C A WD ;, ; D ; for A Y .

Furthermore, we introduce some useful notions and notation related to vector-
valued functions. To Y , quasiordered by the proper convex cone C , we adjoin the
greatest and the smallest elements 1 WD C1, 1 (… Y ), thereby obtaining Y WD
Y [ f1; C1g. We consider that y C 1 D 1,
1 D 1, and 1 C y C 1
for all y 2 Y and
2 RC . Of course, if f W X ! Y , the domain of f is
dom f WD fx 2 X j f .x/ 2 Y [ f1gg;
the sublevel and strict sublevel sets of f of height y 2 Y are levf .y/ WD fx 2
X j f .x/ C yg and lev< f .y/ WD fx 2 X j y f .x/ 2 int C g, and the epigraph
of f is epi f WD f.x; y/ 2 X Y j f .x/ C yg. With such an f we associate
the set-valued map Ff;C W X Y whose graph is epi f . So, Ff;C .x/ D f .x/ C C
for every x 2 dom Ff;C D dom f , epi f D epi Ff;C , levf .y/ D levFf;C .y/, and

f .y/ D levFf;C .y/ for every y 2 Y ; in particular, Ff;C D Ff;C C .
lev< <
Chapter 2
Order Relations and Ordering Cones
In this chapter, first, we give an introduction to order relations and cone properties.
Then we present a detailed overview of solution concepts in vector-valued as well as
set-valued optimization. We introduce and discuss the following solution concepts
for set-valued optimization problems:
• solution concepts based on vector approach,
• solution concepts based on set approach,
• solution concepts based on lattice structure.
Furthermore, we present the embedding approach by Kuroiwa and show
how it is possible to transform a set-valued optimization problem into a vector
optimization problem using this embedding approach. Solution concepts for
set-valued optimization problems with respect to abstract preference relations
and for set-valued problems with variable order structure are studied. Moreover,
we introduce approximate solutions of set-valued optimization problems. Finally,
relationships between different solution concepts are studied.
2.1 Order Relations
In this section, our objective is to study some useful order relations. We begin by
recalling that given a nonempty set M , by M M we represent the set of ordered
pairs of elements of M , that is,
M M WD f.x1 ; x2 / j x1 ; x2 2 M g:
The following definition gives the notion of an order relation.

Definition 2.1.1. Let M be a nonempty set and let R be a nonempty subset of
M M . Then R is called an order relation (or a binary relation) on M and the

DOI 10.1007/978-3-642-54265-7__2
12 2 Order Relations and Ordering Cones
pair .M; R/ is called a set M with order relation R. The containment .x1 ; x2 / 2 R
will be denoted by x1 Rx2 . The order relation R is called:
(a) reflexive if for every x 2 M , we have xRx;
(b) transitive if for all x1 ; x2 ; x3 2 M , the relations x1 Rx2 and x2 Rx3 imply that
x1 Rx3 ;
(c) antisymmetric if for all x1 ; x2 2 M , the relations x1 Rx2 and x2 Rx1 imply
that x1 D x2 .
Moreover, an order relation R is called a preorder on M if R is transitive,
a quasiorder if R is reflexive and transitive and a partial order on M if R
is reflexive, transitive, and antisymmetric. In all the three cases, the containment
.x1 ; x2 / 2 R is denoted by x1 R x2 , or simply by x1 x2 if there is no risk of
confusion. The binary relation R is called a linear or total order if R is a partial
order and any two elements of M are comparable, that is
(d) for all x1 ; x2 2 M either x1 R x2 or x2 R x1 .
Furthermore, if each nonempty subset M 0 of M has a first element x 0 (meaning
that x 0 2 M 0 and x 0 R x 8 x 2 M 0 /; then M is called well-ordered.
We recall Zermelo’s theorem: For every nonempty set M there exists a partial
order R on M such that .M; R/ is well-ordered.
An illustrative example of a relation is M WD f.x; x/ j x 2 M g which
is reflexive, transitive, and antisymmetric, but it satisfies (d) only when M is a
singleton.
We recall that the inverse of the relation R M M is the relation
R 1 WD f.x1 ; x2 / 2 M M j .x2 ; x1 / 2 Rg;
and if S is a relation on M , then the composition of R and S is the relation
S ı R WD f.x1 ; x3 / j 9 x2 2 M j .x1 ; x2 / 2 R; .x2 ; x3 / 2 S g:
Using these two notations, the conditions (a), (b), (c), and (d) are equivalent to
M R, R ı R R, R \ R 1 M and R [ R 1 D M M , respectively.
Definition 2.1.2. Let R be an order relation on the nonempty set M and let M0
M be nonempty. An element x0 2 M0 is called a maximal (minimal) element of
M0 relative to R if for every x 2 M0 ;
x0 Rx ) xRx0 .xRx0 ) x0 Rx/: (2.1)
The collection of all maximal (minimal) elements of M0 with respect to (w.r.t.

for short) R is denoted by Max.M0 ; R/ (Min.M0 ; R/).
Note that x0 is a maximal element of M0 w.r.t. R if and only if x0 is a minimal
element of M0 w.r.t. R 1 , and hence Max.M0 ; R/ D Min.M0 ; R 1 /.
2.1 Order Relations 13
Remark 2.1.3. 1. If the order relation R in Definition 2.1.2 is antisymmetric, then

x0 2 M0 is maximal (minimal) if and only if for every x 2 M0
x0 Rx ) x D x0 . xRx0 ) x0 D x/: (2.2)
2. If R is an order relation on M and ; ¤ M0 M , then R0 WD R \ .M0 M0 / is

an order relation on M0 . In such a situation, the set M0 will always be endowed
with the order structure R0 if not stated explicitly otherwise. If R is a preorder
(partial order, linear order) on M , then R0 is a preorder (partial order, linear
order) on M0 . Therefore, x0 is a maximal (minimal) element of M0 relative to R
iff x0 is a maximal (minimal) element of M0 relative to R0 .
In the following, we give some examples to illustrate the above notions.
Example 2.1.4. (1) Assume that X is a nonempty set and M WD P.X / represents
the collection of subsets of X . Then the order relation R WD f.A; B/ 2 M M j
A Bg is a partial order on M . However, if X contains at least two elements,
then R is not a linear order.
(2) Assume that N is the set of non-negative integers and
RN WD f.n1 ; n2 / 2 N N j 9 p 2 N W n2 D n1 C pg:
Then N is well-ordered by RN . Note that RN defines the usual order relation on

N, and n1 RN n2 will always be denoted by n1 n2 or, equivalently, n2
n1 .
(3) Let R be the set of real numbers and let RC WD Œ0; 1Œ be the set of non-negative
real numbers. The usual order relation on R is defined by
R1 WD f.x1 ; x2 / 2 R R j 9 y 2 RC W x2 D x1 C yg:
Then R1 is a linear order on R, but R is not well-ordered by R1 . In the

following, the fact x1 R1 x2 will always be denoted by x1 x2 or, equivalently,
x2
x1 .
(4) Given n 2 N, n
2, we consider the binary relation Rn on Rn defined by
Rn WD f.x; y/ 2 Rn Rn j 8 i 2 1; n W xi yi g;
where x D .x1 ; : : : ; xn /, y D .y1 ; : : : ; yn / and 1; n WD fi 2 N j 1

i ng. Then Rn is a partial order on Rn , but Rn is not a linear order.
For example, the elements e1 and e2 are not comparable w.r.t. Rn , where
ei WD .0; : : : ; 0; 1; 0; : : : ; 0/ 2 Rn . As usual, by ei we denote the vector whose
entries are all 0 except the i th one, which is 1).
Remark 2.1.5. Every well-ordered subset W of R (equipped with its usual partial
order defined above) is at most countable. Indeed, every element y 2 W , except
the greatest element w of W (provided that it exists), has a successor s.y/ 2 W .
Clearly, if y; y 0 2 W , y < y 0 , then s.y/ y 0 . Therefore, fixing qy 2 Q such that
y < qy < s.y/ for y 2 W n fwg, we get an injective function from W n fwg into Q,
and so W is at most countable.
We emphasize that even when R is a partial order on M , a nonempty subset M0
of M may have zero, one, or several maximal elements, but if R is a linear order,
then every subset has at most one maximal (minimal) element.
Definition 2.1.6. Let ; ¤ M0 M and let R be an order relation on M . Then:
1. M0 is lower (upper) bounded (w.r.t. R) if there exists a 2 M such that aRx
(xRa) for every x 2 M0 . In this case, the element a is called a lower (upper)
bound of M0 (w.r.t. R).
2. If, moreover, R is a partial order, we say that a 2 M is the infimum (supremum)
of M0 if a is a lower (upper) bound of M0 and for any lower (upper) bound a0 of
M0 we have that a0 Ra (aRa0 ).
In set-valued optimization, the existence of maximal elements w.r.t. order
relations is an important problem. For this, the following Zorn’s lemma (or Zorn’s
axiom) plays a crucial role.
Axiom 2.1.7 (Zorn) Let .M; / be a reflexively preordered set. If every nonempty
totally ordered subset of M is upper bounded, then M has maximal elements.
We recall that given a linear space X , a nonempty set M X is affine (or a

linear manifold) if
x C .1
/y 2 M for all x; y 2 M and
2 R. A nonempty
set C of X is called convex if Œx; y WD f
x C .1
/y j
2 Œ0; 1g C for all
x; y 2 C . By convention the empty set ; is considered to be affine and convex. It
is obvious that a linear subspace is affine and an affine set is convex. Moreover, any
intersection of linear subspaces, affine sets, or convex sets is a linear subspace, an
affine set, or a convex set, respectively. These properties allow us to introduce the
linear hull, the affine hull, and the convex hull of a nonempty set A X as being,
respectively,
\
lin A WD fY X j A Y; Y linear subspaceg ;
\
aff A WD fM X j A M; M linear manifoldg ;
\
conv A WD fC X j A C; C convex setg :
Clearly, for X D Rn and R D Rn (from Example 2.1.4 (4)), we have
8 x1 ; x2 2 X; 8
2 R W x1 Rx2 ; 0
)
x1 R
x2 ; (2.3)
8 x1 ; x2 ; x 2 X W x1 Rx2 ) .x1 C x/R.x2 C x/: (2.4)
It is easy to find examples of relations satisfying (2.4). In fact, a nonempty

relation R on the linear space X satisfies (2.4) if and only if there exists ; ¤ D X
such that R D RD , where
2.1 Order Relations 15
RD WD f.x1 ; x2 / 2 X X j x2 x1 2 Dg:
Moreover, RD is reflexive if and only if 0 2 D, and RD is transitive if and only

if D C D D.
Definition 2.1.8. Let R be an order relation on the linear space X ; we say that R
is compatible with the linear structure of X if (2.3) and (2.4) hold.
In linear spaces, a large number of relations R can be defined by cones which are
compatible with the linear structure of the space. For this we first give the following:
Definition 2.1.9. A nonempty set C X is a cone if for every x 2 C and for
every
2 RC we have
x 2 C . Clearly, if C is a cone, then 0 2 C . The cone C is
called
(a) convex if for all x1 ; x2 2 C we have x1 C x2 2 C ,
(b) nontrivial or proper if C ¤ f0g and C ¤ X ,
(c) reproducing if C C D X ,
(d) pointed if C \ .C / D f0g.
Clearly, the cone C satisfies condition (b) in the definition above iff, C is a
convex set.
In the following, we collect a few examples of cones.
Example 2.1.10. (1) Let
RnC WD fx 2 Rn j xi
0 8 i 2 1; ng D fx 2 Rn j .0; x/ 2 Rn g: (2.5)
RnC is obviously a cone in the linear space Rn , which fulfills all the conditions
of Definition 2.1.9.
(2) Let C Œ0; 1 be the linear space of all real functions defined and continuous on
the interval Œ0; 1 R. Addition and multiplication by scalars are defined, as
usual, by
.x C y/.t/ D x.t/ C y.t/; .

x/.t/ D
x.t/ 8 t 2 Œ0; 1
for x; y 2 C Œ0; 1 and

2 R: Then
CC Œ0; 1 WD fx 2 C Œ0; 1 j x.t/

0 8 t 2 Œ0; 1g (2.6)
is a convex, nontrivial, pointed, and reproducing cone in C Œ0; 1. Note that the
set
Q WD fx 2 CC Œ0; 1 j x is nondecreasingg (2.7)
is also a convex, nontrivial, and pointed cone in the space C Œ0; 1, but it doesn’t
satisfy condition (c) from Definition 2.1.9: Q Q is the proper linear subspace
of all functions with bounded variation of C Œ0; 1.
(3) Consider the set C Rn defined by
C WD fx D .x1 ; : : : ; xn /T 2 Rn j x1 > 0; or
x1 D 0; x2 > 0; or
:::
x1 D D xn1 D 0; xn > 0; or
x D 0g:
Then the cone C satisfies all the conditions of Definition 2.1.9.

In the following result, we characterize compatibility between linear and order
relations:
Theorem 2.1.11. Let X be a linear space and let C be a cone in X . Then the
relation
RC WD f.x1 ; x2 / 2 X X j x2 x1 2 C g (2.8)
is reflexive and satisfies (2.3) and (2.4). Moreover, C is convex if and only if RC
is transitive, and, respectively, C is pointed if and only if RC is antisymmetric.
Conversely, if R is a reflexive relation on X satisfying (2.3) and (2.4), then C WD
fx 2 X j 0Rxg is a cone and R D RC .
Proof. See [214, Theorem 2.1.13]. t
u
The above result shows that when ; ¤ C X; the relation RC defined by (2.8)
is a reflexive preorder iff C is a convex cone, and RC is a partial order iff C is a
pointed convex cone.
We note that RRnC D Rn (defined in Example 2.1.4 (4)), while the relation
RC with C Rn defined in Example 2.1.10 (3) is a linear order, called the
lexicographic order on Rn .
Let Y be a linear topological space, partially ordered by a proper pointed convex
closed cone C Y .
Denote this order by “C ”. Its ordering relation is described by
y1 C y2 if and only if y2 y1 2 C for all y1 ; y2 2 Y: (2.9)
In the sequel, we omit the subscript C as no confusion occurs.

As usual, we denote by
C C WD fy 2 Y j y .y/
0 8y 2 C g
the continuous positive dual cone of C , and by
C # WD fy 2 C C j y .y/ > 0 8 y 2 C n f0gg
the quasi-interior of C C .
2.2 Cone Properties Related to the Topology and the Order 17
We recall that the interior and the closure of the subset A of the topological
space .X; / are defined, respectively, by
[
int A WD fD X j D A; D openg;
\
cl A WD A WD fB X j A B; B closedg:
Clearly, int A is open and cl A is closed.
2.2 Cone Properties Related to the Topology and the Order
We discuss now the connections between topology and order. Unlike the notion
of an ordered linear space (i.e., a linear space equipped with a compatible reflexive
preorder), the notion of an ordered topological linear space does not demand for
any direct relation between the order and the involved topology. However, because
a compatible reflexive preorder on a linear space is defined by a convex cone, it
is customary to ask that the cone defining the order be closed, have nonempty
interior, or be normal. Before introducing the notion of a normal cone, we recall
that a nonempty set A of the linear space X is full with respect to the convex cone
C X if A D ŒAC , where
ŒAC WD .A C C / \ .A C /:
Note that ŒAC is full w.r.t. C for every set A X .

Definition 2.2.1. Let .X; / be a t.v.s. and let C X be a convex cone. Then C is
called normal (relative to ) if the origin 0 2 X has a neighborhood base formed
by full sets w.r.t. C .
In the next result we give several characterizations of normal cones. We are using
the notation NX for the set of balanced neighborhoods of 0 2 X in the t.v.s. .X; /.
Theorem 2.2.2. Let .X; / be a topological linear space and let C X be a
convex cone. Then the following statements are equivalent:
(i) C is normal,
(ii) 8 V 2 NX ; 9 U 2 NX W ŒU C V ,
(iii) for all nets .xi /i 2I ; .yi /i 2I X such that 0 C xi C yi for every i 2 I one
has .yi / ! 0 ) .xi / ! 0,
(iv) cl C is normal.
u
The following corollary is immediate.
Corollary 2.2.3. Let .X; / be a Hausdorff t.v.s. and let C X be a convex cone.
If C is normal, then cl C is pointed, and so C is pointed, too.
Let .X; / be a Hausdorff t.v.s. partially ordered by the convex cone C . We say
that a net .xi /i 2I X is nonincreasing if
8 i; j 2 I W j i ) xj C xi : (2.10)
Given ; ¤ A X , we say that A is lower bounded with respect to C if A

is lower bounded with respect to RC (see Definition 2.1.6). Similarly, a 2 X is a
lower bound (infimum) of A w.r.t. C if a is so for RC . Hence a 2 X is a lower
bound of A w.r.t. C if a C x for every x 2 A. An element a is the infimum of A
w.r.t. C if a is a lower bound and for any lower bound a0 of A we have that a0 C a.
The infimum of A w.r.t. C will be denoted by infC A when it exists.
Proposition 2.2.4. Let .X; / be a Hausdorff t.v.s. partially ordered by the closed
convex cone C . If the net .xi /i 2I X is nonincreasing and convergent to x 2 X ,
then fxi j i 2 I g is bounded below and x D inffxi j i 2 I g.
Proof. See [214, Proposition 2.1.24]. t
u
We emphasize that in ordered topological linear spaces, the classical result
concerning the bounded monotone sequences is not generally true. We consider
the linear space `1 of all bounded sequences x D .x k /k1 R endowed with the
norm kxk D supfjx k j j k D 1; 2; : : :g. In `1 we consider the “usual” partial order
generated by the cone `1C WD fx 2 l
1
j x k
0 8 k
1g; `1
C is a pointed closed
convex cone (even reproducing and with nonempty interior).
Example 2.2.5 (Peressini [475, p. 91]). The sequence fxn g `1 , defined by (for
n fixed)

1 if 1 k n;
xnk D
0 if k > n;
is nonincreasing w.r.t. C , and inffxn j n

1g D e 0 WD e where e D .1; 1; 1; : : :/ 2
`1 . But kxn e 0 k D 1 for every n
1. Consequently, fxn gn1 does not converge
to its infimum.
We also recall that a cone C that partially orders a Hausdorff linear topological
space .X; / is said to be Daniell if any nonincreasing net having a lower bound
-converges to its infimum (see Jahn [292, p.29], Luc [402, p. 47], Borwein [67]).
In the following, let us recall some useful notions of cones which play an
important role in proving existence results for solutions of optimization problems
in infinite dimensional spaces.
Definition 2.2.6. Let Y be a Hausdorff topological vector space and C Y a
proper convex cone.
Fig. 2.1 Cone properties C=cl C; Y =Rn

C well-based ⇐= C pointed
Y normed
C nuclear =⇒ C normal
C complete
C Daniell ⇐= C compact base.
(i) C is based if there exists a nonempty convex subset B of C such that C D

RC B (where RC B WD f
b j b 2 B and

0g) and 0 … cl B; the set B is
called a base for C .
(ii) C is called well-based if C has a bounded base.
(iii) Let the topology of Y be defined by a family P of seminorms. C is called
supernormal or nuclear if for each p 2 P there exists y 2 Y , such that
p.y/ hy; y i for all y 2 C ; it holds y 2 C C in this case.
(iv) C is said to be Daniell if any nonincreasing net having a lower bound
converges to its infimum.
(iv) C is said to be regular if any decreasing (increasing) net which has a lower
bound (upper bound) is convergent.
In Fig. 2.1 we give an overview of such additional cone properties and corre-
sponding relations for the case that Y is a Banach space, C a proper and convex
cone in Y .
The following result gives useful information for cones with bases:
Theorem 2.2.7. Let X be a Hausdorff locally convex space and C X a proper
convex cone. Then C has a base if and only if C # ¤ ;.
u
In the following, we collect a few examples of Daniell cones.
Example 2.2.8. 1. We recall that if .x˛ /˛2A is a net which is increasing (decreasing)
in a topological vector space .Y; / ordered by a closed convex cone C and if x is
a cluster point of .x˛ /, then x D sup˛2A x˛ (x D inf˛2A x˛ ) (see Peressini [475,
Proposition 3.1]). Therefore, any regular cone is Daniell.
2. If .Y; jj jj/ is a Banach lattice, that is, Y is a Banach space, vector lattice and the
norm is absolute, i.e., jjxjj D jj jxj jj for any x 2 Y , then the cone YC D fy 2
Y j y
0g is Daniell if Y has weakly compact intervals.
3. A convex cone with a weakly compact base is a Daniell cone.
The following result connects some useful cones.
Proposition 2.2.9 (Isac [280]). Let .Y; P/ be a Hausdorff locally convex space
and C Y a proper convex cone. Then
C well-based H) C nuclear H) C normal:

If Y is a normed space, then
C nuclear H) C well-based:
Remark 2.2.10. Among the classical Banach spaces their usual positive cones
are well-based only in l 1 and L1 .˝/ (but l 1 is not an Asplund space (see
Definition 3.5.3)).
Let Y be a topological vector space over R. Assume .Y; C / is simultaneously a
vector lattice with the lattice operations x 7! x C , x 7! x , x 7! jxj, .x; y/ 7!
supfx; yg and .x; y/ 7! inffx; yg.
Definition 2.2.11. A set A Y is called solid, if x 2 A and jyj jxj implies
y 2 A. The space Y is called locally solid, if it possesses a neighborhood of 0
consisting of solid sets.
Lemma 2.2.12. The following properties are equivalent:
(i) Y is locally solid.
(ii) C is normal, and the lattice operations are continuous.
In order to derive optimality conditions or duality statements in general spaces
(cf. Chaps. 8, 12), the ordering cone is often required to have a nonempty interior.
Therefore, in the following, we give some examples of convex cones with nonempty
interior.
Example 2.2.13. 1. Any closed convex cone C in the Euclidean space .Rn ; h; i/
such that C is self-adjoint (i.e., C D C C ) has a nonempty interior.
2. Consider the space of continuous functions C Œa; b with the norm jjxjj D
supfjx.t/j j t 2 Œa; bg. Then the cone of positive functions in C Œa; b
C Œa; bC WD fx 2 C Œa; b j 8t 2 Œa; b W x.t/

0g
has a nonempty interior.

3. Let Y D l 2 .N ; R/ with the well-known structure of a Hilbert space. The convex
cone
1
X
Cl 2 WD fx D fxi gi 1 j x1
0 and xi2 x12 g
i D2
has a nonempty interior

1
X
int Cl 2 WD fx D fxi gi 1 j x1 > 0 and xi2 < x12 g:
i D2
4. Let l 1 be the space of bounded sequences of real numbers, equipped with the
norm jjxjj D supn2N fjxn jg. The cone
l 1 C WD fx D fxn gn2N j xn
0 for any n 2 Ng
has a nonempty interior (cf. Peressini [475], p. 186).

5. Let C 1 Œa; b be the real vector space formed by all real continuously
differentiable functions defined on Œa; b (a; b 2 R; a < b), equipped with
the norm
Z b Z b
jjf jj1 WD f .f .t// d t C
2
.f 0 .t//2 d tg1=2
a a
for any f 2 C 1 Œa; b. Using a Sobolev’s imbedding theorem, we can show that
the natural ordering cone
C 1 Œa; bC WD ff 2 C 1 Œa; b j f

0g
has a nonempty interior. The proof is based on some technical details (cf. da
Silva [532]).
6. About the locally convex spaces, we put in evidence the following result. If .Y; /
is a real locally convex space, then for every closed convex pointed cone C Y ,
with nonempty interior, there exists a continuous norm jj jj on Y such that C
has a nonempty interior in the normed space .Y; jj jj/.
Proof. Take y0 2 int C and A WD .y0 C /\.C y0 /. Then A is a closed convex
and balanced set with 0 2 int A such that the Minkowski functional pA W Y ! R
defined by
pA .y/ WD infft > 0 j y 2 tAg
is a seminorm. Because int A D core A D fy 2 Y j pA .y/ < 1g A D fy 2

Y j pA .y/ 1g (see Proposition 6.2.1), pA is also continuous. Take y 2 Y with
pA .y/ D 0. Then y 2 n1 A for every n
1, whence n1 y0 ˙ y 2 C for such
n. It follows that ˙y 2 cl C D C , and so y D 0: Hence kk WD pA is a norm
and A D B.Y;kk/ , and so y0 2 intkk C: t
u
Finally, we give an example of a normed (vector) space (n.v.s.) where the natural
ordering cone has a nonempty interior as well as the Daniell property.
Example 2.2.14 (see Jahn [293]). Consider the real linear space L1 .˝/ of all
(equivalence classes of) essentially bounded functions f W ˝ ! R (; ¤ ˝ Rn )
measurable with the norm jj jjL1 .˝/ given by
jjf jjL1 .˝/ WD ess supx2˝ fjf .x/jg for all f 2 L1 .˝/:
The ordering cone
L1 .˝/C WD ff 2 L1 .˝/ j f .x/

0 almost everywhere on ˝g
has a nonempty interior and is weak Daniell.

2.3 Convexity Notions for Sets and Set-Valued Maps
Throughout this section X; Y are real topological vector spaces.

Definition 2.3.1. Let A X be a nonempty set. We say that A is ˛-convex, where
˛ 2 0; 1Œ, if ˛x C .1 ˛/y 2 A for all x; y 2 A. The set A is mid-convex if A is
1
2
-convex. The set A is nearly convex if A is ˛-convex for some ˛ 2 0; 1Œ. The set
A is closely convex if cl A is convex. The empty set is ˛-convex for all ˛ 2 0; 1Œ and
closely convex (and so nearly convex).
Of course, A is convex if and only if A is ˛-convex for every ˛ 2 0; 1Œ. Moreover,
if T W X ! Y is a linear operator and A X , B Y are ˛-convex (nearly convex,
convex), then T .A/ and T 1 .B/ are ˛-convex (nearly convex, convex), too.
Some properties of nearly convex sets are mentioned in the next result (see [214,
Proposition 2.4.3, Corollary 2.4.4]).
Proposition 2.3.2. Let A X be a nonempty nearly convex set. Then
(i) cl A is convex.
(ii) If x 2 icr A and y 2 A, then Œx; y A. Moreover, if x 2 int A and y 2 A,
then Œx; yŒ int A.
(iii) If int A ¤ ;, then int A is convex and icr A D int A.
(iv) If A is open or closed, then A is convex.
Definition 2.3.3. Let C Y be a convex cone. We say that A Y is C -˛-convex
if A C C is ˛-convex; A is nearly C -convex if A C C is nearly convex; A is closely
C -convex if A C C is closely convex. Moreover, A is closely c-C -convex (nearly
C -subconvexlike in [601]) if cl .P.A C C // is convex; A is ic-C -convex (see [517])
if int .P.A C C // is convex and P.A C C / cl .int .P.A C C ///.
The next result, stated essentially in [79, Lemma 2.5], proves to be useful in the
following sections.
Lemma 2.3.4. Assume that C Y is a convex cone with int C ¤ ; and let A Y .
Then
cl.A C C / D cl.cl A C C / D cl .A C int C / ; (2.11)

A C int C D cl A C int C D int.A C C / D int.cl A C C / D int.cl.A C C //:
(2.12)
Therefore, cl.A C C / is convex iff A C int C is convex.

Proof. The equalities in (2.11) follow immediately from the known relation
cl.A C B/ D cl.cl A C B/ D cl.cl A C cl B/; (2.13)
valid for all subsets A; B Y , and the fact that cl C D cl.int C /:

2.3 Convexity Notions for Sets and Set-Valued Maps 23
Note that A C int C is open being the union [a2A .a C int C / of open sets. The
inclusions A C int C cl A C int C and int.A C C / int.cl.A C C // are obvious.
Take y 2 cl A and k 2 int C . Since k int C 2 NY .0/, we have that A \ .y C k
int C / ¤ ;, whence y C k 2 A C int C ; hence cl A C int C A C int C , and so
the first equality in (2.12) is true.
The inclusion A C int C int.A C C / is obvious because A C int C is open.
Fix k 0 2 int C and take y 2 int.A C C /. Then there exists ˛ > 0 such that
y ˛k 0 2 A C C , whence y 2 A C C C int C D A C int C . It follows that
int.A C C / A C int C , and so the second equality in (2.12) is true. The third
equality in (2.12) follows immediately from the first two equalities.
Clearly, cl.A C C / D cl.A C C / C C ; using the first three equalities in (2.12)
we get
int.cl.A C C // D int.cl.A C C / C C / D cl.A C C / C int C

D .A C C / C int C D A C int C:
If cl.A C C / is convex then A C int C D int.cl.A C C // is convex. Conversely,

if A C int C is convex, then cl.A C C / D cl .A C int C / is convex. The proof is
complete. t
u
It is worth observing that
P.A C C / D PA C C; P.A C int C / D PA C int C: (2.14)
Moreover, if int C ¤ ;, using (2.14) and (2.12) we get
int.P.A C C // D PA C int C D int.cl.P.A C C ///: (2.15)
In the next result we establish some relationships between the C -convexity

notions above.
Proposition 2.3.5. Let A Y . The following assertions hold:
(i) Let ˛ 2 0; 1Œ. Then A is C -˛-convex iff ˛A C .1 ˛/A A C C:
(ii) A is closely C -convex iff
A C .1
/A cl.A C C / for all
2 0; 1Œ:
(iii) A is closely C -convex and int C ¤ ; iff
9k 2 int C; 8˛ > 0; 8
2 0; 1Œ W ˛k C
A C .1
/A A C C: (2.16)
(iv) If A is nearly C -convex then A is closely C -convex.

(v) If A is closely C -convex then A is closely c-C -convex.
(vi) A is ic-C -convex iff A is closely c-C -convex and
int .cl .P.A C C /// D int .P.A C C // ¤ ;: (2.17)
(vii) Assume that int C ¤ ;. Then A is ic-C -convex iff A is closely c-C -convex.
Proof. All the assertions a clearly true if A is empty. Therefore, we assume that
A ¤ ;:
(i) The assertion is (almost) evident.
(ii) Assume that A is closely C -convex and take
2 0; 1Œ. Since cl.A C C / is
convex we get

A C .1
/A
cl.A C C / C .1
/ cl.A C C / D cl.A C C /:
Conversely, assume that

A C .1
/A cl.A C C / for all
2 0; 1Œ. Taking

2 0; 1Œ, we have that
.A C C / C .1
/.A C C / cl.A C C / C C D
cl.ACC /. Using (2.13) we get
cl.ACC /C.1
/ cl.ACC / cl.ACC /,
and so cl.A C C / is convex.
(iii) Assume that (2.16) holds; then clearly int C ¤ ;. Take y1 ; y2 2 A and
2
0; 1Œ. Then n1 k C
y1 C .1
/y2 2 A C C for every n 2 N ; taking the
limit we get
y1 C .1
/y2 2 cl.A C C /, and so A is closely C -convex.
Assume now that A is closely C -convex and int C ¤ ;. Consider k 2
int C . Take ˛ > 0, y1 ; y2 2 A and
2 0; 1Œ. Then
y1 C .1
/y2 2
cl.A C C /. Then using (2.12),
˛k C
y1 C .1
/y2 2 cl.A C C /
C int C D .A C C / C int C D A C int C A C C:
(iv) Assume that A is nearly C -convex. Then, using (i), A C C is nearly convex.
Then, by Proposition 2.3.2 (i) we obtain that cl.A C C / is convex, that is A is
closely C -convex.
(v) Clearly, cl .P.A C C // D cl .P.cl.A C C /// : Since cl.A C C / is convex,
from the preceding relation we obtain that cl .P.A C C // is convex, that is A
is c-C -convex.
(vi) Assume that A is ic-C -convex. From the definition of the ic-C -convexity we
have that B WD int .P.A C C // is nonempty and convex, and P.A C C /
cl B. It follows that int.cl B/ D B P.A C C / and cl.P.A C C //
cl B, and so cl.P.A C C // D cl B is convex. Therefore, A is closely c-
C -convex. Moreover, since B is open, convex and nonempty, we have that
int .cl.P.A C C /// D int.cl B/ D B. Therefore, (2.17) holds.
Assume now that A is closely c-C -convex and (2.17) holds. Then C WD
cl.P.A C C // is convex int C D int .P.A C C // DW B ¤ ;. Then clearly B
is convex and cl B D cl .int C / D C P.A C C /. Hence A is ic-C -convex.
(vii) Let int C ¤ ;. Then (2.15) holds and int .P.A C C // ¤ ;. The conclusion
follows using (vi). The proof is complete. t
u
Note that Proposition 2.3.5 (vii) is stated in [601, Theorem 3.1], while the fact
that A is closely c-C -convex if A is ic-C -convex in Proposition 2.3.5 (vi) is proved
in [601, Theorem 3.2]
Let F W X Y . We say that F is ˛-convex (mid-convex, nearly convex,

convex) if graph F is ˛-convex (mid-convex, nearly convex, convex). It is obvious
that if F is ˛-convex (mid-convex, nearly convex, convex), so are dom F; Im F , and
F .x/ for every x 2 X . It is easy to see that F is ˛-convex if and only if

8 x; x 0 2 dom F W ˛F .x/ C .1 ˛/F .x 0 / F ˛x C .1 ˛/x 0 I
in the relation above x; x 0 2 dom F can be replaced by x; x 0 2 X:

To F W X Y we associate the set-valued maps cl F , conv F , convF W X Y
defined by
.cl F /.x/ WD clŒF .x/; .conv F /.x/ WD convŒF .x/;

.convF /.x/ WD convŒF .x/ .x 2 X /:
It is almost obvious that cl F , conv F and convF are ˛-convex (mid-convex, nearly
convex, convex) if F is ˛-convex (mid-convex, nearly convex, convex).
To F W X Y and y 2 Y , where Y is the topological dual of Y , we also
associate
y WD yF W X ! R; y .x/ WD inf fhy; y i j y 2 F .x/g .x 2 X /;

(2.18)
where, as usual, inf ; WD C1; then dom y D dom F for every y 2 Y and
0 D dom F . Clearly, yF D yclF D yconv

F
D yconvF
for every y 2 Y . The
function y (but with sup instead of inf) was introduced in [138], and used (for
example) in [422, 423, 518], too.
Proposition 2.3.6. Let F W X Y .
(i) If F is convex then y is convex for every y 2 Y :
(ii) Assume that Y is a locally convex space. If y is convex for every y 2 Y
then convF is convex.
Proof. (i) Consider x; x 0 2 dom y and ˛ 2 0; 1Œ. Take ; 0 2 R such that
y .x/ < ; y .x 0 / < 0 . Then there exist y 2 F .x/, y 0 2 F .x 0 / such that
hy; y i < , hy 0 ; y i < 0 . Then ˛y C.1˛/y 0 2 F .˛x C .1 ˛/x 0 /, and so
˝ ˛ ˝ ˛
y ˛x C .1 ˛/x 0 ˛y C .1 ˛/y 0 ; y D ˛ hy; y i C .1 ˛/ y 0 ; y
< ˛ C .1 ˛/ 0 :
Letting ! y .x/; 0 ! y .x 0 / we get y .˛x C .1 ˛/x 0 /

˛y .x/ C .1 ˛/y .x 0 /. Hence y is convex.
(ii) Since yF D yconvF
for every y 2 Y , we may (and do) assume that F D
convF . We have that dom F D dom 0 is convex, 0 D dom F being convex.
Assume that F is not convex. Then there exist x; x 0 2 dom F; y 2 F .x/, y 0 2
F .x 0 / and ˛ 2 0; 1Œ such that z WD ˛y C .1 ˛/y 0 … F .˛x C .1 ˛/x 0 / DW

A. Since A is a nonempty closed convex set, there exists y 2 Y such that
hz; y i < inf fhv; y i j v 2 Ag D y .˛x C .1 ˛/x 0 /. Since
˝ ˛
˛y .x/ C .1 ˛/y .x 0 / ˛ hy; y i C .1 ˛/ y 0 ; y D hz; y i

< y ˛x C .1 ˛/x 0 ;
we get the contradiction that y is not convex. Hence F is convex. t

u
Let C Y be a convex cone. We say that F is C -˛-convex (C -mid-convex,
C -nearly convex, C -convex) if the set-valued map
FC W X Y; FC .x/ WD F .x/ C C;
is ˛-convex (mid-convex, nearly convex, convex). Of course, F is C -˛-convex if

and only if

8 x; x 0 2 dom F W ˛F .x/ C .1 ˛/F .x 0 / F ˛x C .1 ˛/x 0 C C:
Note that sometimes graph FC is denoted by epiC F , or simply epi F , and is

called the epigraph of F .
Corollary 2.3.7. Let F W X Y and C Y be a convex cone.
(i) If F is C -convex then y is convex for every y 2 C C :
(ii) Assume that Y is a locally convex space and y is convex for every y 2 C C .
Then convFC is convex; in particular, if F .x/ C C is closed and convex for
every x 2 X , then F is C -convex.
Proof. Of course, F is C -convex if and only if FC is convex. Let us set Qy WD
yconvF
. Note that Q y D yF D y for y 2 C C , while for y 2 Y n C C ,
Q y .x/ D C1 for x 2 dom F and Qy .x/ D 1 for x 62 dom F . Hence Qy is
convex for every y 2 Y if and only if y is convex for every y 2 C C . The
conclusion follows applying Proposition 2.3.6 to FC : t
u
Of course, Proposition 2.3.6 can be obtained from Corollary 2.3.7 taking
C D f0g. Corollary 2.3.7 can be found, essentially, in [138, Proposition 1.6]
and [518, Lemma 3].
The sublevel set of F of height y (w.r.t. C ) is the set
levF .y/ WD fx 2 X j F .x/ \ .y C / ¤ ;gI
when int C ¤ ; we also consider the strict sublevel set of F of height y (w.r.t. C )
defined by
F .y/ WD fx 2 X j F .x/ \ .y int C / ¤ ;g:

lev<
F W
In this way we get the sublevel and strict sublevel set-valued maps levF ; lev<
Y X.
We say that F is C -˛-quasiconvex (C -mid-quasiconvex, C -nearly quasi-
convex, C -quasiconvex) if for every z 2 Y the sublevel set levF .z/ is ˛-convex
(mid-convex, nearly convex, convex). An equivalent definition of C -˛-quasi-
convexity is that

8 x; x 0 2 dom F W .F .x/ C C / \ .F .x 0 / C C / F ˛x C .1 ˛/x 0 C C:
Notice that F is C -˛-quasiconvex whenever

8x; x 0 2 dom F W F .x/ F ˛x C .1 ˛/x 0 C C or

F .x 0 / F ˛x C .1 ˛/x 0 C C:
Note also that F is C -˛-quasiconvex (C -mid-quasiconvex, C -nearly quasiconvex,

C -quasiconvex) whenever F is C -˛-convex (C -mid-convex, C -nearly convex, C -
convex).
The set-valued map F is C -convexlike if
8x1 ; x2 2 X; 8y1 2 F .x1 /; 8y2 2 F .x2 /; 8

2 0; 1Œ;
9x3 2 X W
y1 C .1
/y2 2 F .x3 / C C;
or, equivalently, F .X / C C is convex, that is F .X / is C -convex. Of course, if F is

C -convex then F is C -convexlike.
Li and Chen [387] (see also [602]) say that F is C -subconvexlike if
9k 2 int C; 8˛ > 0; 8x; x 0 2 X; 8

2 0; 1Œ W ˛kC
F .x/C.1
/F .x 0/ F .X /CC:
Using Proposition 2.3.5 (iii), F is C -subconvexlike iff int C ¤ ; and F .X / is

closely C -convex.
We say that f W X ! Y is C -˛-convex (C -mid-convex, C -nearly convex,
C -convex, C -˛-quasiconvex, C -mid-quasiconvex, C -nearly quasiconvex, C -
quasiconvex) if the set-valued map Ff;C is C -˛-convex (C -mid-convex, C -nearly
convex, C -convex, C -˛-quasiconvex, C -mid-quasiconvex, C -nearly quasiconvex,
C -quasiconvex); in particular, f is C -convex if and only if

8 x; x 0 2 X; 8 ˛ 2 Œ0; 1 W f ˛x C .1 ˛/x 0 C ˛f .x/ C .1 ˛/f .x 0 /:
If f is C -˛-convex (C -mid-convex, C -nearly convex, C -convex), then dom f

is so, and f is C -˛-quasiconvex (C -mid-quasiconvex, C -nearly quasiconvex,
C -quasiconvex).
2.4 Solution Concepts in Vector Optimization
In this section, we first recall concepts of Pareto minimal points, weakly and
properly minimal points, then we introduce the concept of Q-minimal points and
establish relations among them.
Unless otherwise mentioned, in the following we consider a linear topological
space Y , partially ordered by a proper pointed convex closed cone C and a
nonempty set A Y .
We introduce the following sets of Pareto minimal points (Pareto maximal points,
respectively) of A with respect to C :
Definition 2.4.1 (Pareto Minimal (Maximal) Points). Consider
Min.A; C / WD fy 2 A j A \ .y C / D fygg: (2.19)
An element y 2 Min.A; C / is called a Pareto minimal point of A with respect

to C .
Furthermore, consider
Max.A; C / WD fy 2 A j A \ .y C C / D fygg: (2.20)
An element y 2 Max.A; C / is called a Pareto maximal point of A with respect

to C .
Moreover, in order to describe weak minimality we will study the following
solution concept in Y . Many solution procedures for vector optimization problems
generate weakly minimal elements.
Definition 2.4.2 (Weakly Minimal (Maximal) Points). Suppose that int C ¤ ;.
Consider
WMin.A; C / WD fy 2 A j A \ .y int C / D ;g: (2.21)
An element y 2 WMin.A; C / is called a weakly minimal point of A with respect

to C . Furthermore, consider
WMax.A; C / WD fy 2 A j A \ .y C int C / D ;g: (2.22)
An element y 2 WMax.A; C / is called a weakly maximal point of A with respect

to C .
Moreover, we introduce the concept of strongly minimal points:
Definition 2.4.3. Consider
StrMin.A; C / WD fy 2 A j A y C C g: (2.23)
2.4 Solution Concepts in Vector Optimization 29
An element y 2 StrMin.A; C / is called a strong minimal point of A with respect

to C .
In the following, we introduce different concepts of properly minimal points.
Properly minimal points are important in the proofs of many theoretical asser-
tions because corresponding scalarizing functionals (see Sect. 5.1) are strictly
C -monotone. The first concept for proper minimality in the following definition (cf.
Ha [228]) is based on scalarization by means of (strictly C -monotone) functionals
y 2 C #.
Definition 2.4.4 (Properly Minimal Points).
(a) Suppose that C # ¤ ; and consider
S-PMin.A; C / WD fy 2 A j 9y 2 C # ; 8y 2 A W y .y/ y .y/g:

(2.24)
An element y 2 S-PMin.A; C / is called a S-properly minimal point of A

w.r.t. C .
(b) Let
Hu-PMin.A; C / WD fy 2 A j .cl conv coneŒ.A y/ [ C / \ .C / D f0gg:
An element y 2 Hu-PMin.A; C / is called a Hurwicz properly minimal point

of A w.r.t. C .
(c) Assume that Y is a n.v.s.; y is a Hartley properly minimal point of A w.r.t.
C (y 2 Ha-PMin.A; C /) if y 2 Min.A; C / and there exists a constant M > 0
such that, whenever there is
2 C C with
.y y/ > 0 for some y 2 A, one
can find 2 C C with

.y y/=k
k M..y y/=kk/:
(d) Consider
Be-PMin.A; C / WD fy 2 A j cl coneŒ.A y/ C C \ .C / D f0gg:
y 2 Be-PMin.A; C / is called a Benson properly minimal point of A w.r.t. C .

(e) Consider
Bo-PMin.A; C / WD fy 2 A j cl cone.A y/ \ .C / D f0gg:
y 2 Bo-PMin.A; C / is called a Borwein properly minimal point of A w.r.t. C .

(f) Consider
GHe-PMin.A; C / WD fy 2 A j 9 proper convex pointed cone D with C n

f0g int D such that .A y/ \ .intD/ D ;g.
y 2 GHe-PMin.A; C / is called a Henig global properly minimal point of A
w.r.t. C .
(g) Suppose that Y is a n.v.s., C has a base and consider
He-PMin.A; C / WD fy 2 A j 9 " > 0 such that cl cone.A y/ \ . C

"BY / D ;g:
y 2 He-PMin.A; C / is called a Henig properly minimal point of A w.r.t. to C .

(h) Assume that Y is a n.v.s. Consider
Sup-PMin.A; C / WD fy 2 A j 9 > 0 such that cl cone.A y/ \ .BY C /

BY g:
y 2 Sup-PMin.A; C / is called a super efficient point of A w.r.t. C .

For the notions of minimal points in Definition 2.4.4, we refer to [292, 293, 402]
and [228]. The concepts of Henig proper minimality and Henig global proper
minimality have been presented in [242]. The above definition of Henig properly
minimal points can be found in [71, 627]; see also [228, 242, 618, 619]. For an
equivalent definition of Henig properly minimal points by means of a functional
from C # the reader is referred to [619]. We note that positive proper minimality
has been introduced by Hurwicz [11], and super efficiency has been introduced by
Borwein and Zhuang [71]. We refer the reader to [219] for a survey and materials
on proper efficiency.
In the sequel, when speaking of weakly minimal points (resp. S-properly minimal
points) we mean that int C (resp. C # ) is nonempty, when speaking of Henig minimal
points we mean that C has a base and when speaking that C has a bounded base
we mean that is bounded.
Let B Y be a convex set such that 0 … cl B (that is B is a base for cone B);
we set
NYB WD N B
WD fV 2 NY j V convex, V \ B D ;g I (2.25)
clearly, NYB ¤ ;. For V 2 N B

we set
PVB WD cone.B C V /I
then PVB is a proper convex cone with int PVB D P.B C int V / ¤ ;:
Similar to the case of normed vector spaces, if is a base of C , we set
He-PMin.A; / WD fy 2 A j 9V 2 NY W Œcl cone.A y/ \ .V / D ;g:
Observe that

He-PMin.A; / D fy 2 A j 9V 2 NY W .A y/ \ PV D f0gg (2.26)

D fy 2 A j 9V 2 NY W .A y/ \ int PV D ;g: (2.27)
Indeed, Œcl cone.A y/ \ .V / D ; ) .A y/ \ cone.V / D f0g and

.A y/ \ cone.V / D f0g ) Œcl cone.A y/ \ .int V / D ;:
Moreover, the super efficiency in the case in which Y is a locally convex space
is defined by
Sup-PMin.A; C / WD fy 2 A j 8V 2 NY ; 9U 2 NY W cl cone.A y/ \
.U C / V g:
The following two Propositions 2.4.5 and 2.4.6 are shown under weaker assump-
tions concerning the cone C Y (Y a linear topological space), namely that C is a
proper convex cone, for the corresponding solution concepts.
Proposition 2.4.5. Let A Y be nonempty. Then
T˚
(i) StrMin.A; C / arg minA y j y 2 C C n 0 , with equality if Y is a l.c.s.
and C is closed. S ˚
(ii) S-PMin.A; C / D arg minA y j y 2 C # D S-PMin.A C C; C /:
(iii) If int C ¤ ;, then Min.A; C / WMin.A; C / and
WMin.A; C / D A \ WMin.A C C; C / D A \ bd.A C C / (2.28)

S˚
arg minA y j y 2 C C n f0g ; (2.29)
with equality if A is closely C -convex.

(iv) We have that GHe-PMin.A; C / Min.A; C / and
S
GHe-PMin.A; C / D fWMin.A; D/ j D 2 DC g
D GHe-PMin.A C C; C / S-PMin.A; C / (2.30)
with equality if A is closely C -convex, where
DC WD fD Y j D proper pointed convex cone with C n f0g int Dg:

(2.31)
(v) Assume that is a base of C . Then
S
He-PMin.A; / D fWMin.A; D/ j D 2 D g D He-PMin.A C C; /
(2.32)
S
farg minA y j y 2 Y ; inf y . / > 0g ; (2.33)
with equality if A is closely C -convex, where
D WD fPV j V 2 NY g: (2.34)
(vi) If is a base of C then D DC ; consequently, He-PMin.A; /

GHe-PMin.A; C /:
Proof. (i) Let y 2 StrMin.A; C /, that is y 2 A y C C . Then clearly
StrMin.A; C / D fyg and the T˚inclusion holds. Assume that Y is a l.c.s. and
C is closed, and take y 2 arg minA y j y 2 C C . Then y 2 A and
hy y; y i
0 for all y 2 A and y 2 C C . By the bipolar theorem we get
y y 2 C CC D cl C D C , whence A y C C:
(ii) The first equality is given by the definition of S-PMin.A; C /. The inclusion
S-PMin.A; C / S-PMin.A C C; C / is obvious. If y 2 S-PMin.A C C; C /,
then y D a C c for some a 2 A, c 2 C , and there exists y 2 C # such that
ha C c; y i hy; y i for all y 2 A C C . In particular ha C c; y i ha; y i ;
whence hc; y i 0. It follows that c D 0, and so y D a 2 A. Consequently,
y 2 S-PMin.A; C /:
(iii) Assume that int C ¤ ;. Taking into account that int C C n f0g, the inclusion
Min.A; C / WMin.A; C / follows.
For y 2 Y , using Lemma 2.3.4, we have
.Ay/\. int C / D ; , y … ACint C , y … ACC Cint C , y … int.ACC /:
It follows that
y 2 WMin.A; C / , Œy 2 A; .A C C y/ \ . int C / D ;
, y 2 A \ WMin.A; C /
, Œy 2 A; y … int.A C C / , y 2 A \ bd.A C C /:
Hence the equalities in (2.28) hold.

Take y 2 arg minA y for some y 2 C C n f0g. Because int C
fy 2 Y j hy; y i > 0g, we obtain that .A y/ \ . int C / D ;; and so
y 2 WMin.A; C /. Hence the inclusion in (2.28) holds.
Assume now that A is closely C -convex and y 2 WMin.A; C / Œ
WMin.A C C; C /: Then .A C C / \ .y int C / D ;; it follows that
cl.A C C / \ .y int C / D ;: Because cl.A C C / is convex, by a separation
theorem (see [293, Theorem 3.16]) there exists y 2 Y n f0g such that
hy C v; y i
hy v0 ; y i for all y 2 A and v; v0 2 C . It follows that
y 2 C C and hy y; y i
0 for every y 2 A, and so y 2 C C n f0g
and y 2 arg minA y :
(iv) The first equality in (2.30) is just the definition of GHe-PMin.A; C /:

Take y 2 GHe-PMin.A; C /; then y 2 WMin.A; D/ for some D 2 DC , and
so .A y/ \ . int D/ D ;. Because C C int D D int D (for every D 2 DC ),
we clearly have that GHe-PMin.A; C / GHe-PMin.A C C; C /:
Take y 2 GHe-PMin.A C C; C /; then y D y C v for some y 2 A, v 2 C .
Assuming that v ¤ 0, we get the contradiction v 2 .C n f0g/ \ .A C C y/
. int D/ \ .A C C y/ D ;. Hence y D y 2 A:
Take y 2 S-PMin.A; C /. Then there exists y 2 C # with 0
hy y; y i hy C v y; y i for all y 2 A, v 2 C: Take D WD
f0g [ fy 2 Y j hy; y i > 0g. Then D is a pointed convex cone with
int D D fy 2 Y j hy; y i > 0g. It follows that .A C C y/ \ . int D/ D ;,
and so y 2 GHe-PMin.A; C /. Hence S-PMin.A; C / GHe-PMin.A; C /:
Assume that A is closely C -convex and take y 2 GHe-PMin.A; C /. Then
there exists D 2 DC such that .A y/ \ . int D/ D ;. It follows that
y 2 A \ WMin.A; D/. Since A is closely C -convex and C D, A is closely
D-convex. From (iii) we get y 2 D C nf0g such that y 2 arg minA y . Because
D C n f0g C # , we obtain that y 2 S-PMin.A; C /:
(v) The first equality in (2.32) is given in (2.27). For the equality He-PMin.A; / D
He-PMin.ACC; / use a similar argument to that used in (iv) (possibly taking
into account that D DC ).
Take y 2 arg minA y for some y 2 Y with 2 WD inf y . / > 0 and
set V WD fy 2 Y j jhy; y ij < g 2 NY . Clearly, D WD PV 2 D .
Because y 2 D C n f0g, from (iii) we get y 2 WMin.A; D/, and so y 2
He-PMin.A; /. Hence the inclusion in (2.33) holds.
The proof of the equality in (2.33) for A closely C -convex is similar to the
proof of the corresponding equality in (2.30).
(vi) The arguments used at the beginning of the proof of (v) show that D DC :
t
u
Proposition 2.4.6. Let A Y be nonempty.
(i) One has
S-PMin.A; C / Hu-PMin.A C C; C / Hu-PMin.A; C / Be-PMin.A; C /

D Be-PMin.A C C; C / (2.35)
with Hu-PMin.A; C / D Be-PMin.A; C / if A is closely C -convex, and
Be-PMin.A; C / D Bo-PMin.A C C; C / Bo-PMin.A; C / Min.A; C /;

(2.36)
GHe-PMin.A; C / Be-PMin.A; C /: (2.37)
Moreover, if Y has the property that for any closed convex cone K Y there
exists y 2 Y such that hy; y i > 0 for every y 2 K n.K/ (for example if Y
is a separable normed vector space) then S-PMin.A; C / D Hu-PMin.A; C /:
(ii) Assume that is a base of C and Y is a locally convex space. Then

Sup-PMin.A; C / He-PMin.A; /. Moreover, if is bounded, then
Sup-PMin.A; C / D He-PMin.A; /:
(iii) Assume that is a compact base of C and Y is a locally convex space. Then
S-PMin.A; C / D Hu-PMin.A; C / and He-PMin.A; / D GHe-PMin.A; C /:
Moreover, if A is closely C -convex, then S-PMin.A; C / D Be-PMin.A; C /,

while if A is closely convex then S-PMin.A; C / D Bo-PMin.A; C /:
(iv) Assume that StrMin.A; C / ¤ ;. Then StrMin.A; C / D Min.A; C /. If C is
closed (or, more generally, cl C \ .C / D f0g), then
StrMin.A; C / D Hu-PMin.A; C / D Hu-PMin.A C C; C /I (2.38)
if C # ¤ ; then
StrMin.A; C / D S-PMin.A; C / D GHe-PMin.A; C /; (2.39)
and StrMin.A; C / D He-PMin.A; / for every base of C if, furthermore,

Y is a locally convex space.
Proof. (i) Take y 2 S-PMin.A; C /. Then y 2 A and there exists y 2 C # such
that hy; y i hy; y i for all y 2 A. It follows that 0 hy; y i for all y 2
E1 WD cl conv cone.A C C y/, and so E1 \ .C / D f0g, because y 2 C # .
Hence the first inclusion in (2.35) holds.
Take y 2 Hu-PMin.ACC; C / . ACC /. Hence y D yCv with y 2 A, v 2
C . Then v 2 .A C C y/ \ .C / .cl conv cone.A C C y// \ .C / D
f0g. Therefore, v D 0, and so y 2 A. It follows that y 2 Hu-PMin.A; C /:
Take y 2 Hu-PMin.A; C /. Then .A y/ [ C E2 WD cl
conv cone Œ.A y/ [ C , whence .A y/ C C E2 . It follows that
F WD cl cone.A C C y/ E2 , and so y 2 Be-PMin.A; C /. Assuming
that A is closely C -convex and y 2 Be-PMin.A; C /, we have that F is a
convex cone and F \ .C / D f0g: Since .A y/ [ C A C C y, it follows
that E2 F , and so y 2 Hu-PMin.A; C /:
The equalities in (2.35) and (2.36) follow directly from the definitions of the
corresponding sets.
Take y 2 Bo-PMin.A C C; C /. Then y D y C v for some y 2 A, v 2 C .
Then v 2 .A C C y/ \ .C / .C / \ cl cone .A C C y/ D f0g.
Hence y 2 A, and so y 2 Bo-PMin.A; C /:
Take y 2 Bo-PMin.A; C /. Since cl cone .A y/ A y, it follows that
y 2 Min.A; C /:
Take y 2 GHe-PMin.A; C / D GHe-PMin.A C C; C /. Then there exists
D 2 DC such that .A C C y/ \ . int D/ D ;, and so ŒRC .A C C y/ \
. int D/ D ;, whence Œcl cone.A C C y/ \ . int D/ D ;. Hence y 2
Be-PMin.A; C /. Therefore, (2.37) holds.
Assume now that for any closed convex cone K Y there exists y 2 Y
such that hy; y i > 0 for every y 2 K n .K/, and take y 2 Hu-PMin.A; C /.
Then E3 \ .C / D f0g, or, equivalently, .C n f0g/ \ .E3/ D ;, where E3 WD
cl conv cone Œ.A y/ [ C . Then there exists y 2 Y such that hy; y i > 0
for every y 2 E3 n.E3 /. Because C E3 , if y 2 C nf0g then y 2 E3 n.E3 /,
and so hy; y i > 0. Hence y 2 C # . Therefore, y 2 S-PMin.A; C /:
(ii) Assume that is a base of C and Y is a locally convex space. Because is
a base of C , there exists V0 2 NYc such that .2V0 / \ D ;, or, equivalently,
V0 \ .V0 / D ;. Take y 2 Sup-PMin.A; C /, that is y 2 A and for every
V 2 NY there exists U 2 NY such that K \ .U C / V . Therefore, there
exists U0 2 NYc such that U0 V0 and ŒK \ .U0 / K \.U0 C / V0 .
It follows that K \ .U0 / V0 \ .U0 / V0 \ .V0 / D ;. Hence
y 2 He-PMin.A; /:
Assume, moreover, that is bounded and take y 2 He-PMin.A; /.
Then there exists V0 2 NY such that K \ .V0 / D ;. Suppose that
y … Sup-PMin.A; C /. Then there exists V1 2 NYc such that for every U 2 NYc
one has K \ .U C / 6 V0 . Hence, for every U 2 NYc there exists yU 2 U ,
tU 2 RC , zU 2 such that yU tU zU 2 K n V0 . Clearly, tU > 0 for every U 2
NYc with U V0 ; otherwise we get the contradiction yU 2 .K n V0 / \ V0 D ;:
Taking p WD pV0 the Minkowski functional of V0 , p is a continuous seminorm
with int V0 D fy 2 Y j p.y/ < 1g (see Proposition 6.2.1). Clearly,
tU1 yU zU 2 K. Because K \ .V0 / D ; we get tU1 yU … V0 for
U V0 , and so tU1 p.yU / D p.tU1 yU /
1 for such U . It follows that
tU p.yU / for U V0 . Hence .tU /U 2NYc ! 0. Since .yU /U 2NYc ! 0 and
.zU /U 2NYc is bounded, it follows that .yU tU zU /U 2NYc ! 0, contradicting
the fact that yU tU zU 2 K n V0 for every U . This contradiction shows that
y 2 Sup-PMin.A; C /:
(iii) Assume that is a compact base of C and Y is a locally convex space.
Take y 2 Hu-PMin.A;
C /, that is y 2 A and E4 \ .C / D f0g, where
E4 WD cl conv cone .A y/ [ C : It follows that E4 \ . / D ;. Using a
separation theorem, there exist y 2 Y and ˛; ˇ 2 R such that
˝ ˛
8y 2 .A y/ [ C; v0 2 ; t 2 RC W t hy; y i
˛ > ˇ
v0 ; y :
(2.40)
It follows that ˛ 0, 2 WD inf y > 0; whence y 2 C # , and hy y; y i
0 for every y 2 A. Hence y 2 S-PMin.A; C /: Taking into account (2.35) we

obtain that S-PMin.A; C / D Hu-PMin.A; C /:
Take y 2 GHe-PMin.A; C /. There exists D 2 DC such that .A y/ \
. int D/ D ;. Clearly, int D. Since is compact and Y is a l.c.s.,
there exists V 2 NY such that C V int D. Taking D 0 WD PV we have
that D 0 2 D and y 2 WMin.A; D 0 /, and so y 2 He-PMin.A; /. Hence
He-PMin.A; / D GHe-PMin.A; C /:
Assume now that A is closely C -convex and take y 2 Be-PMin.A; C /,

that is y 2 A and cl cone.A C C y/ \ .C / D f0g. Since A is closely
C -convex, we obtain that cl cone .A C C y/ is a closed convex cone. From
our hypothesis we have that cl cone .A C C y/ \ . / D ;. Using a
separation theorem, there exist y 2 Y and ˛; ˇ 2 R such that (2.40) holds
with y 2 A C C y instead of y 2 .A y/ [ C . It follows that ˛ 0;
2 WD inf y > 0, whence y 2 C # , and hy y; y i
0 for every y 2 A.
Hence y 2 S-PMin.A; C /. Hence Be-PMin.A; C / D S-PMin.A; C /:
Assume now that A is closely convex and take y 2 Bo-PMin.A; C /, that is
y 2 A and
cl cone .A y/ \ .C / D f0g;
whence cl cone .A y/ \ . / D ;. Since A is closely convex,

cl cone .A y/ is convex. Using a separation theorem, as above, we get
y 2 S-PMin.A; C /. Hence S-PMin.A; C / D Bo-PMin.A; C /:
(iv) Assume now that StrMin.A; C / ¤ ;. Because C is pointed we have that
StrMin.A; C / D fyg for some y 2 A; hence A y C C . Clearly,
.A y/ \ .C / C \ .C / D f0g, and so y 2 Min.A; C /. Conversely,
take y 2 Min.A; C /I then f0g D .A y/ \ .C / 3 y y, and so
y D y 2 StrMin.A; C /. Therefore, StrMin.A; C / D Min.A; C /:
Since A y C , we have that .A y/ [ C D .A C C y/ [ C D C:
Assume, moreover, that C is closed; then
cl conv cone Œ.A C C y/ [ C D C;
and so y 2 Hu-PMin.A; C /. Then (2.38) follows from Min.A; C / D fyg and

relations (2.35), (2.36).
Assume that C # ¤ ;. Because y 2 A y C C , clearly, arg minA y D fyg
for every y 2 C # . Hence S-PMin.A; C / D fyg. The second equality in (2.39)
follows using Proposition 2.4.5 (iv).
Assume that Y is a l.c.s. and is a base of C . Since 0 … cl , there exists
y 2 Y with inf y > 0. Hence y 2 C # , and so arg minA y D fyg.
Using Proposition 2.4.5 (v) we have that y 2 S-PMin.A; C /. Using (2.39) and
Proposition 2.4.5 (vi) we get He-PMin.A; / D fyg: t
u
It is worth observing that for y 2 Y n f0g and C WD Ky WD f0g [
fy 2 Y j hy; y i > 0g we have that C # D Py and S-PMin.A; C / D Min.A; C / D
arg minA y ; and so all the efficiency sets used in Propositions 2.4.5, 2.4.6,
excepting StrMin, reduce to arg minA y . Moreover, StrMin.A; C / ¤ ; if and only
if y has a unique minimum point on A:
The inclusions in assertions (i), and (iii), (iv) of Proposition 2.4.5, as well as
the corresponding equalities for C closed or A convex, respectively, are established
in [175, Theorem 3.1]. The most part of the inclusions in assertions (i) and (iii)
of Proposition 2.4.6, as well as the mentioned equalities, can be found [392]; note
that the hypothesis that is compact can be replaced by the fact that is weakly
compact (for this just apply the corresponding result for the weak topology on
Y ). Assertion (ii) of Proposition 2.4.6 can be found in [618], while the equality
S-PMin.A; C / D Hu-PMin.A; C / in Proposition 2.4.6 (i) is obtained in [271,
Theorem V.2.4]; the case of separable normed vector spaces in (i) is based on
the following extension of the Krein–Rutman theorem obtained by Hurwicz [271,
Lemma V.2.2].
Theorem 2.4.7. Let .Y; kk/ be a separable normed vector space and C Y be
a closed convex cone. Then there exists y 2 Y such that hy; y i > 0 for every
y 2 C n .C /:
Proof. If C n.C / D ; (that is C is a linear space) we can take y D 0. So, assume
that C n .C / 6D ;: Because C CC D C , C C is not trivial, too. The set C1 WD
C C \ UY is a weakly closed subset of UY (hence C1 is w -compact). Because Y
is separable, the weak topology on UY is metrizable (see [108, Theorem V.5.1]),
and so C1 (being w -compact) is w -separable. Let A D fy1 ; y2 ; : : :g C1 be
w -dense in C1 : Take
1
X 1
y WD y I
2k k
kD1
the series is strongly convergent because it is absolutely convergent and Y is a

Banach space. Clearly, y 2˝ C1 ˛ C C . Assume that there exists y 2 C n.C / such
that hy; y i D 0. Because y; yk
0 for every k
1, we obtain that hy; y i D 0
for every y 2 A: With the set A being w -dense in C1 , we obtain that hy; y i D 0
for every y 2 C1 , and so hy; y i D 0
0 for every y 2 C C D RC C1 .
Therefore, we get the contradiction y 2 C CC D C: t
u
In order to describe weak and proper minimality in a unified way, we use the
notation of Q-minimal points (compare Ha [228]).
Definition 2.4.8. Assume that D Y is a proper cone with nonempty interior and
put Q WD int D. We say that y is a Q-minimal point of A (y 2 QMin.A; C /) if
A \ .y Q/ D ;
or, equivalently,
.A y/ \ .Q/ D ;:
In the paper by Gerstewitz and Iwanow [197] properly minimal elements are
defined using a set Q Y with 0 2 bd Q and cl Q C .C n f0g/ int Q. This
approach is related to the well-known concept of dilating cone (or a dilation) of C :
Definition 2.4.9. Suppose that D Y is a proper cone with nonempty interior and
put Q WD int D. Q is said to be a dilation of C , or dilating C if it contains C n f0g.
Remark 2.4.10. Makarov and Rachkovski [409] studied more detailed some con-
cepts of proper efficiency and introduced the notion of B-efficiency, i.e., efficiency
w.r.t. a family of dilations of C . Namely, given B 2 F .C /, where F .C / is
the class of families of dilations of C , y is said to be a B-minimal point of A
(y 2 B Min.A; C /) if there exists B 2 B such that
.A y/ \ .B/ D ;:
It has been established that Borwein proper efficiency, Henig global proper
efficiency, Henig proper efficiency, super efficiency and Hartley proper efficiency
are B-efficiency with B being appropriately chosen family of dilating cones.
The reader will see that in contrast with B-efficiency, the concept introduced in
Definition 2.4.8 includes not only some concepts of proper efficiency among which
are these ones considered in [409] but also the concepts of strong efficiency and
weak efficiency.
In order to study the relationships between weakly / properly minimal points and
Q-minimal points let Y be a n.v.s. and as before a base of C . Setting
ı WD ı WD d.0; / D inffkk j 2 g > 0;
for each 0 < < ı, we can associate to C a convex, pointed and open set V ,
defined by
V WD cone. C BVY /: (2.41)
For each scalar " > 0, we also associate to C an open set C."/
C."/ WD fy 2 Y j d.y; C / < "d.y; C /g:
We are going to show that the weakly / properly minimal points introduced in
Definitions 2.4.2 and 2.4.4 are in fact Q-minimal points (Definition 2.4.8) with Q
being appropriately chosen sets. The following result is shown in [228].
Theorem 2.4.11. (a) y 2 WMin.A; C / iff y 2 QMin.A; C / with Q D intC .
(b) y 2 S-PMin.A; C / iff y 2 QMin.A; C / with Q D fy 2 Y j y .y/ > 0g and
y 2 C #.
(c) y 2 Hu-PMin.A; C / iff y 2 QMin.A; C /, with Q D Y n cl conv coneŒ.A
y/ [ C .
(d) y 2 Be-PMin.A; C / iff y 2 QMin.A; C /, with Q D Y n cl coneŒ.A y/ C
C .
(e) y 2 Ha-PMin.A; C / iff y 2 QMin.A; C / with Q D C."/ for some " > 0.
(f) y 2 Bo-PMin.A; C / iff y 2 QMin.A; C /, with Q being some dilation of C .
(g) y 2 GHe-PMin.A; C / iff y 2 QMin.A; C /, with Q D int D, being some
dilation of C , where D is a proper pointed convex cone in Y .
(h) (supposing that Y is a n.v.s.) y 2 He-PMin.A; C / iff y 2 QMin.A; C / with

Q D V and is some scalar satisfying 0 < < ı D d.0; /.
(i) (supposing that Y is a n.v.s. and C has a bounded base ) y 2 Sup-PMin.A/
iff y 2 QMin.A; C / with Q D V and is some scalar satisfying 0 < < ı D
d.0; /.
Proof. Using Definitions 2.4.4 and 2.4.8 one can easily prove the assertions (a)–
(d) and (g). The assertions (e)–(f) are formulated in a slightly different form as
established by Makarov and Rachkovski [409].
We prove now the assertion (h), namely, we show that y 2 He-PMin.A; C / iff
there is a scalar with 0 < < ı such that
.A y/ \ .V / D ;: (2.42)
Recall that by definition, y 2 He-PMin.A; C / iff
cl cone.A y/ \ . C "BY / D ;: (2.43)
It is also known [619] that y 2 He-PMin.A; C / iff
.A y/ \ .S n / D f0g (2.44)
for some integer n 2 N , where S n D cl cone. C ı=.2n/BY /: Now, suppose that

y 2 He-PMin.A; C /. Then (2.43) holds. Without loss of generality we can assume
that 0 < " < ı. We show that (2.43) holds with D ". Suppose to the contrary that
there is y 0 2 Ay such that y 0 2 V" . Clearly, y 0 2 cl cone.Ay/\.cl cone. C
"BY //. On the other hand, as 0 < D " < ı and by the definition of ı, 0 … V" .
Hence y 0 ¤ 0. This is a contradiction to (2.43). Next, suppose that (2.42) holds for
some . Let n be an integer satisfying n 1 > ı=.2/ or ı=.2n 2/ < . By (2.42)
we have
.A y/ \ .Vı=.2n2/ / .A y/ \ .V / D ;:
Then .A y/ \ .Vı=.2n2/ [ f0g/ D f0g. On the other hand, [619, Lemma 2.1]
states that if .A y/ \ .Vı=.2n2/ [ f0g/ D f0g, then .A y/ \ .S n / D f0g.
Thus, (2.44) holds and therefore, y 2 He-PMin.A; C /, as it was to be shown.
To complete the proof note that the last assertion (i) of this theorem follows from
(h) and the assertion (ii) in Proposition 2.4.6. t
u
Remark 2.4.12. The assertion (h) in the above theorem is inspired by the definition
of Henig properly minimal points for sets in locally convex spaces given by Gong
in [213]. One can deduce that any Henig properly minimal point is a global Henig
properly minimal point.
Furthermore, we mention solution concepts for vector optimization problems
introduced by ElMaghri–Laghdir[175] where it is not supposed that the ordering
cone C Y is pointed and closed. This concept is based on a generalization of

the concept of dilations of a cone C Y (see Definition 2.4.9). In the text below
we follow the presentation in [175]. We assume that X; Y; Z are topological vector
spaces and S is a subset of X . For the following notations and results the lineality
of C , defined by
l.C / WD C \ .C /;
is very important. Of course, C is pointed if l.C / D f0g.

Unlike the assumptions made before we assume in the sequel in this section that
Y is ordered by the proper convex cone C (C is proper if C ¤ l.C /, or equivalently,
C is not a linear subspace of Y ).
Furthermore, using l.C / we introduce
C & WD fy 2 Y j hy; y i > 0 8y 2 C n l.C /g D .C n l.C //# :
Hence, if C is pointed then C & D C # . Note that Theorem 2.4.7 gives sufficient
conditions for C & ¤ ;.
We use the notations y 5 y 0 if y 0 y 2 C ; y < y 0 if y 0 y 2 int C and
furthermore, y y 0 if y 0 y 2 C n l.C /:
In the following we consider a proper vector-valued objective function f W X !
Y , S X and use the notation
f .S / WD ff .x/ j x 2 S \ dom f g Y:
Consider now the vector optimization problem
minimze f .x/ subject to x 2 S: (VP)
Using the lineality of C and without assuming the pointedness and closedness
of C we introduce the following solution concepts for (VP). These concepts are
extensions of the solution concepts introduced in Definitions 2.4.1, 2.4.2 and 2.4.4.
Definition 2.4.13. x 2 S \ dom f is
• strongly l(C)-minimal if f .x/ 5 f .x/ for all x 2 S , or equivalently
f .S / f .x/ C C;
• Pareto l(C)-minimal if f .x/ 5 f .x/ ) f .x/ 5 f .x/ for all x 2 S , or

equivalently
f .S / \ .f .x/ .C n l.C /// D ;;
• weakly l(C)-minimal if f .x/ 6< f .x/ for all x 2 S , or equivalently

f .S / \ .f .x/ int C / D ;;
• l(C)-properly minimal if there exists D Y a proper convex cone with C n

l.C / int D such that x is efficient with respect to D, f .S / \ .f .x/ D n
l.D// D ;:
Definition 2.4.14. Suppose that D Y is a proper convex cone with nonempty
interior. Put Q WD int D. Q is said to be a generalized dilation of C or generalized
dilating C if it contains C n l.C /.
Set
QC WD fD Y j D is a proper convex cone with C n l.C / int Dg :
Lemma 2.4.15. Let C Y be a proper convex cone. Then QC ¤ ; if and only if

C & ¤ ;. Moreover, if C & ¤ ; then
C C int D D int D 8D 2 QC (2.45)
and
C & D [D2QC .D C n f0g/: (2.46)
Proof. Assume that QC ¤ ; and take D 2 QC . Since D is proper, int D ¤ Y:

Take q0 2 Y n int D. By a separation theorem there exists y 2 Y such that
hq0 ; y i 0 < hy; y i for all y 2 int D. It follows that y 2 C & :
Conversely, assume that C & ¤ ; and take y 2 C & : Consider D WD fy 2 Y j
hy; y i
0g. From the very definition of C & we have that C n l.C / fy 2 Y j
hy; y i > 0g D int D:
Let us prove (2.45). First note that C cl .C n cl C /. Indeed, there exists k0 2
C n l.C /. Let k 2 C . Then k C
k0 2 C C .C n l.C // D C n l.C / for every
> 0. The claim follows taking the limit for

! 0. Then, taking D 2 QC and
using repeatedly [79, Lemma 2.5] we get
int D C C int D D cl.C n l.C // C int D D .C n l.C // C int D int D;
and so (2.45) holds.

Take now D 2 QC ; then C n l.C / int D, and so C & D .C n l.C //#
.int D/# D D C n f0g. Hence the inclusion holds in (2.46). Let now y 2 C & .
Then D WD fy 2 Y j hy; y i
0g 2 QC and y 2 D C n f0g. It follows that (2.46)
holds. t
u
For A Y we set
Es .A/ WD EsC .A/ WD fa 2 A j A a C C g;

Ee .A/ WD EeC .A/ WD fa 2 A j A \ .a .C n l.C /// D ;g;
Ew .A/ WD EwC .A/ WD fa 2 A j A \ .a int C / D ;g;
˚
Ep .A/ WD EpC .A/ WD [ EeD .A/ j D 2 QC :
Since int C C n l.C / int D, we have that
Ep .A/ Ee .A/ Ew .A/; (2.47)
the last inclusion makes sense for int C ¤ ;. It is worth to observe that
˚
Ep .A/ D [ EwD .A/ j D 2 QC : (2.48)
Indeed, the inclusion in (2.48) follows from the last inclusion in (2.47). Take
a 2 EwD .A/ for some D 2 QC and consider D 0 WD f0g [ int D. Then D 0 2 QC
0 0 0
and a 2 EwD .A/ because D 0 D. Since EwD .A/ D EeD .A/, we obtain that
a 2 Ep .A/:
Lemma 2.4.16. E .A/ D A\E .ACC / for 2 fs; e; w; pg. Moreover, if l.C / D
f0g, that is C is pointed, then E .A/ D E .A C C / for 2 fs; e; pg. Generally,
Ew .A/ D A \ bd.A C C / (for int C ¤ ;/; and so Ew .A/ and Ew .A C C / (D
bd.A C C /) might be different (take A D f0g).
Proof. For D s use the fact that A a C C ) A A C C a C C . For D e
use the fact that C n l.C / D C C .C n l.C //, while for D w use the fact that
C C int C D int C:
For D p one uses (2.45). If a 2 Ep .A/ then there exists D 2 QC such
that A \ .a int D/ D ;: Hence 0 … A a C int D D .A C C / a C int D
by (2.45), and so a 2 Ep .A C C /. Conversely, let a 2 A \ Ep .A C C /. Then
0 … .A C C / a C int D D A a C int D, and so a 2 Ep .A/:
Assume that l.C / D f0g. We have to prove that E .A C C / A for 2
fs; e; pg. Assume that a C k 2 E .A C C / for some a 2 A and k 2 C n f0g. For
D s we get a C k 5 a, whence the contradiction k 2 C . For D e we get
the contradiction a 2 .A C C / \ .a C k .C n f0g/. For D p, we have that
Ep .A/ Ee .A/ A: t
u
Generally, Es .A/ D ;:
Lemma 2.4.17. If Es .A/ ¤ ; then Es .A/ D Ee .A/. Moreover, if C & ¤ ; then
Es .A/ D Ep .A/ D Ee .A/:
Proof. Because 0 … C n l.C / D C C .C n l.C // we have that
C \ . .C n l.C /// D ;:
2.5 Vector Optimization Problems with Variable Ordering Structure 43
Fix a 2 Es .A/. Let a 2 Es .A/. Then A a C C , and so
A\.a.C nl.C /// .aCC /\.a.C nl.C /// D aCŒC \ ..C n l.C /// D ;:
Hence a 2 Ee .A/, and so Es .A/ Ee .A/:

Take now a 2 Ee .A/ A a C C . Hence a D a C k for some k 2 C . If
k … l.C / then a 2 A \ ..a .C n l.C ///, contradicting the fact that a 2 Ee .A/.
Hence k 2 l.C /, and so A a C C D a k C C a C C , which shows that
a 2 Es .A/:
Assume, moreover, that C & ¤ ;, and take y 2 C & : It follows that D WD fy 2
Y j hy; y i
0g 2 QC . Using (2.45) we have that C \ . int D/ D ;, and so
A \ .a int D/ .a C C / \ .a int D/ D a C .C \ . int D// D ;I
therefore, a 2 Ep A. Hence Es .A/ Ep .A/. Since always Ep .A/ Ee .A/ we get

the conclusion. t
u
The corresponding sets of solutions for (VP) are denoted by Es .f; S /; Ee .f; S /,
Ew .f; S /, Ep .f; S /, respectively. More precisely,
E .f; S / WD fx 2 S \ dom f j f .x/ 2 E .f .S \ dom f //g ; 2 fs; e; w; pg:
2.5 Vector Optimization Problems with Variable Ordering

Structure
Yu introduced in [606] nondominated solutions of vector optimization problems

with variable ordering structure based on general domination set mappings, compare
also Chen, Huang, Yang [91]. Vector optimization problems with variable ordering
structure are studied intensively by Eichfelder in [162, 164–166], Eichfelder,
Ha [168] where corresponding solution concepts, characterizations by scalarization
methods, optimality conditions and numerical procedures are presented. Eich-
felder [163, 166] gives a very detailed overview on solution concepts for vector
optimization problems with variable ordering structure and presents a complete
characterization of these solution concepts (see also Eichfelder, Kasimbeyli [170]
and Eichfelder, Gerlach [167]).
Let X and Y be Banach spaces, ; ¤ S X , f W X ! Y and let C W X Y
be a set-valued map such that for each x 2 X , C.x/ is a nonempty convex set with
0 2 bd C.x/.
We consider the following vector optimization problems with variable ordering
structure
v-minimize f .x/ subject to x 2 S:

The solution concept for this problem is given in the following definition (com-
pare [91, Definition 1.15]):
Definition 2.5.1 (v-Minimal Points, Weakly v-Minimal Points). Let C W X Y
be a set-valued map with 0 2 bd C.x/, C.x/ a convex set for all x 2 X , S X and
f W S ! Y . An element x 2 S is said to be a v-minimal point of f w.r.t. C./ if
.f .S / f .x// \ ..C.x/ n f0g// D ;:
The set of all f .x/ with x a v-minimal point of f w.r.t. C./ is denoted by
Min.f .S /; C.//.
Suppose that for all x 2 X , int C.x/ ¤ ;. An element x 2 S is said to be a
weakly v-minimal point of f w.r.t. C./ if
.f .S / f .x// \ . int C.x// D ;:
The set of all f .x/ with x a weakly v-minimal point of f w.r.t. C./ is denoted by
WMin.f .S /; C.//.
Remark 2.5.2. For further solution concepts, especially nondominated elements, of
vector optimization problems with variable ordering structure see Eichfelder [166].
In the case that we ask for v-minimal points of f w.r.t. C./ the vector
optimization problem with variable ordering structure is given by
Min.f .S /; C.//: (VPv )
When we are looking for weakly v-minimal points of f w.r.t. C./ we study the
problem
WMin.f .S /; C.//:
The following relationships between v-minimal solutions of the vector optimiza-

tion problem (VPv ) and solutions of suitable scalarized problems are shown by
Chen, Huang and Yang [91, Theorem 2.18].
Theorem 2.5.3. Consider the vector optimization problem with variable ordering
structure (VPv ), where C W X Y is a set-valued map such that for each x 2 X ,
C.x/ is a convex subset of Y with 0 2 bd C.x/ and int C.x/ ¤ ;. Then:
(a) Let x 2 S . Suppose that there exists y 2 Y with y .y/ > 0 for all y 2
C.x/ n f0g such that x 2 S is a minimal solution of the scalar optimization
problem
min y .f .x//: (Py )

x2S
2.6 Solution Concepts in Set-Valued Optimization 45
Then x is a v-minimal point of f w.r.t. C./ concerning the vector optimization

problem (VPv ).
(b) Let f .S / be a convex subset of Y and x a v-minimal point of the problem
(VPv ). Then there exists y 2 Y satisfying y .y/ > 0 for all y 2 int C.x/,
such that x is a minimal solution of the scalar optimization problem .Py /.
Remark 2.5.4. Characterizations of solutions of general vector optimization prob-
lems with variable ordering structure by means of nonlinear scalarizing functionals
are given by Eichfelder in [163, 164, 166] and by Eichfelder, Ha [168].
2.6 Solution Concepts in Set-Valued Optimization
Unless otherwise mentioned, let Y be a linear topological space, partially ordered

by a proper pointed convex closed cone C . Let P.Y / D 2Y be the power set of Y .
We consider a set-valued optimization problem with a general geometric con-
straint:
minimize F .x/ subject to x 2 S; (SP)
where S is a subset of X , X is a linear space and the cost mapping F W S

Y is a set-valued mapping. As already introduced, we use the notations F .S / D
[x2S F .x/ and dom F D fx 2 S j F .x/ ¤ ;g.
In Sects. 2.6.1, 2.6.2 and 2.6.3 we introduce different solution concepts for the
problem .SP/. Furthermore, in Sect. 2.6.4 we present the embedding approach by
Kuroiwa [353, 354, 357], in Sect. 2.6.5 we discuss solution concepts with respect
to abstract preference relations by Bao and Mordukhovich [28], in Sect. 2.6.6
we introduce solution concepts for set-valued optimization problems with variable
ordering structure, in Sect. 2.6.7 we study approximate solutions of set-valued
optimization problems and finally, in Sect. 2.7 we discuss relations between the
solution concepts.
2.6.1 Solution Concepts Based on Vector Approach
First, we introduce a solution concept where “minimization” in .SP/ is to be

understood with respect to the partial order C defined in (2.9). In contrast to
single-valued functions, for every x 2 dom F there are many distinct values y 2 Y
such that y 2 F .x/. Hence, in the first approach, when studying minimizers
of a set-valued mapping, we fix one element y 2 F .x/, and formulate the
following solution concept based on the concept of Pareto minimality introduced
in Definition 2.4.1.
Definition 2.6.1 (Minimizer of (SP)). Let x 2 S and .x; y/ 2 graph F . The pair
.x; y/ 2 graph F is called a minimizer of the problem (SP) if y 2 Min F .S /; C .
Furthermore, also the other notions of weakly / properly minimal points for sets
(see Definitions 2.4.2 and 2.4.4) naturally induce corresponding notions of weak /
proper minimizers to the corresponding set optimization problems (see Ha [228]).
Definition 2.6.2 (Weak Minimizer of (SP)). Let x 2 S and .x; y/ 2 graph F .
.x; y/ 2 graph F is called a weak minimizer of the problem (SP) if y 2
The pair
WMin F .S /; C .
Let D Y be as before (see Definition 2.4.8) a proper cone with nonempty
interior and Q WD int D.
Definition 2.6.3 (Q-Minimizer of (SP)). Consider the set-valued optimization
problem (SP). Let x 2 S and .x; y/ 2 graph F . We say that .x; y/ is an
S-proper minimizer (Hurwicz proper minimizer, Hartley proper minimizer,
Benson proper minimizer, Borwein proper minimizer, Henig global proper
minimizer, Henig proper minimizer, super minimizer and Q-minimizer, respec-
tively) of .SP / if y is an S-properly minimal (Hurwicz properly minimal, Hartley
properly minimal, Benson properly minimal, Borwein properly minimal, Henig
global properly minimal, Henig properly minimal, super efficient and Q-minimal,
respectively) point of F .S /, i.e., y 2 S-PMin.F .S /; C / ( y 2 Hu-PMin.F .S /; C /,
y 2 Ha-PMin.F .S /; C /, y 2 Be-PMin.F .S /; C /,
y 2 Bo-PMin.F .S /; C /, y 2 GHe-PMin.F .S /; C /, y 2 He-PMin.F .S /; C /,
y 2 Sup-PMin.F .S /; C /, y 2 Q-Min.F .S /; C /, respectively).
Moreover, especially in Chaps. 8 and 15 we study set-valued optimization
problems, where the set-valued objective map F W X Y is to be maximized
over the feasible set S X (X is a linear space)
maximize F .x/ subject to x 2 S: (SPmax )
Analogously to Definitions 2.6.1 and 2.6.2 we now introduce maximizers and

weak maximizer of .SPmax /.
Definition 2.6.4 (Maximizer of .SPmax /). Let x 2 S and .x; y/ 2 graph F .
The pair .x; y/ 2 graph F is called a maximizer of the problem (SPmax ) if
y 2 Max F .S /; C .
Definition 2.6.5 (Weak Maximizer of .SPmax /). Consider the set-valued opti-
mization problem .SPmax /. Let x 2 S and .x; y/ 2 graph F . We say that .x; y/
is a weak maximizer of .SPmax / if y is a weakly maximal point of F .S /, i.e.,
y 2 WMax.F .S /; C /.
Furthermore, we consider set-valued optimization problems with a special
structure concerning the restrictions, namely inequality restrictions: Let X , Y , Z
be real locally convex Hausdorff spaces, Y , Z be ordered by proper pointed closed
convex cones C , K, respectively, F W X Y and G W X Z. Under these
assumptions we study a set-valued optimization problem of the following form (see

Tasset [570]):
minimize F .x/ subject to x 2 S; (SPT )
where M X is a set satisfying M dom F \ dom G and
S WD fx 2 M j G.x/ \ .K/ ¤ ;g: (2.49)
For set-valued problems .SPT / with a feasible set S given by (2.49) we derive
duality assertions in Sect. 8.1 using the following solution concept with respect to
the quasi(-relative) interior of a cone C Y .
Let B Y be a nonempty convex set; the quasi interior of B is
qi B WD fy 2 B j cl .RC .B y// D Y g
and the quasi-relative interior of B is
qri B WD fy 2 B j cl .RC .B y// is a linear spaceg :
Because cl .RC .B y// cl affB y, we have that cl affB D Y whenever qi B ¤

;. In fact we have
0 2 qi.B B/ ” cl affB D Y H) qi B D qri B: (2.50)
It is worth to observe that for y0 2 B we have that
y0 … qri B ” 9y 2 Y W inf y .B/

hy0 ; y i < sup y .B/I (2.51)
in particular,
y0 2 B n qri B H) 9y 2 Y n f0g W inf y .B/ D hy0 ; y i : (2.52)
Note that in the above implications we do not assume that qri B ¤ ;. Note
that (2.51) covers [72, Theorem 2.7].
Observe also that for B D C a convex cone,
y 2 qi C; y 2 C C n f0g H) hy; y i > 0: (2.53)
Indeed, if hy; y i D 0 then hy 0 y; y i

0 for every y 0 2 C , and so hy 00 ; z i
0
for every y 00 2 cl .RC .C y// D Y: We get so the contradiction y D 0:
Using these notations Tasset [570] introduced the following solution concept for
the set-valued problem .SPT / with restrictions given by (2.49) .
Definition 2.6.6 (Quasi-Weak Minimizer of .SPT /). Assume qi C ¤ ; and con-

sider the set-valued optimization problem .SPT / with restrictions given by (2.49).
Let x 2 S and .x; y/ 2 graph F . The pair .x; y/ is called a quasi-weak minimizer
of the problem (SPT ) with restrictions given by (2.49) if F .S / \ .y qi C / D ;,
and we denote this by y 2 Min.F .S /; qi C /:
2.6.2 Solution Concepts Based on Set Approach
Although the concept of a minimizer of the set-valued problem .SP/ given in

Definitions 2.6.1 and 2.6.3 is of mathematical interest, it cannot be often used in
practice. It is important to mention that a minimizer .x; y/ depends on only certain
special element y of F .x/ and other elements of F .x/ are ignored. In other words,
an element x 2 S for that there exists at least one element y 2 F .x/ which is a
Pareto minimal point (Definition 2.4.1) of the image set of F even if there exist
many bad elements in F .x/, is a solution of the set-valued optimization problem
.SP/. For this reason, the solution concepts introduced in Sect. 2.6.1 are sometimes
improper.
In order to avoid this drawback it is necessary to work with practically relevant
order relations for sets. This leads to solution concepts for set-valued optimization
problems based on comparisons among values of the set-valued objective map F .
First, we will introduce several order relations that are used in order to formulate
corresponding solution concepts for the set-valued problem .SP/. The set less order
relation sC is introduced independently by Young [605] and Nishnianidze [443]
(cf. Eichfelder, Jahn [169]) for the comparison of sets:
Definition 2.6.7 (Set Less Order Relation).
Let C Y be a proper closed convex and pointed cone. Furthermore, let
A; B 2 P.Y / be arbitrarily chosen nonempty sets. Then the set less order relation
is defined by
A sC B W” A B C and A C C B:
Remark 2.6.8. Of course, we have
A B C ” 8a 2 A 9b 2 B W a C b
and
A C C B ” 8b 2 B 9a 2 A W a C b:
Kuroiwa [347, 349, 351] introduced the following order relations:

Definition 2.6.9 (Lower (Upper) Set Less Order Relation). Let C Y be a

proper closed convex and pointed cone. Furthermore, let A; B 2 P.Y / be arbitrary
nonempty sets. Then the lower set less order relation lC is defined by
A lC B W” A C C B
and the upper set less order relation uC is defined by
A uC B W” A B C:
The lower set less order relation lC is illustrated in Fig. 2.2 and the upper set less
order relation uC in Fig. 2.3.
Remark 2.6.10. There is the following relationship between the lower set less order
relation lC and the upper set less order relation uC :
A lC B W” ACC B ” B A.C / ”W B uC A ” .B/ uC .A/:
Remark 2.6.11. It is easy to see that A lC B is equivalent to
A C C B C C:
Furthermore, A uC B is equivalent to
A C B C:
Fig. 2.2 Lower set less order

relation 4lC
B
Fig. 2.3 Upper set less order

relation 4uC
A
It is important to mention that
A lC B and B lC A ” A C C D B C C:
Under our assumption that C is a pointed closed convex cone it holds Min.A C
C; C / D Min.A; C / and Min.B C C; C / D Min.B; C / such that we get
A lC B and B lC A H) Min.A; C / D Min.B; C /:
Under the additional assumptions A Min.A; C / C C and B Min.B; C / C C

(domination property, see [47, 400]) we have
Min.A; C / D Min.B; C / ” A C C D B C C
and so
A lC B and B lC A ” Min.A; C / D Min.B; C /:
Similarly,
A uC B and B uC A ” A C D B C
and because of Max.A C; C / D Max.A; C / and Max.B C; C / D Max.B; C /

it holds
A uC B and B uC A H) Max.A; C / D Max.B; C /:
Under the additional assumption A Max.A; C / C and B Max.B; C / C it

holds
Max.A; C / D Max.B; C / ” A C D B C
and so
A uC B and B uC A ” Max.A; C / D Max.B; C /:
In interval analysis there are even more order relations in use, like the certainly
less order relation cC (Kuroiwa [347–349, 351], compare Eichfelder, Jahn [169]):
Definition 2.6.12 (Certainly Less Order Relation cC ). For arbitrary nonempty
sets A; B 2 P.Y / the certainly less order relation cC is defined by
A cC B W” .A D B/ or .A ¤ B; 8a 2 A 8b 2 B W a C b/:
An illustration of the certainly less order relation cC is given in Fig. 2.4.
Fig. 2.4 Certainly less order

relation 4cC
B
Fig. 2.5 Possibly less order

p
relation 4C
A B
p
Moreover, the possibly less order relation C (Kuroiwa [348, 349, 351]) is given
in the following definition:
p
Definition 2.6.13 (Possibly Less Order Relation C ). For arbitrary nonempty
p
sets A; B 2 P.Y / the possibly less order relation C is defined by
p
A C B W” .9a 2 A; 9b 2 B W a C b/:
p
The possibly less order relation C is illustrated in Fig. 2.5.
Remark 2.6.14. It is clear that A cC B implies
9a 2 A such that 8b 2 B W a C b: (2.54)
Moreover, (2.54) implies A lC B (see Definition 2.6.9) such that
A cC B H) A lC B:
Furthermore, A lC B implies
9a 2 A; 9b 2 B such that a C b: (2.55)

Taking into account Definition 2.6.13, we have

p
A cC B H) A lC B H) A C B: (2.56)
Remark 2.6.15. The relation A cC B implies
9b 2 B such that 8a 2 A W a C b: (2.57)
Moreover, (2.57) implies A uC B (see Definition 2.6.9) such that
A cC B H) A uC B:
Furthermore, A uC B implies
9a 2 A; 9b 2 B such that a C b; (2.58)
such that we get

p
A cC B H) A uC B H) A C B
taking into account Definition 2.6.13.

Furthermore, the minmax less order relation m
C is introduced for sets A, B
belonging to
F WD fA 2 P.Y / j Min.A; C / ¤ ; and Max.A; C / ¤ ;g:
Note that for instance in a topological real linear space Y for every compact set
in P.Y / minimal and maximal elements exist.
Definition 2.6.16 (Minmax Less Order Relation). Let A; B be sets belonging to
F . Then the minmax less order relation m
C is defined by
A m
C B W” Min.A; C / C Min.B; C / and Max.A; C / C Max.B; C /:
s s
The minmax certainly less order relation mc

C is introduced in the next definition:
Definition 2.6.17 (Minmax Certainly Less Order Relation). For arbitrary

A; B 2 F the minmax certainly less order relation mc
C is given by
A mc
C B W” .A D B/ or .A ¤ B; Min.A; C / C Min.B; C /
c
and Max.A; C / cC Max.B; C //:
Finally, we introduce the minmax certainly nondominated order relation mn

C (see
Jahn, Ha [295]).
Definition 2.6.18 (Minmax Certainly Nondominated Order Relation). For arbi-

trary nonempty A; B 2 P.Y / the minmax certainly nondominated order relation
mn
C is defined by
A mn
C B W” .A D B/ or .A ¤ B; Max.A; C / C Min.B; C //:
s
The set less order relation sC and the order relations lC , uC , mC , C and
mc
C are preorders. If C denotes one of these order relations, then we can define
mn
optimal solutions with respect to the preorder C and the corresponding set-valued
optimization problem is given by
C minimize F .x/; subject to x 2 S; (SP C )
where we assume again (compare (SP)) that Y is a linear topological space, partially
ordered by a proper pointed convex closed cone C , S is a subset of X , X is a linear
space, F W X Y .
Definition 2.6.19 (Minimal Solutions of (SP C ) w.r.t. the Preorder C ). An
element x 2 S is called a minimal solution of problem (SP C ) w.r.t. the preorder
C if
F .x/ C F .x/ for some x 2 S H) F .x/ C F .x/:
Remark 2.6.20. When we use the set relation lC introduced in Definition 2.6.9
in the formulation of the solution concept, i.e., when we study the set-valued
optimization problem .SP lC /, we observe that this solution concept is based
on comparisons among sets of minimal points of values of F (see Definition 2.4.1).
Furthermore, considering the upper set less order relation uC (Definition 2.6.9), i.e.,
considering the problem .SP uC / we recognize that this solution concept is based
on comparisons of maximal points of values of F (see Definition 2.4.1).
When x 2 S is a minimal solution of problem .SP lC / there does not exist
x 2 S such that F .x/ is strictly smaller than F .x/ with respect to the set order lC .
In the following we give three examples (see Kuroiwa [347]) of set-valued opti-
mization problems in order to illustrate the different solution concepts introduced in
Definitions 2.6.1 and 2.6.19.
Example 2.6.21. Consider the set-valued optimization problem
minimize F1 .x/; subject to x 2 S;
with X D R, Y D R2 , C D R2C , S D Œ0; 1 and F1 W S Y is given by

Œ.1; 0/; .0; 1/ if x D 0
F1 .x/ WD
Œ.1 x; x/; .1; 1/ if x 2 .0; 1;
where Œ.a; b/; .c; d / is the line segment between .a; b/ and .c; d /.
Only the element x D 0 is a minimal solution in the sense of Definition 2.6.19

w.r.t. lC . However, all elements .x; y/ 2 graph F1 with x 2 Œ0; 1, y D .1
x; x/ for x 2 .0; 1 and y D .1; 0/ for x D 0 are minimizers of the set-valued
optimization problem in the sense of Definition 2.6.1. This example shows that the
solution concept with respect to the set relation lC (see Definition 2.6.19) is more
natural and useful than the concept of minimizers introduced in Definition 2.6.1.
Example 2.6.22. Now we discuss the set-valued optimization problem
minimize F2 .x/; subject to x 2 S;

.1; 13 /; . 13 ; 1/ if x D 0
F2 .x/ WD
Œ.1 x; x/; .1; 1/ if x 2 .0; 1:
The set of minimal solutions in the sense of Definition 2.6.19 w.r.t. lC is the
interval Œ0; 1, but the set of minimizers in the sense of Definition 2.6.1 is given by
f.x; y/ 2 graph F2 j x 2 .0; 1; y D .1 x; x/g:
Here we observe that x D 0 is a lC -minimal solution but the set F2 .x/ (x D 0)
has no Pareto minimal points.
Example 2.6.23. In this example we are looking for minimal solutions of a
set-valued optimization problem with respect to the set relation uC introduced in
Definition 2.6.9.
uC minimize F3 .x/; subject to x 2 S; (SP uC )

ŒŒ.1; 1/; .2; 2/ if x D 0
F3 .x/ WD
ŒŒ.0; 0/; .3; 3/ if x 2 .0; 1;
where ŒŒ.a; b/; .c; d / WD f.y1 ; y2 / j a y1 c; b y2 d g.

Then a minimal solutions of .SP uC / in the sense of Definition 2.6.19 is only
x D 0. On the other hand, x 2 .0; 1 are not minimal solutions of .SP uC / in the
sense of Definition 2.6.19, but for all x 2 .0; 1 there are y 2 F3 .x/ such that .x; y/
are minimizers in the sense of Definition 2.6.1.
Further relationships between different solution concepts in set-valued optimiza-
tion are discussed in Sect. 2.7.
Applications of solution concepts based on set approach introduced in this
section are given in Sect. 15.4 concerning robustness, in Sect. 1.1 concerning
game theory. Furthermore, in Sect. 8.2 we present duality assertions for the primal
problem .SP lC /.
2.6.3 Solution Concepts Based on Lattice Structure
We recall in this section the concept of an infimal set (resp. supremal set), which is
due to Nieuwenhuis [442], was extended by Tanino [563], and slightly modified with
respect to the elements ˙1 by Löhne and Tammer [397]. We will shortly discuss
the role of the space of self-infimal sets, which was shown in [397] to be a complete
lattice. As we will see in Sect. 15.1, this complete lattice is useful for applications
of set-valued approaches in the theory of vector optimization, especially in duality
theory.
First, we recall the definitions of lower and upper bounds as well as the infimum
and supremum of a subset of a partially ordered set. Consider a partially ordered set
.Y; / and A Y . As already introduced in Definition 2.1.6, an element l 2 Y is
called a lower bound of A if l y for all y 2 A. Furthermore, an element u 2 Y is
called an upper bound of A if u
y for all y 2 A. Using lower and upper bounds,
the infimum and supremum for a subset A of a partially ordered set .Y; / is defined
in Definition 2.1.6. An element l 2 Y is called greatest lower bound or infimum of
A Y if l is a lower bound of A and for every other lower bound l of A it holds
l l. If the infimum of A exists we use the notation l D inf A for it. Analogously,
we define the least upper bound or supremum of A Y and denote it by sup A.
Based on the definition of the infimum and supremum we introduce the notion of
a complete lattice that is important for the approach in this section.
Definition 2.6.24. A partially ordered set .Y; / is called a complete lattice if the
infimum and supremum exist for every subset A Y .
A characterization of a complete lattice based on the existence of the infimum of
subsets A Y is given by Löhne [395, Proposition 1.6].
Proposition 2.6.25. A partially ordered set .Y; / is a complete lattice if and only
if the infimum exists for every subset A Y .
Let us give some examples for complete lattices (compare [395]).
Example 2.6.26. It is well known that R D R [ f˙1g equipped with the natural
order relation provide a complete lattice.
Example 2.6.27. Consider a nonempty set Y and let P.Y / D 2Y be the power set
of Y . .P.Y /; / is a complete lattice. The infimum and supremum of a nonempty
subset A P.Y / are described by
[ \
inf A D A; sup A D A: (2.59)
A2A A2A
If A is empty, we put sup A D Y and inf A D ;. Y 2 P.Y / is the least

element and ; 2 P.Y / is the greatest element in .P.Y /; /.
Example 2.6.28. Consider a linear space Y and let C .Y / be the family of all convex
subsets of Y . .C .Y /; / provides a complete lattice. The infimum and supremum
of a nonempty subset A C .Y / is described by
[ \
inf A D conv A; sup A D A: (2.60)
A2A A2A
If A is empty, we put again sup A D Y and inf A D ;:

Example 2.6.29. Consider a topological space Y and let F .Y / be the family of all
closed subsets of Y . .F .Y /; / provides a complete lattice. By
[ \
inf A D cl A; sup A D A (2.61)
A2A A2A
the infimum and supremum of a nonempty subset A F .Y / are given.

If A is empty, we put again sup A D Y and inf A D ;:
Results concerning the infimum and supremum in the space of upper closed sets
are given in Proposition 2.6.40.
In the sequel, in this section we assume that .Y; / is a partially ordered linear
topological space, where the order is induced by a proper pointed convex cone C
satisfying ; ¤ int C ¤ Y . Here we do not assume that C is closed. However,
in Sect. 15.1 we will give a reformulation of a vector optimization problem as I -
valued problem, where the closedness of C is important (compare Proposition 2.17
in [395] and Sect. 2.7). We write y y 0 iff y 0 y 2 C and y < y 0 iff y 0 y 2 int C .
We denote by Y WD Y [ f1g [ fC1g the extended space, where the ordering
is extended by the convention
8y 2 Y W 1 y C1:
The linear operations on Y are extended by the following calculus rules in analogy
to that ones stated for the extended real space R:
0 .C1/ D 0; 0 .1/ D 0;
8˛ > 0 W ˛ .C1/ D C1;
8˛ > 0 W ˛ .1/ D 1;

8y 2 Y W y C .C1/ D C1 C y D C1;
8y 2 Y [ f1g W y C .1/ D 1 C y D 1:
The extended space Y is not a linear space.

In the following definition we introduce the upper closure of A Y (see [395,
397]) that is important for the formulation of the solution concept.
Definition 2.6.30. The upper closure (with respect to C ) of A Y is defined to

be the set
8
<Y if 1 2 A
Cl C A WD ; if A D fC1g
:
fy 2 Y j fyg C int C A n fC1g C int C g otherwise.
We have [395, Proposition 1.40]

8
<Y if 1 2 A
Cl C A WD ; if A D fC1g (2.62)
:
cl ..A n fC1g/ C C / otherwise.
As introduced in Definition 2.4.2, the set of weakly minimal points of a subset

A Y (with respect to C ) is defined by
WMin.A; C / D fy 2 A j A \ .fyg int C / D ;g ;
and the set of weakly maximal points of A is defined by
WMax.A; C / D fy 2 A j A \ .fyg C int C / D ;g :
If A ¤ ;, A Y we have [442, Theorem I-18]
WMin.Cl C A; C / D ; ” A C int C D Y ” Cl C A D Y:
In order to formulate set-valued optimization problems where the solution

concept is based on the lattice structure we introduce the notion of an infimal
set for a subset of Y (see Nieuwenhuis [442], Tanino [563, 566] and Löhne,
Tammer [397]).
Definition 2.6.31 (Infimal Set). The infimal set of A Y (with respect to C) is
defined by
8
< WMin.Cl C A; C / if ; ¤ Cl C A ¤ Y
Inf A WD f1g if Cl C A D Y
:
fC1g if Cl C A D ;:
Remark 2.6.32. We see that the infimal set of a nonempty set A Y (with respect
to C ) coincides with the set of weakly minimal elements of the set cl .A C C / with
respect to C (WMin.cl .A C C /; C /), if cl .A C C / ¤ Y . Note that if A Y then
WMin.A; C / D A \ Inf A.
By our conventions, Inf A is always a nonempty set. Clearly, if 1 belongs
to A, we have Inf A D f1g, in particular, Inf f1g D f1g. Furthermore, it
holds Inf ; D Inf fC1g D fC1g and Cl C A D Cl C .A [ fC1g/ and hence
Inf A D Inf.A [ fC1g/ for all A Y .
Considering the set of weakly maximal points WMax.A; C / of a set A Y with

respect to C (see Definition 2.4.2) we define analogously the lower closure Cl A
and the supremal set Sup A of a set A Y . It holds
Sup A D Inf.A/:
The following assertions were proved by Nieuwenhuis [442] and, in an extended

form, by Tanino [563].
Proposition 2.6.33. For A; B Y with ; ¤ Cl C A ¤ Y and ; ¤ Cl C B ¤ Y it
holds
(i) Inf A D fy 2 Y j y 62 A C int C; fyg C int C A C int C g,
(ii) A C int C D B C int C ” Inf A D Inf B,
(iii) A C int C D Inf A C int C ,
(iv) Cl C A D Inf A [ .Inf A C int C /,
(v) Inf A, .Inf A int C / and .Inf A C int C / are disjoint,
(vi) Inf A [ .Inf A int C / [ .Inf A C int C / D Y .
Proposition 2.6.34. For A Y it holds
(i) Inf Inf A D Inf A, Cl C Cl C A D Cl C A, Inf Cl C A D Inf A, Cl C Inf A D
Cl C A,
(ii) Inf.Inf A C Inf B/ D Inf.A C B/,
(iii) ˛ Inf A D Inf.˛A/ for ˛ > 0.
Proposition 2.6.35. Let Ai Y for i 2 I , where I is an arbitrary index set. Then
it holds
[ [
(i) Cl C Ai D Cl C Cl C Ai ,
[i 2I [ i 2I
(ii) Inf Ai D Inf Inf Ai .
i 2I i 2I
S Cl C fC1g D ; and Cl C A D Cl S
Proof. (i) As C .A n fC1g/ we can assume that
C1 62 i 2I Ai . We also assume f1g 62 i 2I Ai , because the statement is
otherwise obvious.
So we have
! !
[ [ [ [
ClC Ai D cl Ai C C D cl cl.Ai C C / CC D ClC ClC Ai :
i 2I i 2I i 2I i 2I
(ii) Follows from (i) and Proposition 2.6.33 (iv). t

u
In order to formulate set-valued optimization problems in an appropriate form for
deriving duality assertions (see Sects. 8.3 and 15.1) we introduce in the following
definitions the hyperspaces of upper closed sets and self-infimal sets.
Using the upper closure of A Y (Definition 2.6.30) we introduce the
hyperspace of upper closed sets (compare Löhne, Tammer [397]).
Definition 2.6.36 (Hyperspace of Upper Closed Sets).

The family F WD FC .Y / of all sets A Y with
Cl C A D A
is called the hyperspace of upper closed sets.

In F we introduce an addition ˚F W F F ! F , a multiplication by non-
negative real numbers ˇF W IRC F ! F and an order relation 4F as follows:
A ˚F B WD cl .A C B/;
˛ ˇF A WD Cl C .˛ A/
A 4F B W ” A B:
We use the rule 0 ; D f0g. This implies 0 ˇF ; D Cl C f0g D cl C .

Furthermore, using the infimal set of A Y (Definition 2.6.31) we introduce
the hyperspace of self-infimal sets (compare [397]).
Definition 2.6.37 (Hyperspace of Self-Infimal Sets). The family I WD IC .Y /
of all self-infimal subsets of Y , i.e., all sets A Y satisfying
Inf A D A
is called hyperspace of self-infimal sets.

In I we introduce an addition ˚I W I I ! I , a multiplication by non-
negative real numbers ˇI W IRC I ! I and an order relation 4I as follows:
A ˚I B WD Inf.A C B/;
˛ ˇI A WD Inf.˛ A/
A 4I B W ” Cl C A Cl C B:
Note that the definition of ˚I is based on the inf-addition in Y . As a consequence

we obtain f1g ˚I fC1g D fC1g. Of course, for all A 2 I we get 0 ˇI A D
Inff0g D bd C . In the space of self-supremal sets the sup-addition in Y is the
underlying operation (Fig. 2.6).
Lemma 2.6.38. For A; B 2 I with ; ¤ Cl C A ¤ Y and ; ¤ Cl C B ¤ Y we
have
A 4I B ” A \ .B C int C / D ;:
Proof. If Cl C A D ; or Cl C B D ; the proof is immediate, such that we assume

that Cl C A ¤ ; and Cl C B ¤ ;.
Fig. 2.6 The addition and the ordering in I for C D IR2C
Suppose that A 4I B. Taking into account the definition of 4I it holds

Cl C B Cl C A. By Proposition 2.6.33 (iv) we have Cl C B D B [ .B C int C /.
This yields .B [ .B C int C // \ .A int C / D ; because of Cl C B Cl C A.
Therefore A \ .B C int C / D ;.
Conversely, if A \ .B C int C / D ; then B C int C A C int C . Hence Cl C B
Cl C A and so A 4I B. t
u
Proposition 2.6.39. The spaces .F ; ˚F ; ˇF ; / and .I ; ˚I ; ˇI ; 4I / are iso-

morphic and isotone. The corresponding bijection is given by
j W F ! I; j. / D Inf. /; j 1 . / D Cl C . /:
Proof. By Proposition 2.6.34 (i), j is a bijection between F and I .

For A1 ; A2 2 F , we have j.A1 / ˚I j.A2 / D j.A1 ˚ A2 /. This follows from
Proposition 2.6.34 (ii).
Similarly, we can easily verify that for ˛
0 and A; B 2 F we have
˛ ˇI j.A/ D j.˛ ˇF A/ and A B ” j.A/ 4 j.B/: t

u
Proposition 2.6.40. .F ; / and .I ; 4/ are complete lattices. For nonempty
subsets A F and B I the infimum and supremum can be expressed by
[ \
inf A D cl A; sup A D A;
A2A A2A
[ \
inf B D Inf Cl C B; sup B D Inf Cl C B:
B2B B2B
Proof. For the space .F ; / the statements are obvious and for .I ; 4/ they follow
from Proposition 2.6.39. t
u
Fig. 2.7 The infimum and supremum in I for C D IR2C
As usual, if A F and B I are empty we define the infimum (supremum) to

be the largest (smallest) element in the corresponding complete lattice, i.e., inf A D
;, sup A D Y , inf B D fC1g and sup B D f1g.
It follows the main result of this section, which shows that the infimum as well
as the supremum in I can be expressed in terms that frequently are used in vector
optimization (compare [442], [145, 563, 566]), but up to now not in the framework
of complete lattices (see Fig. 2.7).
Theorem 2.6.41. For nonempty sets B I it holds
[ [
inf B D Inf B; sup B D Sup B:
B2B B2B
S S
Proof. (i) ItSholds inf B S
D Inf B2B Cl C B D Inf Cl C B2B Cl C B D
Inf Cl C B2B B D Inf B2B B.
(ii) We have to show that
[ \
Sup B D Inf Cl C B:
B2B B2B
Then the assertion follows with Proposition 2.6.40.

S S
a) If fC1g 2 B we have C1 2 B2B B and hence Sup T B2B B D fC1g.
On the other hand, since Cl C fC1g D ;, we have Inf B2B Cl C B D Inf ; D
fC1g.
b) Let fC1g 62 B but f1g 2 B. If f1g is the only element in B the assertion
is obvious, otherwise we can omit this element without changing the expressions.
c) Let fC1g 62 B and f1g 62 B. Then, B Y and ; ¤ Cl C B ¤ Y for all
B 2 B, i.e., we can use the statements of Proposition 2.6.33. Define the sets
!
[ [
V WD .B int C / D B int C
B2B B2B
and
\
W WD Cl C B:
B2B
We show that V \W D ; and V [W D Y . Assume there exists some y 2 V \W .

Hence there is some BN 2 B such that y 2 BN int C \Cl C BN D ;, a contradiction.
Let y 2 Y n W (we have W ¤ Y , because otherwise it holds Cl C B D Y for
all B 2 B and hence f1g 2 B). Then there exists some BN 2 B such that
y 62 Cl C B.N By Proposition 2.6.33 (iv), (vi) we obtain y 2 BN int C V .
S
If Cl V D Y we have W D ;, hence Sup B2B B D Sup V D fC1g D
Inf ; D Inf W . Otherwise, we have ; ¤ Cl V ¤ Y and ; ¤ Cl C W ¤ Y . By
Proposition 2.6.33, we obtain
[
Sup B D fy 2 Y j y 62 V; fyg int C V g
B2B
D fy 2 Y j y 2 W; .fyg int C / \ W D ;g
D WMin.W; C / D WMin.Cl C W; C / D Inf W
and so the proof is completed. t

u
We will see in Sect. 15.1 that the infimum/supremum in I is closely related
to the solution concepts of vector optimization because the infimal/supremal set is
closely related to the set of weakly minimal/maximal elements (see also Sect. 2.7).
We next show some calculus rules in the hyperspace I of self-infimal sets given
by Löhne, Tammer [398] and Löhne [395, Proposition 1.56].
Proposition 2.6.42. For subsets A ; B I we have
(i) inf A ˚I B D inf A ˚I inf B
(ii) sup A ˚I B 4 sup A ˚I sup B.
Proof. (i) If A D ;, it holds inf A ˚I B D inf A D fC1g and so
inf A ˚I B D inf A ˚I inf B D fC1g:
Otherwise, we have
[ [
inf A ˚I B D Inf A ˚I B D Inf ACB
A2A ;B2B A2A ;B2B
[ [ [ [
D Inf. AC B/ D Inf A ˚I Inf B
A2A B2B A2A B2B
D inf A ˚I inf B: (2.63)

(ii) For all A 2 A and B 2 B we have A ˚I B 4 sup A ˚I sup B. As I

is a complete lattice we can take the supremum on the left-hand side of the
inequality. This yields (ii).
t
u
In the same manner like F and I we define the space F ˘ of lower closed
subsets of Y and the space S of self-supremal subsets of Y , where we underlie
the sup-addition in Y in the latter case.
Using the infimal set given by Definition 2.6.31 we formulate set-valued
optimization problems based on lattice structure:
For an arbitrary set S , consider the following set-valued optimization problem
[
Inf F .x/ (SPY )
x2S
with a set-valued objective map F W S Y . We denote the domain of F W S

Y by Dom F WD fx 2 S j F .x/ ¤ ; and F .x/ ¤ fC1gg.
Next, we introduce a solution concept for set-valued optimization problems in
the hyperspace of self-infimal sets .I ; ˚I ; ˇI ;4I / (see Definition 2.6.37 and
Definition 2.6.31 for infimal sets).
S
Definition 2.6.43 (Infimal Set of .SPY /). The set PN WD Inf x2S F .x/ is the
solution set of the set-valued optimization problem .SPY /.
S S
Since Inf x2S F .x/ D Inf x2S Inf F .x/, .SPY / can be expressed as an I -
valued problem; without loss of generality we can assume that the sets F .x/ are
self-infimal, i.e., F W S ! I . Thus we consider the following problem.
[
PN D Inf F .x/ D inf F .x/: (SP I )
x2S
x2S
Furthermore, in Sect. 8.3.1 we derive duality assertions for F -valued optimiza-

tion problems. We consider the space .F ; / (see Definition 2.6.36), where the
order relation is given by
A 4F B W ” A B
for subsets A, B of F .
In order to formulate the F -valued optimization problem we study an objective
map F W X ! F , i.e., the objective function values of F are subsets of
the hyperspace of upper closed sets introduced in Definition 2.6.36. Using these
notations we study the F -valued problem
[
PN WD inf P .x/ D cl F .x/: (SP F )
x2X
x2X
Remark 2.6.44. In Sect. 8.3 we will show duality assertions for F -valued (I -
valued, respectively) primal problems and corresponding dual problems based on
conjugation as well as Lagrangian technique. Furthermore, in Sect. 15.1 we will use
the lattice approach for deriving duality assertions for vector optimization problems.
Remark 2.6.45. In the special case of single-valued functions F D f W S ! Y
the problems (SP I ) and (SP F ) coincide.
In Sect. 14.2 we will present an algorithm for solving set-valued optimization
problems where the objective map has a polyhedral convex graph. There we will
need the following notions and assertions concerning the infimum and supremum
in a subspace of the hyperspace of upper closed sets .F ; ˚F ; ˇF ; / (see
Definition 2.6.36), see Löhne [395].
Definition 2.6.46. The subspace of all closed convex subsets of an extended
partially ordered linear topological space Y is given by
Fconv WD fA F j 8
2 .0; 1/ W
ˇ A ˚ .1
/ ˇ A D Ag:
The space Fconv can be characterized using the convex hull (compare
Löhne [395, Proposition 1.59]).
Proposition 2.6.47. Assume that Y is a linear topological space ordered by a
proper pointed convex cone C Y with int C ¤ ; and let F D FC .Y /. Then
Fconv D fA Y j ClC conv A D Ag:
Proof. Taking into account
A D ClC conv A ” A D conv A ^ A D ClC A ”
8
2 Œ0; 1 W A D ClC .
A C .1
/A/ ”
8
2 Œ0; 1 W A D
ˇ A ˚ .1
/ ˇ A t
u
Proposition 2.6.48. .Fconv ; / is a complete lattice. For nonempty subsets A

Fconv the infimum and supremum can be expressed by
[ \
inf A D cl conv A; sup A D A:
A2A A2A
Proof. For all A 2 A , where A is a nonempty subset of Fconv we get with

Proposition 2.6.47 A D ClC conv A. This yields
[ [ [
cl conv A D cl conv ClC conv A D cl conv cl conv.A C C /
A2A A2A A2A
[ [
D cl conv .A C C / D cl conv A CC
A2A A2A
[
D ClC conv A:
A2A
S T
So we get cl conv
T A2A A 2 Fconv . Since A2A A is convex and upper closed, we
can conclude A2A A 2 Fconv . So we get the assertions of the proposition. t
u
2.6.4 The Embedding Approach by Kuroiwa
An important approach for deriving optimality conditions and algorithms for

solving set-valued optimization problems is based on the introduction of an
embedding space into which the set-valued optimization problem is embedded
(see Kuroiwa [353, 354]). With this approach Kuroiwa [354] defines notions of
directional derivatives for set-valued maps and derives corresponding necessary and
sufficient optimality conditions (compare Sect. 12.10). In this section we present
results given by Kuroiwa in [353, 354, 357].
Let Y be a n.v.s., let C be a proper closed convex pointed cone in Y with int C ¤
; and int C C ¤ ;.
The set relation cl C discussed in this section is defined as follows (cf. Defini-
tion 2.6.9): For A, B Y ,
C B W” cl.A C C / B:
A cl (2.64)
We consider minimal solutions with respect to the quasi-order cl

C given in (2.64)
in the sense of Definition 2.6.19, and the corresponding set-valued optimization
problem is given by
C minimize F .x/;
cl subject to x 2 S; (SPA cl
C)
where F W S Y is a set-valued objective mapping and S is a set.

We are looking for minimal solutions of (SPA clC ) in the sense of Defini-
tion 2.6.19, i.e., for elements x 2 S with
F .x/ cl
C F .x/ for some x 2 S H) F .x/ cl
C F .x/:
As already mentioned, a subset A of Y is said to be C -convex if ACC is convex.

Furthermore, A Y is said to be C C -bounded if hy , Ai is bounded from below
for any y 2 C C , where C C is the positive dual cone of C .
Let G be the family of all nonempty C -convex and C C -bounded subsets of
Y . In the following, we introduce a process of construction of a normed space V
into which G is embedded. This approach goes back to Kuroiwa and Nuriya [357].
At first, we introduce an equivalence relation on G 2 : For all

.A; B/; .D; E/ 2 G 2 ,
.A; B/ .D; E/ W” cl.A C E C C / D cl.B C D C C /:
The quotient space G 2 = is denoted by V , where
V WD fŒA; Bj.A; B/ 2 G 2 g;
where ŒA, B WD f.D; E/ 2 G 2 j.A; B/ .D; E/g .

Furthermore, we define addition and scalar multiplication on the quotient space
V as follows:
ŒA; B C ŒD; E WD ŒA C D; B C E ;

Œ
A;
B if

0

ŒA; B WD
Œ.
/B; .
/A if
< 0:
Then .V ; C; / is a vector space over R.

In order to introduce an order relation on V we define the following subset of
V by
.C / WD fŒA; B 2 V jB cl
C Ag: (2.65)
It is easy to see that .C / is a pointed convex cone in V .

Using the pointed convex cone .C /, we define an order relation .C / on V
as follows:
ŒA; B .C / ŒD; E W” ŒD; E ŒA; B 2 .C /:
Then, .V , C, , .C / / is an ordered vector space over R .

Let a function ' from G to V be given by
'.A/ WD ŒA; f0g for all A 2 G ;
then
C B , '.A/ .C / '.B/;
A cl
for any A, B 2 G .
By using this function ', the set optimization problem (SPA cl C ) can be
transformed into a vector optimization problem in the following sense: If F is
a map from S to G , then x 2 S is a minimal solution of (SPA cl
C ) if and only if
' ı F .S / \ .' ı F .x/ .C // D f' ı F .x/g: (2.66)
Formula (2.66) means that ' ı F .x/ is a Pareto minimal point of ' ı F .S / with
respect to .C /, i.e., ' ı F .x/ 2 Min.' ı F .S /; .C // (see Definition 2.4.1).
Finally, we introduce a norm j j in G 2 = . Consider c 2 int C and a weak*
compact base W WD fy 2 C C j hy ; ci D 1g of C C , then for each ŒA; B 2 V ,
j ŒA; B j WD sup j infhy ; Ai infhy ; Bij;

y 2W
is well-defined. Furthermore, let
V .W / WD fŒA; B 2 V j j ŒA; B j < C1g;
then .V .W /; j j/ is a normed vector space, and .C / is closed in .V .W /; j j/ .

In Sect. 12.10 we derive necessary and sufficient conditions for solutions of
(SPA cl
C ) using this embedding approach.
2.6.5 Solution Concepts with Respect to Abstract Preference

Relations
In this section we present a solution concept for set-valued optimization problems

with geometric constraints useful in welfare economics (introduced by Bao and
Mordukhovich [28]).
minimize F .x/ subject to x 2 S; (SPA )
where the cost mapping F W X Y is a set-valued mapping, X and Y are Banach

spaces and S is a subset of X .
The “minimization” in (SPA ) is understood with respect to a certain preference
relation on Y. This general (abstract) preference relation on Y is defined as follows
(see Sect. 2.1 or [431, Subsection 5.3.1]): For a given nonempty subset R Y Y ,
one says that y 1 is preferred to y 2 (we write y 1 y 2 ) if .y 1 ; y 2 / 2 R.
Following Bao and Mordukhovich [28] we will study a preference on Y directly
in terms of a given preference mapping L W Y Y instead of a preference on Y
described via a subset R Y Y . The level-set mapping L W Y Y associated
to the preference relation is defined by
L.y/ WD fu 2 Y j u yg: (2.67)

This means that u 2 Y is preferred to y if u 2 L.y/.

An abstract preference has to satisfy some requirements in order to be useful
in optimization and applications, especially in economics and engineering. In [431,
Definition 5.55] three properties imposed in order to postulate the notion of closed
preference relations as follows:
• The preference relation is nonreflexive, this means that .y; y/ … R for all
y 2 Y;
• Given some y 2 Y (a local minimizer in the sequel), the preference is locally
satiated around y in the sense that y 2 cl L.y/ for all y in some neighborhood
of y.
• The preference is almost transitive meaning that
Œu 2 L.y/; v 2 cl L.u/ H) v 2 L.y/: (2.68)
Especially, in the study of vector optimization problems the almost transitivity

property is widely used (under different names) ; see, e.g., [41, 431, 433, 624, 625]
and the references therein. However, the almost transitivity property turns out to be
rather restrictive, in contrast to the first two properties of formulated above. In
particular, for the Pareto minimality (compare Definition 2.4.1) defined by
y 1 y 2 W” y 2 y 1 2 C n f0g (2.69)
via an ordering cone C Y , the preference is almost transitive if and only if C

is convex and pointed; see [431, Proposition 5.56]. For the lexicographical order on
Rq and other natural preference relations important in vector optimization and its
applications including those to welfare economics (see Sect. 15.3) this property of
the ordering cone is not fulfilled.
Bao and Mordukhovich [28] developed an approach to preference relations in
set-valued optimization that is motivated by applications to models in welfare
economics and allows to avoid the almost transitivity condition (2.68), which does
not usually hold for many economies. They required the most natural local satiation
property of the preference mapping L formulated at the reference optimal solution.
The following solution concept is introduced by Bao and Mordukhovich [28,
Definition 3.1]:
Definition 2.6.49 (Fully Localized Minimizers for Constrained Set-Valued
Optimization Problems). Consider the problem (SPA ). Let .x; y/ 2 graph F
and x 2 S . Then
• .x; y/ is called a fully localized weak minimizer for (SPA ) if there exist
neighborhoods U of x and V of y such that there is no y 2 F .S \ U / \ V
preferred to y, i.e.,
F .S \ U / \ L.y/ \ V D ;: (2.70)
• .x; y/ is called a fully localized minimizer for (SPA ) if there exist neighbor-
hoods U of x and V of y such that there is no y 2 F .S \ U / \ V with y ¤ y
and y 2 cl L.y/, i.e.,
F .S \ U / \ cl L.y/ \ V D fyg: (2.71)
• .x; y/ is called a fully localized strong minimizer for (SPA ) if there exist
neighborhoods U of x and V of y such that there is no .x; y/ 2 graph F \
.U V / with .x; y/ ¤ .x; y/ satisfying x 2 S and y 2 cl L.y/, i.e.,
graph F \ .S cl L.y// \ .U V / D f.x; y/g: (2.72)
In Sect. 12.11, Theorem 12.11.3 we derive first order necessary conditions

for fully localized minimizers. Furthermore, we study applications in welfare
economics in Sect. 15.3.
Remark 2.6.50. It is easy to see that each fully localized strong minimizer for (SPA )
is also a fully localized minimizer for (SPA ). Furthermore, each fully localized
minimizer for (SPA ) is also a fully localized weak minimizer for (SPA ). In the
case S D X in Definition 2.6.49, we speak about the corresponding fully localized
minimizers for the mapping F.
Remark 2.6.51. For all of the concepts in Definition 2.6.49 (see Bao and Mor-
dukhovich [28, Definition 3.1]) the underlying feature is that one introduces the
image localization of minimizers in constructions (2.70), (2.71), (2.72). These
concepts introduced in [28, Definition 3.1] are different from the concepts discussed
before even for minimal points and weakly minimal points of single-valued
objectives F WD f W X ! Y and allow to study local Pareto-type optimal
allocations of welfare economics (see Sect. 15.3). The concept of (global or
local) strong minimizers was first time introduced in [28, Definition 3.1] for
set-valued optimization problems; it is related to Khan’s notion of strong Pareto
optimal allocations for models of welfare economics (compare Khan [324]) and the
corresponding relationships established in Sect. 15.3.
Example 2.6.52. In this example we will see that a fully localized strong minimizer
may not provide a partially localized (i.e., with V D Y ) minimum or weak
minimum in (2.70) and (2.71). To see this, we consider a set-valued mapping
F W R R defined by
8
< fxg if x < 0
F .x/ WD f0; 1g if x D 0
:
fx C 1g if x > 0;
with respect to the usual order on R generating the level sets L.y/ D .1; y/,
it is easy to see that condition (2.72) is fulfilled at .0; 1/ with U D . 12 ; 12 / and
V D . 12 ; 32 / but conditions (2.70) and (2.71) do not hold with V D R.
Example 2.6.53. Now, we will see that the localized minimizers and weak
minimizers in Definition 2.6.49 are identical in the case of scalar set-valued
optimization with Y D R and L.y/ D .1; y/, but they may be quite different in
the vector-valued case. For example, if we consider an objective map F W R R2
with F .x/ R2 n int R2 and the usual weak preference on R2 with the level sets
L.y/ D y int R2C (see Definition 2.6.2), we have that .0; 0/ 2 RR2 is a localized
weak minimizer for F, but it is not a localized minimizer for this mapping. Note
finally that localized strong minimizers reduce to standard isolate minimizers for
scalar single-valued optimization problems.
2.6.6 Set-Valued Optimization Problems with Variable

Ordering Structure
In the book by Chen, Huang and Yang [91] set-valued optimization problems with
variable ordering structure are introduced, where the solution concept is related
to the solution concept for vector optimization problems with variable ordering
structure given in Definition 2.5.1.
Let X and Y be Banach spaces , S X be nonempty. Furthermore, let C W X
Y be a cone-valued mapping. We assume that for every x 2 X , the set C.x/ is a
proper closed convex cone with nonempty interior int C.x/.
We consider a set-valued objective mapping F W X Y and a set-valued
optimization problem with variable ordering structure
v-minimize F .x/ subject to x 2 S; (SPv )
where “v-minimize” stands for problems with variable ordering structure with
respect to a cone-valued mapping C W X Y in the following sense:
Definition 2.6.54 ((Weak) v-Minimizer of .SPv /). Let x 2 S and y 2 F .x/.
(a) The pair .x; y/ is called a v-minimizer of .SPv / if
.F .S / y/ \ .C.x// D f0g:
(b) The pair .x; y/ is called a weak v-minimizer of .SPv / if
.F .S / y/ \ . int C.x// D ;:
In the following we will show that a set-valued optimization problem .SPv / can
be transformed into an equivalent vector-optimization problem in the sense that their
v-minimal solution pairs are identical.
Definition 2.6.55. A cone-valued map C W X Y is pointed on S X if the
cone [x2S C.x/ is pointed, i.e.,
.[x2S C.x// \ . [x2S C.x// D f0g:
Remark 2.6.56. A cone-valued map C W X Y is pointed on S if and only if
8x1 ; x2 2 S W C.x1 / \ .C.x2 // D f0g:
For deriving the relationships between v-minimizers of a set-valued problem with

variable ordering structure .SPv / in the sense of Definition 2.6.54 and v-minimal
points of a corresponding vector optimization problem with an objective function
f W X ! Y (see Definition 2.5.1) we need a certain monotonicity property
concerning the set-valued map C with respect to f .
Definition 2.6.57. Let f W X ! Y be a vector-valued function and C W X Y be
a cone-valued mapping. The cone-valued mapping C is called weakly f-monotone,
if for all x1 ; x2 2 X , c1 2 C.x1 /,
f .x1 / f .x2 / 2 c1 C C.x2 / H) C.x2 / C.x1 /:
The following relationships between v-minimal points of a vector optimization

problem (VPv ) and v-minimizers of a set-valued optimization problem (SPv ) are
shown by Chen, Huang and Yang [91, Proposition 2.63].
Proposition 2.6.58. Let S X and C W X Y be a pointed cone-valued
mapping on S . Furthermore, let f W X ! Y be a vector-valued function and
F W X Y be given by
F .x/ D f .x/ C C.x/ .x 2 X /: (2.73)
(a) Suppose that C is weakly f -monotone. If x 2 S is a v-minimal point of the

vector optimization problem (VPv ):
Min.f .S /; C.//;
then .x; f .x// is a v-minimizer of the set-valued optimization problem (SPv ):
v-minimize F .x/ subject to x 2 S:
(b) If .x; y/ is a v-minimizer of the set-valued optimization problem .SPv /, then x

is a v-minimal point of the vector optimization problem .VPv / and y D f .x/.
Proof. First, we show that .a/ holds. Consider a v-minimal point x 2 S of the
problem .VPv /. Then
8x 2 S W f .x/ f .x/ … C.x/ n f0g:

This yields
8x 2 S W .f .x/ f .x/ C C.x// \ .C.x/ n f0g/ D ;: (2.74)
Indeed, if there exists xQ 2 S with
Q f .x/ C C.x//
.f .x/ Q \ .C.x/ n f0g/ ¤ ;;
then there exists c 2 C.x/ and c ¤ 0 such that
c 2 f .x/
Q f .x/ C C.x/:
Q
This means
f .x/ f .x/
Q 2 c C C.x/:
Q
Taking into account the weak f-monotonicity of C , we get
Q f .x/ 2 c C.x/:
f .x/
Furthermore, since C is pointed and c ¤ 0, it follows that
Q f .x/ 2 C.x/ n f0g;

f .x/
in contradiction to the assumption that x is a v-minimal point of the problem .VPv /.

Taking into account (2.73) and (2.74) we get
8y 2 F .x/; x 2 S W y f .x/ … C.x/ n f0g:
Hence, .x; f .x// is a v-minimizer of the set-valued optimization problem .SPv /.

Now, we will prove .b/. Let us assume that .x; y/ is a v-minimizer of the set-
valued optimization problem .SPv /. Then,
y 2 F .x/ D f .x/ C C.x/
and
8y 2 F .S / W y y … C.x/ n f0g: (2.75)
Of course, it holds y D f .x/. We have to show that x is a v-minimal point of

the vector optimization problem .VPv /. Contrarily, suppose that x is not a v-minimal
point of the vector optimization problem .VPv /. Then, for some element xQ 2 S nfxg,
Q f .x/ 2 C.x/ n f0g:

f .x/
So we get
Q y 2 C.x/ n f0g;
f .x/
because of f .x/ D y, in contradiction to (2.75). Hence, x is a v-minimal point of

the vector optimization problem .VPv /. t
u
Remark 2.6.59. In the book by Chen, Huang, Yang [91, Theorem 2.64] necessary
conditions for weak v-minimizers of (SPv ) are shown using the contingent
derivative.
Remark 2.6.60. Necessary and sufficient optimality conditions in form of the
Fermat rule for nondominated solutions of unconstrained set-valued optimization
problems with variable ordering structure and the Lagrange multiplier rule for
the constrained set-valued problems with variable ordering structure are given by
Eichfelder and Ha in [168].
2.6.7 Approximate Solutions of Set-Valued Optimization

Problems
In this section we introduce a concept of approximate solutions in set-valued

optimization. Approximate solutions are of interest from the theoretical as well as
computational point of view. Especially, in order to formulate set-valued versions of
Ekeland’s variational principle (compare Chap. 10) and a subdifferential variational
principle for set-valued mappings (see Sect. 12.9) one is dealing with approximate
solutions.
We consider a set-valued optimization problem:
minimize F .x/ subject to x 2 X; (SP)
where X is a linear space, Y is a linear topological space, C Y is a proper closed

convex cone and the cost mapping F W X Y is a set-valued mapping.
The following concepts for approximate solutions of the set-valued problem
.SP/ was given by Bao and Mordukhovich [27, Definition 3.4] and is related to
minimizers introduced in Definition 2.6.1.
Definition 2.6.61 (Approximate Minimizers of Set-Valued Optimization Prob-
lems). Let x 2 X and .x; y/ 2 graph F . Then:
(i) Consider " > 0 and k 0 2 C n f0g. The pair .x; y/ 2 graph F is called an
approximate "k 0 -minimizer for F if
y C "k 0 … y C for all y 2 F .x/ with x ¤ x:

(ii) Consider " > 0 and k 0 2 C n f0g. The pair .x; y/ 2 graph F is called a strict
approximate "k 0 -minimizer for F if there is a number 0 < "Q < " such that
.x; y/ is an approximate "Qk 0 -minimizer of this mapping.
In Sect. 12.9, Theorem 12.9.1, we will show necessary conditions for strict
approximate "k 0 -minimizers of F .
2.7 Relationships Between Solution Concepts
In this section we study the relationships between different solution concepts in

set-valued optimization. Furthermore, we discuss the special case that the objective
map is single-valued.
closed cone C , X a linear space, S a subset of X and F W X Y . We consider the
set-valued optimization problem (SP):
minimize F .x/ subject to x 2 S: (2.76)
In the formulation of the solution concepts based on set approach the underlying
space is a linear topological space Y whereas the extended space Y WD Y [f1g[
fC1g is considered in the formulation of the solution concepts based on lattice
approach in order to work with infimum and supremum.
Remark 2.7.1. The differences between the solution concepts based on
set-approach in Definition 2.6.19 and the solution concepts based on vector
approach in Definition 2.6.1 are already discussed in Examples 2.6.21, 2.6.22
and 2.6.23.
Remark 2.7.2. In the special case of single-valued functions F D f W X ! Y
the concept of minimizers of the set-valued problem (SP) (see Definition 2.6.1)
coincides with the solution concept for Pareto minimal points of f .S / with respect
to C introduced in Definition 2.4.1: .x; f .x// 2 graph f is a minimizer in the sense
of Definition 2.6.1 if and only if f .x/ is a Pareto minimal point of f .S / with respect
to C , i.e., f .x/ 2 Min.f .S /; C /.
Remark 2.7.3. In the special case of single-valued functions F D f W X ! Y
the concept of minimal solutions of the problem .SP / (see Definition 2.6.19)
w.r.t. the order relations introduced in Definition 2.6.9 coincides with the solution
concept for Pareto minimal points given in Definition 2.4.1.
In the following we consider a linear topological S space Y , a linear space X ,
S X , a set-valued map F W X Y , F .S / D x2S F .x/ ¤ ; and a proper
pointed closed convex cone C Y with int C ¤ ;. The relationship between the
infimal set of F .S / (Definition 2.6.31) and weak minimizers of F .S / in the sense
of Definition 2.6.2 is given in the next proposition.
2.7 Relationships Between Solution Concepts 75
Proposition 2.7.4. Under the assumption that F .S / D cl.F .S / C C / we get
WMin.F .S /; C / D Inf F .S /:
Proof. Taking into account the assumption F .S / D cl.F .S /CC / and the definition
of the infimal set we get
WMin.F .S /; C / D WMin.cl.F .S /CC /; C / D WMin.ClC F .S /; C / D Inf F .S /:
The proof is completed. t

u
The assertion of Proposition 2.7.4 says that the solution concept for I -valued
problems coincides with the set WMin.F .S /; C / in Definition 2.6.3 for weak
minimizers.
Furthermore, if we assume that F .S / C C is closed we get the following
assertion.
Proposition 2.7.5. Under the assumption that F .S / C C is closed we get
Inf F .S / D WMin.F .S / C C; C /:
Proof. Because of the closedness of F .S / C C we get
WMin.F .S / C C; C / D WMin.cl.F .S / C C /; C / D Inf F .S /;
taking into account Definition 2.6.31. t

u
Corollary 2.7.6. Assuming that F .S / C C is closed and WMin.F .S / C C; C / D
WMin.F .S /; C / we get
Inf F .S / D WMin.F .S /; C /:
In Sect. 15.1 we will use methods of set-valued optimization for deriving

duality assertions for vector optimization problems. The relationships between
vector optimization problems and I -valued problems are discussed by Löhne and
Tammer [397] and in a comprehensive and detailed way by Löhne [395].
closed cone C , X a linear space, S a subset of X and f W X ! Y a vector-valued
function. We consider the vector optimization problem
Min.f .S /; C /: (VOP)
In Sect. 15.1 we will see that it is very useful to assign to (VOP) a corresponding
I -valued problem such that one can use the complete lattice structure of
.I ; WDI /, where .I ; / is defined with respect to the ordering cone C of
the vector optimization problem.
For a given vector-valued function f W X ! Y we put
f W X ! I ; f .x/ WD Infff .x/g
and assign to (VOP) the I -valued problem
minimize f subject to x 2 S: (VOPI )
Problem (VOPI ) is said to be the I -extension of the vector optimization problem

(VOP) (see Löhne [395]). The lattice extension of the vector optimization problem
(VOP) allows us to handle the problem in the framework of complete lattices. For
this extension it is important that the ordering cone C is closed as we will see in the
proof of the following proposition.
The following assertion is shown by Löhne [395, Proposition 2.17].
Proposition 2.7.7. For all x; u 2 X it holds
f .x/ C f .u/ ” f .x/ f .u/:
Proof. Consider y D f .x/ and z D f .u/. Let Inffyg Inffzg, then ClC fyg
ClC fzg. With (2.62) we can conclude z 2 cl.fzg C C / cl.fyg C C /. Because of
the assumption that C is closed, we get z 2 fyg C C . This means y C z. The
opposite inclusion is obvious. t
u
As a direct consequence of Proposition 2.7.7 we get corresponding assertions
concerning the solutions of (VOP) and (VOPI ) (see Löhne [395, Proposition 2.18]).
Finally, it is important to mention the following references. In the paper by
Hernández, Jiménez, Novo [244], Benson proper efficiency in set-valued opti-
mization is discussed. Hernández, Jiménez, Novo study in [245] weak and proper
efficiency in set-valued optimization. Flores-Bazán, Hernández characterize effi-
ciency without linear structure in [189]. Moreover, Hernández, Rodríguez-Marín,
Sama describe solutions of set-valued optimization problems [253]. Furthermore,
in Hernández, Rodríguez-Marín [250] certain existence results for solutions of set
optimization problems are derived.
Chapter 3
Continuity and Differentiability
In this chapter we present continuity notions for set-valued mappings and

corresponding properties under convexity assumptions. Furthermore, we introduce
Lipschitz properties for single-valued and set-valued maps. Concepts of generalized
differentiability and corresponding calculus rules are recalled.
3.1 Continuity Notions for Set-Valued Maps
In this section X and Y are separated (in the sense of Hausdorff) topological spaces
and F W X Y is a set-valued map. When mentioned explicitly, Y is a separated
topological vector space (s.t.v.s.). The presentation of this section follows that in
[214, Section 1.2.5] where one can find the proofs for the results which are only
stated.
Definition 3.1.1. Let x0 2 X . We say that
(a) F is upper continuous (u.c.) at x0 if
8 D Y; D open, F .x0 / D; 9 U 2 NX .x0 /; 8 x 2 U W F .x/ D;

(3.1)
i.e., F C1 .D/ is a neighborhood of x0 for each open set D Y such that

F .x0 / D;
(b) F is lower continuous (l.c.) at x0 if
8 D Y; D open, F .x0 / \ D ¤ ;; 9 U 2 NX .x0 /; 8 x 2 U W F .x/ \ D ¤ ;;

(3.2)
i.e., F 1 .D/ is a neighborhood of x0 for each open set D Y such that F .x0 /\
D ¤ ;.

DOI 10.1007/978-3-642-54265-7__3
78 3 Continuity and Differentiability
(c) F is continuous at x0 if F is u.c. and l.c. at x0 .

(d) F is upper continuous (lower continuous, continuous) if F is so at every
x 2 XI
(e) F is lower continuous at .x0 ; y0 / 2 X Y if
8 V 2 NY .y0 /; 9 U 2 NX .x0 /; 8 x 2 U W F .x/ \ V ¤ ;:
It follows from the definition that x0 2 int.dom F / and y0 2 cl .F .x0 // if F is

lower continuous at .x0 ; y0 / and F is lower continuous at x0 2 dom F if and only
if F is lower continuous at every .x0 ; y/ with y 2 F .x0 /; moreover, F is lower
continuous at every x0 2 X n dom F . If x0 2 X n dom F , then F is u.c. at x0 if and
only if x0 2 int.X n dom F /. So, if F is u.c. then dom F is closed, while if F is l.c.
then dom F is open.
The next result is an immediate consequence of the definitions.
Proposition 3.1.2. (i) F is upper continuous if and only if F C1 .D/ is open for
every open set D Y I
(ii) F is lower continuous if and only if F 1 .D/ is open for every open set D Y .
The limit inferior of F at x0 2 X is defined by
lim inf F .x/ WD fy 2 Y j 8 V 2 NY .y/; 9 U 2 NX .x0 /; 8 x 2 U W F .x/ \ V ¤ ;g;

x!x0
while the limit superior of F at x0 2 X is defined by
lim sup F .x/ WDfy 2 Y j 8 V 2 NY .y/; 8 U 2 NX .x0 /; 9 x 2 U W F .x/ \ V ¤ ;g

x!x0
\
D cl F .U / ;
U 2NX .x0 /
where U WD U n fx0 g for U 2 NX .x0 /.

Note that x0 is an accumulation point of dom F if lim supx!x0 F .x/ ¤ ;, while
x0 2 int .dom F [ fx0 g/ if lim infx!x0 F .x/ ¤ ;; if x0 is an isolated point of
X (i.e., fx0 g 2 NX .x0 /), then lim infx!x0 F .x/ D Y . Of course, if x0 is an
accumulation point of dom F , then
lim inf F .x/ lim sup F .x/;

x!x0 x!x0
both of them being closed sets.

Sometimes in the definitions of lim infx!x0 F .x/ and lim supx!x0 F .x/ one
takes x 2 dom F . Note that this situation reduces to the preceding one by
considering the restriction of F at fx0 g [ dom F .
Recall first that if .X; d / is a metric space and A; B X , then the excess of A
over B is
3.1 Continuity Notions for Set-Valued Maps 79
e.A; B/ D sup dist.x; B/ if A; B ¤ ;; e.;; B/ D 0; e.A; ;/ D 1 if A ¤ ;;

x2A
where dist.x; A/ WD infa2A d.x; a/; in particular, dist.x; ;/ D 1. It is simple to

show that for a nonempty and compact set A X and an open set D X , if
A D, then there exists " > 0 such that A" WD fx 2 X j dist.x; A/ < "g D fx j
A \ B.x; "/ ¤ ;g D, where B.x; "/ WD fx 0 2 X j d.x; x 0 / < "g.
In particular cases for X or/and Y one has useful characterizations for the
elements of lim infx!x0 F .x/ and lim supx!x0 F .x/.
Proposition 3.1.3. Let x0 2 X and y 2 Y .
(i) y 2 lim infx!x0 F .x/ if and only if
8 X n fx0 g .xi /i 2I ! x0 ; 9 .x'.j / /j 2J ; 9 Y .yj /j 2J ! y;

8 j 2 J W yj 2 F .x'.j / /;
and y 2 lim supx!x0 F .x/ if and only if
9 X n fx0 g .xi /i 2I ! x0 ; 9 Y .yi /i 2I ! y; 8 i 2 I W yi 2 F .xi /:
(ii) If X and Y are first countable, then y 2 lim infx!x0 F .x/ if and only if
8X n fx0 g .xn / ! x0 ; 9Y .yn / ! y; 9 n0 2 N; 8 n

n0 W yn 2 F .xn /;
and y 2 lim supx!x0 F .x/ if and only if
9 X n fx0 g .xn / ! x0 ; 9 Y .yn / ! y; 8 n 2 N W yn 2 F .xn /:
(iii) If .Y; / is a metric space, then
y 2 lim inf F .x/ , lim dist.y; F .x// D 0;

x!x0 x!x0
y 2 lim sup F .x/ , lim inf dist.y; F .x// D 0:

x!x0 x!x0
In (i) above .xi /i 2I , .yj /j 2J ; .yi /i 2I denote nets, while .x'.j / /j 2J denotes a
subnet of .xi /i 2I , that is ' W J ! I is such that for every i 2 I there exists ji 2 J
with '.j /
i for all j
ji :
We have the following characterization of upper continuity in a special case.
Proposition 3.1.4. Suppose that .Y; / is a metric space and F .x0 / is compact.
Then F is u.c. at x0 if and only if limx!x0 e.F .x/; F .x0 // D 0.
In the next results we provide other characterizations for upper and lower
continuity at a point.
Proposition 3.1.5. Let X , Y be first countable (in particular, let X , Y be metric

spaces) and x 2 dom F . The following assertions are equivalent:
(i) the set-valued map F is u.c. at xI
(ii) for every closed set E Y with F .x/ \ E D ;, there exists U 2 NX .x/ such
that F .U / \ E D ;I
(iii) for every closed set E Y and every sequence X .xn / ! x with F .xn / \
E ¤ ; for n 2 N, we have F .x/ \ E ¤ ;I
(iv) for every open set D Y with F .x/ D, and every sequence X .xn / ! x
there exists nD 2 N such that F .xn / D for n
nD I
(v) for all sequences .xn / X , .yn / Y with .xn / ! x and yn 2 F .xn / n F .x/
for n 2 N, there exists a subsequence .ynk / ! y 2 F .x/.
If X , Y are not first countable, the conditions (i)–(iv) remain equivalent by
replacing sequences by nets.
Proposition 3.1.6. Let .x0 ; y0 / 2 X Y . The following statements are equivalent:
(i) the set-valued map F is lower continuous at .x0 ; y0 /I
(ii) for every net .xi /i 2I ! x0 there exist a subnet .x'.j / /j 2J of .xi / and a net
.yj /j 2J ! y0 such that yj 2 F .x'.j / / for j 2 J .
(iii) y0 2 lim infx!x0 F .x/ \ cl F .x0 /.
Suppose now that X , Y are first countable. Then (i) is equivalent to
(iv) for every sequence X .xn / ! x0 there exist a sequence Y .yn / ! y0
and n0 2 N such that yn 2 F .xn / for every n
n0 .
From the preceding result we obtain that
F is lower continuous at x0 , F .x0 / lim inf F .x/:

x!x0
Definition 3.1.7. We say that

(i) F is closed if graph F is a closed subset of X Y I
(ii) F is closed at x 2 X if for every net ..xi ; yi //i 2I graph F converging to
.x; y/ we have that y 2 F .x/;
(iii) F is closed-valued if F .x/ is closed for every x 2 X ;
(iv) F is compact at x 2 X if for every net ..xi ; yi //i 2I graph F with .xi / ! x,
there exists a subnet .y'.j / /j 2J converging to some y 2 F .x/.
Of course, F is closed if and only if F is closed at every x 2 X ; moreover, if
F is closed, then F is closed-valued. If F is compact at x, then F is closed at x;
moreover, if F is closed (compact) at x, then F .x/ is closed (compact). Note also
that F is compact at x 2 X n dom F if and only if x 2 int.X n dom F /, but F may
be closed at some x 2 cl.dom F / n dom F .
Proposition 3.1.8. Assume that ; ¤ A dom F and F is u.c. at every x 2 A.
(i) If A is compact and F .x/ is compact for every x 2 A, then F .A/ is compact.
(ii) If A is connected and F .x/ is connected for every x 2 A, then F .A/ is

connected.
Recall that the nonempty subset A of the topological space .X; / is connected
if A \ D1 \ D2 ¤ ; whenever D1 ; D2 are open subsets of X such that A \ D1 ¤ ;,
A \ D2 ¤ ; and A D1 [ D2 . Recall that the connected subsets of R are the
intervals, while in a topological vector space every convex set is connected.
When Y is a metric space or a topological vector space, Proposition 3.1.8(i)
follows also from the next result.
Proposition 3.1.9. Let x 2 X . The following assertions hold:
(i) F is compact at x if and only if F .x/ is compact and F is u.c. at x.
(ii) If .Y; / is a metric space, then F is compact at x if and only if
limx 0 !x e.F .x 0 /; F .x// D 0 and F .x/ is compact.
(iii) Let Y be first countable. If F is compact at x, then for every sequence
..xn ; yn //n2N graph F with .xn / ! x, there exists a subsequence .ynk / !
y 2 F .x/. If X is first countable and Y is a metric space, the converse is true.
Related to closedness of set-valued maps we have the following result.
Proposition 3.1.10. Let x 2 X . The following assertions hold:
(i) F is closed at x if and only if for every y 2 Y n F .x/ there exist U 2 NX .x/
and V 2 NY .y/ such that F .U / \ V D ;.
(ii) F is closed at x if and only if lim supx 0 !x F .x 0 / F .x/ D cl F .x/.
(iii) Suppose that X , Y are first countable. Then F is closed at x if and only if for
every sequence graph F ..xn ; yn //n2N ! .x; y/ one has y 2 F .x/.
(iv) If F is u.c. at x and F .x/ is compact, then F is closed at x.
(v) If Y is regular (in particular, if Y is a metric space or a topological vector
space), F is u.c. at x, and F .x/ is closed, then F is closed at x.
The next result gives sufficient conditions for the upper continuity of the
intersection of two set-valued maps.
Proposition 3.1.11. Consider F1 ; F2 W X Y and F .x/ D F1 .x/ \ F2 .x/ for
every x 2 X . Then F is u.c. at x0 2 X if one of the following conditions holds:
(i) F1 is closed at x0 , F2 is u.c. at x0 , and F2 .x0 / is compact;
(ii) Y is normal (in particular, Y is a metric space), F1 ; F2 are u.c. at x0 , and
F1 .x0 /, F2 .x0 / are closed.
Recall that NY denotes the class of balanced neighborhoods of the origin of the
t.v.s. Y .
Definition 3.1.12. Let Y be a topological vector space and x0 2 X . We say that
(a) F is Hausdorff upper continuous (H-u.c.) at x0 if
8 V 2 NY ; 9 U 2 NX .x0 /; 8 x 2 U W F .x/ F .x0 / C V: (3.3)

(b) F is Hausdorff lower continuous (H-l.c.) at x0 if
8 V 2 NY ; 9 U 2 NX .x0 /; 8 x 2 U W F .x0 / F .x/ C V: (3.4)
(c) F is Hausdorff continuous at x0 if F is H-u.c. and H-l.c. at x0 .

(d) F is Hausdorff upper continuous (Hausdorff lower continuous, Hausdorff
continuous) if F is so at every x 2 X .
The above definition can be given when Y is a metric space, too; just replace
V 2 NY by " > 0 and F .x0 / C V by .F .x0 //" . Of course, if Y is a metric space,
then F is H-u.c. at x0 if and only if limx!x0 e .F .x/; F .x0 // D 0, and F is H-lower
continuous at x0 if and only if limx!x0 e .F .x0 /; F .x// D 0.
Concerning the continuity of the sum of set-valued maps and of the multiplication
by scalars, we have the following result.
Proposition 3.1.13. Let Y be a topological vector space, F; F1 ; F2 W X Y ,
x0 2 X , and ˛ 2 R.
(i) If F is u.c. (resp. lower continuous, H-u.c., H-l.c) at x0 , then ˛F is u.c. (resp.
l.c., H-u.c., H-l.c.) at x0 .
(ii) If F1 and F2 are l.c. (resp. H-u.c., H-l.c.) at x0 , then F1 CF2 is l.c. (resp. H-u.c.,
H-l.c.) at x0 .
Note that we have not a similar result to Proposition 3.1.13(ii) for upper
continuity. Also note that if F is u.c. at x0 , then F is H-u.c. at x0 ; the converse
implication is true when F .x0 / is compact. On the other hand, if F is H-l.c. at x0 ,
then F is l.c. at x0 , the converse being true if F .x0 / is compact.
We can characterize Hausdorff upper and lower continuities by using nets, and
even sequences when X and Y are first countable.
Proposition 3.1.14. Suppose that Y is a topological vector space and x 2 X .
Then:
(i) F is H-lower continuous at x if and only if for all nets .xi /i 2I X with
.xi / ! x and .y i /i 2I F .x/ there exist a subnet .x'.j / / and a net .yj /j 2J
such that yj y '.j / ! 0 and yj 2 F .x'.j / / for all j 2 J I
(ii) F is H-u.c. at x if and only if for every net ..xi ; yi //i 2I graph F with .xi / !
x there exists a subnet .y'.j / /j 2J and a net .y j /j 2J F .x/ such that y'.j /
y j ! 0.
Suppose now that X and Y are first countable.
(iii) F is H-lower continuous at x if and only if for every sequence .xn /n2N X
with .xn / ! x and every sequence .y n / F .x/ there exists a sequence .yn /
such that yn y n ! 0 and yn 2 F .xn / for all n
n0 I
(iv) F is H-u.c. at x if and only if for every sequence ..xn ; yn //n2N graph F with
.xn / ! x there exists a sequence .y n / F .x/ such that yn y n ! 0.
Proposition 3.1.15. Suppose that F is upper continuous and closed-valued. Then

graph F is closed. The same conclusion holds if Y is a topological vector space and
F is H-u.c. instead of being u.c.
Note that requiring the upper continuity of F only on dom F is not sufficient for
the conclusion of the preceding result.
Let .An /n2N P.Y /. Recall that
lim inf An WD fy 2 Y j 9 .yn / ! y such that yn 2 An for n

n0 g;
n!1
lim sup An WD fy 2 Y j 9 .ynk / ! y such that ynk 2 Ank for k 2 NgI

n!1
of course, lim infn!1 An lim supn!1 An . We say that .An / converges in

the sense of Kuratowski–Painlevé to A Y if lim supn!1 An A
lim infn!1 An .
Having A; An Y , n 2 N, and taking X WD N [ f1g R (X endowed
with the topology induced by that of R) we may consider the set-valued map
F W X Y defined by F .n/ D An for n 2 N and F .1/ D A. When Y is
first countable, lim infn!1 An is exactly lim infn!1 F .n/, and lim supn!1 An is
lim supn!1 F .n/.
Definition 3.1.16. Suppose that Y is a separated topological vector space and C
Y is a convex cone. We say that F is C -u.c.,C -l.c.,H-C -u.c., or H-C -l.c. at x0 2 X
if relation (3.1), (3.2), (3.3) or (3.4) holds with F .x/ D C C , F .x/ \ .D C / ¤
;, F .x/ F .x0 / C V C C , F .x0 / F .x/ C V C C instead of F .x/ D,
F .x/ \ D ¤ ;, F .x/ F .x0 / C V , F .x0 / F .x/ C V , respectively; similarly
for F C -l.c. at .x0 ; y0 /. F is C -continuous (C -Hausdorff continuous) at x0 if F
is C -u.c. and C -l.c. (H-C -u.c. and H-C -l.c.) at x0 .
Remark 3.1.17. As in Proposition 3.1.13 one can prove that having the set-valued
maps F; F1 ; F2 W X Y that are C -l.c. (resp. C -H-u.c., C -H-l.c.) at x0 2 X , Y
being a t.v.s., C Y a convex cone, and ˛ 2 RC , then F1 C F2 and ˛F are C -l.c.
(resp. C -H-u.c., C -H-l.c.) at x0 .
When P C is another convex cone, if F is P -l.c. (P -u.c., H-P -l.c., or H-P -
u.c.) at x0 , then F is C -l.c. (C -u.c., H-C -l.c., or H-C -u.c.) at x0 . Note also that F is
C -l.c., H-C -u.c., or H-C -l.c. if and only if FC is l.c., H-u.c., or H-l.c., respectively,
but such an equivalence is not true for upper continuity. In fact, we have that F is
C -u.c. at x0 if FC is u.c. at x0 , but the converse is not true even if F D FC .
Definition 3.1.18. When Y is a topological vector space, we say that F is
uniformly C -l.c. at x0 on A if A F .x0 / and
8 V 2 NY ; 9 U 2 NX .x0 /; 8 x 2 U W A F .x/ C V C C:
Hence F is uniformly C -l.c. at x0 on F .x0 / exactly when F is C -H-l.c. at x0 :

Remark 3.1.19. Clearly, if F is C -u.c. (resp. C -H-l.c.) at x0 2 X then F is C -H-

u.c. (resp. C -l.c.) at x0 . If F is C -l.c. at .x0 ; y/ for every y 2 A F .x0 / and A is
compact, then F is uniformly C -l.c. at x0 on A.
Proposition 3.1.20. If F .x0 / is compact, then F is C -u.c. (resp. C -H-l.c.) at x0 2
X provided F is C -H-u.c. (resp. C -l.c.) at x0 :
Proof. The case x0 62 dom F is trivial. So let x0 2 dom F be such that F .x0 / is
compact.
First assume that F is C -H-u.c. at x0 . Consider D Y an open set with F .x0 /
D. Then there exists V 2 NY such that F .x0 / C V D. Since F is C -H-u.c. at
x0 , there exists U 2 NX .x0 / such that F .x/ F .x0 / C V C C D C C for every
x 2 U , and so F is C -u.c. at x0 :
Assume now that F is C -l.c. at x0 and take V 2 NY . There exists V0 2 NY such
that V0 C V0 V . Since F .x0 / is compact, there exists a finite set Y0 F .x0 /
such that F .x0 / Y0 C V0 . Since F is C -l.c. at x0 , for any y 2 Y0 there exists
Uy 2 N .x0 / such that F .x/ \ .y C V0 C / ¤ ;. Take U WD \y2Y0 Uy ; then
U 2 N .x0 /. Take x 2 U and y 2 F .x0 /. Then y D y0 C v for some y0 2 Y0 and
v 2 V0 . Since y0 2 Y0 and x 2 U Uy0 , we have that F .x/ \ .y0 C V0 C / ¤ ;,
that is y0 C v0 c 2 F .x/ for some v0 2 V0 and c 2 C , whence y D y0 C v 2
F .x/ V0 C C C V0 F .x/ C V C C . Hence F is C -H-l.c. at x0 : t
u
With any set-valued map F W X Y we associate its closure F W X Y
defined by F .x/ WD cl .F .x// for x 2 X . Of course, dom F D dom F . Continuity
properties of F and F are deeply related.
Proposition 3.1.21. Let x 2 X and y 2 Y . The following assertions hold:
(i) F is l.c. at .x; y/ if and only if F is l.c. at .x; y/.
(ii) F is l.c. at x if and only if F is l.c. at x.
(iii) If Y is a topological vector space and C Y is a convex cone, then F is
H-C -u.c. (H-C -l.c.) at x if and only if F is H-C -u.c. (H-C -l.c.) at x.
(iv) If F is u.c. at x and either Y is normal or Y is regular and cl F .x/ is compact,
then F is u.c. at x.
(v) If F is u.c. at x and either Y is regular or cl F .x/ is compact, then F is closed
at x.
(vi) F is closed at x if and only if F .x/ is closed and F is closed at x.
(vii) If Y is a metric space, then F is compact at x if and only if F .x/ is closed and
F is compact at x.
Note that if F is u.c. at x, it does not follow that F is u.c. at x (but the implication
is true if F .x/ is closed). Take F W R R, F .x/ D x; x C 1Œ; F is not u.c. at 0,
but F is.
From now on (in this section), Y is a separated topological vector space and
C Y is a convex cone. Recall that the convex cone C determines a quasi-order
C defined by y1 C y2 if y2 y1 2 C .
Definition 3.1.22. We say that

(i) F is C -lower semicontinuous (C -l.s.c. for short) if levF .y/ WD fx 2 X j
F .x/ \ .y C / ¤ ;g is closed for every y 2 Y ;
F .y/ WD fx 2 X j
(ii) F is C -upper semicontinuous (C -u.s.c. for short) if lev<
F .x/ \ .y int C / ¤ ;g is open in X for every y 2 Y .
Note that levF .y/ D .FC /1 .y/ and lev< 1
F .y/ D .Fint C / .y/ for every
1
y 2 Y , so that F is C -l.s.c. iff .FC / is closed-valued, and F is C -u.s.c. iff
.Fint C /1 is open valued. Moreover, if y1 C y2 , then levF .y1 / levF .y2 /
and lev< F .y1 / levF .y2 /. Furthermore, when int C ¤ ; we have that
<
int dom.levF / D dom.lev< F /.

The following result holds.
Proposition 3.1.23. The following assertions hold:
(i) If F is C -u.c. and C is closed then F is C -lower semicontinuous.
(ii) If epi F is closed, then F is C -lower semicontinuous.
(iii) Suppose that Y is a locally convex space, C is closed, int C ¤ ;, F is C -lower
semicontinuous, and either F .x/ is weakly compact or F .x/ C C is closed for
every x 2 X . Then epi F is closed. Moreover, if y0 ; y1 2 Y are such that
y1 y0 2 int C and levF .y1 / is compact, then levF .y0 / is compact and levF
is upper continuous at y0 .
Proof. (i) Take y 2 Y and x 2 X n levF .y/. Then F .x/ \ .y C / D ; (or,
equivalently, F .x/ Y n .y C /). Since C is closed and F is C -u.c. at x, there
exists U 2 N .x/ such that F .x 0 / \ .y C / D ; for every x 0 2 U . This means
that U X n levF .y/, whence levF .y/ is closed.
For the proof of (ii) and (iii) see [214, Proposition 2.5.22]. t
u
Proposition 3.1.24. Consider the set-valued map F W X Y .
(i) Suppose that C is closed and P C is another convex cone. If F is P -upper
continuous, then F is C -lower semicontinuous.
(ii) Suppose that int C ¤ ;; then F is C -lower continuous if and only if F is
C -upper semicontinuous.
From the preceding result we obtain that F is C -l.s.c. and C -l.s.c. when F is
upper continuous (and C is closed).
Proposition 3.1.25. Suppose that int C ¤ ; and y 2 dom.levF /. If levF is lower
continuous at y, then levF .y/ cl lev< F .y/ . Conversely, if X is a topological
vector
space
(or a metric space), levF .y/ is totally bounded, and levF .y/
cl lev<
F .y/ , then lev F is Hausdorff lower continuous at y. Moreover, if levF .y/
is compact, then levF is lower continuous at y.
Definition 3.1.26. We say that F W X Y is weakly C -upper bounded on a set
A X if there exists y 2 Y such that A levF .y/:
Proposition 3.1.27. Consider the following assertions:

(i) int.epi F / ¤ ;;
(ii) there exists x0 2 int.dom F / such that F is weakly C -upper bounded on a
neighborhood of x0 ;
(iii) there exists .x0 ; y0 / 2 graph F such that FC is l.c. at .x0 ; y0 /.
Then (i) ) (ii); moreover, if int C ¤ ; then (iii) ) (ii) ) (i).
Proof. (i) ) (ii) Take .x0 ; y0 / 2 int.epi F /. Then there exists U 2 NX .x0 / and
V 2 NY such that U .y0 CV / epi F ; clearly x0 2 int.dom F / and U fy0 g
epi F . It follows that U levF .y0 /, and so (ii) holds.
Assume now that int C ¤ ;:
(iii) ) (ii) Take y 2 int C ; then V WD .C y/ \ .y C / 2 NY . Because FC is
l.c. at .x0 ; y0 /, there exists U 2 NX .x0 / such that .y0 C V / \ .F .x/ C C / ¤ ;,
or, equivalently, y0 2 F .x/ C V C C for every x 2 U . We obtain that y0 C y 2
F .x/ C C for every x 2 U , and so U levF .y0 C y/. Therefore, (ii) holds.
(ii) ) (i) There exist y 2 Y and U 2 NX .x0 / such that U levF .y/. Then
U fyg epi F , whence U .y C C / epi F . It follows that int.epi F / ¤ ;.
t
u
The continuity properties for vector-valued functions and those of set-valued
maps are related. We give first some continuity notions for extended vector-valued
functions.
Definition 3.1.28. Let f W X ! Y be proper and x0 2 X .
(i) f is C -lower continuous (C -l.c.) at x0 if
8y 2 Y; y f .x0 /; 8V 2 NY ; 9U 2 NX .x0 /;
8x 2 U \ dom f W f .x/ 2 y C V C C I (3.5)
f is C -lower continuous (C -l.c.) if f is C -l.c. at any x0 2 X:

(ii) f is C -upper continuous (C -u.c.) at x0 2 dom f if
8 V 2 NY ; 9 U 2 NX .x0 /; 8 x 2 U W f .x/ 2 f .x0 / C V C I
f is C -upper continuous (C -u.c.) if f is C -u.c. at any x0 2 dom f:

(iii) f is C -lower semicontinuous (C -l.s.c.) if levf .y/ is closed for every y 2 Y .
(iv) f is C -upper semicontinuous (C -u.s.c.) if lev< f .y/ is open for every y 2 Y .
Note that for x0 2 dom f condition (3.5) is equivalent to
8V 2 NY ; 9U 2 NX .x0 /; 8x 2 U \ dom f W f .x/ 2 f .x0 / C V C C;

(3.6)
while for f .x0 / D C1 (3.5) is equivalent to
8y 2 Y; 8V 2 NY ; 9U 2 NX .x0 /; 8x 2 U \ dom f W f .x/ 2 y C V C C:

(3.7)
Recall that Ff;C W X Y is the set-valued map associated to f W X ! Y , that

is, graph Ff;C D f.x; y/ 2 X Y j f .x/ C yg.
Consider x0 2 dom f . If f is C -u.c. at x0 , then x0 2 int.dom f /. On the other
hand, if Ff;C is closed at x0 , then C is closed; if C is closed and Ff;C is u.c. at x0 ,
then Ff;C is closed at x0 . Moreover, f is f0g-u.c. at x0 , f is continuous at x0 ,
x0 2 int.dom f / and f is f0g-l.c. at x0 .
Proposition 3.1.29. Let f W X ! Y be proper and x0 2 dom f . Then:
(i) f is C -l.c. at x0 , Ff;C is C -u.c. at x0 , Ff;C is H-u.c. at x0 .
(ii) f is C -u.c. at x0 , Ff;C is l.c. at x0 , Ff;C is H-l.c. at x0 .
When Y D R and C D RC we have more refined statements; Ff WD Ff;RC in
this case.
Proposition 3.1.30. Let f W X ! R with the associated set-valued map Ff W
X R whose graph is epi f .
(i) Suppose that x0 2 dom f (D dom Ff ). Then f is l.c. at x0 , Ff is u.c. at x0
, Ff is H-u.c. at x0 .
(ii) Suppose that x0 2 X . Then Ff is H-l.c. at x0 ) f is u.c. at x0 , Ff is l.c. at
x0 ; moreover, if either f .x0 / > 1 or X is a topological vector space and f
is convex, then f is u.c. at x0 , Ff is H-l.c. at x0 .
Proposition 3.1.31. (i) If f W X ! Y is a proper C -l.c. function and C is a
proper closed convex cone, then epi f is closed.
(ii) Let fi W Xi ! Y .i 2 f1; 2g/ be proper and C -l.c., where .Xi ; i / are
topological spaces and f W X1 X2 be defined by f .x1 ; x2 / WD f1 .x1 / C
f2 .x2 /. Then f is C -l.c.
(iii) Let f W .Z; / ! Y be a proper C -l.c. function and g W .X; / ! .Z; / be
continuous. Then f ı g is C -l.c.
(iv) Let Z be a topological vector space ordered by the convex cone P , f W Z !
Y be proper, C -l.c. and increasing such that for every y 2 Y there exists
z 2 Z with f .z/
y, and g W .X; / ! Z be proper and P -l.c. Then f ı g
is C -l.c., where f .C1/ WD C1:
Proof. (i) Let .x; y/ 2 X Y n epi f . We have the following three situations: a)
x … cl.dom f /, b) x 2 cl.dom f / n dom f , and c) x 2 dom f:
a) In this case there exists U 2 NX .x/ such that U \ dom f D ;, and so
.U Y / \ epi f D ;. Hence .x; y/ … cl.epi f /:
b) In this case f .x/ D C1. Because Y ¤ C D \fC C V j V 2 NY g,
there exist y 0 2 Y and V 0 2 NY such that y 0 … C C V 0 . Then, there exists
V 2 NY such that V C V V 0 . Using (3.7), there exists U 2 NX .x/
such that f .x 0 / 2 y y 0 C V C C for every x 0 2 U \ dom f . Then
.U .y C V // \ epi f D ;. In the contrary case, there exists x 0 2 U
such that f .x 0 / 2 y C V C . Since f .x 0 / 2 y y 0 C V C C , we get
y 0 2 V V C C C C V 0 C C , a contradiction. Hence .x; y/ … cl.epi f /:
c) In this case we have that y … f .x/ C C . Since C is closed, as above, there

exists V 2 NY such that .y C V / \ .f .x/ C V C C / D ;. Using (3.6), there
exists U 2 NX .x/ such that f .x 0 / 2 f .x/ C V C C for every x 0 2 U \
dom f . It follows that .U .y CV //\epi f D ;, and so .x; y/ … cl.epi f /:
(ii) Let first x D .x1 ; x2 / 2 dom f D dom f1 dom f2 and V 2 NY ; there exists
V 0 2 NY such that V 0 C V 0 V: Since fi is C -l.c. at xi , there exists Ui 2
NXi .xi / such that fi .xi0 / 2 fi .xi / C V 0 C C for every xi0 2 Ui .i 2 f1; 2g/.
Therefore, for .x10 ; x20 / 2 .U1 U2 / \ dom f we have that
f .x10 ; x20 / D f1 .x10 / C f2 .x20 / 2 f1 .x1 / C V 0 C C C f2 .x2 /

CV 0 C C f .x1 ; x2 / C V C C:
Hence f is C -l.c. at x:
Let now x 2 cl.dom f / n dom f D .cl.dom f1 / cl.dom f2 // n .dom f1
dom f2 /. We can have one of the following situations: a) x1 2 cl.dom f1 / n
dom f1 and x2 2 cl.dom f2 / n dom f2 , b) x1 2 cl.dom f1 / n dom f1 and x2 2
dom f2 , c) x1 2 dom f1 and x2 2 cl.dom f2 / n dom f2 :
a) Let y 2 Y , V 2 NY , and take V 0 2 NY such that V 0 C V 0 V . Then for
i 2 f1; 2g;
9Ui 2 NXi .xi /; 8xi0 2 Ui \ dom fi W fi .xi0 / 2 12 y C V 0 C C;
and so
9U W D U1 U2 2 NX1 X2 .x/; 8x 0

D .x10 ; x20 / 2 U \ dom f W f .x 0 / 2 y C V C C: (3.8)
Hence f is C -l.c. at x:
b) Let y 2 Y , V 2 NY , and take V 0 2 NY such that V 0 C V 0 V . Then
9U1 2 NX1 .x1 /; 8x10 2 U1 \ dom f1 W f1 .x10 / 2 y f2 .x2 / C V 0 C C;

9U2 2 NX2 .x2 /; 8x20 2 U2 \ dom f2 W f2 .x20 / 2 f2 .x2 / C V 0 C C:
It follows that (3.8) holds, and so f is C -l.c. at x:

c) This case follows from b) interchanging 1 and 2:
(iii) We have that dom.f ı g/ D g1 .dom f /. Let x 2 X and set z WD g.x/. If
f .z/ D 1, it is nothing to do.
Assume that f .z/ 2 Y and take V 2 NY . Because f is C -l.c. at z, there
exists W 2 NZ .z/ such that f .z0 / 2 f .z/ C V C C for every z0 2 W \ dom f .
By the continuity of g at x, there exists U 2 NX .x/ such that g.x 0 / 2 W for
every x 0 2 U . It follows that .f ı g/.x 0 / D f .g.x 0 // 2 f .g.x// C V C C for
every x 0 2 U \ dom f ı g. Hence f ı g is C -l.c. at x in this case.
Assume now that f .z/ D 1 and take y 2 Y , V 2 NY : Because f is

C -l.c. at z, there exists W 2 NZ .z/ such that f .z0 / 2 y C V C C for every
z0 2 W \ dom f . Continuing as in the previous case, we get U 2 NX .x/ such
that g.x 0 / 2 W for every x 0 2 U . It follows that .f ı g/.x 0 / D f .g.x 0 // 2
y C V C C for every x 0 2 U \ dom.f ı g/. Hence f ı g is C -l.c. at x in this
case, too.
(iv) Consider x0 2 X and set y0 WD .f ı g/.x0 / 2 Y [ fC1g, z0 WD g.x0 / 2
Z [ fC1g. Assume first that z0 2 Z. Let V 2 NY and y 2 Y , with y D
y0 if y0 2 Y . Because f is C -l.c. at z0 , there exists W 2 NZ such that
f .z/ 2 y C V C C for every z 2 .z0 C W / \ dom f . Because g is P -l.c.
at x0 , there exists U 2 NX .x0 / such that g.x/ 2 z0 C W C P for every
x 2 U \ dom g. Take x 2 U \ dom.f ı g/ U \ dom g. Then g.x/
P z for
some z 2 z0 C W , and so f .g.x//
C f .z/ because f is increasing. It follows
that z 2 .z0 C W / \ dom f and so f .g.x// 2 f .z/ C C y C V C C C C D
y C V C C . Therefore, f ı g is C -l.c. at x0 :
Assume now that z0 D C1; consequently, y0 D C1. Let V 2 NY and
y 2 Y . By hypothesis, there exists z 2 Z such that f .z/
y. Because f
is C -l.c. at z, there exists W 2 NZ such that f .z0 / 2 y C V C C for every
z0 2 .z C W / \ dom f . Because g is P -l.c. at x0 and z P g.x0 / D C1, there
exists U 2 NX .x0 / such that g.x/ 2 zCW CP for every x 2 U \dom g. Take
x 2 U \ dom.f ı g/ U \ dom g. Then g.x/
P z0 for some z0 2 z C W ,
and so f .g.x//
C f .z0 /. It follows that z0 2 .z C W / \ dom f and so
f .g.x// 2 f .z0 / C C y C V C C C C D y C V C C . Therefore, f ı g is
C -l.c. at x0 . Hence f ı g is C -l.c. t
u
We are now interested in continuity properties of the composition of two set-
valued maps and the composition of a function with a set-valued map. So consider
another topological space U , the functions f W X U ! Y , g W X ! Y ,
and the set-valued maps W U X , F W X U Y and G W X
Y ; we associate the set-valued maps f; g; F; G W U Y defined by
.f/.u/ WD f ..u/ fug/, .g/.u/ WD g ..u//, .F/.u/ D F ..u/ fug/,
and .G/.u/ WD G ..u//.
Proposition 3.1.32. Let f; g and ; F; G be as above and u0 2 dom .
(i) If is u.c. at u0 and g is C -l.c. on .u0 /, then g is C -u.c. at u0 . Moreover,
if .u0 / is compact, then g is H-C -u.c. at u0 .
(ii) If is u.c. at u0 , F is C -u.c. at .x; u0 / for all x 2 .u0 /, and .u0 / is compact,
then F is C -u.c. at u0 . If f is C -l.c. on .u0 /, then f is H-C -u.c. at u0 .
(iii) If is l.c. at .x0 ; u0 / 2 graph and F is C -l.c. at .x0 ; u0 I y0 /, then F is
C -l.c. at .u0 ; y0 /. In particular, if f is C -u.c. at .x0 ; u0 /, then f is C -l.c. at
.u0 ; y0 /, where y0 D f .u0 ; x0 /.
(iv) If is compact at u0 and F is closed at .x; u0 / for all x 2 .u0 /, then F is
closed at u0 .
(v) If is compact at every u 2 U and G is C -l.s.c., then G is C -l.s.c., too.
Suppose now that X is a topological vector space, too.
(vi) If is H-u.c. at u0 and f is equi-C -l.c. on .u0 / fu0 g, then f is H-C -u.c.
at u0 .
(vii) If is H-l.c. at u0 and f is equi-C -u.c. on .u0 / fu0 g, then f is H-C -l.c.
at u0 .
Detailed references for the most part of the results in this section are
mentioned at the end of [214, Section 2.5]; we mention only some missing
references in [214, Section 2.5] and those for the results not mentioned there.
So, Propositions 3.1.23, 3.1.24, 3.1.25 are stated in [184, Proposition 2.3] and [185,
Proposition 3.1]. Proposition 3.1.27 is essentially established in Song [535]. The
continuity notions for vector-valued functions are in accordance with those of Penot
and Théra [473]. The statement of Proposition 3.1.31 is adapted from [611].
3.2 Continuity Properties of Set-Valued Maps Under

Convexity Assumptions
Throughout this section X and Y are topological vector spaces, C Y is a convex

cone and W X Y is a set-valued map.
Proposition 3.2.1. Let x0 2 dom .
(i) Assume that (a) is nearly convex, and (b) there exist U 2 NX .x0 / and B Y
a bounded set such that .x/ \ .B C / ¤ ; for every x 2 U . Then (c) is
C -lower continuous at x0 .
(ii) Conversely, assume that (c) and (d) B C 2 NY .0/ for some bounded set
B Y hold. Then (b) holds.
Proof. (i) Because isSnearly convex, there exists ˛ 2 0; 1Œ such that is C -˛-
convex. Let ˛ WD n0 ˛n , where
˛0 WD f0; 1g; ˛nC1 WD f˛t C .1 ˛/s j t; s 2 ˛n g 8 n

0:
As seen in [214, Lemma 2.4.2], one has that cl ˛ D Œ0; 1:

Let y0 2 .x0 / and V 2 NY .0/. Since B is bounded, By0 is also bounded,
and so there is some
2 ˛ n f0; 1g such that
.B y0 / V . Fix u 2 U 0 WD
U x0 2 NX .0/; because .x0 C u/ \ .B C / ¤ ;, there is some b 2 B and
v 2 C such that b v 2 .x0 C u/. By the ˛-convexity of we obtain that
.1
/y0 C
.b v/ 2 .1
/ .x0 / C
.x0 C u/ .x0 C
u/ C C;
whence y0 C
.b y0 / 2 .y0 C V / \ . .x0 C
u/ C C /. Hence .y0 C V / \
. .x/ C C / ¤ ; for every x 2 x0 C
U 0 .2 N .x0 //, and so is C -lower
continuous at x0 .
3.2 Continuity Properties of Set-Valued Maps Under Convexity Assumptions 91
(ii) By (d) we have that V WD B C 2 NY .0/. Take y0 2 .x0 /. Because is

C -lower continuous at x0 , there exists U 2 N .x0 / such that . .x/ C C / \
.y0 C V / ¤ ; for every x 2 U . It follows that .x/ \ .y0 C B C / ¤ ;
for x 2 U , which proves that verifies condition (b) (y0 C B being bounded).
The proof is complete. t
u
Corollary 3.2.2. Assume that is nearly convex. Then (ii) ) (iii) in Proposi-
tion 3.1.27 holds. Therefore, if int C ¤ ;, then conditions (i), (ii) and (iii) in
Proposition 3.1.27 are equivalent.
Proof. If (ii) of Proposition 3.1.27 holds, then condition (b) in Proposition 3.2.1
holds, and so is C -lower continuous at x0 by the latter result. Taking y0 2 F .x0 /,
(iii) in Proposition 3.2.1 holds. The proof is complete. t
u
Note that the condition (d) in Proposition 3.2.1 is verified if int C ¤ ; or if Y is a
normed vector space. Moreover, if int C ¤ ; then condition (b) in Proposition 3.2.1
is equivalent to the weak C -upper boundedness of on a neighborhood of x0 :
As in [25] and [26], we say that E Y is quasi (C -lower) bounded if there
exists a bounded set B Y such that E B C C ;
Theorem 3.2.3. Let be C -nearly convex. Suppose that is C -lower continuous
at .x0 ; y0 / 2 graph . Then is C -lower continuous at every x 2 int.dom / D
core.dom /. Moreover, is H-C -lower continuous at every x 2 int.dom / for
which .x/ is quasi-bounded.
The next result is related to the boundedness condition of the preceding theorem.
Proposition 3.2.4. Let be C -nearly convex. Suppose that x0 2 int.dom / is
such that .x0 / is quasi-bounded. Then .x/ is quasi-bounded for every x 2
dom .
As an application of the preceding result we obtain that f W X ! R is finite-
valued on dom f whenever f is nearly convex and finite-valued at x0 2 int.dom f /.
Proposition 3.2.5. Assume that is C -lower continuous at .x0 ; y0 / 2 graph
and C -nearly convex. If .x0 / is quasi-bounded, then is H-C -u.c. at x0 .
The preceding results lead to the following important theorem.
Theorem 3.2.6. Let be a C -nearly convex. Suppose that is C -lower continu-
ous at .x0 ; y0 / 2 graph and .x0 / is quasi-bounded. Then is H-C -continuous
on int.dom /.
The next result corresponds to another well-known result of convex analysis. It
is an immediate consequence of Proposition 3.2.1 and Theorem 3.2.6.
Corollary 3.2.7. Let be a C -nearly convex. Suppose that there exist x0 2
int.dom /, B Y a bounded set, and U 2 NX such that .x0 / B C C
and .x0 C u/ \ .B C / ¤ ; for every u 2 U . Then is H-C -continuous on
int.dom /.
From Corollary 3.2.7, also taking into account Proposition 3.1.30, we obtain that
f W X ! R is (finite-valued and) continuous on int.dom f / whenever f is nearly
convex, f .x0 / 2 R, and f is bounded above on a neighborhood of x0 .
Theorem 3.2.8. Let be nearly convex. Assume that is lower continuous at
some .x0 ; y0 / 2 graph . Then is closed at any x 2 int.dom / with .x/
closed.
Proof. Using Theorem 3.2.3 we have that is lower continuous at any x 2
int.dom /. Because is nearly convex, there exists ˛ 2 0; 1=2 such that
is ˛-convex.
Fix x 2 int.dom / such that .x/ is closed. Take y1 2 Y n .x/. We claim
that there exists y2 2 .x/ such that ˛y1 C .1 ˛/y2 … .x/: In the contrary case
for every z 2 .x/ one has ˛y1 C .1 ˛/z 2 .x/. Inductively we obtain that
.1 .1 ˛/n/y1 C .1 ˛/n z 2 .x/ for every n
1, and so y1 2 cl .x/ D .x/,
a contradiction. Consider y2 2 .x/ with ˛y1 C .1 ˛/y2 … .x/. Because .x/
is closed, there exists V 2 NY .0/ such that
Œ˛y1 C .1 ˛/y2 C ˛V C .1 ˛/V \ .x/ D ;: (3.9)
Because is lower continuous at .x; y2 /, there exists U 2 NX such that .x 0 / \

.y2 C V / ¤ ; for every x 0 2 x C U . We claim that
Œ.x C U / .y1 C V / \ graph D ;; (3.10)
and so .x; y1 / … cl.graph /: Otherwise, let .x; O y/

O 2 Œ.x C U / .y1 C V / \
graph and take x 0 WD .1 ˛/1 .x ˛ x/. O Since xO 2 x C U , we have that
x 0 2 1˛
1
x 1˛
˛
x 1˛
˛
U x C U . Therefore, there exists y 0 2 .x 0 / \ .y2 C V /.
On one hand, because is ˛-convex, we obtain that ˛.x; O y/O C .1 ˛/.x 0 ; y 0 / 2
0 0
graph , and so ˛ yO C .1 ˛/y 2 .˛ xO C .1 ˛/x / D .x/. On the other hand,
˛ yO C.1˛/y 0 2 ˛.y1 CV /C.1˛/.y2 CV /, contradicting (3.9). Therefore, (3.10)
holds.
Let now ..xi ; yi //i 2I graph converge to .x; y/; clearly, .x; y/ 2
cl.graph /. Assuming that y … .x/, by the argument above, we obtain the
contradiction .x; y/ … cl.graph /. Hence y 2 .x/, and so is closed at x: u t
In fact, semicontinuous nearly convex set-valued maps are very close to convex
set-valued maps.
Proposition 3.2.9. Suppose that W X Y is closed-valued and nearly convex.
If dom is convex (for example, if dom is closed or open) and is H-u.c. at
every point of dom , then is convex.
From Theorem 3.2.6 one gets the following characterization of the continuity of
nearly convex proper functions.
Corollary 3.2.10. Let f W X ! Y be a nearly convex proper function and x0 2

int.dom f /. Suppose that C is normal. Then: f is C -u.c. at x0 , f is lower
continuous at x0 , f is H-continuous at every x 2 int.dom f / , f is continuous
on int.dom f / , f is continuous at x0 .
Another characterization of the continuity of a convex function is provided in the
next result.
Proposition 3.2.11. Let C be normal and f W X ! Y be a proper convex
operator. Then:
(i) f is continuous at x0 2 dom f iff
9z 2 Y; 8V 2 NY ; 9U 2 NX ; 8x 2 x0 C U W f .x/ 2 z C V C: (3.11)
Consequently, if g W X ! Y is another proper convex operator such that

g f and f is continuous at x0 2 dom f , then g is continuous at x0 , too.
(ii) If
9 B Y; B bounded, 9U 2 NX W f .x0 C U / B C; (3.12)
then f is continuous at x0 . If int C ¤ ; and f is continuous at x0 2 dom f

then (3.12) holds (even with B a singleton).
Proof. (i) Let us set
N WD fV 2 NY j V D .V C C / \ .V C /g I
because C is normal, N is a base of neighborhoods of the origin of Y:

Assume that f is continuous at x0 2 int.dom f /. Then
8 V 2 NY ; 9 U 2 NX W f .x0 C U / f .x0 / C V;
and so, taking z WD f .x0 /, (3.11) holds. Conversely, assume that (3.11) holds.
Replacing if necessary f by fQ defined by fQ.x/ WD f .x0 C x/ f .x0 /, we
assume that x0 D 0 and f .0/ D 0. Consider V 2 NY ; there exists V1 2 N
such that V1 C V1 V . Since V1 2 NY , there exists
2 0; 1 such that

z 2 V1 . From (3.11) there exists U 2 NX such that f .x/ 2 z C V1 C for
every x 2 x0 C U . Taking into account that V1 is balanced, f is convex and
f .0/ D 0, for x 2 U we have that f .
x/
f .x/, and so
f .
x/ 2
f .x/ C
z C
V1 C V1 C V1 C V C I
moreover,
1
0 D f .0/ D f 2
x C 12 .
x/ 12 f .
x/ C 12 f .
x/;
and so
f .
x/ 2 f .
x/ C C C V C C D V C C
because
x 2 U . Therefore, f .u/ 2 .V C / \ .V C C / D V for every
u 2
U , and so f is continuous at 0 D x0 :
(ii) Assume that (3.12) holds. As above we assume that x0 D 0 and f .0/ D 0.
Therefore, there exist U0 2 NX and B Y bounded such that f .U0 /
B C . Consider V 2 N . Since B is bounded, there exists
2 0; 1 such that

B V . Taking into account that V is balanced, f is convex and f .0/ D 0,
for x 2 U0 we have that f .
x/
f .x/, and so f .
x/ 2
f .x/ C

B
C C V C . Therefore, (3.11) holds with z WD 0:
Assume that int C ¤ ; and take z 2 int C ; then z C is a neighborhood of
0 2 Y . Since f is continuous at x0 , there exists U 2 NX such that f .x0 CU /
f .x0 / C z C . Therefore, (3.12) holds with B WD ff .x0 / C zg: t
u
Recall that the Minkowski (gauge) functional associated to a subset A of a linear
space E is the functional

infft > 0 j x 2 tAg if x ¤ 0;
A W E ! R; A .x/ WD
0 if x D 0:
It is clear that A is positively homogeneous and x 2 tA for all t > A .x/ whenever
A D Œ0; 1A. Moreover, A is convex (hence sublinear) if A is convex. In the case
in which A is an absorbing set, A is finite-valued and it coincides with pA as
introduced in Proposition 6.2.1.
Lemma 3.2.12. Let X , Y be topological vector spaces, f W X ! Y be a proper
convex function and x0 2 dom f . Suppose that for some U 2 NX and some
balanced set V Y one has
8 x 2 x0 C U W f .x/ 2 f .x0 / C V C: (3.13)
Then
8 x 2 x0 C U W ŒV C .f .x/ f .x0 // pU .x x0 /: (3.14)
Proof. Replacing f by g W X ! Y , g.x/ D f .x0 C x/ f .x0 /, we may suppose

that x0 D 0 and f .x0 / D 0. Therefore f .x/ 2 V C for every x 2 U . Since
0 D f 12 x 0 C 12 .x 0 / 12 f .x 0 / C 12 f .x 0 / for every x 0 2 X , we obtain that
f .x/ 2 f .x/ C C V C C C C D V C C for x 2 U , and so f .x/ 2 ŒV C
for all x 2 U:
From Proposition 6.2.1 we have that fx 2 X j pU .x/ < 1g U fx 2
X j pU .x/ 1g. Let x 2 U ., x 2 U / and set 1
t WD pU .x/
0.
Take s WD 1 if t D 1 and s 2 t; 1 if t < 1. Then x 0 WD s 1 x 2 U . Since f
is convex and s 1, we obtain that f .x/ sf .x 0 / C .1 s/f .0/ 2 sŒV C . It

follows that f .˙x/ 2 sŒV C C . Since f .x/ 2 f .x/ C C , we obtain that
f .x/ 2 ŒsŒV C C D sŒV C . Therefore, ŒV C .f .x// t D pU .x/, and so (3.14)
holds. t
u
Under slightly stronger conditions than those in Lemma 3.2.12 we have a
stronger conclusion.
Proposition 3.2.13. Let X , Y be topological vector spaces, f W X ! Y be a
proper convex function and x0 2 dom f . Suppose there exist a closed and convex
U 2 NX and a balanced convex set V Y such that condition (3.13) is satisfied.
Then for every 2 0; 1Œ,
1C
8 x; x 0 2 x0 C U W ŒV C f .x/ f .x 0 / pU .x x 0 /: (3.15)
1
Proof. As in the proof of Lemma 3.2.12, we may assume that x0 D 0 and

f .x0 / D 0. Of course, by Proposition 6.2.1 and Remark 6.2.6, pU is a continuous
semi-norm with U D fx j pU .x/ 1g and int U D fx j pU .x/ < 1g. Let 2 0; 1Œ
and consider x; x 0 2 U int U . Observe that jpU .tx C .1 t/x 0 / pU .x 0 /j
jtj pU .x x 0 /, and so

pU tx C .1 t/x 0 pU .x 0 / C jtj pU .x x 0 / C jtj pU .x x 0 / 8t 2 R:
Let us set
A WD ft 2 R j zt WD tx C .1 t/x 0 2 U g; t0 WD sup A:
Clearly, A is an interval and t0 2 1; 1 (because x 2 int U ); moreover, t0 D 1

if pU .x x 0 / D 0: We claim that t0
pU 1 0
.xx 0 / if pU .x x / > 0. Indeed, let
' W R ! R, '.t/ D pU .tx C .1 t/x 0 /. Since pU is convex and continuous, so
is '. Moreover, '.t/
tpU .x x 0 / pU .x 0 /, and so limt !1 '.t/ D 1. As
'.1/ D pU .x/ < 1, there exists t > 1 such that '.t / D 1. It follows that
'.t/ < 1 for t 2 Œ1; t/, and so Œ1; t / A. Since pU .zt / D 1 and t.x x 0 / D zt x 0 ,
we get t pU .x x 0 / D pU .zt x 0 /
pU .zt / pU .x 0 /
1 , and so t0
pU 1
.xx 0 /
as claimed. For t 2 1; t0 Œ we have that x D .1 t 1 /x 0 C t 1 zt , and so f .x/
.1 t 1 /f .x 0 / C t 1 f .zt /. Using Lemma 3.2.12, for all s 2 1; 1Œ, t 2 1; t0 Œ and
0 > we have

f .x/ f .x 0 / 2 t 1 f .zt / f .x 0 / C t 1 sŒV C C 0 ŒV C C
t 1 .s C 0 /ŒV C C;
and so, because ˛ŒV C ˇŒV C for 0 < ˛ < ˇ, f .x/f .x 0 / 2 t 1 .sC0 /ŒV C C
for all t 2 0; t0 Œ, s > 1, 0 > . It follows that
1C
8t > pU .x x 0 / W f .x/ f .x 0 / 2 tŒV C C:
1
Changing the roles of x and x 0 , and taking into account that ŒŒV C C D ŒV C , we
get
1C
8t > pU .x x 0 / W f .x/ f .x 0 / 2 tŒV C I (3.16)
1
therefore, (3.15) holds. t

u
In the case in which X , Y are normed vector spaces Proposition 3.2.13 can be
used to get the local lipschitzianity of the proper convex function f W X ! Y (see
Theorem 3.3.4).
The results presented without proofs are quoted from the book [214]; see there
the bibliographical comments. Proposition 3.2.1 extends [260, Proposition 3.8].
Theorem 3.2.8 is a slight generalization of Theorem 1 in [486]. The proof of
Proposition 3.2.11 follows that given in [609]. The function A is introduced
and studied for general sets in [474]. Lemma 3.2.12, Proposition 3.2.13 and
Theorem 3.3.4 are slight generalizations of Proposition 2.3 and Corollary 2.4 (c), (d)
in [67]; their proofs mimics the proofs of Lemma 2.2.8, Theorem 2.2.11 and
Corollary 2.2.12 in [614].
3.3 Lipschitz Properties for Single-Valued and Set-Valued

Maps
It is well known from scalar optimization that one needs certain assumptions
concerning the objective function as well as the feasible set, especially in order to
show the stability of solutions with respect to perturbations or rates of convergence
in numerical procedures (compare Klatte, Kummer [329]). Under the assumption
of classical Lipschitz continuity (see Definition 3.3.1) it is possible to show such
results. An important characteristic of the classical Lipschitz continuity (Defini-
tion 3.3.1) in comparison with the general continuity for single-valued functions is
a linear rate of continuity quantified by some modulus (Lipschitz constant L).
Under certain Lipschitz properties characterizations of well-posedness and
sensitivity analysis for constraint systems as well as variational systems are
given in Chap. 4 of the book by Mordukhovich [430]. Furthermore, Lipschitz
continuity is important for deriving calculus rules in generalized differentiation (see
Sects. 3.4, 3.5).
Definition 3.3.1. Consider a map f W S X ! Y , where X and Y are n.v.s.. The
mapping f is Lipschitz continuous around x 2 X (or Lipschitzian around x 2 X )
if there is a neighborhood U of x and a constant L
0 such that
3.3 Lipschitz Properties for Single-Valued and Set-Valued Maps 97
jjf .x/ f .u/jj Ljjx ujj for all x; u 2 S \ U: (3.17)
Remark 3.3.2. In Definition 3.3.1 we can take f W X ! Y and S D dom f .

It is known that a proper convex function f W R ! R is locally Lipschitzian
on int.dom f / (see [614, Theorem 2.1.5]). In [407, Theorem 3.1] a corresponding
result is shown for C -convex vector-valued functions in the finite dimensional
setting. Recall that a vector-valued function f W S ! Y , S X convex, is called
C -convex (C is a proper convex cone in Y ), if for every x; y 2 S ,
2 Œ0; 1, one
has
f .
x C .1
/y/ 2
f .x/ C .1
/f .y/ C:
Theorem 3.3.3. Suppose that C Rq is a proper convex cone, cl C is pointed,

S Rn is a proper convex set. Furthermore, assume that f W S ! Rq is C -convex.
Then f is locally Lipschitzian on the relative interior of S .
A more general result is given by Borwein [67, Corollary 2.4] for normed vector
spaces X , Y and a normal cone C Y ; we give a proof using Proposition 3.2.13.
Theorem 3.3.4. Let X , Y be normed vector spaces, f W X ! Y be a proper
convex function. Assume that f is continuous at some x0 2 dom f and the cone C
is normal. Then f is locally Lipschitz on int.dom f /. Moreover, if
8x 2 x0 C RUX W kf .x/ f .x0 /k M; (3.18)
for some R; M > 0, then

M R C r
8x; x 0 2 x0 C rUX W f .x/ f .x 0 / x x 0 ;
R Rr
where r 2 0; RŒ and WD sup fkyk j y 2 ŒUY C g :
Proof. Assume that (3.18) holds. Take U WD RUX 2 NX , V WD M UY 2 NY
and WD r=R 2 0; 1Œ. Clearly, (3.13) holds. Observe that ŒV C M UY , pU D
R1 kk and ŒV C D pŒV C
pM UY D .M /1 kk. Using Proposition 3.2.13, for
x; x 0 2 x0 C rU we get
1C
.M /1 f .x/ f .x 0 / ŒV C f .x/ f .x 0 / pU .x x 0 /
1
1 C 1
D R x x 0 :
1
Hence (3.16) holds.
Because the continuity of f at some point of dom f implies the continuity of f
on int.dom f / (see Corollary 3.2.10), the conclusion follows. t
u
In order to show a subdifferential chain rule (Lemma 3.5.14) or the relationship

between the coderivative of a vector-valued function and the subdifferential of its
scalarization ([430, Theorem 3.28], compare Sect. 3.5) we assume that a scalar
function is strictly Lipschitzian in the following sense:
Definition 3.3.5. Suppose that X and Y are n.v.s. and f W X ! Y is a
single-valued map. f is called strictly Lipschitzian at x if f is Lipschitzian around
x and there exists a neighborhood V of the origin in X such that the sequence
f .xk C tk v/ f .xk /
yk WD ; k 2 N;
tk
contains a norm convergent subsequence whenever v 2 V , xk ! x, and tk # 0.

Remark 3.3.6. If Y is finite-dimensional, the class of strictly Lipschitzian mappings
reduces to the class of locally Lipschitzian mappings f W X ! Rq . This is not the
case for dim Y D C1, see [430, Example 1.35]. Every at x strictly differentiable
map is strictly Lipschitzian at this point.
In the following we study natural extensions of Lipschitz continuity to set-valued
mappings. These Lipschitzian properties will be used for deriving generalized dif-
ferential calculus and optimality conditions (see Chap. 12). Especially in Sect. 12.8,
we need an extension of Lipschitzian continuity in order to prove necessary
conditions for minimizers of the set-valued optimization problem .SP /.
In the following we present several notions of Lipschitz-like set-valued map-
pings. We study set-valued mappings F W X Y , where X and Y are n.v.s., with
dom F ¤ ;.
Definition 3.3.7. Consider nonempty subsets U X and V Y . A set-valued
map F W X Y is called Lipschitz-like on U relative to V if there exist a
number L
0 such that for every x; u 2 U one has:
F .x/ \ V F .u/ C L kx uk BY ;
where BY denotes the closed unit ball of Y:

Remark 3.3.8. If F .U / \ V D ; then F is Lipschitz-like on U relative to V .
If F W X Y is Lipschitz-like on U relative to V and F .U / \ V ¤ ;, then
U dom F .
Definition 3.3.9. Consider a set-valued map F W X Y and a point .x; y/ 2
graph F . F is called Lipschitz-like around .x; y/ with modulus L
0 if there are
neighborhoods U of x and V of y s.t. for every x; u 2 U one has:
F .x/ \ V F .u/ C L kx uk BY :
The infimum of all such moduli L is called the exact Lipschitzian bound of F
around .x; y/ and is denoted by Lip F .x; y/.
Remark 3.3.10. The local Lipschitz-like property is also known as the pseudo-
Lipschitzian property or the Aubin property of multifunctions.
The case V D Y in Definitions 3.3.7 and 3.3.9 leads us to the following
definition:
Definition 3.3.11. Consider a nonempty set U X .
(i) A set-valued map F W X Y is called Lipschitz continuous on U if there
exist a number L
0 such that for every x; u 2 U one has:
F .x/ F .u/ C L kx uk BY :
(ii) F is called Lipschitzian around x with the exact bound Lip F .x/ if V D Y
Remark 3.3.12. If F W X Y is Lipschitz on U one has U \ dom F D ; or
U dom F .
Remark 3.3.13. It is easy to see that for single-valued mappings f W S X ! Y
all the properties in Definitions 3.3.9 and 3.3.11 reduce to the classical Lipschitz
continuity given in Definition 3.3.1.
Remark 3.3.14. The local Lipschitz-like property of a set-valued mapping F W
X Y (Definition 3.3.9) can be viewed as a localization of the Lipschitzian
behavior not only relative to a point of the domain but also relative to a particular
point of the image y 2 F .x/.
Remark 3.3.14 leads us to the following characterization in terms of the local
Lipschitz continuity of the scalar distance function d.u; A/ WD infv2A jju vjj (u 2
Y , A Y ), where we use the convention inf ; WD C1, to the moving set F .x/
with respect to both variables .x; y/ (see Mordukhovich [430, Theorem 1.41]).
Theorem 3.3.15 (Scalarization of the Lipschitz-Like Property). Consider a set-
valued mapping F W X Y and .x; y/ 2 graph F . The following properties are
equivalent:
(a) F is Lipschitz-like around .x; y/.
(b) The scalar function W X Y ! R defined by (with the convention inf ; D
C1)
.x; y/ WD d.y; F .x// WD inf jjy vjj

v2F .x/
is Lipschitzian around .x; y/.

Remark 3.3.16. The local properties in Definitions 3.3.9 and 3.3.11 are stable /
robust with respect to small perturbations of the reference points. These properties
hold for F if and only if they hold for the mapping F W X Y with F .x/ WD
cl F .x/.
Remark 3.3.17. From Definition 3.3.11, (i), it follows that the Lipschitz continuity
of F on U is equivalent to
8x; u 2 U W haus.F .x/; F .u// Ljjx ujj;
where haus.A1 ; A2 / is the Hausdorff distance (or Pompieu-Hausdorff distance)

between A1 ; A2 Y that is defined by
haus.A1 ; A2 / WD inff
0 j A1 A2 C BY ; A2 A1 C BY g:
It is important to mention that the Hausdorff distance furnishes a metric on the

space of all nonempty and compact (even closed and bounded) subsets of Y . Thus,
if a set-valued map F W X Y is compact-valued, its Lipschitz continuity in the
sense of Definition 3.3.11 is equivalent to the Lipschitz continuity of a single-valued
map x ! F .x/ (see Definition 3.3.1) from X to the space of all nonempty, compact
subsets of Y equipped with the Haussdorff metric.
In the following we explain some relationships between the local Lipschitzian
(Definition 3.3.11, (ii)) and Lipschitz-like properties (Definition 3.3.7) of multi-
functions. It follows directly from the definitions that, if F is Lipschitzian around
x 2 dom F , then it is Lipschitz-like around .x; y/ for every y 2 F .x/ with
Lip F .x/
supfLip F .x; y/ j y 2 F .x/g: (3.19)
The converse of this assertion holds with the equality in (3.19) under some
additional assumptions concerning F . The following theorem is shown by Mor-
dukhovich [430, Theorem 1.42] under the assumption that F W X Y is compact
around x 2 dom F , i.e., there exist a neighborhood U of x such that cl F .U /
is compact, that is F .U / is relatively compact. Furthermore, we assume that F is
closed at x, i.e., for every y … F .x/ there are neighborhoods U of x and V of y
such that F .x/ \ V D ; for all x 2 U . F is closed at x in this sense if F is closed
at x in the sense of [214, Definition 2.5.7 (ii)]. F is closed at x implies that F is
closed-valued at x, i.e., F .x/ is closed.
In the following theorem we discuss the assumptions for the equivalence of the
properties that F is Lipschitzian around x (Definition 3.3.11, (ii)) and that F is
Lipschitz-like around .x; y/ (Definition 3.3.9).
Theorem 3.3.18 (Lipschitz Continuity of Locally Compact Multifunctions).
Let F W X Y be closed at x 2 dom F and compact around this point.
Then F is Lipschitzian around x if and only if it is Lipschitz-like around .x; y/
for every y 2 F .x/. In this case we have
Lip F .x/ D maxfLip F .x; y/ j y 2 F .x/g < C1:
The proof of this result is given by Mordukhovich [430, Theorem 1.42].

Let us now consider the case of a mapping F W X Y between Asplund spaces

(see Definition 3.5.3) with the range space Y partially ordered by a proper, closed,
convex and pointed cone C Y . Recall that the epigraph of F with respect to the
ordering cone C is given by
˚ ˇ
epi F D .x; y/ 2 X Y ˇ y 2 F .x/ C C ;
and the epigraphical multifunction of F denoting by EF W X Y is defined by

˚ ˇ
EF .x/ WD y 2 Y ˇ y 2 F .x/ C C with graph EF D epi F: (3.20)
In Sect. 12.8, Theorem 12.8.9 we will show necessary conditions for minimizers
of set-valued optimization problems under the assumption that the set-valued
mapping F W X Y is epigraphically Lipschitz-like in the following sense:
Definition 3.3.19. A set-valued mapping F W X Y is epigraphically Lipschitz-
like (ELL) around .x; y/ with modulus L
0 if there are neighborhoods U of x
and V of y such that
EF .x/ \ V EF .u/ C Lkx ukBY for all x; u 2 U;
where BY stands for the closed unit ball of Y .

In other words, F is ELL at .x; z/ if its epigraphical multifunction EF is
Lipschitz-like at that point (Definition 3.3.9). This property is known also as the
Aubin property, the pseudo Lipschitzian property; cf. [431, 499]. It agrees with
the classical local Lipschitzian behavior in the case of single-valued functions,
and reduces to the standard Hausdorff of local Lipschitzian property for set-valued
mappings when V D Y .
For deriving necessary optimality conditions using the approximate subdifferen-
tial by Ioffe (see Definition 3.6.1) we need the following Lipschitz property:
Definition 3.3.20. A function f W X ! Y is said to be strongly compactly
Lipschitzian at x if there exist a multifunction R W X Comp.Y /, where
Comp.Y / is the collection of all nonempty kk-compact subsets of Y , and a function
r W X X ! RC satisfying the following properties:
(i) lim r.x; h/ D 0.
x!x
h!0
(ii) There is > 0 such that for all h 2 BX , x 2 x C BX , and all t 2 .0; /,
one has

t 1 f .x C th/ f .x/ 2 R.h/ C khkr.x; th/BX :
(iii) R.0/ D f0g and R is upper semicontinuous.

The strongly compactly Lipschitzian property agrees with the usual Lipschitzian
continuity when Y is finite dimensional.
3.4 Clarke’s Normal Cone and Subdifferential
In this section we recall the concept of the Clarke subdifferential (see [100])
assuming that X is a Banach space. Suppose that f W X ! R is Lipschitzian
around x 2 X with f .x/ 2 R.
Definition 3.4.1. The Clarke generalized subdifferential of f at x is the set
@C f .x/ WD fx 2 X j 8v 2 X W hx ; vi f 0 .xI v/g; (3.21)
where f 0 .xI v/ is the generalized directional derivative of f at x in the direction

v2X
f .y C tv/ f .y/
f 0 .xI v/ WD lim sup :
y!x; t !C0 t
Definition 3.4.2. The Clarke normal cone to S at x 2 S is described via the

weak closure of the cone spanned on the generalized gradient of the distance
function

NC .xI S / D clw [
0
@C d.xI S /:
Now, assume that f is l.s.c. on X . The Clarke subdifferential of f at x is the set
@C f .x/ WD fx 2 X j .x ; 1/ 2 NC ..x; f .x//I epif /g: (3.22)
Remark 3.4.3. If f is Lipschitzian around x 2 X with f .x/ 2 R the definitions

of Clarke’s subdifferential given by (3.21) and (3.22) are equivalent (see [430,
Section 1.4.3]).
The indicator function of a set S X is the function S W X ! R defined by
S .x/ WD 0 for x 2 S and S .x/ WD C1 for x 2 X n S .
The above normal cone to a set can also be equivalently defined using indicator
function S associated with this set as follows
NC .xI S / D @C S .x/:
The following result is given by Clarke [100, Proposition 2.3.1].

Proposition 3.4.4. Consider f W X ! R. Assume that f is l.s.c. on its domain,
x 2 X with f .x/ 2 R. Then, if f is Lipschitz around x we have
3.5 Limiting Cones and Generalized Differentiability 103
@C .f /.x/ D @C f x/:
Moreover, we introduce the Clarke coderivative for a set-valued mapping F W

X Y acting between Banach spaces X and Y :
Definition 3.4.5. Let F W X Y and .x; y/ 2 graph F . Assuming that F is
closed, the Clarke coderivative DC F .x; y/ of F at .x; y/ is a set-valued map from
Y to X defined by
DC F .x; y/.y / WD fx 2 X j .x ; y / 2 NC ..x; y/I graph F /g:
3.5 Limiting Cones and Generalized Differentiability
We will study generalized differentiability properties in Asplund spaces (see

Definition 3.5.3), a special class of Banach spaces, because it is possible to show
fundamental results like extremal principles (see Sect. 5.5) in these spaces (see [342,
343, 430, 432]). Furthermore, several calculus rules in generalized differentiation
theory presented in this section are working in Asplund spaces. This is important
for deriving optimality conditions based on limiting subdifferentials in Sect. 12.8.
We begin by recalling the following:
Definition 3.5.1. For a set-valued map F W X X acting between a Banach
space X and its dual X we denote the sequential Kuratowski-Painlevé upper
limit with respect to the norm topology on X and weak* topology on X by
Lim sup F .x/;

x!x0
which is given as
w
Lim sup F .x/ WD fx 2 X j 9 sequences xk ! x0 ; and xk ! x ;
x!x0
with xk 2 F .xk /; 8k D 1; 2; : : :g;
w
where ! denotes convergence in the weak-star topology of X .
Let X be a Banach space and ˝ be a subset of X . The Fréchet normal cone to
˝ at x 2 ˝ is defined by
n ˇ hx ; u xi o
ˇ
NO .xI ˝/ WD x 2 X ˇ lim sup 0 ; (3.23)
˝ ku xk
u
!x
˝
where u
! x means u ! x with u 2 ˝. This definition can be extended:
Definition 3.5.2 (Generalized Normals). Let ˝ X be a nonempty subset of a
Banach space X . Consider x 2 ˝ and "
0. Then
n ˇ hx ; u xi o
ˇ
NO " .xI ˝/ WD x 2 X ˇ lim sup " (3.24)
˝ ku xk
u
!x
defines the set of "-normals to ˝ at x.

If x … ˝ we put NO " .xI ˝/ WD ; for all "
0.
In the sequel, we present definitions and properties of the basic generalized
differential constructions held in the Asplund space setting and enjoying a full
calculus (see Mordukhovich [425], [426], [427] and Kruger, Mordukhovich [342]).
Definition 3.5.3. A Banach space X is called Asplund space if every convex
continuous function W U ! R defined on an open convex subset U of X is
Fréchet differentiable on a dense subset of U .
Remark 3.5.4. The class of Asplund spaces is quite broad including every reflexive
Banach space and every Banach space with a separable dual; in particular, c0 and
l p , Lp Œ0; 1 for 1 < p < C1 are Asplund spaces, but l 1 and l 1 are not Asplund
spaces. It has been comprehensively investigated in geometric theory of Banach
spaces, and largely employed in variational analysis; see, e.g., [66, 430, 431].
In the following we suppose that ˝ is a subset of an Asplund space X .
Definition 3.5.5. Consider x 2 ˝. ˝ is said to be closed around x 2 ˝ if there is
a neighborhood U of x such that ˝ \ U is a closed set.
Given x 2 ˝. Assume that ˝ is closed around x 2 ˝.
Definition 3.5.6. The (basic, limiting, Mordukhovich) normal cone to ˝ at x 2
˝ is defined by
N.xI ˝/ WD Lim supx!x NO .xI ˝/

n ˇ o (3.25)
ˇ ˝ w
! x; xk ! x with xk 2 NO .xk ; ˝/ ;
D x 2 X ˇ 9 xk
where Lim sup stands for the sequential Painlevé-Kuratowski outer limit of Fréchet
normal cones to ˝ at x as x tends to x.
If x … ˝ we put N.xI ˝/ WD ;.
Note that, in contrast to (3.24), the basic normal cone (3.25) is often nonconvex
enjoying nevertheless full calculus, and that both the cones (3.25) and (3.23) reduce
to the normal cone of convex analysis when ˝ is convex.
We define the (basic, limiting, Mordukhovich) subdifferential of an extended-
real-valued function through basic normals to its epigraph.
3.5 Limiting Cones and Generalized Differentiability 105
Definition 3.5.7. Consider a function f W X ! R and a point x 0 2 X with

jf .x 0 /j < C1. The (basic, limiting, Mordukhovich) subdifferential of f at
x 0 is defined by
@f .x 0 / WD fx 2 X j .x ; 1/ 2 N..x 0 ; f .x 0 //I epi f /g:
The elements of @f .x 0 / are called (basic, limiting) subgradients of f at x 0 .

We put @f .x 0 / WD ; if jf .x 0 /j D C1.
We recall the notions of coderivatives and subdifferentials for set-valued map-
pings (see [24–27, 426, 427]). Given a set-valued mapping F W X Y with its
graph
˚ ˇ
graph F D .x; y/ 2 X Y ˇ y 2 F .x/
between Asplund spaces X and Y .

Definition 3.5.8. The (basic, normal, Mordukhovich) coderivative mapping
D F .x; y/W Y X of F at .x; y/ 2 graph F is defined by
˚ ˇ
D F .x; y/.y / WD x 2 X ˇ .x ; y / 2 N .x; y/I graph F g; (3.26)
We omit y D f .x/ in (3.26) if F D f W X ! Y is single-valued. If f W X ! Y

happens to be strictly differentiable at x (which is automatic when it is C 1 around
this point), then
˚
D f .x/.y / D rf .x/ y for all y 2 Y :
Furthermore, we introduce the Fréchet coderivative mapping of F at .x; y/ 2

graph F (see Mordukhovich [426], [427] and Bao, Mordukhovich [27]) using the
Fréchet normal cone given by (3.23).
Definition 3.5.9. The Fréchet coderivative mapping DO F .x; y/W Y X of F
at .x; y/ 2 graph F is defined by
˚ ˇ
DO F .x; y/.y / WD x 2 X ˇ .x ; y / 2 NO .x; y/I graph F g; (3.27)
Remark 3.5.10. It is important to mention (see Bao, Mordukhovich [27]) that both
coderivative set-valued mappings D F .x; y/W Y X and DO F .x; y/W Y
X are positively homogeneous functions of y 2 Y reducing to the single-valued
adjoint derivative linear operator
DO f .x/ D frf .x/ y g; D f .x/ D frf .x/ y g for all y 2 Y ;

(3.28)
provided that f is Fréchet differentiable at x for the first equality and strictly
differentiable at this point for the second one, i.e.,
f .x/ f .u/ hrf .x/; u xi

lim D 0;
x;u!x jjx ujj
which is automatic when f 2 C 1 around x.

Using coderivatives of the epigraphical multifunction (see (3.20)), we can
define appropriate extensions of the subdifferential notion from extended-real-
valued function to vector-valued and set-valued mappings with values in partially
ordered spaces. The following notions of subdifferentials for set-valued mappings
F W X Y acting between Asplund spaces X and Y are first introduced in [25]
and then further developed in [24, 26, 27]; cf. the epi-coderivatives for set-valued
mappings in [621] and references therein.
Definition 3.5.11. Given .x; y/ 2 graph F , the (limiting, basic, normal) subdif-
ferential of F at .x; y/ in direction y 2 Y is defined by
@F .x; y/.y / WD D EF .x; y/.y /; (3.29)
where EF is the epigraphical multifunction of F (see (3.20)).

As pointed in [27], we always have the following implication
D EF .x; y/.y / ¤ ; H) y 2 N.0I C /: (3.30)
Consequently, the requirement that y 2 N.0I C / is redundant for the definitions

of the subdifferential for set-valued mappings in [24–27]; in particular, [25,
Definition 2.1]. As usual, when F D f W X ! Y is a single-valued vector-valued
function, we omit y from the notation of subdifferentials. The subdifferential of f
at x in (3.29) becomes
@f .x/.y / D D Ef .x/.y /;
and when F D f W X ! .1; 1 is a lower semicontinuous extended-real-

valued function, the subdifferential (3.29) with ky k D 1 agrees with the (basic,
limiting, Mordukhovich) subdifferential; see [430] and Definition 3.5.7. By [430,
Theorem 3.28], the relationship between the coderivative of a vector-valued function
and the subdifferential of its scalarization
˝ ˛
D f .x/.y / D @hy ; f i.x/ with hy ; f i.x/ WD y ; f .x/ (3.31)
holds provided that f is strictly Lipschitzian at x in the sense of Definition 3.3.5.

Furthermore, we will use the Fréchet subdifferential F at .x; y/ 2 graph F
in Sect. 12.9, Theorem 12.9.1, for deriving necessary conditions for approximate
solutions of a set-valued optimization problem.
Definition 3.5.12. For .x; y/ 2 graph F , the Fréchet subdifferential of F at
.x; y/ in direction y 2 Y is defined by
3.6 Approximate Cones and Generalized Differentiability 107
O .x; y/ WD fx 2 X j x 2 DO EF .x; y/.y /; y 2 N.0I C /; jjy jj D 1g:

@F
(3.32)
Observe that the condition y 2 N.0I C / in (3.32) is not a restriction; in fact, it

automatically follows from the inclusion x 2 DO EF .x; y/.y /.
For the sake of compact exposition, we will not recall all the calculus rules for
normal cones and coderivatives. However, we will refer to them in Mordukhovich’s
books [430, 431].
We need the following calculus rules given by Mordukhovich in [430] for
proving necessary optimality conditions in Sect. 12.8. The next lemma follows from
[430, Theorem 3.36 ] taking into account [430, Theorem 1.26 and Corollary 1.81].
Lemma 3.5.13. Let X be an Asplund space and let x0 2 X . Let fi W X ! R,
i D 1; 2; be proper lower-semicontinuous functions and one of these is Lipschitzian
around x0 . Then one has
@.f1 C f2 /.x0 / @f1 .x0 / C @f2 .x0 /:
The next result is shown in [430, Theorem 3.41 and Corollary 3.43]:
Lemma 3.5.14. Assume that X and Y are Asplund spaces. Let f W X ! Y be
strictly Lipschitzian at x0 and W Y ! R be Lipschitzian around f .x0 /. Then one
has
[
@. ı f /.x0 / @hy ; f i.x0 /:
y 2@.f .x0 //
3.6 Approximate Cones and Generalized Differentiability
Assume that X is a Banach space and f W X ! R [ fC1g is l.s.c. on X and

x 2 X with f .x/ 2 R. Ioffe [275, 277] (compare Mordukhovich [425] for the finite
dimensional case) introduced the approximate subdifferential of f at x:
Definition 3.6.1. The approximate subdifferential of f at x is the set
@A f .x/ WD \L2F Lim sup @

" fyCL .y/;
.";y/!.C0;x/
where F is the collection of all finite dimensional subspaces of X , fyCL .u/ D f .u/
if u 2 y C L and fyCL .u/ D C1 otherwise, and furthermore, for "
0
@
" fyCL .y/ W D fx 2 X j 8v 2 X W hx ; vi
"kvk C lim inf t 1 ŒfyCL .y C tv/ fyCL .y/g:

t !C0
The approximate (Ioffe) normal cone to S X at x 2 S is given by
NA .xI S / D [
>0
@A d.xI S /:
The above normal cone to a set can also be equivalently defined using the
indicator function S associated with this set as follows
NA .x; S / D @A S .x/:
Definition 3.6.2. Let F W X Y , X and Y are Banach spaces and .x; y/ 2

graph F . Assuming that F is closed, the approximate (Ioffe) coderivative
DA F .x; y/ of F at .x; y/ is a set-valued map from Y to X defined by
DA F .x; y/.y / WD fx 2 X j .x ; y / 2 NA ..x; y/I graph F /g:
Among all the properties of the above subdifferentials and coderivative, let us
recall those ones which will be used in the sequel. The following proposition is
shown for Banach spaces X and Y (cf. [228, Proposition 21.2]).
Proposition 3.6.3. Let f and g be functions from X to R [ fC1g which are l.s.c.
on their domains, let F be a set-valued closed map from X to Y . We assume that
x 2 domf or x 2 domF .
(i) @A f .x/ @C f .x/ and these subdifferentials coincide with the subdifferential
of convex analysis (@f .x/ D fx 2 X j 8v 2 X W hx ; v xi f .v/
f .x/g) if f is convex and Lipschitz around x.
(ii) If f attains a local minimum at x, then 0 2 @A f .x/.
(iii) @A .f C g/.x/ @A f .x/ C @A g.x/ for x 2 dom f \ dom g if f is Lipschitz
around x.
(iv) DA F .x; y/.y / DC F .x; y/.y / for .x; y/ 2 graph F and y 2 Y .
Chapter 4
Tangent Cones and Tangent Sets
Tangent cones of first-order and tangent cones and tangent sets of higher-order
play a very important role in set-valued optimization. For instance, derivatives
and epiderivatives of set-valued maps are commonly defined by taking tangent
cones and tangent sets of graphs and epigraphs of set-valued maps. Moreover,
properties of tangent cones and tangent sets are quite decisive in giving calculus
rules for derivatives and epiderivatives of set-valued maps. Furthermore, optimality
conditions in set-valued optimization are also most conveniently expressed by using
tangent cones and tangent sets. Sensitivity analysis, constraints qualifications, and
many other issues in set-valued optimization heavily rely on tangent cones and
tangent sets.
This chapter is a thorough introduction to tangent cones of first-order and tangent
sets and tangent cones of second-order. We collect definitions, properties, and
examples of some of the most commonly used tangent cones of first-order and
tangent sets and tangent cones of second-order. To be specific, we study six of the
most commonly used tangent cones of first-order, six of their modified analogues,
and give numerous equivalent characterizations of some of them. We also study the
notion of the local cone approximation from an abstract point of view. This notion
not only illustrates the key features of the tangent cones, but it also suggests new
ways of generating tangent cones. We investigate convex cones which lie in between
the contingent cone and the Clarke’s tangent cone, which are two of the most popular
tangent cones of first-order. In this discussion, we study recession tangent cones, the
Treiman tangent cone, the prototangent cone, and the quasi-strict tangent cone. We
study six most commonly used second-order tangent sets and a related notion of
second-order generalized tangent sets. We study, in sufficient detail, second-order
asymptotic tangent cones and second-order projective tangent cones and collect
their intrinsic features. Useful connections among various tangent cones and tangent
sets are thoroughly investigated. A brief description of some higher-order tangent
cones and higher-order tangent sets is also given. The results of this chapter will
play an important role in several other chapters of this book.

DOI 10.1007/978-3-642-54265-7__4
110 4 Tangent Cones and Tangent Sets
For simplicity in the presentation, we have chosen to work in the setting of a

normed space or a Banach space. However, almost all the results are either directly
valid in more general spaces or have their analogous statements valid in general
spaces.
Finally, we recall that given a normed space Z and an element z 2 Z, by U.z/,
we denote an open ball centered at z and by U" .z/, we denote the open ball of radius
" centered at z. By BZ (or simply B), we represent the unit ball in the space Z. We
also recall that P WD ft 2 Rj t > 0g, and, if not specified otherwise, the notation
ftn g # 0 means that ftn g P and tn ! 0. Moreover, given a nonempty set S of a
S
normed space Z, the notion x ! xN means that x ! xN with x 2 S .
4.1 First-Order Tangent Cones
A large number of first-order tangent cones have been proposed in the literature.
For instance, Ioffe [276], Penot [463], and Ward [592] present a unified framework
for defining a large number of tangent cones of first-order. An interesting survey
article that collects many tangent cones of first-order is by Palata [449], see also
[140, 307]. However, not all of the known tangent cones of first-order have been
very useful. It turns out the tangent cones with most useful properties are actually
the most commonly used ones. In this section, we study only the most commonly
used tangent cones of first-order. To be specific, we will study the radial tangent
cone, the feasible tangent cone, the contingent cone, the interiorly contingent cone,
the adjacent cone, and the interiorly adjacent cone. We give useful properties of
these six cones and give various characterizations of some of them. We collect
numerous examples from the literature to illustrate these notions. This section
follows the exposition of the excellent papers by Elster and Thierfelder [178] and
Giorgi and Guerraggio [208], and the monograph by Bazarra and Shetty [38], among
others. A few examples are taken from [38]. Some related results can be found in
[131, 254, 255, 338, 362, 363].
4.1.1 The Radial Tangent Cone and the Feasible Tangent Cone
We focus on the following two notions of tangent cones of first-order.

Definition 4.1.1. Let X be a normed space, let S be a subset of X , and let xN 2 X .
1. The radial tangent cone R.S; x/ N of S at xN is the set of all x 2 X for which
there exists a sequence ftn g P with tn # 0 such that xN C tn x 2 S for every
n 2 N.
4.1 First-Order Tangent Cones 111
Fig. 4.1 Feasible tangent

cone in Example 4.1.6
Fig. 4.2 Feasible tangent

cone in Example 4.1.9
N of S at xN is the set of all x 2 X such that

2. The feasible tangent cone IR.S; x/
for every ftn g P with tn # 0, there exists an index m 2 N satisfying that
xN C tn x 2 S for every n
m.
Remark 4.1.2. The radial tangent cone and the feasible tangent cone are the
simplest examples of first-order tangent cones. Clearly these cones are independent
of the topology of X .The feasible tangent cone is also known as the cone of feasible
directions, the cone of admissible directions, the cone of linear tangent vectors,
the cone of inner tangent vectors, the cone of radial interior displacements, the
weak feasible tangent cone, and the radial cone (see [85, 178, 584]).
The following result gives properties of the radial tangent cone and the feasible
tangent cone.
Theorem 4.1.3. Let X be a normed space, let S; S1 ; S2 be subsets of X , and let
N xN 1 ; xN 2 2 X . Then:
x;
N D X nIR.X nS; x/
1. R.S; x/ N
N D X nR.X nS; x/
2. IR.S; x/ N
N R.S; x/
3. IR.S; x/ N
4. xN 2 S ” 0 2 R.S; x/ N ” 0 2 IR.S; x/ N
5. S1 S2 H) R.S1 ; x/ N R.S2 ; x/ N
6. S1 S2 H) IR.S1 ; x/ N IR.S2 ; x/ N
7. R.S1 [ S2 ; x/
N D R.S1 ; x/ N [ R.S2 ; x/ N
8. IR.S1 \ S2 ; x/
N D IR.S1 ; x/ N \ IR.S2 ; x/ N
9. R.S1 S2 ; .xN 1 ; xN 2 // R.S1 ; xN 1 / R.S2 ; xN 2 /
10. IR.S1 S2 ; .xN 1 ; xN 2 // D IR.S1 ; xN 1 / IR.S2 ; xN 2 /.
Proof. A proof can be deduced from the arguments used in the proof of
Theorem 4.2.10. u
t
We collect a few examples of the radial tangent cone and the feasible tangent
cone.
Example 4.1.4. Let S1 D Q and let S2 D fRnQg [ f0g. Clearly S1 \ S2 D f0g and
we obtain that R.S1 \ S2 ; 0/ D f0g, R.S1 ; 0/ D R, and R.S2 ; 0/ D R. This reveals
the fact that the intersection formula R.S1 \ S2 ; x/N D R.S1 ; x/ N \ R.S2 ; x/,
N is, in
general, false.
˚
Example 4.1.5. For S R2 given by S D .x; y/j y D x 3 , the feasible tangent
cone IR.S; x/N of S at xN D .0; 0/ is the cone given by IR.S; x/
N D f.0; 0/g. Note that
if we exclude .0; 0/ from the set S , then the feasible tangent cone is empty.
˚
Example 4.1.6.
For S R2 given by S D .x; y/j 1 x 1; y
x 1=3 ;
y
x , the feasible tangent cone IR.S; x/ N of S at xN D .0; 0/ is given by
N D f.x; y/j x
0; y
xg.
IR.S; x/
Example 4.1.7. For S R2 given by S D f.x; y/j y
jxjg, the feasible tangent
N of S at xN D .0; 0/ is given by IR.S; x/
cone IR.S; x/ N D f.x; y/j y
jxjg.
Example 4.1.8. For S R2 given by S D f.x; y/j y > x 2 g, the feasible tangent
cone IR.S; x/N of S at xN D .0; 0/ is given by IR.S; x/
N D f.x; y/j y > 0g.
˚ S
Example
˚ 4.1.9. For S R given
2
by S D .x; y/j 0 x 2; 0 y 3
.x; y/j .x 2/2 C .y 1/2 1 , the feasible tangent cone IR.S; x/N of S at xN D
N D f.x; y/j x < 0g [ f.0; 0/g.
.3; 1/ is given by IR.S; x/
Remark 4.1.10. Examples 4.1.7 and 4.1.8 reveal an important fact: The feasible
tangent cone, in general, is neither closed nor open. This is also true for the radial
tangent cone.
4.1.2 The Contingent Cone and the Interiorly Contingent Cone
In this subsection, we study the following two notions of tangent cones of first-order.
Definition 4.1.11. Let X be a normed space, let S be a subset of X , and let
xN 2 X .
1. The contingent cone T .S; x/ N of S at xN is the set of all x 2 X for which there
are sequences ftn g P and fxn g X with tn # 0 and xn ! x satisfying that
xN C tn xn 2 S for every n 2 N.
2. The interiorly contingent cone IT.S; x/ N of S at xN is the set of all x 2 X such
that for all sequences ftn g P and fxn g X with tn # 0 and xn ! x, there
exists an integer m 2 N satisfying that xN C tn xn 2 S for every n
m.
The notion of the contingent cone is commonly attributed to G. Bouligand (see
[75]) who coined the phrase contingent to generalize the setting when there is one
and only one tangent half-line at every point of a given set. Independently of [75],
F. Severi (see [527]) also proposed the notion of the contingent cone. Recently,
in an interesting paper [143], S. Dolecki and G. Greco trace the evolution of the
concepts of differentiability, tangency and related notions (such as total differential,
contingent, paratangent cones, etc.) to Giuseppe Peano (see also [49, 142]). The
sequential definition of the interiorly contingent cone, which is also known as the
cone of interior displacements, can be traced back to the monograph by Flett [187,
Section 4.5]. Penot [463] also gives the above sequential definition of the interiorly
contingent cone. The contingent cone has been recognized under numerous names
such as the Bouligand tangent cone, the tangent cone, the cone of adherent
displacements, the outer tangent cone, among others, see [53, 262].
The contingent cone has been used extensively in optimization theory and
variational analysis and numerous characterizations for this notion have been
proposed in the literature. In our next result, we show the equivalence among
some of the most commonly used definitions of the contingent cone. We follow the
scheme of the excellent paper by Giorgi and Guerraggio [208] and the monograph
by Bazarra and Shetty [38], where further details can be found.
The following is an equivalence of different characterizations of the contingent
cone.
Theorem 4.1.12. Let X be a normed space, let S be a subset of X , and let xN 2 cl S .
Then, for i 2 f1; : : : ; 16g, we have T .S; x/
N D Ti .S; x/,
N where Ti .S; x/
N are given
as follows:
N D
T1 .S; x/ ˚ 2 X j 9 ftn g P 9 fxn g S such that xn ! x;
fx N tn .xn x/
N !xg
N D
T2 .S; x/ x 2 X j 9 ftn g P 9 fxn g S such that tn # 0; tn1 .xn x/N !x
N D
T3 .S; x/ fx 2 X j 8 t > 0 9 s 2 .0; t/ 9 xO 2 Ut .0/ such that xN C s.x C x/O 2 Sg
N D
T4 .S; x/ fx 2 X j 8 U.x/ 8 t > 0 9 s 2 .0; t/ 9 xO 2 U.x/ such that xN C s xO 2 S g
N D
T5 .S; x/ ( 2 X j 8 U.x/ 8 t > 0 9 s 2 .0; t/ such that .xN C sU.x// \ S!!
fx ¤ ;g
\ [ \
N D x 2 X j 8 U.x/ 8 U.0/ we have xN C U.0/
T6 .S; x/ tU.x/
t >0
S ¤; ( !
[ \ \
N D x 2 X j 8 U.x/ 8 U.x/
T7 .S; x/ N we have xN C tU.x/ N
U.x/
) t >0
S ¤;
1
\
N D
T8 .S; x/ cl cone .S \ U1=n .x/
N x/
N
\
nD1
N D
T9 .S; x/ cl cone .S \ U.x/
N x/
N
N
U.x/

N
T10 .S; x/ D x 2 X j 9 t > 0 9 fxn g S with xn ¤ x;
N xn ! xN such that

xn xN S
! tx f0g
kxn xkN
N
T11 .S; x/ D x 2 X j 9 fxn g with xN C xn 2 S; xn ! 0; 9 t > 0 such that
[
xn
! tx f0g
kxn k
S xN
N D lim sup
T12 .S; x/
t
t #0
d.xN C tx; S /
N D x 2 X j lim inf
T13 .S; x/ D0
t #0 t
N D
T14 .S; x/ x 2 X j 9 z./ W R ! X with lim z.r/ D x such that 8 t > 0 9
r!0
s 2 .0; t/ with xN C sz.s/ 2 S

N D x 2 X j 9 tn # 0 9 fxn g S such that xn D xN C tn x C tn ".tn /
T15 .S; x/

where lim ".t/ D 0
t !0

xO xN x
N D x 2 X j 8 " > 0 9 xO 2 U" .x/
T16 .S; x/ N \ S such that
kxO xk
[ N kxk
<" f0g.
Proof. We divide the proof into several parts.

1. We begin by showing that T .S; x/ N D T1 .S; x/.N Let x 2 T .S; x/ N be arbitrary.
Then there are sequences ftn g P and fxn g X such that tn # 0, xn ! x and
1
xN C tn xn 2 S . Define sn D and xO n D xN C tn xn . Then fsn g P, fxO n g S
tn
with xO n ! xN and sn .xO n x/
N D xn ! x. Therefore, x 2 T1 .S; x/ N and hence
we have shown that T .S; x/ N T1 .S; x/.
N For the converse, let x 2 T1 .S; x/ N be
arbitrary. Then there are sequences ftn g P and fxn g S such that xn ! xN
1
and tn .xn x/
N ! x. Clearly tn ! 1. By setting sn D and xO n D tn .xn x/, N
tn
we obtain that xO n ! x with xN C sn xO n 2 S . Consequently, x 2 T .S; x/.
N
2. Since the identity T1 .S; x/ N D T2 .S; x/ N is trivial, we have obtained that
N D T2 .S; x/.
T .S; x/ N
3. In view of the fact that
S xN \ [
S xN
lim sup D C U.0/ ;
t #0 t t
U.0/;s>0 t 2.0;s/
N T5 .S; x/
the equivalence among T4 .S; x/, N and T12 .S; x/ N is evident.
4. The equivalence among T10 .S; x/,N T11 .S; x/
N and T16 .S; x/ N is also easy to prove.
N D T9 .S; x/
5. Noticing that the identity T8 .S; x/ N is trivial, we proceed to prove
N D T9 .S; x/.
that T1 .S; x/ N Let x 2 T1 .S; x/ N be arbitrary. Then, there exist
ftn g P and fxn g S such that xn ! xN and tn .xn x/ N ! x. Due to the
convergence xn ! x, N there exists n1 2 N such that xn 2 S \ U.x/ N for every
n > n1 . Consequently, tn .xn x/N 2 cone .S \ U.x/ N x/ N for every n > n1 ,
which ensures that x D limn!1 tn .xn x/ N 2 cl cone.S \TU.x/ N x/.
N Since
N we
this argument is valid for every U.x/, T have T 1 .S; N
x/ U.x/N cl cone.S \
U.x/N x/.
N For the converse, let x 2 U.x/ N cl cone.S \ U. N
x/ N
x/. Given
n 2 N, we choose U1=n .x/. N Then the above containment implies that x is the
limit of elements tm .xm x/,
N where tm > 0 and xm 2 S \ U1=n .x/. N We choose
mn with kx tmn .xmn x/kN < n1 . By varying n, we construct the sequences
ftmn g P, fxmn g S such that tmn .xmn x/N ! x. Therefore, x 2 T1 .S; x/. N
6. We claim that T1 .S; x/ N D T15 .S; x/. N Let x 2 T1 .S; x/. N Then there are
sequences ftn g P, fxn g S such that xn ! xN and tn .xn x/ N ! x. In
other words, x D tn .xn x/N " tn1 , where " W R ! X is a function such that
".t/ ! 0 as t ! 0. Since xn ! xN and lim .tn .xn x// N exists, we deduce that
n!1
tn ! 1. Defining sn D 1=tn , we obtain that xn D xN C sn x C sn ".sn /, ensuring
N T15 .S; x/.
T1 .S; x/ N For the converse, assume that there exists fxn g S such
that xn D xN C sn x C sn ".sn /, which implies that sn1 .xn x/ N ! x and by
setting tn D s1n , we deduce that x 2 T1 .S; x/.
N
7. We next claim that T .S; x/ N T6 .S; x/.N Let x 2 T .S; x/ N be arbitrary.
Then there exists ftn g P and fxn g S such that tn # 0, xn !
x and xN C tn xn 2 S . Since tn # 0, there exists an index m 2 N
such that tn xn 2 U.0/ for every n
m. Furthermore, since xn ! x,
we also have that tn xn 2 tn U.x/ for someTU.x/.
S Therefore, T for suitable
n 2 N, we have N
x C
T S n nt x 2 xN
T C U.0/ t >0 tU.x/ S . In other
words, xN C U.0/ t >0 tU.x/ S ¤ ;, which confirms that N
T .S; x/
N
T6 .S; x/.
8. Let us now prove that T6 .S; x/ N T16 .S; x/.N Let x 2 T6 .S; x/ N be arbitrary.
Then for every U" .0/, we can find 0 ¤ xO 2 U.0/ such that xN C xO D xM 2 S . It
suffices to show that

xM xN x
< ":
kxM xk
N kxk
S
Since xO 2 t >0 tU.x/, for every U.x/ or xO D t.x C z/, with z sufficiently
small. Therefore,

xM xN xO O O
D kxkkxk xkxk D
xkxk C zkxk xkx C sk ;
kxM xk
N kxk kxkkxk
O kx C zkkxkk
which can be made smaller than ".

9. We now claim that T16 .S; xN T13 .S; x/. N Let x 2 T16 .S; x/ N be arbitrary.
Then xQ D tx with t 2 R and kxk Q D 1 also belongs to T16 .S; x/. N
Therefore
there exists a sequence
fz n g S with z
n ! N
x such that lim
n!1
kzn xkN 1 .zn x/N xQ D 0, or limkhk!0 inf khk1 .z x/ N xQ D 0.
Since limkhk!0 inf h1 kxN C khkxQ zk D 0, we obtain that xQ 2 T13 .S; x/, N
implying that x 2 T13 .S; x/. N
10. We next show that T13 .S; x/ N T1 .S; x/. N Let x 2 T13 .S; x/ N be arbitrary. Then
1
for every ftn g P with tn # 0, we have infx2S O h n k xN C hn x xk O < tn .
Therefore, there exists fxO n g S such that h1 n k xN C h n x O
x n k < tn .
From, limn!1 h1 n . N
x C h n x O
x n / D 0, or its equivalent expression
h1
n .xN C hn x xO n / D o.1/, it follows that x D o.1/ C h1 n .x O n x/,
N
and therefore, limn!1 h1 n . O
x n N
x/ D x with O
x n ! N
x. Consequently,
N T1 .S; x/.
T13 .S; x/ N
11. We claim that T15 .S; x/ N T3 .S; x/. N Let x 2 T15 .S; x/ N be arbitrary. Then there
are sequences ftn g P and fxn g such that xN C tn x C tn xn 2 S . Therefore,
for every ı > 0, there are scalars n1 > 0 and n2 > 0 such that for every n
with n
maxfn1 ; n2 g, we have jtn j < ı and kxn k < ı. Consequently, for
nN D maxfn1 ; n2 g, tn D
and xnN D z we obtain xN C
x C
z 2 S with

2 .0; ı/ and kzk < ı: This ensures that x 2 T3 .S; x/. N
12. Next we show that T3 .S; x/ N T4 .S; x/. N Let x 2 T3 .S; x/ N be arbitrary. Let
U .x/ be a neighborhood and let " > 0 be fixed but arbitrary. For ı D minf; "g,
there is a positive number t < ı " and an element z 2 Uı .0/ U .0/ for
which xN C tx C tz D xN C t.x C z/ 2 S . Therefore, x 2 T4 .S; x/. N
13. Our next claim is that T12 .S; x/ N T10 .S; x/. N Let x 2 T12 .S; x/ N be arbitrary.
We consider a sequence of neighborhoods of 0 and a sequence f
n g such that

n ! 0: We construct sequences fxn g S , ftn g ! 0, and fzn g ! 0 such
that x D tn1 .xn x/ N C zn , implying that tn1 .xn x/ N D x zn ! x, which
proves that x 2 T10 .S; x/. N
14. We claim that T11 .S; x/ N T14 .S; x/. N Let x 2 T11 .S; x/ N be arbitrary. Then
there exists fxn g ! 0 for which xn D ˛xkxn k C o.kxn k/. By hypothesis, we
have xN C xn 2 S and hence xN C ˛xkxn k C o.kxn k/ 2 S . Then, for every
> 0
and t 2 .0;
/ t D ˛kxn k with n large enough, for which xN C t.t/ 2 S with
.t/ given by Œx C o.1/.
15. Finally, the inclusion T11 .S; x/ N T .S; x/ N is easy to prove. t
u
Remark 4.1.13. The above characterizations have their own advantages. For
instance, it follows immediately from the definition of T1 .S; x/ N that if xN is an
N D f0g. In fact, in this case, as required by T1 .S; x/,
isolated point then T1 .S; x/ N
xn D xN is the only sequence in S converging to x.
N Furthermore, the cone T5 .S; x/N
reduces to the empty-set when xN is an isolated point of S . Therefore, to be

able to provide a reasonable comparison, following the approach of [207], the
element f0g has been added to this cone. We remark that T1 .S; x/ N was used in
[1, 34, 36, 37, 262], among others. Characterizations T2 .S; x/N and T15 .S; x/N were
proposed by Bazaraa, Goode, and Nashed [36]. Characterizations T8 .S; x/ N (or
N is due to Varaiya [582]. The monographs [254] and [255] by M. R.
T9 .S; x//
N R. T. Rockafellar in [491] used T .S; x/
Hestenes contains T10 .S; x/. N and T5 .S; x/.
N
Definition of T11 .S; x/N is given in Canon et. al [83]. Hiriart-Urruty [262] also
N Federer [182] contain T16 .S; x/.
uses T6 .S; x/. N A more detailed remark on the use
of these characterizations can be found in [207]. Some other works dealing with
various aspects of contingent cones are [58, 59, 64, 113, 119, 131, 192, 261, 308, 309,
413, 435, 436, 453, 463–466, 468–470, 501, 519, 571, 577–579, 592].
The following result from [178] gives characterizations of the interiorly contin-
gent cone.
Then, for i 2 f1; 2g, we have IT.S; x/
N D IT i .S; x/,
N where IT i .S; x/
N are given as
follows:
N D fx 2 X j 9 U.x/ 9 t > 0 8 s 2 .0; t/ 8 z 2 U.x/ we have xN C sz 2 S g
IT 1 .S; x/
( ! )
[ \
N D x 2 X j 9 U.x/
IT 2 .S; x/ N 9 U.x/ such that xN C tU.x/ U.x/N S .
t >0
Proof. The equivalence between IT 1 .S; x/ N and IT.S; x/

N is given in Flett [187,
Section 4.5] and is easy to show. Here we will only prove the equivalence
N and IT 2 .S; x/.
between IT 1 .S; x/ N Let x 2 IT 1 .S; x/
N be Sarbitrary. Then there are
a neighborhood U.x/ and a scalar t > 0 such that xN C s2.0;tT / sU.x/ S . We
can find neighborhoods U.x/ N Sand V .x/ TU.x/ such that U.
S N
x/ .xN C tV .x// D
; which implies that xN C s>0 sV .x/ N xN C s2.0;t / sU.x/ and, in
U.x/
S T
particular, xN C s>0 sV .x/ N S , confirming that x 2 IT 2 .S; x/.
U.x/ N We
N IT 2 .S; x/.
have shown that IT 1 .S; x/ N
For the converse,
S x 2 IT 2 .S;
let N Then there are neighborhoods U.x/ and
Tx/.
N satisfying xN C t >0 tU.x/
U.x/ N S . Since there exist a scalar t > 0
U.x/
and a neighborhood V .x/ U.x/ such that for every s 2 .0; t/, we have xN C
sV .x/ U.x/,N and consequently, for every s 2 .0; t/, we have xN C sV .x/ S ,
which ensures that x 2 IT 1 .S; x/.
N The proof is complete. t
u
We next collect some properties of the contingent cone and the interiorly
contingent cone.
N xN 1 , xN 2 2 X . Then:
x,
N D X nT .X nS; x/
1. IT.S; x/ N
N D X nIT.X nS; x/
2. T .S; x/ N
N is an open cone
3. IT.S; x/
4. N is a closed cone
T .S; x/
5. N D T .S; x/
T .cl S; x/ N
6. N D IT.S; x/
IT.int S; x/ N
7. xN 2 cl S ” 0 2 T .S; x/ N
8. x 2 int S ” 0 2 IT.S; x/ N
9. IT.S; x/N IR.S; x/ N R.S; x/ N T .S; x/ N
10. S1 S2 H) IT.S1 ; x/ N IT.S2 ; x/ N
11. S1 S2 H) T .S1 ; x/ N T .S2 ; x/ N
12. IT.S1 \ S2 ; x/N D IT.S1 ; x/ N \ IT.S2 ; x/ N
13. T .S1 [ S2 ; x/
N D T .S1 ; x/ N [ T .S2 ; x/ N
14. IT.S1 S2 ; .xN 1 ; xN 2 // D IT.S1 ; xN 1 / IT.S2 ; xN 2 /
15. T .S1 S2 ; .xN 1 ; xN 2 // T .S1 ; xN 1 / T .S2 ; xN 2 /.
Proof. The proof can be derived from the arguments supplied in the proof of
Theorem 4.2.10. u
t
We now give a few examples of the contingent cone and the interiorly contingent
cone.
N for
Fig. 4.3 T .S; x/
Example 4.1.16
N for
Fig. 4.4 T .S; x/
Example 4.1.17
Example 4.1.16. For S R2 given by S D f.x; y/ 2 j y

x 3 g, the contingent
N of S at xN D .0; 0/ is given by T .S; x/
cone T .S; x/ N D f.x; y/j y
0g.
Example 4.1.17. For S R2 given by S D f.x; y/j x 2 C jxj C y D 0g, the
N of S at xN D .0; 0/ is given by T .S; .0; 0// D f.x; y/j y C
contingent cone T .S; x/
jxj D 0g.
Example 4.1.18. For the sets S1 ; S2 R given by
S1 WD fx 2 Rj x D 22n ; n 2 Ng [ f0g;
S2 WD fx 2 Rj x D 22nC1 ; n 2 Ng [ f0g;
we have T .S1 ; 0/ D T .S2 ; 0/ D RC , whereas, .1; 1/ … T .S1 S2 ; .0; 0//, which

shows that the inclusion T .S1 S2 ; .xN 1 ; xN 2 // T .S1 ; xN 1 /T .S2 ; xN 2 /, is, in general,
false (see [60]).
Example 4.1.19. For the sets S1 ; S2 R2 given by
˚
S1 D .x; y/j y
x 2 ;
˚
S2 D .x; y/j y x 2 ;
we have
T .S1 \ S2 ; .0; 0// D f.0; 0/g;

T .S1 ; .0; 0// \ T .S2 ; .0; 0// D f.x; y/ 2 R2 j y D 0g;
showing that the intersection formula T .S1 \ S2 ; x/

N D T .S1 ; x/
N \ T2 .S; x/,
N is, in
general, false.
The following definition gives a useful way of extending the scope of tangent
cones:
Definition 4.1.20. Let X be a reflexive Banach space, let S X , and let xN 2 S .
The weak contingent cone TW .S; x/ N of S at xN is the collection of all x 2 S such
that there are a sequence fxn g S with xn ! xN and a sequence ftn g P such that
tn .xn x/
N converges weakly to x.
Remark 4.1.21. The notion of weak contingent cone was proposed by Bazaraa,
Goode, Nashed and Shetty [37]. Borwein [59] conducted a thorough study of an
extension of the above notion where he employed a general topology that subsumed
the weak as well as the weak topology.
Fig. 4.5 Failure of the

intersection formula
T .S1 \ S2 ; x/
N D
N \ T2 .S; x/
T .S1 ; x/ N in
Example 4.1.19
4.1.3 The Adjacent Cone and the Interiorly Adjacent Cone
This subsection deals with the following two notions of tangent cones.
xN 2 S .
1. The adjacent cone A.S; x/ N of S at xN is the set of all x 2 X such that for every
sequence ftn g P with tn # 0, there exists fxn g X with xn ! x satisfying
xN C tn xn 2 S for every n 2 N.
2. The interiorly adjacent cone IA.S; x/N of S at xN is the set of all x 2 X for which
there exists ftn g P with tn # 0 such that for every fxn g X with xn ! x,
there exists an index m 2 N satisfying xN C tn xn 2 S for every n
m.
Remark 4.1.23. The adjacent cone has also been referred to as derivable cone,
intermediate cone, cone of attainable directions, Nagumo cone, and Ursescu
cone, among others. The interiorly adjacent cone is also known as the cone of quasi-
interior directions, see [268].
The following result gives some of the most commonly used definitions of the
adjacent cone:
For i 2 f1; : : : ; 5g, we have A.S; x/ N where Ai .S; x/
N D Ai .S; x/, N are given as
follows:

N D x 2 X j 9 " W RC ! X; lim ".t/ D 0; 9 t > 0 such that xN C sxC
A1 .S; x/
t !0

s".s/ 2 S; 8 s 2 .0; t/
˚
N D x 2 X j 9 W RC ! X; 9 t > 0 such that .0/ D x;
A2 .S; x/ N 0 .0/ D x;
.s/ 2 S; 8 s 2 .0; t/g
N D fx 2 Xj 8 U.x/ 9 t > 0 8 s 2 .0; t/ 9 xO 2 U.x/ such that xN C s xO 2 S g
A3 .S; x/

d.xN C tx; S /
N D x 2 X j lim
A4 .S; x/ D0
t !0 t
S xN
N D lim inf
A5 .S; x/ .
t #0 t
Proof. The proof, which is based on [38, 207], is divided into several parts.
1. We begin by showing that A1 .S; x/ N D A2 .S; x/.
N Let x 2 A2 .S; x/N be arbitrary.
Then from the expression 0 .0/ D x, we obtain that lims#0 s 1 ..s/ x/
N D x,
which implies that .s/ xN D sx C s".s/, where ".s/ ! 0 as s ! 0. In other
words, .s/ D xN C sx C s".s/ with .s/ 2 S for every s 2 .0; t, with suitable
t. Therefore x 2 A1 .S; x/.
N For the converse, assume that x 2 A1 .S; x/.
N Then we
have .s/ D xN C sx C s".s/, for the defined s and ". Therefore, xN D .0/ and
lims!0 s 1 ..s/ x/ N D x C lims!0 ".s/ D x, and consequently x 2 A2 .S; x/.N
2. The identity A.S; x/N D A2 .S; x/
N is immediate.
3. The proofs for A1 .S; x/ N A3 .S; x/
N and A3 .S; x/
N A4 .S; x/N follow from the
definition.
4. We next prove that A4 .S; x/ N A2 .S; x/. N For x 2 A4 .S; x/, N we have
limh!0C h1 d.xN C hx; S / D 0. Consequently, for every " > 0 there exists
ı > 0 such that h1 .d.xN C hx; S // < " for every h < ı. Therefore, for every
h 2 .0; ı/, there is an element y 2 S such that h1 .xN C hx y.h// < " or
equivalently limh!0C h1 .xN C hx y.h// D 0. It follows that xN C hx C o.h/ D
y.h/ 2 S . By taking .h/ D xN C hx C o.h/ continuously extended to Œ0; ı/, we
get that x 2 A2 .S; x/.
N t
u
The characterizations given above have been known for quite some time.
Recently, in an interesting paper, Jiménez and Novo [306] proposed two new
characterizations of the adjacent cone by using the notion of slowly convergent
sequences. In the following, by ˝ we represent the set of all strictly decreasing
sequences ftn g of positive real numbers converging to 0 such that tn1tn
! 1. In
[306], the following two characterizations of the adjacent cone were proposed:
For i 2 f6; 7g, we have A.S; x/
N D Ai .S; x/,
N where Ai .S; x/
N are given as follows:
N WD fx 2 X j 9 ftn g ˝; 9 fxn g ! 0; such that xN C tn x C tn xn 2 S;

A6 .S; x/
for every n 2 Ng
N WD fx 2 X j 9 ftn gg ˝; 9 fxn g ! x; such that xN C tn xn 2 S;
A7 .S; x/
for every n 2 Ng:
Proof. See Jiménez and Novo [306]. t

u
We note that the interiorly adjacent cone can equivalently be expressed as
follows:
N D fx 2 X j 9 U.x/ 8 t > 0 9 s 2 .0; t/ 8 xO 2 U.x/; we have xN C s xO 2 S g :

IA.S; x/
The following theorem collects some of the main features of the adjacent cone
and the interiorly adjacent cone.
x,
1. N D X nIA.X nS; x/
A.S; x/ N
2. N D X nA.X nS; x/
IA.S; x/ N
A.S; x/
4. N is an open cone
IA.S; x/
5. N D A.S; x/
A.cl S; x/ N
6. IA.int S; x/ D IA.S; x/ N
7. x 2 cl S ” 0 2 A.S; x/ N
8. x 2 int S ” 0 2 IA.S; x/ N
9. N A.S; x/
IR.S; x/ N T .S; x/ N
10. N IA.S; x/
IT.S; x/ N R.S; x/ N
11. S1 S2 H) A.S1 ; x/ N A.S2 ; x/ N
12. S1 S2 H) IA.S1 ; x/ N IA.S2 ; x/ N
13. A.S1 S2 ; .xN 1 ; xN 2 // D A.S1 ; xN 1 / A.S2 ; xN 2 /
14. A.S1 S2 ; .xN 1 ; xN 2 // D A.S1 ; xN 1 / A.S2 ; xN 2 /
15. IA.S1 S2 ; .xN 1 ; xN 2 // IA.S1 ; xN 1 / IA.S2 ; xN 2 /.
Proof. The proof can be derived from the proof of Theorem 4.2.10. t
u
We now give a few examples of the adjacent cone and the interior adjacent cone.
˚
Example 4.1.27. For S R2 given by S D .x; y/j y D x 3 , the adjacent cone
N of S at xN D .0; 0/ is given by A.S; x/
A.S; x/ N D f.x; y/j y D 0g.
˚ S˚
S R given by S D .x; y/j x
0; y
x
2 3
Example 4.1.28. For .x; y/j
x 0; y
x , the adjacent cone A.S; x/ N of S at xN D .0; 0/ is given by
N D f.x; y/j y
maxf0; xgg.
A.S; x/
N for
Fig. 4.6 A.S; x/
Example 4.1.28
4.2 Modified First-Order Tangent Cones 123
N for
Fig. 4.7 A.S; x/
Example 4.1.30
˚
Example 4.1.29. For S R2 given by S D .x; y/j y D x sin.x 1 /; x ¤ 0 [
f.0; 0/g, the adjacent cone A.S; x/
N of S at xN D .0; 0/ is given by A.S; x/
N D
f.x; y/j jyj jxjg.
Example 4.1.30. For S R2 defined by

˚ [ 2 1
S D .x; y/j x
0; y
x 2
.x; y/j x 0; x y x ;
3 3
N of S at xN D .0; 0/ is given by
the adjacent cone A.S; x/
[ 2 1

N D f.x; y/j x
0; y
0g
A.S; x/ .x; y/j x 0; x y x :
3 3
4.2 Modified First-Order Tangent Cones
All the first-order tangent cones discussed so far in this chapter are, in general, non-
convex. A quite easy fix, as done in some of the earlier works, where convexity of
the tangent cone was required, is to consider the so-called pseudo tangent cone
˘.S; x/N by taking the convex hull of the contingent cone T .S; x/N of S at x:
N
N D conv .T .S; x//

˘.S; x/ N :
It turns out that the convexity achieved through the above procedure of taking the
convex hull of the contingent cone is not optimal, as the pseudo tangent cone is not
easily analytically tractable. Furthermore, it often gives too large an approximation
of the underlying set. In the following, we discuss another procedure of getting
convex approximations of sets by considering the modified analogues of the cones
studied above. These cones are always convex. The starting point of this idea was the
influential work of F. H. Clarke who introduced the notion of the so-called Clarke’s
cone (see Clarke [100]). As we will see below, the Clarke’s cone is a modified
version of the adjacent cone. The foregoing discussion is based on the excellent
paper [178].
4.2.1 The Modified Radial and the Modified Feasible Tangent

Cones
In this section, we deal with the following two modified tangent cones of first-order.
1. The modified radial tangent cone RM .S; x/ N of S at xN is the collection of all
x 2 X such that for every neighborhood V .x/ N and for every scalar t > 0, there
exist a scalar s 2 .0; t/ and an element xO 2 V .x/ N \ S [ fxg N satisfying that
xO C sx 2 S .
2. The modified feasible tangent cone IRM .S; x/ N of S at xN is the collection of all
x 2 X for which there exist a neighborhood V .x/ N and a scalar t > 0 such that
for every s 2 .0; t/ and for every xO 2 V .x/
N \ S [ fxg,
N we have xO C sx 2 S .
N is referred to as the modified cone of feasible
Remark 4.2.2. In [178], IRM .S; x/
directions. This cone also appears in the works of R. T. Rockafellar (see [492,493]).
The following result collects some of the useful features of the modified radial
tangent cone and the modified feasible tangent cone.
N
Theorem 4.2.3. Let X be a normed space, let S; S1 ; S2 be subsets of X , and let x,
xN 1 , xN 2 2 X . Then:
1. N D X nIRM .X nS; x/
RM .S; x/ N
2. N D X nRM .X nS; x/
IRM .S; x/ N
3. N is a convex cone
IRM .S; x/
4. N is the complement of a convex cone
RM .S; x/
5. N IR.S; x/
IRM .S; x/ N R.S; x/ N RM .S; x/ N
6. xN 2 S ” 0 2 IRM .S; x/ N ” 0 2 RM .S; x/ N
7. IRM .S1 \ S2 ; x/
N IRM .S1 ; x/ N \ IRM .S2 ; x/ N
8. RM .S1 [ S2 ; x/
N RM .S1 ; x/ N [ R.S2 ; x/ N
9. IRM .S1 S2 ; .xN 1 ; xN 2 // D IRM .S1 ; xN 1 / IRM .S2 ; xN 2 /
10. RM .S1 S2 ; .xN 1 ; xN 2 // RM .S1 ; xN 1 / RM .S2 ; xN 2 /.
u
4.2.2 The Modified Contingent and the Modified Interiorly

Contingent Cones
Here we study the following two modified tangent cones of first-order.


1. The modified contingent cone TM .S; x/ N of S at xN is the collection of all x 2 X
such that for every neighborhood V .x/, N every neighborhood V .x/, and every
scalar t > 0, there are a scalar s 2 .0; t/, an element xO 2 V .x/
N \ S [ fxgN and an
element xM 2 V .x/ so that we have xO C s xM 2 S .
2. The modified interiorly contingent cone IT M .S; x/ N of S at xN is the collection
of all x 2 X for which there are a neighborhood V .x/, N a neighborhood V .x/,
and a scalar t > 0 such that for every s 2 .0; t/ for every xM 2 V .x/ and for every
xO 2 V .x/
N \ S [ fxg,N we have xO C s xM 2 S .
N is referred to as the modified cone of interior
Remark 4.2.5. In [178], IT M .S; x/
displacements. The cone IT M .S; x/N is also known as the cone of epi Lipschitzian
directions (see [492, 493]), and the hypertangent cone (see [524]).
The following sequential definition of the hypertangent cone has been used quite
often:
S
N D fx 2 X j 8 fxn g ! x;
IT M .S; x/ N 8 ftn g # 0; 8 fzn g ! x; we have
xn C tn zn 2 S; for large ng: (4.1)
The following result collects some of the useful features of the modified
contingent cone and of the modified interiorly contingent cone.
N
Theorem 4.2.6. Let X be a normed space, let S; S1 ; S2 be subsets of X , and let x,
xN 1 , xN 2 2 X . Then:
1. N D X nTM .X nS; x/
IT M .S; x/ N
2. N D X nIT M .X nS; x/
TM .S; x/ N
TM .S; x/
4. N is an open cone
IT M .S; x/
5. N is a convex cone
IT M .S; x/
6. N is the complement of a convex cone
TM .S; x/
7. N IRM .S; x/
IT M .S; x/ N IR.S; x/ N R.S; x/ N RM .S; x/ N TM .S; x/
N
8. N IT.S; x/
IT M .S; x/ N IR.S; x/ N R.S; x/ N T .S; x/ N TM .S; x/
N
9. xN 2 cl S ” 0 2 TM .S; x/ N
10. x 2 int S ” 0 2 IT M .S; x/ N
11. N IT M .S; x/
IT M .int S; x/ N
12. N TM .S; x/
TM .cl S; x/ N
13. IT M .S1 \ S2 ; x/
N IT M .S1 ; x/ N \ IT M .S2 ; x/N
14. TM .S1 [ S2 ; x/
N TM .S1 ; x/ N [ TM .S2 ; x/ N
15. IT M .S1 S2 ; .xN 1 ; xN 2 // D IT M .S1 ; xN 1 / IT M .S2 ; xN 2 /
16. TM .S1 S2 ; .xN 1 ; xN 2 // TM .S1 ; xN 1 / TM .S2 ; xN 2 /.
u
4.2.3 The Modified Adjacent and the Modified Interiorly

Adjacent Cones
Now we investigate the following modified tangent cones of first-order.

1. The modified adjacent cone AM .S; x/ N of S at xN is the collection of all x 2 X
such that for every neighborhood V .x/, there are a neighborhood V .x/ N and a
scalar t > 0 satisfying that for every scalar s 2 .0; t/ and for every element
xO 2 V .x/
N \ S [ fxg,
N there exists xM 2 V .x/ with xO C s xM 2 S .
2. The modified interiorly adjacent cone IAM .S; x/ N of S at xN is the collection
of all x 2 X for which there exists a neighborhood V .x/ so that for every
neighborhood V .x/N and every scalar t > 0, there are a scalar s 2 .0; t/ and an
element xO 2 V .x/
N \ S [ fxg
N such that for every xM 2 V .x/, we have xO C s xM 2 S .
The following result gives some characterizations of the above notion.
For i 2 f1; : : : ; 7g, we have AM .S; x/N D Ci .S; x/, N where Ci .S; x/
N are given as
follows:
( )
d.xO C tx; S / d.x;
O S/
N D x 2 X j lim sup
C1 .S; x/ D0
O x;
x! N t #0 t
N D fx 2 X j 8 fxn g ! x;
C2 .S; x/ N with fxn g cl S; 8 ftn g # 0; 9 fzn g ! x
such that xn C tn zn 2 S g

d.xN C h C
x/; S
N D x 2 X j lim lim inf
C3 .S; x/
t !0
!0
\ [ \
S xN
N D
C4 .S; x/ U.x/
N
t
N x2S \U.x/;
U.0/
#0; U.x/ N t 2.0;
/
S xN
C5 .S; xN D lim inf
S
t #0; x !xN
t
N D fx 2 X j 8 " > 0; 9 ˛; ˇ > 0; 8 y 2 S \ U˛ .x/;
C6 .S; x/ N 8h 2 .0; ˇ/; 9 x 2
U" .x/ with y C hx 2 S g
( )
d.xO C
x; S /
N D x 2 X j lim
C7 .S; x/ D0 .
S

#0; xO !xN

Proof. A proof of this result can be found in the excellent paper Giorgi and
Guerraggio [208]. t
u
Remark 4.2.9. Definition 4.2.7 was proposed by [178]. However, Theorem 4.2.8
shows that when xN 2 cl S , AM .S; x/
N coincides with the celebrated Clarke’s tangent
cone proposed by F. H. Clarke and used extensively in optimization and variational
analysis. In fact, Giorgi and Guerraggio [208] show that if xN 2 cl S , then AM .S; x/
N
N are the same, however, if xN … cl S , then they are two different cones.
and C6 .S; x/
In fact, if xN … cl .S /, then C6 .S; x/

N coincides with X whereas AM .S; x/
N D ;.
N which was originally introduced by F. H. Clarke in a finite-
The notion C1 .S; x/,
dimensional setting, was adopted in general spaces by Hiriart-Urruty [262], who
N and proved its equivalence with C1 .S; x/.
also proposed C2 .S; x/ N
N and IAM .S; x/
In the following AM .S; x/ N will be denoted by C.S; x/ N and
N respectively.
IC.S; x/,
We give some features of the Clarke’s tangent cone and the modified interiorly
adjacent cone.
x,
1. IC.S; x/N D X nC.X nS; x/ N
2. C.S; x/N D X nIC.X nS; x/ N
3. C.S; x/N is a closed cone
4. IC.S; x/N is an open cone
5. C.S; x/N is a convex cone
6. IC.S; x/N is the complement of a convex cone
7. xN 2 cl S ” 0 2 C.S; x/ N
8. xN 2 int S ” 0 2 IC.S; x/ N
9. N D IC.S; x/
IC.int S; x/ N
10. C.cl S; x/N D C.S; x/ N
11. N C.S; x/
IRM .S; x/ N A.S; x/N
12. IA.S; x/N IC.S; x/ N RM .S; x/N
13. C.S1 S2 ; .xN 1 ; xN 2 // D C.S1 ; xN 1 / C.S2 ; xN 2 /
14. IC.S1 S2 ; .xN 1 ; xN 2 // IC.S1 ; xN 1 / IC.S2 ; xN 2 /.
Proof. The assertions (1) through (4) follow directly from the definition. We will
N is a cone, to prove that it is convex, it suffices to show that
prove (5). Since C.S; x/
for arbitrarily chosen elements x1 ; x2 2 C.S; x/,
N we have x1 C x2 2 C.S; x/. N Let
U.x1 C x2 / be an arbitrary neighborhood of x1 C x2 . We will show that there are a
neighborhood V .x/ N and a scalar t > 0 such that for every element xO 2 V .x/ N \S [
fxg
N and for every scalar s 2 .0; t/, there exists xM 2 U.x1 C x2 / with xO C s xM 2 S , or
equivalently,
S xO
\ U.x1 C x2 / ¤ ;: (4.2)
s
Let U.x1 / and U.x2 / be neighborhoods of x1 and x2 such that
U.x1 / C U.x2 / U.x1 C x2 /: (4.3)
Since x1 2 C.S; x/,

N there are a neighborhood V1 .x/
N and a scalar t1 > 0 such that
S xO 1
\ U.x1 / ¤ ; for every xO 1 2 V1 .x/
N \ S [ fxg
N for every s1 2 .0; t1 /:
s1
(4.4)
Analogously, since x2 2 C.S; x/,

N there are a neighborhood V2 .x/
N and a scalar
t2 > 0 such that
S xO 2
\ U.x2 / ¤ ; for every xO 2 2 V2 .x/
N \ S [ fxg
N for every s2 2 .0; t2 /:
s2
(4.5)
N and a scalar t > 0 such that
We choose a neighborhood V .x/
V .x/ N
N V1 .x/; (4.6a)
t minft1 ; t2 g; (4.6b)
N C sU.x1 / V2 .x/
V .x/ N for every s 2 .0; t/: (4.6c)
Then, for every xO 2 V .x/

N \ S [ fxgN and for every s 2 .0; t/, there exists a vector
xQ 1 2 U.x1 / such that xO C s xQ 1 2 S and xO C s xQ 1 2 V2 .x/. Using (4.5), we confirm
that there exists a vector xQ 2 2 U.x2 / such that xO C s xQ 1 C s xQ 2 2 S . This, in view
of (4.3), implies that xQ 1 C xQ 2 2 U.x1 C x2 /, confirming that (4.2) holds. Therefore,
x1 C x2 2 C.S; x/. N This establishes the convexity of the cone C.S; x/. N
Having proved (5), the assertion (6) is a direct consequence.
We next prove (7). Let xN 2 cl S be arbitrary. Then, for every t > 0, we have
S \ .xN C tU.0// ¤ ;; (4.7)
where U.0/ is arbitrary. We set t D 1 and V .x/ N D xN C U.0/, then pick xO 2

N \ S [ fxg
V .x/ N and s 2 .0; t/.
N then by (4.7), we have s 1 .S x/
Note that if xO D x, N \ U.0/ ¤ ;. On the other
O then xO 2 S and we have 0 2 s 1 .S x/
hand, if xN ¤ x, N \ U.0/ ¤ ;, confirming
that 0 2 C.S; x/.
N Therefore, we have shown that xN 2 cl S implies that 0 2 C.S; x/.N
For the converse, we assume that 0 2 C.S; x/. N Then, for every U.0/, there exists a
scalar s > 0 such that s 1 .S x/
O \ U.0/ ¤ ;, which confirms that x 2 cl S .
A proof of (8) is a consequence of the duality relationship given in (1) and (2).
The proof of (9) and (10) are similar, so we will only prove (10). Let x 2
C.cl S; x/N be arbitrary. Then for each neighborhood U.x/ there are a neighborhood
N and a scalar t > 0 such that
V .x/
.xO C sU.x// \ cl S ¤ ; for every xO 2 V .x/

N \ cl S [ fxg;
N for every s 2 .0; t/:
Since the set U.x/ is open, the above expression implies that .xO C sU.x// \S ¤
; for every xO 2 V .x/ N \ S [ fxg,N and for every s 2 .0; t/, confirming that
x 2 C.S; x/.N Therefore, we have shown that C.cl S; x/ N C.S; x/. N To prove
the converse inclusion, we pick an arbitrary x 2 C.S; x/. N Let U.x/ be a given
neighborhood of x. Then, for U1 .x/ D 21 .U.x/x/Cx, there are a neighborhood
N and a scalar t > 0 such that xO C sU1 .x/ \ S ¤ ; for every xO 2 V .x/
V .x/ N \ [fxg
N
and for every s 2 .0; t/.
4.3 Miscellaneous Properties of First-Order Tangent Cones 129
We choose xO 2 V .x/ N \ cl S [ fxgN and s 2 .0; t/. Then, for each U.0/, there is an
element xM such that xM 2 .U.0/ C x/ O \ .V .x/ N \ S [ fxg/
N and .xM C sU.x// \ S ¤ ;,
which implies that .xO CsU1 .x/CU.0//\S ¤ ; for every U.0/, and .xO CsU.x//\
cl S cl .xO C sU1 .x// \ cl S ¤ ;, confirming that x 2 C.cl S; x/. N
Properties (11) and (12) follow from the definitions of the tangent cones involved.
We next prove (13). Let .x1 ; x2 / 2 C.S1 S2 ; .xN 1 ; xN 2 // be arbitrary. Therefore,
for every neighborhood U.x1 / and U.x2 /, there are neighborhoods V .xN 1 / and V .xN 2 /
and a scalar t > 0 satisfying that for every scalar s 2 .0; t/ and for every element
.xO 1 ; xO 2 / 2 V .xN 1 / V .xN 2 / \ S [ f.xN 1 ; xN 2 /g there exist xM 1 2 U.x1 / and xM 2 2 U.x2 /
with .xO 1 ; xO 2 / C s.xM 1 ; xM 2 / 2 S1 S2 . In other words, for every neighborhood U.x1 /
and U.x2 / there are neighborhoods V .xN 1 / and V .xN 2 / and a scalar t > 0 satisfying
that, for every scalar s 2 .0; t/ and for every element xO 1 2 V1 .xN 1 /\S1 [ xN 1 and xO 2 2
V2 .xN 2 /\S2 [ xN 2 there exist xM 1 2 U.x1 / and xM 2 2 U.x2 / satisfying that xO 1 Cs xM 1 2 S1
and xO 2 C s xM 2 2 S2 , confirming that .x1 ; x2 / 2 C.S1 ; xN 1 / C.S2 ; xN 2 /. This proves
that C.S1 S2 ; .xN 1 ; xN 2 // C.S1 ; xN 1 / C.S2 ; xN 2 /. The converse inclusion follows
by reversing the arguments.
Finally, analogous arguments lead to a proof of property (14). t
u
In the following, we give a few simple examples of the Clarke tangent cone.
Example 4.2.11. For S R2 given by S D f.x; y/j x 2 CjxjCy D 0g, the Clarke’s
tangent cone of S at .0; 0/ is given by C.S; .0; 0// D f.0; 0/g.
Remark 4.2.12. The above example depicts a situation in which the Clarke’s
tangent cone is trivially convex and the contingent cone is nonconvex (See
Example 4.1.17).
Example 4.2.13. For S R2 given by S D f.x; y/j y log.jxj C 1/ 0g,
the Clarke’s tangent cone of S at .0; 0/ is given by C.S; .0; 0// D f.x; y/j y C
jxj 0g.
The convexity of the Clarke’s tangent cone is obtained by varying the direction x.
Unfortunately, as a consequence, the isotone property is lost, as shown below:
Example
˚ 4.2.14. For X D R2 , we ˚define two sets S1 and S2 in X S by S1 D
.x; y/ 2 R2 j y
x 2 and S2 D .x; y/ 2 R2 j y x 2 sin x1
0 f.0; 0/g.
Then S1 S2 , and C.S1 ; .0; 0// D R RC , and C.S2 ; .0; 0// D f.x; y/j y
jxjg,
showing that C.S1 ; .0; 0// 6 C.S2 ; .0; 0//.
4.3 Miscellaneous Properties of First-Order Tangent Cones
In this section, we give some relationships among the contingent cone, the adjacent
cone, and the Clarke’s cone which are among the most commonly used tangent
cones. We also give some formulae based on a combination of different cones, and
give a few additional examples.
Fig. 4.8 The Clarke’s

tangent cone in
Example 4.2.13
We begin with the following interesting example (see [267]).

Example 4.3.1. For S R2 defined by S D S1 [ S2 [ S3 , where
˚
S1 D .x1 ; x2 / 2 R2 j .x1 C 1/2 C x22 D 1; x1 2 Œ2; 0; x2 0 ;
˚
S2 D .x1 ; x2 / 2 R2 j 2 x1 0 ;
˚
S3 D .x11 ; 0/ 2 R2 j x1 2 N ;
we have
˚ [˚
T .S; .0; 0// D .x1 ; x2 / 2 R2 j x1 D 0; x2 0 .x1 ; x2 / 2 R2 j x2 D 0 ;
˚ [˚
A.S; .0; 0// D .x1 ; x2 / 2 R2 j x1 D 0; x2 0 .x1 ; x2 / 2 R2 j x1 0; x2 D 0 ;
C.S; .0; 0// D f.0; 0/g;
which shows that all the three cones, in general, are different.
To obtain the desired features of these cone the following notions have been
introduced:
Definition 4.3.2. Let X be a normed space, let S X be nonempty, and let xN 2 S .
N if T .S; x/
The set S is said to be tangentially regular at x, N D C.S; x/,N that is,
the contingent cone and the Clarke’s tangent cone coincide. Moreover, the set S is
N if T .S; x/
said to be derivable at x, N D A.S; x/,
N that is, the contingent cone and the
adjacent cone coincide.
Example 4.3.3. For a set S R2 given by S WD f.x; y/ 2 R2 j x 2 C y 2 2jxjg,
we have T .S; .0; 0// D R2 , however, C.S; .0; 0// D f.0; 0/g. This shows that the
tangential regularity concept is not related to the convexity of the contingent cone
(see [60]).
4.3 Miscellaneous Properties of First-Order Tangent Cones 131
We have the following useful result which is quite easy to prove:

Proposition 4.3.4. Let X and Y be normed spaces, let S be a subset of X , and let
xN 2 S be arbitrary. Assume that ` W X ! Y is a linear and continuous map. Then
the following inclusions hold:
N T .`.S /; `.x//;
`.T .S; x// N (4.8)
N A.`.S /; `.x//:
`.A.S; x// N (4.9)
Our next result is a collection of intersection formulas relating a few cones (see
[577]):
Theorem 4.3.5. Let X be a normed space, let S1 and S2 be two nonempty subsets
of X , and let xN 2 cl S1 \ cl S2 . Then the following relationships hold:
1. N \ IA.S2 ; x/
IT.S1 ; x/ N IA.S1 \ S2 ; x/
N
2. N \ A.S2 ; x/
IT.S1 ; x/ N A.S1 \ S2 ; x/
N
3. N \ T .S2 ; x/
IT.S1 ; x/ N T .S1 \ S2 ; x/
N
4. N \ A.S2 ; x/
IA.S1 ; x/ N T .S1 \ S2 ; x/.
N
Proof. The proof can be obtained by using the definition of the cones involved. u
t
We have another simple result connecting the three cones:
Theorem 4.3.6. Let X be a normed space, let S be a nonempty subset of X , and
let xN 2 cl S . Then the following inclusions hold:
N C C.S; x/
1. T .S; x/ N T .S; x/N
2. A.S; x/N C C.S; x/
N A.S; x/ N
N C C.S; x/
3. IT M .S; x/ N IT M .S; x/.
N
Proof. Let x 2 T .S; x/N and z 2 C.S; x/N be arbitrary. Then, by the definition of
the contingent cone, there are sequences ftn g P and fxn g X with tn # 0 and
xn ! z such that yn WD xN C tn xn 2 S for every n 2 N. Clearly, zn ! x.
N We recall
S
that due to the containment z 2 C.S; x/N for every fyn g X with yn ! x and for
every fsn g P with sn ! 0, there exists fzn g with zn ! z such that yn C sn zn 2 S
for every n 2 N. This implies that xN C tn .xn C zn / D yn C tn zn 2 S , confirming
that x C z 2 T .S; x/.
N This proves the first inclusion. The proof of the second and
the third inclusions can be given analogously (see also Theorem 4.12.8). t
u
We have the following useful consequence of the above result:
Corollary 4.3.7. Let X be a normed space, let S X , and let xN 2 S . If
N ¤ ;, then
IT M .S; x/
N D int .C.S; x//

IT M .S; x/ N :
N if IT M .S; x/
Remark 4.3.8. A set S is called epi-Lipschitzian at x, N ¤ ;.
Rockafellar [492] showed that in a finite dimensional setting any set S is
N ¤ ;. He also gave an example to

epi-Lipschitzian if and only if int .C.S; x//
N ¤ ; does not imply that S
show that in general spaces, the condition int .C.S; x//
is epi-Lipschitzian.
Several authors have studied the relationship between the Clarke’s tangent
cone and limit inferior of contingent cones at neighboring points. An important
contribution was made by Penot [461] (see also Cornet [113], Wang [589]) who
showed that for closed sets in finite dimensional setting the Clarke’s cone is
equal to the limit inferior of contingent cones. Treiman [574] extended this result
for epi-Lipschitzian closed sets in Banach spaces. Many related results and their
consequences can also be found in [18, 267].
In the following, we give a result (without proof) by Borwein and Strojwas [70,
Theorem 4.1]:
Theorem 4.3.9. Let X be a locally convex Hausdorff space, let S X be
compactly epi-Lipschitzian at xN 2 S . Suppose that one of the following assumptions
holds:
N
1. S is tangentially regular at x.
2. S is closed relative to some neighborhood of xN and X is a Banach space.
3. S is weakly compact relative to some neighborhood of xN and X is a normed
space.
Then:
lim inf T .S; x/ D C.S; x/:

N
S
x !xN
4.4 First-Order Tangent Cones on Convex Sets
On convex sets, the tangent cones of first-order have simpler structure and nice
properties. The highlight of this section is the relationship among the tangent cones
on convex sets.
4.4.1 Connections Among First-Order Tangent Cones

on Convex Sets
We recall that in Sect. 4.1 and Sect. 4.2, we studied the following tangent cones:
N D fx 2 X j 8 t > 0 9 s 2 .0; t/ such that xN C sx 2 S g;

R.S; x/
N D fx 2 X j 9 t > 0 8 s 2 .0; t/ we have xN C sx 2 S g;
IR.S; x/
4.4 First-Order Tangent Cones on Convex Sets 133
N D fx 2 X j 8 U.x/ 8 t > 0 9 s 2 .0; t/ 9 xO 2 U.x/ such that

T .S; x/
xN C s xO 2 S g ;
N D fx 2 X j 9 U.x/ 9 t > 0 8 s 2 .0; t/ 8 z 2 U.x/ we have xN C sz 2 S g ;
IT.S; x/
N D fx 2 X j 8 U.x/ 9 t > 0 8 s 2 .0; t/ 9 xO 2 U.x/ such that
A.S; x/
xN C s xO 2 S g ;
N D fx 2 X j 9 U.x/ 8 t > 0 9 s 2 .0; t/ 8 xO 2 U.x/ we have
IA.S; x/
xN C s xO 2 S g ;
N D fx 2 X j 8 V .x/
RM .S; x/ N 8 t > 0 9 s 2 .0; t/ 9 xO 2 V .x/
N \ S [ fxg
N such that
xO C sx 2 S g;
N D fx 2 X j 9 V .x/
IRM .S; x/ N 9 t > 0 8 s 2 .0; t/ 8 xO 2 V .x/
N \ S [ fxg
N we have
xO C sx 2 S g;
N D fx 2 X j 8 V .x/ 8 V .x/
TM .S; x/ N 8 t > 0 9 s 2 .0; t/9 xO 2 V .x/
N \ S [ fxg
N
9 xM 2 V .x/ W xO C s xM 2 S g;
N D fx 2 X j 9 V .x/ 9 V .x/
IT M .S; x/ N 9 t > 0 8 s 2 .0; t/ 8 xO 2 V .x/
N \ S [ fxg
N 8
xM 2 V .x/ we get xO C s xM 2 S g;
N D fx 2 X j 8 U.x/ 9 V .x/
C.S; x/ N 9 t > 0 8 s 2 .0; t/ 8 xO 2 V .x/
N \ S [ fxg
N 9
xM 2 U.x/ W xO C s xM 2 S g;
N D fx 2 X j 9 U.x/ 8 V .x/
IC.S; x/ N 8 t > 0 9 s 2 .0; t/ 9 xO 2 V .x/
N \ S [ fxg
N 8
xM 2 V .x/ W xO C s xM 2 S g:
The following relationships hold among the above cones for xN 2 cl .S / X :
N IT.S; x/
IT M .S; x/ N IA.S; x/
N IC.S; x/
N

N IR.S; x/
IRM .S; x/ N R.S; x/
N RM .S; x/
N

N A.S; x/
C.S; x/ N T .S; x/
N TM .S; x/;
N
where the cones in the first row are open and the cones in the third row are closed.
It turns out that this relationship simplifies significantly on convex sets. The
following result specifies the connections among various tangent cones of first-
order.
Theorem 4.4.1. Let X be a normed space, let S X be convex, and let x 2 cl S .

Then:
N D IT.S; x/
1. IT M .S; x/ N D IA.S; x/
N D cone .int S x/.
2. IR.S; x/N D R.S; x/
N D cone .S x/.
N
3. C.S; x/N D A.S; x/
N D T .S; x/
N D cl cone .S x/.
N
Proof. We begin by showing that when the set S is convex, the following inclusions
hold
N D IA.S; x/;
IT.S; x/ N (4.10)
N D R.S; x/;
IR.S; x/ N (4.11)
N D T .S; x/:
A.S; x/ N (4.12)
Let x 2 IA.S; x/N be arbitrary. Then there are a neighborhood U.x/ and a scalar
N
t > 0 such that xCtU.x/ S . Since the set S is convex, it follows that xCsU.x/
N
S for every s 2 .0; t/ which confirms that x 2 IT.S; x/. N The converse inclusion
follows from the properties mentioned above. This proves (4.10). Identity (4.11)
can be shown is an analogous manner.
Let x 2 T .S; x/N D T .cl S; x/.
N Then for each neighborhood U.x/, there exists
a scalar t > 0 such that .xN C tU.x// \ cl S ¤ ;, which, in view of the convexity,
of the set S , yields .xN C sU.x// \ cl S ¤ ; for every s 2 .0; t/, proving that
x 2 A.cl S; x/N D A.S; x/.N Since A.S; x/N T .S; x/,N we proved (4.12).
Next we claim that
N D IT M .S; x/;
IT.S; x/ N (4.13)
N D C.S; x/:
A.S; x/ N (4.14)
Let x 2 IT.S; x/.

N Then, there are a neighborhood U.x/ and a scalar t > 0 such
that for every s 2 .0; t, we have xN C sU.x/ S . For some 2 .0; 1/, we define
N WD xN C t.U.x/ x/, and U1 .x/ WD x C .1 /.U.x/ x/. Since U.x/ is
V1 .x/
convex, we have
N C tU1 .x/ D xN C t.U.x/ x/ C tx C t.1 /.U.x/ x/

V1 .x/
xN C tx C t.U.x/ x/ D xN C tU.x/ S;
S
and using the convexity of the set S , we have V1 .x/N \ S fxg N C sU1 .x/ S for
every, s 2 .0; t. Therefore, y 2 IC.S; x/,
N proving that IT.S; x/ N IT M .S; x/.
N Due
N IT M .S; x/,
to IT.S; x/ N we have proven (4.13).
For (4.14), we assume that y 2 A.S; x/. N Let U.x/ be an arbitrary neighborhood
of x. For some 2 .0; 1/, we define U1 .x/ WD x C .1 /.U.x/ x/. By the
N there exists a scalar t > 0 such that .xCsU1 .x//\S ¤ ;
definition of cone A.S; x/,
for every s 2 .0; t. We define a neighborhood V1 .x/N by V1 .x/
N WD x Cs.U.x/x/.
Then, for every xO 2 V1 .x/,
N we have
.xO C tU.x// \ S .xO C t.U.x/ x/ C tx C t.1 /.U.x/ x//

\S .xN C tU1 .x// \ S ¤ ;;
and using the convexity of the set S once again, we deduce that .xO C sU.x// \ S ¤
; for every xO 2 V1 .x/ N \ S [ fxg, N and for every s 2 .0; t, which proves that
x 2 IC.S; x/.
N We have shown that IA.S; x/ N IC.S; x/,
N and since the converse is
always true, we established (4.14).
Finally, for xN 2 cl S , we claim that
N D cone .S x/;
IR.S; x/ N (4.15)
N D cone .int S x/;
IT.S; x/ N (4.16)
N D cl cone .S x/:
T .S; x/ N (4.17)
The identity (4.15) is a consequence of the definition. Therefore, we have
N D IT.int S; x/
IT.S; x/ N IR.int S; x/
N D cone.int S x/:
N
Let x 2 cone.int S x/.N Then there exists a scalar t > 0 and an element xO 2
int S such that x D t.xO x/.
N Therefore, for U D t.int S x/, N which is an open
neighborhood
1 of x, we have xN C sU D xN C st.int S N
x/ S for every s 2
0; t , confirming that x 2 IT.S; x/.N Therefore, we have shown that IT.S; x/ N
cone.int S x/,
N and this proves (4.16).
For (4.17), we note that due to (4.15), we get T .S; x/ N cl IR.S; x/ N D
cl cone .S x/.
N To prove the converse, we assume that x 2 T .S; x/. N Then, for
any neighborhood U.x/, there exists a scalar t > 0 such that .xN C tU.x// \ S ¤ ;,
implying U.x/ \ cone.S x/ N ¤ ;, and hence x 2 cl cone .S x/. N t
u
Let us now recall another notion of a cone:
Definition 4.4.2. Let X be a normed space, let S X , and let xN 2 S . The closed
N
radial cone R.S; N is the collection of all x 2 X for which there are sequences
x/
ftn g P and fxn g X such that xN C tn xn 2 S .
Remark 4.4.3. Taa [548] traces this notion back to Bouligand [74]. It is clear that
N R.S;
T .S; x/ N N Furthermore, T .S; x/
x/. N
N D R.S; N if the set S is convex,
x/,
justifying its name (see (4.11) and (4.12)).
Our objective now is to study cones generated by taking a conical hull of set.
The importance of such cones and their connection to tangent cones is evident from
Theorem 4.4.1.
Let X be a linear topological space and let K be a convex cone with nonempty
interior. Let xN 2 X be a point. By K.x/,
N we denote the closure of the convex hull
of K C x, N that is,
N D cl cone .K C x/
K.x/ N D cl f
.k C xj
N k 2 K;

0/g :
Clearly, cone .K C x/
N K C f
xj
N

0g. In particular, if xN 2 K, then the
equality holds, or
N D cl .K C f
xj
K.x/ N

0g/ whenever xN 2 K: (4.18)
We can repeat the above construction. With another element zN 2 X , we denote

N zN/ the closure of the conical hull of K.x/
by K.x; N C zN, that is,
N zN/ D cl cone .K.x/

K.x; N C zN/ D cl f
.k.x/
N C zN/ j k.x/
N 2 K.x/;
N

0g :
We have the following result (see [45]):

Proposition 4.4.4. Let xN 2 K and let zN 2 K.x/.
N Then
N zN/ D cl .K C f
xN C Nzj

0;
0g/ :
K.x;
Proof. We set fxg

N WD f
xj
N

0; g and fNzg WD f
Nzj

0g, and apply (4.18) twice
to obtain that
N zN/ D cl .K.x/
K.x; N C zN/ D cl .cl .K C x/
N C zN/ : (4.19)
Now, setting A WD K C fxg

N C fNzg, we have int.A/ cl .K C fxg/
N C fNzg cl .A/.
Noting that A is convex, and by the fact that K is solid, A is solid as well. Using
cl .int.A// D cl .A/, we deduce that cl .cl .K C fxg/
N C fNzg/ D cl .K C fxgN C fNzg/.
This proves (4.19) and the proof is complete. t
u
We have the following important result based on the above calculation (see [45]):
Proposition 4.4.5. Let W be a topological space and let g W W ! Rn . We set
K D RnC and define I.x/ N WD fi 2 f1; 2; : : : ; ngj gi .x/
N D 0/g and J.x;
N d / WD fi 2
N gi0 .x/.d
I.x/j N / D 0g, where d 2 W . Then:
N 2 K, then
1. If g.x/

RC if i 2 I.x/
N
N D ˘inD1 Si
K.g.x// where Si D
R otherwise:
N 2 K, and g0 .x/.d

2. If g.x/ N / 2 K.g.x//,
N then

RC if i 2 J.x;
N d/
N g.x/d
K.g.x/; N /D ˘inD1 Si where Si D
R otherwise:
The first part of this result says that g0 .x/.d

N / 2 K.g.x//
N if and only if
gi0 .x/.d
N 0 for i 2 I.x/.
/ N
We conclude this section by another result concerning the interior of the conical
hulls:
Proposition 4.4.6. Let xN 2 K and let zN 2 K.x/.
N Then,
N D int.K/ C f
xj
int.K.x// N

0g; (4.20)
N zN// D int.K/ C f
xN C Nzj

0;
0g:
int.K.x; (4.21)
Proof. By (4.18), cone .K C x/N D K C f

xjN

0g. It is known that int.K.x//
N D
int.K/ C f
xjN

0g, see [628, Theorem 2.2]. Therefore, (4.20) follows from the
observation that int cl A D int A, where A is a solid convex set. Using (4.20) we
deduce (4.21). The proof is complete. t
u
4.4.2 Properties of First-Order Tangent Cones on Convex Sets
We begin by recalling that given a nonempty set S of a normed space X with X as

the topological dual of X , the normal cone to S at xN 2 S is given by:
N.x; N D fx 2 X j hx ; x xi
N S / D .S x/ N 0; for every x 2 S g :
The following well-known result computes the contingent cone on a constraint

set.
Theorem 4.4.7. Let X and Y be Banach spaces, let L X and M Y be closed
and convex, let A W X ! Y be linear and continuous, and let K WD fx 2 Lj A.x/ 2
M g D L \ A1 .M / be nonempty, that is, 0 2 A.L/ M . Then for any xN 2 K, we
have
T .K; x/ N \ A1 .T .M; A.x///;

N T .L; x/ N
N.x; N L/ C A N.A.x/;
N K/ N.x; N M /:
If we assume that
0 2 int.M A.L//; (4.22)
then for every xN 2 K D A1 .M / \ L, the following intersection formulas hold:
T .K; x/ N \ A1 .T .M; A.x///;

N D T .L; x/ N
N.x; N L/ C A N.A.x/;
N K/ D N.x; N M /:
Proof. See [18]). t

u
The following two corollaries are immediate:

Corollary 4.4.8. Let X and Y be Banach spaces, let L X be closed and convex,
and let A W X ! Y be linear and continuous. Then for any yN 2 int .A.L// and for
any xN 2 L \ A1 .y/,
N we have
T .L \ A1 .y/;
N x/N D T .L; x/
N \ Ker AI
N L \ A1 .y//
N.x; N L/ C Im A :
N D N.x;
Corollary 4.4.9. Let X and Y be Banach spaces and let K X and L Y be

closed and convex. If 0 2 int.K L/, then for any xN 2 K \ L, we have
T .K \ L; x/
N D T .K; x/
N \ T .L; x/;
N
N K \ L/ D N.x;
N.x; N K/ C N.x;
N L/:
The constraint qualification (4.22) plays an important role in the proof. However,
there are instances when this condition is violated but the intersection formula
remains valid. This situation is depicted in the following example given by
Křivan [341]:
Example 4.4.10. For X D Y D R2 , let K be a triangle with vertices .0; 0/, .1; 1/
and .1; 1/. By choosing L D K and A.x/ D x, we observe that 0 … int.K L/.
However, the intersection formula holds:
T .K; .0; 0// \ T .L; .0; 0// D T .K \ L; .0; 0//:
In the following, our objective is to give an extension of Theorem 4.4.7 by

relaxing the condition (4.22). This extension is motivated by an interesting work of
Křivan [341], who provided a sufficient condition in terms of the support functions
for the contingent cone to the intersection of two closed and convex subsets of a
Hilbert space to be equal to the intersection of contingent cones to these sets. He
also showed that for Hilbert spaces the surjectivity condition, which is essential to
prove the same formula in the context of reflexive Banach spaces, follows from his
condition (see Aubin and Cellina [15] and Aubin and Ekeland [16]).
We recall the following result:
Theorem 4.4.11. Let X; Y be separated locally convex spaces, f 2 .X /, g 2
.Y /, and A 2 L .X; Y / be such that A.dom f / \ dom g ¤ ;. If
.f C g ı A/ .x / D min ff .x A y / C g .y / j y 2 Y g ; 8x 2 X ;
(4.23)
then
@.f C g ı A/.x/ D @f .x/ C A .@g.Ax// ; 8x 2 X: (4.24)

Proof. The proof can be found in Hirriart-Urruty [263]. We provide a proof for
completeness.
Clearly, f C g ı A 2 .X /. The inclusion in (4.24) holds always.
Let x 2 @.f C g ı A/.x/. Then x 2 dom.f C g ı A/ D dom f \ A1 .dom g/
and
.f C g ı A/.x/ C .f C g ı A/ .x / D hx; x i :
From (4.23) we get y 2 Y .f C g ı A/ .x / D f .x A y / C g .y /.

It follows that
˝ ˛ ˝ ˛
f .x/ C f .x A y / x; x A y C g.Ax/ C g .y / Ax; y D 0:
Since f .x/ C f .x A y / hx; x A y i

0 and g.Ax/ C g .y /
hAx; y i
0 (by the Young–Fenchel inequality), we obtain that f .x/ C f .x
A y / hx; x A y i D 0 and g.Ax/ C g .y / hAx; y i D 0, and so
x A y 2 @f .x/ and y 2 @g.Ax/, whence x D x A y C A y 2
@f .x/ C A .@g.Ax//. The conclusion follows. t
u
We have the following extension of Theorem 4.4.7:
Theorem 4.4.12. Let X and Y be normed vector spaces, f 2 .X /, g 2 .Y /
and A 2 L .X; Y / be such that A.dom f / \ dom g ¤ ;. If there exists c > 0 such
that
.f C g ı A/ .x / D inf ff .x A y / C g .y / j y 2 Y ; ky k
c kx kg ; 8 x 2 X ; (4.25)
then (4.24) holds.

Proof. It is sufficient to observe that (4.23) holds. Because f , g are w -lsc and A
is w -w continuous we obtain that the mapping Y 3 y 7! f .x A y / C
g .y / 2 RN is w -lsc. Since the set U WD fy 2 Y j ky k c kx kg is w -
compact, there exists some yN 2 U Y such that
inf ff .x A y / C g .y / j y 2 Y ; ky k c kx kg
D f .x A yN / C g .yN /: (4.26)
Since one has always
.f C g ı A/ .x / inf ff .x A y / C g .y / j y 2 Y g 8x 2 X ;
from (4.25) and (4.26) we get that (4.23) holds. The conclusion follows using
Theorem 4.4.11. t
u
Recall that in a normed space X with X as its dual, the support function of
K X is given by
K .x / WD sup hx ; xi x 2 X :
x2K
The following result of Křivan [341] is a direct consequence of the above result:
Theorem 4.4.13. Let X; Y be normed vector space, L X , M Y be convex
sets and A 2 L .X; Y / such that L \ A1 .M / ¤ ;. Assume that there exists c > 0
such that
L\A1 .M / .x / D inf fL .x A y / C M .y / j y 2 Y ;

ky k c kx kg ; 8x 2 X : (4.27)
Then the following identities hold:
N.x; L \ A1 .M // D N.x; L/ C A .N.Ax; M // ; 8x 2 L \ A1 .M /;

(4.28)
T .L \ A1 .M /; x/ D T .L; x/ \ A1 .T .M; Ax// ; 8x 2 L \ A1 .M /:
(4.29)
Proof. Take f WD L and g WD M . Then f 2 .X /, g 2 .Y / and A.dom f / \

dom g D A.L/ \ M ¤ ;. Then f C g ı A D L\A1 .M / , f D L , g D M ,
.f C g ı A/ D L\A1 .M / , @f .x/ D N.x; L/ for x 2 L, @g.y/ D N.y; M / for
y 2 M and @.f C g ı A/.x/ D N.x; L \ A1 .M // for x 2 L \ A1 .M /. Applying
Theorem 4.4.12 we get (4.28).
The relation (4.29) follows immediately from (4.28) taking into account that for
(convex) cones K1 ; K2 U one has .K1 C K2 / D K1 \ K2 , while for the
(convex) cone K U and A 2 L .U; V / one has .A.K// D A1 .K / without
any supplementary condition, where U; V are separated locally convex spaces. u t
The following useful intersection formula is a direct consequence of the above
result.
Theorem 4.4.14. Let X be a normed space and let L X and K Y be two
nonempty, closed, and convex sets. Assume that there exists a constant c > 0 such
that for every x 2 X we have
˚
L\A1 .K/ .x / D inf L x z C K .z /j z 2 Y ; kz k ckx k :
(4.30)
Then, for every xN 2 L \ K, the following intersection formula holds:
T .L \ K; x/
N D T .L; x/
N \ T .K; x/:
N
Example 4.4.15. This example is a continuation of Example 4.4.10. That is, we

have X D R2 , and K is a triangle with vertices .0; 0/, .1; 1/ and .1; 1/, and
L D K. Since
B..0; 0/; 1/ ŒN..0; 0/; K/ \ B..0; 0/; 1/ C ŒN..0; 0/; L/ \ B..0; 0/; 1/ ;
it follows that for every x 2 X , there exists x1 2 .N.0; K/ \ B..0; 0/; 1// and
x2 2 .N.0; L/ \ B..0; 0/; 1// such that x D kx kx1 C kx kx2 . We set u D
kx kx1 . Since u 2 N.0; K/ and x u 2 N.0; L/, we deduce that
inf fK .w / C L .x w /j w 2 X; kw k kx kg K u L .x u / D 0:
Furthermore, it follows from the convexity of the support function K\L that
K\L .x / K\L .u / C K\L .x u / K .u / C L .x u /:
Since K\L .x / D 0 for every x 2 X , it follows that
0 D K\L .x / D inf fK .v / C L .x v /j v 2 X; kv k kx kg ;
confirming that (4.30) holds, and therefore, T .L \ K; .0; 0// D T .L; .0; 0// \
T .K; .0; 0//.
The following result, whose proof can be found in Křivan [341], connects the
two constraint qualification conditions, namely, (4.22) and (4.27) in a Hilbert space
setting.
Proposition 4.4.16. Let X and Y be Hilbert spaces and let L X and K Y be
nonempty, closed, and convex with L B.0; p/ for some p > 0. Let A W X ! Y
be a continuous and linear map. Assume that (4.22) holds. Then, there exists c > 0
such that (4.30) is satisfied.
Conditions ensuring (4.27) in a finite-dimensional setting are also given by
Křivan [341].
The following result, whose proof can be found in [16, Page 174] and [18,
Page 141], gives some additional properties of the contingent cone on a convex set.
Clearly, these stronger properties are a byproduct of the convexity of the underlying
N D N.x;
sets. In the following, we set T .S; x/ N S / .
Theorem 4.4.17. Let X , X1 ; : : : ; Xn , and Y be Banach spaces, let S; S1 ; S2 ; : : : ; Sn
N xN 1 , xN 2 ; : : : xN n be given elements. Then the following
be closed, convex sets, and let x,
properties hold:
1. If xN 1 2 S1 X and xN 2 2 S2 X , then
T .S1 C S2 ; xN 1 C xN 2 / D cl .T .S1 ; xN 1 / C T .S2 ; xN 2 //;

N.xN 1 C xN 2 ; S1 C S2 / D N.xN 1 ; S1 / \ N.xN 2 ; S2 /:
2. If A W X ! Y is linear and continuous and xN 2 S X , then
N D cl .A.T .S; x///

T .A.S /; A.x// N
N A.S // D .A /1 .N.x;
N.A.x/; N S //:
3. If xN i 2 Si Xi for i 2 f1; : : : ; ng, are convex, then
T .S1 S2 : : : Sn ; .xN 1 ; xN 2 ; : : : ; xN n // D T .S1 ; xN 1 / T .S2 ; xN 2 / : : : T .Sn ; xN n /;

N..xN 1 ; xN 2 ; : : : ; xN n /; S1 S2 : : : Sn / D N.xN 1 ; S1 / N.xN 2 ; S2 / : : : N.xN n ; Sn /:
4. Assume that for closed and convex sets Si X , for i 2 f1; : : : ; ng and for
xN 2 \niD1 Si , there exists k > 0 such that for every xi with kxi k k, we have
\niD1 .Si xi / ¤ ;. Then
T .S1 \ S2 \ : : : \ Sn ; x/
N D T .S1 ; x/
N \ : : : \ T .Sn; x/
N
N S1 \ S2 \ : : : \ Sn / D N.x;
N.x; N S1 / C : : : C N.x;
N Sn /:
We note that Bazaraa, Goode and Nashed [36] showed that if S1 and S2 are
closed and convex subsets of Rn such that ri.S1 / \ ri.S2 / ¤ ;, then for every
xN 2 cl S1 \ cl S2 , we have
N \ T .S2 ; x/
T .S1 ; x/ N D T .S1 \ S2 ; x/:
N (4.31)
We now collect a few characteristic examples of the contingent cones:

Example 4.4.18. Let X be a Hilbert space and let S X be a closed and convex
cone. Then
N S / D S \ fxg
N.x; N ?:
Consequently, x 2 T .S; x/
N if and only if
hs; xi 0 for every s 2 S with hs; xi

N D 0:
N S / D S ?.
N D S and N.x;
In particular, if S is a closed subspace, then T .S; x/
P
Example 4.4.19. Let ˙ D fx 2 RnC j niD1 xi D 1g be the simplex and let I.x/ N D
fi 2 f1; : : : ; ng with xN i D 0g. Then
x 2 T .RnC ; x/
N ” xi
0 for every i 2 I.x/;
N
X
n
x 2 T .˙n ; x;
N / ” xi
0 for every i 2 I.x/
N and xi D 0:
i D1
4.5 First-Order Local Cone Approximation 143
4.5 First-Order Local Cone Approximation
Having seen several interesting examples of tangent cones of first-order, our goal
in this section is to study an axiomatic approach for local cone approximation of
sets. The primary objective of this approach to single out the common features
shared by the most frequently used tangent cones. This approach also gives ways
to construct new local cone approximations from the known ones. This section is
based on papers by K. H. Elster and J. Thierfelder (see [179–181]), and it also
benefits from Castellani and Pappalardo [85], Dolecki [140, 141], and Ward [592].
For the following notion of abstract local cone approximation, we recall that the
recession cone S1 of a set S is the collection of all z such that for every r
0, we
have S C rz S .
Definition 4.5.1. Given a normed space X , a set-valued map K W 2X X ! 2X
is called a local cone approximation, if it associates to each set S X and each
point xN 2 X , a cone K.S; x/
N X satisfying the following properties:
1. N D K.S x;
K.S; x/ N 0/,
2. N D K.S \ U.x/;
K.S; x/ N x/ N for every U.x/,
N
3. N D ; if xN 62 cl S ,
K.S; x/
4. N D X if x 2 int S ,
K.S; x/
5. N D K.`.S /; `.x//,
`.K.S; x// N where ` W X ! X is a linear homomorphism,
6. S1 .K.S; x//
N 1 , where S1 and .K.S; x// N 1 are the recession cones of S and
N
K.S; x/.
Remark 4.5.2. In the above definition, property (1) demands the invariance of the
cone approximation with respect to contemporaneous shifting of the set S and the
N In other words, the cone approximation K.S; x/
point x. N and an approximation
of S xN at the origin are identical. Properties (2), (3), and (4) depict the local
nature of the local cone approximation. Axiom (5) shows the invariance of the cone
approximation with respect to a linear homomorphism such as rotation or reflection.
Condition (6) gives a relationship between the recession cone of S and K.S; x/. N
This property is crucial in deriving some conclusions regarding the so-called K-
epiderivatives.
The following theorem confirms that the notion of local cone approximation is
quite rich:
let xN 2 X . Then, the radial tangent cone, the feasible tangent cone, the contingent
cone, the interiorly contingent cone, the adjacent cone, the interiorly adjacent cone,
the modified radial tangent cone, the modified feasible tangent cone, the modified
contingent cone, the modified interiorly contingent cone, the Clarke’s cone, and the
modified interiorly adjacent cone are all local cone approximations.
Proof. We begin by noting that the first four properties either follow from the
definition or can be obtained from the features of these cones given in the previous
sections. Therefore, it suffices to prove the last two properties. We will give a proof
of these two properties only for the Clarke’s cone C.S; x/. N Similar arguments can
be used to develop detailed proofs for other cones.
We begin by showing that `.C.S; x// N D C.`.S /; `.x//.N Assume that x 2
N is arbitrary. This implies that `1 .x/ 2 C.S; x/.
`.C.S; x// N Let U.x/ be an arbitrary
neighborhood of x. Then, due to the continuity of `, there is a neighborhood
U.`1 .x// of `1 .x/ such that `.U.`1 .x/// U.x/. Since `1 .x/ 2 C.S; x/, N
there are a neighborhood V .x/ N and a scalar t > 0 such that for every s 2 .0; t/ and
for every xO 2 V .x/
N \ S [ x, N we have S s xO \ U.`1 .x// ¤ ;.
Also, there are a neighborhood V .`.x// N and a scalar t > 0 such that for every
s 2 .0; t/ and for every xO 2 V .`.x//
N \ `.S / [ f`.x/g,
N we have
S `1 .x/
O
\ U.`1 .x// ¤ ;:
s
Using `.U.`1 .x/// U.x/, we obtain
`.S / xO `.S / xO
\ U.x/ \ `.U.`1 .x/// ¤ ;;
s s
which ensures that x 2 C.`.S /; `.x// N and hence proving the inclusion
N C.`.S /; `.x//.
`.C.S; x// N
N be arbitrary. We claim that `1 .x/ 2
For the converse, let x 2 C.`.S /; `.x//
N Due to the continuity of ` , for a neighborhood U.`1 .x//, there exists a
C.S; x/. 1
neighborhood U.x/ such that
`1 .U.x// U.`1 .x//: (4.32)
Using the definition of the Clarke’s tangent cone, there are a neighborhood
N and a scalar t > 0 such that for every s 2 .0; t/ and for every xO 2
V .`.x//
N \ `.S / [ f`.x/g,
V .`.x// N we have
`.S / xO
\ U.x/ ¤ ;:
s
Also, there are a neighborhood V .x/N and a scalar t > 0 such that for every
s 2 .0; t/ and for every xO 2 V .x/
N \ S [ fxg,
N we have
`.S / `.x/
O
\ U.x/ ¤ ;;
s
which, by using (4.32), we obtain
S xO S xO
\ U.`1 .x// \ `1 .U.x// ¤ ;
s s
confirming that `1 .x/ 2 C.S; x/:

N
4.5 First-Order Local Cone Approximation 145
Finally, we proceed to prove the inclusion S1 .C.S; x// N 1 . We begin by

choosing z 2 S1 . Therefore, for every r
0, we have S Crz S . Let x 2 C.S; x/N
be arbitrary. Then for every neighborhood U.x/, there exists a neighborhood V .x/N
and a scalar t > 0 such that for every xO 2 V .x/
N \ S [ fxg
N and for every s 2 .0; t/,
we have
S xO
\ U.x/ ¤ ;:
s
For x C z, by taking U.x C z/ WD U.x/ C z, we obtain

S xO S xO
\ U.x C z/ D z \ U.x/ C z;
s s

S sz xO
D \ U.x/ C z;
s

S xO
\ U.x/ C z;
s
¤;
which proves that x C z 2 C.S; x/

N and hence z 2 .C.S; x//
N 1. t
u
The following result shows that all the conditions of the local cone approximation
are vital.
Theorem 4.5.4. For each axiom of the local cone approximation given in Defini-
tion 4.5.1, there exists a local cone approximation which violates this axiom and
satisfies all the other axioms. In other words, none of the axioms can be expressed
by the others.
Proof. We will construct local cone approximations violating one axiom and
verifying all the others. Let X be a normed space, let S X , and xN 2 X . Let
N be a local cone approximation.
K.S; x/
N WD K.S; x/
1. The local cone approximation defined by K1 .S; x/ N C xN violates the
first property K.S x;
N 0/ D K.S; x/
N but satisfies all the other conditions.
2. The local cone approximation defined by

cone .S x/
N if x 2 int S;
N WD
K2 .S; x/
; if x 2 X n int S;
N D K.S \ U.x/;
generally violates the second property K.S; x/ N x/N but satisfies
all the other axioms.
3. The local cone approximation defined by K3 .S; x/N WD X violates the third
property but satisfies all the other axioms.
4. The local cone approximation defined by K4 .S; x/ N WD ; violates the fourth

property but satisfies all the other axioms.
N WD
5. Let A be a nonempty cone. The local cone approximation defined by K5 .S; x/
N C A generally violates the fifth axiom but satisfies all the other
K.S; x/
conditions. In fact,
N D K.`.S /; `.x//
K5 .`.S /; `.x// N C A D `.K.S; x//
N C A;
N D `.K.S; x//
`.K5 .S; x// N C A/ D `.K.S; x//
N C `.A/:
Clearly, the first four axioms hold. The proof of the sixth axiom is as follows:
Let z 2 S1 . Then, we have .K.S; x//N 1 . Furthermore, K5 .S; x/Cz
N D K.S; x/CN
A C z K.S; x/ N C A D K5 .S; x/.N
6. The local cone approximation defined by
8
< ; if x … cl S;
N D X if
K6 .S; x/ x 2 int S;
:
f0g elsewhere;
violates the sixth property for the case when X D R2 , S D R2C and xN D .0; 0/.
In fact, we have
K6 .S; .0; 0// D f.0; 0/g;

.K6 .S; .0; 0///1 D f.0; 0/g;
and this completes the proof. t

u
The following results gives ways of constructing new notions of local cone
approximations:
let xN 2 X . Let I be a finite index set. Let K.S; x/
N and Ki .S; x/,
N with i 2 I , be local
cone approximations. Then each of the following relationships define a local cone
approximation:
1. N
int K.S; x/
2. N
cl K.S; x/
3. conv K.S; x/ N
4. X
[ nK.X nS; N
x/
5. N
Ki .S; x/
i 2I
\
6. N
Ki .S; x/
i 2I
N
7. ˙i 2I Ki .S; x/.
4.6 Convex Subcones of the Contingent Cone 147
Proof. We begin with a proof of part (1). The first four properties of Definition 4.5.1
being easy to verify, we give a proof for the last two properties. Let ` be a linear
homomorphism. Then:
N D int `.K.S; x//

int K.`.S /; `.x// N `.int K.S; x//:
N
Furthermore, since `1 is continuous, we have
N D int `1 .K.`.S /; `.x///

int K.S; x/ N `1 .int K.`.S /; `.x///;
N
N int K.`.S /; `.x//,

implying that `.int K.S; x// N and the property is proved.
N 1
For the sixth property of Definition 4.5.1, it suffices to note that .K.S; x//
.int K.S; x//1 .
The proofs of the statements that cl K.S; x/ N and conv K.S; x/ N are local cone
approximations can be made by analogous arguments.
For showing that X nK.X nS; x/N is a local cone approximation, we only prove
the sixth axiom of Definition 4.5.1. Due to the identity S1 D .X nS /1 , we infer
that
S1 D .X nS /1 .K.X nS; x//

N 1 D .X nK.X nS; x//
N 1:
The proof for the other properties is also easy to prove. t

u
Before any further advancement, we recall the following definition:
Definition 4.5.6. Given a normed space X , a local cone approximation K W 2X
X ! X is called isotone, if for each set A B X and for each xN 2 X , we have
N K.B; x/.
K.A; x/ N
The following result gives a useful implication of an isotone local cone approxi-
mation.
Theorem 4.5.7. Let X be a normed space, let S be a subset of X , and let
xN 2 X . Given a local cone approximation K.S; x/, N the local cone approximation
M.S; x/ N WD X nK.X nS; x/ N is isotone and forms a disjunction preserving pair with
K.; /. That is, for every A; B; X and every xN 2 X with A \ B D ;, it holds that
K.A; x/N \ M.B; x/ N D ;.
4.6 Convex Subcones of the Contingent Cone
As we have observed in the previous sections, two of the most commonly used
tangent cones, namely the contingent cone and the Clarke’s tangent cone, have
their own merits. The contingent cone is isotone, a property that plays a vital
role in deriving necessary optimality conditions and has numerous other useful
implications. On the other hand, the convexity of the Clarke’s tangent cone is a
remarkably useful analytical tool. As it has already been stated, convexity on the
contingent cone can be imposed by taking the convex hull of the contingent cone.
However, this process results in a cone that is often too large to give much useful
information as a local approximation tool. Therefore, it is of importance to ask
whether there are closed and convex cones that lie between the Clarke’s tangent cone
and the contingent cone. To answer this and study related questions, we begin by
investigating a local cone approximation K.; / which is convex-valued and satisfies
the following chain of inclusions
N K.S; x/
C.S; x/ N T .S; x/;
N (4.33)
where S is a nonempty subset of a normed space X and xN 2 X .

Convex and closed cones K.; / satisfying (4.33) have been known to play an
important role in nonsmooth analysis and optimization and we anticipate their use
in set-valued optimization.
An important question pertaining to the above discussion is whether we can find
a cone that contains the most important features of the contingent cone and the
Clarke’s cone, namely, it is isotone and convex both.
The following impossibility theorem of Ward [590] gives a negative answer to
this question:
Theorem 4.6.1. There is no mapping K W 2R Rn ! 2R that has all the following
n n
properties:
1. K is isotone
2. K is convex
3. K.S; x/ N for every S Rn and xN 2 cl S
N T .S; x/
4. S K.S; x/,
N whenever S is a one dimensional subspace of Rn and xN 2 S .
Proof. Consider the subsets of Rn defined by S1 D R f0g and S2 D f0g R.
Note that if K possessed the third property, then K.S1 [ S2 ; .0; 0// S1 [ S2 . On
the other hand, if K is isotone and possessed the fourth property, then S1 [ S2
K.S1 ; 0/ [ K.S2 ; 0/ D K.S1 [ S2 ; .0; 0//. Therefore, K.S1 [ S2 ; .0; 0// D S1 [ S2 ,
if K is isotone and satisfies the last two properties. Evidently, such a mapping is
bound to fail to be convex. t
u
Having obtained a negative answer, our next genuine step is to ask whether
there are closed and convex cones satisfying (4.33) which yield more accurate
approximation of a set than the Clarke’s tangent cone and at the same time are
also equipped with its analytical strengths? In the following, we will show that an
answer to this question is affirmative. We intend to discuss a few cones that satisfy
(4.33) and have significantly useful properties. Our treatment of this subject is based
on D. Ward’s seminal contributions [590, 591, 593, 597].
We begin by noticing that a simple way to obtain a closed and convex cone is by
taking the recession cone of a closed cone. Given a nonempty subset S of a normed
space X and an element xN 2 cl S , we define the recession contingent cone and the
recession adjacent cone as follows:
N D fx 2 X j T .S; x/
T1 .S; x/ N C x T .S; x/g;
N (4.34)
N D fx 2 X j A.S; x/
A1 .S; x/ N C x A.S; x/g:
N (4.35)
It follows from the above definition that
C.S; x/ N T .S; x/;

N T1 .S; x/ N
N A1 .S; x/
C.S; x/ N A.S; x/;
N
and it can be shown that the above chain of inclusions are, in general, strict (see
[414]).
The following example shows that despite of the well-known inclusion
N T .S; x/,
A.S; x/ N T1 .S; x/
N the inclusion A1 .S; x/ N is, in general, false.
Example 4.6.2. Define f W R ! R by

0 if x D 0;
f .x/ D
2.nC1/ if 2.nC1/ jxj < 2n ; n D 0; ˙1; ˙2; : : : :
Then, for S D epi .f / and xN D .0; 0/, we have
N D f.x; y/j y
jxjg ;
T .S; x/

jxj
N D .x; y/j y

A.S; x/ ;
2
and consequently
N D f.x; y/j y
jxjg ;
T1 .S; x/

jxj
N D .x; y/j y
A1 .S; x/ :
2
In the following, we give the two cones, both introduced by Penot [463].
Definition 4.6.3. Let X be a normed space, let S be a subset of X , and let xN 2 S .
1. The prototangent cone P .S; x/ N of S at xN is the collection of all x 2 X such
that for every sequence .xn ; tn / ! .x; N 0C / with fxn g S and tn1 .xn x/ N
convergent, there exists fzn g X converging to x with xn C tn zn 2 S for every
n 2 N.
2. The interiorly prototangent cone P .S; x/ N of S at xN is the collection of all x 2
N 0C / with fxn g S and tn1 .xn x/
X such that for every sequence .xn ; tn / ! .x; N
convergent, for every fzn g X converging to x, we have xn C tn zn 2 S , for
sufficiently large n 2 N.
It follows from the above definition that
N P .S; x/
C.S; x/ N A1 .S; x/
N A.S; x/
N T .S; x/;
N
N D T .S; x/,
and, if, A.S; x/ N then
P .S; x/ N D T1 .S; x/:

N D A1 .S; x/ N
The following result, which follows directly from the definitions of the involved
tangent cones, gives some useful information:
Proposition 4.6.4. Let X and Y be normed spaces, let S1 X and S2 Y be two
subsets, and let xN 2 cl S1 and yN 2 cl S2 . Then
N A1 .S2 ; y/
A1 .S1 ; x/ N D A1 .S1 S2 ; .x;
N y//;
N
N P .S2 ; y/
P .S1 ; x/ N P .S1 S2 ; .x;
N y//:
N
Note that often product formulas are given by combining two tangent cones. For
instance, it has been known that the following inclusion
N A.S2 ; y/
T .S1 ; x/ N T .S1 S2 ; .x;
N y//;
N (4.36)
N cannot be combined with any

holds. The following example shows that T1 .S; x/
N
other cone to produce an analogue of such formulas. This deficiency of T1 .S; x/
limits is usefulness.
Example 4.6.5. For the two sets S1 and S2 given by
S1 D fx 2 Rj x D 22n ; n D 1; 2; 3; : : :g [ f0g;
S2 D fx 2 Rj x D 22nC1 ; n D 1; 2; 3; : : :g [ f0g;
we have
T .S1 ; 0/ D T .S2 ; 0/ D RC ;
A.S1 ; 0/ D A.S2 ; 0/ D f0g;
T1 .S1 S2 ; .0; 0// D f.0; 0/g;
and hence, the inclusion T1 .S1 ; 0/ K.S2 ; 0/ T1 .S1 S2 ; .0; 0// fails for
K D T; T1 ; A; A1 ; P or C .
The following two results shed some light on the behaviour of the tangent cones
under linear and continuous maps.
Proposition 4.6.6. Let X and Y be normed spaces, let S X with xN 2 cl S , and

let L W X ! Y be a linear and continuous map. Then the following implications
hold:
N L.A.S; x//
A.L.S /; L.x// N H) N A1 .L.S /; L.x//;
L.A1 .S; x// N
N L.T .S; x//
T .L.S /; L.x// N H) N T1 .L.S /; L.x//:
L.T1 .S; x// N
Proof. For x 2 A1 .S; x/,

N we set z D L.x/. Let v 2 A.L.S /; L.x//N be arbitrary.
By hypothesis v D L.u/ for some u 2 A.S; x/. N Therefore, z C v D L.x C u/,
and since x C u 2 A.S; x/, N we get z C v 2 L.A.S; x// N A.L.S /; L.x//.N
Consequently, z D A1 .L.S /; L.x//N and the first implication is established. The
second implication can be proved in an analogous fashion. t
u
Proposition 4.6.7. Let X and Y be normed spaces, let S X , let xN 2 cl S , and let
L W X ! Y be a linear and continuous map. Assume that the following condition
N 0C / such that fvn g L.S / and tn1 .vn L.x//
holds: For every .vn ; tn / ! .L.x/; N
converges, there are fun g S with vn D L.un / and tn1 .un x/
N that are convergent.
Then the following inclusions hold:
N P .L.S /; L.x//;
L.P .S; x// N
N K1 .L.S /; L.x//:
L.K1 .S; x// N
Proof. Let z 2 L.P .S; x//

N be arbitrary. Then z D L.x/ with x 2 P .S; x/.N Assume
N 0C /, with fzn g L.S / and tn1 .zn L.x//,
that .zn ; tn / ! .L.x/; N is convergent. By
the hypothesis, there exists fxn g S with zn D L.xn / and convergent tn1 .xn x/.
N
Then there exists un ! x such that xn C tn un 2 S . Hence L.un / ! L.x/ and zn C
tn L.un / D L.xn C tn un / 2 L.S /. Consequently, z 2 P .L.S /; L.x// N and the first
inclusion is verified. The second inclusion follows the same lines of argument. u t
In the following, we study another closed convex cone and its “interiorly”
counterpart.
Definition 4.6.8. Let X be a Banach space, and let S be a subset of X that is closed
near xN 2 S .
1. The Treiman tangent cone E.S; x/ N to S at xN is the set of all x 2 T .S; x/
N such
N such that for all xO 2 ŒU.x/
that for every ";
> 0, there exists U.x/ N \ S nfxg,
N
we have
S \ fxO C .0;
kxO xkU
N " .x/g ¤ ;: (4.37)
2. The interiorly Treiman tangent cone IE.S; x/ N to S at xN is the set of all

x 2 T .S; x/N for which there are ";
> 0 such that for every t 2 .0;
/,
N satisfying that for every xO 2 ŒU.x/
there exists U.x/ N \ S nfxg,
N we have
xO C tkxO xkU
N " .x/ S .
The Treiman tangent cone was introduced by Treiman [545] whereas the
interiorly Tangent cone was introduced by Ward [593]. J. Treiman gave several
equivalent characterizations of the Treiman cone. In particular, he showed that
(4.37) can be replaced by the following condition:
S \ fxO C
kxO xkU
N " .x/g ¤ ;: (4.38)
We now give two characterizations of the Trieman tangent cone proposed by

Ward [593]:
Proposition 4.6.9. Let X be a Banach space and let S X be closed near xN 2
bd.S /. Then
N D fx 2 T .S; x/j
E.S; x/ N 8";
> 0; 9 U.x/
N such that (4.37) holds
8xO 2 ŒU.x/
N \ bd.S / nfxgg:
N (4.39)
N
Proof. Let T be the set on the right hand side of (4.39). Since S is closed near x,
N such that U.x/
there exists U.x/ N \ bd.S / U.x/ N \ S . It then becomes evident that
E.S; x/N T . For the converse, let x 2 X nE.S; x/.
N Then there exists ";
> 0 such
N there exists xO 2 .U.x/
that for every U.x/, N \ S /nfxg
N satisfying that
S \ fxO C .0;
kxO xkU
N " .x/g D ;:
Clearly such an element xO must belong to bd.S /, so x 2 X nfT g. This implies

that T E.S; x/
N and (4.39) holds. The proof is complete. t
u
N follows easily from Proposi-
The following sequential version of E.S; x/
tion 4.6.9.
Proposition 4.6.10. Let X be a Banach space and let S X be closed near xN 2 S .
Then the Trieman tangent cone E.S; x/N is the collection of all z 2 T .S; x/
N such
that for all fxn g bd.S / nfxg
N with xn ! x, N there exists a subsequence fxm g
and there exist sequences ftm g P and fzm g X with tm # 0, zm ! z and
xm C tm kxn xkz
N m 2 S.
Before any advancement, we also recall the following version of prototangent
cone proposed by Michel and Penot [420] (see [545] and Definition 4.6.3).
Definition 4.6.11. Let X be a normed space, let S be a subset of X , and let xN 2 S .
The P-tangent cone PM.S; x/ N of S at xN is the collection of all x 2 X such that
N 0C / with fxn g S and tn1 .xn x/
for every sequence .xn ; tn / ! .x; N convergent,
there exists a subsequence fxm g of fxn g and a sequence fzm g converging to x with
xm C tm zm 2 S .
We have the following results concerning the properties of the above cones (see
[593]).
Theorem 4.6.12. Let X be a Banach space, let S; S1 ; S2 be subsets of X which are

closed near xN 2 S \ S1 \ S2 . Then
1. N is a closed, convex cone
E.S; x/
2. N is an open convex cone
E.S; x/
3. N E.S; x/
C.S; x/ N PM.S; x/ N T .S; x/
N
4. N IE.S; x/
IC.S; x/ N E.S; x/N
5. N D PM.S; x/
E.S; x/ N provided that X D Rn
6. N \ IE.S2 ; x/
E.S1 ; x/ N E.S1 \ S2 ; x/
N
7. N C IE.S; x/
E.S; x/ N IE.S; x/
N
8. E.S; xN D cl IE.S; x//
N provided that IE.S; x/
N ¤;
9. IE.S; xN D int E.S; x//
N provided that IE.S; x/
N ¤;
10. N \ E.S2 ; x/
E.S1 ; x/ N E.S1 \ S2 ; x/
N provided that E.S1 ; x/
N \ IE.S2 ; x/
N ¤ ;.
Proof. A proof of the first property can be found in [545] whereas the next four
properties follow from the definition. We remark that the convexity of IE.S; x/ N
directly follows from the inclusion E.S; x/ N C IE.S; x/ N IE.S; x/N proved in the
following.
We give a proof for the inclusion E.S1 ; x/N \ IE.S2 ; x/
N E.S1 \ S2 ; x/. N Let
x 2 E.S1 ; x/ N \ IE.S2 ; x/
N be arbitrary. Then x 2 A.S1 ; x/N \ A.S2 ; x/
N due to the
fact that E.; / A.; / (see [593]). Therefore, x 2 A.S1 \ S2 ; x/
N implying that
x 2 T .S1 \ S2 ; x/.
N Now let " > 0 and
> 0 be given. Then there exist
2 2 .0;
/
and "2 2 .0; "/ such that for each t 2 .0;
2 /, there exists Ut .x/
N satisfying that for
all xO 2 .Ut .x/
N \ S2 /nfxg,
N we have
N "2 .x/ S2 :
xO C tkxO xkU (4.40)
Let t 2 .0;
2 / be arbitrary. By (4.38), there exists U.x/
N with U.x/
N Ut .x/
N such
that for each xO 2 .U.x/N \ S1 /nfxg,
N there exists z 2 U"2 .x/ satisfying
xO C
kxO xkz
N 2 S1 : (4.41)
N \ S1 \ S2 /nfxg.
Now, let xO 2 .U.x/ N Then there exists z 2 U"2 .x/ satisfying
the above equation. In addition xO C tkxO xkzN S2 by (4.40). Therefore, x 2
E.S1 \ S2 ; x/.
N
N C IE.S; x/
For the inclusion E.S; x/ N IE.S; x/,N let x 2 E.S; x/ N and z 2
N Without any loss of generality, we assume that x and z are both nonzero.
IE.S; x/.
It is easy to show that x C z 2 K.S; x/. N Since z 2 IE.S; x/, N there exists " 2
.0; minfkxk; kzkg/ and
> 0 such that for each t 2 .0;
/, there exists U.x/
N with
xO C tkxO C xk.x
N C U" .0// S , for every xO 2 .U.x/
N \ S /nfxg.
N Choose
1 2 .0;
/
such that
i1 "
" 1 h "

1 < min kxk C ; kxk C .kzk C "/ :
4 4 2
Then, for each t 2 .0;

1 /, there exists Ut .x/
N such that for every xO 2 .Ut .x/ N \
S /nfxg,N we have xO C tkxO xk.zN C U" .0// S . Moreover, there exists U.x/ N with
U.x/ N Ut .x/
N such that for every xO 2 .U.x/N \ S /nfxg
N we have xO C tkxO xk.x
N C
U
˚ " .0// U t . N
x/.
Since x 2 E.S;
N
x/, there exists U 1 . N
x/ with U 1 . N
x/ U. N
x/ and
xO C tkxO xkN x C U"=4 .0/ \ S ¤ ;, for every xO 2 .U1 .x/ N \ S /nfxg.N
Now let t 2 .0;
1 /, xO 2 .U1 .x/ N \ S /nfxg,
N and w 2 U"=4 .0/. There exists
w1 2 U"=4 .0/ such that xQ WD xO C tkxO xk.xN C w1 / 2 S \ Ut .x/. N We also have
xQ ¤ x, N since
kxQ xk N
1 kxO xk.kxk
N
kxO xk N C kw1 k/ D kxO xk.1
N
1 /.kxk C kw1 k/ > 0
by our choice of
1 . Therefore, xQ C tkxQ xk.z
N C U" .0// S , or in other words,
xO C tkxO xk.x
N C w1 / C tkxQ C xk.z
N C U" .0// S:
Finally, we define u such that kxQ xk.z

N C u/ D kxO xk.z
N C w w1 / which
implies that
N 1 f.kxO xk
u D kxQ xk N kxQ xk/x
N C kxO xk.w
N w1 /g:
To finish the proof, it suffices to show that kuk < ". For this, we first note that
j kxO xk
N kxQ xk
N j kxQ xk
O tkxO xk.kxk
N C kw1 k/:
Then by our choice of

1 , we have
N 1 kxO xkftkzk.kxk
kuk kxQ xk N C kw1 k/ C kw w1 kg
h " i 1 n " "o
1
1 kxk C
1 kxk C kzk C < ";
4 4 2
and we conclude that x C z 2 IE.S; x/. N This at once confirms the eighth and the
ninth properties.
For the final property, let x 2 E.S1 ; x/
N \ E.S2 ; x/N and y 2 E.S1 ; x/
N \ IE.S2 ; x/
N
be arbitrary. By using the convexity of cones, we get x C ty 2 E.S1 ; x/ N \ IE.S2 ; x/
N
for every t > 0. Then, x C ty 2 E.S1 \ S2 ; x/ N for every t > 0. Since the cone E.; /
is closed, we deduce that x 2 E.S1 \ S2 ; x/.N t
u
N D PM.S; x/
The following example shows that E.S; x/ N is false in infinite
dimensions (see [545]):
Example 4.6.13. Let X be an infinite dimensional Banach space and let
fx0 ; x1 ; x2 : : :g be a sequence of elements of norm 1 such that for all m; n we
have kxn xm k > ı > 0. Define
nx o
S WD Œ0; 1/ fx0 g [ [1
n
nD1 :
n
Then S is a closed set such that
T .S; 0/ D PM.S; 0/ D Œ0; 1/ fx0 g;

E.S; 0/ D C.S; 0/ D f0g:
We give another example to show that the Treiman tangent cone may be strictly
larger than the Clarke’s tangent cone.
Example 4.6.14. For S R2 defined by S D f.x; y/j y 0g [ f.x; y/j y D x 2 g,
we have
E.S; .0; 0// D PM.S; .0; 0// D f.x; y/j y 0g;

C.S; .0; 0// D f.x; 0/j x 2 Rg:
The following result gives a useful property of the Treiman tangent cone:
Theorem 4.6.15. Let X and Y be Banach spaces and let L W X ! Y be a linear
and continuous mapping. Let S X be closed near xN 2 S . Assume that for each
sequence fwn g bd .L.S //nfL.x/gN with wn ! L.x/, N there are a subsequence
fwm g and a sequence fxm g bd.S /nfxg
N such that L.xm / D wm with xm ! xN and
fkxm xk=kw
N m L.x/kg
N is bounded above. Then:
N E.L.S /; L.x//:
L.E.S; x// N
Proof. Let z 2 L.E.S; x//N with z D L.x/ with x 2 E.S; x/. N Then x 2 T .S; x/N
N Let fwn g be a sequence in bd.L.S //nfL.x/g
and hence z 2 T .L.S /; L.x//. N that
converges to L.x/.N By hypothesis, there are a subsequence fwm g and a sequence
fxm g bd.S /nfxg N such that L.xm / D wm , xm ! x, N and the sequence fkxm
N
xk=kw m L.x/kg
N is bounded from above. Since x 2 E.S; x/, N the sequence fxn g
may be chosen so that there exist sequences ftm g and fzm g with tm # 0, zm ! x, and
xm C tm kxm xkzN m 2 S . Then wm C tm kxm xkL.z
N m / 2 L.S / and L.zm / ! z,
since L is linear and continuous. We define
kxm xk
N
sm WD tm :
kwm L.x/k
N
Since tm ! 0 and fkxm xk=kw

N m L.x/kg
N is bounded, we deduce that sm ! 0
and that wm C sm kwm L.x/kL.z
N m / 2 S which confirms that z 2 E.L.S /; L.x//
N
and the proof is complete. t
u
We conclude this section by discussing the notion of quasi-strict tangent cone of
Penot [463].
xN 2 S .
N of S at xN is the collection of all

1. The quasi-strict tangent cone Q.S; x/
x 2 X such that for every sequence .xn ; tn / ! .x;
N 0C / with fxn g S and
lim tn1 .xn x/
N 2 A.S; x/,
N there exists fzn g X converging to x with
n!1
xn C tn zn
2 S for every n 2 N.
2. The interiorly quasi-strict tangent cone IQ.S; x/ N of S at xN is the collection
of all x 2 X such that for every sequence .xn ; tn / ! .x;N 0C / with fxn g S
and lim tn1 .xn x/ N 2 A.S; x/,
N for every fzn g X converging to x, we have
n!1
xn C tn zn 2 S , for sufficiently large n 2 N.
We have the following results concerning the properties of the above cone (see
[463, 593]).
Theorem 4.6.17. Let X be a Banach space, let S; S1 ; S2 be subsets of X which are
closed near xN 2 S \ S1 \ S2 . Then
1. Q.S; x/N is a closed, convex cone
2. N E.S; x/
C.S; x/ N Q.S; x/ N A.S; x/
N T .S; x/
N
3. N \ Q.S2 ; x/
Q.S1 ; x/ N Q.S1 \ S2 ; x/
N provided that Q.S1 ; x/
N \ IQ.S2 ; x/
N ¤;
4. N \ A.S2 ; x/
A.S1 ; x/ N D A.S1 \ S2 ; x/
N provided that Q.S1 ; x/
N \ IQ.S2 ; x/
N ¤ ;.
Proof. The proof is very similar to arguments used above and can be found in [463,
593]. t
u
We also state the following analogue of Theorem 4.6.15.
Theorem 4.6.18. Let X and Y be Banach spaces, let L W X ! Y be a linear and
continuous mapping, let S X , and let xN 2 S . Assume that A.L.S /; L.x//
N
N Then L.Q.S; x//
L.A.S; x//. N Q.L.S /; L.x//.N
Proof. See [590, 593]. t
u
4.7 First-Order Inversion Theorems and Intersection

Formulas
We now proceed to give inversion theorems which play a fundamental role in

deriving calculus rules for the derivatives of set-valued maps. For this, we first
collect a few definitions and results:
Definition 4.7.1. Let X be a Banach space and let S X be nonempty.
1. The set S is said to be closed near x 2 cl S , if U" .x/ \ S is closed for some
" > 0.
N it is called epi-Lipschitz-like at x,
2. If S is closed near x, N if there are > 0, a
convex set Q with polar set Qo weak locally compact, and ˇ > 0 such that
S \ U .x/ C tQ S for every t 2 .0; ˇ/.
4.7 First-Order Inversion Theorems and Intersection Formulas 157
Remark 4.7.2. Note that if X is finite-dimensional, then any locally closed set is
epi-Lipschitz-like by setting Q D f0g. See [62, 63] for an excellent introduction to
epi-Lipschitz-like sets.
Definition 4.7.3. Let X and Y be normed spaces and let g W X ! Y be a given
map.
1. The map g is called strictly differentiable at xN 2 X , if there is a continuous
N W X ! Y such that
linear map Dg.x/
g.x C ty/ g.x/

lim D Dg.x/.
N y/;
N for every yN 2 X:
N y;0
.x;y;t /!.x; N C/ t
2. The map g is called Hadamard differentiable at xN 2 X , if
g.xN C ty/ g.x/

N
lim D Dg.x/.
N y/;
N for every yN 2 X:
N C/
.y;t /!.y;0 t
The following important result is a special case of Borwein and Strojwas [65,
Theorem 4.1]. This result, which unifies the finite-dimensional and Banach space
case treated separately by Borwein [60], is taken from Ward [590].
Theorem 4.7.4. Let X and Y be Banach spaces, and let g W X ! Y be strictly
N for some
> 0, where xN 2 g 1 .0/ \ S and S X is
differentiable on U
.x/,
N Suppose
epi-Lipschitz -like at x.
N D Y:
Dg.x/C.S; x/ (4.42)
N there exists
Then there are k > 0 and ı > 0 such that for each x 2 S \ Uı .x/,
z 2 S \ g1 .0/ with
kx zk kkg.x/k:
With the above preparation, we give the following inversion theorem taken from
[590]:
Theorem 4.7.5. Let X and Y be Banach spaces, and let g W X ! Y be strictly
N for some
> 0, where xN 2 g 1 .0/ \ S and S X is epi-
differentiable on U
.x/
N Assume that (4.42) holds. Then the following inclusions hold:
Lipschitz -like at x.
C.S; x/ N 1 .0/ C.S \ g 1 .0/; x/;

N \ Dg.x/ N (4.43)
A.S; x/ N 1 .0/ A.S \ g 1 .0/; x/;
N \ Dg.x/ N (4.44)
1 1
N \ Dg.x/
T .S; x/ N .0/ T .S \ g .0/; x/;
N (4.45)
P .S; x/ N 1 .0/ P .S \ g 1 .0/; x/;
N \ Dg.x/ N (4.46)
T1 .S; x/ N 1 .0/ T1 .S \ g 1 .0/; x/;

N \ Dg.x/ N (4.47)
A1 .S; x/ N 1 .0/ A1 .S \ g 1 .0/; x/:
N \ Dg.x/ N (4.48)
Proof. The inclusions for the Clarke’s tangent cone and the contingent cone are
given in [60]. We prove the inclusion for A1 and the remaining inclusions can
be proved by using similar arguments. Let x 2 A1 .S; x/ N \ Dg.x/ N 1 .0/ and
1
z 2 A.S \ g .0/; x/ N be chosen arbitrarily. It is sufficient to show that x C z 2
A.S \ g 1 .0/; x/. N and z 2 A.g1 .0/; x/
N Since A.; / is isotone, z 2 A.S; x/ N
1
Dg.x/N .0/. Therefore, for an arbitrary sequence ftn g P with tn ! 0, there exists
fwn g with wn ! .x C z/ and xN C tn wn 2 S . Since g is strictly differentiable at x,
N
we have
g.xN C tn wn / g.x/
N
! Dg.x/.x
N C z/ D 0:
tn
Now, Theorem 4.7.4 implies that there exists un 2 S \ g 1 .0/ with

tn1
.un xN tn wn / ! 0. By setting vn D tn1 .un x/,
N we note that vn ! .x C z/
and xN C tn vn D un 2 S \ g 1 .0/. Consequently x C z 2 A.S \ g 1 .0/; x/
N and the
proof is complete. t
u
The following inversion, which is recently given by Durea and Strugariu [155],
is based on the notion of metric subregularity which is an extension of the metric
regularity, see [10, 146, 155, 622]. We remark that a detailed study of such results
under the notion of metric regularity have been carried out by Borwein [60],
Cominetti [103] and Jourani [309], among others.
Theorem 4.7.6. Let X and Y be Banach spaces, let S X and Q Y be closed
sets, let g W X ! Y be a continuously Fréchet differentiable map, and let xN 2
S \ g 1 .Q/. Assume that f W X Y ! Y defined by f .x; y/ WD g.x/ y is
N g.x/;
metrically subregular at .x; N 0/ with respect to S Q in the following sense:
N and U .g.x//
There exist a constant k > 0 and neighborhoods U .x/ N such that for
every .x; y/ 2 ŒU .x/
N U .g.x//
N \ S Q, we have

d .x; y/; f 1 .0/ \ .S Q/ kkg.x/ yk: (4.49)
Then the following formulas hold:

N 1 .A .Q; g.x///
N \ Dg.x/
T .S; x/ N T S \ g 1 .Q/; xN ;

N 1 .T .Q; g.x///
N \ Dg.x/
A .S; x/ N T S \ g 1 .Q/; xN ;

N 1 .A .Q; g.x///
N \ Dg.x/
A .S; x/ N D A S \ g 1 .Q/; xN :
Proof. In view of the metric subregularity hypothesis, there are a number k > 0
N and U .g.x//
and neighborhoods U .x/ N such that for every .x; y/ 2 ŒU .x/ N
N \ S Q, (4.49) holds.
U .g.x//
4.7 First-Order Inversion Theorems and Intersection Formulas 159
We choose x 2 T .S; x/ N 1 .A.Q; g.x///

N \ Dg.x/ N so that x 2 T .S; x/ N and
N
Dg.x/.x/ 2 A.Q; g.x//.N By using the definitions of the contingent cone and the
adjacent cone, there are sequences ftn g P, fxn g X and fzn g Y such that
tn # 0, xn ! x and zn ! Dg.x/.x/N with xN C tn xn 2 S and g.x/
N C tn zn 2 Q. For
sufficiently large n 2 N, (4.49) with .x; y/ D .xN C tn xn ; g.x/
N C tn zn / ensures that
there exists .pn ; qn / 2 S Q with g.pn / D qn such that
k.xN C tn xn ; g.x/
N C tn zn / .pn ; qn /k < kkg.xN C tn xn / g.x/
N tn zn k C tn2 :
Consequently, for sufficiently large n 2 N, we have pn 2 S \ g 1 .Q/ and
kxN C tn xn pn k < kkg.xN C tn xn / g.x/

N tn zn k C tn2
which implies that

pn xN g.xN C tn xn / g.x/
N
C x < k z
n C tn :
t n tn
n
Noticing that
g.xN C tn xn / g.x/
N
! Dg.x/.x/;
N
tn
we deduce that wn WD tn1 .pn x/ N ! x and the proof of the first inclusion is
complete. The remaining parts follow by using the properties of the contingent cone
and the adjacent cone. t
u
For the sake of completeness, we also state the following inversion theorem from
[18, Page 147]. As it can be seen that (4.50) is used to prove that the inverse of a
certain map is pseudo-Lipschitz and hence it can be obtained by the arguments used
in the proof of Theorem 4.7.6.
Theorem 4.7.7. Let X and Y be Banach spaces, let S X and Q Y be closed
sets, let g W X ! Y be a continuously differentiable around xN 2 S \ g 1 .Q/.
Assume that either Y is finite-dimensional and
N
Dg.x/.C.S; N C.Q; g.x//
x// N D Y; (4.50)
N and for
or, there are c > 0, ˛ 2 Œ0; 1Œ and > 0 such that for every x 2 S \ U .x/
every y 2 Q \ U .g.x//,
N
BY Dg.x/ .A.S; x/ \ cBX / T .Q; y/ C ˛BY :

Then the following relations hold:

N 1 .T .Q; g.x///
N \ Dg.x/
A .S; x/ N T S \ g 1 .Q/; xN ;

N 1 .C .Q; g.x///
N \ Dg.x/
C .S; x/ N C S \ g 1 .Q/; xN ;

N 1 .A .Q; g.x///
N \ Dg.x/
A .S; x/ N D A S \ g 1 .Q/; xN :
The following two corollaries of the above result are immediate:

Corollary 4.7.8. Let X and Y be Banach spaces, let Q Y be closed, let g W
X ! Y be a continuously differentiable around xN 2 g 1 .Q/. Assume that either Y
is finite-dimensional and
N C C.Q; g.x//
Im .Dg.x// N D Y;
or, there are c > 0, ˛ 2 Œ0; 1Œ and > 0 such that for every x 2 U .x/
N and for
every y 2 Q \ U .g.x//,
N
BY cDg.x/ .BX / C T .Q; y/ C ˛BY :

N 1 .T .Q; g.x///
Dg.x/ N D T g 1 .Q/; xN ;

N 1 .A .Q; g.x///
Dg.x/ N D A g 1 .Q/; xN ;

N 1 .C .Q; g.x///
Dg.x/ N C g 1 .Q/; xN :
Corollary 4.7.9. Let X be a Banach space, let S1 ; S2 X be closed sets, and let
xN 2 S1 \ S2 . Assume that either X is finite-dimensional and
N C.S2 ; x/
C.S1 ; x/ N DX
or there are c > 0, ˛ 2 Œ0; 1Œ and > 0 such that for every x 2 S1 \ U .x/
N and for
every y 2 S2 \ U .x/,
N
BY .A.S1 ; x/ \ cBX / T .S2 ; y/ C ˛BY :
N \ T .S2 ; x/
A .S1 ; x/ N T .S1 \ S2 ; x/
N ;
N \ C .S2 ; x/
C .S1 ; x/ N C .S1 \ S2 ; x/
N ;
N \ A .S2 ; x/
A .S1 ; x/ N D A .S1 \ S2 ; x/
N :
In finite dimensions, an inversion theorem can be given involving the basic

normal cone (see [596, Theorem 2.4]):
4.8 Expressions of the Contingent Cone on Some Constraint Sets 161
Theorem 4.7.10. Let S Rn be locally closed at xN 2 Rn and let g W Rn ! Rm be

strictly differentiable at xN 2 S \ g1 .0/. If
0 2 y Dg.x/
N C N.S; x/
N H) y D 0; (4.51)
then the following formulas hold:

N 1 .0/ D T S \ g 1 .0/; xN ;
N \ Dg.x/
T .S; x/

N 1 .0/ D A S \ g 1 .0/; xN :
N \ Dg.x/
A .S; x/
4.8 Expressions of the Contingent Cone on Some Constraint

Sets
In this section, we give explicit formulas for the computation of the contingent cone
on certain constraint sets. Similar results can be derived by using the inversion
theorem studied in the previous sections, however, by taking into account the
specific structure of the constraint sets, the conditions can be significantly relaxed.
Let X be a Hilbert space, let I1 D f1; 2; : : : ; mg and I2 D fm C 1; m C 2; : : : ; sg
be index sets, and let hi W X ! R with i 2 I1 [ I2 be maps defining the sets of
equality and inequality constraints by
Si D fx 2 X j hi .x/ D 0g for i 2 I1 ;
Sj D fx 2 X j hj .x/
0g for j 2 I2 ;
and, in a combined form,
\
s
˚
S WD Si D x 2 X j hi .x/ D 0; for i 2 I1 ; and hj .x/
0; for j 2 I2 :
i D1
(4.52)
We aim to give formulas for the contingent cone for the constraint set P \ S at
any xN 2 P \ S , where P X is an abstract constraint set. For this, we first derive
expressions for the contingent cone for S in terms of the sets Si with i 2 I1 [ I2 and
then extend the formulas to the set S \ P . This section is based on the interesting
work by Di [134] where more details can be found.
We will recall some notions of derivatives and auxiliary results. We begin with
the following:
Definition 4.8.1. Let X be a Hilbert space, and let g W X ! R be a given function.
N if there exists an
We say that the map g admits first-order Fréchet derivative at x,
element rg.x/N 2 X such that
g.xN C tx/ g.x/

N
lim D hrg.x/;
N xi; (4.53)
t !0 t
and the limit is uniform with respect to x in bounded sets.

By Riesz’s theorem, first-order Fréchet derivatives can be seen as a linear map
from X to R.
We now give a formula for the contingent cone to a single equality and inequality
constraint:
Proposition 4.8.2. Let X be a Hilbert space and let h W X ! R be continuous
around xN 2 X and differentiable at x.
N Assume that h.x/N D 0 and rh.x/
N ¤ 0.
Then:
1. For S D fx 2 X j h.x/ D 0g, we have
N D fx 2 X j hrh.x/;
T .S; x/ N xi D 0g:
2. For S D fx 2 X j h.x/
0g, we have
N D fx 2 X j hrh.x/;
T .S; x/ N xi
0g:
We also recall the following notion:

Definition 4.8.3. Let X be a Hilbert space, let S X be closed, and let xN 2 S .
The strict normal cone NO .x;
N S / of S at xN is the collection of all elements u 2 X
such that
lim sup hu; z xi

N 0: (4.54)
N x¤z2S
z!x; N
The strict normal cone is known to have the characterization given by:
NO .x;
N S / D T .S; x/N C.
We next recall the following alternative theorem:
Lemma 4.8.4. Let X be a Hilbert space and let A 2 L .Rm ; X /. Assume that 1
and 2 are two closed, convex cones in Rm such that 2 1 and int.2 / ¤ ;.
Then exactly one of the following systems has a solution:
1. AT Ax 2 int.2 / for x 2 1 such that jxj 1I
2. Ap D 0 2 X , for 0 ¤ p 2 2 .
Recall that an m-dimensional regular box in Rm is a nonempty, closed,
polyhedron with 2m vertices and 2m faces such that it is linearly homomorphic
to the standard box Œ0; 1m .
We now recall the following generalized mean-value theorem that uses the above
notion:
Lemma 4.8.5. Let g D .g1 ; : : : ; gm / be maps given on an open set containing an

m-dimensional regular box B in Rn . Let ˘i1 and ˘i2 be its i -th pair of opposite
sides. If for each i D 1; 2; : : : ; m, gi .x/ > ai for x 2 ˘11 and gi .x/ < ai for
x 2 ˘12 , then there exists z 2 int.B/ with g.z/ D a WD .a1 ; a2 ; : : : ; am /.
The following fundamental theorem on systems of linear equations will also be
used shortly:
Lemma 4.8.6. Let X be a Hilbert space, let x; ai 2 X and let bi 2 R, where
i 2 f0; 1; : : : ; sg. Then for an inequality ha0 ; xi b0 to be a consequence of a
consistent system
hai ; xi D bi for i 2 I1 hai ; xi bi for i 2 I2 ; (4.55)
it is necessary and sufficient to have the following condition: There exists

D
X
s X
s
.
1 ; : : : ;
s /, with
i
0 for i 2 I2 such that a0 D
i ai , and
i b i b 0 .
i D1 i D1
Proof. See Zeidler [617]. t

u
In order to compute the contingent cone to the constraint set S at xN 2 S , it is
convenient to divide the inequality constraints into active and non-active types as
follows:
I2D D fi 2 I2 j hi .x/
N D 0g;
I2> D fi 2 I2 j hi .x/
N > 0g:
We also define
WD f
2 Rs j
i
0; for i 2 I2D and
i D 0 for i 2 I2> g ; (4.56)
and note that
D f
2 Rs j
i D 0; for i 2 I1 and
i
0 for i 2 I2D g : (4.57)
Finally, we recall the following technical result (see [134]):

Lemma 4.8.7. Let v1 ; v2 ; : : : ; vs be elements in a Hilbert space X . If there is no
0 ¤
2 such that
X
s

i vi D 0; (4.58)
i D1
then there exists an m-dimensional regular box B in Rs , a linearly bijective image

of Œ0; 1m , with the j th pair of opposite sides ˘j1 and ˘j2 , such that for j 2 I1 , we
have
* +
X
s
vj ; ˛i vi > 0 for ˛ 2 ˘j1 ;
i D1
* +
X
s
vj ; ˛i vi < 0 for ˛ 2 ˘j2 ;
i D1
Ps
and for j 2 I2D , we have hvj ; i D1 ˛i vi i > 0, for ˛ in all sides of B, where
˛ D .˛1 ; ˛2 ; : : : ; ˛s / 2 Rs .
We are now ready to give the following formula, originally proved by Di [134]
(see [135, 136]).
Theorem 4.8.8. Let X be a Hilbert space, let S X be given by (4.52), and let
xN 2 S . Assume that the following Mangasarian-Fromovitz constraint qualification
holds: Let h D .h1 ; h2 ; : : : ; hs / be continuous around xN and Fréchet differentiable
Xs
N The only
2 , where is given by (4.56), such that
at x.
i rhi .x/
N D 0 is
i D1
D 0. Then the following formula holds:
\
s
˚
N D
T .S; x/ N D x 2 X j rhT .x/.x/
T .Si ; x/ N 2 D rh.x/
N 1
i D1
ˇ
ˇ hrhi .x/;
N xi D 0; for i 2 I1
D x 2 X ˇˇ ;
N xi
0; for i 2 I2D
hrhi .x/;
where is given by (4.57) and rh.x/

N 1 is an inverse image of rh.x/
N D
N : : : ; rhs .x//.
.rh1 .x/; N
\
s
Proof. It suffices to show that N T .S; x/,
T .Si ; x/ N as the converse inclusion
i D1
always holds and the other formulas follow from Proposition 4.8.2.
\s
Let x 2 N be arbitrary. Then, using Proposition 4.8.2, it follows that
T .Si ; x/
i D1
hrhi .x/;
N xi D 0 for i 2 I1 ;
hrhi .x/;
N xi
0 for i 2 I2D :
By the Mangasarian-Fromovitz constraint qualification and Lemma 4.8.7, there

exists
n an m-dimensional regular
o box B in Rs with m pairs of opposite sides
˘j ; ˘j j j D 1; : : : ; m such that for j 2 I1 , we have
1 2
* +
X
s
rhj .x/;
N ˛i rhi .x/
N >0 for ˛ 2 ˘j1 I (4.59a)
i D1
* +
X
s
rhj .x/;
N ˛i rhi .x/
N <0 for ˛ 2 ˘j2 ; (4.59b)
i D1
and for j 2 I2D , we have

* +
X
s
rhj .x/;
N ˛i rhi .x/
N > 0 for ˛ in all sides of the box B: (4.60)
i D1
For n 2 N, we define
!
1X
s
xn;t .˛/ D xN C t x C ˛i rhi .x/
N ; (4.61)
n i D1
N for any j 2 I1 or j 2
and by using the Fréchet differentiability of functions at x,
I2D , we have
* +
1X
s
hj .xn;t .˛// D hj .x/
N C t rhj .x/;
N xC ˛i rhi .x/
N
n i D1
!!
1X
s
Co t x C ˛i rhi .x/
N
n i D1
* +
1X
s
D t rhj .x/;
N xC ˛i rhi .x/
N C o.t/:
n i D1
Furthermore, for an inactive constraint, that is, for j 2 I2> or j such that
N > 0, it follows from the continuity of hj around xN that for sufficiently small
hj .x/
t, we have hj .xn;t .˛// > 0 for ˛ 2 B. Also from (4.60) and (4.61), we can choose
a sufficiently small t, denoted by tn , such that for all j 2 I1 , we have
hj .xn;tn .˛// > 0 for ˛ 2 ˘j1 ;

hj .xn;tn .˛// < 0 for ˛ 2 ˘j2 ;
and for all j 2 I2D , we have
hj .xn;tn .˛//
0 for ˛ 2 B:
By using Lemma 4.8.5, there exists ˛O n 2 B such that
hj .xn;tn .˛O n // D 0 for j 2 I1 (4.62a)

hj .xn;tn .˛O n //
0 for j 2 I2 ; (4.62b)
which implies that xn WD xn;tn .˛O n / 2 S and since tn1 .xn x/

N ! x, we deduce
that x 2 T .S; x/.
N t
u
Remark 4.8.9. Ledzewicz and Walczak [370] proved the above result in a Banach
N See
space setting by imposing an additional condition that h is Lipschitz around x.
also [371].
Since the strict normal cone is the polar of the contingent cone, we have the
following result:
Corollary 4.8.10. Under the assumptions of Theorem 4.8.8, we have
( )
X
s
O N S/ D
N .x;
i rhi .x/j
N
2 D rh.x/:
N
i D1
Having obtained expressions for the contingent cone on the set of equality and
inequality constraints, we now proceed to incorporate an abstract constraint set:
Theorem 4.8.11. Let X be a finite-dimensional Hilbert space, let P X be closed
and convex, let S X be given by (4.52), and xN 2 P \S . Assume that the following
constraint qualification holds: Let h D .h1 ; h2 ; : : : ; hs / be continuous around xN and
N Assume that there is no 0 ¤
2 , where is given by
Fréchet differentiable at x.
Xs
(4.56), such that N 2 NO .x;

i rhi .x/ N P /. Then:
i D1
\
s
T .S \ P; x/
N D N \ T .P; x/:
T .Si ; x/ N (4.63)
i D1
\
s
Proof. Let x 2 N \ T .P; x/
T .Si ; x/ N be arbitrary. Then there exist sequences
i D1
fxn g P and ftn g P such that tn # 0, zn D tn1 .xn x/
N ! x, and
N xi D 0
hrhi .x/; for i 2 I1
hrhi .x/;
N xi
0 for i 2 I2D :
We define ˇn WD kzn xk and note that ˇn ! 0 and xn D xN C tn x C

tn ˇn xQ n for some xQ n with jxQ n j 1. Similar to the proof of Theorem 4.8.8, there
exists an m-dimensional regular box B in Rs with m pairs of opposite sides
n o
˘j1 ; ˘j2 j j D 1; : : : ; m and with 2m vertices such that j 2 I1 , we have
* +
X
s
rhj .x/;
N ˛i rhi .x/
N >0 for ˛ 2 ˘j1 I (4.64a)
i D1
* +
X
s
rhj .x/;
N ˛i rhi .x/
N <0 for ˛ 2 ˘j2 ; (4.64b)
i D1
and for j 2 I2D , we have

* +
X
s
rhj .x/;
N ˛i rhi .x/
N > 0 for ˛ in all sides of the box B: (4.65)
i D1
We define
ˇ
ˇ N D 0 for j 2 I1
hw; rhj .x/i
W D w 2 X ˇˇ jwj 1; : (4.66)
N
0 for j 2 I2D
hw; rhj .x/i
Given n 2 N, ˛ 2 B, w 2 W , and t > 0, we define

!
1 X
s
1
xn;t .˛; w/ D xN C t tn x C ˇn 2
˛i rhi .x/
N C ˇn w ; 3
(4.67)
i D1
N for any j 2 I1 or
and by using the Fréchet differentiability of functions hi at x,
j 2 I2 , we have
* +
1 X
s
hj .xn;t .˛; w// D hj .x/
N C t tn hrhj .x/;
N xi C t tn ˇn rhj .x/;
N 2
˛i rhi .x/
N
i D1
ˇ ˇ!
ˇ 1 X
s ˇ
ˇ 1
ˇ
C t tn hrhj .x/;
N wi C o t tn ˇx C ˇn2 ˛i rhi .x/
N C ˇn3 wˇ ;
ˇ ˇ
i D1
which is positive for j 2 I2> , ˛ 2 B and w 2 W when t is small enough, say t sn ,

N > 0I and is positive for j 2 I2D , ˛ 2 B and w 2 W when t is small
because hj .x/
enough, say t sn , because hj .x/N D 0 and hrhj .x/;
N wi
0, and equals
* + ˇ ˇ!
X
s ˇ 1 X
s ˇ
1
ˇ 1
ˇ
2
N
t tn ˇn rhj .x/; ˛i rhi .x/
N C o t tn ˇx C ˇn
2
˛i rhi .x/
N C ˇn wˇ ;
3
ˇ ˇ
i D1 i D1
for j 2 I1 , ˛ 2 B, w 2 W and for all t, because hj .x/

N D 0, hrhj .x/;
N xi D 0 and
hrhj .x/;
N wi D 0: Therefore, for j 2 I1 , we have
hj .xn;sn .˛; w// > 0 for ˛ 2 ˘j1 ; and w 2 W; (4.68a)
hj .xn;sn .˛; w// < 0 for ˛ 2 ˘j2 and w 2 W; (4.68b)
and for all j 2 I2 , we have
hj .xn;sn .˛; w//

0 for ˛ 2 B; and w 2 W: (4.69)
Assuming that there exists an ˛Q 2 B such that the compact, convex set
Q D fxn;sn .˛; w//j w 2 W g

Sn;sn .˛/
is disjoint with P , then by the separation theorem, we know that there exists a
Q from P , that is,
hyperplane strongly separating Sn;sn .˛/
hun ; yi hun ; xn;sn .˛;

Q w/i; (4.70)
for every w 2 W and y 2 P where un ¤ 0. Without any loss of generality, we

assume that kun k D 1 and un ! u. This means that in the limit, hu; y xi N 0 for
every y 2 P , confirming that u 2 NO .x; N P /. On the other hand, given w 2 W such
that kwk D 1, we define zn .sn / WD sn xn C .1 sn /x,
N vn .w/ WD xn;sn .˛;
Q w/ zn .sn /,
and get zn .sn / D xN C sn .xn x/
N D xN C sn tn vn with
!
1 X
s
1
vn .w/ D sn tn x vn C ˇk 2
˛Q i rh.x/
N C ˇn w
3
i D1
and by (4.70), we have

vn .w/
un ;
0; (4.71)
jvn .w/j
Since kx vn k D ˇn , we deduce that kvvnn .w/

.w/k ! 0. Therefore, by taking the limit
in (4.71), we have hu; wi
0, implying that u 2 W . By employing Lemma 4.8.6,
X s
there exists a nonzero vector
2 such that N D u 2 NO .x;

i rhi .x/ N P /, which,
i D1
however, contradicts the imposed constraint qualification, and hence for any ˛ 2 B,
we must have Sn;sn \ P ¤ ;. Therefore, for any vertex ˛i 2 B there exists wi 2 W
such that zn;i D xn;sn .˛i ; wi / 2 P . Now, zn;1 ; : : : zn;2m form a new m-dimensional
regular box B.Q Therefore, we apply the generalized mean value theorem to BQ which
is a subset of P since P is convex. From (4.69) and (4.70), we can apply generalized
mean value theorem to B. Q Consequently, there exists xO n 2 BQ \ P such that
hj .xO n / D 0 for j 2 I1 and hj .xO n /

0; forj 2 I2 ;
4.9 Second-Order Tangent Sets 169
that is, xO n 2 P \ S . It can be shown that .tn sn /1 .xO n x/

N ! x which confirms that
x 2 T .P \ S; x/ N because xO n 2 Bn . Since the converse inclusion is always true, the
u
Since the Clarke normal cone and the strict normal cone on a convex set P
coincide, we get:
Corollary 4.8.12. Under the assumptions of Theorem 4.8.11, we have
NO .x; N C NO .x;
N P \ S / D rh.x/ N P /: (4.72)
Proof. We will show that the right side is closed. Assume that zn WD xn C yn with
xn 2 rh.x/ N and yn 2 NO .x;N P /, and zn ! zN. Since rh.x/ N is a closed
set in a finite-dimensional subspace, if fxn g is bounded, then we can assume that
xn ! xQ 2 rh.x/, N and hence yn D zn xn ! zN xQ 2 NQ .x; N P /, that is, zN
belong to the right side of the equality. If xn is not bounded, then we can assume
xn
that kxn k ! 1 and ! xQ ¤ 0. Therefore, the limit of
kxn k
zn xn yn
D C
kxn k kxn k kxn k
contradicts the basic constraint qualification and the proof is complete. t

u
4.9 Second-Order Tangent Sets
In this section, we study second-order tangent sets, which are natural extensions
of some of the tangent cones of first-order investigated in the previous sections. In
set-valued optimization, second-order derivatives and second-order epiderivatives
of set-valued maps are defined by using second-order tangent sets and second-order
tangent cones. These second-order notions are also fundamental in giving second-
order optimality conditions and second-order sensitivity analysis, among others. We
note that Hoffmann and Kornstaed [265] were among the first ones to conduct a
systematic study of second-order local approximation (see also [192]). The seminal
contributions of Penot [465, 466, 468] and Ward [594, 596]) and the monographs by
Bonnans and Shapiro [53] and Rockafellar and Wets [499] serves as an excellent
introduction to this subject. The survey article by Giorgi, Jiménez, and Novo [209]
provided a good guideline for arranging the properties of the second-order tangent
sets. Some arguments that we use can also be found in Castellani [84], Castellani
and Pappalardo [86]. See also [104–107].
4.9.1 Second-Order Radial Tangent Set and Second-Order

Feasible Tangent Set
In this subsection, we focus on the following two notions of second-order tangent

sets:
N uN / 2
Definition 4.9.1. Let X be a normed space, let S be a subset of X , and let .x;
X X.
1. The second-order radial tangent set R2 .S; x; N uN / of S at .x;
N uN / is the set of
all x 2 X for which there exists a sequence ftn g P with tn # 0 such that
xN C tn uN C tn2 x=2 2 S for every n 2 N.
2. The second-order feasible tangent set IR2 .S; x; N uN / of S at .x;
all x 2 X such that for every ftn g P with tn # 0, there is an m 2 N satisfying
that xN C tn uN C tn2 x=2 2 S for every n
m.
An equivalent formulation of the above two notions is as follows:
N uN / D fx 2 X j 8 t > 0 9 s 2 .0; t/ such that xN C s uN C s 2 x=2 2 S g;

R2 .S; x;
N uN / D fx 2 X j 9 t > 0 8 s 2 .0; t/ we have xN C s uN C s 2 x=2 2 S g:
IR2 .S; x;
The following theorem collects some properties of the second-order radial

tangent set and the second-order feasible tangent set.
Theorem 4.9.2. Let X be a normed space, let S; S1 ; S2 be nonempty subsets of X ,
let uN 2 X , and let xN 2 cl S \ cl S1 \ cl S2 . Then the following properties hold:
1. N 0/ D R.S; x/
R2 .S; x; N
2. N 0/ D IR.S; x/
IR2 .S; x; N
3. N 0/ R2 .S; x;
IR2 .S; x; N 0/
4. IR2 .X nS; x;
N uN / D X nR2 .S; x; N uN /
5. 0 2 IR2 .S; x;
N uN / ” uN 2 IR.S; x/ N
6. 0 2 R2 .S; x;
N uN / ” uN 2 R.S; x/ N
7. S1 S2 H) R2 .S1 ; x; N uN / R2 .S2 ; x;N uN /
8. S1 S2 H) IR .S1 ; x; 2
N uN / IR2 .S2 ; x;
N uN /
9. R2 .S1 [ S2 ; x;
N uN / D R2 .S1 ; x; N uN / [ R2 .S1 ; x;
N uN /
10. IR2 .S1 \ S2 ; x;
N uN / D IR2 .S1 ; x;N uN / \ IR2 .S1 ; x;
N uN /.
Proof. The proof is based on arguments similar to the one given for the first-order
analogues. t
u
4.9.2 Second-Order Contingent Set and Second-Order

Interiorly Contingent Set
Our focus is on the following notions of second-order tangent sets.

N uN / 2
X X.
1. The second-order contingent set T 2 .S; x; N uN / of S at .x;N uN / is the set of all
x 2 X for which there are a sequence ftn g P with tn # 0 and a sequence
fxn g X with xn ! x such that xN C tn uN C tn2 xn =2 2 S for every n 2 N.
2. The second-order interiorly contingent set IT 2 .S; x; N uN / of S at .x; N uN / is the
set of all x 2 X such that for every sequence ftn g P with tn # 0 and every
sequence fxn g X with xn ! x there exists an index m 2 N satisfying that
xN C tn uN C tn2 xn =2 2 S for every n
m.
An equivalent formulation of the above two notions reads:
N uN / D fx 2 X j 8 U.x/ 8 t > 0 9 s 2 .0; t/ 9 z 2 U.x/ with

T 2 .S; x;
xN C t uN C t 2 z=2 2 S g;
N uN / D fx 2 X j 9 U.x/ 9 t > 0 8 s 2 .0; t/ 8 z 2 U.x/ we have
IT 2 .S; x;
xN C s uN C s 2 z=2 2 S g:
We now give other equivalent characterizations of the second-order contingent

set.
Theorem 4.9.4. Let X be a normed space, let S be subset of X , let uN 2 X , and
N uN / D Ti2 .S; x;
let xN 2 cl S . Then, for i 2 f1; : : : ; 5g, we have T 2 .S; x; N uN /, where
2
Ti .S; x;N uN / are given as follows:
( )
x n xN tn N
u
N uN / D x 2 X j 9 ftn g # 0 9 fxn g S such that
T12 .S; x; 1 2
!x ,
˚ 2 tn
N uN / Dn x 2 X j 9 ftn g # 0 such that d.xN C tn uN C tn2 x=2; S / D o.tn2 / ,o
T22 .S; x;
T32 .S; x;
N uN / D x 2 Xj 9 fxn g S 9 ftn g # 0 such that xn D xN C tn uN C tn2 x=2 C o.tn2 / ,

d.xN C t uN C t 2 x=2; S /
N uN / D x 2 X j lim inf
T42 .S; x; D 0 ,
t #0 t2
S xN t uN
N uN / D lim sup
T52 .S; x; 1 2
.
t #0 2t
Proof. The proof is based on standard arguments and is omitted. t

u
The following result collects some properties of the second-order contingent sets
and the second-order interiorly contingent sets. (In the proofs, we drop the factor
1=2, see Sect. 4.10.)
Theorem 4.9.5. Let X be a normed space, let S; S1 ; S2 be nonempty subsets of X ,
let uN 2 X , and let xN 2 cl S \ cl S1 \ cl S2 . Then the following properties hold:
N 0/ D T .S; x/
1. T 2 .S; x; N
N 0/ D IT.S; x/
2. IT 2 .S; x; N
3. T 2 .X nS; x; N uN / D X nIT 2 .S; x; N uN /

4. 0 2 T .S; x;
2
N uN / H) uN 2 T .S; x/ N
5. 0 2 IT 2 .S; x; N uN / (H uN 2 IT.S; x/ N
6. N uN / is closed
T 2 .S; x;
7. N uN / D T 2 .cl S; x;
T 2 .S; x; N uN /
8. 2
IT .S; x; N uN / is open
9. N uN / D IT 2 .int S; x;
IT 2 .S; x; N uN /
10. 2
T .S; x; N uN / ¤ ; H) uN 2 T .S; x/ N and hence uN 62 T .S; x/
N H)
N uN / D ;
T 2 .S; x;
11. N uN / ¤ X H) uN 62 IT.S; x/
IT 2 .S; x; N and hence uN 2 IT.S; x/
N H)
N uN / D X
IT 2 .S; x;
12. N uN / IR2 .S; x;
IT 2 .S; x; N uN / R2 .S; x; N uN / T 2 .S; x;N uN /
13. S1 S2 H) T .S1 ; x; 2
N uN / T .S2 ; x;
2
N uN /
14. S1 S2 H) IT 2 .S1 ; x; N uN / IT 2 .S2 ; x; N uN /
15. T .S1 [ S2 ; x;
2
N uN / D T .S1 ; x;
2
N uN / [ T 2 .S2 ; x;
N uN /
16. IT 2 .S1 \ S2 ; x; N uN / D IT 2 .S1 ; x;
N uN / \ IT 2 .S2 ; x;N uN /
17. N uN / C t uN T 2 .S; x;
T 2 .S; x; N uN /; for every t 2 R
18. N uN / C t uN IT 2 .S; x;
IT 2 .S; x; N uN /; for every t 2 R.
Proof. The proof is based on arguments similar to the one given for the first-order
analogues. For the sake of completeness, we give proofs of a few properties. The
first five properties are entirely analogous to the first-order case. We note that
the closedness of T 2 .S; x; N uN / follows from the openness of IT 2 .S; x; N uN / and the
duality relationship between them. To prove that IT 2 .S; x; N uN / is open, we choose
an arbitrary x 2 IT 2 .S; x; N uN /. Let xO 2 U" .x/ with " > 0. Due to the containment
x 2 IT 2 .S; x; N uN /, there exists t > 0 such that for each s 2 .0; t/ and for each
xQ 2 Ut .x/, we have xN C s uN C s 2 xQ 2 S . For " < t, we choose "N < t " which
implies that U"N .x/O Ut .x/ and for each s 2 .0; "N/, we have xN C s uN C s 2 U"N .x/O S.
This confirms that U"N .x/ IT 2 .S; x; N uN /.
It is evident that if uN … T .S; x/ N then T 2 .S; x;N uN / D ;. This fact further implies
that if uN 2 IT.S; x/ N then IT .S; x;
2
N uN / D X . In fact, this follows from the chain of
implications that uN 2 IT.S; x/ N implies uN … T .X nS; x/ N which further implies that
T 2 .X nS; x;N uN / D ; and hence IT 2 .S; x; N uN / D X . Proofs of some the properties can
also be deduced from the proof of Theorem 4.10.3 t
u
The following example shows that the inclusions stated above are, in general,
strict.
SS S
Example 4.9.6. For S R2 given by S D f.x; y/j y > 0g n2N f. n ; 0/g
1
f.0; 0/g, we have
IT 2 .S; .1; 0/; .0; 0// D int S I

R2 .S; .1; 0/; .0; 0// D int S [ f.0; 0/gI
IR2 .S; .1; 0/; .0; 0// D int S [ f.x; 0/j x

0gI
T 2 .S; .1; 0/; .0; 0// D cl S:
We give a simple example (see [53]).

Example 4.9.7. For the subsets S1 and S2 given by
1
[ .2nC1/ 2n [
S1 WD 2 ;2 f0gI
nD0
1
[ 2n .2n1/ [
S2 WD 2 ;2 f0g;
nD0
we have S WD S1 [ S2 D Œ0; 1 and T 2 .S1 ; 0; 1/ D T 2 .S2 ; 0; 1/ D T 2 .S; 0; 1/ D R.
In contrast with first-order tangent cones, the second-order analogues given here,
are not, in general, cones or even connected sets (see Example 4.10.8) and they
could be empty.
We conclude this subsection by the following simple example (see [53]):
Example 4.9.8. Let X be a normed space and let S X be closed and convex.
Then
8 2
N h/ R; if k > 0
< T .S; x;
T 2 .S RC ; .x;
N 0/; .h; k// D T 2 .S; x;
N h/ RC ; if k D 0
:
;; if k < 0:
4.9.3 Second-Order Adjacent Set and Second-Order Interiorly

Adjacent Set
Our focus here is on the following two notions of second-order tangent sets.
N uN / 2
X X.
1. The second-order adjacent set A2 .S; x; N uN / of S at .x;
N uN / is the set of all x 2 X
such that for every sequence ftn g P with tn # 0, there exists a sequence
fxn g S with xn ! x satisfying that xN C tn uN C tn2 xn =2 2 S for every n 2 N.
2. The second-order interiorly adjacent set IA2 .S; x; N uN / of S at .x;N uN / is the set
of all x 2 X such that there exists a sequence ftn g P with tn # 0 so that for
every sequence fxn g S with xn ! x there exists an index m 2 N satisfying
that xN C tn uN C tn2 xn =2 2 S for every n
m.
An equivalent formulation of the above second-order tangent sets reads:
N uN / D fx 2 X j 8 U.x/ 9 t > 0 8 s 2 .0; t/ 9 z 2 U.x/ with

A2 .S; x;
xN C t uN C t 2 z=2 2 S g;
N uN / D fx 2 X j 9 U.x/ 8t > 0 9 s 2 .0; t/ 8 z 2 U.x/ we have
IA2 .S; x;
xN C s uN C s 2 z=2 2 S g:
We now give other equivalent characterizations of the second-order adjacent set.

Theorem 4.9.10. Let X be a normed space, let S be nonempty subset of X , let
uN 2 X , and let xN 2 cl S . Then for i 2 f1; 2; 3g, we have A2 .S; x; N uN / D A2i .S; x;
N uN /,
2
where Ai .S; x; N uN / are given as follows:
( )
xn xN tn N
u
N uN / D x 2 X j 8 ftn g # 0 9 fxn g S such that
A21 .S; x; 1 2
!x
t
( )2 n
d.xN C t uN C t 2 x=2; S /
N uN / D x 2 X j lim sup
A22 .S; x; D0
t #0 t2
S xN t uN
N uN / D lim inf
A23 .S; x; 1 2
.
t #0 t
2
Proof. The proof is based on standard arguments and hence omitted. t
u
The following result collects some properties of the second-order adjacent sets
and second-order interiorly adjacent sets.
Theorem 4.9.11. Let X be a normed space, let S; S1 ; S2 be nonempty subsets of
X , let uN 2 X , and let xN 2 cl S \ cl S1 \ cl S2 . Then the following properties hold:
1. N 0/ D A.S; x/
A2 .S; x; N
2. N 0/ D IA.S; x/
IA2 .S; x; N
3. A2 .X nS; x; N uN / D X nIA2 .S; x; N uN /
4. 0 2 A2 .S; x; N uN / H) uN 2 A.S; x/ N
5. 0 2 IA2 .S; x; N uN / (H uN 2 IA.S; x/ N
6. N uN / is closed
A2 .S; x;
7. N uN / D A2 .cl S; x;
A2 .S; x; N uN /
8. 2
IA .S; x;N uN / is open
9. N uN / D IA2 .int S; x;
IA2 .S; x; N uN /
10. 2
A .S; x;N uN / ¤ ; H) uN 2 A.S; x/ N and hence uN 62 A.S; x/
N H)
N uN / D ;
A2 .S; x;
11. N uN / ¤ X H) uN 62 IA.S; x/
IA2 .S; x; N and hence uN 2 IA.S; x/
N H)
N uN / D X
IA2 .S; x;
12. N uN / IR2 .S; x;
IT 2 .S; x; N uN / A2 .S; x; N uN / T 2 .S; x;
N uN /
13. 2
IT .S; x;N uN / IA .S; x;
2
N uN / R .S; x;
2
N uN / T 2 .S; x;
N uN /
14. S1 S2 H) A2 .S1 ; x; N uN / A2 .S2 ; x; N uN /
15. S1 S2 H) IA2 .S1 ; x; N uN / IA2 .S2 ; x; N uN /
4.10 Generalized Second-Order Tangent Sets 175
16. IA2 .S1 [ S2 ; x;

N uN / D IA2 .S1 ; x;
N uN / [ IA2 .S2 ; x;
N uN /
17. A .S1 [ S2 ; x;
2
N uN / A .S1 ; x;
2
N uN / [ A2 .S2 ; x;
N uN /
18. N uN / C t uN A2 .S; x;
A2 .S; x; N uN /; for every t 2 R
19. 2
N uN / C t uN IA2 .S; x;
IA .S; x; N uN /; for every t 2 R.
The following is an analogue of Example 4.9.7 (see [53]).
Example 4.9.12. For the subsets S1 and S2 given by
1
[ [
S1 WD 22nC1 ; 22n f0gI
nD0
1
[ [
S2 WD 2.2n1/ ; 22n f0g;
nD0
we have S WD S1 [ S2 D Œ0; 1, A2 .S1 ; 0; 1/ D A2 .S2 ; 0; 1/ D ;, and A2 .S; 0; 1/ D

R, showing that the inclusion A2 .S1 [ S2 ; x;
N uN / A2 .S1 ; x;
N uN / [ A2 .S2 ; x;
N uN / is, in
general, strict.
The following example, given by Penot [467], shows that T 2 .; ; / is larger than
A .; ; /.
2
Example 4.9.13. Let frn g P be a sequence such that for some c 2 .0; 1/,
rnC1 crn , for each n 2 N and let S R2 be given by S WD
f.x; y/j y
.rn C rnC1 x rn rnC1 for every n 2 N/g. Then, it turns out that
.0; 2/ 2 T 2 .S; .0; 0/; .1; 0//, however, .0; 2/ 62 A2 .S; .0; 0/; .1; 0//.
4.10 Generalized Second-Order Tangent Sets
In this section, we study an extension of the second-order contingent sets proposed

by Cambini, Martein, and Vlach [82]. This generalization not only gives an
interesting insight into the second-order contingent set but it has also served as a
motivation to introduce the important notion of second-order asymptotic tangent
cones which will be studied in the next section. In this section, which is based on
[82], we mainly focus on the generalized contingent sets of second-order, a notion
that corresponds to the second-order contingent set. However, analogues of other
second-order tangent sets can also be defined and studied.
For later use, we formulate the following condition: Let f˛n g P, fˇn g P,
and fxn g X be arbitrary sequences and k be a nonnegative number satisfying
ˇn
˛n ! 1; ˇn ! 1; ! k; (4.73a)
˛n
xn ! xN ˛n .xn x/
N ! uN ˇn Œ˛n .xn x/
N uN ! x: (4.73b)
We begin with the following:

Definition 4.10.1. Let X be a normed space, let S X , let uN 2 X , let xN 2 S , and
let k
0.
1. The second-order indexed contingent set Tk2 .S; x; N uN / of S at .x;
all x such that there are sequences f˛n g P, fˇn g P, and fxn g S satisfying
(4.73). The generalized second-order contingent set TO 2 .S; x; N uN / of S at .x;
N uN /
is then defined by
[
TO 2 .S; x;
N uN / WD N uN /:
Tk2 .S; x;
k>0
2. The second-order indexed adjacent set A2k .S; x; N uN / of S at .x;

all x such that for all sequences f˛n g P and fˇn g P satisfying (4.73a), there
exists fxn g S satisfying (4.73b). The generalized second-order adjacent set
AO2 .S; x;
N uN / of S at .x;
N uN / is then defined by
[
AO2 .S; x;
N uN / WD N uN /:
A2k .S; x;
k>0
Remark 4.10.2. We emphasize that in the above definition, the sequences f˛n g and
fˇn g depend on k. Furthermore, to define the second-order indexed contingent set
T12
.S; x; N uN /, we replace the requirement that ˇ˛nn ! 1 by ˇ˛nn ! 0.
N uN / of S at .x;
The same argument applies to other sets.
We now give basic properties of the second-order indexed contingent set and
the generalized second-order contingent set. Similar results for other generalized
adjacent sets can be given.
Theorem 4.10.3. Let X be a normed space, let S be a subset of X , let uN 2 X , and
let xN 2 S . Then:
1. For each k > 0, we have
2 2
N uN / D
T 2 .S; x; N uN /:
T .S; x;
k k
2. Let t 2 R and k
0 be arbitrary. Then x C t uN 2 Tk2 .S; x; N uN /, whenever x 2
N uN /.
Tk2 .S; x;
3. Assume that X is finite-dimensional. Then 0 2 T02 .S; x; N uN / Moreover,
N uN / D ; implies that there exists uN 2 T02 .S; x;
T 2 .S; x; N uN / such that uN ¤ 0
and hNu; d i D 0.
4. T12
N uN / ¤ ; if and only if 0 2 T 2 .S; x;
.S; x; N uN /.
Proof. We begin with a proof of the first part. Let x 2 T 2 .S; x; N uN / be arbitrary.
Then there exist sequences f n g P and fzn g X such that n # 0, zn ! x, and
xN C n uN C 12 n2 zn 2 S for every n 2 N. Let k > 0 be a given number. We define
˛n WD 1n , ˇn WD kn , and xn WD xN C n uN C 12 n2 zn , and note that the sequences

f˛n g P and fˇn g P satisfy (4.73a). Furthermore,

1 1 1
˛n .xn x/
N D n uN C n2 zn D uN C n zn ! uN ;
n 2 2

k 1 k k
ˇn Œ˛n .xn x/
N uN D uN C n zn uN D zn ! x;
n 2 2 2
which means that k2 x 2 Tk2 .S; x; N uN / and hence proving the inclusion T 2 .S; x; N uN /
2 2
T
k k .S; N
x; uN /.
For the converse, let z 2 k2 Tk2 .S; x; N uN /, where k > 0, be arbitrary. Then there
2
exists x 2 Tk2 .S; x; N uN / with z D x confirming that there exist sequences f˛n g P,
k
fˇn g P, and fxn g S such that (4.73) holds. We define sequences f n g P and
fzn g P by n WD ˛1n , and zn WD 2˛n Œ˛n .xn x/ N uN , and note that f n g P, n #
0 and xN C n uN C 12 n2 zn DW xn 2 S . Since ˇ˛nn ! k1 , and ˇn Œ˛n .xn x/ N uN ! x, we
obtain zn D ˇn Œ˛n .xn x/
2˛n
N uN ! k x D z, giving x 2 T .S; x;
2 2
N uN /. Therefore,
2
T .S; x;N uN / k Tk .S; x;
2 2
N uN / and the proof of the first part is complete.
Let us now proceed to prove the second part. For an arbitrary k > 0 and t 2 R,
let x 2 Tk2 .S; x; N uN /. Then there are sequences f˛n g P, fˇn g P, and fxn g S
such that (4.73) holds. We define sequences f˛O n g, fˇOn g and fxO n g by ˛O n WD ˛n C kt ,
ˇOn WD ˇn , and xO n WD xn , and note that, without any loss of generality, we can
assume that f˛O n g P and fˇOn g P, as they are positive for sufficiently large n.
O
Clearly, ˛O n ! 1, ˇOn ! 1, ˇn ! k, and xO n ! x.
˛O n
N Furthermore, we have

t
˛O n .xO n x/
N D ˛n .xn x/
N C .xn x/
N ! uN ;
k
and

O t
ˇn Œ˛O n .xO n x/
N uN D ˇn ˛n .xn x/N uN C .xn x/N
k
t
D ˇn Œ˛n .xn x/
N uN C ˇn .xn x/;
N
k
and by ˇn .xn x/ N ! k uN , we get ˇOn Œ˛O n .xO n x/

N D ˇ˛nn ˛n .xn x/ N uN ! .x C t uN /
proving x C t uN 2 Tk .S; x;
2
N uN /.
We now take k D 0. Then for x 2 T02 .S; x; N uN /, there are sequences f˛n g P,
fˇn g P, and fxn g S such that (4.73) holds with k D 0. We define sequences
f˛O n g, fˇOn g, and fxO n g by ˛O n WD ˛n C tˇ˛nn , ˇOn WD ˇn , xO n WD xn , and assume,
without any loss of generality, that f˛O n g P and fˇOn g P as they are positive
O
for sufficiently large n. Clearly, ˛O n ! 1, ˇOn ! 1, ˇn ! 0, and xO n ! x,
˛O n
N and
t˛n
˛O n .xO n x/
N D ˛n .xn x/
N C .xn x/;
N
ˇn
ˇOn Œ˛O n .xO n x/
N uN D ˇn Œ˛n .xn x/
N uN C t˛n .xn x/;
N
t
which, due to the fact that ! 0, implies that ˛O n .xO n x/ N ! uN and
ˇn
ˇOn Œ˛O n .xO n x/
N uN ! .x C t uN /, and hence x C t uN 2 T02 .S; x;
N uN /. This completes
the proof of the second part.
For the third part, we assume that we assume that X is finite-dimensional. Let
T 2 .S; x; N uN / ¤ ;. Then for x 2 T 2 .S; x;N uN /, there are sequences f˛n g P, fˇn g
P, and fxn g S such that
ˇn
!2 ˛n .xn x/
N ! uN ˇn Œ˛n .xn x/
N uN ! x:
˛n
p
We define f˛O n g P, fˇOn g P, and fxO n g S by ˛O n WD ˛n , ˇOn WD ˛ n , and
xO n WD xn and note that
ˇOn p
!0 ˛O n .xO n x/
N ! uN ˇOn Œ˛O n .xO n xN uN / D ˛n Œ˛n .xn x/
N uN ! 0;
˛O n
which confirms that 0 2 T02 .S; x; N uN /.

Now assume that T 2 .S; x; N uN / D ;. Then, for an arbitrary 0 ¤ uN 2 T .S; x/, N
there are sequences f n g P with n ! 1 and fxn g S with xn ¤ xN such
that n .xn x/N ! uN . For ˛n WD kxkN uk
N , we have ˛n .xn x/
n xk
N ! uN . We now define
zn WD ˛n .xn x/
N uN and ˇn WD kzn k . Without loss of generality, we may assume that
1
the sequence 2˛n zn is not bounded because otherwise we could find its convergent
sequences. The limit of such a subsequence would belong to T 2 .S; x; N uN /, which
would contradict our assumption. It follows that, without loss of generality, we may
assume that ˛n kzn k ! 1. Therefore, ˇ˛nn D ˛n kz 1
nk
! 0, and we conclude that
ˇn zn D kzn k has a subsequence that converges to some z 2 T02 .S; x;
zn
N uN / with kzk D
1. From the identity zn C uN D ˛n .xn x/, N we have kzn k CkNuk C2hzn ; uN i D ˛n2 kxn
2 2
N D kNuk . Therefore, kzn k C 2hzn ; uN i D 0 and taking the limit appropriately, we

xk 2 2
obtain hz; uN i D 0. (See also [209] for the alternative arguments given for this proof.)
It remains to show that 0 2 T02 .S; x; N uN / for the present case. For this, we define,
p
ˇOn WD ˇn , ˛O n WD ˛n , xO n WD xn . Then,
p p
ˇOn ˇn 1 kzn k
D D p D !0
˛O n ˛n ˛n zn ˛n kzn k
p ˇn zn
ˇOn Œ˛O n .xn x/
N uN D ˇn zn D p ! 0:
ˇn
N uN /. It
Let us now prove the last part. We begin with assuming that 0 2 T 2 .S; x;
follows that there exists f˛n g P, fˇn g P, and fxn g S such that
ˇn
˛n ! 1 ˇn ! 1 ! 2 ˛n .xn x/
N ! uN ˇn .˛n .xn x/
N uN / ! 0:
˛n
N uN , ˇOn WD
Define zn WD ˛n .xn x/ 1
kzn k
, and ˛O n WD ˛n xO n D xn . Since ˛n zn D
Ǒn
˛n
ˇ z
ˇn n n
! 0, we have D ˛n kz
˛O n
1
nk
! 1, and hence from ˇOn Œ˛O n .xO n / uN D
kzn k ! z, we conclude that z 2 T1 N uN /.
zn 2
.S; x;
Now let z 2 T1 .S; x;
2
N uN /. Then there are sequences f˛n g P, fˇn g P and
fxn g S such that
ˇn
˛n ! 1 ˇn ! 1 ! 1 ˛n .xn x/
N ! uN ˇn .zn / ! z;
˛n
ˇn
where zn D ˛n .xn x/
N uN . Clearly, ˇn zn D 2˛n .2˛n zn / ! z. Furthermore, since
ˇn
2˛n ! 1, we obtain that 2˛n zn ! 0 2 T .S; x;
N uN /. The proof is complete.
2
t
u
As a consequence of the above result, we have:
Corollary 4.10.4. Let X be a normed space, let S be a subset of X , let uN 2 X , and
let xN 2 S . Then:
N uN / D T 2 .S; x;
T22 .S; x; N uN /;
N uN / TO 2 .S; x;
T 2 .S; x; N uN /;
[k
TO 2 .S; x;
N uN / D N uN /:
T 2 .S; x;
2
k>0
In the following, we give a few examples to illustrate the notion of the gener-
alized contingent sets. Our first example shows that the inclusion T 2 .S; x; N uN /
TO 2 .S; x;
N uN / is, in general, strict.
˚
Example 4.10.5. For X D R2 , we define S D .x1 ; x2 / 2 R2 j x2 D x12 . We
set xN D .0; 0/, and uN D .1; 0/. For x 2 T 2 .S; x; N uN /, there are sequences
f n g P and fxn g X with xn D .xn1 ; xn2 / such that n ! 0, xn ! x,
and xN C n uN C 12 n2 xn 2 S , where the containment xN C n uN C 12 n2 xn 2 S
2 2
implies that 12 n2 xn2 D n C 12 n2 xn1 , and hence xn2 D 2 1 C 12 n xn1 ! 2,
Fig. 4.9 Contingent sets in

Example 4.10.5
Fig. 4.10 Contingent sets in

Example 4.10.7
because
˚ fxn1 g is convergent and
n ! 0. We have shown that T .S; x;
2
N uN /
.x1 ; x2 / 2 R j x2 D 2; x1 2 R . To complete the argument, it suffices to show that
2
x D .x1 ; x2 / D .0; 1/ belongs to TO 2 .S; x;

N uN /. For this, let f˛n g P be an arbitrary
sequence such that ˛n ! 1. Set ˇn D ˛n and define fxn g R2 by xn1 D ˛1n , and
xn2 D ˛12 , so that fxn g S with xn ! x. N Moreover,
n
˛n .xn1 xN 1 / D ˛n xn1 D 1;
1
˛n .xn2 xN 2 / D ˛n xn2 D ;
˛n
ˇn Œ˛n .xn1 xN 1 / uN 1 D ˇn 0 D 0;
ˇn
ˇn Œ˛n .xn2 xN 2 / uN 2 D 0 D 1;
˛n
4.11 Second-Order Asymptotic Tangent Cones 181
N uN / TO 2 .S; x;
confirming that x 2 T12 .S; x; N uN /. In fact, it can shown that
N uN / D f.x1 ; 2/j x1 2 Rg;

T 2 .S; x;
N uN / D f.x1 ; k/j x1 2 Rg;
Tk2 .S; x;
N uN / D f.x1 ; 0/j x1 2 Rg;
T02 .S; x;
2
T1 N uN / D ;:
.S; x;
p
Example 4.10.6. For S WD f.x1 ; x2 / 2 R2 j x2 D x1 x1 ; x1
0g, we set xN D
.0; 0/, uN D .1; 0/, and k > 0. Then,
N uN / D ;;
T 2 .S; x;
N uN / D ;;
Tk2 .S; x;
N uN / D f.x1 ; x2 /j x2
0g;
T02 .S; x;
2
T1 N uN / D ;:
.S; x;
Example 4.10.7. For S WD f.x1 ; x2 / 2 R2 j x2 D x14 g, we set xN D .0; 0/, uN D .1; 0/,
and k > 0. Then,
N uN / D f.x1 ; x2 /j x2 D 0g;
T 2 .S; x;
N uN / D f.x1 ; x2 /j x2 D 0g;
Tk2 .S; x;
N uN / D f.x1 ; x2 /j x2 D 0g;
T02 .S; x;
2
T1 N uN / D f.x1 ; x2 /j x2
0g:
.S; x;
The next example shows that the second-order contingent set is not even a
connected set:

Example
2 4.10.8.
Define a set S R2 by S WD f.x; y/j y
x 2 [
x
y for x < 0g. Then for xN D .0; 0/, we have T .S; .0; 0// D R2 and
for uN ˛ D .˛; 0/ with ˛ 2 RC, we have

N uN ˛ / D f.x; y/j y
˛ 2 [ y
˛ 2 g:
T12 .S; x;
4.11 Second-Order Asymptotic Tangent Cones
All the second-order tangent sets studied in the previous section are, in general, not
cones. In this section, we collect a related notion of second-order asymptotic tangent
cones, which, as evident from the name, are always cones. We remark that probably,
Fig. 4.11 Non-connected

contingent set of
second-order for
Example 4.10.8
Ledzewicz and Schaettler [364] were the first ones to introduce second-order notions
of local approximation which were actually cones. In fact, there are a few other
authors who introduced second-order conical notions of local approximation. Some
of these notions will be studied in later part of this chapter.
4.11.1 Second-Order Asymptotic Feasible Tangent Cone

and Second-Order Asymptotic Radial Tangent Cone

N uN / 2
X X.
1. The second-order asymptotic radial tangent cone R2 .S; x; N uN / of S at .x;
N uN /
is the set of all x 2 X for which there exist sequences f.sn ; tn /g P P with
sn
.sn ; tn / # 0 and # 0 such that xN C sn uN C sn tn x 2 S for every n 2 N.
tn
2. The second-order asymptotic feasible tangent cone IR2 .S; x; N uN / of S at .x;
N uN /
is the set of all x 2 X such that for every sequence f.sn ; tn /g P P with
sn
.sn ; tn / # 0 and # 0, there exists an index m 2 N satisfying that xN C sn uN C
tn
sn tn x 2 S for every n
m.
Equivalently,
s
R2 .S; x;
N uN / D fx 2 Xj 8 t > 0 9 s; r 2 .0; t/ with < t such that xN C s uN C srx 2 S g;
r
s
N uN / D fx 2 Xj 9 t > 0 8 s; r 2 .0; t/ with < t we have xN C s uN C srx 2 S g:
IR2 .S; x;
r
The following theorem collects some properties of the second-order asymptotic

radial tangent cone and the second-order asymptotic feasible tangent cone.

1. R2 .S; x;
N 0/ D R.S; x/ N
2. IR.S; x;
N 0/ D IR.S; x/ N
3. IR2 .X nS; x;
N uN / D X nR2 .S; x; N uN /
4. 0 2 IR2 .S; x;
N uN / ” uN 2 IR.S; x/ N
5. 0 2 R2 .S; x;
N uN / ” uN 2 R.S; x/ N
6. S1 S2 H) R2 .S1 ; x; N uN / R2 .S2 ; x;
N uN /
7. S1 S2 H) IR .S1 ; x;2
N uN / IR2 .S2 ; x;
N uN /
8. R2 .S1 [ S2 ; x;
N uN / D R2 .S2 ; x;
N uN / [ R2 .S1 ; x;
N uN /
9. IR2 .S1 \ S2 ; x;
N uN / D IR2 .S1 ; x;
N uN / \ IR2 .S2 ; x;
N uN /.
Proof. The proof is a direct consequence of the definition of the cones involved. u
t
4.11.2 Second-Order Asymptotic Contingent Cone

and Second-Order Asymptotic Interiorly Contingent
Cone
Our focus is on the following two notions of second-order asymptotic tangent sets.
N uN / 2
X X.
1. The second-order asymptotic contingent cone T2 .S; x; N uN / of S at .x; N uN / is
the set of all x 2 X for which there are a sequence f.sn ; tn /g P P with
sn
.sn ; tn / # 0 and # 0 and a sequence fxn g X with xn ! x satisfying that
tn
xN C sn uN C sn tn xn 2 S for every n 2 N.
2. The second-order asymptotic interiorly contingent cone IT2 .S; x; N uN / of S at
N uN / is the set of all x 2 X such that for every sequence f.sn ; tn /g P P with
.x;
sn
.sn ; tn / # 0 and # 0 and every sequence fxn g X with xn ! x, there exists
tn
an index m 2 N satisfying that xN C sn uN C sn tn xn 2 S for every n
m.
Equivalently,
N uN / D fx 2 X j 8 U.x/ 8 t > 0 9 r; s 2 .0; t/ with

T2 .S; x;
s
< t 9 z 2 U.x/ with xN C s uN C srz 2 S g;
r
IT2 .S; x;
N uN / D fx 2 X j 9 U.x/ 9 t > 0 8 s; r 2 .0; t/ / with
s
< t 8 z 2 U.x/ we have xN C s uN C srz 2 S g:
r
Remark 4.11.4. The second-order asymptotic contingent cone is exactly the object
N uN / which was proposed in Definition 4.10.1. Consequently, some of its
T0 .S; x;
properties which are given in Theorem 4.11.6, follow from Theorem 4.10.3.
The following theorem collects some of the other commonly used characteriza-
tions of the second-order asymptotic contingent cone.
Theorem 4.11.5. Let X be a normed space, let S be subset of X , let uN 2 X , and
N uN / D T2i .S; x;
let xN 2 cl S . Then, for i 2 f1; : : : ; 5g, we have T2 .S; x; N uN /, where
Ti .S; x;
2
N uN / are given by

sn
T21 .S; x;
N uN / D x 2 X j 9 .sn ; tn / # .0; 0/ with # 0; 9 fxn g S such that
tn

xn x sn uN
!x
sn tn
( )
d.xN C s uN C stx; S /
T2 .S; x;
2
N uN / D x 2 X j lim infs D0
.s;t /#.0;0/; t #0 st

sn
T3 .S; x;
2
N uN / D x 2 X j 9 .sn ; tn / # .0; 0/; # 0; 9 fxn g ! x;
tn

tn2
xN C sn uN C xn 2 S
sn

ˇn
T24 .S; x;
N uN / D x 2 X j 9 .˛n ; ˇn / ! .1; 1/; # 0; 9 fxn g S; such that
˛n

ˇn .˛n .xn x/ N uN / ! x
S xN s uN
N uN / D
T25 .S; x; lim sup :
.s;t /#.0;0/; s
t #0
st
Proof. The proof is based on standard arguments. t

u
The following theorem collects some properties of the second-order asymptotic
contingent cone and second-order asymptotic interiorly contingent cone.
1. N 0/ D T .S; x/
T2 .S; x; N
2. N 0/ D IT.S; x/
IT2 .S; x; N
3. T2 .X nS; x;N uN / D X nIT2 .S; x; N uN /
4. 0 2 T2 .S; x;N uN / H) uN 2 T .S; x/ N
5. 0 2 IT2 .S; x;N uN / (H uN 2 IT.S; x/ N
6. T2 .S; x;
N uN / is a closed cone
7. T2 .S; x;
N uN / D T2 .cl S; x;N uN /
8. IT2 .S; x;
N uN / is an open cone
9. IT2 .S; x;
N uN / D IT2 .int S; x;N uN /
10. T .S; x;
2
N uN / ¤ ; H) uN 2 T .S; x/ N and hence uN 26 T .S; x/
N H)
T2 .S; x;
N uN / D ;
11. IT2 .S; x;
N uN / ¤ X H) uN 62 IT.S; x/ N and hence uN 2 IT.S; x/
N H)
T2 .S; x;
N uN / D X
12. S1 S2 H) T2 .S1 ; x; N uN / T2 .S2 ; x;
N uN /
13. S1 S2 H) IT .S1 ; x; 2
N uN / IT2 .S2 ; x;N uN /
14. IT2 .S1 \ S2 ; x;N uN / D IT2 .S2 ; x;
N uN / \ IT2 .S1 ; x;
N uN /
15. T2 .S1 [ S2 ; x;
N uN / D T2 .S1 ; x;
N uN / [ T2 .S2 ; x;
N uN /
16. T2 .S; x;
N uN / C t uN T2 .S; x;
N uN /; for every t 2 R
17. IT .S; x;
2
N uN / C t uN IT2 .S; x;N uN /; for every t 2 R.
Proof. The proof is based on standard arguments used earlier in this chapter. t
u
Remark 4.11.7. The authors in [209] give a nice proof of the fact that, in a finite
N uN / is equivalent to uN 2 T .S; x/.
dimensional setting, 0 2 T2 .S; x; N
The following example, originally given in [209] shows that, in general,
N uN / and T2 .S; x;
T 2 .S; x; N uN / are different objects:
Example 4.11.8. For S R2 given by S D f.x; y/ 2 R2 j y
x 2 g [ f.x; y/ 2
R2 jy D x 2 g, and for xN D .0; 0/ and uN D .1; 0/, we have
N uN / D f.x; y/ 2 R2 j y
2 [ y D 2g;
T 2 .S; x;
T2 .S; x;
0g:
4.11.3 Second-Order Asymptotic Adjacent Cone

and Second-Order Asymptotic Interiorly Adjacent Cone
Our focus is on the following two notions:

N uN / 2
X X.
1. The second-order asymptotic adjacent cone A2 .S; x; N uN / of S at .x;N uN / is the
set of all x 2 X such that for every sequence f.sn ; tn /g P P with .sn ; tn / # 0
sn
and # 0, there exists a sequence fxn g X with xn ! x satisfying that
tn
2. The second-order asymptotic interiorly adjacent cone IA2 .S; x; N uN / of S at
N uN / is the set of all x 2 X such that there exists a sequence f.sn ; tn /g P P
.x;
sn
with .sn ; tn / # 0 and # 0 so that for every sequence fxn g X with xn ! x,
tn
there exists an index m 2 N satisfying that xN C sn uN C sn tn xn 2 S for every
n
m.
Equivalently,
A2 .S; x;
N uN / D fx 2 X j 8 U.x/ 9 t > 0 8 s; r 2 .0; t/ / such that
s
< t 9 z 2 U.x/ with xN C s uN C srz 2 S g;
r
IA2 .S; x;
N uN / D fx 2 X j 9 U.x/ 9 t > 0 8 s; r 2 .0; t/ / with
s
< t 8 z 2 U.x/ we have xN C s uN C srz 2 S g:
r
The following result collects some of the other commonly used definitions of the
second-order asymptotic adjacent cones.
Theorem 4.11.10. Let X be a normed space, let S be a subset of X , let uN 2 X ,
and let xN 2 cl S . Then for i 2 f1; : : : ; 5g, we have A2 .S; x;
N uN / D A2i .S; x;
N uN /, where
Ai .S; x;
N uN / are given as follows:

sn
N uN / D x 2 X j 8 .sn ; tn / # .0; 0/ with
A21 .S; x; # 0; 9 fxn g S such that
tn

xn x sn uN
!x
sn tn
( )
d.xN C tn uN C stx; S /
A2 .S; x;
2
N uN / D fx 2 X j lim sup D0
.s;t /#.0;0/; st #0 st
sn
A23 .S; x;
N uN / D fx 2 X j 8 .sn ; tn / # .0; 0/ with # 0 9fxn g ! x such that
tn
xN C sn uN C sn1 tn2 xn 2 S g

ˇn
A24 .S; x;
N uN / D x 2 X j 8 .˛n ; ˇn / ! .1; 1/ with # 0 9fxn g S such that
˛n

ˇn .˛n .xn x/N uN / ! x
S xN s uN
A25 .S; x;
N uN / D lim inf :
.s;t /#.0;0/; st #0 st
The following result collects some properties of the second-order asymptotic

adjacent cones and the second-order asymptotic interiorly adjacent cones.
1. A2 .S; x;
N 0/ D A.S; x/ N
2. IA2 .S; x;
N 0/ D IA.S; x/ N
3. A2 .X nS; x;
N uN / D X nIA2 .S; x;
N uN /
4.12 Miscellaneous Properties of Second-Order Tangent Sets and Second-. . . 187
4. 0 2 A2 .S; x;N uN / H) uN 2 A.S; x/ N

5. 0 2 IA2 .S; x;N uN / (H uN 2 IA.S; x/ N
6. A2 .A; x;
N uN / is a closed cone
7. A2 .S; x;
N uN / D A2 .cl S; x;
N uN /
8. IA .S; x;
2
N uN / is an open cone
9. IA2 .S; x;
N uN / D IA2 .int S; x;
N uN /
10. A .S; x;
2
N uN / ¤ ; H) uN 2 A.S; x/ N and hence uN 62 A.S; x/ N H)
A2 .S; x;
N uN / D ;.
11. IA2 .S; x;
N uN / ¤ X H) uN 62 IT.S; x/ N and hence IA2 .S; x;
N uN / ¤ X H)
IA .S; x;
2
N uN / D X .
12. S1 S2 H) A2 .S1 ; x; N uN / A2 .S2 ; x;N uN /
13. S1 S2 H) IA .S1 ; x; 2
N uN / IA2 .S2 ; x;
N uN /
14. A2 .S; x;
N uN / C t uN A2 .S; x;
N uN /; for every t 2 R
15. IA2 .S; x;
N uN / C t uN IA2 .S; x;N uN /; for every t 2 R.
Proof. The proof is based on standard arguments used earlier in this chapter. t
u
In the above, we have studied second-order extensions of the first-order tangent
cones. In fact, it is possible to modify some of these definitions to obtain second-
order versions of the first-order modified tangent cones. We explain this by
giving the following definition of the second-order Clarke’s tangent set given by
Ward [595]:
Definition 4.11.12. Let X be a normed space, let S X , and let .x; N uN / 2 S X .
The second-order Clarke’s tangent cone C 2 .S; x; N uN / is the set of all x 2 X such
that for every ftn g P with tn # 0 and for every fzn g S with zn ! x, N there exists
a sequence fxn g X such that xn ! x and zn C tn uN C tn2 xn 2 S for every n 2 N.
Clearly, it is possible to obtain variants of the other first-order modified tangent
cones.
4.12 Miscellaneous Properties of Second-Order Tangent Sets

and Second-Order Asymptotic Tangent Cones
In this section, we collect useful properties of the second-order tangent cones and
the second-order tangent sets studied in the previous sections.
We begin with the following result that shows that, on convex sets, the relation-
ship among some of the second-order tangent sets and second-order tangent cones
is quite elegant:
Theorem 4.12.1. Let X be a normed space, let S X be convex, let xN 2 S , and
let uN 2 T .S; x/.
N Then:
N uN / C T .T .S; x/;
1. T 2 .S; x; N uN / T 2 .S; x;
N uN / T .T .S; x/;
N uN /
2
2. A .S; x;N uN / C T .T .S; x/;
N uN / A2 .S; x;
N uN / T .T .S; x/;
N uN /
3. T2 .S; x;
N uN / C T .T .S; x/;
N uN / T2 .S; x;N uN /
4. A.A.S; x/;N uN / D T .T .S; x/;N uN / D cl cone Œcone .S x/ N uN
5. N uN / A.A.S; x/;
A2 .S; x; N uN /
6. N uN / D T .T .S; x/;
T 2 .S; x; N uN / if 0 2 T 2 .S; x;N uN /
7. 2
A .S; x;N uN / D T .S; x;
2
N uN / D T .T .S; x/;
N uN / if 0 2 A2 .S; x; N uN /
8. T2 .S; x;
N uN / D cl cone Œcone .S x/ N uN , provided that T2 .S; x; N uN / ¤ ;
9. N uN / T2 .S; x;
T 2 .S; x; N uN /, provided that T2 .S; x; N uN / ¤ ;
10. N uN / C ˇT 2 .S; x;
˛T 2 .S; x; N uN / T 2 .S; x;
N uN /, for any ˛
0 and ˇ
0
11. N uN / D cl .T .S; x/
T .T .S; x/; N C LinfNug/ where LinfNug WD ft uN j t 2 Rg, the
linear space generated by uN .
N uN / be arbitrary. Then there are sequences ftn g P and
Proof. Let x 2 T 2 .S; x;
fxn g X with tn # 0 and xn ! x such that xN C tn uN C 12 tn2 xn 2 S for every n 2 N.
Next we choose an element z 2 S and positive scalars ˛; ˇ 2 P arbitrarily. Since
S is convex, for sufficiently large n 2 N, we have xN C ˇtn .z x/ N 2 S . Using the
convexity of S once again, for sufficiently large n 2 N, we also have

˛tn 1 ˛tn
1 xN C tn uN C tn2 xn C ŒxN C ˇtn .z x/
N 2S
2 2 2
which implies that
1
xN C tn uN C tn2 Œxn C ˛ Œˇ.z x/
N uN C ˛tn xn 2 S
2
and hence x C ˛ Œˇ.z x/ N uN 2 T 2 .S; x; N uN /. Since S is convex, we have

N D cl [ˇ>0 ˇ.S x/
T .S; x/ N , and this, in view of the closedness of T 2 .S; x;
N uN /,
ensures that x C ˛ ŒT .S; x/
N uN T 2 .S; x;
N uN /.
Furthermore, since T .S; x/N is also convex, by repeating the same arguments
again, we deduce that x C T .T .S; x/;
N uN / T 2 .S; x; N uN / proving that
N uN / C T .T .S; x/;
T 2 .S; x; N uN / T 2 .S; x;
N uN /:
For the second inclusion, we note that, since S xN C T .S; x/,

N we have

1 1
d xN C tn uN C tn2 x; S
d tn uN C tn2 x; T .S; x/
N ;
2 2

and hence x 2 T 2 .S; x; N uN / implies that d uN C t2n x; T .S; x/
N D o.tn /, confirming
that x2 2 T .T .S; x/;
N uN /. Hence x 2 T .T .S; x/;
N uN / due to the fact that T .T .S; x/;
N uN /
is a cone.
All the remaining parts can be proved by similar arguments. For instance, for
N uN / D T .T .S; x/;
T 2 .S; x; N uN /, we note that if 0 2 T 2 .S; x; N uN /, then we have the
N uN / T 2 .S; x;
inclusion T .T .S; x/; N uN /. Since the converse inclusion holds, we
easily deduce the identity. Finally, the last part follows from the fact if K is a closed
and convex cone and wN 2 K, then T .K; w/ N D cl .K C Linfwg/
N (see [53]). The proof
is complete. t
u
Remark 4.12.2. The second inclusion was given by Cominetti [103]. The first two
inclusions, the sixth, the seventh, and the last part is given in [53]. See also [209,
304, 305]. The second to last part is given by Di [134]. We note that the last part
relates nicely Proposition 4.4.4.
It is natural to consider second-order analogues by taking tangent cones of
tangent cones. The following interesting example, originally given by Jourani [309],
shows that, in general, the second-order tangent sets are different from the tangent
cone of a tangent cone.
˚
Example 4.12.3. For S D .x; y/ R2 j x ¤ 1; .x C 1/y D 1 , we set xN D
.0; 1/ and uN D .1; 1/. Then,
N uN / D f.x; y/ 2 R2 j y D x C 1g;
A2 .S; x;
N uN / D f.x; y/ 2 R2 j y D xg:
A.A.S; x/;
˚
Example 4.12.4. For S D .x; y/ 2 R2 j x > 1; .x C 1/y
1 , we set xN D
.0; 1/ and uN D .1; 1/. Then,
x C 1g;
A2 .S; x;
xg:
A.A.S; x/;
Remark 4.12.5. If the underlying set S is convex, then A2 .S; x;N uN / is convex.
N uN / may fail to be convex even for a convex set S . See [53,
However, T 2 .S; x;
Example 3.3.1]
Our next result is a collection of intersection formulas relating a few cones:
Theorem 4.12.6. Let X be a normed space, let S1 and S2 be two nonempty subsets
of X , let xN 2 cl S1 \ cl S2 , and let uN 2 X . Then the following relationships hold:
1. N uN / \ IA2 .S2 ; x;
IT 2 .S1 ; x; N uN / IA2 .S1 \ S2 ; x;
N uN /
2. 2
IT .S1 ; x;N uN / \ A .S2 ; x;
2
N uN / A2 .S1 \ S2 ; x;
N uN /
3. N uN / \ T 2 .S2 ; x;
IT 2 .S1 ; x; N uN / T 2 .S1 \ S2 ; x;
N uN /
4. 2
IA .S1 ; x;N uN / \ A .S2 ; x;
2
N uN / T 2 .S1 \ S2 ; x;
N uN /.
Proof. The proof can be obtained by using the definition of the sets involved. t
u
The following result collects a useful feature of the second-order contingent cone
and the second-order adjacent cone (see [594]).
N uN / 2
Theorem 4.12.7. Let X be a normed space, let S be a subset of X , and let .x;
S X . Let ` W X ! Y be linear and continuous. Then the following inclusions hold:
N uN // T 2 .`.S /; `.x/;
`.T 2 .S; x; N `.Nu//; (4.74)
2
N uN // A .`.S /; `.x/;
`.A .S; x; 2
N A.Nu//: (4.75)
Proof. Following directly from the definitions of the involved sets. t

u
We have the following auxiliary result which will be used shortly (see
Ward [594]):
Theorem 4.12.8. Let X be a normed space, let S be a subset of X , let xN 2 S and
let uN 2 X . Then the following inclusions hold:
N uN / C C.S; x/
T 2 .S; x; N T 2 .S; x;
N uN /
N uN / C C.S; x/
A2 .S; x; N A2 .S; x;
N uN /
N uN / C IT M .S; x/
A2 .S; x; N IT 2 .S; x;
N uN /:
Proof. Let x 2 T 2 .S; x; N uN / and y 2 C.S; x/ N be arbitrary. Then, due to the

containment x 2 T 2 .S; x; N uN /, there are sequences ftn g P and fxn g X such
t2
that tn # 0, xn ! x and zn WD xN C tn uN C 2n x 2 S for every n 2 N. Since fzn g S
and zn ! x, N the containment y 2 C.S; x/ N ensures that there exists a sequence
fyn g Y with yn ! y such that for every sequence fsn g P with sn # 0, we have
t2 t2
zn C sn yn 2 S for every n 2 N. Since, xN C tn uN C 2n .xn C yn / D zn C 2n yn 2 S , we
at once deduce that x C y 2 T 2 .S; x; N uN /. This proves the first inclusion. The proof
of the second inclusion is obtained by similar arguments.
For the final inclusion, let x 2 A2 .S; x; N uN / and y 2 IT M .S; x/
N be arbitrary. We
set z WD x C y. Let ftn g P with tn # 0 and fzn g X with zn ! z be arbitrary
sequences. Since x 2 A2 .S; x; N uN /, there exists a sequence fyn g X with yn ! x
such that xn WD xN C tn uN C tn2 yn =2 2 S for every n 2 N. We set un D zn yn and
note that un ! y, and since y 2 IT M .S; x/, N we have
tn2 t2 t2
xN C tn uN C zn D xN C tn uN C n .yn C un / D xn C n un 2 S
2 2 2
N uN /.
for n sufficiently large. Consequently x C y 2 IT 2 .S; x; t
u
We now state a useful consequence of the above result.
Proposition 4.12.9. Let X be a normed space, let S be a subset of X , let xN 2 S
and let uN 2 X . Assume that IT M .S; x/
N ¤ ; and A2 .S; x;
N uN / ¤ ;. Then

N uN / D cl IT 2 .S; x;
A2 .S; x; N uN / ; (4.76)

N uN / D int A2 .S; x;
IT 2 .S; x; N uN / : (4.77)
The following intersection formula for second-order contingent sets is due to

Ward [594]:
Theorem 4.12.10. Let X be a normed space, let fS1 ; S2 ; : : : ; Sn g be a family of
\
n
nonempty subsets of X , let xN 2 Si , and let uN 2 X . Assume that
i D1
N \ IT M .S2 ; x/
C.S1 ; x/ N \ \ IT M .Sn ; x/
N ¤ ;: (4.78)
Then the following intersection formulas hold:
N uN / \ A2 .S2 ; x;
T 2 .S1 ; x; N uN / \ \ A2 .Sn ; x;
N uN / T 2 .S1 \ S2 \ \ Sn ; x;
N uN /I
(4.79)
N uN / \ A2 .S2 ; x;
A2 .S1 ; x; N uN / \ \ A2 .Sn ; x;
N uN / D A2 .S1 \ S2 \ \ Sn ; x;
N uN /:
(4.80)
N uN / D T 2 .Si ; x;
Furthermore, if A2 .Si ; x; N uN / for every i D 2; 3; : : : ; n, then
N uN / \ T 2 .S2 ; x;
T 2 .S1 ; x; N uN / \ \ T 2 .Sn ; x;
N uN / D T 2 .S1 \ S2 \ \ Sn ; x;
N uN /:
(4.81)
Proof. To prove (4.79), we first note that the following inclusion is always true:
N uN / \ IT 2 .S2 ; x;
T 2 .S1 ; x; N uN / \ \ IT 2 .Sn ; x;
N uN / T 2 .S1 \ S2 \ \ Sn ; x;
N uN /:
(4.82)
Let x 2 T 2 .S1 ; x; N uN / \ A2 .S2 ; x;

N uN / \ \ A2 .Sn ; x;
N uN / and y 2 C.S1 ; x/
N \
N \ \ IT M .Sn ; x/
IT M .S2 ; x/ N be arbitrary. By using Theorem 4.12.8 and the fact
that C.S1 ; x/N and IT M .Si ; x/
N are cones, we deduce that for every t > 0,
x C ty 2 T 2 .S1 ; x;
N uN / \ IT 2 .S2 ; x;
N uN / \ \ IT 2 .Sn ; x;
N uN /;
and by using (4.82), we get
x C ty 2 T 2 .S1 \ S2 \ \ Sn ; x;
N uN /;
which confirms that

y 2 cl T 2 .S1 \ S2 \ \ Sn ; x;
N uN / D T 2 .S1 \ S2 \ \ Sn ; x;
N uN /:
Inclusion (4.79) and equality (4.80) follow from analogous arguments. t

u
In the above result, (4.78) played a crucial role involving convex cones. In the
following, we state an analogue of the above result for the case when the underlying
sets are closed and convex and all but one are solid. The proof of this result relies on
the Robinson’s constraint qualification (see [53, Proposition 3.36]) and hence makes
good use of the imposed conditions:
Theorem 4.12.11. Let X be a Banach space and let fS1 ; S2 ; : : : ; Sn g be a family
of closed and convex subsets of X . Assume that there exists zN 2 S1 such that zN 2
int.Si /, for i 2 f2; : : : ; ng. Assume that xN 2 S1 \ S2 \ \ Sn and uN 2 T .S1 ; x/
N \
N \ \ T .Sn ; x/.
T .S2 ; x/ N Then:
T .S1 \ S2 \ \ Sn ; x/
N D T .S1 ; x/
N \ T .S2 ; x/
N \ \ T .Sn ; x/;
N (4.83)
T 2 .S1 \ S2 \ \ Sn ; x;
N uN / T 2 .S1 ; x;
N uN / \ T 2 .S2 ; x;
N uN / \ \ T 2 .Sn ; x;
N uN /;
(4.84)
A .S1 \ S2 \ \ Sn ; x;
2
N uN / D A .S1 ; x;
2
N uN / \ A .S2 ; x;
2
N uN / \ \ A .Sn ; x;
2
N uN /:
(4.85)
4.13 Second-Order Inversion Theorems
In this section, we give some second-order inversion theorems which will play
a fundamental role in deriving calculus rules for the second-order graphical and
epigraphical derivatives. In the following, we give two results. The first result,
Theorem 4.13.1, is by Cominetti [103, Theorem 3.1] where the proof shows the
importance of the metric regularity. The second result is taken from [155], where a
more general notion of metric subregularity is used.
We begin with the following result due to Cominetti [103].
Theorem 4.13.1. Let X and Y be Banach spaces, let S X and Q Y be
nonempty, closed and convex, let g W X ! Y is continuous on S and twice
differentiable at xN 2 S with g.x/
N 2 Q, and let uN 2 X . Assume that the following
Robinson’s constraint qualification holds:
0 2 corefDg.x/.S
N x/
N .Q g.x//g:
N (4.86)
Then the following identities hold:
A.S \ g 1 .Q/; x/
N D A.S; x/ N 1 ŒA.Q; g.x//
N \ Dg.x/ N ; (4.87)
A2 .S \ g 1 .Q/; x;
N uN / D A2 .S; x; N 1
N uN / \ Dg.x/
2
A .Q; g.x/; N Dg.x/.N
N u// D 2 g.x/.N
N u; uN / : (4.88)
Proof. Since (4.87) follows from (4.88), it is sufficient to prove the latter. The map
g being continuously differentiable, is Lipschitz continuous near x,N and hence for
some k
0 and every x near x,N we have
4.13 Second-Order Inversion Theorems 193
d.g.x/; Q/ k d.x; g 1 .Q// k d.x; S \ g 1 .Q//:
Then for an arbitrary x 2 A2 .S \ g 1 .Q/; x;

N uN /, we have

1 2
d g xN C t uN C t x ; Q D o.t 2 /:
2
By using the Taylor expansion, we have

1 1
g xN C t uN C t 2 x D g.x/
N C tDg.x/.N
N u/ C t 2 Dg.x/.x/
N N u; uN / C o.t 2 /;
C D 2 g.x/.N
2 2
implying that
N
Dg.x/.x/ C D 2 g.x/.N
N u; uN / 2 A2 .Q; g.x/;
N Dg.x/.N
N u//;
which, when combined with the inclusion A2 .S \ g 1 .Q/; x;

N uN / A2 .S; x;
N uN /,
yields
A2 .S \ g 1 .Q/; x;
N uN / A2 .S; x;
N uN / \ Dg.x/ N 1
2
N Dg.x/.N
A .Q; g.x/; N u// D 2 g.x/.N
N u; uN / :
For the converse inclusion, we note that for some k

0 and all x 2 S near x,
N
we have (see [103, Corolary 2.2])
d.x; S \ g 1 .Q// kd.g.x/; Q/: (4.89)

Now, if x 2 A2 .S; x; N 1 A2 .Q; g.x/;
N uN / \ Dg.x/ N Dg.x/.N
N u// D 2 g.x/.N
N u; uN / ,
then

1 2
d g.x/N C tDg.x/.N
N u/ C t Dg.x/.x/ N 2
N u; uN / ; Q o.t 2 /
C D g.x/.N
2
and hence by choosing xt ! x with xN C t uN C 12 tn2 xt 2 S , we get

1
d g xN C t uN C t 2 xt ; Q o.t 2 /:
2
Therefore, by using (4.89), we deduce that

1 2 1
d xN C t uN C tn xt ; S \ g .Q/ o.t 2 /;
2
which gives that x 2 A2 .S \ g 1 .Q/; x;

N uN /. The proof is complete. t
u
We remark that Robinson’s constraint qualification 0 2 corefDg.x/.S

N x/
N
.Q g.x//g
N can equivalently be written as
N
Dg.x/.IR.S; N IR.Q; g.x//
x// N D Y;
where IR.A; zN/ D [t >0 t.A zN/ is the feasible tangent cone to A at zN. If either Y is
finite dimensional or int.Q/ ¤ ;, this condition is equivalent to the condition
1
N
Dg.x/ N S // \ .N.g.x/;
.N.x; N Q// D f0g;
where N.; / is the classical normal cone.

For the case of mathematical programming where g W Rn ! Rp Rq with
g D .g1 ; g2 ; : : : ; gp ; h1 ; : : : ; hq /;
S D Rn ;
Q D Rp f0Rq g;
we have
˚
S \ g 1 ..y; z/ C Q/ D x 2 Rn j gi .x/ yi hj .x/ D zj ;
and Robinson’s constraint qualification becomes the Mangasarian-Fromowitz con-

straint qualification. That is:
p
1. frhj .x/g
N j D1 is linearly independent.
2. There exists v 2 Rn such that
hrhj .x/;
N vi D 0 for every j;
hrgj .x/;
N vi < 0 for all i such that gi .x/
N D 0:
In [53], an excellent discussion of various equivalent formulations of these

constraint qualifications and their limitations can be found.
Furthermore, under the setting of Theorem 4.13.1, with g W X ! Y twice
continuously differentiable, for the particular case when S D X and Q is a closed
and convex, under the following Robinson’s constraint qualification
0 2 int Œg.x/
N C Dg.x/X
N Q ; (4.90)
Bonnans and Shapiro [53, Page 167] gave the following formulas:

T 2 .g 1 .Q/; x; N 1 T 2 .Q; g.x/;
N uN / D Dg.x/ N Dg.x/.N
N u// D 2 g.x/.N
N u; uN / ; (4.91)

A2 .g 1 .Q/; x; N 1 A2 .Q; g.x/;
N u// D 2 g.x/.N
N u; uN / : (4.92)
4.13 Second-Order Inversion Theorems 195
Several particular cases of Theorem 4.13.1 and formulas (4.91)–(4.91) have

been given in the literature. For instance, Theorem 12.9 of the book by Ben-Israel,
Ben-Tal, and Zlobec [42] gives (4.88) for the case S D X and Q D f0g. For this
case, (4.88) reads:
A2 .g 1 .0/; x;
N uN / D fx 2 X j Dg.x/.x/
N C D 2 g.x/.N
N u; uN / D 0g:
Furthermore, Kawasaki [317] proved analogous result for the case when S D X
and Q D f0g K, where K is a closed, convex, and solid cone. For the particular
case when S D X and Q is a convex polyhedra, it was shown by Rockafellar [494]
that

A2 .g 1 .Q/; x; N 1 A.A.Q; g.x//;
N u// D 2 g.x/.N
N u; uN / ;
which is equivalent to the given result as
N Dg.x/.N
A.A.Q; g.x//; N Dg.x/.N
N u// D A2 .Q; g.x/; N u//
when Q is a polyhedra.
We conclude this discussion by a remark that the above results also generalizes
the following well-known inversion theorem (see [524]):
Theorem 4.13.2. Let X and Y be Banach spaces, let Q be a nonempty, closed,
and convex subset of Y , let g W X ! Y be continuous on X and continuously
differentiable at xN 2 g 1 .Q/. Assume that Robinson’s constraint qualification
(4.90) holds. Then
T .g 1 .Q/; x/ N 1 .T .Q; x//:

N D Dg.x/ N (4.93)
We note that the above result subsumes the following classical tangent space
theorem:
Theorem 4.13.3. Let X and Y be Banach spaces, let g W X ! Y be continuous on
X and continuously differentiable at xN 2 g 1 .0/. If Dg.x/
N is surjective, then
N D Ker.Dg.x//:
T .Ker.g/; x/ N
It is evident from the proof of Theorem 4.13.1 that the imposed conditions make
a good use of the metric regularity. In our next result, we give an inversion theorem
under the weaker notion of metric subregularity (see [155]).
Theorem 4.13.4. Let X and Y be Banach spaces, let S X and Q Y be
closed sets, let g W X ! Y be a twice continuously differentiable map, and let
xN 2 S \ g 1 .Q/. Assume that f W X Y ! Y defined by f .x; y/ D g.x/ y is
N g.x/;
metrically subregular at .x; N 0/ with respect to S Q in the following sense:
N and U .g.x//
There exist a constant k > 0 and neighborhoods U .x/ N such that for
every .x; y/ 2 ŒU .x/N U .g.x//

N \ S Q, (4.49) holds. Then for any uN 2 X , we
have

1 1 2
2
T .S; x;N uN / \ Dg.x/
N 2
N Dg.x/.N
A .Q; g.x/; N u// D g.x/.N N u; uN /
2

T 2 S \ g 1 .Q/; x;N uN ;

1
A2 .S; x; N 1 T 2 .Q; g.x/;
N uN / \ Dg.x/ N Dg.x/.N
N u// D 2 g.x/.NN u; uN /
2
1

T S \ g .Q/; x;
2
N uN ;

1 1 2
2
A .S; x;N uN / \ Dg.x/
N 2
N Dg.x/.N
A .Q; g.x/; N u// D g.x/.N N u; uN /
2

A2 S \ g 1 .Q/; x; N uN :
Proof. It follows from the metric subregularity hypothesis that there are a number
N and U .g.x//
k > 0 and neighborhoods U .x/ N such that for every .x; y/ 2 ŒU .x/
N
N \ S Q, we have
U .g.x//

d .x; y/; f 1 .0/ \ .S Q/ kg.x/ yk: (4.94)
Let

1 1 2
x 2 T .S; x;
2
N uN / \ Dg.x/
N 2
N Dg.x/.N
A .Q; g.x/; N u// D g.x/.N
N u; uN /
2
be arbitrary. Therefore, x 2 T 2 .S; x;

N uN / and
1
N
Dg.x/.x/ C D 2 g.x/.N
N u; uN / 2 A2 .Q; g.x/;
N Dg.x/.N
N u//:
2
By using the definitions of the second-order adjacent set and the second-order
contingent set, there are sequences ftn g, fxn g, and fzn g such that tn # 0, xn ! x,
with xN C tn uN C tn2 xn 2 S , and

1 2
zn ! Dg.x/.x/
N C D g.x/.N
N u; uN /
2
N C tn Dg.x/.N
with g.x/ N u/ C tn2 zn 2 Q.
Equation (4.94) with
.x; y/ D .xN C tn uN C tn2 xn ; g.x/

N C tn Dg.x/.N
N u/ C tn2 zn /
4.14 Expressions of the Second-Order Contingent Set on Specific Constraints 197
ensures that for sufficiently large n 2 N, there exists .pn ; qn / 2 S Q with g.pn / D
qn such that
k.xN C tn uN C tn2 xn ; g.x/

N C tn Dg.x/.N
N u/ C tn2 zn / .pn ; qn /k
< kkg.xN C tn uN C tn2 xn / g.x/
N tn Dg.x/.N
N u/ tn2 zn k C tn3 :
Consequently, pn 2 S \ g 1 .Q/ and

pn xN tn uN g.xN C tn uN C tn2 xn / g.x/
N tn Dg.x/.N
N u/

xn < k zn
t 2 t 2 C tn :
n n
Since
g.xN C tn uN C tn2 xn / g.x/

N tn Dg.x/.N
N u/ 1
! Dg.x/.x/
N C D 2 g.x/.N
N u; uN /;
tn2 2
we deduce that
pn xN tn uN
wn D !x
tn2
with
xN C tn uN C tn2 wn D pn 2 S \ g 1 .Q/
and hence the proof of the first inclusion is complete. The remaining parts follow
by using the properties of the second-order contingent set and the second-order
adjacent set. t
u
4.14 Expressions of the Second-Order Contingent Set

on Specific Constraints
In this section, we give explicit formulas for the computation of the second-order
contingent set on certain constraint sets. This section also contains some characteris-
tic examples of second-order contingent sets. We begin with the following definition
proposed by Di [134].
Definition 4.14.1. Given a Hilbert space X , a map g W X ! R is said to admit a
second-order Fréchet derivative at xN if there exists an element r 2 g.x/
N 2 L .X; X /
such that
g.xN C tx/ g.x/N thrg.x/;

N xi
lim D hx; r 2 g.x/.x/i;
N (4.95)
t #0 .t 2 =2/
and the limit is uniform with respect to x in bounded sets, where rg.x/
N is the first-
order derivative.
Remark 4.14.2. As noted by Di [134], the above notion of second-order Fréchet
differentiability is not a local property, and it does not imply that the function is
strictly differentiable at that point, nor continuously differentiable around the point,
nor Lipschitz around the point.
For the second-order contingent set, we have the following analogue of Proposi-
tion 4.8.2:
Proposition 4.14.3. Let X be a Hilbert space and let h W X ! R be continuous
around xN 2 X and differentiable at x.
N Let h.x/
N D 0, let rh.x/
N ¤ 0, and let h be
N
second-order Fréchet differentiable at x.
1. For S D fx 2 X j h.x/ D 0g and uN 2 T .S; x/,
N we have
˚ ˝ ˛
N uN / D x 2 X j uN ; r 2 h.x/.N
T 2 .S; x; N u/ C hrh.x/;
N xi D 0 :
2. For S D fx 2 X j h.x/
0g and uN 2 T .S; x/,
N we have
˚ ˝ ˛
N uN / D x 2 X j uN ; r 2 h.x/.N
N xi
0 :
Proof. We will only prove the case of the equality constraint. Let x 2 T 2 .S; x; N uN /.
t2
Then there exist fxn g S and ftn g # 0 such that xn D xN C tn uN C 2n x C o.tn2 /. By
N we get
the differentiability of h at x,
0 D h.xn / h.x/ N

t2
D rh.x/;N tn uN C x C o.t 2
n/
2

1 t2 t2
C tn uN C n x C o.tn2 /; r 2 h.x/
N tn uN C n x C o tn2
2 2 2
2
t2
Co tn uN C n x C o.tn2 /
2
tn2 ˝ ˛
D N xi C uN ; r 2 h.x/.N
hrh.x/; N u/ C o.tn2 /;
2
˝ ˛
N xi C uN ; r 2 h.x/.N
which confirms that hrh.x/; N u/ D 0 and hence
˚ ˝ ˛
N uN / x 2 X j uN ; r 2 h.x/.N
N xi D 0 :
For the converse, assume that for some x we have

˝ ˛
N xi C uN ; r 2 h.x/.N
hrh.x/; N u/ D 0:
Then, for any n 2 N, we define

t t
yn .t/ WD xN C t uN C x C rh.x/
N ;
2 n

t t
zn .t/ WD xN C t uN C x rh.x/
N :
2 n
N we have
Since h has the second-order Fréchet derivative at x,

t t
h.yn .t// h.x/
N D t rh.x/; N uN C x C rh.x/ N
2 n

t2 t t t t
C n uN C x C rh.x/; N r 2 h.x/
N uN C x C rh.x/
N
2 2 n 2 n
2
t2 t2
C o t uN C x C rh.x/ N
2 n
t2
D hrh.x/;N rh.x/i
N
n
t2 ˝ ˛
C N xi C uN ; r 2 h.x/.N
hrh.x/; N u/ C o.t 2 /;
2
t2
D N rh.x/i
hrh.x/; N C o.t 2 /; (4.96)
n
and analogously
t2
h.zn .t// h.x/
N D N rh.x/i
hrh.x/; N C o.t 2 /: (4.97)
n
By employing the uniform convergence in Fréchet differentiability, there exists
a t, denoted by tn , with tn # 0 such that (4.96) is positive and (4.97) is negative.
Therefore, by the intermediate value theorem, along the line segment between yn .tn /
and zn .tn /, there exists xn such that h.xn / h.x/
N D 0, confirming that xn 2 S .
Clearly, we also have
xn xN tn uN
1 2
!x
t
2 n
which ensures that x 2 T 2 .S; x;

u
In the following result, we give explicit expressions for the second-order
contingent set on the set of equality and inequality constraints:
Theorem 4.14.4. Let X be a Hilbert space, let S X be the constraint defined by
(4.52), and xN 2 S be arbitrary. Assume that the following Mangasarian-Fromovitz
constraint qualification holds: Let h D .h1 ; h2 ; : : : ; hs / be continuous around xN

N The only
2 , where is given by (4.56), such
and Fréchet differentiable at x.
Xs
that
i rhi .x/
N D 0, is
D 0. Assume that hi W X ! R admits second-order
i D1
Fréchet derivatives at xN for i 2 f1; 2; : : : ; sg and uN 2 T .S; x/.
N Then the following
formula holds:
\
s
N uN / D
T 2 .S; x; N uN /
T .Si ; x;
i D1
˚ ˝ ˛
D x 2 X j uN ; r 2 h.x/.N N 2
N u/ C hx; rh.x/i
ˇ˝ ˛ ˝ ˛
ˇ uN ; r 2 hi .x/.N
N u/ C x; r 2 hi .x/
N D 0; for i 2 I1
ˇ
D x 2 X ˇ˝ ˛ ˝ ˛ :
uN ; r 2 hi .x/.N
N u/ C x; r 2 hi .x/
N
0; for i 2 I2D
Proof. It suffices to show that
\
s
N uN / T .S; x;
T .Si ; x; N uN /;
i D1
\
s
as the converse inclusion holds universally. Let x 2 N uN / be arbitrary. For
T .Si ; x;
i D1
n 2 N, we define
!
t X
s
t
xn;t .˛/ D xN C t uN C x C ˛i rhi .x/
N ; (4.98)
2 n i D1
N we get
and by using the Frećhet differentiability of hi at x,
* +
t X
s
t
hj .xn;t .˛// D hj .x/ N uN C x C
N C t rhj .x/; ˛i rhi .x/
N
2 n i D1
*
t X
s
t2 t
C uN C x C ˛i rhi .x/;
N r 2 hj .x/
N
2 2 n i D1
!+
t X
s
t
uN C x C r˛i hi .x/
N
2 n i D1
!!2
t X
s
t2
C o t uN C x C ˛i rhi .x/
N
2 n i D1
* +
˝ ˛ t2 X s
D hj .x/
N C t rhj .x/;
N uN C rh.x/;
N ˛i rh.x/
N
n i D1
t 2 ˝ ˛ ˝ ˛
C rhj .x/;
N x C uN ; r 2 hj .x/.N
N u/ C o.t 2 /:
2
Consequently,
* +
t2 X s
hj .xn;t .˛// D N
rhj .x/; ˛i rhi .x/
N C o.t 2 /
n i D1
for j 2 Ii because hj .x/ N D 0, hrhj .x/;

N uN i D 0, and hrhj .x/;
N xi C
hNu; r 2 hj .x/.N
N u/i for such j . Moreover,
* +
t2 X s
hj .xn;t .˛//
N
rhj .x/; ˛i rhi .x/
N C o.t 2 /
n i D1
for j 2 I2D because hj .x/ N

0, hrhj .x/; N uN i
0, and hrhj .x/; N xi C
hNu; r 2 hj .x/.N
N u/i
0 for such j , and hj .xn;t .˛// > 0 for j 2 I2> if t is sufficiently
small. By the same reasoning as in the proof of Theorem 4.8.8, there exist tn and ˛n
such that ftn g # 0 and xn;t .˛n / satisfies (4.62). Therefore, xn WD xn;t .˛n / 2 S . It
can be checked that
xn xN tn uN
tn2
! x;
2
and hence it follows that x 2 T 2 .S; x;

u
The following result is an analogous of Theorem 4.14.4 for the second-order
contingent sets and can be proved by employing similar arguments (see Di [134]):
Theorem 4.14.5. Assume that the conditions of Theorem 4.8.11 hold. In addition,
assume that hi W X ! R admits second-order Fréchet derivatives at xN for i 2
f1; : : : ; sg and uN 2 T .P \ S; x/.
N Then:
\
s \
T 2 .P \ S; x;
N uN / D N uN /
T .Si ; x; N uN /
T 2 .P; x;
i D1
˚
D x 2 T 2 .P; x;
N uN /j hNu; r 2 h.x/i N 2 :
N C hx; rh.x/i
(4.99)
4.15 Miscellaneous Second-Order Tangent Cones
In this section, we study a few other notions of second-order tangent cones.
4.15.1 Second-Order Tangent Cones of Ledzewicz

and Schaettler
Probably, Ledzewicz and Schaettler [364] were the first to introduce second-order
notions of local approximation which were actually cones. We recall the two notions
proposed in [364].
Definition 4.15.1. Let X be a Banach space, let S X , let xN 2 S , and let uN 2 X .
1. The second-order feasible cone FC2 .S; x; N uN / of S at xN in the direction uN is the
collection of all pairs .x; / 2 X RC for which there exist an "0 > 0 and a
p
neighborhood V .x/ such that for every " 2 .0; "0 /, we have xN C" C"2 V .x/ N
S.
2. The second-order incident cone TC2 .S; x; N uN / of S at xN in the direction uN is the
collection of all pairs .x; / 2 X RC for which there exists an "0 > 0 such that
for every " 2 .0; "0 /, there exists an element r."/ 2 X of o."2 / satisfying that
p
xN C " C "2 x C r."/ 2 S .
Remark 4.15.2. In [364], the second-order incident cone is called the second-
order tangent cone. It is evident that TC2 .S; x; N uN / is a cone, that is, for any x 2
N uN / and any
> 0, it holds that
x 2 TC2 .S; x;
TC2 .S; x; N uN /. The second-order
feasible cone, FC2 .S; x;N uN /, on the other hand, is an open cone. It is also easy to
see that if uN 2 IT.S; x/,
N then FC2 .S; x; N uN / D X .0; 1/. The interiorly contingent
cone, studied earlier in this chapter, is called the first-order feasible cone in [364].
We now give a few examples to illustrate the above notions (see [364, 365]):
˚
Example 4.15.3. For S R2 given by S WD .x; y/j x
0; x 2 y 2x 2 , we
have IT.S; .0; 0// D ;, and FC2 .S; .0; 0/; .1; 0// D f..x; y/; / 2 R3 j 0 <
< y < 2 g.
Example 4.15.4. For S R2 given by S WD f.x; y/j y
jxjg, we obtain that
IT.S; .0; 0// D int S , and FC2 .S; .0; 0/; .1; 1// D f..x; y/; / 2 R3 j > 0;
x < yg.
˚
Example 4.15.5. For S R2 given by S WD .x; y/j x 2 C y 2 D 1 , a
˚
simple computation leads to TC2 .S; .1; 0/; .0; 1// D ..x; y/; / 2 R3 j > 0 ,
C 2x D 0g.
We now specify the above cones and their duals on some specific sets. For proofs
see [364].
4.15 Miscellaneous Second-Order Tangent Cones 203
Proposition 4.15.6. Let X be a Banach space, let g W X ! R be twice Fréchet

differentiable at xN with g 0 .x/
N ¤ 0. Let uN satisfy g 0 .x/.N
N u/ D 0. Then, for S WD
fx 2 X j g.x/ 0g, the second-order feasible cone FC2 .S; x; N uN / and its dual
N uN /C are given by:
FC2 .S; x;

1
N uN / D .x; / 2 X RC j g 0 .x/.x/
FC2 .S; x; N C g 00 .x/.N
N u; uN / < 0 ;
2
C
˚
2
N uN / D .
; / 2 X RC j 9 ˛1 0 ˛2
0 such that
FC .S; x;

g 0 .x/
N 0 ˛1
D 1 00 :
N u; uN / 1
2 g .x/.N ˛2
Let X and Y be Banach spaces, let g W X ! Y be a map, and let S WD fx 2

X j g.x/ D 0g. If g is continuously Fréchet differentiable in a neighborhood of
xN 2 S , and if the derivative g0 .x/
N is onto, that is, g is the so-called regular map at
N then by the classical Lusternik theorem, we have
x,
˚
N D x 2 X j g 0 .x/.x/
T .S; x/ N D 0 D Ker g 0 .x/
N : (4.100)
Among many generalization of the Lusternik theorem, the one given by

Avakov [20] deals with non-regular operators that are twice Fréchet differentiable
N Assuming that Im.g 0 .x//
at x. N is closed in Y , for uN 2 X , we define a linear map
N uN / W X ! Y Y nIm.g 0 .x//
G.x; N by the formula

N uN /.x/ D g 0 .x/.N
x ! G.x; N u/; g 00 .x/.N
N u; x/ ; (4.101)
where W Y ! Y nIm.g 0 .x// N is the quotient map. Then, under the conditions
that G.x;N uN /.Nu/ D 0 and Im.G.x; N uN // D Im.g0 .x//
N Y nIm .g 0 .x//,
N it holds that
uN 2 T .S; x/.
N
Using the above result, the following characterization of the second-order tangent
cone has been proved by Ledzewicz and Schaettler [364].
Proposition 4.15.7. Let X and Y be Banach spaces, let g W X ! Y be three
times Fréchet differentiable at xN 2 S , let Im.g0 .x// N be closed, let g 0 .x/.N
N u/ D 0
00 0
N u; uN / 2 .Im/.g .x//,
and g .x/.N N and let Im.G.x; N uN // D Im .g 0 .x//N Y nIm.g 0 .x//.
N
Then:

1
2
TC .S; x; N uN / D .x; / 2 X RC j g 0 .x/.x/
N C g 00 .x/.N N u; uN / D 0;
2

00 1 000 0
N u; x/ C g .x/.N
g .x/.N N u; uN ; uN / 2 Im.g .x//
N ;
6

N uN /C D .
; / 2 X R j 9 y1 2 Im.g 0 .x//
TC 2 .S; x; N ; y2 2 Im.g 0 .x//
N ? and
r
0 such that
D g 0 .x/
N y1 C .g 00 .x/.N
N u; // y2

1 00 1 00
D y1 ; g .x/.N
N u; uN / C y2 ; g .x/.N N u; uN /
2 2

1
C y2 ; g 000 .x/.N
N u; uN ; uN / C r :
6
4.15.2 Projective Tangent Cones of Second-Order
Our focus in this section is on the following notion of projective second-order

tangent sets introduced by Penot [468] in a higher-order setting. See also [465, 466,
469].
Definition 4.15.8. Let X be a normed space, let S be a subset of X , let uN 2 X , and
let xN 2 cl S .
1. The second-order projective contingent cone P2 .S; x; N uN / of S at .x;
N uN / is the
set of all .x; t/ 2 X R such that there are a sequence f.sn ; tn /g P P with
sn
.sn ; tn / ! .0C ; t/ and # 0 and a sequence fxn g S with xn ! x satisfying
tn
s2
that, for every n 2 N, we have xN C sn uN C 2tnn xn 2 S .
2. The second-order projective adjacent cone PO 2 .S; x; N uN / of S at .x;
N uN / is the
set of all .x; t/ 2 X R such that for every sequence f.sn ; tn /g P P with
sn
.sn ; tn / ! .0C ; t/ and # 0, there exists a sequence fxn g S with xn ! x
tn
s2
satisfying that for every n 2 N, we have xN C sn uN C 2tnn xn 2 S .
Remark 4.15.9. In the above definitions, if instead of the convergence xn ! x, we
assume that xn converges weakly to x, we obtain the notions of the second-order
weak projective contingent cone and the second-order weak projective adjacent
N uN / and PO 2W .S; x;
cone, denoted by P2W .S; x; N uN /, respectively.
We recall that the projective space P .X / associated with the space X is the set
of equivalence classes of pairs .x; s/ 2 X RC for the relation
.x; s/ .x;
O sO / if .x;
O sO / D .tx; ts/ for some t > 0:
We shall denote by P the quotient mapping P W X R ! P .X / and call it the

projective projection. Evidently, P .X / can be decomposed as the union
P .X / D X1 [ X0 ;
where X1 is the image of X f1g and X0 is the image of X f0g.

Therefore, P2 .S; x; N uN / can be split into two parts, namely the union of
P.T 2 .S; x;
N uN // f1g and P.T2 .S; x;
N uN // f0g where
4.15 Miscellaneous Second-Order Tangent Cones 205

2
N uN / D x 2 X j 9 ftn g # 0; 9 fxn g ! x; so that
T .S; x;

t2
xN C tn uN C n xn 2 S; 8 n 2 N ;
2

tn
T2 .S; x;
N uN / D x 2 X j 9 f.tn ; rn g # .0; 0/; with ! 0 9 fxn g ! x; with
rn

t2
xN C tn uN C n xn 2 S; 8 n 2 N ;
2rn
are the second-order contingent set and the second-order asymptotic contingent cone
studied earlier in this chapter.
It has been noted that the second-order contingent set can be empty, and,
seemingly to remedy this situation, has been one of the motivations behind
the second-order projective contingent cone. The following result clarifies this
statement:
Theorem 4.15.10. Let X be a reflexive Banach space, let S be a subset of X , let
xN 2 cl S and let uN 2 T .S; x/.
N Then P2W .S; x;
N uN / is nonempty.
Proof. We show that either T 2 .S; x; N uN / or T2 .S; x;
N uN / is nonempty where these
cones are considered to be equipped with the weak topology. Due to the containment
uN 2 T .S; x/,
N there exists a sequence ftn g 2 P with tn # 0 such that the sequence
fsn g given by sn WD tn1 d.xN C tn uN / converges to 0. Since 0 2 T 2 .S; x; N uN / if
sn D 0 for infinitely many n, we may assume that sn > 0 and set rn D 2stnn , and
wn D sn1 tn1 .zn xN tn uN /, where zn 2 S is such that kxN C tn uN zn k 2sn tn . Then
t2
rn1 tn D 2sn ! 0 and xN C tn uN C 2rnn wn D zn 2 S .
Taking a subsequence if necessary, we may assume that frn g ! r 2 Œ0; 1 and
fwn g has a weak limit u in 2BX . If r D 1, setting wN n D rn1 wn , we get fwN n g ! 0
N uN /. If r 2 P, the same choice shows that r 1 w 2 T 2 .x;
and 0 2 T 2 S; x; N uN /. Finally,
if r D 0, we have w 2 T2 .S; x; N uN /. t
u
In the following, we collect some of the useful properties of the projective
contingent cones of second-order and for their proof we refer the reader to the
original contribution of Penot [468].
Proposition 4.15.11. Let X and Y be reflexive Banach spaces, let g W X ! Y
be a mapping of class C 2 on an open subset S0 of X . Let S be a subset of X , let
xN 2 S0 \ S and let C be a subset of Y with g.S / C . Then for each uN 2 X , and
.w; r/ 2 P2 .S; x;
N uN /, we have
.g 1 .x/.w/
N C rg 00 .x/.N N g 0 .x/.N
N u; uN // 2 P2 .C; g.x/; N u//:
Proposition 4.15.12. Let X be a reflexive Banach space, let S be a convex subset

of X , let xN 2 S , and let uN 2 T .S; x/.
N Then, for any z 2 T .T .S; x/;
N uN /, .w; r/ 2
N uN /, we have .w C z; r/ 2 P2 .S; x;
P2 .S; x; N uN /.
For many other results on the above cones, the reader is referred to Penot [468].
4.15.3 Second-Order Tangent Cone of N. Pavel
We now study a notion that was introduced by Pavel [455] and later modified by
Pavel and Ursescu [457], Pavel, Huang, and Kim [456], and by Constantin [104–
106].
Definition 4.15.13. Let X be a normed space, let S X , and let xN 2 S . The
N is the collection of all x 2 X such that
second-order derivable cone T 2 .S; x/
there exists uN 2 X with

1 t2
lim 2 d xN C t uN C x; S D 0: (4.102)
t #0 t 2
It follows that (4.102) is equivalent to that uN 2 T .S; x/

N and for some ".t/ D o.t/,
we have
t2
xN C t uN C .x C ".t// 2 S: (4.103)
2
We have the following simple observation:
Proposition 4.15.14. Let X be a normed space, let S X , and let xN 2 S . Then,
N is a cone.
T 2 .S; x/
p
Proof. Let x 2 T 2 .S; x/
N and let
> 0. By taking t D
t in (4.102), we get
p t 2
1
lim d xN C t
N
u C .
x/ ; S D 0; (4.104)
t #0 t2 2
which at once ensures that

x 2 T 2 .S; x/.
N t
u
We have the following characteristic example (see [456]):
Example 4.15.15. Let X be a Hilbert space and let S WD fx 2 X j jxj D rg, and let
xN 2 S . Then the second-order derivable cone is given by T 2 .S; x/
N D fx 2 X j 9 uN 2
X such that hNu; xi
N D 0; jNuj2 C hx;
N xi D 0g.
We have the following useful result (see [456]):
Proposition 4.15.16. Let X be a normed space, let Y be a finite dimensional space,
and let g W X ! Y be a twice Fréchet differentiable function. Then, for S defined
by S WD fx 2 X j g.x/ D 0g, the second-order derivable cone at xN 2 S is given by
˚
N D x 2 X j 9 uN with g.x/
T 2 .S; x/ N D 0; hg 0 .x/;
N uN i D 0;

hg 00 .x/.N
N u/; uN i C hg 0 .x/;
N xi D 0 :
4.16 Second-Order Local Approximation 207
4.15.4 Connections Among the Second-Order Tangent Cones

p
As noted by Penot [468], by setting t D s r , we note that for r > 0, .w; r/ 2
N uN / if and only if .w; r/ 2 PO 2 .S; x;
TC2 .S; x; N uN / if and only if r 1 w 2 A2 .S; x;N uN / WD
2
lim inft !0C 2t .S xN t uN /. However, for .w; 0/ 2 TC 2 .S; x; N uN / if and only if
w 2 A.S; z/ WD lim inft !0C t 1 .S x/, N the first-order adjacent cone, and there is
no relation with the case .w; 0/ 2 PO 2 .S; x; N uN /.
It is also evident that TC2 .S; x; N uN / T 2 .S; x/ N RC . In fact, if .x; / 2
p
N uN /, then we have xN C t uN C t 2 .x C ".t// where ".t/ ! 0 as t ! 0.
TC2 .S; x;
Then, by noting that
p p t2
xN C t uN C t 2 .x C ".t// D xN C t uN C .2x C 2".t// 2 S;
2
we deduce that 2x 2 T 2 .S; x/,

N and since T 2 .S; x/
N is a cone, we conclude that
x 2 T .S; x/.
2
N
4.16 Second-Order Local Approximation
This section is devoted to a second-order local approximation of sets, which is

a natural generalization of the first-order local cone approximation studied in
Sect. 4.5. Our exposition of this section is based on the work by Castellani and
Pappalardo [86].
We begin by introducing the following notion of second-order local approxima-
tion:
Definition 4.16.1. Given a normed space X , a set-valued map K 2 W 2X X X !
2X is called a local approximation of second-order if it associates, to each set
S X and each pair .x;N uN / 2 X X , a set K 2 .S; x;
N uN / X such that
N uN / D K 2 .S; x;
t 2 K 2 .S; x; N t uN /
for every t
0 and satisfies the following properties:
1. N uN / D K 2 .S x;
K 2 .S; x; N 0; uN /
2. 2
K .S; x;N uN / D K 2 .S \ U.x/; N x; N uN /, for every U.x/
N
3. N uN / D ; if xN 62 cl S
K 2 .S; x;
4. N uN / D X if xN 2 int S .
K 2 .S; x;
5. N uN // D K 2 .`.S /; `.x/;
`.K 2 .S; x; N `.Nu// for any linear homomorphism ` W X !
X
6. S1 K 2 .S; x; N uN / 1 for every xN 2 cl S , and for every uN 2 cl K 2 .S; x;
N 0/.
Evidently, all the axioms of the above definition are satisfied when
N uN / D K.S; x/;
K 2 .S; x; N
where K.S; x/ N is a local cone approximation of first-order introduced in Sect. 4.5.

By following the arguments given for the local cone approximation of first-order,
it can be shown that the above axioms are independent or each other. That is, no
axiom can be expressed by the others.
Furthermore, it can be shown that the second-order radial tangent set, the second-
order interiorly radial tangent set, the second-order contingent set, the second-order
interiorly contingent set, the second-order adjacent set, and the second-order
interiorly adjacent set are all local approximations of second-order. We prove this
statement only for the second-order contingent set in the following result (for
simplicity, we drop the factor 1=2):
Theorem 4.16.2. Let X be a normed space, let S be a subset of X , let xN 2 S and
N uN / is a local approximation of
uN 2 X . The second-order contingent set T 2 .S; x;
second-order in the sense of Definition 4.16.1.
N uN / D
Proof. The first four axioms of Definition 4.16.1 and the fact that t 2 T 2 .S; x;
N t uN / for every t
0, are quite easy to verify and can also be seen from
T 2 .S; x;
the properties of the second-order contingent set stated earlier. Therefore, we will
only verify the last two axioms. We begin with a proof of the fifth axiom. Let x 2
N `.Nu// be arbitrary, where ` W X ! X is a linear homomorphism.
T 2 .`.S /; `.x/;
Then for every t > 0, there are xO 2 Ut .x/ and s 2 .0; t/ such that
N C s`.Nu/ C s 2 xO 2 `.S /;
`.x/
which, due to the continuity of `1 , is equivalent to the containment that
xN C s uN C s 2 `1 .x/
O 2 S:
By using the continuity of `1 once again, we deduce that for each t > 0 there
are xQ 2 Ut .`.x// and s 2 .0; t/ such that xN C s uN C s 2 xQ 2 S . This confirms
that x 2 `.T 2 .S; x; N uN // and hence we have shown that T 2 .`.S /; `.Nu/; `.x// N
2
`.T .S; x; N uN //. The converse inclusion is obtained by simply employing the used
arguments in a reverse order.
For the final axiom, let z 2 S1 and x 2 T 2 .S; x; N uN / be arbitrary. We will need to
show that for every t > 0 there are s 2 .0; t/ and zN with kNzk t such that xN C s uN C
s 2 .x C z C zN/ 2 S . We clearly have the existence of s 2 .0; t/ and zN with kzk t
such that xN C s uN C s 2 .x C zN/ 2 S which implies that .xN C s uN C s 2 .x C zN// C s 2 z 2 S .
u
The following result gives a mechanism to construct new notions of local
approximation of second-order from the known ones:
4.16 Second-Order Local Approximation 209
Theorem 4.16.3. Let X be a normed space, let S be a subset of X , and let .x; N uN / 2
X X . Let K 2 .S; x;
N uN / be a local approximation of second-order in the sense of
Definition 4.16.1. Then L2 .S; x; N uN / defined by the following relationship is also a
local approximation of second-order in the sense of Definition 4.16.1:
N uN / D X nK 2 .X nS; x;
L2 .S; x; N uN /:
Furthermore, if K 2 .; ; / is isotone, then for every S1 ; S2 X and every .x;

N uN / 2
X X , the following implication holds:
S1 \ S2 D ; H) K 2 .S1 ; x;
N uN / \ L2 .S2 ; x;
N uN / D ;: (4.105)
N uN /.
Proof. We will show that the six axioms of Definition 4.16.1 hold for L2 .S; x;
1. The first axiom holds due to the fact that
L2 .S x;
N 0; uN / D X nK 2 .X nS x;
N 0; uN / D X nK 2 .X nS; x;
N uN / D L2 .S; x;
N uN /:
N we have
2. For the second axiom, we note that for an arbitrary U.x/,
L2 .S \ U.x/;
N x;N uN / D X nK 2 .X n.S \ U.x/;
N x;N uN /
D X nK 2 .X nS [ .X nU.x//;
N x; N uN /
2
D X n K ..X nS \ U.x//N [ ..X nU.x//
N \ U.x//;
N x; N uN /
D X nK 2 .X nS \ U.x/;
N x;N uN /
D X nK 2 .QnS; x;
N uN /
D L2 .S; x;
N uN /:
3. Note that if xN 2 int S then xN … cl .X nS / and consequently
L2 .S; x; N uN / D X n; D X:
4. The proof for this axiom is analogous to the proof of the third axiom.
5. For a linear homomorphism ` W X ! X , we have
N `.Nu// D X nK 2 .X n`.S /; `.x/;

L2 .`.S /; `.x/; N `.Nu//
D `.X /n.K 2 .`.X nS /; x;
N uN //
D `.K 2 .X nS; x;
N uN //
D `.L2 .S; x;
N uN //:
6. Since .X nS /1 D S1 , we have

S1 D .X nS /1 K 2 .X nS; x;
N uN / 1

D X nK 2 .X nS; x;
N uN / 1 D L2 .S; x;N uN / 1 :
Therefore, all the six axioms have been verified and the proof is complete. u
t
N uN / is built from K 2 .S; x;
Note that if L2 .S; x; N uN / using (4.105) then we have
X nL2 .S; x;
N uN /:
4.17 Higher-Order Tangent Cones and Tangent Sets
In the previous sections, we witnessed numerous second-order tangent cones and

second-order tangent sets. These objects are natural extensions of tangent cones
of first-order. It turns out that all the second-order tangent cones and the second-
order tangent sets were either initially defined as higher-order approximation tools
or have a natural higher-order extension. See Constantin [104–106], Furukawa and
Yoshinaga [192], Hoffmann and Kornstaedt[265], Ledzewicz and Schättler [365,
365], Penot [468] and the cited references therein. In recent years, numerous
authors have focused on higher-order tangent sets and used them to define higher-
order derivatives and epiderivatives of set-valued maps. Higher-order tangent sets
have also been used to give higher-order optimality conditions and higher-order
sensitivity analysis. See Anh, Khanh, and Tung [8], Li, Sun, and Zhu [382],
Wang, Li, and Teo [586], and the cited references therein. Some interesting results
for nonsmooth optimization problems can be found in Ginchev, Guerraggio, and
Rocca [205, 206]. Although higher-order derivatives, epiderivatives, optimality
conditions, and sensitivity analysis are of interest, they typically follow the same
pattern of their second-order counterparts, and, in general, all the results which hold
for second-order study have their higher-order analogues. Therefore, in this work,
we will not dwell on the higher-order tangent sets and tangent cones but, for the
sake of completeness, we collect the definitions of higher-order extensions of the
most commonly used second-order tangent sets.
Definition 4.17.1. Let X be a normed space, let S be a nonempty subset of X , let
xN 2 cl .S /, and let uN 1 ; uN 2 ; : : : ; uN m1 be elements of X .
1. The m-th order radial tangent set Rm .S; x; N uN ; : : : ; uN m1 / of S at
N uN ; : : : ; uN m1 / is the set of all x 2 X for which there exists a sequence
.x;
ftn g P with tn # 0 such that
xN C tn uN C tn2 uN 2 C C tnm1 uN m1 C tnm x 2 S for every n 2 N:

4.17 Higher-Order Tangent Cones and Tangent Sets 211
2. The m-th order feasible tangent set IRm .S; x; N uN ; : : : ; uN m1 / of S at

N uN ; : : : ; uN m1 / is the set of all x 2 X such that for every ftn g P with
.x;
tn # 0, there is an m 2 N satisfying that
xN C tn uN C tn2 uN 2 C C tnm1 uN m1 C tnm x 2 S for every n

m:
3. The m-th-order contingent set T m .S; x; N uN ; : : : ; uN m1 / of S at .x;

N uN ; : : : ; uN m1 /
is the set of all x 2 X for which there are a sequence ftn g P with tn # 0 and a
sequence fxn g X with xn ! x such that
xN C tn uN C tn2 uN 2 C C tnm1 uN m1 C tnm xn 2 S for every n 2 N:
4. The m-th-order interiorly contingent set IT m S; x; N uN ; : : : ; uN m1 / of S at

N uN ; : : : ; uN m1 / is the set of all x 2 X such that for every sequence ftn g P
.x;
with tn # 0 and every sequence fxn g X with xn ! x there exists an index
m 2 N satisfying that
xN C tn uN C tn2 uN 2 C C tnm1 uN m1 C tnm xn 2 S for every n

m:
5. The m-th order adjacent set Am .S; x;N uN ; : : : ; uN m1 / of S at .x;

N uN ; : : : ; uN m1 / is
the set of all x 2 X such that for every sequence ftn g P with tn # 0, there
exists a sequence fxn g S with xn ! x satisfying that
xN C tn uN C tn2 uN 2 C C tnm1 uN m1 C tnm xn 2 S for every n 2 N:
6. The m-th order interiorly adjacent set IAm .S; x; N uN ; : : : ; uN m1 / of S at

N uN ; : : : ; uN m1 / is the set of all x 2 X such that there exists a sequence
.x;
ftn g P with tn # 0 so that for every sequence fxn g S with xn ! x there
exists an index m 2 N satisfying that
xN C tn uN C tn2 uN 2 C C tnm1 uN m1 C tnm xn 2 S for every n

m:
Clearly, when m D 2, we recover the second-order tangent sets (without the

factor 1=2). It is evident that the properties discussed for the second-order objects
can be extended to the higher-order tangent sets.
Chapter 5
Nonconvex Separation Theorems
In this chapter we introduce nonlinear scalarization methods that are very important
from the theoretical as well as computational point of view. We introduce differ-
ent scalarizing functionals and discuss their properties, especially monotonicity,
continuity, Lipschitz continuity, sublinearity, convexity. Using these nonlinear
functionals we show nonconvex separation theorems. These nonlinear functionals
are used for deriving necessary optimality conditions for solutions of set-valued
optimization problems in Sect. 12.8 and in different proofs, especially in the proof
minimal point theorems in Chap. 10. Moreover, we study characterizations of
solutions of set-valued optimization problems by means of nonlinear scalarizing
functionals. Finally, we present the extremal principle by Kruger and Mordukhovich
and discuss its relationship to separation properties of nonconvex sets. This extremal
principle will be applied in Sect. 12.9 for deriving a subdifferential variational
principle for set-valued mappings and in Sect. 12.11 in order to prove a first
order necessary condition for fully localized minimizers of set-valued optimization
problems.
5.1 Separating Functions and Examples
Separating functionals play an important role in many fields of mathematics;

especially, in order to prove minimal point theorems (see Chap. 10) and optimality
conditions (see Chap. 12). In this section we introduce an axiomatic approach for
scalarization by means of (in general nonlinear) functionals. We consider a linear
topological space Y , k 0 2 Y n f0g, a proper set D Y and a separating functional
' W Y ! R [ f˙1g having some of the following properties:
.Asf 1/ The function ' is D-monotone, i.e., y; w 2 Y , y 2 w D implies
' .y/ ' .w/.
.Asf 10 / The function ' is strictly C -monotone, i.e., y; w 2 Y , y 2 w .D n f0g/
implies ' .y/ < ' .w/.

DOI 10.1007/978-3-642-54265-7__5
214 5 Nonconvex Separation Theorems
.Asf 2/ The function ' is convex, i.e., for all y 1 ; y 2 2 Y ,

2 Œ0; 1 it holds
'.
y 1 C .1
/y 2 /
'.y 1 / C .1
/'.y 2 /:
.Asf 20 / The function ' is sublinear, i.e., ' is subadditive (for all y 1 ; y 2 2 Y
it holds
'.y 1 C y 2 / '.y 1 / C '.y 2 //
and positively homogenous (for all y 2 Y and

0 it holds '.
y/ D
'.y/).
.Asf 200 / The function ' is linear.
.Asf 3/ The function ' enjoys the property

8s 2 R; 8y 2 Y W ' y C sk0 D ' .y/ C s: (5.1)
.Asf 4/ The function ' is lower continuous.

.Asf 40 / The function ' is continuous.
Examples for functions satisfying the axioms given above are listed in the following:
Example 5.1.1. Assume that A is a proper closed subset of Y . Let k 0 2 Y n f0g
be such that A Œ0; C1/k 0 A. Consider the function ' W Y ! R [ f˙1g,
defined by
˚
'A;k 0 .y/ WD inf t 2 R j y 2 tk0 C A : (5.2)
We use the convention inf ; D C1. Then it holds dom 'A;k 0 D Rk 0 C A (see [214,
Theorem 2.3.1]).
If A D D is a proper closed convex cone and k 0 2 int D, the function (5.2) fulfills
.Asf 1/, .Asf 20 / .Asf 3/ and .Asf 40 / (see [214, Corollary 2.3.5] and Corollary 5.2.8 of
this section).
If A is a proper closed convex subset of Y with A Œ0; C1/k 0 A and D Y
with A D A, the functional (5.2) fulfills .Asf 1/, .Asf 2/, .Asf 3/ and .Asf 4/.
These properties of the functional (5.2) are shown in [214, Theorem 2.3.1] (see
Theorems 5.2.3, 5.2.6).
Example 5.1.2. The scalarization method by Pascoletti and Serafini [454] for
a vector optimization problem
minimize f .x/ subject to x 2 S; (VP)
(where Y D Rm , D D Rm C , S R , f .x/ D .f1 .x/; : : : ; fm .x// and fi W S ! R

n T
for all i D 1; : : : ; m) given by

5.1 Separating Functions and Examples 215
min t (5.3)
subject to the constraints
f .x/ 2 a C tr D;
x 2 S;
t 2 R;
with parameters a; r 2 Rm , r 2 int Rm

C , satisfies the axioms .Asf 1/, .Asf 2/, .Asf 3/,
.Asf 40 /.
Example 5.1.3. The following functional was introduced by Hiriart-Urruty [262]:
For a nonempty subset A of a Banach space Y with A ¤ Y , the oriented distance
function A W Y ! R is given as
A .y/ WD d.y; A/ d.y; Y n A/:
Zaffaroni has shown (see [607, Proposition 3.2]):

(i) A is Lipschitz of rank 1.
(ii) If A is convex, then A is convex and if A is a cone, then A is positively
homogeneous.
(iii) Assume that A D D is a proper closed convex cone. If y 1 ; y 2 2 Y with
y 1 y 2 2 A, then A .y 1 / A .y 2 /.
The functional A satisfies the axioms .Asf 1/, .Asf 20 /, .Asf 40 / if A D D is a
proper closed convex cone. Further results concerning the functional A are given
in Sect. 5.3.
Example 5.1.4. Certain nonlinear functionals are used in Mathematical Finance in
order to express a risk measure (for example a valuation of risky investments) with
respect to an acceptance set B Y . Let Y be a linear space of random variables,
and let ˝ be a set of elementary events (a set of all possible states of the future).
Then a future payment of an investment is a random variable y W ˝ ! R. Positive
payments in the future are wins, negative ones are losses. If no investment is being
done, then y takes on the value zero. In order to valuate such an investment, we
need to valuate random variables by comparing them. To do that, we introduce an
ordering relation that is induced by a set B Y . Artzner, Delbaen, Eber and Heath
proposed in [12] axioms for a cone B Y of random variables that represent
acceptable investments:
(A1) fy 2 Y jy.!/
0 .! 2 ˝/g B; B \ fy 2 Y jy.!/ < 0 .! 2 ˝/g D ;,
(A2) B C B B.
Cones B Y satisfying the axioms (A1) and (A2) of acceptable investments can
be used in order to introduce a preference relation on Y . The decision maker prefers
y1 to y2 (changing from y2 to y1 is an acceptable risk) if and only if y1 y2 is an
element of B, i.e.,
y1
B y2 ” y1 y2 2 B:
Artzner, Delbaen, Eber and Heath [12] (compare Heyde [256]) introduced coherent
risk measures W Y ! R. In the papers by Artzner, Delbaen, Eber and Heath [12]
and Rockafellar, Uryasev and Zabarankin [488, 498] (compare also Rockafellar,
Uryasev [496, 497]) the following properties of coherent risk measures are
supposed:
.P1/ .y C tk0 / D .y/ t,
.P 2/ .0/ D 0 and .
y/ D
.y/ for all y 2 Y and
> 0,
.P 3/ .y 1 C y 2 / .y 1 / C .y 2 / for all y 1 ; y 2 2 Y ,
.P 4/ .y 1 / .y 2 / if y 1
B y 2 .
The sublevel set L .0/ D B of to the level 0 is a convex cone and corresponds
to the acceptance set B. It can be shown (see [498]) that a coherent risk measure
admits a representation as
.y/ D infft 2 R j y C tk0 2 Bg: (5.4)
Furthermore, Heyde [256] has shown that a coherent risk measure can be identified
with the functional 'B;k 0 .y/ (see (5.2)) by
'B;k 0 .y/ D .y/:
We get corresponding properties .Asf 1/, .Asf 20 /, .Asf 3/ for the functional .y/
like in Example 5.1.1 for the functional 'A;k 0 depending from the properties of the
set A D B, i.e., of the acceptance set B in Mathematical Finance.
Examples for coherent risk measures are the conditional value at risk (see
[191, Section 4.4, Definition 4.43]) and the worst-case risk measure (see [191,
Example 4.8]).
Example 5.1.5 (Value at Risk). Let ˝ be a fixed set of scenarios. A financial
position is described by a mapping x W ˝ ! R and x belongs to a given class
X of financial positions. Here x.!/ is the discounted net worth of the position at
the end of the trading period if the scenario ! 2 ˝ is realized. Now, the aim is to
quantify the risk of x by some real number .x/. Assume that X is the linear space
of bounded measurable functions (containing the constants) on some measurable
space .˝; S /. Furthermore, let P be a probability measure on .˝; S /. A position
x is considered to be acceptable if the probability of a loss is bounded by a given
level
2 .0; 1/, i.e., if P Œx < 0
. The corresponding monetary risk measure
V @R
, defined by
V @R
.x/ WD inffm 2 R j P .m C x < 0/
g
is called Value at Risk at level

. V @R
is the smallest amount of capital which,
if added to x and invested in the risk-free asset, keeps the probability of a negative
5.2 Nonlinear Separation 217
outcome below the level

. The Value at Risk only controls the probability of a loss.
It is important to mention that the Value at Risk does not capture the size of such a
loss if it occurs.
V @R
is positively homogeneous but in general it is not convex (see Föllmer and
Schied [191], Example 4.11), this means that .Asf 2/ and .Asf 20 / are not fulfilled.
Example 5.1.6 (Worst-Case Risk Measure). Consider the worst-case risk measure
max defined by
max .x/ WD inf x.!/ for all x 2 X ;

!2˝
where ˝ is a fixed set of scenarios, x W ˝ ! R, x belongs to a given class X of

financial positions and X is the linear space of bounded measurable functions on
some measurable space .˝; S /. The value max .x/ is the least upper bound for the
potential loss that can occur in each scenario. max is a coherent risk measure (see
Föllmer and Schied [191, Example 4.8]) such that we get the properties mentioned
in Example 5.1.4.
5.2 Nonlinear Separation
In order to show minimal-point theorems and corresponding variational principles

in Chap. 10 or optimality conditions in Chap. 12 we use a scalarization method by
means of certain nonlinear functionals. In this section we discuss useful properties
of these functionals (cf. Göpfert, Riahi, Tammer, Zălinescu [214] and Tammer,
Zălinescu [560]).
5.2.1 Construction of Scalarizing Functionals
Having a nonempty subset A of a real linear space Y and an element k 0 ¤ 0

of Y , Gerstewitz [196] introduced the function (5.5) (see Fig. 5.1) as scalarizing
functional in vector optimization. In the papers by Gerstewitz, Iwanow [197], Gerth,
Weidner [198] and Göpfert, Tammer, Zălinescu [215] the essential properties for
the usage of the functional (5.5) as scalarizing functional in vector optimization
are derived, especially monotonicity properties. The construction in (5.5) was
mentioned by Krasnosel’ski [340] (see Rubinov [511]) in the context of operator
theory. Furthermore, the functional (5.5) is used by Pascoletti, Serafini [454] as
scalarizing functional in multicriteria optimization.
'A WD 'A;k 0 W Y ! R; 'A;k 0 .y/ WD infft 2 R j y 2 tk0 C Ag; (5.5)

Fig. 5.1 Level sets of the function 'A;k0 from (5.5), where A D K D R2C and k 0 2 int K hold
where, as usual, inf ; WD C1 (and sup ; WD 1); we use also the convention
.C1/ C .1/ WD C1.
This function was used by many authors, mainly for scalarization of vector
optimization problems. Luenberger [408, Definition 4.1] considered
.gI y/ WD inff 2 R j y g 2 Y g;
the corresponding function being called the shortage function associated to the
production possibility set Y Rm and g 2 Rm C n f0g. The case when g D
.1; : : : ; 1/ was introduced earlier by Bonnisseau and Cornet [54]. A similar function
is introduced in [408, Definition 2.1] under the name of benefit function.
More recently such a function was considered in the context of mathematical
finance beginning with Artzner et. al. [12]; see Heyde [256] (see Example 5.1.4).
Under the name of topical function such functions were studied by Singer and
his collaborators (see [510]). We discuss many important properties of 'A;k 0
in Sect. 5.2.2. Moreover, we study local continuity properties in Sect. 5.2.4.
Bonnisseau and Crettez [55] obtained local Lipschitz properties for 'A;k 0 (called
Luenberger shortage function in [55]) in a very special case, more general results
are given by Tammer and Zălinescu [560]. Of course, 'A;k 0 is a continuous sublinear
functional if A is a proper closed convex cone and k 0 2 int A (see Corollary 5.2.8)
and so 'A;k 0 is Lipschitz continuous. Such Lipschitz properties of 'A;k 0 are of
interest also in the case when A Y is an arbitrary (convex) set and the interior
of the usual ordering cone in Y is empty like in mathematical finance where the
acceptance sets are in function spaces as Lp and the corresponding risk measures
are formulated by means of 'A;k 0 (see e.g. Föllmer and Schied [190, 191]).
5.2.2 Properties of Scalarization Functions
Throughout this section Y is a separated locally convex space and Y is its

topological dual, C Y is a proper closed convex cone, k 0 2 C n .C / and
A Y is a nonempty set.
The recession cone of the nonempty set A X is the set
A1 WD fu 2 X j x C tu 2 A 8x 2 A; 8t 2 RC g:
It follows easily that A1 is a convex cone; A1 is also closed when A is closed. If

A is a closed convex set then A1 D \t 2P t.A a/, where P WD 0; C1Œ and a 2 A
(A1 does not depend on a 2 A).
Furthermore, we assume that A satisfies the following condition (see also [55]):
(A1) A is closed, satisfies the free-disposal assumption A C D A, and A 6D Y .
We shall use also the (stronger) condition:
(A2) A is closed, satisfies the strong free-disposal assumption A .C n f0g/ D
int A, and A 6D Y .
Because A C D A [ .A .C n f0g//, we have that .A2/ ) .A1/. Moreover,
the condition A .C n f0g/ D int A is equivalent to A .C n f0g/ int A.
Remark 5.2.1. Assume that the nonempty set A satisfies assumption .A2/. Then C
is pointed, that is, C \ .C / D f0g, and A Pk 0 int A for k 0 2 C n f0g.
The last assertion is obvious. For the first one, assume that k 2 C \ .C / n f0g.
Take a 2 bd A . A/; such an a exists because A ¤ Y . Then a0 WD a k 2 int A
A, and so a D a0 .k/ 2 int A, a contradiction.
Remark 5.2.2. When A satisfies condition (A1) or (A2) with respect to C and k 0 2
C n.C / then A satisfies condition (A1) or (A2), respectively, with respect to RC k 0 .
In fact in many situations it is sufficient to take C D RC k 0 for some k 0 2 Y n f0g.
In such a situation (A1) [resp. (A2)] means that A is a closed proper subset of Y and
A RC k 0 D A [resp. A Pk 0 int A].
The free-disposal condition A D A C shows that C A1 . As observed
above A1 is also closed because A is closed. Hence A1 is the largest closed
convex cone C verifying the free-disposal assumption A D A C .
The aim of this section is to find a suitable functional ' W Y ! R and conditions
such that two given nonempty subsets A and H of Y can be separated by '.
To A Y satisfying (A1) and k 0 2 C n .C / we associate the function 'A;k 0
defined in (5.5). We consider the set
A0 WD f.y; t/ 2 Y R j y 2 tk0 C Ag:
The assumption on A shows that A0 is of epigraph type, i.e. if .y; t/ 2 A0 and

t 0
t, then .y; t 0 / 2 A0 . Indeed, if y 2 tk0 C A and t 0
t, since
tk0 C A D t 0 k 0 C A .t 0 t/k 0 t 0 k 0 C A;
(because of (A1)) we obtain that .y; t 0 / 2 A0 . Also observe that A0 D T 1 .A/,

where T W Y R ! Y is the continuous linear operator defined by T .y; t/ WD tk0 C
y. So, if A is closed (convex, cone), then A0 is closed (convex, cone). Obviously, the
domain of 'A is the set Rk 0 C A and A0 epi 'A cl A0 (because A0 is of epigraph
type), from which it follows that A0 D epi 'A if A is closed, and so 'A is a lower
semicontinuous (l.s.c.) function.
In the next results we collect several useful properties of 'A (compare Göpfert,
Riahi, Tammer, Zălinescu [214] and Tammer, Zălinescu [560]).
Theorem 5.2.3. Assume that C Y is a proper closed convex cone, k 0 2 C n.C /
and A Y is a nonempty set. Furthermore, suppose
(A1) A is closed, satisfies the free-disposal assumption A C D A, and A 6D Y .
Then 'A (defined in (5.5)) is l.s.c., dom 'A D Rk 0 C A,
fy 2 Y j 'A .y/
g D
k 0 C A 8
2 R; (5.6)
and
'A .y C
k 0 / D 'A .y/ C
8 y 2 Y; 8
2 R: (5.7)
Moreover,
(a) 'A is convex if and only if A is convex; 'A .
y/ D
'A .y/ for all
> 0 and
y 2 Y if and only if A is a cone.
(b) 'A is proper if and only if A does not contain lines parallel to k 0 , i.e.,
8 y 2 Y; 9 t 2 R W y C tk0 … A: (5.8)
(c) 'A is finite-valued if and only if A does not contain lines parallel to k 0 and
Rk 0 C A D Y: (5.9)
(d) Let B Y ; 'A is B-monotone if and only if A B A.

(e) 'A is subadditive if and only if A C A A.
Proof. We have already observed that dom 'A D Rk 0 C A and 'A is l.s.c. when A
is closed. From the definition of 'A the inclusion in (5.6) is obvious, while the
converse inclusion is immediate, taking into account the closedness of A. Formula

(5.7) follows easily from (5.6).
(a) Since the operator T defined above is onto and epi 'A D T 1 .A/, we have that
epi 'A is convex (cone) if and only if A D T .epi 'A / is so. The conclusion
follows.
(b) We have
'A .y/ D 1 , y 2 tk0 C A 8t 2 R , fy C tk0 j t 2 Rg A:
The conclusion follows.

(c) The conclusion follows from (b) and the fact that dom 'A D Rk 0 C A.
(d) Suppose first that A B A and take y1 ; y2 2 Y with y2 y1 2 B. Let t 2 R
be such that y2 2 tk0 C A. Then y1 2 y2 B tk0 C .A B/ tk0 C A, and so
'A .y1 / t. Hence 'A .y1 / 'A .y2 /. Assume now that 'A is B-monotone and
take y 2 A and b 2 B. From (5.6) we have that 'A .y/ 0. Since y .y b/ 2
B, we obtain that 'A .y b/ 'A .y/ 0, and so, using again (5.6), we obtain
that y b 2 A.
(e) Suppose first that A C A A and take y1 ; y2 2 Y . Let ti 2 R be such that yi 2
ti k 0 C A for i 2 f1; 2g. Then y1 C y2 2 .t1 C t2 /k 0 C .A C A/ .t1 C t2 /k 0 C A,
and so 'A .y1 C y2 / t1 C t2 . It follows that 'A .y1 C y2 / 'A .y1 / C 'A .y2 /.
Assume now that 'A is subadditive and take y1 ; y2 2 A. From (5.6) we have
that 'A .y1 /; 'A .y2 / 0. Since 'A is subadditive, we obtain that 'A .y1 C y2 /
'A .y1 /C'A .y2 / 0, and so, using again (5.6), we obtain that y1 Cy2 2 A. u t
Remark 5.2.4. Under assumption (A1), from Theorem 5.2.3 we get that 'A is l.s.c.,
A D fy 2 Y j 'A .y/ 0g; int A fy 2 Y j 'A .y/ < 0g; (5.10)
and so
bd A D A n int A fy 2 Y j 'A .y/ D 0g: (5.11)
In general the inclusion in (5.11) is strict.

Example 5.2.5. Consider C WD R2C , k 0 WD .1; 0/ and
A WD . 1; 0 1; 0/ [ .Œ0; C1Œ 1; 1/ :
Then 'A .u; v/ D 1 for v 1, 'A .u; v/ D u for v 2 .1; 0 and 'A .u; v/ D C1
for v > 0. In particular, 'A .0; 1/ D 1 and .0; 1/ 2 bd A (see Fig. 5.2).
Theorem 5.2.6. Assume that C Y is a proper closed convex cone, k 0 2 C n.C /
and A Y is a nonempty set. Furthermore, suppose
int A, and A 6D Y .
Fig. 5.2 y 2 bd A with

'A .y/ D 1 in Example
5.2.5
Then (5.6), (5.7), (a), (b), (c) from Theorem 5.2.3 hold, and moreover
(f) 'A is continuous and
fy 2 Y j 'A .y/ <

g D
k 0 C int A; 8
2 R; (5.12)
fy 2 Y j 'A .y/ D
g D
k 0 C bd A; 8
2 R: (5.13)
(g) If 'A is proper, then
'A is B-monotone , A B A , bd A B A:
Moreover, if 'A is finite-valued, then
'A strictly B-monotone , A .B n f0g/ int A , bd A .B n f0g/ int A:
(h) Assume that 'A is proper; then
'A is subadditive , A C A A , bd A C bd A A:
Proof. Suppose now that (A2) holds.

(f) Let
2 R and take y 2
k 0 C int A. Since y
k 0 2 int A, there exists " > 0
such that y
k 0 C "k 0 2 A. Therefore 'A .y/
" <
, which shows
that the inclusion always holds in (5.12). Let
2 R and y 2 Y be such
that 'A .y/ <
. There exists t 2 R, t <
, such that y 2 tk0 C A. It follows
with (A2) that y 2
k 0 C A .
t/k 0
k 0 C int A. Therefore (5.12) holds,
and so 'A is upper semicontinuous. Because 'A is also lower semicontinuous,
we have that 'A is continuous. From (5.6) and (5.12) we obtain immediately
that (5.13) holds.
(g) Let us prove the second part, the first one being similar to that of (and partially
proved in) Theorem 5.2.3 (d). So, let 'A be finite-valued.
Assume that 'A is strictly B-monotone and take y 2 A and b 2 B n f0g.
From (5.6) we have that 'A .y/ 0, and so, by hypothesis, 'A .yb/ < 0. Using
(5.12) we obtain that y b 2 int A. Assume now that bd A .B n f0g/ int A.
Consider y1 ; y2 2 Y with y2 y1 2 B n f0g. From (5.13) we have that y2 2
'A .y2 /k 0 Cbd A, and so y1 2 'A .y2 /k 0 Cbd A.B nf0g/ 'A .y2 /k 0 Cint A.
From (5.12) we obtain that 'A .y1 / < 'A .y2 /. The remaining implication is
obvious.
(h) Let 'A be proper. One has to prove bd A C bd A A ) 'A is subadditive.
Consider y1 ; y2 2 Y . If fy1 ; y2 g 6 dom 'A , there is nothing to prove; hence let
y1 ; y2 2 dom 'A . Then, by (5.13), yi 2 'A .yi /k 0 C bd A for i 2 f1; 2g, and so
y1 C y2 2 .'A .y1 / C 'A .y2 //k 0 C .bd A C bd A/ .'A .y1 / C 'A .y2 //k 0 C A.
Therefore 'A .y1 C y2 / 'A .y1 / C 'A .y2 /. t
u
When k 0 2 int C we get an additional important property of 'A (see also
Theorem 5.2.12).
Corollary 5.2.7. Assume that C Y is a proper closed convex cone, k 0 2 int C
and A Y satisfies condition (A1). Then 'A is finite-valued and continuous.
Proof. Because k 0 2 int C we have that Rk 0 C C D Y . From Theorem 5.2.3 (c) it
follows that
dom 'A D A C Rk 0 D A C C Rk 0 D A C Y D Y:
Assuming that 'A is not proper, from Theorem 5.2.3 (c) we get y C Rk 0 A for
some y 2 Y . Then Y D y C Rk 0 C A C D A, a contradiction. Hence 'A is
finite-valued.
Moreover, we have that A Pk 0 A int C int.A C / D int A. Applying
Theorem 5.2.6 (f) for C replaced by RC k 0 we obtain that 'A is continuous. t
u
From the preceding results we get the following particular case.
Corollary 5.2.8. Let C Y be a proper closed convex cone and k 0 2 int C . Then
'C W Y ! R; 'C .y/ WD infft 2 R j y 2 tk0 C g
is a well-defined continuous sublinear function such that for every

2 R,
fy 2 Y j 'C .y/
g D
k 0 C; fy 2 Y j 'C .y/ <
g D
k 0 int C:
Moreover, 'C is strictly .int C /-monotone.

Proof. The assertions follow using Theorem 5.2.6 and Corollary 5.2.7 applied for
A WD C . For the last part note that C C int C D int C . t
u
Now all preliminaries are done, and we can prove the following nonconvex
separation theorem.
Theorem 5.2.9 (Nonconvex Separation Theorem). Let A Y be a closed
proper set with nonempty interior, H Y a nonempty set such that H \ int A D ;.
Let C Y be a proper closed convex cone and k 0 2 int C . Furthermore, assume
int A, and A 6D Y .
Then 'A defined by (5.5) is a finite-valued continuous function such that
'A .x/
0 > 'A .y/ 8 x 2 H; 8 y 2 int AI (5.14)
moreover, 'A .x/ > 0 for every x 2 int H .

Proof. By Corollary 5.2.7 'A is a finite-valued continuous function. By Theo-
rem 5.2.6 (f) we have that int A D fy 2 Y j 'A .y/ < 0g, and so (5.14) obviously
holds.
Take y 2 int H I then there exists t > 0 such that y tk0 2 H . From (5.7) and
(5.12) we obtain that 0 'A .y tk0 / D 'A .y/ t, whence 'A .y/ > 0. t
u
Of course, if we impose additional conditions on A, we have additional properties
of the separating functional 'A (see Theorems 5.2.3 and 5.2.6).
5.2.3 Continuity Properties
If A is a proper closed subset of Y (hence ; ¤ A ¤ Y ) and A Pk 0 int A,

applying Theorem 5.2.6 for C WD RC k 0 we obtain that 'A is continuous (on Y )
and (5.13) holds. In the next result we characterize the continuity of 'A at a point
y0 2 Y (compare Tammer and Zălinescu [559, 560, 614]).
Proposition 5.2.10. Assume that C Y is a proper closed convex cone, k 0 2 C n
.C / and A Y is a nonempty set satisfying condition (A1). Then the function 'A
is (upper semi-) continuous at y0 2 Y if and only if y0 'A .y0 /; C1Œ k 0 int A.
Proof. If 'A .y0 / D C1 it is clear that 'A is upper semicontinuous at y0 and the
inclusion holds. So let 'A .y0 / < C1.
Assume first that 'A is upper semicontinuous at y0 . Let
2 'A .y0 /; C1Œ. Then
there exists a neighbourhood V of y0 such that 'A .y/ <
for every y 2 V . It
follows that for y 2 V we have y 2
k 0 C A, that is, V
k 0 C A. Hence
y0 2
k 0 C int A, whence y0
k 0 2 int A.
Assume now that y0 'A .y0 /; C1Œ k 0 int A and take 'A .y/ <
< C1.
Then, by our hypothesis, V WD
k 0 C A is a neighbourhood of y0 and from the
definition of 'A we have that 'A .y/
for every y 2 V . Hence 'A is upper
semicontinuous at y0 . t
u
Corollary 5.2.11. Under the hypotheses of Proposition 5.2.10 assume that 'A is
continuous at y0 2 bd A. Then 'A .y0 / D 0.
Proof. Of course, 'A .y0 / 0. If 'A .y0 / < 0, from the preceding proposition we
obtain the contradiction y0 D y0 0k 0 2 int A. t
u
5.2.4 Lipschitz Properties
The primary goal of this section is to study local Lipschitz properties of the
functional 'A;k 0 under as weak as possible assumptions concerning the subset
A Y and k 0 2 Y (compare Tammer and Zălinescu [560]).
When A is a convex set, as noticed above, 'A is convex. In such a situation from
the continuity of 'A at a point in the interior of its domain one obtains the local
Lipschitz continuity of 'A on the interior of its domain (if the function is proper).
Moreover, when A D C and k 0 2 int C then (it is well known that) 'A is a
continuous sublinear function, and so 'A is Lipschitz continuous.
Recently in the case Y D Rm and for C D Rm C Bonnisseau–Crettez [55] obtained
the Lipschitz continuity of 'A around a point y 2 bd A when k 0 is in the interior
of the Clarke tangent cone of A at y. The (global) Lipschitz continuity of 'A can be
related to a result of Gorokhovik–Gorokhovik [217] established in normed vector
spaces as we shall see in the sequel.
Theorem 5.2.12. Assume that C Y is a proper closed convex cone, k 0 2 C n
.C / and A Y is a nonempty set satisfying condition (A1).
(i) One has
'A .y/ 'A .y 0 / C 'C .y y 0 / 8y; y 0 2 Y: (5.15)
(ii) If k 0 2 int C then 'A is finite-valued and Lipschitz on Y .

Proof. (i) By Theorem 5.2.3 (applied for A and A WD C , respectively) we have
that 'A and 'C are lower semicontinuous functions, 'C being sublinear and
proper.
Let y; y 0 2 Y . If 'A .y 0 / D C1 or 'C .y y 0 / D C1 it is nothing to
prove. In the contrary case let t; s 2 R be such that y y 0 2 tk0 C and
y 0 2 sk0 C A. Then, taking into account assumption .A1/
y 2 tk0 C C sk0 C A D .t C s/k 0 C .A C / D .t C s/k 0 C A:
It follows that 'A .y/ t C s. Passing to infimum with respect to t and s

satisfying the preceding relations we get (5.15).
(ii) Assume that k 0 2 int C . Let V Y be a symmetric closed and convex
neighbourhood of 0 such that k 0 C V C and let pV W Y ! R be the
Minkowski functional associated to V ; then pV is a continuous seminorm and
V D fy 2 Y j pV .y/ 1g. Let y 2 Y and t > 0 such that y 2 tV. Then

t 1 y 2 V k 0 C , whence y 2 tk0 C . Hence 'C .y/ t. Therefore,
'C .y/ pV .y/. This inequality confirms that .Rk 0 C D/ dom 'C D Y .
Moreover, since 'C is sublinear we get 'C .y/ 'C .y 0 / C pV .y y 0 /
and so
ˇ ˇ
ˇ'C .y/ 'C .y 0 /ˇ pV .y y 0 / 8y; y 0 2 Y; (5.16)
that is, 'C is Lipschitz.

By Corollary 5.2.7 we have that 'A is finite-valued (and continuous). From
(5.15) we have that 'A .y/ 'A .y 0 / 'C .y y 0 / pV .y y 0 /, whence
(interchanging y and y 0 )
ˇ ˇ
ˇ'A .y/ 'A .y 0 /ˇ pV .y y 0 / 8y; y 0 2 Y: (5.17)
Hence 'A is Lipschitz continuous (on Y ). t

u
Note that the condition A .C n f0g/ int A does not imply that 'A is proper.
Example 5.2.13. Take A WD f.x; y/ 2 R2 j y
jxj1 g, with the convention
01 WD C1, and C WD RC k 0 with k 0 WD .0; 1/. Then A .C n f0g/ D int A and
'A .0; 1/ D 1.
Note that, with our notation, [55, Proposition 7] asserts that 'A;k 0 is finite and
locally Lipschitz provided Y D Rn , C D RnC and k 0 2 int C , which is much less
than the conclusion of Theorem 5.2.12(ii).
Of course, in the conditions of Theorem 5.2.12(ii) we have that k 0 2 int A1
because C A1 . In fact we have also a converse of Theorem 5.2.12(ii).
Proposition 5.2.14. Assume that C Y is a proper closed convex cone, k 0 2
C n .C / and A Y is a nonempty set satisfying condition (A1). If 'A is finite-
valued and Lipschitz then k 0 2 int A1 .
Proof. By hypothesis there exists a closed convex and symmetric neighbourhood V
of 0 such that (5.17) holds. We have that A D fy 2 Y j 'A .y/ 0g. Let y 2 A,
v 2 V and ˛
0. Then
'A .y C ˛.v k 0 // 'A .y C ˛v/ ˛ 'A .y/ C ˛pV .v/ ˛ 0
because V D fy 2 Y j pV .y/ 1g. Hence V k 0 A1 , which shows that

k 0 2 int A1 . t
u
Corollary 5.2.15. Under the assumptions of Proposition 5.2.14, the function 'A is
finite-valued and Lipschitz if and only if k 0 2 int A1 .
Proof. The necessity is given by Proposition 5.2.14. Assume that k 0 2 int A1 .
Taking C WD A1 , using Theorem 5.2.12(ii) we obtain that 'A is finite-valued and
Lipschitz. t
u
If int C ¤ ; and k 0 … int C , 'C is not finite-valued, and so it is not Lipschitz.

One may ask if the restriction of 'C at its domain is Lipschitz. The next examples
show that both situations are possible.
Example 5.2.16. Take C D R2C and k 0 D .1; 0/. We have that 'C .y1 ; y2 / D y1
for y2 0, 'C .y1 ; y2 / D C1 for y2 > 0, and so 'C jdom 'C is Lipschitz.
˚
Example 5.2.17. Take C WD .u; v; w/ 2 R3 j v; w
0; u2 vw and k 0 WD
.0; 0; 1/; then
8
< C1 if y > 0 or Œy D 0 and x ¤ 0;
'C .x; y; z/ D z if x D y D 0;
:
z x =y if y < 0:
2
It is clear that the restriction of 'C at its domain is not continuous at .0; 0; 0/ 2
dom 'C and the restriction of 'C at the interior of its domain is not Lipschitz.
However, 'C is locally Lipschitz on the interior of its domain.
The last property mentioned in the previous example is a general one for 'A when
A is convex.
Proposition 5.2.18. Let A be a proper closed subset of Y and k 0 2 Y nf0g be such
that A RC k 0 D A. If A is convex, has nonempty interior, and does not contain
any line parallel with k 0 (or equivalently k 0 … A1 ), then 'A is locally Lipschitz on
int.dom 'A / D Rk 0 C int A.
Proof. Because A does not contain any line parallel with k 0 , 'A is proper (see
Theorem 5.2.3 taking into account assumption .A1/). We know that dom 'A D
Rk 0 C A, and so int.dom 'A / D int.Rk 0 C A/ D Rk 0 C int A (see, e.g., [614,
Exercise 1.4]). On the other hand it is clear that A fy 2 Y j 'A .y/ 0g. Since
int A ¤ ;, we have that 'A is bounded above on a neighbourhood of a point, and
so 'A is locally Lipschitz on int.dom 'A / D Rk 0 C int A (see e.g. [614, Corollary
2.2.13]). t
u
We have seen in Theorem 5.2.12 that 'A is Lipschitz even if A is not convex
when k 0 2 int C . So, in the sequel we are interested by the case in which A is not
convex, k 0 … int C and A does not contain any line parallel with k 0 .
Note that for A not convex and y 2 int.dom 'A / we can have situations in which
'A is not continuous at y or 'A is continuous but not Lipschitz around y.
Example 5.2.19. Take C WD R2C , k 0 WD .1; 0/ and
A1 WD . 1; 0 1; 1/ [ .Œ0; 1 1; 0/

A2 WD f.a; b/ j a 2 0; 1Œ; b a2 g [ . 1; 0 1; 1/ :
Then
8 8
< C1 if v > 1; < C1 if v > 1;
'A1 ;k 0 .u; v/ D u if 0 < v 1; 'A2 ;k 0 .u; v/ D u if 0 < v 1;
: : p
u 1 if v 0; u v if v 0:
It is clear that .0; 0/ 2 int.dom 'A1 / but 'A1 is not continuous at .0; 0/, and .0; 0/ 2
int.dom 'A2 /, 'A2 is continuous at .0; 0/ but 'A2 is not Lipschitz at .0; 0/.
In what concerns the Lipschitz continuity of 'A around a point y 2 dom 'A in
finite dimensional spaces this can be obtained using the notion of epi-Lipschitzianity
of a set as introduced by Rockafellar [491] (see also [493]). We extend this notion
in our context. We say that the set A Y is epi-Lipschitz at y 2 A in the direction
v 2 Y n f0g if there exist " > 0 and a (closed convex symmetric) neighbourhood V0
of 0 in Y such that
8y 2 .y C V0 / \ A; 8w 2 v C V0 ; 8
2 Œ0; " W y C
w 2 A: (5.18)
Note that (5.18) holds for v D 0 if and only if y 2 int A. Moreover, if y 2 int A
then A is epi-Lipschitz at y 2 A in any direction.
Theorem 5.2.20. Let A be a proper closed subset of Y and k 0 2 Y n f0g be such
that A RC k 0 D A. Assume that y0 2 Y is such that 'A .y0 / 2 R. Then 'A is
finite and Lipschitz on a neighbourhood of y0 if and only if A is epi-Lipschitz at
y WD y0 'A .y0 /k 0 in the direction k 0 .
Proof. Using (5.7) we get 'A .y/ D 0. Recall also that A D fy 2 Y j 'A .y/ 0g
and the finite values of 'A are attained (because A is closed).
Assume that there exist a closed convex symmetric neighbourhood V of 0 in
Y and p W Y ! R a continuous seminorm such that 'A is finite on y0 C V and
j'A .y/ 'A .y 0 /j p.y y 0 / for all y; y 0 2 y0 C V . Taking into account (5.7), we
have that 'A is finite on y C V and
ˇ ˇ
ˇ'A .y/ 'A .y 0 /ˇ p.y y 0 / 8y; y 0 2 y C V:
Take V0 WD fy 2 13 V j p.y/ 1g and " 2 0; 1 such that "k 0 2 V0 . Let us

show that (5.18) holds with v replaced by k 0 . For this take y 2 .y C V0 / \ A,
w 2 k 0 C V0 and
2 Œ0; ". Then y
k 0 y 2 V0 C V0 V and y C
w y D
y
k 0 y C
.w C k 0 / 2 V0 C V0 C V0 V , and so
'A .y C
w/ 'A .y
k 0 / C p.
.w C k 0 // D 'A .y/
C
p.w C k 0 /

.p.w C k 0 / 1/ 0:
Hence y C
w 2 A.
Assume now that (5.18) holds with v replaced by k 0 . Let r 2 0; " be such that
2r.1 C p.k 0 // < 1, where p WD pV0 . Of course, fy j p.y/
g D
V0 for every
> 0 and if p.y/ D 0 then y 2

V0 for every
> 0. Set
M WD fy 2 y C rV 0 j j'A .y/j p.y y/gI
of course, y 2 M . We claim that M D y C rV 0 . Consider y 2 M , w 2 V0 and
2 Œ0; r. Setting y 0 WD y 'A .y/k 0 2 A, we have that 'A .y 0 / D 0 and

p.y 0 y/ p.y y/ C j'A .y/j p.k 0 / r 1 C p.k 0 / < 1
2 1; (5.19)
and so, by (5.18), y 0 C

.w k 0 / 2 A; hence 'A .y 0 C
w/
.
Take v 2 rV 0 . On one hand one has

1
'A .y 0 C v/ D 'A y 0 C p.v/ v p.v/
p.v/
if p.v/ > 0, and 'A .y 0 C v/ D 'A .y 0 C

.
1 v//
for every
2 0; r, whence
'A .y 0 C v/ 0 D p.v/. Therefore, 'A .y 0 C v/ p.v/.
On the other hand, assume that 'A .y 0 Cv/ < p.v/. Because 2r.1Cp.k 0 // < 1,
there exists t > 0 such that r C .t C r/p.k 0 / 1=2 and 'A .y 0 C v/ < p.v/ t DW
t 0 < 0. It follows that y 0 C v t 0 k 0 2 A. Moreover, taking into account (5.19),
p.y 0 Cvt 0 k 0 y/ p.y 0 y/Cp.v/C.t Cp.v//p.k 0 / 1=2Cr C.t Cr/p.k 0 / 1;
and so y 0 C v t 0 k 0 2 .y C V0 / \ A. Using (5.18), if p.v/ > 0 then

1
y 0 C tk0 D y 0 t 0 C p.v/ k 0 D y 0 C v t 0 k 0 C p.v/ k 0 v 2 A;
p.v/
while if p.v/ D 0 then

y 0 C .1 /tk0 D y 0 C v t 0 k 0 C t k 0 . t/1 v 2 A
for WD minf 21 ; "t 1 g. We get the contradiction 0 D 'A .y 0 / t < 0 in the first
case and 0 D 'A .y 0 / t.1 / < 0 in the second case. Hence 'A .y 0 C v/ 2 R
and j'A .y 0 C v/ 'A .y 0 /j p.v/ for every v 2 rV 0 , or equivalently,
'A .y C v/ 2 R; j'A .y C v/ 'A .y/j p.v/ 8v 2 rV 0 : (5.20)
When y WD y 2 M , from (5.20) we get y C rV 0 M , and so M D y C rV 0

as claimed. Moreover, if y; y 0 2 y C 12 rV 0 , then y 2 M and y 0 D y C v for
some v 2 rV 0 ; using again (5.20) we have that j'A .y 0 / 'A .y/j p.y 0 y/. The
conclusion follows. t
u
The next result is similar to Corollary 5.2.11.
Corollary 5.2.21. Let A be a proper closed subset of Y and k 0 2 Y n f0g be such
that A RC k 0 D A. Consider y 2 bd A. If A is epi-Lipschitz at y in the direction
k 0 then 'A .y/ D 0.
Proof. Consider " 2 0; 1Œ and V0 provided by (5.18) with v WD k 0 . Assume that
'A .y/ ¤ 0. Then there exists t > 0 such that tpV0 .k 0 / " and y WD y C tk0 2 A.
Taking
WD t in (5.18) we obtain that yCt.k 0 CV0 / D yCtV 0 A, contradicting
the fact that y 2 bd A. t
u
Corollary 5.2.22. Let A be a proper closed subset of Y and k 0 2 Y n f0g be such
that ARC k 0 D A. Assume that dim Y < C1 and y 2 bd A. Then 'A is finite and
Lipschitz on a neighbourhood of y if and only if k 0 2 int C.A; y/, where C.A; y/
is the Clarke’s tangent cone of A at y (see Definition 4.2.7).
Proof. By [493, Theorem 2], k 0 2 int C.A; y/ if and only if A is epi-Lipschitz at y
in the direction k 0 . The conclusion follows from Corollary 5.2.11, Theorem 5.2.20
and Corollary 5.2.21. t
u
The fact that 'A is Lipschitz on a neighbourhood of y under the condition k 0 2
˝ 0 ˛ 6] in the case Y D R (and C D RC ).
m m
int C.A; y/ is obtained in [55, Proposition

Consider y 2 Y such that k ; y ¤ 0, H WD ker y and take
'0 W H ! R; '0 .z/ WD 'A .z/;
that is, '0 D 'A jH . Since 'A is l.s.c., so is '0 . Then any y 2 Y can be written
uniquely as z tk0 with z 2 H and t 2 R. So, by (5.7), 'A .y/ D 'A .z tk0 / D
'0 .z/ t. Using (5.10) we obtain that A D fz tk0 j .z; t/ 2 epi '0 g. Conversely, if
g W H ! R is a l.s.c. function and A WD fz tk0 j .z; t/ 2 epi gg, then A is a closed
set with A RC k 0 D A and '0 D g. Therefore, the closed set A with the property
A RC k 0 D A is uniquely determined by a l.s.c. function '0 W H ! R. Moreover,
for y D z t k 0 we have that 'A is finite (resp. continuous) at y if and only if '0
is finite (resp. continuous) at z. Moreover, because Y D H C Rk 0 and the sum is
topological (that is, the projection onto H parallel to Rk 0 is continuous), we have
that 'A is finite and Lipschitz continuous on a neighbourhood of y if and only if '0
is finite and Lipschitz continuous on a neighbourhood of z. Similarly, 'A is finite
and Lipschitz continuous if and only if '0 is finite and Lipschitz continuous.
Note that for Y a normed vector space in [217] one says that A is (globally)
epi-Lipschitz in the direction e 2 Y n f0g if there exist a closed linear subspace
H of codimension 1 with e 62 H and a Lipschitz function g W H ! R such that
A D fy C ˛e j y 2 H; ˛ 2 R; g.y/ ˛g; A is epi-Lipschitz if there exists
e 2 Y n f0g such that A is epi-Lipschitz in the direction e. The main result of [217]
asserts that the proper closed set A Y is epi-Lipschitz in the direction e if and
only if e 2 int A1 , and so A Y is epi-Lipschitz if and only if int A1 ¤ ;.
The discussion above shows that not only the main theorem of [217] can be
obtained from Corollary 5.2.15, but this one extends the main theorem of [217] to
locally convex spaces.
These properties of the functional 'A;k 0 are used for deriving minimal point
theorems (cf. [214, 235]), optimality conditions (see [29]) and duality asser-
tions(cf. [214]).
5.2.5 The Formula for the Conjugate and Subdifferential

of 'A for A Convex
The results of this section (except the second part of Corollary 5.2.24) were
established in several papers; we give the proofs for reader’s convenience. The
formula for the conjugate of 'A is derived by [488, Theorem 3] and [512, Theorem
2.2]. Results concerning the subdifferential of 'A are given in [156, Theorem 2.2,
Lemma 2.1]. In the statements below we use some usual notation from convex
analysis. So, having X a separated locally convex space with topological dual X
and a convex function f W X ! R, the conjugate of f is the function f W X ! R
defined by f .x / WD sup fhx; x i f .x/ j x 2 X g and its subdifferential at
x 2 X with f .x/ 2 R is the set @f .x/ WD fx 2 X j hx 0 x; x i
f .x 0 /f .x/ 8x 2 X g; @f .x/ WD ; if f .x/ … R. Recall that the indicator function
of a set A X is the function A W X ! R defined by A .x/ WD 0 for x 2 A and
A .x/ WD C1 for x 2 X n A, while the support of A is the function A WD .A / .
When A is nonempty, the domain of A is a convex cone which is called the barrier
cone of A and is denoted by bar A. Moreover, the normal cone of A at a 2 A is the
set N.A; a/ WD @A .a/. Suppose that the assumptions of Sect. 5.2.2 are fulfilled. As
before, we again use the notations
C C D fy 2 Y j hk; y i
0 8k 2 C g;
C # D fy 2 Y j hk; y i > 0 8k 2 C n f0gg
for the positive dual cone of C and the quasi interior of C C , respectively.
Proposition 5.2.23. Assume that A Y is closed and convex with A C D A and
k 0 … A1 . Then 'A 2 .Y /, that is, 'A is a proper l.s.c. convex function,
˝ ˛
A .y / if y 2 bar A; k 0 ; y D 1;
'A .y / D (5.21)
C1 otherwise,
˝ ˛ ˝ ˛
and @'A .y/ fy 2 bar A j k 0 ; y D 1g fy 2 C C j k 0 ; y D 1g for every
y 2 Y.
Proof. From Theorem 5.2.3 we have that 'A 2 .Y /. Consider y 2 Y . Then
'A .y / D sup fhy; y i 'A .y/ j y 2 Y g

˚
D sup hy; y i t j y 2 Y; t 2 R; y 2 tk0 C A
˚˝ ˛
D sup tk0 C a; y t j y 2 Y; t 2 R; a 2 A
˝ ˛
D supfha; y i j a 2 Ag C supft. k 0 ; y 1/ j t 2 Rg:
Hence (5.21) holds.

Since @f .y/ dom f for every proper function f W Y ! R and every y 2 Y ,

the first estimate for @'A .y/ follows. Moreover, because A D A C we have
A D AC D A C C D A C C C , and so bar A C C .
The estimate bar A C C becomes more precise when C D A1 ; in fact
one has .A1 /C D clw .bar A/. Indeed, from [614, Exercise 2.23] we have that
A1 D .A /1 D dom A D dom A , whence .A1 /C D .A1 / D clw .dom A / , and
so .A1 /C D clw .dom A /.
Using Proposition 5.2.23 one deduces the expression of @'A (see also [156,
Theorem 2.2] for Y a normed vector space).
Corollary 5.2.24. Assume that A is convex and k 0 … A1 . Then for all y 2 Y
one has
˝ ˛
@'A .y/ D fy 2 bar A j k 0 ; y D 1; hy; y i 'A .y/
hy; y i 8y 2 Ag:
(5.22)
Moreover, if A .C n f0g/ int A holds then @'A .y/ C # for every y 2 Y .
Proof. Fix y 2 Y . If y … dom 'A then both sets in (5.22) are empty. Let y 2
dom 'A . Then, of course, y 'A .y/k 0 2 A. If y 2 @'A .y/ then 'A .y/C'A .y / D
hy; y i. Taking into account (5.21) we obtain that
˝ ˛
y 2 bar A; k 0 ; y D 1 and hy; y i 'A .y/
hy; y i 8y 2 A; (5.23)
that is, the inclusion holds in (5.22).

˝ 0 Conversely,
˛ if y 2 Y˝ is such that (5.23)
˛
holds, since y 'A .y/k 2 A and k ; y D 1, we obtain that y 'A .y/k 0 ; y D
0
A .y /, which shows that 'A .y/ C 'A .y / D hy; y i. Hence y 2 @'A .y/.
Therefore, (5.22) holds.
Assume now A .C n f0g/ int A, and take y 2 @'A .y/. Hence y 2 dom 'A .
Consider k 2 C nf0g. Since .y k/y D k 2 .C nf0g/, by Theorem 5.2.12 (iv),
we have that hk; y i 'A .y k/ 'A .y/ < 0, that is, hk; y i > 0. Therefore,
y 2 C #.
5.3 Scalarizing Functionals by Hiriart-Urruty and Zaffaroni
In this section, we characterize Q-minimal points of a subset A of a Banach

space Y with respect to a proper pointed closed convex cone C Y (elements
of QMin.A; C /, see Definition 2.4.8) using a scalarization by means of Hiriart-
Urruty’s function and establish properties of subdifferentials of this function in some
special cases.
In Sect. 12 we derive necessary and sufficient conditions for Q-minimal
points/Q-minimizers (see Definitions 2.4.8, 2.6.3) in the form of being a solution
for a scalar optimization problem. Here the scalar function is the oriented distance
function introduced by Hiriart-Urruty [262] (see Example 5.1.3). In this section
5.3 Scalarizing Functionals by Hiriart-Urruty and Zaffaroni 233
we suppose that Y is a Banach space and C Y a proper closed convex and

pointed cone. Recall that for ; ¤ A Y and Y n A ¤ ;, this function is defined by
A .y/ WD d.y; A/ d.y; Y n A/:
Remark 5.3.1. Of course, if int A D ;, it holds d.y; Y n A/ D 0 and so for all

y2Y
A .y/ D d.y; A/;
i.e., Hiriart-Urruty’s function at y 2 Y coincides with the distance from y to A.

Furthermore, if A.¤ Y / is convex and int A ¤ ; we have A .y/ D cl A .y/ for
all y 2 Y . Indeed, on the one hand side we have obviously d.y; A/ D d.y; cl A/
and d.y; Y n A/ D d.y; cl.Y n A// D d.y; Y n int A/. On the other hand,
d.y; Y n cl A/ D d.y; cl.Y n cl A// D d.y; Y n int.cl A// D d.y; Y n int A/;
taking into account int.cl A/ D int A since A is convex, and so d.y; Y n A/ D

d.y; Y n cl A/.
It is well known that this function A has very good general properties (see
Zaffaroni [607]) and we list below for the reader’s convenience the properties we
shall use in the sequel.
Proposition 5.3.2. Assume that A is a nonempty subset of Y and Y n A ¤ ;.
(i) A is Lipschitz of rank 1.
(ii) Y nA D A .
(iii) A is convex if A is convex and A is concave if A is reverse convex, i.e.,
Y n A is convex.
(iv) A .y/ < 0 iff y 2 int A, A .y/ D 0 iff y 2 bd A and A .y/ > 0 iff
y 2 Y n int A.
(v) Suppose that A is convex and has a nonempty interior, and x 2 bd A. Then
@A .x/ N.xI A/ n f0g
where N.xI A/ is the normal cone in the sense of convex analysis of A at x.

In the following proposition we present a result by Ha [228, Proposition 21.10]
that characterizes Q-minimal points introduced in Definition 2.4.8 via scalarization
by means of the functional A .
Proposition 5.3.3. Suppose that A Y is a nonempty set and Q WD int D for a
proper cone D Y with nonempty interior. Then, a 2 QMin.A; C / if and only if
the function Q . a/ attains its minimum at a, that is
Q .a a/
Q .0/ D 0; 8a 2 A: (5.24)
Proof. Observe first that 0 2bd.Q/ because of Definition 2.4.8 and Q .0/ D 0
taking into account Proposition 5.3.2 (iv). Now, let a 2 QMin.A; C /. By Defi-
nition 2.4.8 we have .A a/ Y n .Q/. Because of Proposition 5.3.2 (iv),
Q .a a/
0 for all a 2 A, i.e., (5.24) holds. Next, suppose that (5.24) holds.
If a … QMin.A; C / then there is a0 2 A such that a0 a 2 Q. As the set Q
is open it holds Q .a0 a/ < 0 since Proposition 5.3.2 (iv), a contradiction to
(5.24). Thus, a 2 QMin.A; C /. t
u
For deriving optimality conditions it is of interest to know more about the
subdifferential of Q . a/ at a 2 QMin.A; C /. The results presented in
the following proposition are shown by Ha [228, Proposition 21.11] and play an
important role in formulating Fermat rules and Lagrange multiplier rules.
Proposition 5.3.4. (i) @ int C .0/ C C n f0g.
(ii) Let D be a proper cone with int D ¤ ; and Q D int D a convex dilation of C .
Then
@Q .0/ C # :
(iii) Assume that C has a bounded base , then it holds @V .0/ int C C for
V given by (2.41).
(iv) Suppose that A has a nonempty interior, then it holds
@A Y ncl conv coneŒ.Aa/[C .0/ C C n f0g;
where @A denotes the approximate subdifferential introduced in Defini-

tion 3.6.1.
Proof. (i) We get the result applying Proposition 5.3.2 (v) to A D int C and
take account of N.0I int C / D C C .
(ii) Applying Proposition 5.3.2 (v) to A D C we get @Q .0/ N.0I Q/nf0g.
Now take y 2 @Q .0/ and k 2 C n f0g Q. We have to show that
y .k/ > 0 for all k 2 C n f0g. As y ¤ 0 because of Proposition 5.3.2 (v),
there is y 2 Y such that y .y/ > 0. On the other hand, since Q is open
and k 2 Q, there exists a scalar t > 0 such that k C ty 2 Q. Hence,
y .k C ty/ 0 and y .k/
ty .y/ > 0. Thus, y 2 C # .
(iii) Suppose that is bounded and denote
ı WD supfkk j 2 g < C1:
Let y 2 @V .0/. It is known that y 2 int C C if and only if y is uniformly

positive on C in the sense that there exists a scalar ˛ > 0 such that y .k/
˛kkk for all k 2 C n f0g. Let k 2 C n f0g be an arbitrary vector. As is

a base of C , there exist a scalar t > 0 and 2 such that k D t. Since
C BV Y V , it follows that
5.3 Scalarizing Functionals by Hiriart-Urruty and Zaffaroni 235
k C tBVY D t. C BV Y / V ;
i.e., the open ball centered at k with the radius t is contained in V .
Therefore, we get
d.k; Y n .V //
t:
On the other hand, t D .kkk=kk/

.=ı/kkk. Hence,
d.k; Y n .V //
.=ı/kkk:
Of course, V .0/ D 0. Because of y 2 @V .0/, the definition of the

convex subdifferential yields
y .k/ V .k/ V .0/ D d.k; V / d.k; Y n .V //

D d.k; Y n .V // .=ı/kkk
such that we have hy ; ki

.=ı/kkk. This means that y is uniformly
positive on C , or y 2 int C C as it was to be shown.
(iv) By the relationship between approximate subdifferential and the Clarke subd-
ifferential (see Proposition 3.6.3 (i)), it is sufficient to show that
@C Y ncl conv coneŒ.Aa/[C .0/ C C n f0g:
For the simplicity, we denote Q WD Y n V , where V WD cl conv cone

Œ.A a/ [ C . Note that V is a closed convex cone with a nonempty interior
and C V . We have to show that
@C Q .0/ C C n f0g:
By Proposition 5.3.2 (ii), we have Q .0/ D Y nV .0/ D V .0/. The
properties of the Clarke subdifferential and the subdifferential of convex
analysis yield
@C .V /.0/ D @C V .0/ D @V .0/:
Applying Proposition 5.3.2 (v) to the closed convex cone V which has a
nonempty interior gives @V .0/ N.0I V / n f0g. Furthermore, since
C V we get N.0I V / C C . Therefore, we get @V .0/ C C n f0g,
which yields @C Q .0/ C C n f0g. t
u
Remark 5.3.5. Note that the properties described above for the oriented distance
function are similar to those required for risk measures used in mathematical finance
[256, 488, 497] (see Example 5.1.4).
5.4 Characterization of Solutions of Set-Valued Optimization

Problems by Means of Nonlinear Scalarizing Functionals
5.4.1 An Extension of the Functional 'A
In this section we assume that Y is a linear topological space ordered by a proper

convex closed pointed cone C Y with nonempty interior, A Y , k 0 2 int C .
We consider the functional 'A;k 0 given by (5.5). Now, we take d 2 Y and A D
d C C . Then, the functional 'A;k 0 has the form
'd;C;k 0 .y/ D infft 2 R j y 2 tk0 C d C C g; y 2 Y: (5.25)
As usual, the family of all nonempty subsets of Y is denoted by P.Y /.

Replacing d in (5.25) by a set D 2 P.Y /, we get a functional 'D;C;k 0 W Y !
R [ f1g defined by
'D;C;k 0 .y/ WD infft 2 R j y 2 tk0 C D C C g; y 2 Y: (5.26)
Setting A D D C C the functional 'D;C;k 0 coincides with the functional 'A;k 0 given
by (5.5). Obviously, we have for any y 2 Y
'D;C;k 0 .y/ D inf f'd;C;k 0 .y/g: (5.27)

d 2D
Definition 5.4.1. A set D 2 P.Y / is called C -proper, if D C C ¤ Y .

We denote the family of all C -proper subsets of Y by PC .Y /.
In this section we discuss an extension of the functional 'D;C;k 0 W Y ! R[f1g
defined by (5.26) to a function ˚k 0 W PC .Y /2 ! R [ fC1g (see Definition 5.4.5)
that is very useful in set-valued optimization where the solution concept is based on
set approach (compare Sect. 2.6.2).
The results in this section are derived by Hernández and Rodríguez-Marín [252].
Based on a corresponding scalarization technique by means of the functional ˚k 0
Hernández and Rodríguez-Marín have shown a characterization of solutions of set-
valued optimization problems where the solution concept is given by the lower set
less order relation lC (compare Definition 2.6.9).
It is easy to check that the set-relation l defined below in (5.28) is an
equivalence relation on P.Y /, such that we can consider an order on the family
of equivalence classes which are denoted by Œl :
A l B W” A lC B and B lC A: (5.28)
Obviously, A 2 ŒBl if and only if A C C D B C C (compare Definition 2.6.9 and

Remark 2.6.11).
5.4 Characterization of Solutions of Set-Valued Optimization Problems by. . . 237
We say that A 2 P.Y / is C -closed if A C C is a closed set and that A is C -

bounded if for each neighborhood U of zero in Y there is some positive number t
such that A tU C C . Furthermore, we say that A is C -compact if any cover of A
of the form fU˛ C C j U˛ are openg admits a finite subcover. Every C -compact set
is C -closed and C -bounded (see Luc [402, Proposition 3.3]).
The following characterization of C -proper sets is given in [252, Lemma 2.16].
Lemma 5.4.2. Let D 2 P.Y /. D is C -proper if and only if 'D;C;k 0 .y/ > 1 for
every y 2 Y .
From Lemma 5.4.2 we get that 'D;C;k 0 ./ is a real-valued function, if D 2
PC .Y /.
From the properties (5.6), (5.12), (5.13) of the functional 'A;k 0 given by (5.5) in
Theorems 5.2.3, 5.2.6 we get the corresponding assertions concerning 'D;C;k 0 in the
following lemma:
Lemma 5.4.3. Let D 2 PC .Y / and t 2 R. Then, for any y 2 Y it holds:
(i) 'D;C;k 0 .y/ < t ” y 2 tk0 C D C int C ,
(ii) 'D;C;k 0 .y/ t ” y 2 tk0 C cl.D C C /,
(iii) 'D;C;k 0 .y/
t ” y … tk0 C D C int C ,
(iv) 'D;C;k 0 .y/ D t ” y 2 tk0 C bd.D C C /,
(v) 'D;C;k 0 .y/ > t ” y … tk0 C cl.D C C /.
From the definition of the functional in (5.26) we get the following assertions:
Proposition 5.4.4. Let D; B 2 PC .Y / and y 2 Y . Then
(i) If D is C -closed, it holds 'D;C;k 0 .y/ D mind 2D f'd;C;k 0 .y/g.
(ii) If D lC B, then 'D;C;k 0 .y/ 'B;C;k 0 .y/. In particular, if D 2 ŒBl then
'D;C;k 0 .y/ D 'B;C;k 0 .y/.
Now, we study an extension of the functional (5.26) defined from Y to R [ f1g
to a function defined from PC .Y /2 to R [ fC1g.
Definition 5.4.5. Consider a fixed element k 0 2 int C . Let the function ˚k 0 W
PC .Y /2 ! R [ fC1g defined by
˚k 0 .B; D/ WD sup f'B;C;k 0 .d /g for .B; D/ 2 PC .Y /2 : (5.29)

d 2D
From Lemma 5.4.3, (ii), we get
˚k 0 .B; D/ t ” D tk0 C B C C (5.30)
if B is a C -closed set and t 2 R.

Using this assertion we get an equivalent formulation of (5.29) in the following
proposition:
Proposition 5.4.6. Let B 2 PC .Y / be a C -closed set and D 2 PC .Y /. Assume

˚k 0 .B; D/ < C1. Then we get the following expression for ˚k 0 :
˚k 0 .B; D/ D minft 2 R j D tk0 C B C C g:
Proof. Consider the set
T WD ft 2 R j ˚k 0 .B; D/ tg:
Under the assumption ˚k 0 .B; D/ < C1 and taking into account (5.30) we get
min t D ˚k 0 .B; D/
t 2T
and the proof is completed. t

u
Lemma 5.4.7. Consider B 2 PC .Y /. Then B is C -bounded if and only if
˚k 0 .C; B/ < C1.
Proof. [“)”:] Assume that B is C -bounded, then for the neighborhood of zero
U WD k 0 C int C
there exists ˛ > 0 such that B ˛.k 0 C int C / C C ˛k 0 C C . Thus, for each
b 2 B it holds 'C;C;k 0 .b/ ˛. Hence, ˚k 0 .C; B/ ˛.
[“(”:] Conversely, suppose that ˚k 0 .C; B/ < C1, then there exists r 2 R such
that ˚k 0 .C; B/ < r. Namely, for each b 2 B we have
'C;C;k 0 .b/ < r:
Taking into account Lemma 5.4.3 (i), the above inequality can be written as b 2
rk0 C int C , that is
B rk0 C int C: (5.31)
Let U 0 be a neighborhood of zero. Chose

> 0 such that rk0 2
U 0 , then we get
together with (5.31) that
B
U 0 C C;
which implies that B is C -bounded. t

u
Theorem 5.4.8. Let B be a C -bounded set and D 2 PC .Y /. Then D is C -
bounded if and only if ˚k 0 .B; D/ < C1.
Proof. [“)”:] Suppose that D is C -bounded. Then, because of (5.27), for every
d 2 D we have
'B;C;k 0 .d / D inf f'b;C;k 0 .d /g: (5.32)

b2B
It is easy to prove that 'b;C;k 0 .d / D '0;C;k 0 .d b/ and '0;C;k 0 .d b/ D 'C;C;k 0 .d b/.
Hence the equality (5.32) can be written as
'B;C;k 0 .d / D inf f'C;C;k 0 .d b/g:

b2B
Then, for any fixed b0 2 B we obtain
'B;C;k 0 .d / 'C;C;k 0 .d b0 / for all d 2 D:
This yields
˚k 0 .B; D/ sup f'C;C;k 0 .d b0 /g D ˚k 0 .C; D b0 /:

d 2D
Because of the C -boundedness of D it follows ˚k 0 .C; D b0 / < C1 taking

into account Lemma 5.4.7. Therefore, ˚k 0 .B; D/ < C1.
[“(”:] Conversely, we suppose ˚k 0 .B; D/ D r with r 2 R, then for each
d 2 D we get 'B;C;k 0 .d / r. Let ˛ > 0, then 'B;C;k 0 .d / < r C ˛ and applying
Lemma 5.4.3 (i) it follows d 2 .r C ˛/k 0 C B C C for all d 2 D, thus
D .r C ˛/k 0 C B C C: (5.33)
Let U be a neighbourhood of zero. Since B is C -bounded we get that the set

.r C ˛/k 0 C B is also C -bounded. Then, there exists
> 0 such that
.r C ˛/k 0 C B
U C C:
Taking into account (5.33) this inclusion yields
D
U C C:
This means that D is C -bounded. t

u
In Sect. 5.4.2 we will give characterizations of solutions of set-valued optimiza-
tion problems where the solution concept is given by the lower set less order relation
lC by means of the functional ˚k 0 . In order to prove such assertions we need certain
monotonicity properties of the functional ˚k 0 .
Definition 5.4.9. Let S P.Y /. A given functional ' W P.Y / ! R is said to
be lC -increasing on S if
A; B 2 S; A lC B H) '.A/ '.B/:

The following result follows directly from Proposition 5.4.4, (ii).

Theorem 5.4.10. Let D 2 PC .Y /, then ˚k 0 .; D/ is lC -increasing on PC .Y /.
Furthermore, the following assertion holds (see Hernández and Rodríguez-Marín
[252, Theorem 3.10]).
Theorem 5.4.11. Let B 2 PC .Y / be a C -closed set. Then the following state-
ments are true:
(i) ˚k 0 .B; B/ D 0,
(ii) If B 2 ŒDl , then ˚k 0 .B; D/ D ˚k 0 .D; B/ D 0,
(iii) B lC D if and only if ˚k 0 .B; D/ 0.
5.4.2 Characterization of Solutions of Set-Valued Optimization

Problems with Lower Set Less Order Relation lC
by Scalarization
In this section we assume that Y is a linear topological space partially ordered by

a proper convex closed pointed cone C Y with nonempty interior. Furthermore,
let S be a set and F W S Y is a set-valued map with domain S . We consider a
set-valued optimization problem where the solution concept is given with respect to
the lower set less order relation lC (see Definitions 2.6.9 and 2.6.19):
lC minimize F .x/ subject to x 2 S: (SP lC )
Let M .PC .Y /; R/ be the set of all functions from PC .Y / to R.

The next results are shown by Hernández and Rodríguez-Marín [252, Theo-
rem 4.1 and Corollary 4.3]. First, we present certain kinds of separation theorems.
Theorem 5.4.12. Consider the problem (SP lC ). Assume that F is C -closed and
C -bounded valued on S and x 0 2 S . Then, x 0 is a minimal solution of (SP lC )
if and only if there exists a functional ' 2 M .PC .Y /; R/ which is lC -increasing
on PC .Y / such that the following statements are true:
(i) If x 2 S and F .x/ 2 ŒF .x 0 /l then '.F .x// D 0.
(ii) If x 2 S and F .x/ … ŒF .x 0 /l then '.F .x// > 0.
(iii) If D 2 PC .Y / and D lC F .x 0 / then '.D/ 0.
Proof. Assume that x 0 is a minimal solution of (SP lC ). We fix k 0 2 int C .
Taking into account Theorem 5.4.10, the functional
'./ WD ˚k 0 .; F .x 0 // 2 M .PC .Y /; R/

is lC -increasing on PC .Y /. So it is sufficient to show that ' satisfies the conditions

.i /, .ii/ and .iii/. The conditions .i / and .iii/ are consequences of Theorem 5.4.11
.ii/ and .iii/, respectively.
Now, we will show that .ii/ holds. x 0 is a minimal solution of (SP lC ).
Then, for each x 2 S with F .x/ … ŒF .x 0 /l we have F .x/ 6lC F .x 0 /. Hence,
by Theorem 5.4.11 .iii/,
'.F .x// D ˚k 0 .F .x/; F .x 0 // > 0:
Conversely, we assume that .i /, .ii/ and .iii/ are true for some ' 2
M .PC .Y /; R/ which is lC -increasing on PC .Y /. We have to prove that x 0
is a minimal solution of (SP lC ). Indeed, if we suppose that x 0 is not a minimal
solution of (SP lC ), there exists x 0 2 S such that F .x 0 / … ŒF .x 0 /l and
F .x 0 / lC F .x 0 /:
Thus, by .ii/, '.F .x 0 // > 0. On the other hand, by .iii/, it follows '.F .x 0 // 0
which is a contradiction. u
t
For the special case of a vector optimization problem we get from Theo-
rem 5.4.12 well known characterizations of solutions by means of a sustainable
scalarization method.
Consider the vector optimization problem
Min.f .S /; C /; (VOP)
where f W S ! Y is a single-valued map. We use the solution concept introduced

In the assertion of the next corollary we use (strictly) increasing functionals ' W
Y ! R in the following sense (compare .Asf 1/ and .Asf 10 /): ' is called increasing,
if y; w 2 Y , w y 2 C implies '.y/ '.w/. Furthermore, ' is called strictly
increasing, if y; w 2 Y , w y 2 int C implies '.y/ < '.w/.
Corollary 5.4.13. Consider the problem (VOP). Then, f .x 0 / 2 Min.f .S /; C / if
and only if there exists a continuous functional ' W Y ! R that is increasing and
strictly increasing, such that:
(i) '.f .x 0 // D 0.
(ii) If x 2 S and f .x/ ¤ f .x 0 /, then '.f .x// > 0.
(iii) If y 2 f .x 0 / C , then '.y/ 0.
Now, we derive sufficient conditions for minimal solutions of (SP lC ) using
the functional ˚k 0 .B; / (see (5.29)). In order to show these conditions we introduce
the following notations.
Definition 5.4.14. A net fD˛ j ˛ 2 I g from P.Y / is said to be lC -decreasing

(with respect to lC ) and we write fD˛ g #l if D˛ lC Dˇ for each ˛; ˇ 2 I and
ˇ < ˛.
For a given x 0 2 S we introduce
L.x 0 / WD fx 2 S j F .x/lC F .x 0 /g:
Definition 5.4.15. Let x 0 2 X and k 0 2 int C . F is said to be ˚k 0 -stable at x 0 if

for any fx˛ g L.x 0 / and t 2 R verifying fF .x˛ /g #l and ˚k 0 .F .x 0 /; F .x˛ // D t
for all ˛ imply that there exists x 2 S such that F .x/lC F .x 0 / C tk0 .
The following existence result for minimal solution of (SP lC ) in the sense of
Definition 2.6.19 is shown by Hernández and Rodríguez-Marín [252, Theorem 5.3].
Theorem 5.4.16. Consider the problem (SP lC ). Let x 0 2 S and k 0 2 int C .
Assume that F is C -closed valued, F .x 0 / is C -bounded, F is ˚k 0 -stable at x 0
and [x2L.x0 / F .x/ is C -compact. Then there exists x 2 L.x 0 / and x is a minimal
solution of (SP lC ).
Furthermore, Hernández and Rodríguez-Marín [252, Theorem 5.4] proved the
following existence result using Cantor’s intersection theorem.
Theorem 5.4.17. Consider the problem (SP lC ). Assume .X; d / is a complete
metric space , S X , F is C -bounded valued and for each x 2 S the set L.x/ is
closed. Assume that x 0 2 S and that the following conditions are satisfied:
(i) [x2L.x 0 / F .x/ is C -bounded,
(ii) there are an element k 0 2 int C and a C -bounded set B Y such that
x 1 ; x 2 2 S; F .x 1 /lC F .x 2 / H) d.x 1 ; x 2 /˚k 0 .B; F .x 1 // ˚k 0 .B; F .x 2 //:
Then there exists x 2 L.x 0 / with x a minimal solution of (SP lC ).

Proof. Suppose that x 0 is not a minimal solution of (SP lC ). Taking into account
Definition 2.6.19 there exists x 2 L.x 0 / with
F .x/lC F .x 0 / and F .x 0 /6lC F .x/:
Because of assumption (ii) it follows
0 < ˚k 0 .B; F .x// ˚k 0 .B; F .x 0 //:

S
On the other hand, since x2L.x 0 / F .x/ and F .x 0 / are C -bounded sets, by
Theorem 5.4.8
0 < sup f˚k 0 .B; F .x//g ˚k 0 .B; F .x 0 // < C1: (5.34)

x2L.x0 /
For z 2 S we introduce v.z/ WD supx2L.z/ f˚k 0 .B; F .x//g. From (5.34) we get
that there exists an element x1 2 L.x 0 / with
v.x 0 / ˚k 0 .B; F .x1 // 21 :
If x1 2 L.x 0 / and x1 is not a minimal solution of (SP lC ), then there exists
x2 2 L.x1 / such that
v.x1 / ˚k 0 .B; F .x2 // 22 :
In this way we obtain a sequence fxn gn2N from L.x 0 / such that for each n 2 N it
holds L.xnC1 / L.xn /.
Moreover, since for each n 2 N the set L.xn / is closed and
v.xn / ˚k 0 .B; F .xnC1 // 2.nC1/; (5.35)
it follows fdiam L.xn /g ! 0. Indeed, if that is not the case, then there exists ı > 0
with diam L.xn / > ı for each n 2 N. By triangle inequality of d.; /, for fixed
n 2 N, there exists z0 2 L.xn / such that d.z0 ; xn / > ı2 and taking into account
assumption (ii), we get
ı
< d.z0 ; xn / ˚k 0 .B; F .z0 // ˚k 0 .B; F .xn //: (5.36)
2
On the other hand, we have
˚k 0 .B; F .z0 // ˚k 0 .B; F .xn // v.xn / ˚k 0 .B; F .xn //
and as xn 2 L.xn1 /, then v.xn / v.xn1 /. Therefore, because of (5.35) and (5.36)
we get
ı
< d.z0 ; xn / v.xn1 / ˚k 0 .B; F .xn // 2n
2
and for n ! C1 the above inequality implies ı < 0, a contradiction.
So, because for every n 2 N the set L.xn / is closed, .X; d / is a complete metric
space, L.xnC1 / L.xn / and diam L.xn / ! 0 we get from Cantor’s intersection
theorem
\
L.xn / D fxg:
n2N
Consequently, x 2 L.x 0 / and L.x/ D fxg. Hence, x is a minimal solution of

(SP lC ) and the proof is completed. u
t
The following corollary is a direct consequence of Theorem 5.4.17.
Corollary 5.4.18. Consider the vector optimization problem
Min.f .S /; C /; (VOP)
where f W S ! Y , C is a proper pointed closed convex cone in the linear

topological space Y . Assume .X; d / is a complete metric space , S X and for
each x 2 S the set L.x/ is closed. Suppose that x 0 2 S and the following conditions
are satisfied:
(i) f .L.x 0 // is C -bounded,
(ii) there are an element k 0 2 int C and a C -bounded set B Y such that
x 1 ; x 2 2 S; f .x 1 /2f .x 2 /C H)d.x 1 ; x 2 /'B;C;k 0 .f .x 1 //'B;C;k 0 .f .x 2 //:
Then there exists x 2 L.x 0 / with f .x/ 2 Min.f .S /; C /.

Remark 5.4.19. Kuroiwa [347, Theorems 4.1 and 4.2] obtained existence results
for minimal solutions of (SP lC ) under similar conditions to those supposed
in Theorems 5.4.16 and 5.4.17. However, the assumptions in Theorems 5.4.16
and 5.4.17 are less restrictive because there are no continuity assumptions.
5.5 The Extremal Principle
In the book by Mordukhovich [430] a general concept of set extremality (see

Mordukhovich [425, 427], Kruger, Mordukhovich [342], Kruger [343]) is pre-
sented. Its relationships to separation properties of nonconvex sets are discussed
and corresponding applications for deriving optimality conditions in constrained
optimization are developed in [425, 430]. In this section we will present these
fundamental results, especially we will formulate several versions of the extremal
principle. These extremal principles will be applied in Sect. 12.9 for deriving
a subdifferential variational principle for set-valued mappings (Theorem 12.9.1)
and in Sect. 12.11 in order to prove a first order necessary condition for fully
localized minimizers of set-valued optimization problems in the sense of Defini-
tion 2.6.49.
First, we introduce local extremal points and extremal systems (see [430,
Definition 2.1]).
Definition 5.5.1. Consider nonempty subsets ˝1 ; ; ˝n (n
2) of a Banach
space X . Let x be a common point of these sets. This point x is called a local
extremal point of the set system f˝1 ; ; ˝n g if there are sequences faik g X ,
k!1
i D 1; ; n, and a neighborhood V of x such that aik ! 0 and
5.5 The Extremal Principle 245
\
n
.˝i aik / \ V D ; for all large k 2 N:
i D1
In this case f˝1 ; ; ˝n ; xg is said to be an extremal system in X .

Remark 5.5.2. For the case of two closed sets ˝1 and ˝2 in the Banach space X we
say that a point xN 2 ˝1 \ ˝2 is locally extremal for the set system f˝1 ; ˝2 g if there
is a neighborhood V of xN such that for any " > 0 we can find an element a 2 "BX
with
˝1 \ .˝2 C a/ \ V D ;: (5.37)
It is important to mention that the condition ˝1 \ ˝2 D fxg does not necessary

imply that x is a local extremal point of f˝1 ; ˝2 g.
Mordukhovich [430, Definition 2.5] introduced the following three basic ver-
sions of the extremal principle in Banach spaces and explained that they can be
considered as a kind of local separation of nonconvex sets around extremal points.
In the following definition the set of "-normals NO " .xI ˝/ (see Definition 3.5.2), the
Fréchet normal cone NO .xI ˝/ (see (3.23)) and the limiting normal cone N.xI ˝/
(see Definition 3.5.6) (with X Banach space, ˝ X , x 2 ˝) are used.
Definition 5.5.3. Let f˝1 ; ; ˝n ; xg be an extremal system in X . Then
(a) f˝1 ; ; ˝n ; xg satisfies the "-extremal principle if for every " > 0 there are
xj 2 ˝j \ .x C "BX / and xj 2 X such that
xj 2 NO " .xj I ˝j / j D 1; ; n; (5.38)

x1 C C xn D 0; jjx1 jj C C jjxn jj D 1: (5.39)
(b) f˝1 ; ; ˝n ; xg satisfies the approximate extremal principle if for every

" > 0 there are xj 2 ˝j \ .x C "BX / and xj 2 X and
xj 2 NO .xj I ˝j / C "BX j D 1; ; n; (5.40)
such that (5.39 ) holds.

(c) f˝1 ; ; ˝n ; xg satisfies the exact extremal principle if there are basic
normals
xj 2 N.xI ˝j /; j D 1; ; n; (5.41)
such that (5.39 ) holds.

The "- (approximate, exact, respectively) extremal principle holds in the space
X if it holds for every extremal system f˝1 ; ; ˝n ; xg in X , where all the sets ˝j
(j D 1; ; n) are closed around x.
Remark 5.5.4. Taking into account NO .xI ˝/ C "BX NO " .xI ˝/ it is clear that
the "-extremal principle follows from the approximate extremal principle for any
extremal system f˝1 ; ; ˝n ; xg in a Banach space X .
Remark 5.5.5. The assertions in the extremal principles provide necessary condi-
tions for local extremal points of set systems and can be considered as generalized
Euler equations in an abstract geometric setting. Furthermore, the extremal princi-
ples are related to local separation of nonconvex sets. Indeed, if we consider the
exact extremal principle for two sets ˝1 and ˝2 , then (5.39) and (5.41) reduce to
the fact that there exists an element x 2 X with
0 ¤ x 2 N.xI ˝1 / \ .N.xI ˝2 //: (5.42)
Such a kind of extremal principle we also get using the separation theorem for
nonconvex sets (Theorem 5.2.9) where the nonlinear separating functional is given
by (5.5). Applying Theorem 5.2.9 to an extremal system f˝1 ; ˝2 ; xg and taking into
account the properties of the subgradients of the nonlinear separating functional (see
Lemma 12.8.1 (ii) and (iii)) we get a condition corresponding to (5.42) following
the line of the proof of Theorem 12.8.2.
Remark 5.5.6. If we assume that ˝1 and ˝2 are convex, (5.42) is equivalent to
8 x1 2 ˝1 ; 8x2 2 ˝2 W hx ; x1 i hx ; x2 i;
which is nothing else then the classical separation property for two convex sets.
Mordukhovich [430, Theorem 2.20] has shown the following fundamental
assertion in an Asplund space X .
Theorem 5.5.7. Suppose that X is an Asplund space. Then the approximate
extremal principle holds in X .
Remark 5.5.8. Taking into account Remark 5.5.4 we get from Theorem 5.5.7 that
the "-extremal principle holds in an Asplund space X (see [430, Theorem 2.20]).
We will use Theorem 5.5.7 for the case of two subsets ˝1 and ˝2 of an Asplund
space X in Sects. 12.9 and 12.11. In the case of a system f˝1 ; ˝2 g, a locally
extremal point xN 2 ˝1 \ ˝2 is given by (5.37). Then we get the following assertion
from Theorem 5.5.7 (see [27]).
Theorem 5.5.9 (The Extremal Principle). Let X be an Asplund space and xN be
a local extremal point of the set system f˝1 ; ˝2 g, where both ˝1 and ˝2 are closed
around x.N Then for every " > 0 there are
xj 2 ˝j \ .xN C "BX / and xj 2 NO .xj I ˝j / j D 1; 2;

5.5 The Extremal Principle 247
satisfying the relations
1 " jjx1 jj C jjx2 jj 1 C " and jjx1 C x2 jj ":
As explained in Remark 5.5.5, the extremal principle can be considered as a

nonconvex variational counterpart of the classical separation principle for convex
sets, similar like the separation theorems for nonconvex sets (see Theorem 5.2.9
in Sect. 5.2). It plays in fact a fundamental role in variational analysis similar to
that played by the convex separation and Bishop-Phelps theorems under convexity
assumptions; see the books by Mordukhovich [430] and [431] for more details and
numerous applications.
In order to derive necessary optimality conditions for minimizers of set-valued
optimization problems with general geometric constraints by using suitable rules
of subdifferential/coderivative calculus we need certain additional assumptions,
especially we assume normal compactness properties of sets and mappings, which
are automatically fulfilled in finite dimensional spaces while are indispensably
needed in infinite-dimensional spaces due to the natural lack of compactness therein.
Among the advantages of the underlying sequential normal compactness (SNC)
properties recalled below one gets important SNC calculus results ensuring the
preservation of these properties under various operations and describing broad
classes of sets and mappings in infinite dimensions for which these properties are
fulfilled; see Mordukhovich [430] for more details, discussions, and applications.
These properties are used in the framework of Asplund spaces, and so the given
definitions are specified to this setting; compare [430] for appropriate modifications
in general Banach spaces.
Recall that a set S X is sequentially normally compact (SNC) at x 2 S
S w
if for any sequences xk ! x and xk ! 0 with xk 2 NO .xk I S /, k 2 N, we have
jjxk jj ! 0 as k ! C1. A set-valued mapping F W X Y is SNC at
.x; y/ 2 graph F if its graph is SNC at this point.
For deriving the necessary optimality conditions in Theorem 12.11.3 and corre-
sponding applications in welfare economics (see Sect. 15.3) the following extension
of Theorem 5.5.9 is important (see [430, Lemma 5.58]). The result is shown under
certain partially sequentially normally Q compactness conditions. Recall that a set
˝ X in the product space X D niD1 Xi is partially sequentially normally
compact (PSNC) at x 2 cl ˝ with respect to fXi j i 2 I g as I f1; ; ng (or
with respect to the indices I ) if for any sequences of elements .xk ; xk / 2 X X
with xk D .x1k ; ; xnk / and xk D .x1k

; ; xnk / satisfying
˝
xk ! x and xk 2 NO .xk I ˝/ for all k 2 N (5.43)
the following implication holds

w
xik ! 0; i 2 I; and jjxik jj ! 0; i 2 f1; ; ng n I ) jjxik jj ! 0; i 2 I:
The strong PSNC property of ˝ at x with respect to fXi j i 2 I g with

I f1; ; ng says that for any sequences f.xk ; xk /g satisfying (5.43) we have
the implication

w
xik ! 0; i 2 f1; ; ng ) jjxik jj ! 0; i 2 I:
Both PSNC and strong PSNC properties do not depend on the product structure
and reduce to the sequentially normally compact (SNC) property of ˝ at x in
the extreme case of I D f1; ; ng. Sufficient conditions ensuring the fulfillment
of these properties and their preservation under various operations are given in
Mordukhovich [430]. Especially, a set ˝ is SNC at x 2 ˝ if it is compactly epi-
Lipschitzian (CEL) around this point. Both properties are fulfilled for a broad class
of Banach spaces; see, e.g., [66, 278, 430] for more details and references.
The following formulation of the (extended) extremal principle is given in Bao
and Mordukhovich [28, Section 2] (compare [431, Lemma 5.58]) and will be
applied in Sect. 12.11 in order to show first order necessary conditions for localized
minimizers in Theorem 12.11.3 as well as in Sect. 15.3 for deriving an extended
welfare theorem (see Theorem 15.3.5).
Theorem 5.5.10 (Extended Extremal Principle).
Q set system f˝1 ; ˝2 g, where both ˝1 and ˝2
Let x be a local extremal point of the
are closed around x in the product niD1 Xi of Asplund spaces Xi , i D 1; ; n.
Take two index sets I; J f1; ; ng with I [ J D f1; ; ng and I \ J D ;.
Furthermore, suppose that one of the following PSNC conditions is satisfied for ˝1
and ˝2 :
• ˝1 is PSNC at x with respect to I and ˝2 is strongly PSNC at x with respect
to J ;
• ˝1 is strongly PSNC at x with respect to I and ˝2 is PSNC at x with respect
to J .
Then there exists an element x 2 X such that
0 ¤ x 2 N.xI ˝1 / \ .N.xI ˝2 //:
Proof. Consider the extremal system f˝1 ; ˝2 ; xg. Under the given assumptions
we
Q get with QTheorem 5.5.7 that the approximate extremal principle holds in
. i 2I Xi /. j 2J Xj / with the corresponding index sets I and J therein satisfying
the condition I \ J D ;. Following the line of the proof in [431, Lemma 5.58] we
get the desired result. t
u
Chapter 6
Hahn-Banach Type Theorems
In this chapter we present a generalization of the Hahn–Banach–Kantorovich

extension theorem to K-convex set-valued maps, as well as Yang’s extension
theorem. We also present classical separation theorems for convex sets, the core
convex topology on a linear space, and a criterion for the convexity of the cone
generated by a set.
In the sequel X , Y , Z are real linear spaces. Recall that the class of linear
operators from X into Y is denoted by L.X; Y /; we set X 0 WD L.X; R/. In the
case in which X is a topological vector space, as usual, we denote by X its
topological dual; of course, X X 0 . Moreover, if X and Y are topological vector
spaces, L .X; Y / denotes the set of continuous linear operators from X into Y . We
consider K Z a proper (i.e. f0g ¤ K ¤ Z) convex cone. The cone K induces a
quasi-order on Z, denoted K or simply if there is no risk of confusion. So, for
z1 ; z2 2 Z one has z1 z2 (or equivalently z2
z1 ) if z2 z1 2 K. We denote by
core A the algebraic interior (or core) of A X and by icr A the relative algebraic
interior (or intrinsic core) of A, that is icr A is the algebraic interior of A with respect
to the affine hull aff A of A. It follows that core A ¤ ; if and only if aff A D X and
icr A ¤ ;; in this case (that is aff A ¤ X ) icr A D core A. Of course, if A X is
an affine set (that is, aff A D A) then icr A D A. Recall that for A a convex set one
has
a 2 core A , Œ8x 2 X; 9
2 P W a C
x 2 A ; (6.1)
and
a 2 icr A , Œ8x 2 A; 9
2 P W .1 C
/a
x 2 A ; (6.2)
where P WD 0; 1Œ R. When A is convex we have
a0 2 icr A; a 2 A;
2 0; 1Œ ) .1
/a0 C
a 2 icr AI (6.3)

DOI 10.1007/978-3-642-54265-7__6
250 6 Hahn-Banach Type Theorems
in particular, icr A is a convex set. Indeed, take a0 2 icr A, a 2 A and

2 0; 1Œ, and
set a WD .1
/a0 C
a. Consider x 2 A. Because a0 2 icr A, by (6.2), there exists
> 0 such that x WD .1 C /a0 x 2 A. Then ˛ WD .1
/=.1 C
/ 2 0; 1Œ,
WD ˛ > 0, and .1 C /a x D ˛x C .1 ˛/a 2 A. Using again (6.2) we get
a 2 icr A.
When K0 WD core K ¤ ;, we use z1 < z2 (or z2 > z1 ) when z2 z1 2 K0 , and
z1 6< z2 (or z2 6> z1 ) if z2 z1 … K0 .
As in Sect. 2.1, we extend Z to Z WD Z [ f1; C1g, ˙1 … Z, .˙1/ D
1, and consider that 1 z C1 (even 1 < z < C1 if K0 ¤ ;) for
all z 2 Z; moreover, z C .˙1/ WD ˙1, 0 .C1/ WD C1, 0 .1/ WD 0,
t .˙1/ WD ˙1 for all z 2 Z and t 2 P.
For A; B Z and a relation on Z (for example can be , <, 6<) we write
AB whenever ab for all a 2 A and b 2 B. If a 2 X , we write aB if fagB.
Again, as in Sect. 2.1, for a function f W X ! Z , dom f WD fx 2 X j f .x/ ¤
C1g and epi f WD f.x; z/ 2 X Z j f .x/ zg; f is proper if dom f ¤ ;
and f does not take the value 1, while f is convex if epi f is convex. We say
that f W X ! Z is sublinear if f is proper, f .0/ D 0, f .tx/ D tf .x/ and
f .x C x 0 / f .x/ C f .x 0 / for all x; x 0 2 X and t 2 P; hence epi f is a convex
cone containing the origin when f is sublinear.
In the sequel we identify a set-valued map W E F (that is, a function
W E ! 2F ) with its graph graph WD f.x; y/ 2 E F j y 2 .x/g. So, several
times in the sequel, having a subset A of E F we interpret A as the graph of a
set-valued map from E to F , and so A.x/ WD fy 2 F j .x; y/ 2 Ag.
6.1 The Hahn–Banach–Kantorovich Theorem
In this section the proper convex cone K Z is such that .Z; K/ has the least
upper bound property (l.u.b.p., for short), i.e. every nonempty and upper bounded
set has a least upper bound; of course, in this situation every nonempty and lower
bounded set has a greatest lower bound. If the nonempty set B Z is bounded
from below (resp. above), we denote by inf B (sup B) a greatest lower (resp. least
upper) bound of B; if B is not bounded below (resp. above) we set inf B WD 1
(resp. sup B WD C1). We also use the notation inf ; WD C1 and sup ; WD 1.
We first give an extension theorem for linear mappings bounded above by convex
set-valued maps. As we shall see below, the result is equivalent to the classical
Hahn–Banach–Kantorovich theorem when the intrinsic core of the domain of the
involved set-valued map is nonempty.
Theorem 6.1.1. Let W X Z be a K-convex set-valued map, X0 X a linear
subspace and T0 2 L.X0 ; Z/. Suppose that 0 2 icr .dom X0 / and T0 x .x/
for all x 2 X0 .\ dom /. Then there exists T 2 L.X; Z/ such that T jX0 D T0
and Tx .x/ for all x 2 X .
6.1 The Hahn–Banach–Kantorovich Theorem 251
Proof. Consider X WD aff .dom X0 /. Because 0 2 icr .dom X0 /

dom X0 , X is a linear space.
The case X D X ; hence 0 2 dom .dom X0 /. Consider the set
˚
F WD .E; T / j X0 E D span E X; T 2 L.E; Z/;

T jX0 D T0 ; Tx .x/ 8x 2 E I
clearly, .X0 ; T0 / 2 F . For .E1 ; T1 /; .E2 ; T2 / 2 F we set .E1 ; T1 / .E2 ; T2 / if

E1 E2 and T2 jE1 D T1 . Clearly, is a partial order on F . Take C WD f.Ei ; Ti / j
i 2 I g a chain in F .I ¤ ;/. Define E WD [i 2I Ei and T W E ! Z defined
by Tx WD Ti x if x 2 Ei . Because C is a chain, E is a linear subspace of X with
X0 E. Moreover, T is a well defined linear operator and Tx z for .x; z/ 2
\ .E Z/; this is valid because x being in E, there exists i 2 I with x 2 Ei ,
and so .x; z/ 2 \ .Ei Z/ and Tx D Ti x. It follows that .E; T / 2 F and .E; T /
is an upper bound for C . Using Zorn’s lemma, .F ; / has maximal elements; let
.E; T / 2 F be maximal. If E D X the linear operator T verifies the conditions in
the conclusion of the theorem.
Assume that E ¤ X . Then there exists x 2 X n E. Because X0 E we have
that dom X0 dom E, and so 0 2 core .dom E/. Set E1 WD E C Rx.
Clearly E E1 , the inclusion being strict. Since 0 2 core .dom E/, there
exists
2 P such that ˙
x 2 dom E, and so there exist x1 2 E and z1 2 Z
such that .x1 C
x; z1 / 2 and x2 2 E and z2 2 Z such that .x2
x; z2 / 2 .
Set A WD epi . It follows that the sets
ˇ
z1 Tx1 ˇˇ
B1 WD x
ˇ 1 2 E; z1 2 Z;
1 2 P W .x1 C
1 x; z1 / 2 A

1
D fz1 Tx1 j x1 2 E; z1 2 Z W .x1 C x; z1 / 2 PAg ;
and
ˇ
Tx2 z2 ˇˇ
B2 WD ˇ x2 2 E; z2 2 Z;
2 2 P W .x2
2 x; z2 / 2 A

2
D fTx2 z2 j x2 2 E; z2 2 Z W .x2 x; z2 / 2 PAg ;
are nonempty (and convex) sets. Moreover,
8 b1 2 B1 ; 8b2 2 B2 W b1
b2 :
Indeed, let b1 2 B1 and b2 2 B2 ; then b1 D z1 Tx1 and b2 D Tx2 z2 with

x1 ; x2 2 E, z1 ; z2 2 Z and .x1 C x; z1 /; .x2 x; z2 / 2 PA. It follows that .x1 C
x2 ; z1 C z2 / 2 PA (because this set is a convex cone), and x1 C x2 2 E. From the
hypothesis, we have that T .x1 C x2 / z1 C z2 , which is equivalent to b2 b1 .
It follows that B2 is bounded from above and any element of B1 is an upper bound
for B2 . Take z 2 Z a least upper bound of B2 ; then b2 z b1 for all b1 2 B1 and

b2 2 B2 . Let us define T1 by T1 .x C
x/ WD TxC
z (for x 2 E and
2 R); we have
that T1 W E1 ! Z is linear, T1 jE D T and T1 x z for all .x; z/ 2 \.E1 Z/. Let
us prove the last assertion. So, take .x; z/ 2 such that x WD x0 C
x with x0 2 E
and
2 R. If
D 0 then .x; z/ D .x0 ; z/ 2 \ .E Z/, and so T1 x D Tx0 z.
If
< 0 then
2 WD
> 0 and T1 x D Tx0
2 z. Setting b2 WD
1 2 .Tx0 z/ we
have that b2 2 B2 . But
Tx0 z
T1 x z ” Tx0
2 z z ” z ” b2 z:

2
Hence our assertion is true in this case by the choice of z. The case
> 0 is proven
similarly. It follows that .E1 ; T1 / 2 F and .E; T / .E1 ; T1 /. Since .E; T / is
maximal we obtain that .E; T / D .E1 ; T1 /, whence the contradiction E D E1 .
Therefore our assumption that E ¤ X is false.
The case X ¤ X . Of course, dom X . Taking x0 2 X0 \ dom , we have
that X0 D x0 X0 X . Applying the first case we find T W X ! Z a linear
operator such that T jX0 D T and T x z for all .x; z/ 2 X Z. Taking Y a
linear subspace of X such that X D X ˚ Y (that is X D X C Y and X \ Y D f0g)
and T W X ! Z defined by T .x C y/ WD T .x/ for x 2 X , y 2 Y , T verifies the
conclusion. The proof is complete. t
u
Remark 6.1.2. In the proof of Theorem 6.1.1 we used only the fact that P epi
is convex (or, equivalently, RC epi is a convex cone) and not that epi itself is
convex.
Remark 6.1.3. The preceding result is obtained by Brumelle [81, Lemma 1] for
X0 ¤ f0g and convex with 0 2 .0/ and 0 2 core.dom /, and by Nehse [439,
Theorem 2]. Moreover, Chen and Wang [93, Theorem 1] obtained Theorem 6.1.1 for
convex-valued with dom D X and 0 2 .0/. The proof here is from Zălinescu
[615].
Recently, Du (see [147, Theorem 2.1]) obtained the following result (for the case
in which X0 \ core.dom / ¤ ;).
subspace and W X0 Z a .K/-convex set-valued map with dom D X0 .
Suppose that 0 2 icr .dom X0 / and .x/ .x/ for all x 2 X0 . Then there
exists a .K/-convex set-valued map ˚ W X Z with dom ˚ D X such that
˚jX0 D and ˚.x/ .x/ for all x 2 X .
Proof. Consider the set-valued map Q W X Z whose graph is graph Q WD
epiK epiK . Clearly, graph Q D epiK Q is convex, and so Q is K-convex;

moreover, dom Q D dom dom D dom X0 and Q .0/ D [x2X0 .x/

.x/ C K . It follows that 0 2 icr.dom Q / and 0 Q .0/. Applying Theorem 6.1.1
for Q and T0 WD 0 2 L.f0g; Z/, we get T 2 L.X; Z/ such that Tx Q .x/ for
every x 2 X . Hence
Tx Tx0 z z0 8x 2 dom ; 8z 2 .x/; 8x 0 2 X0 ; 8z0 2 .x 0 /: (6.4)
Let X1 X be a linear space such that X D X0 ˚ X1 and define ˚ W X Z

by ˚.x0 C x1 / WD .x0 / C Tx1 for x0 2 X0 and x1 2 X1 . Clearly, dom ˚ D X
and epiK ˚ D epiK C graph T jX1 , and so ˚ is .K/-convex. Fix x 2 dom ,
z 2 .x/ and z0 2 ˚.x/. Then x D x0 Cx1 with x0 2 X0 , x1 2 X1 , and z0 D z0 CTx1
with z0 2 .x0 /. Using (6.4) we get Tx1 z z0 , or, equivalently, z0 z. It follows
that ˚.x/ .x/ for every x 2 X . t
u
Remark 6.1.5. Theorem 6.1.4 looks more general than Theorem 6.1.1. However
it is not possible to obtain Theorem 6.1.1 from Theorem 6.1.4 because instead
of a linear operator T with T jX0 D T0 we get a .K/-convex set-valued map
having as domain the entire space X . However, by the procedure used in the proof
of Theorem 6.1.4 we can obtain Theorem 6.1.1 from its particular case in which
X0 WD f0g.
Corollary 6.1.6. Let P X be a convex cone and let Y X be a linear subspace.
If S0 2 L.Y; Z/ is monotone, that is x1 ; x2 2 Y and x1 P x2 imply S0 x1 S0 x2 ,
and Y P is a linear subspace, then there exists S 2 L.X; Z/ monotone with
S jY D S 0 .
Proof. Consider W X Z with graph WD graph S0 P f0g; clearly, is
K-convex and dom D Y P . Take X0 WD f0g and T0 D 0 2 L.X0 ; Z/. Then
dom X0 D Y P , and so 0 2 icr.dom X0 /. Moreover, if .x; z/ 2 graph \
.X0 Z/, then x D 0 D y v with y 2 Y , v 2 P and z D S0 y. Hence y D v
P 0,
whence z D S0 y
S0 0 D 0. Hence 0 D T0 x z. Applying Theorem 6.1.1 we
get T 2 L.X; Z/ such that Tx z for all .x; z/ 2 graph S0 P f0g, that is,
y 2 Y and v 2 P imply T .y v/ S0 y. In particular, Ty S0 y 2 K \ .K/
for every y 2 Y and Tv
S0 0 D 0 for every v 2 P , that is T is monotone. Let us
consider Y 0 X a linear space such that X D Y ˚ Y 0 and define T0 W X ! Z by
T0 .y C y 0 / WD S0 y Ty for y 2 Y and y 0 2 Y 0 ; then T0 x 2 K \ .K/ for every
x 2 X . Taking S WD T the conclusion follows. t
u
Corollary 6.1.7. Let f W X ! Z be a proper convex operator, X0 be a linear
subspace of X and T0 2 L.X0 ; Z/. Suppose that T0 x f .x/ for every x 2 X0 .
If 0 2 icr.dom f X0 / then there exists T 2 L.X; Z/ such that T jX0 D T0 and
Tx f .x/ for every x 2 X .
Proof. In Theorem 6.1.1 take graph WD epi f ; of course, the hypothesis of the
theorem holds, so that there exists T 2 L.X; Z/ such that T jX0 D T0 and Tx
.x/ for every x 2 X . Since f .x/ 2 .x/ for x 2 dom f D dom and f .x/ D
C1 for x 2 X n dom f , the conclusion follows. t
u
Remark 6.1.8. Corollary 6.1.6 is the implication (LUB))(MP) in [127, Theorem
VI.3.1]. For f sublinear with dom f D X , Corollary 6.1.7 is the well-known
Kantorovich’ generalization of the Hahn–Banach extension theorem, and is the
implication (LUB))(HB) in [127, Theorem VI.3.1]; moreover, Corollary 6.1.7
for f sublinear is equivalent to [412, Theorem 3 (3)], which, at its turn, covers
[412, Theorem 3 (1)]. Corollary 6.1.7 for f with X0 \ core.dom f / ¤ ; is [629,
Theorem 2.1].
Assume that f is sublinear; then condition 0 2 icr.dom f X0 / in Corol-
lary 6.1.7 is equivalent to (H) “X0 Cdom f is a linear subspace of X ”. We emphasize
the importance of condition (H) in Corollary 6.1.7. Without this condition the
conclusion of Corollary 6.1.7 could be false even for Z D R and dim X < 1;
see [534] and [7] for interesting counterexamples.
Using Theorem 6.1.1 we get easily the following separation theorem obtained by
Nehse in [439, Theorem 1].
Proposition 6.1.9. Let A; B X Z be such that RC .A B/ is convex. Assume
that 0 2 icr.PrX .A B// and A.x/ B.x/ for all x 2 PrX .A/ \ PrX .B/. Then
there exist T 2 L.X; Z/ and z0 2 Z such that
8.x1 ; z1 / 2 A; 8.x2 ; z2 / 2 B W z1 Tx1 z0 z2 Tx2 : (6.5)
Proof. Consider the set C WD B A X Z. Then PrX .C / D PrX .A/ PrX .B/
and if .0; z/ 2 C then z D z1 z2 with z1 2 A.x/, z2 2 B.x/ for some x 2
PrX .A/ \ PrX .B/; hence 0 z. Taking X0 WD f0g and T .0/ WD 0, we can apply
Theorem 6.1.1 (see also Remark 6.1.2) for graph WD C and T0 . Therefore, there
exists T 2 L.X; Z/ such that Tx z for every .x; z/ 2 C . It follows that
8.x1 ; z1 / 2 A; 8.x2 ; z2 / 2 B W z1 Tx1 z2 Tx2 :
Taking z0 WD sup fz1 Tx1 j .x1 ; z1 / 2 Ag, the conclusion follows. t

u
Before stating the next result recall that the subdifferential @f .x0 / of the proper
operator f W X ! Z at x0 2 dom f is the set of those T 2 L.X; Z/ such that
8 x 2 X W Tx Tx0 f .x/ f .x0 /:
(In the continuous case, @f .x0 / \ L .X; Y / is denoted by @s f .x0 /; see Sect. 7.1.)
Corollary 6.1.10. Let f W X ! Z be a proper convex operator and consider
x0 2 icr.dom f /. Then @f .x0 / is nonempty.
Proof. Consider g W X ! Z , g.x/ D f .x0 C x/ f .x0 /. Then g is a convex
operator with 0 2 icr.dom g/. Consider now graph D epi g, X0 D f0g and
T0 .0/ WD 0; is a convex set-valued map. As dom D dom g, we have that
0 2 icr.dom X0 /, and, of course, .x; z/ 2 \ .X0 Z/ implies T0 x D 0 z.
Applying Theorem 6.1.1 we get T 2 L.X; Z/ such that Tx z for every .x; z/ 2 .
In particular, if x 2 dom f then .xx0 ; f .x/f .x0 // 2 epi g, whence T .xx0 /
f .x/ f .x0 /. t
u
Note that Corollary 6.1.10 can be viewed as a particular case of the next result;
just take g W X ! Z defined by g.x0 / WD f .x0 / and g.x/ WD C1 for x 2 X n
fx0 g. The next result is the sandwich theorem proved by Zowe in [629, Theorem 3.1]
(see also [610, Corollary 2.6]).
Corollary 6.1.11. Let f , g W X ! Z be proper convex operators. Suppose that
0 2 icr.dom f dom g/ and that f .x/
g.x/ for all x 2 dom f \ dom g. Then
there exists T 2 L.X; Z/ and z0 2 Z such that
8 x 2 X W g.x/ Tx C z0 f .x/:
Proof. Consider W X Z with
graph WD f.x; z/ 2 X Z j f .x/ zg C f.x 0 ; z0 / 2 X Z j g.x 0 / z0 g:
Then is convex and dom D dom f dom g. The hypothesis of Theorem 6.1.1
holds for X0 D f0g and T0 .0/ WD 0. Indeed, if .0; z/ 2 then .0; z/ D .x; z0 / C
.x; z00 /, with f .x/ z0 and g.x/ z00 ; it follows that T0 .0/ D 0 f .x/Cg.x/
z0 C z00 D z. Therefore, there exists T 2 L.X; Z/ such that Tx z for .x; z/ 2 .
In particular, for x1 2 dom f and x2 2 dom g we get T .x1 x2 / f .x1 / C g.x2 /,
which yields g.x2 / Tx2 f .x1 / Tx1 for all x1 2 dom f and x2 2 dom g. It
follows that
sup .g.x2 / Tx2 / inf .f .x1 / Tx1 / :

x2 2dom g x1 2dom f
Taking z0 between these two values, we get the desired conclusion. t

u
Note that Theorem 6.1.1 follows from Corollary 6.1.11 when icr.dom / is
nonempty. To show this we need the following result (see [610, Proposition 1.5
(i)]).
Proposition 6.1.12. Let W X Z be K-convex set-valued map. Assume that
.x0 / is bounded below for some x0 2 icr.dom /. Then
f W X ! Z ; f .x/ WD inf .x/; (6.6)
is a proper convex operator with dom f D dom .

Proof. From the definition of f we have that dom f D dom . Let x 2 dom .
Because dom D PrX .epi / is convex and x0 2 icr.dom /, using (6.2), there
exists
> 0 such that x 0 WD .1 C
/x0
x 2 dom . Fix z0 2 .x 0 / and take
z 2 .x/ arbitrary. Then, by the convexity of epi , we have

1 0
1
x0 ; zC z D .x; z/ C .x 0 ; z0 / 2 epi ;
1C
1C
1C
1C
1 0
and so Z 3 f .x0 /
1C
z C 1C
z. It follows that
.1 C
/
1 f .x0 / z0 z 8z 2 .x/:
Hence f .x/ 2 Z, and so f is proper.

Take now x0 ; x1 2 dom f and
2 0; 1Œ, and set x
WD .1
/x0 C
x1
.2 dom D dom f /. For z0 2 .x0 / and z1 2 .x1 / we have that
.x
; .1
/z0 C
z1 / D .1
/.x0 ; z0 / C
.x1 ; z1 / 2 epi ;
whence f .x
/ .1
/z0 C
z1 . Taking the infimum with respect to z0 2 .x0 /
and z1 2 .x1 / we get f .x
/ .1
/f .x0 / C
f .x1 /. Hence f is
convex. t
u
As mentioned above, Theorem 6.1.1 follows from Corollary 6.1.11 when
icr.dom / is nonempty. Indeed, in this case there exists x0 2 X0 \ icr.dom / [see
Proposition 6.3.2 (iii)]. Taking f defined in (6.6), we have that T0 .x0 / f .x0 / <
C1, and so f .x0 / 2 Z. Using Proposition 6.1.12 we get f .x/ 2 Z for every
x 2 dom . Taking g.x/ WD T0 x for x 2 X0 , g.x/ WD C1 for x 2 X n X0 and
applying Corollary 6.1.11 we get the conclusion of Theorem 6.1.1.
The preceding discussion shows that only the case when the intrinsic core of
dom is empty Theorem 6.1.1 is not covered by known results (however see [412,
Theorem 3]). It is worth observing that under the hypotheses of Theorem 6.1.1 we
have that f .x/ WD inf .x/ 2 Z for every x 2 dom and T0 x f .x/ for every
x 2 X0 \ dom . For this use the conclusion of Theorem 6.1.1. However a direct
proof is not so obvious as that in the case X0 \ icr.dom / ¤ ;.
Let us point out the following consequence of Theorem 6.1.1. Similar to [458,
Definition 1.3], we say that H W X Z is affinelike if there exist T 2 L.X; Z/
and a nonempty convex set M Y such that H.x/ D T .x/ C M for every x 2 X .
Corollary 6.1.13. Let F W X Z be a convex set-valued map, X0 X a
linear subspace and H0 W X0 Z an affinelike set-valued map. Suppose that
0 2 icr .dom F X0 / and F .x/ H0 .x/ K for every x 2 X0 \ dom F . Then
there exists H W X Z an affinelike set-valued map such that H jX0 D H0 and
F .x/ H.x/ K for every x 2 X \ dom F .
Proof. Let T0 2 L.X0 ; Z/ and M Y a nonempty convex set such that H0 .x/ D
T0 .x/ C M for every x 2 X0 . Consider W X Z with
graph WD f.x; z m C k/ j z 2 F .x/; m 2 M; k 2 Kg D epi F f0g M:
One obtains immediately that is convex, dom D dom F and T0 .x/ z for all
.x; z/ 2 \ .X0 Z/. Applying Theorem 6.1.1 we get T 2 L.X; Z/ such that
T jX0 D T0 and T .x/ z for all .x; z/ 2 . Setting H.x/ WD T .x/ C M , the
conclusion follows. t
u
Concerning the affine and affinelike set-valued maps we mention the following
result. The equivalence of (a), (b”) and (c) (in an slightly different form) is
mentioned in [572, 573].
Proposition 6.1.14. Let A X Z. Consider the following assertions.
(a) A is an affine manifold;
(b)
A.x/ C .1
/A.x 0 / A .
x C .1
/x 0 / for all x; x 0 2 X and
2 R;
(b’)
A.x/ C .1
/A.x 0 / A .
x C .1
/x 0 / for all x; x 0 2 PrX .A/ and

2 R;
(b”)
A.x/ C .1
/A.x 0 / D A .
x C .1
/x 0 / for all x; x 0 2 PrX .A/ and

2 R;
(c) there exist linear subspaces X0 X , Z0 Z, a linear map T0 W X0 ! Z and
.x0 ; z0 / 2 X Z such that PrX .A/ D x0 CX0 and A.x/ D T0 .xx0 /Cz0 CZ0
for every x 2 PrX .A/;
(d) A is an affinelike set-valued map on a linear subspace X0 X .
Then (a) , (b) , (b’) , (b”) , (c); if X0 WD PrX .A/ is a linear space then
(a) ) (d).
Proof. We may (and do) assume that .0; 0/ 2 A; otherwise replace A by A.x0 ; z0 /
with .x0 ; z0 / 2 A.
The fact that (a) , (b) , (b’) ( (b”) is simple verification, (b) being a rewriting
of (b’) taking into account the fact that B C ; D ; C B D ; for B Z.
) (c) Because
(a) A is a linear subspace, X0 WD PrX .A/ and Z0 WD A.0/ D
PrZ A \ .f0g Z/ are linear spaces. Take Z1 Z a linear subspace such that
Z D Z0 ˚ Z1 . Observe first that for .x; z/; .x; z0 / 2 A we have .0; z z0 / 2 A,
and so z z0 2 Z0 . Hence, taking T0 W X0 ! Z defined by T0 .x/ WD z1 , where
z D z0 C z1 2 A.x/ with z0 2 Z0 , z1 2 Z1 , T0 is well defined. Since A is a linear
subspace it follows immediately that T0 is a linear operator and A.x/ D T0 .x/ C Z0
for every x 2 X0 .
(c) ) (b”) follows by a simple verification.
(c) ) (d) is obvious when X0 WD PrX .A/ is a linear space. t
u
Observe that (d) ) (a) if and only if the set M in the definition of an affinelike
set-valued map is an affine set.
Note also that the spaces X0 and Z0 are uniquely determined by A; in fact X0 is
the parallel subspace of PrX .A/ and Z0 is the parallel subspace of A.x0 / for some
x0 2 PrX .A/. The equivalences of (a), (b”) and (c) (the last one presented in the
form there exist an affine mapping T0 W PrX .A/ ! Z and a linear subspace Z0 of
Z such that A.x/ D T0 .x/ C Z0 for every x 2 PrX .A/) are provided in [573]. Note
that when Z is ordered by the convex cone K, in [573] one says that the affine set-
valued map A X Z is non-vertical when every two distinct elements from A.x/
are not comparable for each x 2 X , which is equivalent, by [573, Lemma 3.1], to
Z0 \ K D f0g where Z0 is provided by Proposition 6.1.14 (d).
We do not treat here the continuous versions of the preceding results, that is the
case in which X; Z are topological linear space and the involved linear operators
are taken to be continuous. This can be done as in Section 4 of [610].
6.2 Classical Separation Theorems for Convex Sets
We say that the nonempty subsets A; B of X are separated by the hyperplane

Hx ;˛ WD fx 2 X j hx; x i D ˛g, where x 2 X 0 n f0g and ˛ 2 R, if
inf x .A/
˛
sup x .B/ [or inf x .B/
˛
sup x .A/]; the separation
is proper or strict if, moreover, A [ B 6 Hx ;˛ or inf x .A/ > sup x .B/ or
inf x .B/ > sup x .A/], respectively. Note that A and B can be (properly, strictly)
separated (by some hyperplane) if and only if A B and f0g can be (properly,
strictly) separated. Moreover, note that the nonempty subsets A; B of X can be
(properly, strictly) separated if and only if conv A and conv B can be (properly,
strictly) separated.
For deriving separation theorems the Minkowski functional proves to be very
useful.
Proposition 6.2.1. Let A X be such that 0 2 core A and Œ0; 1A D A. Then the
Minkowski functional
pA W X ! R; pA .x/ WD inff˛ > 0 j x 2 ˛Ag;
is positively homogeneous and
fx 2 X j pA .x/ < 1g A fx 2 X j pA .x/ 1g: (6.7)
If, furthermore, A is convex, then A is sublinear and
core A D fx 2 X j pA .x/ < 1g: (6.8)
Proof. For x 2 X set .x/ WD f˛ > 0 j x 2 ˛Ag. Because 0 2 core A, we have

that .x/ ¤ ; for every x 2 X and .0/ D P. Hence pA .0/ D 0 and pA .x/ 2 R
for x 2 X . From the very definition of .x/ we have that .˛x/ D ˛.x/ for all
x 2 X and ˛ 2 P; hence pA is positively homogeneous. Moreover, if ˛ 2 .x/
1
and ˛ ˇ then ˇ 2 .x/. ˛ˇ 2 Œ0; 1, and so, because A D Œ0; 1A, we
Indeed,
1
1 1
have that ˇ x D ˛ˇ ˛ x 2 A; hence ˇ 2 .x/. Hence, if pA .x/ < 1 then
there exists ˛ 2 .x/ with ˛ < 1, and so 1 2 .x/, whence x 2 A; if x 2 A then
1 2 .x/, and so pA .x/ 1. Therefore, (6.7) holds.
Assume, moreover, that A is convex. Since .x/ C .x 0 / .x C x 0 / for all
x; x 0 2 X , we obtain immediately that pA is sub-additive, and so pA is sublinear.
Let x 2 core A. By (6.1) there exists
> 0 such that x C
x 2 A, and so
.1 C
/1 2 .x/. Therefore, pA .x/ .1 C
/1 < 1. Conversely, assume that
6.2 Classical Separation Theorems for Convex Sets 259
a 2 X is such that pA .a/ < 1, and take x 2 X . Then for

> 0 sufficiently small,
using the fact that pA is sublinear, we have that pA .a C
x/ pA .a/C
pA .x/ < 1,
whence a C
x 2 A. Therefore, a 2 core A, and so (6.8) holds. t
u
We provide now an algebraic separation theorem.
Theorem 6.2.2. Let A X be a convex set with icr A ¤ ; and let x0 2 X n icr A.
Then A and fx0 g can be properly separated, that is, there exists x 2 X 0 which is
not constant on A [ fx0 g such that
8x 2 A W hx; x i hx0 ; x i :
Proof. W.l.o.g. we may assume that 0 2 icr A (otherwise we do a translation).

Consider first the case span A D aff A D X . In this situation, A is a convex set
with 0 2 core A. Using Proposition 6.2.1 we have that pA W X ! R is sublinear
and (6.7) holds. Consider X0 WD Rx0 and T0 2 L.X0 ; R/ defined by T0 .
x0 / WD
pA .x0 /; clearly T0 .x/ pA .x/ for every x 2 X0 . By Corollary 6.1.7 we get

x 2 L.X; R/ D X 0 such that x jX0 D T0 and x .x/ D hx; x i pA .x/ for
every x 2 X . Using (6.7), since x0 … core A we have that hx0 ; x i D T0 .x0 / D
pA .x0 /
1 and for x 2 A we get hx; x i pA .x/ 1. Clearly, x is not constant
on A [ fx0 g because x .0/ D 0 < 1 x .x0 /. The conclusion follows in this case.
Let now X0 WD span A D aff A ¤ X . We have two possible situations: a)
x0 2 X0 , and b) x0 … X0 . In case a), from the preceding situation, we get x0 2 X00
which is not constant on A [ fx0 g such that x0 .x/ x0 .x0 / for every x 2 A. Any
x 2 X 0 with x jX0 D x0 does the job. In case b) there exists x 2 X 0 such that
x jX0 D 0 and x .x0 / D 1; this x does the job. t
u
We provide now other variants of the separation theorem.
Theorem 6.2.3. Let A; B X be nonempty convex sets. The sets A and B can be
properly separated in each of the following situations:
(i) core A ¤ ; and B \ core A D ;,
(ii) icr A ¤ ;, icr B ¤ ; and icr A \ icr B D ;,
(iii) 0 … icr.A B/ ¤ ;.
Proof. Using Proposition 6.3.2 (iii) and (iv) we have that (ii) ) (iii) and (i) )
(iii), respectively. Assume that (iii) holds. By the Theorem 6.2.2, f0g and A B
can be separated properly, and so, as noticed above, A and B can be separated
properly. t
u
In the case in which X is a topological linear space one prefers to have hyper-
planes determined by continuous linear functionals when dealing with separation
theorems; in this case we speak about separation by closed hyperplanes. This is
motivated by the following result.
Proposition 6.2.4. Let .X; / be a topological vector space and x 2 X 0 n f0g.
Then the following statements are equivalent:
(i) x is continuous,
(ii) Hx ;˛ is closed for some (for every) ˛ 2 R,
(iii) the interior of the half-space Hx ;˛ WD fx 2 X j hx; x i ˛g is nonempty for
some (for every) ˛ 2 R.
Proof. First observe that Hx ;˛ D x˛ C Hx ;0 and Hx ;˛ D x˛ C Hx ;0 for every
˛ 2 R and x˛ 2 X with hx˛ ; x i D ˛; such an x˛ exists because x ¤ 0. So, in (ii)
and (iii) we can (and do) take ˛ D 0.
(i) ) (ii) is obvious.
(ii) ) (iii) Take x0 2 X with hx0 ; x i < 0. Clearly, x0 … Hx ;0 .D ker x /, and
so, because ker x is closed, there exists a balanced neighborhood U of 0 such that
.x0 CU /\ker x D ;. Assume that x0 CU 6 Hx ;0 . Then there exists u0 2 U such
that x0 C u0 … Hx ;0 , that is hx0 C u0 ; x i > 0. Then t WD hx0 ; x i = hu0 ; x i 2
0; 1Œ, and so tu0 2 U . Because hx0 C tu0 ; x i D 0, we get the contradiction .x0 C
U / \ ker x ¤ ;. Hence x0 2 int Hx ;0 , whence int Hx ;0 ¤ ;.
(iii) ) (i) Let x0 2 int Hx ;0 . Then ˛ WD hx0 ; x i > 0; else x is non-positive
on a neighborhood of the origin, and so x D 0. It follows that 0 2 int Hx ;˛ , and
so there exists a balanced neighborhood U of 0 with U Hx ;˛ . It follows that
jhu; x ij ˛ for every u 2 U , whence jhu; x ij "˛ for every u 2 "U , for every
" > 0. This proves the continuity of x . t
u
Using Theorem 6.2.3 and Proposition 6.2.4 we get the following result on
separation by closed hyperplanes.
Theorem 6.2.5. Let X be a separated locally convex space and let A; B X be
nonempty convex sets.
(i) If int A ¤ ; and B \ int A D ;, or more generally, if 0 … int.A B/ ¤ ;, then
there exists x 2 X n f0g such that sup x .A/ inf x .B/.
(ii) If A is compact, B is closed and A\B D ;, or more generally, if 0 … cl.AB/,
then there exists x 2 X such that sup x .A/ < inf x .B/.
Proof. (i) If int A ¤ ; then core.A B/ D int.A B/ D .int A/ B ¤ ;. So,
let 0 … int.A B/ ¤ ;. Using Theorem 6.2.2 (iii) for A B and x0 WD 0, we
get x 2 X 0 n f0g such that sup x .A B/ 0, that is A B Hx ;0 . From
Proposition 6.2.4 we obtain that x 2 X . The conclusion follows.
(ii) Of course, if A is compact and B is closed, then A B is closed; moreover, if
A \ B D ; then 0 … A B D cl.A B/. So, assume that 0 … cl.A B/. Then
there exists a balanced convex neighborhood U of 0 such that U \.B A/ D ;.
Using (i) we get x 2 X n f0g such that
sup x .U / inf x .B A/ D inf x .B/ sup x .A/:
Since U is a neighborhood of 0 and x ¤ 0 we obtain that sup x .U / > 0, and

so sup x .A/ < inf x .B/. The proof is complete. t
u
6.3 The Core Convex Topology 261
Remark 6.2.6. Observe that in the case in which X is a topological vector space
and A X is a convex set such that 0 2 int A, then pA is a continuous sublinear
function and int A D fx 2 X j pA .x/ < 1g, cl A D fx 2 X j pA .x/ 1g.
6.3 The Core Convex Topology
As seen in the previous sections, we have established several results in real linear
spaces without topology. However, every real linear space X can be endowed with
the strongest locally convex topology c WD P , where P is the family of all the
semi-norms defined on X ; c is called the core convex topology in the sequel. In
the next proposition we collect several results concerning this topology.
Proposition 6.3.1. Let X be a real linear space. Then
(i).X; c / is a (separated) locally convex space.
(ii)X WD .X; c / D X 0 .
(iii)intc C D core C for every convex set C X .
(iv) Every convex function f W X ! R is c -continuous on core.dom f / D
int.dom f /.
(v) Let f W X ! R be a proper convex function. Then the restriction of f to
aff.dom f / is continuous on icr.dom f /. Moreover, f is subdifferentiable on
icr.dom f /.
Proof. (i) Note that fj'j j ' 2 X 0 g P. Because for every x 2 X n f0g there
exists ' 2 X 0 with '.x/ ¤ 0, we obtain that P is sufficient, and so c is
separated.
(ii) We observed above that X X 0 ; let ' 2 X 0 . Then p WD j'j 2 P. Since
j'j p, ' is c -continuous, and so ' 2 X .
(iii) Let C X be a convex set. The inclusion int C core C is well known.
Take x 2 core C . Then A WD C x WD fc x j c 2 C g is absorbing (and
convex). Therefore, A0 WD A \ .A/ is an absorbing, symmetric convex set.
It follows that the Minkowski functional pA0 associated to A0 is a semi-norm
on X . Moreover, core A0 D fx 2 X j pA0 .x/ < 1g A0 . Because pA0 2 P,
pA0 is c -continuous, and so fx 2 X j pA0 .x/ < 1g is c -open and contains 0.
Therefore, 0 2 int A0 int.C x/ D int C x. Hence x 2 int C . Thus we
have the inclusion core C int C , too.
(iv) Let f W X ! R. From (iii) we have that core.dom f / D int.dom f /. It is well
known that a convex function is 1 on icr.dom f / if f takes the value 1
(see e.g. [614, Proposition 2.1.4]). Hence f is continuous on core.dom f / if f
takes the value 1. Thus assume that f is proper and take x0 2 core.dom f /.
Consider A WD fx 2 X j f .x/ f .x0 / C 1g. It is easy to see that x0 2 core A,
that is, x0 2 int A. Hence A is a c -neighborhood of x0 and so f is bounded
above on a neighborhood of x0 . Therefore, f is continuous at x0 (see e.g. [614,
Theorem 2.2.9]).
(v) Let x0 2 icr.dom f /. Taking the usual translation (g.x/ WD f .x0 C x/), we
may assume that x0 D 0. Then X0 WD aff.dom f / is a linear subspace of X .
Take f0 WD f jX0 . Then icr.dom f / D core.dom f0 /, and so 0 2 core.dom f0 /.
By (iv) we have that f0 is c -continuous at 0. By [614, Theorem 2.4.12]
@f .x0 / D @f .0/ ¤ ;. t
u
Note that the subdifferential of a function on X when X is endowed with the core
topology is called (generally) the algebraic subdifferential.
Related to operations with the intrinsic core we mention the following result
which is well known in finite dimensional spaces.
Proposition 6.3.2. (i) Let A X be a convex set and T 2 L.X; Y /. If icr A is
nonempty then icr.T .A// D T .icr A/.
(ii) Let A X and B Y be nonempty sets. Then icr.A B/ D icr A icr B.
(iii) Let A; C X be convex sets such that icr A and icr C are nonempty. Then
icr.A C / D icr A icr C .
(iv) Let A; C X be convex sets such that core A ¤ ;. Then core.A C / D
core A C .
Proof. (i) Let first a 2 icr A and take y 2 T .A/; then y D Tx with x 2 A.
Because a 2 icr A, by (6.2) there exists
> 0 such that .1 C
/a
x 2 A. It
follows that .1 C
/Ta
y 2 T .A/. Since T .A/ is convex, using again (6.2),
we get Ta 2 icr.T .A//. Conversely, fix a 2 icr A and take y 2 icr.T .A//. By
(6.2) we get
> 0 such that y 0 WD .1 C
/y
Ta 2 T .A/, and so y 0 D Tx0
with x 0 2 A. Then y D Tx00 with x 00 WD 1C
1
x 0 C 1C
a. Since A is convex,
a 2 icr A, x 2 A and 1=.1 C
/ 2 0; 1Œ we have that x 00 2 icr A, and so
0
y 2 T .icr A/.
(ii) It is clear that aff.A B/ D aff A aff B. Doing a translation we may assume
that 0 2 A and 0 2 B. In this way, replacing if necessary X by aff A and Y
by aff B, the conclusion reduces to core.A B/ D core A core B, which is
immediate.
(iii) Consider T 2 L.X X; X / defined by T .x; x 0 / WD x x 0 . Applying (ii) to A
and C , then (i) to T and A C we get the conclusion.
(iv) Let a 2 core A and c 2 C . Consider x 2 X . Then there exists
> 0 such that
a C
x 2 A, and so .a c/ C
x 2 A C . Hence a c 2 core.A C /.
Conversely, fix a0 2 core A, c0 2 C , and take u 2 core.A C / D icr.A
C /. Then there exists
> 0 such that .1 C
/u
.a0 c0 / D a c for some
a 2 A and c 2 C . Because 1C

2 0; 1Œ, using (6.3), we get u D 1C
a0 C

1
1C
a 1C
c0 C 1C
c 2 core AC . Hence core.AC / D core AC . u

1
t
Note that assertion (iv) can be found in [562, Theorem 2.1], while assertion (iii)
from the preceding result can be found in [581]. Note also that it is possible to
have convex sets A; B X with icr A or icr B empty but icr.A B/ nonempty.
For this take A WD B WD `C 2 or A WD f.xn /n1 2 `2 j x1 D x2 g and B WD `2 .
C
C C C C C
Since `2 D `2 `2 we have that aff `2 D `2 , and so icr.`2 / D core.`2 /. It
is well known that core.`C C
2 / D ;, and so icr.`2 / D ;. In the second case A is a
6.3 The Core Convex Topology 263
linear subspace, and so icr A D A. In both cases we have that A B D `2 and so

icr.A B/ D `2 ¤ ;.
Another result related to the intrinsic core of convex sets is the following one; the
assertion (ii) for X and Z finite dimensional linear spaces is just [489, Theorem 6.8].
Proposition 6.3.3. Let A X Z be a convex set (set-valued map) and .x0 ; z0 / 2
X Z. Then:
(i) the following statements are equivalent: a) .x0 ; z0 / 2 core A, b) x0 2
core .PrX .A// and z0 2 core .A.x0 //, c) x0 2 core A1 .z0 / and z0 2
core .A.x0 //;
(ii) the following statements are equivalent: a) .x0 ; z0 / 2 icr A, b) x0 2 icr.PrX .A//
and z0 2 icr.A.x0 //.
Proof. First observe that, doing a translation, we may (and do) suppose that
.x0 ; z0 / D .0; 0/. Let us first prove (i).
a) ) c) Let x 2 X ; since .0; 0/ 2 core A, there exists

> 0 such that
.x; 0/ 2
A, whence
x 2 A1 .0/. Therefore 0 2 core A1 .0/ . Similarly, 0 2 core .A.0//.
c) ) b) is obvious (because A1 .0/ PrX .A/).
b) ) a) Let .x; z/ 2 X Z. Since 0 2 core .PrX .A//, there exists
> 0
such that
x 2 PrX .A/, and so there exists z0 2 Z such that .
x; z0 / 2 A. Since
0 2 core .A.0//, there exists > 0 such that .0; .
zz0 // 2 A. Since A is convex,
it follows that

1
.x; z/ D 0; .
z z0 / C
x; z0 2 A:
1C 1C 1C
Therefore .0; 0/ 2 core A.
(ii) a) ) b) Assume that .0; 0/ 2 icr A. Consider first x 2 PrX .A/. Then .x; z/ 2
A for some z 2 Z. It follows that .x 0 ; z0 / WD
.x; z/ 2 A for some
> 0,
and so x 0 D
x 2 PrX .A/. Hence 0 2 icr.PrX .A//. Let now z 2 A.0/, that is,
.0; z/ 2 A. As before,
.0; z/ 2 A for some
> 0, whence
z 2 A.0/. Hence
0 2 icr.A.0//.
b) ) a) Let .x; z/ 2 A. Then x 2 PrX .A/. Since 0 2 icr.PrX .A//, there exists
> 0 such that x 0 WD

x 2 PrX .A/ for some
> 0, and so .x 0 ; z0 / 2 A for some
z0 2 Z. Then

1 0
1
0; z C z D .x 0 ; z0 / C .x; z/ 2 A;
1C
1C
1C
1C
and so z00 WD 1C
1 0
z C 1C
z 2 A.0/. Because 0 2 icr .A.0//, there exists > 0

000 00
such that z WD z 2 A.0/. Taking WD =.1 C
C / 2 0; 1Œ we obtain that
x 0 D
x and
.1 / 0
.1 /

z0 C .1 /z000 D z0 z zD z D
z:
1C
1C
1C
C
Hence
.x; z/ D .x 0 ; z0 / C .1 /.0; z000 / 2 A, which proves that .0; 0/ 2
icr A. t
u
Note that the implications a) ) c) and c) ) b) in (i) are valid for arbitrary
sets A, but generally c) ) a) [and so b) ) a)] is not valid if A is not convex;
take for example A WD .Œ1; 1 f0g/ [ .f0g Œ1; 1/ R2 . Also note that
even for A convex the fact that 0 2 icr.A1 .0// and 0 2 icr.A.0// does not imply
that .0; 0/ 2 icr A; for this take A WD f.x; x/ j x 2 RC g R2 . Moreover, the
assertion (ii) of the preceding proposition cannot be obtained from (i) because, even
for .0; 0/ 2 A, span A is not the product of two linear spaces.
In [572, Lemma 2.1] and [573] the implication b) ) a) of (i) is given in a weaker
form: if core .PrX .A// ¤ ; and core .A.x// ¤ ; for every x 2 core .PrX .A// then
core A ¤ ;.
6.4 Yang’s Generalization of the Hahn–Banach Theorem
In this section the proper (that is, f0g ¤ K ¤ Z) convex cone K has nonempty
algebraic interior core K. Of course 0 … core K, K C core K D core K C core K D
core K and P core K D core K.
subspace and T0 2 L.X0 ; Z/. Suppose that 0 2 icr.dom X0 / and T0 x 6> z for
all .x; z/ 2 \ .X0 Z/. Then there exists T 2 L.X; Z/ such that T jX0 D T0
and Tx 6> z for all .x; z/ 2 .
Proof. We give a proof similar to that of Theorem 6.1.1. Consider X WD
aff .dom X0 /. Because 0 2 icr .dom X0 / dom X0 , X is a linear
space.
The case X D X ; hence 0 2 core .dom X0 /. Consider the set
˚
F WD .E; T / j X0 E D span E X; T 2 L.E; Z/;

T jX0 D T0 ; Tx 6> z 8.x; z/ 2 \ .E Z/ I
clearly, .X0 ; T0 / 2 F . For .E1 ; T1 /; .E2 ; T2 / 2 F we set .E1 ; T1 / .E2 ; T2 / if

E1 E2 and T2 jE1 D T1 ; is a partial order on F . Exactly as in the proof of
Theorem 6.1.1 we have that every chain C in F is bounded above. Using Zorn’s
lemma, .F ; / has maximal elements; let .E; T / 2 F be maximal. If E D X the
linear operator T verifies the conditions in the conclusion of the theorem.
Assume that E ¤ X , and take x 2 X n E. As in the proof of Theorem 6.1.1 we
obtain that the sets
B1 WD fz1 Tx1 j x1 2 E; z1 2 Z W .x1 C x; z1 / 2 PAg ;

B2 WD fTx2 z2 j x2 2 E; z2 2 Z W .x2 x; z2 / 2 PAg
6.4 Yang’s Generalization of the Hahn–Banach Theorem 265
are nonempty (and convex), where A WD epi . We claim that
.B1 C core K/ \ .B2 core K/ D ;:
In the contrary case there exist k1 ; k2 2 core K, x1 ; x2 2 E, z1 ; z2 2 Z such that

.x1 C x; z1 /; .x2 x; z2 / 2 PA and k1 C z1 Tx1 D Tx2 z2 k2 , whence
T .x1 C x2 / > z1 C z2 . But .x1 C x2 ; z1 C z2 / 2 PA and x1 C x2 2 E, contradicting
the hypothesis.
Assume that .B1 C core K/ [ .B2 core K/ D Z. Then B1 B1 C core K and
B2 B2 core K. Indeed, if b1 2 B1 n .B1 C core K/, then b1 2 B2 core K,
whence b1 D b2 k, with b2 2 B2 and k 2 core K. In this situation we get the
contradiction b1 C k=2 D b2 k=2 2 .B1 C core K/ \ .B2 core K/ D ;. Fix
now b1 2 B1 and b2 2 B2 and consider
WD supf 2 Œ0; 1 j 8 t 2 Œ0; W .1 t/b1 C tb2 2 B1 C core Kg:
Set b WD .1 /b1 C b2 and suppose first that b 2 B1 C core K; of course, in this

situation < 1. Then b D b 1 C k, with b 1 2 B1 and k 2 core K. Since k 2 core K,
there exists ı 2 P such that k C .b2 b1 / 2 core K for 2 Œı; ı. It follows that
8 2 Œı; ı W b C .b2 b1 / D b 1 C k C .b2 b1 / 2 B1 C core K:
Taking D minfı; 1 g, we get
8 t 2 Œ0; W Œ1 . C t/ b1 C . C t/b2 2 B1 C core K;
contradicting the choice of . Suppose now that b 2 B2 core K; of course > 0.

In this situation b D b 2 k, with b 2 2 B2 and k 2 core K. Since k 2 core K, there
exists ı 2 P such that k C .b2 b1 / 2 core K for 2 Œı; ı. It follows that
8 2 Œı; ı W b .b2 b1 / D b 2 k .b2 b1 / 2 B2 core K:
Taking D minfı; g, we get the contradiction
Œ1 . / b1 C . /b2 2 B1 C core K:

Therefore .B1 C
core K/ [ .B2 core K/ ¤ Z. Taking z 2 Z n .B1 C core K/ [
.B2 core K/ and E1 D E C Rx, then defining T1 by T1 .x C tx/ D Tx C tz,
T1 W E1 ! Z is a linear extension of T and .x; z/ 2 \ .E1 Z/ implies T1 x 6> z.
Therefore, .E1 ; T1 / 2 F and .E; T / .E1 ; T1 /. Because .E; T / is maximal in F
we obtain that .E; T / D .E1 ; T1 /, whence the contradiction E D E1 3 x. Hence
E D X.
When aff .dom X0 / ¤ X , proceeding as in the second case of the proof of
Theorem 6.1.1 we get the desired conclusion. t
u
Remark 6.4.2. In the proof of Theorem 6.4.1 we used only the fact that P epi
is convex (or, equivalently, RC epi is a convex cone) and not that epi itself is
convex.
The next result was obtained by Thierfelder in [573, Theorem 2.2].
Corollary 6.4.3. Let i W X Z, i 2 f1; 2g, be two convex set-valued maps
and z 2 Z 0 n f0g. Assume that 0 2 icr.dom 1 dom 2 / and inf z .1 .x//
sup z .2 .x// for all x 2 dom 1 \dom 2 . Then there exists T 2 L.X; Z/ such that
inf z .1 .x1 / Tx1 /
sup z .2 .x2 / Tx2 / for all x1 2 dom 1 and x2 2 dom 2 .
Proof. Consider K WD Kz WD fz 2 Z j z .z/
0g. Then core K D fz 2 Z j
z .z/ > 0g, and so 0 6> z if and only if z .z/
0. Consider also the set-valued map
W X Z with graph WD graph 1 graph 2 . Then is convex, and so is
K-convex, dom D dom 1 dom 2 and
.0/ D [x2X .1 .x/ 2 .x// D [x2dom 1 \dom 2 .1 .x/ 2 .x//:
From our hypothesis we have that 0 6> .0/. Taking X0 WD f0g and T0 WD
0 2 L.X0 ; Z/, the hypothesis of Theorem 6.4.1 holds. Therefore, there exists
T 2 L.X; Z/ such that Tx 6> .x/ for every x 2 X . This implies that for
all .x1 ; z1 / 2 graph 1 and .x2 ; z2 / 2 graph 2 we have that z .Tx1 Tx2 /
z .z1 / z .z2 /, or, equivalently, z .z1 Tx1 /
z .z2 Tx2 /. The conclusion
follows. t
u
Corollary 6.4.4. Let p W X ! Z be a sublinear operator, X0 be a linear subspace
of X and T0 2 L.X0 ; Z/. Suppose that T0 x 6> p.x/ for every x 2 X0 . If X0 Cdom p
is a linear subspace then there exists T 2 L.X; Z/ such that T jX0 D T0 and
Tx 6> p.x/ for every x 2 X .
Proof. In Theorem 6.4.1 let us take W X Z with graph D epi p; of course,
the hypotheses of the theorem hold, so that there exists T 2 L.X; Z/ such that
T jX0 D T0 and .x; z/ 2 implies Tx 6> z. Taking z D p.x/, the conclusion
follows. t
u
Before stating the next result recall that the weak subdifferential @w f .x0 / of the
proper operator f W X ! Z at x0 2 dom f is the set of those T 2 L.X; Z/ such
that
8 x 2 X W Tx Tx0 6> f .x/ f .x0 /:
Corollary 6.4.5. Let f W X ! Z be a proper convex operator and consider

x0 2 icr.dom f /. Then @w f .x0 / is nonempty.
Proof. Take g.x/ D f .x0 C x/ f .x/, graph D epi g, X0 D f0g and T0 WD
0 2 L.X0 ; Z/. Of course, if .x; z/ 2 \ .X0 Z/, then x D 0 and z 2 K,
whence T0 x 6> z. Using Theorem 6.4.1 we obtain T 2 L.X; Z/ such that Tx 6> z
for .x; z/ 2 . Taking x 2 dom f , we have that .x x0 ; f .x/ f .x0 // 2 , so

that T .x x0 / 6> f .x/ f .x0 /. u
t
Corollary 6.4.6. Let f; g W X ! Z be proper convex operators. Suppose that
0 2 icr.dom f dom g/ and that f .x/ C g.x/ 6< 0 for all x 2 dom f \ dom g.
Then there exists T 2 L.X; Z/ such that
8 x1 2 dom f; 8 x2 2 dom g W f .x1 / Tx1 6< g.x2 / Tx2 :
Proof. Consider W X Z with
graph WD f.x; z/ 2 X Z j f .x/ zg C f.x 0 ; z0 / 2 X Z j g.x 0 / z0 g:
Then is convex and dom D dom f dom g. Take X0 D f0g and T0 WD

0 2 L.X0 ; Z/. The hypotheses of the preceding theorem hold. Indeed, if .0; z/ 2
then .0; z/ D .x; z0 / C .x; z00 /, with f .x/ z0 and g.x/ z00 ; assuming
that 0 D T0 .0/ > z, from f .x/ C g.x/ z0 C z00 D z and the known relation
K C core K D core K we get the contradiction f .x/ C g.x/ < 0. Hence T0 .0/ 6> z.
Therefore, there exists T 2 L.X; Z/ such that Tx 6> z for .x; z/ 2 . In particular,
for x1 2 dom f and x2 2 dom g we get T .x1 x2 / 6> f .x1 / C g.x2 /, which yields
g.x2 / Tx2 6> f .x1 / Tx1 for all x1 2 dom f and x2 2 dom g. The proof is
complete. t
u
Note that applying Corollary 6.4.6 for f D p and g.x/ D T0 x for x 2 X0 ,
g.x/ D C1 for x … X0 , where p W X ! Z is a sublinear operator and T0 2
L.X0 ; Z/ we get a weaker variant of Corollary 6.4.4.
As mentioned in [603, Lemma 1], Wang [588] obtained Theorem 6.4.1 for X0 D
f0g, T0 D 0 and
a K-convex set-valued map with core.epi / ¤ ; and 0 2
core 1 .z0 / for some z0 2 Z (compare also with Proposition 6.3.3); based on
results in [588], Yang [603, Theorem 1] obtained a weaker form of Theorem 6.4.1:
the interiority hypothesis is stronger [more precisely, core.epi / ¤ ; and X0 \
core.dom / ¤ ; instead of 0 2 icr.dom X0 /] and the conclusion is weaker
[more precisely 0 6< .T T0 /.x/ 6< 0 for all x 2 X0 instead of T jX0 D T0 ].
As in the case Z D R, the extension theorems can be used for separating convex
sets in product spaces. For example, from Theorem 6.4.1 we can deduce the next
separation result.
Proposition 6.4.7. Let A; B X Z be such that RC .A B/ is convex. Assume
that 0 2 icr.PrX .A/ PrX .B// and A.x/ 6< B.x/ for all x 2 PrX .A/ \ PrX .B/.
Then there exists T 2 L.X; Z/ such that
8.x1 ; z1 / 2 A; 8.x2 ; z2 / 2 B W z1 Tx1 6< z2 Tx2 : (6.9)
Proof. Consider the set C WD A B X Z. Then PrX .C / D PrX .A/ PrX .B/
and if .0; z/ 2 C then z D z1 z2 with z1 2 A.x/, z2 2 B.x/ for some x 2
PrX .A/ \ PrX .B/; hence z 6< 0. Taking X0 WD f0g and T0 .0/ WD 0, we can apply
Theorem 6.4.1 (see also Remark 6.4.2) for graph WD C and T0 . Therefore, there
exists T 2 L.X; Z/ such that z 6< Tx for every .x; z/ 2 C . Hence (6.9) holds. t
u
Note that Thierfelder [572, 573] says that A and B are separable by an affine
mapping if there exist T 2 L.X; Z/ and z0 2 Z such that
8.x1 ; z1 / 2 A; 8.x2 ; z2 / 2 B W z1 Tx1 6< z0 6< z2 Tx2 (6.10)
[compare with (6.5)].

One can ask which is the relationship between (6.9) and (6.10). Setting A0 WD
fz1 Tx1 j .x1 ; z1 / 2 Ag, B0 WD fz2 Tx2 j .x2 ; z2 / 2 Bg, condition (6.9) becomes
B0 \ .A0 C core K/ D ;, or equivalently, .B0 core K/ \ .A0 C core K/ D ;,
while condition (6.10) becomes z0 … .B0 core K/ [ .A0 C core K/. So, it is quite
clear that (6.10) does not imply (6.9). (One can take X WD R, Z WD R2 , K WD R2C ,
A WD f.0; .1; 1//g, B WD f.0; .1; 1//g, T WD 0 and z0 WD .2; 2/.)
However, when A; B X Z are convex and core K ¤ ; (which is the case
in Proposition 6.4.7) we have that (6.9) implies (6.10). Indeed, taking A0 and B0
as above we have seen that condition (6.9) becomes B0 \ .A0 C core K/ D ;.
Since A0 and B0 are convex sets and A0 C core K is algebraically open, by the
classic (algebraic) separation theorem we get z 2 Z 0 , z ¤ 0, and 2 R such that
hb; z i ha C k; z i for all a 2 A0 , b 2 B0 , k 2 core K. It follows that
8 k 2 core K W hk; z i > 0; and

8.x1 ; z1 / 2 A; 8.x2 ; z2 / 2 B W hz1 Tx1 ; z i

hz2 Tx2 ; z i :
Taking z0 2 Z such that hz0 ; z i D we have that (6.10) holds.

The previous discussion shows that (6.9) is more adequate as separation property
than (6.10). Another argument is the fact that taking x1 D x2 in (6.9) we recover a
part of the hypothesis of Proposition 6.4.7, which is not the case for (6.10).
However, using directly the classic separation theorem we obtain a stronger
conclusion than that of Theorem 6.4.1 under the same hypotheses.
subspace and T0 2 L.X0 ; Z/. Suppose that 0 2 icr.dom X0 / and T0 x 6> z for
all .x; z/ 2 \ .X0 Z/. Then there exists z 2 Z 0 and T 2 L.X; Z/ such that
T jX0 D T; 8 z 2 core K W hz; z i > 0; and 8 .x; z/ 2 W hTx; z i hz; z i :
In particular Tx 6> z for very .x; z/ 2 . Moreover, T can be defined by Tx D

T0 x0 hx1 ; x i z, where z 2 Z, x 2 X 0 and the linear subspace X1 X with
X D X0 ˚ X1 are fixed, and x D x0 C x1 with x0 2 X0 , x1 2 X1 .
Proof. Assume first that 0 2 core .dom X0 /. Let B D epi graph T0 . It is
obvious that PrX .B/ D dom X0 , and so 0 2 core .PrX .B//. Then B.0/ ¤ ; and
z 2 B.0/ ) 0 6> z. Indeed, if z 2 B.0/ then .0; z/ D .x; z0 / .x0 ; T0 x0 / C .0; k 0 /
for some .x; z0 / 2 , x0 2 X0 and k 0 2 K. It follows that x D x0 and z D

z0 C k 0 T0 x0 6< 0. Since core K ¤ ;, it is clear that core .B.0// ¤ ;. Therefore,
by Proposition 6.3.3 (i), core B ¤ ;. Moreover, .0; 0/ … core B. In the contrary
case, again by Proposition 6.3.3 (i), 0 2 core.B.0//, contradicting the fact observed
above that z 6< 0 for every z 2 B.0/.
Using an algebraic separation theorem we get .x ; z / 2 X 0 Z 0 n f.0; 0/g such
that
8 .x; z/ 2 ; 8x0 2 X0 ; 8k 2 K W hx x0 ; x i C hz T0 x0 C k; z i
0:
(6.11)
First notice that, because 0 2 core .dom X0 /, if z D 0 then x D 0, a

contradiction; therefore, z ¤ 0. Next, from (6.11), it follows that hk; z i
0
for every k 2 K,
8x0 2 X0 W hx0 ; x i C hT0 x0 ; z i D 0; (6.12)

8.x; z/ 2 W hx; x i C hz; z i
0: (6.13)
Of course, hk; z i > 0 for every k 2 core K. Fix z 2 Z such that hz; z i D 1,
X1 X a linear subspace such that X0 ˚ X1 D X and T W X ! Z defined by
Tx WD T0 x0 hx1 ; x i z for x D x0 Cx1 with x0 2 X0 and x1 2 X1 . Then, of course,
T jX0 D T0 and for .x; z/ 2 , taking x0 2 X0 and x1 2 X1 with x D x0 C x1 , we
have
hTx; z i D hT0 x0 hx1 ; x i z; z i D hT0 x0 ; z i hx1 ; x i D hx; x i hz; z i :
If X WD aff.dom X0 / D span.dom X0 / ¤ X we find first x 2

L.X ; R/ and z 2 Z 0 as above and then extend x to an element of X 0 . The proof is
complete. t
u
Notice that using the same hypotheses as in Theorem 6.4.1, the conclusion of
Theorem 6.4.8 is stronger, not only because hTx; z i hz; z i implies Tx 6> z,
but also because we have a very special expression for T . This is quite surprising
because in the proof of Theorem 6.4.8 we utilized the usual separation theorem from
the scalar case which is equivalent to the classical Hahn–Banach theorem, while in
the proof of Theorem 6.4.1 we utilized a similar technique to that used in the proof
of Theorem 6.1.1, that is, for the proof of the Hahn–Banach–Kantorovich theorem.
Furthermore, the technique utilized in the proof of Theorem 6.4.8 is more adequate
for obtaining continuous versions of this theorem.
A continuous version of Theorem 6.4.8 is the next result.
Theorem 6.4.9. Let X; Z be separated locally convex spaces, W X Z be a
K-convex set-valued map, X0 X a linear subspace and T0 2 L .X0 ; Z/. Suppose
that int.epi / ¤ ;, X0 \ int.dom / ¤ ;, and T0 x 6> z for all .x; z/˝ 2 ˛ \
.X0 Z/. If either (a) X0 has a topological supplement, or (b) T0 x D x; x0 z0
for every x 2 X0 with fixed x0 2 X and z0 2 Z, then there exists z 2 Z and
T 2 L .X; Z/ such that
T jX0 D T; 8 z 2 core K W hz; z i > 0; and 8 .x; z/ 2 W hTx; z i hz; z i :
In particular Tx 6> z for very .x; z/ 2 . Moreover, in case (a) T can be defined
by Tx D T0 x0 hx1 ; x i z, where z 2 Z, x 2 X are fixed and x ˝ D x0˛ C x1
with x0 2 X0 , x1 2 X1 , while in case (b) T can be defined by Tx WD x; x0 .z0
hz0 ; z i z1 / hx; x i z1 for x 2 X with fixed x 2 X and z1 2 Z such that
hz1 ; z i D 1.
Proof. In the proof of Theorem 6.4.8, because int.epi / ¤ ;, we have that int B
.D core B/ is nonempty; it follows that x 2 X and z 2 Z . In case (a), because
the projections on X0 and X1 are continuous, taking into account the construction
of T we obtain that T is continuous.
˝ ˛
Assume now that T0 x D x; x0 z0˝ for every x 2 X0 ˛with fixed x0 2 X and
z0 2 Z. From (6.12) we obtain that x; x C ˝hz0 ; z˛i x0 D 0 for every x 2 X0 .
Taking z1 2 Z with hz1 ; z i D 1 and Tx WD x; x0 .z0 hz0 ; z i z1 / hx; x i z1
for x 2 X , we have that T is continuous and Tx D T0 x for x 2 X0 . Moreover, for
.x; z/ 2 we have
hTx; z i D hhx; x0 i z0 .hz0 ; z i hx; x0 i hx; x i/z1 ; z i

D hx; x0 i hz0 ; z i hz0 ; z i hx; x0 i hx; x i
D hx; x i hz; z i ;
the last inequality being obtained using (6.13). t

u
Thierfelder in [572] and [573, Theorem 2.5] obtains similar conclusions for
Theorem 6.4.8 in the following two situations: 1) there exists z strictly positive on
K n f0g such that hz; z i
hT0 x0 ; z i for .x; z/ 2 \ .X0 Z/ and 2) there exists
an algebraically open convex cone P such that K n f0g P and .0/ C P .0/.
Taking P D fz 2 Z j hz; z i > 0g in case 1), in each situation we have that
.0/ \ .P / D ;, and so the conclusion is obtained applying Theorem 6.4.8 for K
replaced by P [ f0g.
Remark 6.4.10. Not that the condition X0 \ int.dom / ¤ ; in Theorem 6.4.9
p To show this take X WD Z WD R, WD f.x; z/ 2 X Z j x
is essential.
0; z
xg, K WD RC , X0 WD f0g, T0 WD 0; in this case all the hypotheses
of Theorem 6.4.9, but X0 \ int.dom / ¤ ;, are verified (compare with [419,
Theorem 2.1]).
Related to the non emptiness of the interior of epi see Proposition 3.1.27 and
Corollary 3.2.2.
6.5 A Sufficient Condition for the Convexity of RC A 271
6.5 A Sufficient Condition for the Convexity of RC A
We have seen that the convexity of epi in Theorems 6.1.1 and 6.4.1 can be
replaced by the convexity of RC epi . In this section we consider A X Z,
where X; Z are real vector spaces. Let us denote by NZ0 a class of nonempty subsets
of Z with the property that for every W 2 NZ0 one has Œ0; 1W D W and there
exists V 2 NZ0 such that V C V W . Of course, NZ0 could contain only the set
f0g, or could be the class of balanced neighborhoods of 0 with respect to a topology
such that .Z; / is a topological vector space.
Theorem 6.5.1. Let A X Z be a nonempty set. Suppose that the following
conditions hold:
(i) 9˛; ˇ > 0; 8u; v 2 A W ˛u C ˇv 2 A,
(ii) 8u 2 A, 8V 2 NZ0 , 9ı 2 0; 1Œ, 8
2 Œı; 1 W
u 2 A C f0g V .
Then
AQ WD \V 2NZ0 .RC A C f0g V / (6.14)
is a convex cone.
Proof. Take ˛ 0 WD ˛.˛ C ˇ/1 and ˇ 0 WD ˇ.˛ C ˇ/1 and define inductively the
sequence of subsets In by
˚
I0 WD f0; 1g; InC1 WD ˛ 0 t C ˇ 0 s j t; s 2 In for n 2 N:
We prove that AQ is convex in several steps.

Step 1: In InC1 for n 2 N and cl I D Œ0; 1, where I WD [n2N In . This is
(almost) obvious.
Step 2: t 2 In , u; v 2 A ) tu C .1 t/v 2 .˛ C ˇ/n A. We prove this claim by
induction. For n D 0 the claim is obvious. Suppose that it holds for n 2 N and
take t 2 InC1 and u; v 2 A. Then t D ˛ 0 t1 C ˇ 0 t2 with t1 ; t2 2 In . So
tu C .1 t/v D ˛ 0 Œt1 u C .1 t1 /v C ˇ 0 Œt2 u C .1 t2 /v :
By our assumption, ti u C .1 ti /v D .˛ C ˇ/n ui with ui 2 A for i 2 f1; 2g,

and so

tu C .1 t/v D .˛ C ˇ/n ˛ 0 u1 C ˇ 0 u2
D .˛ C ˇ/n1 .˛u1 C ˇu2 / 2 .˛ C ˇ/n1 A:
Step 3: t 2 I , u; v 2 A ) tu C .1 t/v 2 RC A. This is obvious by Step 1 and

Step 2.
Step 4: t 2 Œ0; 1, u; v 2 A ) tu C .1 t/v 2 A. Q For t 2 f0; 1g the assertion

is obvious because A A. Q Fix t 2 0; 1Œ and set w WD .1 t/u C tv. For

2 12 ; 1Œ consider t
WD 1
t 2 0; 1Œ and u
WD
t
1 .1 t/u. We have that

w D t
u
C .1 t
/v. Consider V 2 NZ0 and take W 2 NZ0 with W C W V .
As u 2 A, by (ii), there exists ı 2 12 ; 1Œ such that u
2 A C f0g W for every

2 Œı; 1. Since cl I D Œ0; 1, we have that I \ 1 t; 1 ıtŒ ¤ ;, and so there
exists
2 ı; 1Œ such that t
2 I \ 1 t; 1 ıtŒ. Hence u
D uQ C .0; z/ with
uQ 2 A and z 2 W . Using Step 3, we obtain that

w D t
u
C .1 t
/v D t
uQ C .1 t
/v C t
.0; z/;
and so w 2 RC A C f0g 2W RC A C f0g V for every V 2 NZ0 . Hence

w 2 A.Q
Step 5: AQ is a convex cone. Let u 2 AQ and t > 0. For V 2 NZ0 there exists
W 2 NZ0 such that tW V . By the definition of A,Q there exist u0 2 RC A and
0 0
z 2 W such that u D u C .0; z/, and so tu D tu C .0; tz/ 2 RC A C f0g V . It
Q
follows that tu 2 A.
Let now u; v 2 AQ and consider V 2 NZ0 ; take W 2 NZ0 such that W CW CW
V . Since u; v 2 A, Q there exist u0 ; v0 2 A, s; t
0 and z; w 2 W such that u D
0 0
su C.0; z/, v D tv C.0; w/. If sCt D 0 then uCv D 0C.0; zCw/ 2 RC ACf0gCV .
If s C t > 0 then

u C v D .s C t/ s.s C t/1 u0 C t.s C t/1 v0 C .0; z C w/:

Q
Using Step 4 and the first part uQ WD .s C t/ s.s C t/1 u0 C t.s C t/1 v0 2 A.
Therefore, there exists u 2 RC A and wQ 2 W such that uQ D u C .0; w/.
Q It follows
Q 2 RC A C f0g V . Since V 2 NZ0 is arbitrary we
that u C v D u C .0; z C w C w/
Q Therefore, AQ is a convex cone.
have that u C v 2 A. t
u
Note that in the case f0g 2 NZ0 , if A X Z verifies conditions (i) and (ii) in
the preceding theorem then RC A is convex. Observe that A satisfies condition (ii)
if A is open (X; Z being t.v.s.) or 0; 1A D A.
Corollary 6.5.2. Let Z be a topological vector space and let K Z be a proper
convex cone. If A X Z is a nonempty set verifying conditions (i) and (ii) in
Theorem 6.5.1 for NZ0 the class of balanced neighborhoods of 0 2 Z.
(j) Assume that int K ¤ ;. Then
A \ Œf0g . int K/ D ; ) conv A \ Œf0g . int K/ D ;:
(jj) Assume that K is closed. Then
A.0/ K ) .conv A/.0/ K:

6.5 A Sufficient Condition for the Convexity of RC A 273
Q
Proof. By Theorem 6.5.1 we have that conv A A.
(j) Assume that int K ¤ ; and A\Œf0g. int K/ D ;. It is sufficient to show that
AQ \ Œf0g . int K/ D ;. Suppose that .0; k/ 2 AQ \ Œf0g . int K/. Since
k 2 int K, there exists V 2 NZ0 such that k C V int K. On the other hand,
from .0; k/ 2 AQ we get u 2 A, t
0 and z 2 V such that .0; k/ D tuC.0; z/,
whence tu D .0; k z/ 2 f0g . int K/. Since K is proper, we have that
t > 0, whence the contradiction u 2 A \ Œf0g . int K/ D ;.
(jj) Assume that K is closed and A.0/ K. It is sufficient to show that A.0/ Q K.
Q
Assume that z 2 A.0/ n K. Because K is closed, there exists V 2 NZ0 such
that W V and z … W . Then there exist u 2 A, t
0 and z0 2 W such that
.0; z/ D tu C .0; z0 /. Hence tu D .0; z z0 /. Because z … W we have that t > 0.
It follows that t 1 .zz0 / 2 A.0/, and so zz0 2 K. Hence z 2 K CW K CV .
Since V is arbitrary we obtain the contradiction z 2 K. Hence A.0/Q K. u t
Theorem 6.5.1 and its corollary can be found in [613].
Chapter 7
Conjugates and Subdifferentials
To each type of efficiency for optimization problems it is possible to associate

notions of conjugate and subdifferential for vector valued functions or set-valued
maps. In this chapter we study the conjugate and the subdifferential corresponding
to the strong efficiency as well as the subdifferentials corresponding to the weak and
Henig type efficiencies. For the strong conjugate and subdifferential we establish
similar results to those in the convex scalar case, while for the other types of
subdifferential we establish formulas for the subdifferentials of the sum and the
composition of functions and set-valued maps.
7.1 The Strong Conjugate and Subdifferential
Let us begin with the strong case. Recall that having A Y , Y being ordered by
the proper pointed convex cone C , y 2 A is strongly efficient if A y C C ,
that is y y for all y 2 A. Correspondingly, having F W X Y and .x; y/ 2
graph F , T 2 L .X; Y / is a sub-gradient of F at .x; y/ if y T x y Tx for
all .x; y/ 2 graph F . Assuming that .Y; C / has the l.u.b.p. (see Sect. 6.1), if T is
a sub-gradient of F at .x; y/ then 'F W X ! Y defined by 'F .x/ WD inf F .x/ if
x 2 dom F and 'F .x/ WD C1 elsewhere, is well defined, 'F is a proper operator
with dom 'F D dom F and
8x 2 X W Tx T x 'F .x/ 'F .x/I
moreover, 'F .x/ D y. Conversely, if T 2 L .X; Y / verifies the relation above, then
T is a sub-gradient of F at .x; y/. So, in the case of strong efficiency, provided the
l.u.b.p. holds, it is sufficient to study the subdifferential for vector functions.
In the sequel X; Y; Z are topological vector spaces, Y being ordered by the
proper pointed convex cone C . Consider f W X ! Y a proper function.

DOI 10.1007/978-3-642-54265-7__7
276 7 Conjugates and Subdifferentials
The (strong) subdifferential of f at x 2 dom F is the set
@s f .x/ WD fT 2 L .X; Y / j 8x 2 X W Tx T x f .x/ f .x/g I
if x 2 X n dom f we set @s f .x/ WD ;. An element T 2 @s f .x/ is called (strong)

sub-gradient of f at x. Similar to the "-subdifferential in the scalar case, for & 2 Y ,
let us consider the set
@s& f .x/ WD fT 2 L .X; Y / j 8x 2 X W Tx T x f .x/ f .x/ C &g
for x 2 dom f and @s& f .x/ WD ; otherwise. Clearly, @s0 f .x/ D @s f .x/, and & 2 C
if @s& f .x/ ¤ ;; moreover, @s&1 f .x/ @s&2 f .x/ if &1 &2 .
Before continuing the study of the subdifferential of convex operators let us
mention an auxiliary result whose proof is practically the same as in the scalar case
(see [614]). Similar to the scalar case, we denote by .X; Y /, or .X; Y; C / if
we wish to emphasize the ordering cone C , the class of proper convex operators
f W X ! Y .
Remark 7.1.1. Note that the convex function f W X ! Y is proper if f .x0 / 2 Y
for some x0 2 icr.dom f / provided that C has the l.u.b.p. (see also Proposi-
tion 6.1.12).
Indeed, take x 2 dom f . Then there exists > 0 such that x1 WD .1 C /x0

x 2 dom f ; hence x0 D 1C 1
x1 C 1C x, and so f .x0 / 1C
1
f .x1 / C 1C f .x/.
It follows that f .x/ 2 Y .
Lemma 7.1.2. Let f 2 .X; Y /, x 2 dom f , u 2 X and Iu WD ft 2 R j x C tu 2
dom f g, IuC WD Iu \ P.
(i) The mapping 'u W Iu n f0g ! Y , defined by 'u .t/ WD t 1 Œf .x C tu/ f .x/
is increasing.
(ii) Assume that C is pointed, .Y; C / has the l.u.b.p. and x 2 icr.dom f /. Then for
every & 2 C one has that
f .x C tu/ f .x/ C &

8u 2 X W f&0 .x; u/ WD inf 2 Y [ fC1g;
t 2IuC t
with dom f&0 .x; / D RC .dom f x/, and f&0 .x; / is sublinear; in particular,
f&0 .x; u/ 2 Y for every u 2 X if x 2 core.dom f /. Moreover, T 2 @s& f .x/ if
and only if Tu f&0 .x; u/ for every u 2 X .
(iii) In the case & D 0, the conclusion in (ii) holds if C is pointed and any
decreasing sequence in C has infimum.
Proof. The proofs for (i) and (ii) are similar to the scalar case (see [614, Theo-
rem 2.1.13]). In the case (iii) let u 2 P.dom f x/. Since x 2 icr.dom f /,
0 2 int Iu . Take t0 2 Iu \ .1; 0/ and .tn /n1 Iu a decreasing sequence
with limit 0. Then y0 WD 'u .t0 / yn WD 'u .tn / for every n
1 and .yn /n1
7.1 The Strong Conjugate and Subdifferential 277
is decreasing. It follows that y WD infn1 yn exists. Moreover, using again (i) we

obtain that y D inft 2IuC 'u .t/. u
t
Throughout this chapter we use the notation f 0 .x; u/ WD f00 .x; u/.
The result in Lemma 7.1.2 is practically stated in [580, Propositions 3, 4].
Concerning the subdifferentiability of f at x 2 dom f we have the following
result.
Theorem 7.1.3. Assume that f 2 .X; Y / is continuous at x 2 dom f , C is a
pointed normal convex cone and & 2 C . Then @s& f .x/ is an equi-continuous subset
of L .X; Y /. Moreover, if .Y; C / has the l.u.b.p., then @s& f .x/ is nonempty and
˚
8u 2 X W f&0 .x; u/ D max Tu j T 2 @s& f .x/ : (7.1)
Proof. Without loss of generality we may (and do) assume that x D 0 and
f .0/ D 0.
Let V 2 NY be such that V D ŒV C . Then there exists V1 2 NY such that
V1 C V1 V . Since V1 is absorbing, there exists t 2 0; 1Œ such that t& 2 V1 . Since
f is continuous at 0 and f .0/ D 0, there exists U1 2 NX such that f .U1 / V1 .
Then for T 2 @s& f .x/ and x 2 U1 one has Tx 2 f .x/ C & C V1 C & C ,
whence T .tx/ 2 tV1 C t& C V1 C V1 C V C . Therefore, Tu 2 V C
for every u 2 U WD tU 1 2 NX . Because T is linear and U; V are balanced, we get
T .U / .V C / \ .V C C / D V . Therefore, @s& f .x/ is equi-continuous.
Assume now that .Y; C / has the l.u.b.p. Take u0 2 X and set X0 WD Ru0 .
Consider T0 W X0 ! Y , T0 .tu0 / WD tf 0& .0; u0 /. Then T0 is linear and T0 u f&0 .0; u/
for every u 2 X0 . Applying Theorem 6.1.1 for T0 and f&0 .0; / we get T 2 L .X; Y /
such that T jX0 D T0 and Tu f&0 .0; u/ for every u 2 X . Hence Tx f .x/ C &
for every x 2 X . Since f is continuous at 0, using Proposition 3.2.11, relation
(3.11) holds. It follows that (3.11) holds with T instead of f . Applying again
Proposition 3.2.11 we obtain that T is continuous, and so T 2 @s& f .x/. Therefore,
@s& f .x/ ¤ ;. Since the inequality Tu f&0 .x; u/ for u 2 X holds for every
T 2 @s& f .x/, (7.1) holds. t
u
Theorem 7.1.3 extends slightly [580, Theorem 6]. Note that the proof of
Theorem 7.1.3 shows that when C is normal and f is continuous at x 2 dom f ,
any T 2 L.X; Y / such that Tx T x f .x/ f .x/ C & for x 2 X is continuous
(that is any algebraic &-sub-gradient is a topological &-sub-gradient).
A variant of the preceding result is valid in the case in which f jaff.dom f / is
continuous at x 2 dom f . For its proof we need the following result.
Proposition 7.1.4. Assume that X is a locally convex space and C is a pointed
normal convex cone with nonempty interior such that .Y; C / has the l.u.b.p. If X0
X is a linear subspace of X and T0 2 L .X0 ; Y /, then there exists T 2 L .X; Y /
such that T jX0 D T0 .
Proof. Since T0 2 L .X0 ; Y /, T0 is convex and continuous at 0. According to
Proposition 3.2.11 (ii), there exist y 2 Y and p W X ! RC a continuous seminorm
such that [x 2 X0 , p.x/ 1] imply T0 x y. Therefore y

0 and T0 x p.x/y
for x 2 X0 . Taking p W X ! Y , p.x/ D p.x/y for x 2 X , p is a (continuous)
sublinear operator; therefore, applying Theorem 6.1.1 we get T 2 L .X; Y / such
that T jX0 D T0 and Tx p.x/ for all x 2 X . Using again Proposition 3.2.11 (ii),
we have T 2 L .X; Y /. The proof is complete. t
u
Corollary 7.1.5. Assume that X is a locally convex space, f 2 .X; Y / is such
that f jaff.dom f / is continuous at x 2 dom f , and C is a pointed normal convex cone
such that .Y; C / has the l.u.b.p. Then @s& f .x/ ¤ ; for every & 2 C . Moreover,
˚
8u 2 RC .dom f x/; 8& 2 C W f&0 .x; u/ D max Tu j T 2 @s& f .x/ :
Proof. Without loss of generality we may (and do) assume that x D 0 and
f .0/ D 0. Take X0 WD span.dom f / D aff.dom f / ¤ X and consider f0 WD f jX0 .
Applying Theorem 7.1.3, there exists T0 2 @s& f0 .0/. Applying now Proposition 7.1.4
we get T 2 L .X; Y / such that T jX0 D T0 . Since dom f X0 it is clear that
T 2 @s& f .0/. t
u
Consider f W X ! Y a proper function, x 2 dom f and 2 RC . We say that
f is -regular at x if
S˚
8y 2 C C n f0g W @ .y ı f /.x/ D y ı @s& f .x/ j & 2 C; h&; y i D ;
(7.2)
where, as usual, .y ıf /.x/ WD hf .x/; y i for x 2 dom f and .y ıf /.x/ WD C1
otherwise.
Note that the inclusion always holds in (7.2).
We say that f is regular at x 2 dom f if f is 0-regular at x; hence 0-regularity
reduces to the notion introduced by Valadier [580] when x 2 core.dom f /
Theorem 7.1.6. Assume that f 2 .X; Y / is continuous at x 2 dom f and C is
a pointed normal convex cone such that .Y; C / has the l.u.b.p.
(i) If Y is a separated locally convex space such that the intervals Œy; y 0 with
y; y 0 2 Y are relatively weakly compact, then @s& f .x/ is relatively compact in
Lpw .X; Y / for every & 2 C .
(ii) Moreover, assume that C is closed. Then @s& f .x/ is compact in Lpw .X; Y /.
(iii) Besides the hypotheses of (i) and (ii), assume that Y is weakly sequentially
complete; then f is regular at x.
In the result above Y is the space Y endowed with the weak topology .Y; Y /,
and Lpw .X; Y / is the space L .X; Y / endowed with the topology of point-wise
convergence.
Proof. (i) Because f is continuous at x, we have that x 2 int.dom f /. By
Theorem 7.1.3, @s& f .x/ . L .X; Y // is nonempty and equicontinuous. Let
U0 2 NX be such that x C U0 dom f and let u 2 U0 . Then
8T 2 @s& f .x/ W Tu 2 Œf .x/ f .x u/ &; f .x C u/ f .x/ C &;

and so fTu j T 2 @s& f .x/g is relatively weakly compact for every u 2 U0 .

Because U0 is absorbent we obtain that fTu j T 2 @s& f .x/g is relatively weakly
compact for every u 2 X . The conclusion follows using [76, Proposition
III.3.4].
(ii) By Lemma 7.1.2, we have that
˚
@s& f .x/ D \u2X T 2 L.X; Y / j Tu f&0 .x; u/
˚
D \u2X T 2 L.X; Y / j Tu 2 f&0 .x; u/ C :
Since the mapping Lpw .X; Y / 3 T 7! Tu 2 Y is continuous for any

u 2 X and C is (weakly) closed, @s& f .x/ is a closed subset of Lpw .X; Y / (as
intersection of a family of closed sets). Therefore, using (i), @s& f .x/ is compact
in Lpw .X; Y /.
(iii) Let us fix y 2 C C and D 0, and prove the inclusion in (7.2). Since
the mapping Lpw .X; Y / 3 T 7! y ı T 2 .X ; w / is continuous, where
w D .X ; X /, y ı @s f .x/ is w -compact. Hence y ı @s f .x/ X is
nonempty, convex and w -compact (hence w -closed).
Set g WD y ı f . Since g is convex and continuous at x, using Theorem 7.1.3
for Y replaced by R (or [614, Theorem 2.4.98]), we have that
8u 2 X W g 0 .x; u/ D g00 .x; u/ D max fhu; x i j x 2 @g.x/g : (7.3)
Take x 2 X n y ı @s f .x/. Since, as seen above, y ı @s f .x/ is nonempty,

convex and w -closed, there exists u 2 X such that
hu; x i > sup fhu; x i j x 2 y ı @s f .x/g D sup fhT u; y i j T 2 @s f .x/g :

(7.4)
By Theorem 7.1.3, there exists T0 2 @s f .x/ such that T0 u D f 0 .x; u/. We

have

˝ 0
˛ f .x C tn u/ f .x/
hu; y ı T0 i D hT0 u; y i D f .x; u/; y D inf ;y ;
n1 tn
where .tn /n1 P is a decreasing sequence with limit 0 such that x C

tn u 2 dom f for every n
1. By Lemma 7.1.7 (below), the sequence
tn1 .f .x C tn u/ f .x// n1 weakly converges to its infimum. So, because
y is continuous and positive,

f .x C tn u/ f .x/
hT0 u; y i D inf ; y D g 0 .x; u/;
n1 tn
and so, by (7.4), hu; x i > g 0 .x; u/. Taking into account (7.3), we obtain
that x … @g.x/. Therefore, y ı @s f .x/ @.y ı f /.x/. The proof is
complete. t
u
Lemma 7.1.7. Assume that Y is a separated locally convex weakly sequentially

complete space and C Y be normal and closed. If .yn /n1 Y is a
decreasing sequence which is bounded from below, then .yn / converges weakly to
inffyn j n
1g. Therefore, Y is weakly sequentially Daniell.
Proof. Because C is normal, we have that C C is reproducing, that is
C C C C D Y .
For y 2 C C we have that the sequence .hyn ; y i/n1 R is decreasing
and bounded from below, and so .hyn ; y i/ is convergent (in R). Because C C
is reproducing, we obtain that .hyn ; y i/ is convergent for every y 2 Y . In
particular, the sequence .yn /n1 is a weakly Cauchy sequence. Because Y is weakly
sequentially complete, there exists y 2 Y such that .yn /n1 weakly converges to y.
Since hyn ; y i & hy; y i for y 2 C C , we obtain that hyn y; y i
0 for all
y 2 C C and all n
1, whence yn y 2 C CC D cl C D C for every n
1.
Hence y is a lower bound for .yn /n1 . If y 0 2 Y is another lower bound for .yn /n1 ,
then hyn y 0 ; y i
0 for all n
1 and y 2 C C , whence, taking the limit we
get hy y 0 ; y i
0 for every y 2 C C . Hence y y 0 2 C , which shows that
y D inffyn j n
1g. The proof is complete. t
u
Observe that assuming that Y is weakly complete instead of being weakly
sequentially complete in Lemma 7.1.7 we obtain that Y is weakly Daniell.
Throughout the rest of this section we assume that C is a pointed proper normal
convex cone such that .Y; C / has the l.u.b.p.
Definition 7.1.8. Let f W X ! Y . The (strong) conjugate of f is
fs W L .X; Y / ! Y ; fs .T / WD supfTx f .x/ j x 2 X g:
Of course, as in the scalar case, if f .x/ D C1 for every x 2 X then

fs .T / D 1 for every T 2 L .X; Y /, that is fs D 1; if there exists x 2 X
such that f .x/ D 1 then fs D C1. (As in the scalar case, it is possible to have
fs D C1 even if f is C -convex and f .x/ 2 Y for every x 2 X:/ So, it is natural
to consider f a proper operator.
Setting .1/ C .C1/ WD .C1/ C .1/ WD C1, the Fenchel–Young
inequality
8x 2 X; 8T 2 L .X; Y / W f .x/ C fs .T /
Tx;
holds.
Proposition 7.1.9. Let f W X ! Y be proper. Then
(i) fs is a convex operator.
(ii) Let & 2 C and x 2 dom f ; then
T 2@s& f .x/,fs .T /Cf .x/ T xC&; T 2@f .x/ , fs .T /Cf .x/ D T x:
(iii) If X is ordered by a convex cone P and f is .P; C /-increasing then dom fs
LC .X; Y / WD fT 2 L .X; Y / j T .P / C g.
Proof. The proofs for (i) and (ii) are immediate. For (iii) we proceed in the following
way. Fix z0 2 dom f and take T 2 dom fs . Then for z 2 P we have that T .z0 z/
f .z0 z/ C fs .T / f .z0 / C fs .T /, whence Y 3 y0 WD Tz0 f .z0 / fs .T /
Tz DW y. Replacing z by ˛z with ˛ 2 RC , we obtain that y0 ˛y for all ˛ 2 RC .
For ˛ D 0 we get y0 0. Taking ˛ WD ˇ 1 with ˇ 2 P, we get ˇy0 y for every
ˇ 2 P. Clearly, P I WD fˇ 2 R j ˇy0 yg and I is an interval. If 0 2 I then
y 2 C . Else 0 D inf I (and y0 ¤ 0). Take y the least upper bound of fˇy0 j ˇ 2 Pg;
hence y y. Since 0
ˇy0 for ˇ 2 P, we have that y 0. From ˇy0 y for
every ˇ 2 P, we get ˇy0 2y for ˇ 2 P, and so y 2y, that is 0 y. It follows
that y D 0, and so y 2 C . Therefore, T .P / C . t
u
The proof of (iii) uses an idea from the proof of [533, Theorem 1].
Let now Z be another topological vector space and ˚ W X Z ! Y be an
operator. Consider the problem
min ˚.x; 0/ s.t. x 2 X; (P )
called the primal problem. The perturbed problems for (P ) are
min ˚.x; z/ s.t. x 2 X; (Pz )
for z 2 Z; of course, the initial problem (P ) coincides with .P0 /.

To the operator ˚ (or problems (Pz )), we associate the marginal (or value)
function
h W Z ! Y ; h.z/ WD inf f˚.x; z/ j x 2 X g I (7.5)
hence, the value vP of problem (P ) is h.0/. Note that
dom h D PrZ .dom ˚/ D PrZ .epi ˚/;
and so 0 2 dom h iff 0 2 PrZ .epi ˚/ iff there exists x 2 X such that
˚.x; 0/ < C1.
It is worth observing that h is convex when ˚ is convex. Furthermore, for every
T 2 L .Z; Y / one has
hs .T / D sup fTz h.z/ j z 2 Zg D sup fTz infx2X ˚.x; z/ j z 2 Zg

D sup fTz ˚.x; z/ j x 2 X; z 2 Zg D ˚s .0; T /:
Because h.0/
hs .T / D ˚s .0; T / for every T 2 L .Z; Y /, it is natural to
consider as dual problem of (P) the problem
maxŒ˚s .0; T / s:t: T 2 L .Z; Y /: (D)

It follows that vP
vD , that is weak duality holds for problems (P ) and (D).
Of course, vP D vD automatically when vP D 1 or vD D C1. A very important
situation is when vP D vD 2 Y and (D) has optimal solutions, that is the problem
(P ) is stable. Taking into account Proposition 7.1.9, one has that (P ) is stable if and
only if vP 2 Y and @s h.0/ ¤ ;; moreover, in this case, @s h.0/ is the set of solutions
of (D).
In the following result we provide quite general conditions in order to have
vP D vD and (D) have optimal solutions.
Theorem 7.1.10. Let ˚ 2 .X Z; Y / with 0 2 PrZ .dom ˚/. Assume that one of
the following conditions holds:
(i) there exists x0 2 X such that ˚.x0 ; / is finite and continuous at 0I
(ii) X is a locally convex space and there exists x0 2 X such that ˚.x0 ; /jZ0 is
finite and continuous at 0, where Z0 WD span.PrZ .dom ˚//I
(iii) X; Y are Fréchet spaces, Z0 is a barreled space (verified if Z is a Fréchet
space and Z0 is closed), epi ˚ is closed and 0 2 icr.PrZ .dom ˚//.
Then

inf ˚.x; 0/ D max ˚s .0; T / : (7.6)
x2X T 2L .Z;Y /
Moreover, x is an optimal solution for (P ) if and only if there exists T 2 L .Z; Y /

such that .0; T / 2 @s ˚.x; 0/.
Proof. Consider the marginal function h defined by (7.5); because ˚ is convex, h
is convex, too. Since 0 2 PrZ .dom ˚/ D dom h, we have that h.0/ < C1. If
h.0/ D 1 then the conclusion clearly holds because ˚s .0; T / D hs .T / D C1
for every T 2 L .Z; Y /.
Assume that h.0/ 2 Y . Because in anyone of the conditions (i)–(iii)
0 2 icr.PrZ .D//, by Remark 7.1.1, h 2 .Z; Y /.
(i) It is obvious that h ˚.x0 ; /; applying Proposition 3.2.11 we obtain that h is
continuous at 0. Using now Theorem 7.1.3 we have that @s h.0/ ¤ ;. Taking
T0 2 @s h.0/ we have that
vD vP D h.0/ D hs .T0 / D ˚s .0; T0 / vD :
It follows that (7.6) holds.

(ii) As in (i), hjZ0 ˚.x0 ; /jZ0 , and so hjZ0 is continuous at 0 2 dom hjZ0 . Using
now Corollary 7.1.5 we get T0 2 @s h.0/. As in (i), the conclusion holds.
(iii) Let ˚0 WD ˚jX Z0 . Then ˚0 2 .X Z0 / has closed epigraph (because
epi ˚ D epi ˚0 X Z0 Y ). Consider
F W X Y Z0 ; graph F WD f.x; y; z/ 2 X Y Z0 j .x; z; y/ 2 epi ˚0 g :
It is obvious that graph F is a closed convex set and Im F D PrZ .dom ˚/ D

PrZ0 .dom ˚/ Z0 , and so 0 2 core.Im F /. Take .x0 ; y0 / 2 X Y such
that 0 2 F .x0 ; y0 / ., ˚0 .x0 ; 0/ y0 /. Using [576, Theorem] (see also [614,

Theorem 1.3.7]), we obtain that
8W 2 NY ; U 2 NX ; 9V 2 NZ0 W V F ..x0 ; y0 / C U W / :
Therefore,
8W 2NY ; U 2NX ; 9V 2NZ0 ; 8z 2 V; 9x2x0 CU; y2y0 CW W z 2 F .x; y/;
or, equivalently
8W 2NY ; U 2NX ; 9V 2NZ0 ; 8z 2 V; 9x 2 x0 CU W ˚0 .x; z/ 2 y0 CW C:
In conclusion,
9x0 2 X; y0 2 Y; 8W 2 NY ; U 2 NX ; 9V 2 NZ0 ;
8z 2 V; 9x 2 x0 C U W ˚0 .x; z/ 2 y0 C W C: (7.7)
From the previous relation we obtain that
9y0 2 Y; 8W 2 NY ; 9V 2 NZ0 ; 8z 2 V W hjZ0 .z/ 2 y0 C W C:
Applying Proposition 3.2.11, we obtain that h0 is continuous at 0. Using (ii)

we get the conclusion.
Assume that x is an optimal solution for (P ). Then ˚.x; 0/ D infx2X ˚.x; 0/,
and so, by (7.6), there exists T 2 L .Z; Y / such that ˚.x; 0/ D ˚s .0; T /. It
follows that ˚.x; 0/ C ˚s .0; T / D 0 D 0x C T 0, and so .0; T / 2 @s ˚.x; 0/. The
converse implication is immediate. t
u
As in the scalar case, the preceding result gives the possibility to compute
conjugates and subdifferentials for functions obtained by operations that preserve
convexity.
Theorem 7.1.11. Let f 2 .X; Y /, g 2 .Z; Y / and S 2 L .X; Z/. Assume that
one of the following conditions holds:
(i) there exists x0 2 dom f \ S 1 .dom g/ such that g is continuous at Sx0 I
(ii) X is a locally convex space and there exists x0 2 dom f \ S 1 .dom g/ such
that gjaff.dom g/ is continuous at Sx0 I
(iii) X; Y are Fréchet spaces, C is closed, 0 2 icr.dom g S.dom f // and
span.dom g S.dom f // is a barreled space, f and g are C -lower
continuous.
Then
.f C g ı S /s .T / D min ffs .T U ı S / C gs .U / j U 2 L .Z; Y /g (7.8)

for all T 2 L .X; Y /, and

Sn o
@s& .f C g ı S /.x/ D @s&1 f .x/ C @s&2 g.Sx/ ı S j &1 ; &2 2 C; &1 C &2 D &
(7.9)
for all x 2 X and & 2 C .
Proof. Fix T 2 L .X; Y / and U 2 L .Z; Y /. Then
.T U ı S /x f .x/ C fs .T U ı S /; Uz g.z/ C gs .U / 8x 2 X; z 2 Z:
Taking z WD Sx and summing up both inequalities we get
Tx f .x/ C g.Sx/ C fs .T U ı S / C gs .U / 8x 2 X:
This shows that
.f C g ı S /s .T / fs .T U ı S / C gs .U /:
Hence
.f C g ı S /s .T / inf ffs .T U ı S / C gs .U / j U 2 L .Z; Y /g (7.10)
(for arbitrary f , g and S 2 L .X; Z/). It follows that (7.8) holds if T … dom.f C
g ı S /s .
Take T 2 dom.f C g ı S /s and consider ˚ W X Z ! Y defined
by ˚.x; z/ WD f .x/ C g.Sx C z/ Tx. Clearly, ˚ is convex and 0 2
PrZ .dom ˚/ D dom gS.dom f /; hence ˚ 2 .X Z; Y /; moreover Z0 WD span
.PrZ .dom ˚// D span .dom g S.dom f //.
Assume that (i) holds. Then the mapping z 7! ˚.x0 ; z/ D f .x0 / C g.Sx0 C z/
Tx0 is clearly continuous at 0 (because g is continuous at Sx0 ). Hence condition (i)
of Theorem 7.1.10 holds. In a similar way, condition (ii) of Theorem 7.1.10 holds
when (ii) is verified.
Assume that (iii) holds. Set fQ.x; z/ WD f .x/ Tx D .f T /.PrX .x; z// and
Q z/ WD g.Sx C z/ D g..S; IdZ /.x; z//, where .S; IdZ /.x; z/ WD Sx C z. Since
g.x;
T is continuous and f is C -lower continuous we have that f T is C -lower
continuous by Proposition 3.1.31 (ii); since PrX 2 L .X Z; X /, fQ is C -lower
continuous by Proposition 3.1.31 (iii). Similarly, gQ is C -lower continuous because
g is C -lower continuous and .S; IdZ / 2 L .X Z; Z/. Since ˚ D fQ C g, Q ˚ is
C -lower continuous by Proposition 3.1.31 (ii). Using now Proposition 3.1.31 (i) we
have that epi ˚ is closed. Hence condition (iii) in Theorem 7.1.10 holds. Therefore,
(7.6) holds.
Clearly, infx2X ˚.x; 0/ D .f C g ı S /s .T / 2 Y . Using (7.6), there exists
U 2 L .Z; Y / such that .f C g ı S /s .T / D ˚s .0; U /. But
˚s .0; U / D sup fUz f .x/ g.Sx C z/ C Tx j x 2 X; z 2 Zg

˚
D sup U.z0 Sx/ f .x/ g.z0 / C Tx j x 2 X; z0 2 Z
D sup fUzg.z/ j x 2 X; z 2 Zg C sup f.T U ı S /xf .x/ j x 2 X g
D gs .U / C fs .T U ı S /:
Taking into account (7.10) we obtain that (7.8) holds.

Take now T 2 @s&1 f .x/ and U 2 @s&2 g.Sx/, where &1 , &2 2 C . Then
f .x/ C fs .T / Tx C &1 ; g.Sx/ C gs .U / U.Sx/ C &2 ;
and so
.f C g ı S /.x/ C gs .U / .T C U ı S /x C &1 C &2 ;

which proves that T C U ı S 2 @s&1 C&2 .f C g ı S /.x/. Hence the inclusion in
(7.9) holds (for arbitrary f , g and S 2 L .X; Z/).
Conversely, take T 2 @s& .f C g ı S /.x/. By Proposition 7.1.9 one has that
.f C g ı S /.x/ C .f C g ı S /s .T / Tx C &. From (7.8) we get U 2 L .Z; Y /
such that .f C g ı S /s .T / D fs .T U ı S / C gs .U /, and so
f .x/ C g.Sx/ C fs .T U ı S / C gs .U / Tx C & D .T U ı S /x C U.Sx/ C &;
or, equivalently,

f .x/ C fs .T U ı S / .T U ı S /x C g.Sx/ C gs .U / U.Sx/ &:

Taking &1 WD f .x/ C fs .T U ı S / .T U ı S /x 2 C and &2 WD & &1 2
C , we obtain that T U ı S 2 @s&1 f .x/, U 2 @s&2 g.Sx/. Hence, the inclusion in
(7.9) holds, too. t
u
There are two important particular cases of the preceding result: a) X D Z,
S D IdX , and b) f D 0.
Corollary 7.1.12. Let f; g 2 .X; Y /. Assume that one of the following conditions
holds:
(i) there exists x0 2 dom f \ dom g such that g is continuous at x0 I
(ii) X is a locally convex space and there exists x0 2 dom f \ dom g such that
gjaff.dom g/ is continuous at x0 I
(iii) X; Y are Fréchet spaces, C is closed, 0 2 icr.dom f dom g/ and
span.dom f dom g/ is a barreled space, and f; g are C -lower continuous.
Then
8T 2 L .X; Y / W .f C g/s .T / D min ffs .T S / C gs .S / j S 2 L .X; Y /g ;

S˚ s
8x2X; 8&2C W @s& .f C g/.x/ D @&1 f .x/C@s&2 g.x/ j &1 ; &2 2C; &1 C&2 D & :
Corollary 7.1.13. Let g 2 .Z; Y / and S 2 L .X; Z/. Assume that one of the
following conditions holds:
(i) there exists x0 2 S 1 .dom g/ such that g is continuous at Sx0 I
(ii) X is a locally convex space and there exists x0 2 S 1 .dom g/ such that
gjaff.dom g/ is continuous at Sx0 I
(iii) X; Y are Fréchet spaces, C is closed, 0 2 icr.dom gIm S / and span.dom g
Im S / is a barreled space, and g is C -lower continuous.
Then
8T 2 L .X; Y / W .g ı S /s .T / D min fgs .U / j U 2 L .Z; Y /; U ı S D T g ;

8x 2 X; 8& 2 C W @s& .g ı S /.x/ D @s& g.Sx/ ı S:
Theorem 7.1.14. Let Z be ordered by the convex cone P , f 2 .Z; Y / be .P; C /-

increasing, g 2 .X; Y /, and dom f \ g.dom g/ ¤ ;. Assume that one of the
(i) there exists x0 2 g 1 .dom f / . dom g/ such that f is continuous at g.x0 /I
(ii) X is a locally convex space and there exists x0 2 g 1 .dom f / such that
f jaff.dom f / is continuous at g.x0 /I
(iii) X; Y are Fréchet spaces, 0 2 icr.dom f g.dom g// and span.dom f
g.dom g// is a barreled space, C is closed, f , g are C -lower continuous and
for every y 2 Y there exists z 2 Z with f .z/
y.
Then, setting f .C1/ WD C1, one has
8T 2 L .X; Y / W .f ı g/s .T / D min ffs .U / C .U ı g/s .T / j U 2 LC .Z; Y /g ;

(7.11)
where U.C1/ WD C1, and
Sn s o
@s& .f ı g/.x/ D @&1 .U ı g/.x/ j U 2 @s&2 f g.x/ ; &1 ; &2 2 C; &1 C &2 D &
(7.12)
for all x 2 X and & 2 C .
Proof. Using Proposition 7.1.9 (iii) we have that dom fs LC .Z; Y /. Taking
T 2 L .X; Y / and U 2 LC .Z; Y /, we have that
8x 2 X; 8z 2 Z W f .z/ C fs .U /
Uz; .U ı g/.x/ C .U ı g/s .T /
Tx:
Adding side by side both inequalities for z D g.x/ with x 2 dom g, we get
f .g.x// C fs .U / C .U ı g/s .T /

Tx:
It follows that fs .U / C .U ı g/s .T /

.f ı g/s .T /, and so
8T 2 L .X; Y / W .f ı g/s .T / inf ffs .U / C .U ı g/s .T / j U 2 LC .Z; Y /g :

(7.13)
A similar computation shows that the inclusion in (7.12) holds.
In order to prove (7.11) we use Theorem 7.1.10. For this (it is sufficient to) take
T 2 dom.f ı g/s . Consider ˚ W X Z ! Y , ˚.x; z/ D f .g.x/ C z/ Tx. Since
f and g are convex and f is increasing, ˚ is convex. Moreover, PrZ .dom ˚/ D
dom f g.dom g/, and so 0 2 PrZ .dom ˚/.
In case (i) we have that ˚.x0 ; / is continuous at 0 because f is continuous
at g.x0 / 2 dom f . Similarly, in case (ii), ˚.x0 ; /jZ0 is continuous at 0, where
Z0 WD span .PrZ .dom ˚// D span.dom f g.dom g//. In case (iii) we have
that 0 2 icr .PrZ .dom ˚// and Z0 is barreled. We have to prove that epi ˚ is
closed. For this, by Proposition 3.1.31 (i), it is sufficient to show that ˚ is C -lower
continuous. But gQ W X Z ! Y defined by g.x; Q z/ WD g.x/ C z is proper
and C -lower continuous. Using Proposition 3.1.31 (iv), we have that ˚ is C -lower
continuous. Therefore condition (i), (ii) or (iii) of Theorem 7.1.10 is verified if (i),
(ii) or (iii) holds, respectively. Using Theorem 7.1.10 we get U 2 L .Z; Y / such
that infx2X ˚.x; 0/ D ˚s .0; U /, or, equivalently, .f ı g/s .T / D ˚s .0; U /. But
˚s .0; U / D sup fUz f .g.x/ C z/ C Tx j x 2 X; z 2 Zg

˚
D sup U.z0 g.x// f .z0 / C Tx j x 2 dom g; z0 2 Z
D supfTx U.g.x// j x 2 dom gg C supfUz f .z/ j z 2 Zg
D .U ı g/s .T / C fs .U /:
It follows that U 2 dom fs LC .Z; Y /. Using (7.13) we have that (7.11) holds.
Take now T 2 @s& .f ı g/.x/ with x 2 X and & 2 C . Using Proposition 7.1.9 (ii)
we have that .f ı g/.x/ C .f ı g/s& .T / Tx C &; hence x 2 g 1 .dom f / dom g.
From (7.11) we get U 2 LC .Z; Y / such that .f ıg/s& .T / D fs .U /C.U ıg/s .T /,
and so
.f ı g/.x/ C fs .U / C .U ı g/s .T / Tx C &;
whence

f .g.x// C fs .U / U.g.x// C .U ı g/.x/ C .U ı g/s .T / Tx &:
Taking &1 WD .U ı g/.x/ C .U ı g/s .T / Tx 2 C and &2 WD & &1 2 C , we have

that U 2 @s&2 f .g.x// and T 2 @s&1 .U ı g/.x/. Hence (7.12) holds. t
u
The strong conjugate and subdifferential of vector valued functions were
introduced by Raffin [485] and studied by many authors; see the book [358] by
Kusraev and Kutateladze for a detailed list of references. Theorem 7.1.6 is stated by
Valadier in [580, Corollary 7]; its proof and its preliminaries follow the presentation
in [580]. The other results in this section can be found, essentially, in [610] (see
also [609, 611, 616]).
7.2 The Weak Subdifferential
Throughout this section we assume that C Y is a proper pointed convex cone

with nonempty interior; moreover, we use the notation
E w;C .A/ WD E w .A/ WD WMin A WD WMin.A; C /:
Moreover, for & 2 Y , we set
E&w;C .A/ WD E&w .A/ WD fa 2 A j A \ .a & int C / D ;g D A n .& C A C int C /:

(7.14)
Of course, E w .A/ D E0w .A/. Clearly,
E&w .A/ ¤ ; ) & 2 D w WD Y n . int C /:
The next result represents an extension of Proposition 2.4.5 (iii).

Lemma 7.2.1. Let ; ¤ A Y be closely C -convex and y 2 Y . Then A \ .y
int C / D ; iff
9y 2 C C n f0g; 8y 2 A W hy; y i hy; y i : (7.15)
In particular for & 2 Y , y 2 E&w .A/ iff y 2 A and there exists y 2 C C n f0g such
that hy &; y i hy; y i for every y 2 A.
Proof. Since C C int C D int C , we have that
A \ .y int C / D ; , y … A C int C , .A C C / \ .y int C / D ;:
So, assuming that A \ .y int C / D ;, we have that .A C C / \ .y int C / D ;.
Using a separation theorem, there exists y 2 Y n f0g such that hy C v; y i
hy v0 ; y i for all y 2 A, v 2 C , v0 2 int C . It follows that y 2 C C and hy; y i

hy; y i for every y 2 A. Conversely, let y provided by (7.15). Assuming that
y 2 A \ .y int C /, we get the contradiction 0
hy y; y i > 0 because
y y 2 int C and y 2 C C n f0g. t
u
Let us recall now the notion of (approximate) weak subdifferential of set-valued
maps. First note that F C T is C -convex (resp. C -closed) for T 2 L .X; Y /
provided F W X Y is C -convex (resp. C -closed). Let F W X Y ; the weak
subdifferential of F at .x; y/ 2 graph F is the set
@w F .x; y/ WD @w;C F .x; y/ WD fT 2 L .X; Y / j y T x 2 E w .Im.F T //gI

moreover, if x 2 dom F , the weak subdifferential of F at x is the set
S
@w F .x/ WD @w F .x; y/
y2F .x/
7.2 The Weak Subdifferential 289
(see [566, Definition 4.1]). Similarly, if & 2 Y , the &-weak subdifferential of F at

.x; y/ 2 graph F is the set
@w& F .x; y/ WD @w;C

& F .x; y/ WD fT 2 L .X; Y / j y T x 2 E& .Im.F T //gI
w
of course, @w F .x; y/ D @w0 F .x; y/.

ChenTand Jahn [92] have taken as the weak subdifferential of F at x 2 dom F
the set f@w F .x; y/ j y 2 F .x/g.
T Note that y 2 E&w .F .x// if @w& F .x; y/ ¤ ;; consequently, F .x/ D E w .F .x// if
f@ F .x; y/ j y 2 F .x/g ¤ ;.
w
The next result establishes a characterization of the &-weak sub-gradients of a

convex set-valued map at a point in its graph.
Proposition 7.2.2. Let F W X Y be C -convex, .x; y/ 2 graph F , T 2 L .X; Y /
and & 2 Y . Then T 2 @w& F .x; y/ iff there exists y 2 C C n f0g such that
8.x; y/ 2 graph F W hy y T .x x/ C &; y i

0: (7.16)
Proof. By definition, T 2 @w& F .x; y/ iff y T x 2 E&w .Im.F T //. Since F T is

C -convex we have that Im.F T / is C -convex. Using Lemma 7.2.1 we obtain that
T 2 @w& F .x; y/ iff there exists y 2 C C nf0g such that hy; y i
hy T x &; y i
for all y 2 Im.F T /. The conclusion holds. t
u
A simple existence result for the &-weak subdifferential of F at .x; y/ 2 graph F
is the following.
Proposition 7.2.3. Let F W X Y be C -convex, .x; y/ 2 graph F and & 2 Y .
Assume that int.epi F / ¤ ;. If @w& F .x; y/ ¤ ; then .x; y &/ … int.epi F /.
Conversely, if .x; y &/ … int.epi F / and x 2 int.dom F / then @w& F .x; y/ ¤ ;.
Proof. Take T 2 @w& F .x; y/. By Proposition 7.2.2 there exists y 2 C C n f0g
such that (7.16) holds. Then hx x; T y i C hy .y &/; y i
0 for every
.x; y/ 2 epi F . Since .T y ; y / ¤ .0; 0/, it follows that .x; y &/ … int.epi F /.
Conversely, assume that .x; y &/ … int.epi F / and x 2 int.dom F /. Since epi F
is convex, by a separation theorem, there exists .x ; y / 2 X Y n f.0; 0/g such
that
8.x; y/ 2 epi F W hx x; x i C hy .y &/; y i

0: (7.17)
It follows that y 2 C C . Assuming that y D 0, we obtain that hx x; x i

0
for every x 2 dom F . Since x 2 int.dom F / we get x D 0, and so we get the
C
˝ 0 .x ˛; y / D .0; 0/. Therefore, y 2 C n f0g. There existsk 2
0
contradiction int C
such that k ; y D 1. Taking T W X ! Y defined by Tx WD hx; x i k 0 , from
(7.17) we obtain that (7.16) holds, and so T 2 @w& F .x; y/. t
u
Before establishing other existence results and some rules for the &-weak sub-
differential let us recall the scalarization function associated to cl C and k 0 2 int C :
' D 'k 0 W Y ! R; '.y/ WD infft 2 R j y 2 tk0 cl C g: (7.18)
As usual, we use the conventions
inf ; WD C1; sup ; WD 1; .C1/ C .1/ WD .1/ C .C1/ WD C1:
It is known (see Corollary 5.2.8 for C replaced by clC ) that ' is a C -increasing
continuous sublinear function such that
8
2 R W Œ'
D
k 0 cl C; Œ' <
D
k 0 int C; (7.19)
and
8y 2 Y; 8t 2 R W '.y C tk0 / D '.y/ C t:
Moreover, the Fenchel conjugate of ' is given by the following formula (see
Proposition 5.2.23):
˝ ˛
0 if y 2 C C ; k 0 ; y D 1;
' .y / D (7.20)
C1 otherwise.
The next result is immediate from (7.19).

Lemma 7.2.4. Let A Y and y 2 Y . Then A \ .y int C / D ; iff '.y y/
0
for every y 2 A. In particular for & 2 Y , one has y 2 E&w .A/ iff y 2 A and
'.y y C &/
0 for every y 2 A.
Lemma 7.2.5. If A, B Y then inf '.ACB/ inf '.A/Cinf '.B/. In particular,
if inf '.A C B/ 2 R then inf '.A/, inf '.B/ 2 R.
Proof. The assertions are obvious. t
u
Lemma 7.2.6. Let F W X Y and take
'F W X ! R; .'F /.x/ WD inf ' .F .x// :
Then (i) dom 'F D dom F , (ii) 'F is convex provided F is C -convex, (iii) 'F is
lsc provided F is C -closed.
Proof. We may (and do) assume that graph F ¤ ;. Let us set WD 'F .
(i) The assertion is obvious.
(ii) Assume that F is C -convex. Since .x/ D infy2Y ˚.x; y/ for x 2 X , where
˚.x; y/ WD '.y/Cepi F .x; y/, we have that is convex ( being the marginal
function associated to the convex function ˚).
(iii) Let us set
Ar WD fx 2 X j rk0 2 F .x/ C C g D fx 2 X j .x; rk0 / 2 epi F g:

Since epi F is closed, Ar is closed for every r 2 R. Clearly, Ar As if r < s.

Moreover, if x 2 Ar then there exists y 2 F .x/ such that rk0 2 y C C , that is
y 2 rk0 C ; hence .x/ '.y/ r. Therefore, x 2 Œ r. On the other
hand, if r < s and x 2 Œ r then there exists y 2 F .x/ such that '.y/ < s,
and so y 2 sk0 C ; this implies that x 2 As . Therefore, Ar Œ r As
for all r; s 2 R with r < s. Then
Ar Œ r \s>r As \s>r Œ s D Œ r;
and so Œ r D \s>r As is closed. Hence is lsc. t

u
Using functions of the type 'F we derive subdifferentiability criteria for C -
convex set-valued maps. First let us introduce the strong intrinsic core of A X
as being

icr A if aff A is closed,
sic A WD A WD
ic
; otherwise.
Theorem 7.2.7. Let F W X Y be C -convex, x 2 dom F and y 2 E&w .F .x// for

some & 2 f0g [ int C . Assume that one of the following conditions holds:
(i) int.epi F / ¤ ; and x 2 int.dom F /,
(ii) X is Fréchet, F is C -closed and x 2 sic.dom F /.
Then @w& F .x; y/ ¤ ;; more precisely, there exists x 2 X such that the
operator T 2 L .X; Y / defined by T .x/ WD hx; x i k 0 for x 2 X is in @w& F .x; y/,
where k 0 WD & if & 2 int C and k 0 2 int C if & D 0.
Proof. Let ' be defined by (7.18) with k 0 WD & if & 2 int C and k 0 2
int C if & D 0; set WD 1 if & ¤ 00 and WD 0 if & D 0. Because
y 2 Ek w
0 .F .x// (hence F .x/ y C k \ . int C / D ;), we have that
ı WD inf '.F .x/ y/ 0. Consider W X ! R be defined by .x/ WD
inf '.F .x/ y C ık 0 /. Clearly, .x/ D 0. Because F is C -convex, so is the set-
valued map x 7! F .x/ y C ık 0 , and so, by Lemma 7.2.6 (ii), is a convex
function with dom D dom F .
(i) Take .x0 ; y0 / 2 int.epi F /. Then there exists U 2 NX .x0 / such that
y0 2 F .x/ C C for every x 2 U . Since ' is C -increasing, we obtain that
.x/ '.y0 y C ık 0 / for all x 2 U . Since is convex and finite at
x 2 int.dom F / D int.dom /, we obtain that is finite-valued on dom
and continuous on int.dom / (see e.g. [614, Theorem 2.2.9]). In particular,
@ .x/ ¤ ; (see e.g. [614, Theorem 2.4.9]).
(ii) By Lemma 7.2.6 (iii) we have that is lsc, and so 2 .X / because
.x/ 2 R. Since x 2 sic.dom /, @ .x/ ¤ ; (which is a consequence of
[614, Theorem 2.2.12]).
In both cases take x 2 @ .x/. Then hx x; x i .x/ .x/ D .x/ for

every x 2 X . Therefore, hx x; x i '.y y C ık 0 / for all .x; y/ 2 graph F .
Taking Tx WD hx; x i k 0 , we obtain that '.y .y T x/ C ık 0 /
0 for every
y 2 .F T /.X /. Using Lemma 7.2.4 we obtain that y T x 2 Eık w
0 .Im.F T //,
whence T 2 @k 0 F .x; y/ because ık C k .

w 0 0
t
u
As mentioned in [535, Theorem 2.1] (in an equivalent form), int.epi F / ¤ ;
iff there exists y0 2 Y such that F 1 .y0 C / has nonempty interior. Other (not
necessarily independent) sufficient conditions for int.epi F / ¤ ; are provided in
the next result.
Proposition 7.2.8. Let F W X Y and x0 2 int.dom F /. Assume that one of the
(a) there exists y 2 Y such that F 1 .y C / 2 NX .x0 /,
(b) there exists a (topologically) bounded set E Y such that F 1 .E C / 2
NX .x0 /,
(c) there exist U 2 NX .x0 / and a function f W U ! Y such that f .x/ 2 F .x/ for
every x 2 U and f is continuous at x0 ,
(d) there exist U 2 NX .x0 / and a function f W U ! Y such that f .x/ 2 F .x/ for
every x 2 U and f .U / is (topologically) bounded,
(e) there exists y 2 F .x0 / such that F is lower semicontinuous at .x0 ; y/.
Then there exists y0 2 Y such that .x0 ; y0 / 2 int.epi F /.
Proof. It is clear that F 1 .y0 C / 2 NX .x0 / if .x0 ; y0 / 2 int.epi F /. Assume that
(a) holds, that is U WD F 1 .y C / 2 NX .x0 /, and take k 0 2 int C . Then there
exists V 2 NY such that k 0 C V C . Take x 2 U and y 0 2 y0 C V , where
y0 WD y C k 0 . Since .x; y/ 2 epi F and y y 0 , we have that .x; y 0 / 2 epi F ; hence
U .y0 C V / epi F , which proves that .x0 ; y0 / 2 int.epi F /.
Even if int C D ;, the implications (a) ) (b) and (d) ) (b) are obvious.
(b) ) (a) Consider y 0 2 int C . Because y 0 C 2 NY .0/, there exists > 0
such that E .y 0 C / D y C , where y WD y 0 . It follows that F 1 .E C /
F 1 .y C /.
(c) ) (a) Take y 2 f .x0 / C int C ; from the continuity of f at x0 , U0 WD
f 1 .y C / 2 NX .x0 /. Because f 1 .y C / F 1 .y C /, (a) holds.
(e) ) (a) Take y 0 2 y C int C . Because y 0 C 2 NY .y/ and F is lsc at x0 ,
there exists U 2 N .x0 / such that F .x/ \ .y 0 C / ¤ ; for every x 2 U ; hence (a)
holds. t
u
Recall that F is connected at x0 2 X (see [569, Definition 3.4],
[389, Definition 4]) if there exist U 2 NX .x0 / and a continuous function
f W U ! Y such that f .x/ 2 F .x/ for every x 2 U .
Corollary 7.2.9. Let F W X Y be C -convex and x 2 int.dom F / be such that
F .x/ D E w .F .x// and F .x/ C is convex. Assume that either (i) int.epi F / ¤ ;
or (ii) X is Fréchet, F is C -closed and F .x/ is compact. Then there exists T 2
T
y2F .x/ @ F .x; y/ such that Im T R int C . In fact, T can be taken in the form
w
Tx D hx; x i k 0 with x 2 X and k 0 2 int C .

Proof. Let us define H W X Y by H.x/ WD F .x/ F .x/. Clearly, dom H D

dom F . The set-valued map H is C -convex. Indeed, for x; x 0 2 dom H and
2
0; 1Œ we have that

H.x/ C .1
/H.x 0 / D
F .x/ C .1
/F .x 0 / Œ
F .x/ C .1
/F .x/
F .
x C .1
/x 0 / C C ŒF .x/ C
D H.
x C .1
/x 0 / C C:
Because F .x/ D E w .F .x// we have that H.x/ \ . int C / D ŒF .x/ F .x/ \
. int C / D ;, and so 0 2 E w .H.x//.
(i) Fixing y0 2 F .x/ and taking H0 .x/ WD F .x/ y0 , we have that epi F
.0; y0 / D epi H0 epi H , and so int.epi H / ¤ ;.
(ii) Because F .x/ is compact, H is C -closed.
Applying Theorem 7.2.7 we get x 2 X and k 0 2 int C such that the
operator T 2 L .X; Y / defined by Tx D hx; x i k 0 is in @w H.x; 0/. Therefore,
y 0 T .x x/ … . int C / for all .x; y 0 / 2 graph H , that is y y T .x x/ …
. int C / for all .x; y/ 2 graph F and y 2 F .x/. This shows that T 2
@w F .x; y/ for all y 2 F .x/. t
u
The next result will be used for deriving estimates of the weak subdifferential of
the sum of two set-valued maps.
Lemma 7.2.10. Let ˚ W X Z Y be a C -convex set-valued map with 0 2
PrZ .dom ˚/ and y 2 Y be such that .x; 0; y/ 2 graph ˚ ) y … y int C . Set
Z0 WD span .PrZ .dom ˚//. Assume that one of the following conditions holds:
(i) for some k 0 2 int C one has:
9 2 R; 9V 2 NZ .0/; 8z 2 V \ Z0 ; 9x 2 X W ˚.x; z/ \ .y C k 0 C / ¤ ;I
(7.21)
(ii) X; Y; Z are Fréchet spaces, ˚ is C -closed and 0 2 sic .PrZ .dom ˚//.
Then there exists z 2 Z such that .x; z; y/ 2 graph ˚ ) y hz; z i k 0 … y
int C .
Proof. Consider the function W X Z Y ! R defined by .x; z; y/ WD
'.y y/ C epi ˚ .x; z; y/, where ' is defined in (7.18). Clearly, is a proper
convex function and dom D epi ˚. Moreover, PrZ .dom / D PrZ .epi ˚/ D
PrZ .dom ˚/.
In case (i) we have that V \ Z0 fz 2 Z j 9.x; y/ 2 X Y W .x; z; y/ g,
while in case (ii) we have that is lsc. Using [614, Theorem 2.7.1 (i)] in case (i)
and [614, Theorem 2.7.1 (vii)] in case (ii) we obtain that

1 > ˛ WD inf .x; 0; y/ D max .0; z ; 0/: (7.22)
.x;y/2X Y z 2Z
Take .x; y/ 2 X Y . If .x; 0; y/ 62 epi ˚ then .x; 0; y/ D 1 > 0. Let .x; 0; y/ 2

epi ˚; then there exists y 0 C y such that .x; 0; y 0 / 2 graph ˚. It follows that
y 0 … y int C , and so y … y int C . From the properties of ' we get '.y
y/
0, whence .x; 0; y/
0. Hence ˛
0. By (7.22) there exists z 2 Z
such that .0; z ; 0/ D ˛ 0. Because ' is continuous and˝ using˛ e.g. [614,
Theorem 2.8.7] together with (7.20), there exists y 2 C C with k 0 ; y D 1 such
that 0
.0; z ; 0/ D hy; y i C epi ˚ .0; z ; y /. It follows that for .x; z; y/ 2
graph ˚ we have that
hy; y i
epi ˚ .0; z ; y /
hz; z i hy; y i :
Therefore,
˝ ˛
8.x; z; y/ 2 graph ˚ W y y C hz; z i k 0 ; y D hz; z i hy y; y i 0;
which proves that y hz; z i k 0 … y int C if .x; z; y/ 2 graph ˚. t

u
Theorem 7.2.11. Let F1 W X Y , F2 W Z Y be two C -convex set-valued
maps, and S 2 L.X; Z/. Assume that there exist x0 2 dom F1 \ S 1 .dom F2 / and
y0 2 Y such that .Sx0 ; y0 / 2 int.epi F2 /. Consider .x; y 1 / 2 graph F1 , .S x; y 2 / 2
graph F2 , & 2 C and T 2 @w& .F1 C F2 ı S /.x; y 1 C y 2 /. Then there exists z 2 Z
such that T2 2 @w& F2 .S x; y 2 /, where T2 .z/ WD hz; z i k 0 .z 2 Z/ for some given
k 0 2 int C , and T1 WD T T2 ı S 2 @w& F1 .x; y 1 /. Consequently,

@w& .F1 C F2 ı S /.x; y 1 C y 2 / @w& F1 .x; y 1 / C @w& F2 .S x; y 2 / ı S: (7.23)
Proof. Let us fix k 0 2 int C . Consider
˚ W X Z Y; ˚.x; z/ WD F1 .x/ C F2 .Sx C z/ Tx: (7.24)
It follows easily that ˚ is C -convex. Let us set y WD y 1 C y 2 T x &. Clearly,

.x; 0/ 2 dom ˚; moreover, because T 2 @w& .F1 C F2 ı S /.x; y 1 C y 2 /, we have
that .x; 0; y/ 2 graph ˚ ) y … y int C . We wish to show that (7.21) holds. First
observe that
epi ˚ D f.x; z; y1 C y2 / j .x; y1 / 2 epi.F1 T /; .Sx C z; y2 / 2 epi F2 /g :
Fix y00 2 F1 .x0 / Tx0 and take WD 1 C '.y0 C y00 y/. Because .Sx0 ; y0 / 2
int.epi F2 /, there exist V 2 NZ .0/ such that .Sx0 C V / fy0 g epi F2 . For z 2 V
we have that Sx0 Cz 2 Sx0 CV , and so there exist yz 2 F2 .Sx0 Cz/\.y0 C /. Then
y00 Cyz 2 F1 .x0 /CF2 .Sx0 Cz/Tx0 and y00 Cyz y 2 y0 Cy00 y C k 0 C ,
whence y00 Cyz 2 ˚.x0 ; z/\.y C k 0 C /. Therefore, condition (7.21) holds. Using
Lemma 7.2.10 (i) we get z 2 Z such that .x; z; y/ 2 graph ˚ ) y hz; z i k 0 …
y int C .
Let T2 W Z ! Y be defined by T2 z WD hz; z i k 0 ; clearly T2 2 L.Z; Y /. Let

.z; y/ 2 graph F2 . Then y 1 C y T x 2 F1 .x/ C F2 .z/ T x D ˚.x; z S x/,
and so y 1 C y T x hz S x; z i k 0 … y int C . It follows that y T2 z …
.y 2 T2 .S x// & int C for all .z; y/ 2 graph F2 , and so T2 2 @w& F2 .S x; y 2 /.
Let us take .x; y/ 2 graph F1 . Then y C y 2 Tx 2 F1 .x/ C F2 .S x/ Tx D
˚.x; S.x x//, and so y C y 2 Tx hS.x x/; z i k 0 … y int C . It follows that
y T1 x … .y 1 T1 x/& int C for all .x; y/ 2 graph F1 , and so T1 2 @w& F1 .x; y 1 /.
The inclusion (7.23) is now obvious. The proof is complete. t
u
The set-valued map ˚ defined in (7.24) is not necessarily C -closed if F1 and F2
are C -closed. So, we can not apply Lemma 7.2.10 (ii) for getting the corresponding
results in the case 0 2 sic .S.dom F1 / dom F2 /. In the next result we obtain a
similar conclusion as that of Theorem 7.2.11 in the case in which & 2 f0g [ int C .
Theorem 7.2.12. Let X; Y be Fréchet spaces, F1 W X Y , F2 W Z Y be
two C -convex and C -closed set-valued maps, and S 2 L.X; Z/ be such that 0 2
sic .S.dom F1 / dom F2 /. Assume that .x; y 1 / 2 graph F1 , .S x; y 2 / 2 graph F2
and T 2 @w& .F1 C F2 ı S /.x; y 1 C y 2 / for some & 2 f0g [ int C . Then the conclusion
of Theorem 7.2.11 holds.
Proof. Take k 0 WD & if & ¤ 0 and k 0 2 int C if & D 0, and let ' be defined by
(7.18). Because T 2 @w& .F1 CF2 ıS /.x; y 1 Cy 2 / we have that y WD y 1 Cy 2 T x 2
E&w ŒIm.F1 C F2 ı S T /. Using Lemma 7.2.4 we obtain that
' .y y C &/
0 8y 2 .F1 C F2 ı S T /.X /: (7.25)
Setting ı1 WD inf.F1 .x/ y 1 / 0, ı2 WD inf.F2 .S x/ y 2 / 0, from (7.25)

and Lemma 7.2.5 we get
inf ' ..F1 .x/ y 1 / C .F2 .x/ y 2 // .ı1 / C .ı2 /;
where D 1 if & ¤ 0 and WD 0 if & D 0. Hence ı1 ; ı2 2 RC and ı1 C ı .

Consider 1 W X ! R, 2 W Z ! R be defined by 1 .x/ WD inf '.F1 .x/
Tx .y 1 T x/ C ı1 k 0 / and 2 .z/ WD inf '.F2 .z/ y 2 C ı2 k 0 /. Clearly, 1 .x/ D 0
and 2 .S x/ D 0. Because F1 , F2 are C -convex and C -closed, so are the set-valued
maps x 7! F1 .x/ Tx .y 1 T x/ C ı1 k 0 , z 7! F2 .z/ y 2 C ı2 k 0 , and so,
by Lemma 7.2.6, 1 and 2 are lsc convex functions. Because 1 .x/ D 0 and
2 .S x/ D 0 it follows that 1 and 2 are also proper. Clearly, dom 1 D dom F1
and dom 2 D dom F2 . Let x 2 dom 1 \S 1 .dom 2 / D dom F1 \S 1 .dom F2 /.
Because
.F1 C F2 ı S T / .x/ y D .F1 .x/ Tx .y 1 T x// C .F2 .Sx/ y 2 /;
from (7.25) and Lemma 7.2.5 we get .ı1 C ı2 / 1 .x/ C 2 .Sx/ < 1. Hence
0 WD .ı1 C ı2 /
0 and
8x 2 X W 1 .x/ C 2 .S x/ 1 .x/ C 2 .x/ C 0 :

It follows that 0 2 @0 . 1 C 2 ı S /.x/. Because 1 ; 2 2 .X / and 0 2

sic .S.dom 1 / dom 2 /, using [614, Theorem 2.8.3 (vii)] we have that there exist
01 ; 02
0 with 01 C 02 D 0 and z 2 @02 2 .S x/ such that S z 2 @01 1 .x/.
Because z 2 @02 2 .S x/ we have that hz S x; z i 2 .z/ 2 .S x/ C 02 D
0
2 .z/ C 2 for z 2 Z, whence
8.z; y/ 2 graph F2 W hz S x; z i '.y y 2 C ı2 k 0 / C 02 : (7.26)
Let us set T2 z WD hz; z i k 0 and 2 WD ı2 C 02

0. From (7.26) we obtain that
'.y T2 z .y 2 T2 .S x// C 2 k 0 /
0 for all .z; y/ 2 graph F2 , which shows
that .F2 T2 /.Z/ .y 2 T2 .S x// C 2 k 0 \ . int C / D ;. This proves that
T2 2 @2 k 0 w F2 .S x; y 2 / @w F2 .S x; y 2 /. In a similar way we obtain that T1 WD
T T2 ı S 2 @1 k 0 w F1 .x; y 1 / @w F1 .x; y 1 /, where 1 WD ı1 C 01
0; clearly
D 1 C 2 . t
u
The weak subdifferential of a vector function was introduced by Tanino and
Sawaragi in [569]. Then the notion was studied by many other authors (see
[91, 92, 389, 535, 549–551, 566]). The results of this section are mainly extensions
of results of Taa from [549–551]. Theorem 7.2.2 is stated in [550, Theorem 4.1]
in Banach spaces, Lemma 7.2.6 is practically contained in the proof of [551,
Theorem 3.1], Theorem 7.2.7 (i) for & D 0 is stated in [566, Proposition 4.3] and
completed in [535, Proposition 2.4] (see also [537, Proposition 2.3]), Corollary 7.2.9
was obtained in [92, Theorem 7] under more stringent conditions. In the case
X D Z, S D IdX and & D 0 Theorem 7.2.11 was obtained by Lin in
[389, Theorem 3.1] for F2 connected at some x0 2 int.dom F1 / \ int.dom F2 /
and by Taa in [549, Theorem 3.2] for X and Y Banach spaces, with different
proofs; the case & ¤ 0 is considered in [550, Theorem 4.2], but the estimate for
@w& .F1 C F2 /.x; z1 C z2 / is less precise. Theorem 7.2.12 for Z D X , S WD IdX and
& D 0 is practically [551, Theorem 3.1].
7.3 Subdifferentials Corresponding to Henig Proper

Efficiency
Throughout this section Y is a locally convex space, C Y is a proper pointed

convex cone with C # ¤ ; and is a base of C . As in (2.25), we set NY WD
fV 2 NYc j V \ D ;g, where NYc WD fV 2 NY j V convexg. As already
observed, for V 2 NY , PV WD cone. CV / is a proper convex cone with int PV D
P. C int V /. Note that PV 1 PV 2 (and so WMin.AI PV1 / WMin.AI PV2 /) for
V1 ; V2 2 NY with V1 V2 .
Similarly to &-weak efficiency, for ; ¤ A Y and & 2 Y , we define
S˚
E&;C .A/ WD E& .A/ WD E&w;D .A/ j D 2 D . 2 fh; tg/; (7.27)
7.3 Subdifferentials Corresponding to Henig Proper Efficiency 297
where

DC if D h;
D WD

(7.28)
D if D t;
DC and D being defined in (2.31) and (2.34), respectively; D h corresponds

to global Henig efficiency, while D t corresponds to Henig efficiency w.r.t. the
base .
Correspondingly, for 2 fh; tg we define the following &-subdifferentials
.& 2 Y / of the set-valued map F W X Y at .x; y/ 2 graph F :
S˚ w;D
@;C
& F .x; y/WD@& F .x; y/WD @& F .x; y/ j D2D ; @ F .x; y/WD@0 F .x; y/:
Applying the results from the previous section we get easily the following ones.
Proposition 7.3.1. Let F W X Y be C -convex, .x; y/ 2 graph F , & 2 Y and
T 2 L .X; Y /. Then for 2 ft; hg, T 2 @& F .x; y/ iff there exists y 2 C such
that
8.x; y/ 2 graph F W hy y T .x x/ C &; y i

0; (7.29)
where C h WD C # and C t WD fy 2 Y j inf y . / > 0g.

Proof. Let D t (the proof for D h being similar). Take T 2 @t& F .x; y/; hence
there exists D D PV 2 D .V 2 NY / such that T 2 @w;D & F .x; y/. Applying
Proposition 7.2.2 for C replaced by D, there exists y 2 D C n f0g such that (7.29)
holds. It follows that inf y > 0, and so y 2 C t . Conversely, assume that (7.29)
holds for y 2 C t . Consider V WD fy 2 Y j jhy; y ij 12 inf y g. Then clearly
C
V 2 NY and y 2 PV n f0g. Applying again Proposition 7.2.2 for C replaced
by D WD PV we obtain that T 2 @w;D & F .x; y/, and so T 2 @& F .x; y/.
t
t
u
Theorem 7.3.2. Let F W X Y be C -convex and x 2 int.dom F /. Assume that
for some x0 2 X one of the following conditions holds:
(a) there exists a (topologically) bounded set E Y such that F 1 .E C / 2
NX .x0 /,
(b) there exist U 2 NX .x0 / and a function f W U ! Y such that f .x/ 2 F .x/ for
every x 2 U and f is continuous at x0 ,
(c) there exist U 2 NX .x0 / and a function f W U ! Y such that f .x/ 2 F .x/ for
every x 2 U and f .U / is (topologically) bounded,
(d) there exists y0 2 F .x0 / such that F is lower semicontinuous at .x0 ; y0 /.
Then for 2 fh; tg and & 2 C we have that @& F .x; y/ ¤ ; provided y 2
E& .F .x//.
Proof. Consider now D h and & 2 C (the proof for D t being similar).
Because y 2 E& .F .x//, there exists D 2 DC such that y 2 E&w;D .F .x//; hence
C nf0g int D. In case (a), because F 1 .EC / F 1 .ED/ and int D ¤ ;, we

obtain that condition (b) in Proposition 7.2.8 holds. If (b), or (c), or (d) holds, clearly
the condition (c), or (d), or (e) in Proposition 7.2.8 holds, respectively. Therefore,
there exists y0 2 Y such that .x0 ; y0 / 2 int.epiD F / ¤ ;. Using now Theorem 7.2.7
for the present C replaced by D, k 0 D & .2 int D/ and D 1 if & ¤ 0, respectively
k 0 2 int D and D 0 if & D 0, we have that @w;D & F .x; y/ ¤ ;. t
u
Of course, another sufficient condition instead of conditions (a)–(d) in
Theorem 7.3.2 is the fact that int.epiC F / ¤ ;.
Theorem 7.3.3. Let F1 W X Y , F2 W Z Y be two C -convex set-valued maps,
and S 2 L.X; Z/. Assume that there exists x0 2 dom F1 \ S 1 .dom F2 / such that
one of the following conditions holds:
(a) there exists a (topologically) bounded set E Y such that F21 .E C / 2
NZ .Sx0 /,
(b) there exist V 2 NZ .Sx0 / and a function f2 W V ! Y such that f2 .z/ 2 F2 .z/
for every z 2 V and f2 is continuous at Sx0 ,
(c) there exist V 2 NZ .Sx0 / and a function f2 W V ! Y such that f2 .z/ 2 F2 .z/
for every z 2 V and f2 .V / is (topologically) bounded,
(d) there exists y0 2 F2 .Sx0 / such that F2 is lower semicontinuous at .Sx0 ; y0 /.
Consider .x; y 1 / 2 graph F1 , .S x; y 2 / 2 graph F2 and & 2 C . Then

8 2 fh; tg W @& .F1 C F2 ı S /.x; y 1 C y 2 / @& F1 .x; y 1 / C @& F2 .S x; y 2 / ı S:
(7.30)
Proof. Take D t. Let T 2 @t& .F1 C F2 ı S /.x; y 1 C y 2 /. Then there exists
V 2 NY such that T 2 @w;D & .F1 C F2 ı S /.x; y 1 C y 2 /, where D WD PV . As

in the proof of Theorem 7.3.2, we get y0 2 Y such that .Sx0 ; y0 / 2 int.epiD F2 /.

Applying Theorem 7.2.11 for C replaced by D we obtain that T 2 @w;D & F1 .x; y 1 /C
w;D
@& F2 .S x; y 2 / ı S . The conclusion follows. t
u
In the case D t, Proposition 7.3.1 and Theorems 7.3.2, 7.3.3 were obtained in
[386] for & D 0 and in [623] for int C ¤ ; under slightly less general conditions.
7.4 Exact Formulas for the Subdifferential of the Sum

and the Composition
We assume that X , Y , Z are topological vector spaces, Y being ordered by the

proper pointed convex cone C . Besides the notations in (7.14) and (7.27), for A Y
and & 2 Y we set
˚
E&s;C .A/ WD E&s .A/ WD y 2 A j y 5 y 0 C & 8y 2 A :
7.4 Exact Formulas for the Subdifferential of the Sum and the Composition 299
In the sequel, when dealing with 2 fs; w; h; tg we assume:

• C is closed and Y is a locally convex space if D s,
• int C ¤ ; if D w,
• C # ¤ ; if D h,
• is a base of C and Y is a locally convex space if D t.
Remark 7.4.1. Note that for a nonempty set A Y , & 2 Y and 2 fs; w; h; tg
one has
E& .A/ ¤ ; ) & 2 D ;
where
8
<C if D s;
D WD Y n . int C /

if D w; (7.31)
: T
Y n . fint D j D 2 D g/ if 2 fh; tg;
Moreover, for &; & 0 2 Y and 2 fs; w; h; tg one has
& 5 & 0 ) E& .A/ E&0 .A/: (7.32)
Besides C h and C t used in Proposition 7.3.1, we set C s WD C w WD C C n f0g. So

8 C
< C n f0g if 2 fw; sg
C WD C # if D h;
:
fy 2 Y j inf y . / > 0g if D t:
Lemma 7.4.2. Let A Y be a nonempty set and & 2 Y .

(i) One has
T
E&s .A/ D h&; y i - arg minA y : (7.33)
y 2C s
(ii) For 2 fw; h; tg one has

S
h&; y i - arg minA y E& .A/I (7.34)
y 2C
moreover, if A is closely C -convex then equality holds in (7.34).

Proof. The proof of (i) is easy.
(ii) Take D w; hence int C ¤ ;. Take y 2 C w D C C n f0g and y 2 h&; y i-
arg minA y . Then hy &; y i hy; y i for all y 2 A. By Lemma 7.2.1 we obtain
that A \ .y & int C / D ;, and so y 2 E&w .A/.
Conversely, assume that A is closely C -convex and take y 2 E&w .A/. Then
y 2 A and A \ .y & int C / D ;. Using again Lemma 7.2.1, there exists
y 2 C C n f0g D C w such that hy &; y i hy; y i for all y 2 A. It follows that
y 2 h&; y i-arg minA y .
The proof for the case 2 fh; tg is similar. One uses the fact that A is closely
D-convex for D 2 D . t
u
From the very definition of @& f .x/, for every proper function f W X ! Y and
every x 2 X one has
8T 2 L.X; Y /; 8 2 fs; w; h; tg; 8& 2 Y W @& .f C T /.x/ D @& f .x/ C T I

(7.35)
moreover, if S 2 L.Z; X / is an isomorphism (of t.v.s.), then
8z 2 Z; 8 2 fs; w; h; tg; 8& 2 Y W @& .f ı S /.z/ D @& f .S z/ ı S: (7.36)
Proposition 7.4.3. Let f W X ! Y be a proper positively homogeneous operator

(that is f .0/ D 0 and f .tx/ D tf .x/ for t > 0 and x 2 X ). Then

@ f .0/ if & 2 D ;
8 2 fs; w; h; tg W @& f .0/ D
; otherwise
where D is defined by (7.31).

Proof. Consider first D s (hence C is closed). Take T 2 @s& f .0/. Then
.f T /.x/ C & 2 C for every x 2 dom f . Taking x D 0 we get & 2 C D D s .
For x 2 dom f and n 2 N we have that .f T /.nx/ C & 2 C , and so
.f T /.x/ C n1 & 2 C for n
1. Because C is closed we get .f T /.x/ 2 C .
It follows that T 2 @s f .0/. Conversely, take & 2 C and T 2 @s f .0/. Then
.f T /.x/ C & 2 & C C C , and so T 2 @s& f .0/.
Consider D w (hence int C ¤ ;). Take T 2 @s& f .0/. Then .f T /.x/ C & 2
Y n . int C / for every x 2 dom f . Taking x D 0 we get & 2 Y n . int C / D D w .
For x 2 dom f and n 2 N we have that .f T /.nx/ C & 2 Y n . int C /, and
so .f T /.x/ C n1 & 2 Y n . int C / for n
1. Taking the limit for n ! 1
we get .f T /.x/ 2 Y n . int C /. It follows that T 2 @w f .0/. Conversely, take
& 2 Y n . int C / and T 2 @w f .0/. Then .f T /.x/ C & 2 & C ŒY n . int C /
Y n . int C /, and so T 2 @w& f .0/.
The proof of the case 2 fh; tg is similar to the case D w (just use some
D 2 D instead of C ). t
u
Of course, if f W X ! Y is proper and g W X ! Y is defined by g.x/ WD
f .x x0 / C y0 for some x0 2 X and y0 2 Y , then
8x 2 X; 8 2 fs; w; h; tg W @& g.x/ D @& f .x x0 /: (7.37)

Let us set
#C .X; Y / WD # .X; Y / WD @ .0/ . 2 fs; w; h; tg/; (7.38)
where W X ! Y , .x/ WD 0 for every x 2 X .

Proposition 7.4.4. Let # .X; Y / be defined by (7.38). Then

f0g if D s;
# .X; Y / D (7.39)
fT 2 L.X; Y / j ker T \ C ¤ ;g if D fw; h; tg:

Proof. The assertion is obvious for D s.

Take D w; hence int C ¤ ;. Let T 2 @w .0/. Then 0 2 Ew .T .X //, and so
T .X / \ int C D ;. By a separation there exists y 2 Y n f0g such that hTx; y i
hy; y i for all x 2 X and y 2 C . It follows that hx; T y i D hTx; y i D 0 for
every x 2 X , and so T y D 0, and 0 hy; y i for every y 2 C , and so y 2 C C .
Therefore, the inclusion holds in (7.39) holds for D w. Conversely, assume that
T 2 L.X; Y / is such that ker T \ C w ¤ ; and take y 2 ker T \ .C C n f0g/.
Then hTx; y i D 0 hy; y i for all x 2 X and y 2 C . Because y ¤ 0, we have
that hy; y i > 0 for y 2 int C , and so T .X / \ . int C / D ;. Hence T 2 @w .0/.
Therefore, (7.39) holds for D w.
Take 2 fh; tg; hence C ¤ ;. Then
S˚ D S ˚
# .X; Y /D #w .X; Y / j D2D D T 2L.X; Y / j ker T \.D C n f0g/¤;
D2D
˚ S C
D T 2L.X; Y / j ker T \ fD n f0g j D 2 D g ¤ ;
D fT 2 L.X; Y / j ker T \ C ¤ ;g:
u
Corollary 7.4.5. Let T 2 L.X; Y /. Then for every x 2 X we have that

T C # .X; Y / if & 2 D ;
8x 2 X; 8 2 fs; w; h; tg W @& T .x/ D

; otherwise:
Proof. From (7.37) we get @& T .x/ D @& T .0/, while from Proposition 7.4.3 we get
@& T .0/ D @ T .0/ for & 2 D and @& T .0/ D ; if & 2 Y n D . Using (7.35) we
have that @ T .0/ D @ .0/ C T D T C # .X; Y /. The conclusion follows. t
u
In the next result one provides formulas for the approximate subdifferentials
using scalar functions. As usual, for y 2 C C we set y .C1/ D C1.
Theorem 7.4.6. Let f W X ! Y be proper and & 2 Y . Then:
(i) for every x 2 X one has
T ˚
@s& f .x/ D T 2 L.X; Y / j y ı T 2 @h&;y i y ı f .x/ I (7.40)
y 2C s
(ii) for 2 fw; h; tg and x 2 X one has

S ˚
@& f .x/ T 2 L.X; Y / j y ı T 2 @h&;y i y ı f .x/ ; (7.41)
y 2C
with equality if f is C -convex.

Proof. The result is a direct consequence of Lemma 7.4.2 and the definition of
@& f .x/. u
t
Corollary 7.4.7. Let f W X ! Y be a proper C -convex function, x 2 dom f and
2 fw; h; tg.
(i) If @ f .x/ ¤ ; then there exists y 2 C such that y ı f is lsc at x.
(ii) Conversely, assume that X is a Fréchet space and x 2 sic.dom f /. If y ı f is
lsc at x for some y 2 C , then @ f .x/ ¤ ;.
Proof. (i) Because f is C -convex, we have equality in (7.41). Taking T 2
@ f .x/, there exists y 2 C such that y ı T 2 @ .y ı f / .x/, and so y ı f
is lsc at x.
(ii) Because dom y ı f D dom f we have that x 2 sic.dom y ı f /. The
function y ı f being a proper convex function, and lsc at x, applying [359,
Corollary 3.2] we obtain that @.y ı f /.x/ ¤ ;. Take x 2 @.y ı f /.x/
and T0 2 L.X; Y / defined by T0 x WD hx; x i y0 , where y0 2 C is such that
hy0 ; y i D 1 (the existence of y0 is ensured by the fact that y 2 C ). Then
y ı T D x . From (7.41) we obtain that T0 2 @ f .x/ ¤ ;. t
u
The -regularity . 2 RC / of the proper function f W X ! Y at x 2 dom f
was defined in Sect. 7.1:
S˚
8y 2 C C n f0g W @ .y ı f /.x/ D y ı @s& f .x/ j & 2 C; h&; y i D ;
(7.42)
Theorem 7.1.6 provides a sufficient condition for the 0-regularity of f 2 .X; Y / at

x 2 int.dom f /. For example, if Y D Rm and C Rm is a pointed closed convex
cone, any proper C -convex function f W X ! Y is regular at x 2 int.dom f /
provided f is continuous at x.
Remark 7.4.8. Let y 2 C C n f0g be such that there exists k 2 C with hk; y i > 0
(or, equivalently, 9k 2 C W hk; y i D 1). Consider ; ¤ A Y and IA W X ! Y ,
IA .x/ WD 0 for x 2 A, IA .x/ WD C1 for x 2 X n A. Then f WD IA satisfies (7.2)
at x 2 A for every 2 RC . In particular, if for every y 2 C C n f0g there exists
k 2 C with hk; y i > 0 then IA is -regular at x 2 A for every
0.
Indeed, fix x 2 A and 2 RC . As already observed, the inclusion in (7.42)
holds always. Clearly, y ı IA D A , where A is the indicator function of A. Take
x 2 @ A .x/, that is hx x; x i for every x 2 A. Consider T 2 L.X; Y /
defined by Tx WD hx; x i k, where k 2 C is such that hk; y i D 1. Then
Tx T x D hx x; x i k & WD k for every x 2 A. Therefore, T 2 @s& IA .x/.

Since y ı T D x , we have that x 2 y ı @s& IA .x/. Hence (7.42) holds.
For 2 fw; h; tg and 2 RC one says that f is --regular at x 2 dom f if
(7.42) holds for every y 2 C ; f is -regular at x 2 dom f if f is -0-regular
at x.
Of course, if f is -regular at x 2 dom f then f is --regular at x for 2
fw; h; tg.
Let us observe that the inclusion @ .f C g/.x/ @ f .x/ C @ g.x/ for x 2
dom f \dom g, well known in the scalar case, is not true in general for 2 fw; h; tg.
This can be seen taking f D g D . The above inclusion becomes # .X; Y / C
# .X; Y / # .X; Y /. Since # .X; Y / is a cone, the relation above would imply
that # .X; Y / is a convex cone. Taking X WD R, Y WD R2 and C WD R2C , we have
that T; S 2 #w .X; Y / for Tx WD .2x; x/, Sx WD .x; 2x/, but T C S … #w .X; Y /.
The above example shows that we can not expect to obtain equality in Theo-
rem 7.2.11 (for example), even for & D 0 and for (multi) functions with very good
properties if dim Y > 1. For this reason the conclusion of the following result is
important.
Theorem 7.4.9. Let f; g W X ! Y be proper operators, 2 fw; h; tg, and & 2 Y .
Then
S˚
8x 2 X W @& .f C g/.x/ @&1 f .x/ C @s&2 g.x/ j &1 2 Y; &2 2 C; &1 C &2 D & :
(7.43)
Moreover, assume that f; g 2 .X; Y / and either (i) X is a l.c.s., f or g is
continuous at some point in dom f \ dom g, or (ii) f; g are C -lower continuous, X
is a Fréchet space and 0 2 sic.dom f dom g/.
If g is -regular at x 2 dom f \ dom g then
@ .f C g/.x/ D @ f .x/ C @s g.x/: (7.44)
If & ¤ 0 and g is --regular at x 2 dom f \ dom g for every 2 RC then

S˚
@& .f Cg/.x/ D @&1 f .x/C@s&2 g.x/ j &1 2 Y; &2 2 C; &1 C&2 D & : (7.45)
Proof. Take &1 2 Y , &2 2 C with &1 C &2 D & and T 2 @w&1 f .x/, S 2 @s&2 g.x/.
Assume that T C S … @w& .f C g/.x/, and so there exists x 0 2 dom f \ dom g
such that .f .x 0 / C g.x 0 / Tx0 Sx0 / .f .x/ C g.x/ Tx Sx/ 2 .& int C /.
Because S 2 @s&2 g.x/ we have that g.x/ Sx g.x 0 / C Sx0 2 .&2 C /, and so we

get f .x 0 / Tx0 .f .x/ Tx/ 2 &2 & C . int C / C .C / D .&1 int C /,
contradicting the fact that T 2 @w&1 f .x/. Hence (7.43) holds.
A similar argument (just replacing C by some D 2 D ) shows that the inclusion
(7.43) holds in the case 2 fh; tg, too.
Assume now that (i) or (ii) holds. Let & ¤ 0 and g be --regular for every
0
at x 2 dom f \ dom g, and take T 2 @& .f C g/.x/. Since f C g is convex, by
Theorem 7.4.6 there exists y 2 C such that y ı T 2 @h&;y i .y ı .f C g//.x/.
Clearly, y ı .f C g/ D y ı f C y ı g. In case (i) we have that one of the

convex functions y ı f , y ı g is continuous at some point in dom f \ dom g D
dom.y ı f / \ dom.y ı g/. Using Theorem 7.1.11 (i) for Y D R, we get
S
@h&;y i y ıf Cy ıg .x/D @1 .y ıf /.x/C@2 .y ıg/.x/ :
1 ;2 0; 1 C2 Dh&;y i
(7.46)
In case (ii), by Proposition 3.1.31 (iv), y ı f and y ı g are lsc. Applying

Theorem 7.1.11 (iii) for Y D R, we have that (7.46) holds again.
Hence there exist 1 ; 2
0 with 1 C 2 D h&; y i and u 2 @1 .y ı f /.x/,
v 2 @2 .y ıg/.x/ such that y ıT D u Cv . Because g is -2 -regular at x, there

exist &2 2 2 C C with h&2 ; y i D 2 and S 2 @s&2 g.x/ such that v D y ı S .

Then u D y ıT y ıS D y ı.T S / 2 @h&1 ;y i .y ıf /.x/, where &1 WD & &2 .
Using again Theorem 7.4.6 we have that u 2 @&1 f .x/. Therefore, (7.45) holds.
For & D 0 and g -regular at x, the proof is similar (but simpler, because 1 D
2 D 0, &1 D &2 D 0/. t
u
The next result relates @ f .x/ and @s f .x/ for 2 fw; h; tg.
Corollary 7.4.10. Let f 2 .X; Y / and 2 fw; h; tg. If f is -regular at x 2
dom f then
@ f .x/ D @s f .x/ C # .X; Y /:
If f is --regular at x 2 dom f for every 2 RC then

S˚ s
8& 2 Y n f0g W @& f .x/ D # .X; Y / C @& 0 f .x/ j & 0 2 C \ .& D / :
Proof. Apply Theorem 7.4.9 for and f , taking into account Corollary 7.4.5. t
u
The next result deals with the composition of convex functions. We assume that
Z is ordered by a pointed convex cone K.

Theorem
1
g W X ! Z , f W Z ! Y be proper operators with dom f ı
7.4.11. Let
g D g .dom f / ¤ ;, & 2 Y and 2 fw; h; tg. Then
Sn o
@& .f ı g/.x/ @&1 .T ı g/.x/ j &1 2Y; &2 2C; &1 C &2 D &; T 2 @s&2 f .g.x//
(7.47)
for every x 2 X , where f .C1/ WD C1, T .C1/ WD C1.
Moreover, assume that f 2 .Z; Y / is .K; C /-increasing, g 2 .X; Z/ and
either (i) X and Z are l.c.s., f is continuous at some point in dom f \ g.dom g/,
or (ii) f; g are C -lower continuous, dom g is closed, X and Z are Fréchet spaces,
and 0 2 sic.dom f g.dom g//. Consider x 2 dom f ı g.
If f is -regular at g.x/ then

S
@ .f ı g/.x/ D f@ .T ı g/.x/ j T 2 @s f .g.x//g : (7.48)
If & ¤ 0 and f is --regular at g.x/ for every

0 then
Sn o
@& .f ı g/.x/ D @&1 .T ı g/.x/ j &1 2 Y; &2 2 C; &1 C&2 D&; T 2 @s&2 f .g.x//
(7.49)
Proof. For x 2 X n dom f ı g (7.47) holds clearly. Let x 2 dom f ı g, &1 2
Y; &2 2 C with &1 C &2 D & and T 2 @s&2 f .g.x//, S 2 @w&1 .T ı g/.x/. Assume that

S 62 @w& .f ı g/.x/. Then there exists x 0 2 dom f ı g such that f .g.x 0 // Sx0
Œf .g.x// Sx 2 .& int C /. Since T 2 @s&1 f .g.x// we have that T .g.x 0 /
g.x// 2 f .g.x 0 // f .g.x// C &2 C . Summing up both relations we get

T .g.x 0 // Sx0 ŒT .g.x// Sx 2 Œ&1 C .C / C . int C / D &1 int C;
a contradiction because S 2 @w&1 .T ı g/.x/.

The proof for 2 fh; tg is similar (in fact reduces to the preceding case taking
some D 2 D ).
Assume now f 2 .Z; Y / is .K; C /-increasing, g 2 .X; Z/ and (i) or (ii)
holds. Let & ¤ 0 and f be --regular for every
0 at x 2 dom f ı g, and
take T 2 @& .f ı g/.x/. Since f ı g is convex, by Theorem 7.4.6 there exists
y 2 C such that y ı S 2 @ .y ı .f ı g// .x/ D @ ..y ı f / ı g/ .x/, where
WD h&; y i. Since y 2 C C C n f0g, we have that fQ WD y ı f is convex and
increasing. In case (i) fQ is continuous at some point in dom fQ \ g.dom g/, while in
case (ii) fQ is convex and lsc. Using [102, Proposition 4.11] we get

@ fQ ı g .x/ D @ .y ı f / ı g .x/
S˚
D @1 .z ı g/.x/ j z 2 @2 fQ.g.x//; 1 ; 2
0; 1 C 2 D :
Hence there exist 1 ; 2

0 with 1 C 2 D and z 2 @2 fQ.g.x// D @2 .y ı
f /.g.x// such that y ı S 2 @1 .z ı g/.x/. Because f is -2 -regular at g.x/,
there exist &2 2 2 C C with h&2 ; y i D 2 and T 2 @s&2 f .g.x// such that
z D y ı T . Taking &1 WD & &2 , we have that h&1 ; y i D 1 . Hence y ı S 2
@1 .z ı g/.x/ D @1 .y ı .T ı g//.x/, and so, using again Theorem 7.4.6, we obtain
that S 2 @&1 .T ı g/.x/. Hence (7.49) holds.
For D 0 and f -regular at g.x/ the proof is similar (but simpler, because
1 D 2 D 0, &1 D &2 D 0/. t
u
Corollary 7.4.12. Let Y be a l.c.s., T 2 L.X; Z/, f 2 .Z; Y /, x 2 T 1 .dom f /
and 2 fw; h; tg. Assume that either (a) X , Z are l.c.s., f is continuous at Tx for
some x 2 T 1 .dom f /, or (b) X , Z are Fréchet spaces, f C -lower continuous and
0 2 sic.dom f Im T /.
If f is -regular at T x then
@ .f ı T /.x/ D @s f .T x/ ı T C # .X; Y /: (7.50)
If f is --regular at T x for every

0 then
S˚
8& 2 Y nf0g W @& .f ıT /.x/ D # .X; Y /C @s& 0 f .T x/ıT j & 0 2 C \.& D / :
(7.51)
Proof. We prove only the case in which & 2 Y n f0g and f is --regular at T x for
every
0 (the proof in the other case being similar). Using Theorem 7.4.11 and
Corollary 7.4.5 we obtain that
Sn o
@& .f ı T /.x/D @&1 .S ı T /.x/ j &1 2 Y; &2 2 C; &1 C&2 D&; S 2@s&2 f .T x/
Sn o
D S ı T CY .X; Y / j &1 2D ; &2 2C; &1 C&2 D&; S 2@s&2 f .T x/
Sn s o
D# .X; Y /C @&2 f .T x/ ı T j &1 2 D ; &2 2 C; &1 C &2 D & :
Hence (7.51) holds. t

u
The results presented in this section are mainly obtained by El Maghri and
Laghdir in [175] (for & D 0) and [174] (for general &), where the cases D w
and D h (denoted p) were considered. For simplifying the presentation we have
taken the cone C to be pointed.
Chapter 8
Duality
In this chapter we present duality assertions for set-valued optimization problems in

infinite dimensional spaces where the solution concept is based on vector approach,
on set approach as well as on lattice approach. For set-valued optimization problems
where the solution concept is based on vector approach we present conjugate
duality statements. The notions of conjugate maps, subdifferential and a pertur-
bation approach used for deriving these duality assertions are given. Furthermore,
Lagrange duality for set-valued problems based on vector approach is shown.
Moreover, we consider set-valued optimization problems where the solution concept
is given by a set order relation introduced by Kuroiwa and derive corresponding
saddle point assertions. For set-valued problems where the solution concept is based
on lattice structure, we present duality theorems that are based on an consequent
usage of infimum and supremum (in the sense greatest lower and least upper bounds
with respect to a partial ordering). We derive conjugate duality assertions as well as
Lagrange duality statements.
A comparison of different approaches to duality in set-valued optimization is
given at the end of this chapter.
It is an old idea to study additionally to a given optimization problem (p.x/ !
inf with infimal value I) a corresponding dual problem (d.u/ ! sup with supremal
value S; S I ), remember the dual variational principles of Dirichlet and
Thompson or simply the pair of dual programs in linear optimization. The reasons
for the introduction of a useful dual problem are the following:
• The dual problem has (under additional conditions) the same optimal value as
the given “primal” optimization problem, but solving the dual problem could be
done with other methods of analysis or numerical mathematics.
• An approximate solution of the given minimization problem gives an estimation
of the infimal value I from above, whereas an approximate solution of the dual
problem is an estimation of I from below, so that one gets intervals containing I.

DOI 10.1007/978-3-642-54265-7__8
308 8 Duality
• Recalling Lagrange method, saddle points, equilibrium points of two person

games, shadow prices in economics, perturbation methods or dual variational
principles, it becomes clear, that optimal dual variables often have a special
meaning for the given problem.
Of course, the just listed advantages require a skilfully chosen dual program.
Nevertheless, the mentioned points are motivation enough, to look for dual problems
in set-valued optimization with corresponding properties too. There are a lot of
papers, which are dedicated to that aim, also a lot of survey papers (see the books in
[73, 292, 293, 402] and references therein).

on Vector Approach
8.1.1 Conjugate Duality for Set-Valued Problems Based

on Vector Approach
A comprehensive and detailed theory for set-valued conjugate duality in infinite

dimensional spaces is presented in the book by Boţ, Grad and Wanka [73]. In
[73] the notions of conjugate maps, subdifferential and a perturbation approach
introduced by Kawasaki [315], Tanino and Sawaragi [569] are extended and applied.
Assume that F W X Y D Y [f1g[fC1g, X and Y are linear topological
spaces, X and Y the corresponding topological dual spaces. The domain of F is
given by Dom F WD fx 2 X j F .x/ ¤ ; and F .x/ ¤ fC1gg and the graph of F
by graph F WD f.x; y/ 2 X Y j y 2 F .x/g. Suppose that C Y is a proper
pointed closed convex cone. However, it is possible to show the results presented in
this section without the closedness assumption concerning C (see [73]). We denote
the set of linear continuous mappings from X to Y by L .X; Y /. In this section we
consider the set-valued optimization problem
minimize F .x/ subject to x 2 X; (SP)
where we use the notations from Definition 2.4.1 and the solution concept intro-
duced in Definition 2.6.1 adapted for set-valued functions F W X Y . Consider
A Y . In the case that A D ;, take by convention Min.A; C / D fC1g and
Max.A; C / D f1g. The operations with sets in Y are taken like for sets in Y (see
Sect. 2.6.3). It is important to mention that if 1 2 Max.A; C /, then A D f1g
or A D ;. Furthermore, if C1 2 Min.A; C /, one has A D fC1g or A D ;.
In order to construct a dual problem to (SP) using the concept of conjugation we
introduce in the following definition the conjugate map of F , the biconjugate map
of F and the subgradient of F (compare also Sects. 7.1 and 7.2).
8.1 Duality Assertions for Set-Valued Problems Based on Vector Approach 309
Definition 8.1.1. Consider a set-valued map F W X Y .

(a) The set-valued map F W L .X; Y / Y defined by
F .T / WD Max.[x2X ŒT .x/ F .x/; C /
is called the conjugate map of F .

(b) The set-valued map F W X Y defined by
F .x/ WD Max.[T 2L .X;Y / ŒT .x/ F .T /; C /
is called the biconjugate map of F .

(c) The operator T 2 L .X; Y / is said to be a subgradient of F at .x; y/ 2
graph F \ .X Y / if
Tx y 2 Max.[u2X ŒT u F .u/; C /:
The set of all subgradients of F at .x; y/ 2 graph F \ .X Y / is called

the subdifferential of F at .x; y/ 2 graph F \ .X Y / and it is denoted
by @F .xI y/. Moreover, for all x 2 X denote
@F .x/ WD [y2F .x/\Y @F .xI y/:
If for all y 2 F .x/ \ Y it holds @F .xI y/ ¤ ; then F is said to be

subdifferentiable at x.
Concerning the existence of subgradients in the sense of Definition 8.1.1 (c) the
following assertion is given by Boţ, Grad, Wanka [73, Proposition 7.1.4].
Proposition 8.1.2. Let F W X Y be a set-valued map and .x; y/ 2 graph F \
.X Y /. Then @F .xI y/ ¤ ; if and only if y 2 F .x/. Consequently, F is
subdifferentiable at x if and only if F .x/ \ Y F .x/.
In Definition 8.1.1 (a) the conjugate map F is defined for linear continuous
operators T and in Definition 8.1.1 (c) the subgradients of F are linear continuous
operators. An other possibility would be to replace T 2 L .X; Y / by linear continu-
ous functionals x 2 X (see Tanino and Sawaragi [569] for the finite dimensional
case, Boţ, Grad and Wanka [73, Definition 7.1.9] for the infinite dimensional case
and Definition 8.1.3 below). The relationship between subgradients in the sense
of Definition 8.1.1 and the following Definition 8.1.3 is given for T 2 L .X; Y /,
x 2 X and c 2 Y n f0g by
Tx D hx ; xic for x 2 X:
The following definitions of conjugates and subgradients with respect to an

element c 2 Y n f0g are introduced in [73, Definition 7.1.9].
310 8 Duality
Definition 8.1.3. Consider a set-valued map F W X Y and c 2 Y n f0g.

(a) The set-valued map Fc W X Y defined by
Fc .x / WD Max.[x2X Œhx ; xic F .x/; C /;
is called the c-conjugate map of F .

(b) The set-valued map Fc W X Y defined by
Fc .x/ WD Max.[x 2X Œhx ; xic Fc .x /; C /;
is called the c-biconjugate map of F .

(c) The vector x 2 X is called c-subgradient of F at .x; y/ 2 graph F \ .X
Y / if
hx ; xic y 2 Max.[u2X Œhx ; uic F .u/; C /:
The set of all c-subgradients of F at .x; y/ 2 graph F \ .X Y / is called the

c-subdifferential of F at .x; y/ 2 graph F \ .X Y / and it is denoted by
@c F .xI y/. Moreover, for all x 2 X denote
@c F .x/ WD [y2F .x/\Y @c F .xI y/:
If for all y 2 F .x/ \ Y it holds @c F .xI y/ ¤ ; then F is said to be c-

subdifferentiable at x.
Remark 8.1.4. The definition of c-subgradients given above is closed to Defi-
nition 8.4.5 where we use infimal sets instead of the Pareto maximum like in
Definition 8.1.3 (c).
Remark 8.1.5. In Boţ, Grad and Wanka [73] the definition of c-subgradients is used
in order to introduce the property that the primal set-valued optimization problem
is c-stable: The problem (SP) is called c-stable with respect to a certain set-valued
perturbation map if the corresponding minimal value map is c-subdifferentiable at 0.
In the following we present weak and strong conjugate duality statements for
set-valued optimization problems where the solution concept is based on vector
approach. For deriving duality assertions for set-valued optimization problems via
conjugation the following generalization of the Young-Fenchel inequality given by
Boţ, Grad, Wanka [73, Proposition 7.1.1] is important.
Proposition 8.1.6. Let F W X Y be a set-valued map, x 2 X and T 2
L .X; Y /. Then for all y 2 F .x/ and all y 2 F .T / it holds
y C y … Tx .C n f0g/: (8.1)
Consider now F W X Y [fC1g. Let us introduce a set-valued perturbation

map
˚ W X V Y [ fC1g such that ˚.x; 0/ D F .x/ for all x 2 X;
where V is a linear topological space, the so called perturbation space, its

topological dual space is denoted by V . Then the set-valued problem (SP) is
embedded into a family of perturbed problems
Min.˚.X; v/; C /; (SPv )
where v 2 V is the perturbation variable.

Using the conjugate of the perturbation map
˚ W L.X; Y /L.V; Y / Y ; ˚ .T; Q/ D Max.[x2X;v2V ŒTxCQv˚.x; v/; C /
we can attach a set-valued dual problem
maximize ˚ .0; Q/ subject to Q 2 L .V; Y / (DSP)
to the primal problem (SP). The following weak duality statement is shown in Boţ,
Grad, Wanka [73, Theorem 7.1.11].
Theorem 8.1.7 (Weak Duality for (SP) and (DSP)). For all x 2 X and all Q 2
L .V; Y / it holds
˚.x; 0/ \ f˚ .0; Q/ .C n f0g/g D ;:
Proof. Let x 2 X and Q 2 L .V; Y / be arbitrarily chosen. Suppose that ˚.x; 0/ \

f˚ .0; Q/ .C n f0g/g ¤ ;. Then there exists y 2 ˚.x; 0/ \ f˚ .0; Q/ .C n
f0g/g. This means that there exist y 2 F .x/ and y 2 ˚ .0; Q/ with y C y 2
C n f0g. So we get a contradiction to (8.1) in Proposition 8.1.6. t
u
Boţ, Grad, Wanka [73, Corollary 7.1.12] gave the following equivalent formula-
tion of the weak duality assertion in Theorem 8.1.7:
Corollary 8.1.8. For all x 2 X and all Q 2 L .V; Y /, it holds y … y .C n f0g/
whenever y 2 F .x/ and y 2 ˚ .0; Q/.
Furthermore, from the weak duality in Theorem 8.1.7 we get:
Corollary 8.1.9. Let y 2 F .x/ \ f˚ .0; Q/g for x 2 X and Q 2 L .V; Y /.
Then .x; y/ is a minimizer of (SP) and .Q; y/ is a maximizer of (DSP).
Proof. Suppose that .x; y/ is not a minimizer of (SP). Then there exist x 2 X
and y 2 F .x/ D ˚.x; 0/ such that y 2 y .C n f0g/, i.e., y 2 ˚.x; 0/ \
.˚ .0; Q/ .C n f0g// in contradiction to Theorem 8.1.7. Assuming .Q; y/ not
t
being a maximizer of (DSP) we get analogously a contradiction to Theorem 8.1.7.u
312 8 Duality
In order to give a reformulation of the dual problem (DSP) and to formulate a

certain regularity assumption for deriving strong duality statements we introduce
the minimal value map H W V Y [ fC1g defined by
H.v/ WD Min.[x2X ˚.x; v/; C /:
Definition 8.1.10. Let A Y be a given set.

(a) The set Min.A;C / is said to be externally stable if AnfC1g Min.A; C /CC .
(b) The set Max.A;C / is said to be externally stable if Anf1g Max.A; C /C .
This property is also called domination property (see Luc [402, 404]).
For some arbitrary v 2 V one says that H.v/ is externally stable if
Min.[x2X ˚.x; v/; C / is externally stable in the sense of Definition 8.1.10.
Subsequentially we suppose that H.v/ is externally stable for all v 2 V . Then it
holds H .Q/ D ˚ .0; Q/ and the dual problem (DSP) can be equivalently written
in the form
Max.[Q2L .V;Y / ŒH .Q/; C /: (DSP)
Furthermore, the following assertion is shown by Boţ, Grad, Wanka [73,

Lemma 7.1.14].
Lemma 8.1.11. It holds
Max.[Q2L .V;Y / Œ˚ .0; Q/; C / D H .0/:
For deriving strong duality assertions we need a certain kind of regularity

assumption:
Definition 8.1.12. The problem (SP) is called stable with respect to the pertur-
bation map ˚ if the minimal value map H is subdifferentiable at 0 in the sense of
Definition 8.1.1 (c).
It is important to mention that ˙1 … H.0/ under our assumptions.
The following theorem shows the existence of a common element y in the
objective values of the primal and dual set-valued problems. So it can be considered
as a strong duality assertion (see [73, Theorem 7.1.15]).
Theorem 8.1.13 (Strong Duality). If the set-valued problem (SP) is stable in the
sense of Definition 8.1.12, then for each minimizer .x; y/ of the problem (SP) there
exists a maximizer .Q; y/ of the problem (DSP).
Proof. Suppose that (SP) is stable. Then H is subdifferentiable at 0 because of
Definition 8.1.12. This is equivalent to H.0/ H .0/ taking into account
Proposition 8.1.2, as ˙1 … H.0/. Take a minimizer .x; y/ of the problem (SP)

such that y 2 H.0/. From Lemma 8.1.11 we get the existence of Q 2 L.V; Y / such
that y 2 ˚ .0; Q/. Moreover, Corollary 8.1.9 yields that .Q; y/ is a maximizer
of the dual problem (DSP). t
u
Remark 8.1.14. A comparison of different approaches to duality in set-valued
optimization is given in Sect. 8.4. Using the concept of supremum/infimum by
Tanino and Sawaragi (cf. [569]) of a set, defined in terms of the closure of the
set, Lalitha and Arora [360] introduce the notions of conjugate and biconjugate
maps as well as that of subgradients of a set-valued map. Conjugate duality results
are established for a set-valued optimization problem based on vector approach in
[360].
In the paper by Li, Chen and Wu [379] two conjugate dual problems are proposed
by considering the different perturbations for a set-valued vector optimization
problem with explicit constraints. Weak duality assertions, inclusion relations
between the image sets of dual problems, strong duality and stability criteria are
investigated.
8.1.2 Lagrange Duality for Set-Valued Optimization Problems

Based on Vector Approach
In this section we derive duality statements for solutions of set-valued optimization

problems in the sense of Definition 2.6.6. Let X , Y , Z be real locally convex
Hausdorff spaces, Y , Z be ordered by proper pointed closed convex cones C , K,
respectively, F W X Y and G W X Z. Under these assumptions we consider
the set-valued optimization problem (SP) with restrictions given in (2.49) (compare
Tasset [570]):
where M X is a convex set with M dom F \ dom G and
S WD fx 2 M j G.x/ \ .K/ ¤ ;g:
Here the “minimization” is to be understood in the sense of Definition 2.6.6, i.e.,

we are looking for quasi-weak minimizers of (SP). In this section we will show
duality assertions for the problem (SP) and a corresponding dual problem based on
the solution concept introduced in Definition 2.6.6.
Recall that the multifunction F W X Y is C –convex if and only if epi F is
convex (compare Sect. 2.3).
314 8 Duality
We introduce
E WD EC WD .F; G/.M / C .C K/
D f.y C u; z C v/ j x 2 M; y 2 F .x/; z 2 G.x/; u 2 C; v 2 Kg :
If x0 2 S and y0 2 F .x0 / then .y0 ; 0/ 2 E :

Lemma 8.1.15. If F is C –convex and G is K–convex, then E is convex .
Proof. Under the given assumptions .F; G/ W X Y Z defined by .F; G/.x/ WD
F .x/ G.x/ is C K–convex. Since E D epiC K .F; G/; we have that E is
convex.
Theorem 8.1.16. Let F be C –convex and G be K–convex. Assume that
(a) x0 2 S and y0 2 F .x0 /;
(b) .y0 ; 0/ … qri E ;
(c) 0 2 qi.G.M / C K/:
Then there exists y 2 C C n f0g; z 2 K C such that
8.y; z/ 2 .F; G/.M / W hy; y i C hz; z i

hy0 ; y i :
Moreover, z .G.x0 / \ .K// D f0g .

Proof. Because .y0 ; 0/ 2 E n qri E ; by (2.52), there exists .y ; z / 2 Y Z n
f.0; 0/g such that
hy C u; y i C hz C v; z i
hy0 ; y i 8.y; z/ 2 .F; G/.M /; 8.u; v/ 2 C K:
It follows that y 2 C C and z 2 K C : Assume that y D 0: Then hz; z i

0
for every z 2 G.M / C K; and so hz; z i
0 for every z 2 cl .RC .G.M / C K// :
Since 0 2 qi.G.M / C K/ we obtain that z D 0: This contradicts the fact that
.y ; z / ¤ .0; 0/: Hence y ¤ 0:
Let now z0 2 G.x0 / \ .K/: Then 0
hz0 ; z i D hy0 y0 ; y i C hz0 ; z i
0;
and so hz0 ; z i D 0: The proof is complete. t
u
Proposition 8.1.17. Assume that qi C ¤ ;, x0 2 S and y0 2 F .x0 /. Then the
assertion of Theorem 8.1.16 imply that y0 2 Min.F .S /; qi C /: In particular, under
the conditions of Theorem 8.1.16 one has y0 2 Min.F .S /; qi C /:
Proof. Assume that y 2 .F .S / y0 / \ . qi C /: Then there exists x 2 S and y 2
F .x/ such that y D y y0 2 . qi C /: By (2.53) we obtain that hy y0 ; y i < 0:
Since x 2 S; there exists z 2 G.x/ \ .K/: So we get the contradiction 0 >
hy y0 ; y i C hz; z i
0: Hence y0 2 Min.F .S /; qi C /. t
u
In the next theorem we show an assertion concerning the existence of linear
continuous operators related to our duality statements in Theorem 8.1.21.
Theorem 8.1.18. Let F be C –convex, G be K–convex, and qi C ¤ ;. Assume

that
(a) x0 2 S and y0 2 F .x0 /;
(b) .y0 ; 0/ … qri E ;
(c) 0 2 qi.G.M / C K/:
Then there exists T 2 LC WD LC .Z; Y / such that y0 2 .F C TG/.x0 /;
y0 2 Min ..F C TG/.M /; qi C /
and
T .G.x0 / \ .K// D f0g:
Proof. Taking into account Theorem 8.1.16 we get the existence of y 2 C C n f0g
and z 2 K C such that hy; y i C hz; z i
hy0 ; y i for all .y; z/ 2 .F G/.M /.
Since qi C ¤ ;; we have that C cl.qi C /; and so there exists y 2 qi C such that
hy; y i D 1: Let T W Z ! Y be defined by T .z/ WD hz; z i y: Clearly T 2 LC :
From Theorem 8.1.16 we have that T .G.x0 / \ .K// D f0g; and so y0 2 .F C
TG/.x0 / because G.x0 / \ .K/ ¤ ;: Moreover, for x 2 M; y 2 F .x/; z 2 G.x/
we have
hy C T z; y i D hy; y i C hz; z i hy; y i D hy; y i C hz; z i

hy0 ; y i :
Since y 2 C C n f0g; this shows that y C T z y0 … . qi C /; and so

..F C TG/.M / y0 / \ . qi C / D ;: Hence y0 2 Min ..F C TG/.M /; qi C /. u
t
Remark 8.1.19. In comparison with the corresponding result by Tasset [570, The-
orem 3.7] we do not assume y0 2 Min.F .S /; qi C / in Theorem 8.1.18 (compare
with Proposition 8.1.17).
In order to derive Lagrange duality assertions we introduce the set-valued
Lagrangian L W M LC Y for (SP) with restrictions given in (2.49) by
L.x; T / WD F .x/ C TG.x/:
Furthermore, we introduce the set-valued map W LC Y by
.T / WD Min.L.M; T /; qi C / D fy 2 L.M; T / j .L.M; T / y/ \ . qi C / D ;g:
Using these notions we introduce a set-valued dual problem to (SP) given by
maximize .T / subject to T 2 LC : (SD)
In (SD) the “maximization” is to understand in the following sense: We are

looking for a quasi-weak maximizer, i.e., for .T0 ; y0 / 2 graph with
. .LC / y0 / \ qi C D ;: (8.2)
316 8 Duality
Now, we formulate a weak duality assertion.

Theorem 8.1.20 (Weak Duality). For every y 2 .LC / it holds
.F .S / y/ \ . qi C / D ;
or equivalently, .F .S / .LC // \ . qi C / D ;:
Proof. Take x0 2 S and y 2 .T / D Min ..F C TG/.M /; qi C / with T 2 LC .
Because G.x0 / \ .K/ ¤ ; and T 2 LC ; we have that
F .x0 / y F .x0 / C TG.x0 / C C y .F C TG/.M / C C y:
Since C C qi C D qi C and ..F C TG/.M / y/ \ . qi C / D ;; we obtain that

.F .x0 / y/ \ . qi C / D ;: The conclusion follows. t
u
Theorem 8.1.21 (Duality Result). Assume that qi C ¤ ; and the hypotheses of
Theorem 8.1.16 hold. Then
y0 2 Min .F .S /; qi C / ” y0 2 Max . .LC /; qi C / :
Proof. We have seen in Proposition 8.1.17 that under the present conditions
y0 2 Min .F .S /; qi C / : We have to show that y0 2 Max . .LC /; qi C / : We
know already that y0 2 .LC /: Moreover, from Theorem 8.1.20 we have that
. .LC / F .S // \ qi C D ;: In particular, . .LC / y0 / \ qi C D ;; and so
y0 2 Max . .LC /; qi C /. t
u
Furthermore, we will show a Lagrange multiplier rule for Henig global proper
minimizer of (SP) in the sense of Definition 2.6.3. First, we recall some notations.
Let X be a separated locally convex space and K X a convex cone. The convex
set B K is a base of K if 0 … cl B and K D RC B: Then there exists a closed
convex and symmetric neighborhood U of 0 such that .2U /\B D ;: Let CU .B/ WD
RC .U C B/: Clearly, CU .B/ is a pointed convex cone with non empty interior;
moreover, K n f0g int CU .B/: Indeed, let x 2 K n f0g: Then x D tb for some
t > 0 and b 2 B: It follows that x 2 t.B C int U / CU .B/: The claim follows
because t.B C int U / is open.
Set
C .B/ WD fx 2 X j inf x .B/ > 0g:
The following result is shown by Gong [213].

Proposition 8.1.22. Let K be a convex cone with a base B:
(a) For any open balanced convex neighborhood U of zero we have
.CU .B//C n f0g C .B/:
(b) For any x 2 C .B/, there exists an open convex balanced neighborhood U
of zero such that x 2 .CU .B//C :
8.2 Duality Assertions for Set-Valued Problems Based on Set Approach 317
Proof. (a) is immediate. For (b) take x 2 C .B/I then r WD inf x .B/ > 0:
Consider U WD fx 2 X j jhx; x ij r=2g: Then for x 2 CU .B/ we have that
x D t.b C u/ with t
0; b 2 B; u 2 U; and so hx; x i D t .hb; x i C hu; x i/
t.r r=2/ D tr=2

0: Hence x 2 .CU .B//C . t
u
Theorem 8.1.23. Let F be C –convex and G be K–convex. If
(a) x0 2 S and y0 2 F .x0 /;
(b) there exists a symmetric convex neighborhood U of 0 in Y such that .2U /\B D
; and .y0 ; 0/ … qri EU0 , where EU0 denotes the set .F; G/.M / C .CU .B/; K/,
(c) 0 2 qi.G.M / C K/,
then .x0 ; y0 / is a Henig minimizer of (SP) with restrictions given by (2.49) and
(d) there exist y 2 C .B/; z 2 K C such that z .G.x0 / \ .K// D f0g and
hy y0 ; y i C hz; z i
0 8.y; z/ 2 .F; G/.M /: (8.3)
Conversely, if (a) and (d) hold then .x0 ; y0 / is a Henig global proper minimizer
of (SP) in the sense of Definition 2.6.3 and (b) holds.
Proof. Assume that (a), (b) and (c) hold. Applying Theorem 8.1.16 for C
replaced by CU .B/; there exist y 2 .CU .B// n f0g; z 2 K C such that
z .G.x0 / \ .K// D f0g and (8.3) holds. Moreover, .x0 ; y0 / is a quasi-weak
minimal solution of (SP) w.r.t. CU .B/; and so it is a Henig global proper minimizer
of (SP) in the sense of Definition 2.6.3. By Proposition 8.1.22 we have that
y 2 C .B/:
Conversely, assume that (a) and (d) hold. By Proposition 8.1.22 there exists
a symmetric convex neighborhood U of 0 in Y such that .2U / \ B D ; and
y 2 .CU .B// n f0g: It follows that .y0 ; 0/ … qri EU0 ; and so (b) holds. Using
Proposition 8.1.17 we have that .x0 ; y0 / is a quasi-weak minimizer of (SP) w.r.t.
CU .B/; and so it is a Henig global proper minimizer of (SP) in the sense of
Definition 2.6.3. t
u

on Set Approach
In this section we present Lagrange duality statements going back to Kuroiwa [351].
We consider set-valued optimization problems, where the solution concept is given
by the relation lC (see Definition 2.6.9):
A lC B ” A C C B ” 8b 2 B; 9a 2 A such that a C b
for a proper closed convex and pointed cone C Y and arbitrarily chosen sets
A; B Y .
318 8 Duality
Let X be a linear topological space, .Y; C /, .Z; K / linear topological spaces

with the proper pointed closed convex ordering cones C and K, respectively, F W
X Y , and G W X Z, then the primal set-valued optimization problem is
given by
lC minimize F .x/; subject to G.x/ lK 0: (SP lC )
Here the minimization is to understand in the sense of Definition 2.6.19 with respect
to the preorder lC :
• An element x0 2 X is said to be an lK -type feasible point of (SP lC ) if
G.x0 / lK 0.
• An element x0 2 X is said to be an lC -type minimal solution of (SP lC ) if it
is lK -type feasible and if
F .x/ lC F .x0 / H) F .x0 / lC F .x/
for each lK -type feasible x.

It is easy to see that fx 2 X j G.x/ lK 0g dom G and this set coincides with
fx 2 X j G.x/ \ .K/ ¤ ;g, which is the usual inequality restriction in set-valued
optimization. So the problem (SP lC ) can be written as
lC minimize F .x/; subject to x 2 S; (SP lC )
where the feasible set is given by
S WD fx 2 X j G.x/ \ .K/ ¤ ;g:
Consider L .Z; Y / D fT W Z ! Y j T is linear and continuousg, LC .Z; Y / D

fT 2 L .Z; Y / j T .K/ C g, graph G D f.x; z/ 2 X Z j z 2 G.x/g. In order to
construct a dual problem to (SP lC ), we introduce ˚ W L .Z; Y / Y defined by
˚.T / WDlC MinfF .x/ C T .z/ j .x; z/ 2 graph Gg;
where “lC MinfF .x/ C T .z/ j .x; z/ 2 graph Gg” denotes the set of objective
function values F .x/ C T .z/ for minimal solutions .x; z/ 2 graph G with respect to
lC in the sense of Definition 2.6.19.
Now, we introduce a generalized Lagrange function corresponding to (SP lC ).
Definition 8.2.1. For x 2 X and z 2 Z and T 2 L .Z; Y /, let
L.x; z; T / WD F .x/ C T .z/:
This function L is said to be the Lagrange function corresponding to (SP lC ).

We will study saddle points of the Lagrange function L in the following sense:
Definition 8.2.2 (Saddle Points). A triple .x0 ; z0 ; T0 / 2 graph G LC .Z; Y / is
said to by an lC -type saddle point of L if the following two conditions .a/ and
.b/ are satisfied:
(a) L.x; z; T0 /lC L.x0 ; z0 ; T0 /, .x; z/ 2 graph G H) L.x0 ; z0 ; T0 /lC L.x; z; T0 /,
(b) L.x0 ; z0 ; T0 /lC L.x0 ; z0 ; T /; T 2 LC .Z; Y / H) L.x0 ; z0 ; T /lC L.x0 ; z0 ; T0 /.
Remark 8.2.3. The triple .x0 ; z0 ; T0 / satisfies condition .a/ in Definition 8.2.2 if
and only if L.x0 ; z0 ; T0 / 2 ˚.T0 /.
The dual problem to (SP lC ) is the following set-valued optimization problem
lC maximize ˚.T /; subject to T 2 LC .Z; Y /: (SD lC )
For this dual problem we use the following solution concept:

• An element T0 2 L .Z; Y / is said to be an lC -type feasible point of (SD lC )
if T0 2 LC .Z; Y / and ˚.T0 / ¤ ;.
• An element T0 2 L .Z; Y / is said to be an lC -type maximal solution of
(SD lC ) if it is lC -type feasible and if
˚.T0 / lC ˚.T / H) ˚.T / lC ˚.T0 /
for each lC -type feasible T .

Theorem 8.2.4 (Weak Duality Between (SP lC ) and (SD lC )). Let x0 be an
lK -type feasible element of (SPlC ), T1 an lC -type feasible element of (SDlC ),
and .x1 ; z1 / an element of graph G with F .x1 / C T1 .z1 / 2 ˚.T1 /. Then,
F .x0 / lC F .x1 / C T1 .z1 / ) F .x1 / C T1 .z1 / lC F .x0 /:
Proof. Suppose that
F .x0 / lC F .x1 / C T1 .z1 /; (8.4)
G.x0 / lK 0, .x1 ; z1 / 2 graph G and T1 2 LC .Z; Y /. Because of G.x0 / lK 0 we

can find z0 2 G.x0 / such that z0 2 K. So we get from (8.4)
F .x0 / C T1 .z0 / lC F .x1 / C T1 .z1 /:
Then, taking into account F .x1 / C T1 .z1 / 2 ˚.T1 / and Definition 2.6.19 we get
F .x1 / C T1 .z1 / lC F .x0 / C T1 .z0 / lC F .x0 /:
t
u
320 8 Duality
Theorem 8.2.5 (Saddle Point Theorem). Suppose that .Y; C /, .Z; K / are
linear topological spaces, C and K are proper pointed closed convex cones.
Additionally, assume that int C ¤ ;. If .x0 ; z0 ; T0 / 2 graph G LC .Z; Y / is an
lC -type saddle point of L, then it holds
(i) z0 2 K and T0 .z0 / D 0,
(ii) x0 is an lC -minimal solution of (SP lC ),
(iii) T0 is an lC -maximal solution of (SD lC ).
Proof. .i /: In order to show z0 2 K and T0 .z0 / D 0 we first suppose z0 … K.
By a separation argument there exists z0 2 K C such that hz0 ; z0 i > 0. Using
an element c 2 int C we define Tn 2 LC .Z; Y / by
Tn .z/ WD nhz0 ; zic; n D 1; 2;
for each z 2 Z. Then, we can choose an element n0 2 N such that T0 .z0 / 2

Tn .z0 / C for each n
n0 .
Taking into account Definition 8.2.2 (b), from L.x0 ; z0 ; T0 / lC
L.x0 ; z0 ; Tn / we have
L.x0 ; z0 ; Tn / lC L.x0 ; z0 ; T0 /;
hence Tn .z0 / 2 T0 .z0 / C for each n

n0 . This implies nhz0 ; z0 ic D 0 for
all n
n0 , a contradiction to hz0 ; z0 i > 0 and c 2 int C . So we get z0 2 K
which implies
T0 .z0 / 2 C: (8.5)
For the zero element O of LC .Z; Y / it holds
F .x0 / C T0 .z0 / lC F .x0 / C O.z0 / (8.6)
because of (8.5). From (8.6) and Definition 8.2.2 (b) we get
F .x0 / C O.z0 / lC F .x0 / C T0 .z0 /:
This yields T0 .z0 / 2 C such that we have T0 .z0 / D 0 because of (8.5) and C
is pointed.
.ii/: Consider an element x 2 X such that G.x/ lK 0 and
F .x/ lC F .x0 /: (8.7)
For an lK -type feasible point z of (SP lC ) it holds z 2 G.x/ \ .K/ such
that we get
L.x; z; T0 / D F .x/ C T0 .z/ lC F .x0 / C T0 .z0 / D L.x0 ; z0 ; T0 /:

Since .x0 ; y0 ; T0 / is a saddle point, the converse inequality holds. Since z 2

K and T0 .z0 / D 0 we get
F .x0 / lC F .x/ C T0 .z/ lC F .x/:
This implies together with (8.7) that x0 is an lC -minimal solution of

(SP lC ).
.iii/: Taking into account the definition of saddle points, we have L.x0 ; z0 ; T0 / 2
˚.T0 /. Consider T1 2 LC .Z; Y /, .x1 ; z1 / 2 graph G with L.x1 ; y1 ; T1 / 2
˚.T1 / such that
L.x0 ; z0 ; T0 / lC L.x1 ; z1 ; T1 /: (8.8)
From .i / we know T0 .z0 / D 0, such that we get F .x0 / lC F .x1 / C T1 .z1 /.
Furthermore, we have T1 .z0 / 2 C since z0 2 K and T1 2 LC .Z; Y /.
This yields
L.x0 ; z0 ; T1 / D F .x0 / C T1 .z0 / lC F .x1 / C T1 .z1 / D L.x1 ; z1 ; T1 /:
The converse relation is satisfied since L.x1 ; z1 ; T1 / 2 ˚.T1 / such that we

can conclude
L.x1 ; z1 ; T1 / D F .x1 / C T1 .z1 / lC F .x0 / C T1 .z0 / lC F .x0 / D L.x0 ; z0 ; T0 /:
This yields together with (8.8) that T0 is an lC -maximal solution of

(SD lC ). u
t
Corollary 8.2.6. Suppose that the assumptions of Theorem 8.2.5 are fulfilled. Then,
.x0 ; z0 ; T0 / 2 graph G LC .Z; Y / is an lC -type saddle point of L if and only if
the following conditions hold:
(i) L.x; z; T0 / lC L.x0 ; z0 ; T0 /, .x; z/ 2 graph G H) L.x0 ; z0 ; T0 / lC
L.x; z; T0 /,
(ii) z0 2 K and T0 .z0 / D 0.
Proof. First, suppose that .x0 ; z0 ; T0 / is an lC -type saddle point of L, then .i / and
.ii/ hold because of Theorem 8.2.5.
Furthermore, suppose that .x0 ; z0 ; T0 / satisfies the conditions .i / and .ii/.
Because of (i) it is sufficient to show that .b/ of Definition 8.2.2 is fulfilled. If
T 2 LC .Z; Y / such that
L.x0 ; z0 ; T0 /lC L.x0 ; z0 ; T /;
then it follows that 0 D T0 .z0 / 2 T .z0 / C since T0 .z0 / D 0 in (ii) and

T .z0 / D 0 because of z0 2 K and C is pointed, therefore L.x0 ; z0 ; T / D F .x0 / D
L.x0 ; z0 ; T0 /. This yields that .x0 ; z0 ; T0 / is an lC -type saddle point of L. t
u
322 8 Duality
Further results concerning Lagrange duality for set-valued optimization prob-

lems based on the set approach where the solution concept is given by Defini-
tion 2.6.9 are derived in the paper by Hernández, Rodríguez-Marín [251], compare
also Hernández, Rodríguez-Marín [249].

on Lattice Structure
In this section, duality theorems will be presented that are based on an consequent
usage of infimum and supremum (in the sense greatest lower and least upper bounds
with respect to a partial ordering) (see Sect. 2.6.3).
We consider set-valued optimization problems where the objective map takes its
values in F (see Definition 2.6.36) as well as in I (see Definition 2.6.37). In the
book by Rockafellar and Wets [499] set-valued maps F W Rn Rq are considered
but we formulate conjugate duality results only for F -valued maps, i.e., we assume
the upper closedness of the values. However, this is no loss of generality taking into
account that semi-continuity properties are not influenced by the closure operation
to the values of the map F (compare [499, Proposition 4.4, Definition 5.4]).
By the fact F and I being isomorphic and isotone (see Proposition 2.6.39) it is
clear that it is sufficient to prove the duality assertions just for one case, either for
the F -valued case or for the I -valued case. The advantage of the F -valued case
is that the operations, the ordering and the infimum and supremum have an easier
structure, which is beneficial for proofs. In this case we speak about set inclusion
problems (see Löhne [393, 394]). In Sect. 8.3.1 we show Fenchel duality for F -
valued optimization problems, especially that under the usual assumptions weak
as well as strong duality assertions can be obtained. Furthermore, in Sect. 8.3.2
we derive Lagrange duality assertions for the I -valued case. The advantage of
the I -valued case is that it is closely related to vector optimization problems (see
Sect. 15.1).
Let Y be a separated locally convex space ordered by a proper pointed closed
convex cone C Y such that ; ¤ int C ¤ Y and Y D Y [ f1g [ fC1g. Let
Y be the topological dual space and let X be a linear space.
In the subsequent considerations let F WD FC .Y / and I WD IC .Y /.
The main tool for the proofs of the following duality assertions is a scalarization
of the F -valued and I -valued problems depending on a parameter y 2 Y n f0g.
Let us start with some statements concerning a suitable scalarization technique.
The support function A W Y ! R with respect to A Y is defined as
usual by
A .y / WD .y j A/ WD sup hy ; yi ; (8.9)

y2A
8.3 Duality Assertions for Set-Valued Problems Based on Lattice Structure 323
where R is equipped with the sup-addition, that is, 1C.C1/ D C1C.1/ D

1. This ensures that the expression
8y 2 Y W .y j A C B/ D .y j A/ C .y j B/
is valid for all (not necessarily nonempty) sets A; B Y .
8.3.1 Conjugate Duality for F -Valued Problems
We will show conjugate duality assertions for optimization problems with a F -

valued objective function introduced in Sect. 2.6.3. Löhne [394, 395] has shown
a Fenchel duality theorem for optimization problems with a F -valued objective
function and being based on the order relation “set inclusion” and infimal / supremal
sets introduced by Tanino [566], Nieuwenhuis [442]) (compare Sect. 2.6.3). A more
general result for closed (but not necessarily upper closed) sets was recently
obtained in [393, 394].
The notions and solution concepts that we will use in this section are introduced
in Sect. 2.6.3. We consider the space .F ; / (see Definition 2.6.36), where the order
relation is given by
A 4F B W ” A B
for subsets A, B of F (compare Sect. 2.6.3). Furthermore, for deriving the duality
assertions we use the property that .F ; / is a complete lattice (Proposition 2.6.40).
In order to formulate the F -valued optimization problem we study an objective
map F W X ! F , i.e., the objective function values of F are subsets of the
hyperspace of upper closed sets introduced in Definition 2.6.36.
For simplicity we assume in this section X D X D Rn , Z D Z D Rm and
Y D Y D Rq . Suppose that C Rq is a proper pointed closed convex cone.
Remark 8.3.1. The following duality assertions can be extended to more general
spaces than X D Rn , Z D Rm and Y D Rq as long as the corresponding scalar
result, which is used in the proof, is valid in these spaces. Then, one usually has to
modify the constraint qualification.
For deriving a Fenchel dual problem associated to the primal F -valued problem
we introduce the conjugate of a F -valued map with respect to c 2 Y . In difference
to Definition 8.1.3, where a solution concept given by the vector approach is applied
in the definition of a conjugate, we use in the following definition the solution
concept in the complete lattice F .
Definition 8.3.2. Let c 2 Y . The function Fc W X ! F ˘ (where F ˘ is the space
of lower closed subsets of Y ), defined by
324 8 Duality
Fc .x / WD inf fF .x/ hx ; xi fcgg ;

x2X
is said to be the conjugate of F with respect to c 2 Y .

Remark 8.3.3. In order to avoid calculations in the space F ˘ of lower closed
subsets of Rq we will prefer to use the term Fc .x / 2 F rather than Fc .x / 2
F ˘ in the following. Nevertheless a calculus in F ˘ is possible in the same way as in
q q
F , if we replace the inf-addition in R by the sup-addition in R , C by C , and
by in the definition of the order relation. By a consequent usage of the notions in
F ˘ we could express the conjugate as Fc .x / D supx2X fhx ; xi fcg F .x/g ;
where “sup” now means the supremum in F ˘ . For more details on this kind of
duality, using the concept of oriented sets by Rockafellar [489], see [393, 394].
Corresponding conjugate duality statements for I -valued functions are shown by
Löhne [395, Theorem 3.5] in a similar way. This is quite natural because F -valued
and I -valued problems are isomorphic and isotone (compare Proposition 2.6.39).
For given F -valued maps F W X ! F and G W Z ! F , a linear map
A W X ! Z and a vector c 2 Y , let
P WX !F and Dc W Z ! F
be defined, respectively, by

P .x/ D F .x/ ˚ G.Ax/ and Dc .u / D Fc .AT u / ˚ Gc .u / :
We consider the following F -valued optimization problems (compare Proposi-

tion 2.6.40), the primal problem
[
PN WD inf P .x/ D cl P .x/; (SP F )
x2X
x2X
and the dual problem associated to (SP F ) for c 2 Y

\
DN c WD sup Dc .u / D Dc .u /: (SDF F )
u 2Z u 2Z
Note that the convexity of a function F W X ! F is equivalent to the convexity of

the graph of the corresponding set-valued map (compare Sect. 2.3). In particular, if
F and G are convex in problem (SP F ), the values PN and DN c are convex sets.
We define weak and strong duality in the usual way, i.e., by DN c PN (that is
DN c F PN ) and DN c D PN , respectively.
In the following theorem (see Löhne, Tammer [397]) we will show weak duality
for the problems (SP F ) and (SDF F ) as well as strong duality under
convexity and certain regularity assumptions. The next theorem will be used in
the proof of Theorem 15.1.4 in Sect. 15.1 for deriving Fenchel duality for vector
optimization problems and it is a special case of Löhne [394, Theorem 4.3]. .
Theorem 8.3.4 (Weak and Strong Duality for (SP F ) and (SDF F )). The
problems (SP F ) and (SDF F ) (with arbitrary c 2 Y ) satisfy the weak
duality inequality, i.e., DN c PN .
Furthermore, let F and G be convex, let c 2 int C and let the following regularity
assumption be fulfilled:
0 2 ri.dom G A dom F /:
Then we have strong duality, i.e., DN c D PN .

Proof. The weak duality is obvious from the definition. In order to prove the strong
duality assertion we use a scalarization method by the support function A W Rq !
R with respect to A Y given by (8.9). Note that for (not necessarily nonempty)
subsets A; B F we have A C B D A˚B and ˛A D ˛A for ˛ > 0. Moreover,
for all (not necessarily nonempty) sets A F we have inf A D infA2A A
(see e.g. [489, Corollary 16.5.1]).
It holds PN D infx2X P .x/ . By the extended real–valued functions FNy W
X R and N respectively, by FNy .x/ WD
Gy W Z R being defined,

N
y j F .x/ and Gy .u/ WD y j G.u/ this can be rewritten as a collection

of scalar optimization problems

˚
8y 2 Y W .y j PN / D inf FNy .x/ C GN y .Ax/ : (8.10)
x2X
The convexity of F and G implies the convexity of FNy and GN y , respectively.

Clearly, we have dom F D dom FNy and dom G D dom GN y , whence 0 2
ri.dom GN y A dom FNy /. A scalar duality theorem, for instance [69, Theo-
rem 3.3.5], yields that
n o
8y 2 Y W .y j PN / D sup FNy .AT u / GN y .u / ;
u 2Z
where the supremum is attained whenever .y j PN / is finite, i.e.,
8y 2 dom . j PN /; 9Nu 2 Z W .y j PN / D FNy .AT uN / GN y .Nu /:

(8.11)
Let y 2 dom . j PN / be arbitrarily given (hence pN ¤ Y ). Since ; ¤ PN ¤ Y and

PN D Cl C PN , we have dom . j PN / C . By the choice
˝ c 2 ˛int C , it follows that
hy ; ci < 0. Hence, there exists ˛y > 0 such that ˛y y ; c D 1. This can be
rewritten as
˝ ˛
8t 2 R W ˛y y j ft cg D ˛y y ; t c D t: (8.12)
326 8 Duality
For ˛ WD ˛y > 0 we have
.8.11/
˛ y j PN D ˛y j PN D FN˛y

.A uN / GN ˛y
T
u /
.N
˚ ˝ ˛ ˚
D inf AT uN ; x C FN˛y .x/ C inf hNu ; ui C GN ˛y .u/
x2X u2Z
˚ ˇ ˝ ˛ ˇ
D inf ˛y ˇ AT uN ; x fcg ˛y ˇF .x/
(8.12)
x2X
˚ ˇ ˇ
C inf ˛y ˇhNu ; ui fcg ˛y ˇG.u/
u2Z
ˇ
ˇ
ˇ
˚ ˝ T ˛ ˚
D ˛y ˇ inf A uN ; x fcg C F .x/ ˚ inf hNu ; uifcgCG.u/
x2X u2Z
ˇ
D ˛y ˇ Fc .AT uN / ˚ Gc .Nu / D ˛y jDc .Nu /

D ˛ y jDc .Nu / :
We deduce that
8y 2 dom . j PN /; 9Nu 2 Z W .y jDc .Nu // D .y j PN /: (8.13)
It follows that .y j DN c / .y j PN / for all y 2 Y , hence (DN c and PN being
closed and convex) DN c PN . By the weak duality inequality we obtain DN c D PN .u t
Remark 8.3.5. Note that we have not the usual dual attainment assertion, i.e., the
dual value DN c is not attained by a single element uN 2 Z . Instead we have
condition (8.13), which describes the present situation.
Remark 8.3.6. In Theorem 8.3.4 (and its conclusion Theorem 15.1.4) we suppose
the constraint qualification 0 2 ri.dom G A dom F /. In the proof we use this
condition in order to obtain the corresponding condition for the family of scalarized
problems in (8.10). If all these problems are polyhedral, the constraint qualification
can be replaced by dom G \ A dom F ¤ ;, compare e.g. [69, Corollary 5.1.9].
An comprehensive discussion of conjugate duality theory for problems where the
solution concept is based on lattice approach is given in the book by Löhne [395].
8.3.2 Lagrange Duality for I -Valued Problems
The results in this section are derived by Löhne and Tammer [398], Hernández,
Löhne, Rodríguez-Marín, Tammer [247] and Löhne [395].
First, we introduce the I -valued primal problem for that we will derive Lagrange
duality statements.
Throughout this section let Y be an extended locally convex space ordered by

a proper pointed closed convex cone C Y such that ; ¤ int C ¤ Y . Let X be a
linear space.
We put I WD IC .Y / and consider the space .I ; 4I / (see Definition 2.6.37),
where the order relation is given by
A 4I B W ” A 4 B W ” Cl C A Cl C B
for subsets A, B of I (compare Sect. 2.6.3). The property that .I ; 4I / is a

complete lattice (Proposition 2.6.40) will be used intensively for deriving duality
assertions.
In this section we derive Lagrange duality statements for I -valued problems
using the notations and results given in Sect. 2.6.3, especially the order relation
4WD4I .
Given S X a nonempty set and an I -valued map F W S ! I we consider the
following I -valued problem
P WD inf F .x/: (SP I )

x2S
The domain of the I -valued map F is given by Dom F D fx 2 S j F .x/ ¤ ; ,

F .x/ ¤ fC1gg. Note that we have
[
P D inf F .x/ D Inf F .x/;
x2S
x2S
S S
where Inf x2S F .x/ denotes the infimal set of x2S F .x/ (see Theorem 2.6.41).
We emphasize that problem (SP I ) is related with several well-known
problems as we see in Sect. 15.1. Indeed, if Y D R and C D f0g we obtain the
real case (4W D).
In this section we establish several Lagrangian duality problems in the framework
of I -valued problems and give weak and strong duality theorems. To do that we
need to define an I -valued Lagrange dual problem associated to (SP I ).
In a natural way we assign to (SP I ) a dual problem
D WD sup .u/ (SDL I )

u2U
where U is a nonempty set and W U ! I is the dual objective function that is

constructed in the following by a Lagrangian technique.
We say that (SP I ) and (SDL I ) satisfy weak duality if
D 4I P:
Furthermore we speak about strong duality if
D D P:
328 8 Duality
Taking into account that I is a complete lattice and Lemma 2.6.38 we can
rewrite the weak duality property as follows:
Lemma 8.3.7. Consider problems (SP I ) and (SDL I ). The following
conditions are equivalent
(i) D 4I P
(ii) For any x 2 S and u 2 U ,
.u/ 4I F .x/:
If, in addition, F .x/ ¤ 1 and .x/ ¤ 1 for all x 2 S and u 2 U , the above
conditions are equivalent to
(iii) For any x 2 S and u 2 U ,
.F .x/ C int C / \ .u/ D ;:
To establish the strong duality theorem we need the notion of convex I -valued
objective function F W X ! I .
Definition 8.3.8. Let X be real a linear space. The function F W X ! I is convex
if for all t 2 Œ0; 1 and all x 1 ; x 2 2 X it holds

F tx 1 C .1 t/x 2 4 t ˇ F .x 1 / ˚ .1 t/ ˇ F .x 2 /:
We next recall the notion of cone-convexity (compare Sect. 2.3) which is used for
the set-valued constraints: Let X; Z be real linear spaces and let K Z be a proper
closed convex cone. A set–valued map G W X Z is said to be K–convex if

8x 1 ; x 2 2 X; 8t 2 Œ0; 1 W G tx 1 C .1 t/x 2 C K tG.x 1 / C .1 t/G.x 2 /:
Given a set-valued map GW X Z we denote by GI W X ! I the

corresponding I -valued map associated to G as follows: GI .x/ D Inf G.x/ for
x 2 X.
Lemma 8.3.9. Let GW X Z. If G is K-convex then GI is convex.
Proof. Let x 1 ; x 2 2 X and t 2 Œ0; 1. The conclusion follows from

GI tx 1 C .1 t/x 2 4 t ˇ GI .x 1 / ˚ .1 t/ ˇ GI .x 2 /
is equivalent to
Cl C .t .G.x 1 // C .1 t/ .G.x 2 /// Cl C G.tx 1 C .1 t/x 2 /:
t
u
In the following we will derive duality assertions for optimization problems

with I -valued objective function and set-valued constraints. First we formulate the
corresponding assumptions and notations.
Let X be a linear space and let Z be a separated locally convex space. Moreover,
let hZ; Z i be a dual pair. Let F W X ! I , let G W X Z be a set-valued map
with dom G D X and let K Z be a proper closed convex cone. We denote by
L .Z; Y / the set of all linear continuous operators from Z to Y and by LC .Z; Y / D
fT 2 L .Z; Y / j T .K/ C g the subset of all positive operators.
From now on, we suppose that the restrictions in (SP I ) are given in the form
S WD fx 2 X j G.x/ \ K ¤ ;g : (8.14)
Constraints of this type have been considered by many authors, such as Borwein
[56]; Corley [112]; Jahn [290]; Luc [402]; Götz and Jahn [218]; Jahn [293]; Crespi,
Ginchev and Rocca [116]; Boţ and Wanka [52]; Hernández, Löhne, Rodríguez-
Marín, Tammer [248].
Then the I -valued primal problem (SP I ) is rewritten as follows:
PN WD inf F .x/; S WD fx 2 X j G.x/ \ K ¤ ;g : (SP I )

x2S
We first recall a scalar Lagrange duality result with this type of set-valued
constraints. The scalar result is later used in the proof of the strong duality for the
general problem. We consider a scalar problem with an objective f W X ! R:
pO WD inf f .x/; S WD fx 2 X j G.x/ \ K ¤ ;g : O

(P)
x2S
O is defined as
The Lagrangian map associated to (P)
L W X Z ! R; L.x; u / D f .x/ C inf hu ; ui : (8.15)

u2G.x/CK
The dual objective is
W Z ! R; .u / WD inf L.x; u /

x2X
and the dual problem is
dO WD sup .u /: O
(D)
u 2Z
Note that we have

(

inf ff .x/ C inf hu ; uig if u 2 K C
inf L.x; u / D x2X u2G.x/
x2X 1 else .
330 8 Duality
Thus we obtain the following version of the dual problem
dO WD sup .u /: O 0)
(D
u 2K C
Of course it holds weak duality, i.e., dO p.

O Under certain convexity assumptions
and some constraint qualification we get the following strong duality assertion,
which we prove in the same way as [69, Proposition 4.3.5].
O and (D)).
Theorem 8.3.10 (Strong Duality for the Scalar Problems (P) O Let f W
X ! R be convex, let G W X Z be K-convex and let
G.dom f / \ int K ¤ ;: (8.16)
Then, we have strong duality, i.e., dO D p.

O
Proof. The value function is defined by
vWZ!RW v.u/ WD inf ff .x/j G.x/ \ .fug K/ ¤ ;g
O For the conjugate v W Z ! R of v

It is easily shown to be convex and v.0/ D p.
it holds
v .u / D sup f hu ; ui v.u/j u 2 Zg

D sup f hu ; ui f .x/j G.x/ \ .fug K/ ¤ ;; u 2 Z; x 2 dom f g
D sup f hu ; ui f .x/j u 2 G.x/ C K; x 2 dom f g
D inf L.x; u / D .u /:
x2X
It follows v .0/ D dO . We next show that v is lower semi-continuous at 0 (even

continuous). Indeed, by (8.16) there is some xN 2 dom f and some zN 2 int K
such that zN 2 G.x/.
N There exists some neighborhood V of 0 such that fNzg V
int K. Now it follows that f .x/N is an upper bound of v on V . This implies that v
is continuous at 0, see e.g. [172, Lemma 2.1].
If pO D 1, we get pO D dO D 1 from the weak duality. So we can assume
that pO D v.0/ is finite. Thus we have v.0/ D cl v.0/, compare [490, Theorem 4].
By the bipolar theorem, e.g. [172, Proposition 4.1], we have cl v D v . This yields
pO D v.0/ D v .0/ D dO . t
u
Our main aim is to derive duality results for I -valued optimization problems
using the ideas of Sect. 2.6.3. To this end we develop an I -valued version of
Theorem 8.3.10.
We will use the following scalarization method of the I -valued problem
depending on a parameter y 2 C n f0g. For W Y ! R given by (8.9)
and A 2 I , we set
'A .y / WD '.y j A/ WD .y j Cl C A/: (8.17)
For fixed y we get by (8.17) a functional from I to R. The addition, the

multiplication by positive reals, the infimum and the supremum for extended
real-valued functions are defined pointwise. Here we use the inf-addition, i.e.,
1 C 1 D C1 1 D C1.
For fixed A 2 I , we consider 'A to be a function from C n f0g into R, that is,
'A W C n f0g ! R:
For some 2 R we write 'A whenever 'A .y / D for all y 2 C n f0g.

In the next theorem we collect several properties of the scalarization by means of
the functional (8.17) that are useful for the proof of the duality statements (compare
[395, Theorem 1.65]).
Theorem 8.3.11. For any A; B 2 I and ˛ > 0 it holds
(i) ŒCl C A D conv Cl C
A ^ 'A 1 ” A D f1g,

(ii) 'A C1 ” 9y 2 C n f0g W 'A .y / D C1 ” A D fC1g,
(iii) A 4 B H) 'A 'B ,
(iv) Cl C A D conv Cl C A ^ 'A 'B H) A 4 B,
(v) 'A˚B D 'A C 'B ,
(vi) ˛ 'A D '˛ˇA .
For a nonempty subset A I we have
(vii) 'inf A D inf 'A
A2A
(viii) 'sup A
sup 'A
A2A
Proof. (i) As Cl C f1g D Y , we get 'f1g D .y j Y / D 1 for all y 2

C n f0g. On the other hand, 'A 1 implies that
C1 D .y j Cl C A/ C .y j C / D .y j Cl C A C C / D .y j Cl C A/
for all y 2 Y n f0g. It follows Cl C A D cl conv Cl C A D Y and hence A D

f1g.
(ii) It holds Cl C fC1g D ; and so A D fC1g implies 'A .y / D .y j;/ D
C1 for all y 2 C n f0g. If .y jCl C A/ D 1 for some y 2 C n f0g, then
Cl C A D ;. This implies A D fC1g.
(iii) Let A 4 B. If A D f1g or B D fC1g, from (i) and (ii) we get 'A
'B . Otherwise we get Cl C A Cl C B and hence 'A .y / D .y jCl C A/
.y jCl C B/ D 'B .y / for all y 2 Y .
(iv) Let 'A 'B , i.e., for all y 2 Y it holds .y jCl C A/ .y jCl C B/.
As Cl C A is convex and closed we get Cl C A cl conv Cl C B Cl C B and so
A 4 B.
In order to prove the statements (v) to (vii), let y 2 C nf0g be arbitrarily given.
332 8 Duality
(v) If A or B equals fC1g then A ˚ B D fC1g and the statement follows as

Cl C fC1g D ;. If A and B are not fC1g but one of them or both equal f1g then
the result follows from the fact Cl C f1g D X . So we can now assume A; B Y
and in this case we have Cl C A D cl .ACC /. It holds 'ACB .y / D .y j cl .AC
B C C // D .y j cl .A C C // .y j cl .B C C // D 'A .y / C 'B .y /.
(vi) If A D fC1g, then ˛A D fC1g and so ˛'A .y / D '˛ˇA .y / D C1.
If A D f1g, then ˛A D f1g and so ˛'A .y / D '˛ˇA .y / D 1. If A Y ,
then we have ˛ 'A .y / D ˛.y j cl .A C C // D .y j cl .˛A C C // D
'˛ˇA .y /.
(vii) It remains to show the statement for the case fC1g 62 A , because omitting
fC1g does not change anything. If f1g 2 A the equality can be easily shown.
So let A Y for all A 2 A . It holds
ˇ ! ˇ !
ˇ [ ˇ [
ˇ ˇ
'inf A .y / D y ˇ Cl C A D y ˇ cl .A C C /
A2A A2A

D inf y j cl .A C C / D inf 'A .y /:

A2A A2A
(viii) We have sup A < A and hence 'sup A .y /

'A .y / for all A 2 A . Taking
the supremum the statement follows. t
u
The following example shows that statement (viii) in Theorem 8.3.11 does not
hold with equality:
˚
˚ (SeeT [393, Example 1.3.6]). Let C WD ˚ RCT,A1 WD .0; 1/ C
2 T
Example 8.3.12
bd C , A2 D .1; 0/ C bd C , then sup fA1 ; A2 g D .1; 1/ C bd C . For y D
.1; 1/T we get 'A1 .y / D 'A2 .y / D 1 but 'supfA1 ;A2 g .y / D 2.
Coming back to the I -valued optimization problem (SP I ) we introduce a
corresponding Lagrangian map. For a given fixed vector c 2 int C , the Lagrangian
map of problem (SP I ) is defined by
˚
Lc W X Z ! I ; Lc .x; u / WD F .x/ ˚ inf hu ; ui fcg C bd C :
u2G.x/CK
(8.18)
In the special case q D 1, C D RC , c D 1, the Lagrangian coincides with the

Lagrangian of the scalar problem PO in (8.15). For every choice of c 2 int C we
have a different Lagrangian map and a different corresponding dual problem, but
we show that weak duality and strong duality hold for all these problems.
The scalar counterpart of the following result (compare [398], [395, Proposi-
tion 3.23]) is well known.
Proposition 8.3.13. For all x 2 S and all c 2 int C it holds
sup Lc .x; u / D F .x/:

u 2Z
Proof. For arbitrary c 2 int C it holds

˚
sup Lc .x; u / 4 F .x/ ˚ sup inf hu ; ui fcg C bd C
u 2Z u 2Z u2G.x/CK

˚
D F .x/ ˚ sup inf hu ; zi fcg C bd C
u 2Z z2G.x/

˚
C inf hu ; d i fcg C bd C
d 2K
Taking into account

˚ bd C if u 2 K C
inf hu ; d i fcg C bd C D
d 2K f1g otherwise
we get
˚
sup Lc .x; u / 4 F .x/ ˚ sup inf hu ; zi fcg C bd C
u 2Z u 2K C z2G.x/
˚
4 F .x/ ˚ sup inf hu ; zi fcg C bd C
u 2K C z2G.x/\K
˚
4 F .x/ ˚ inf sup hu ; zi fcg C bd C
z2G.x/\K u 2K C
4 F .x/:
Since L.x; 0/ D F .x/, it follows that sup Lc .x; u / D F .x/. t

u
u 2Z
We next define the I -valued Lagrange dual problem associated to (SP I ).

The dual objective function (with respect to c 2 int C ) is defined by
c W Z ! I ; c .u / WD inf Lc .x; u /

x2X
and the I -valued dual problem (with respect to c 2 int C ) associated to (SP I )
is defined by
DN c WD sup c .u /: (SDcL I )

u 2Z
Similarly to the real case we have

(

inf fF .x/ C inf hu ; ui fcg C bd C g if u 2 K C
c .u / D inf Lc .x; u / D x2X u2G.x/
x2X f1g otherwise.
334 8 Duality
Therefore we obtain the following equivalent formulation of the dual prob-

lem (SDcL I )
DN c WD sup c .u /: (SD0c

L I )
u 2K C
Theorem 8.3.14 (Weak Duality for (SP I ) and (SDcL I )). For every c 2
int C , the problems (SP I ) and (SDcL I ) satisfy the weak duality inequality,
i.e., DN c 4 PN .
Proof. Since I is a complete lattice, we immediately have
sup inf Lc .x; u / 4 inf sup Lc .x; u /; (8.19)

u 2Z x2X x2X u 2Z
(even if Lc would be replaced by an arbitrary function from X Z into I ). By

Proposition 8.3.13 we know that inf sup Lc .x; u / 4 PN . t
u
x2X u 2Z
It follows a strong duality assertion under a certain constraint qualification

(compare [398] and [395, Theorem 3.26]).
Theorem 8.3.15 (Strong Duality for (SP I ) and (SDcL I )). Let F be
convex and let G be K-convex, let
G.dom F / \ int K ¤ ;; (8.20)
and let c 2 int C . Then, we have strong duality, i.e., DN c D PN .

Proof. Of course, we have PN ¤ fC1g (by (8.20)). If PN D f1g we obtain
PN D DN c from the weak duality. So we can assume PN Y . As PN is self-infimal, it
is nonempty.
Let y 2 C such that hc; y i D 1. It holds
ˇ
ˇ ˇ
' y ˇˇ hu ; ui fcg C bd C D ' y ˇ hu ; ui fcg C bd C
Th. 8.3.11 (vii)
inf inf
u2G.x/CK u2G.x/CK
ˇ
D inf y ˇ hu ; ui fcg C C
u2G.x/CK
D inf hu ; ui :
u2G.x/CK
It follows that
ˇ
ˇ
'.y jPN / D ' y ˇˇ inf F .x/
G.x/\K¤;
ˇ
D ' y ˇ F .x/
Th. 8.3.11 (vii)
inf
G.x/\K¤;

ˇ
D ' y ˇ F .x/ C
Th. 8.3.10
sup inf inf hu ; ui
u 2Z x2X u2G.x/CK

ˇ
D sup inf ' y ˇ F .x/
u 2Z x2X
ˇ
ˇ
C' y ˇˇ inf hu ; ui fcg C bd C
u2G.x/CK
ˇ
D sup ' y ˇ c .u /
Th. 8.3.11 (v, vii)
u 2Z
ˇ ˇ
' y ˇ sup c .u / D ' y ˇ DN c :
Th. 8.3.11 (viii)

u 2Z
The same is true for arbitrary y 2 C . If y 62 C , we have
'.y j PN / D .y j Cl C PN / D .y j PN CC / D .y j PN /.y j C / D 1:

Therefore we have ' y j PN ' y j DN c for all y 2 Y . As Cl C PN is convex,
Theorem 8.3.11 (iv) yields PN 4 DN c . From the weak duality we get equality. t
u
Note that strong duality (i.e., PN D DN c ) implies that (8.19) is satisfied with
equality.
Proposition 8.3.13 clarifies the relation between the problem (SP I ) and the
Lagrangian only for feasible points x 2 S . It remains the question what happens
if x is not feasible. From special cases in scalar optimization, we expect that
sup Lc .x; u / D fC1g. This can be shown under additional assumptions to the
u 2Z
constraints. Note that the following proposition is not used in the proof of the duality
theorem. A similar result was shown in [393]. The next result goes back to [398] and
[395, Proposition 3.24].
Proposition 8.3.16. Let F W X ! I be a proper function. Assume that the set
G.x/ C K is convex and closed for all x 2 X . Then for all c 2 int C it holds

F .x/ if x2S
sup Lc .x; u / D
u 2Z fC1g else.
Proof. The case x 2 S follows from Proposition 8.3.13. Let x 62 S , i.e., G.x/ \
K D ;. Thus, 0 62 G.x/ C K. Applying the well-known separation result [32,
Theorem 1.15], yields the existence of some uN 2 K C n f0g such that
inf hNu ; ui D inf hNu ; ui sup hNu ; ui > 0: (8.21)

u2G.x/CK u2G.x/ u2K
336 8 Duality
Choose some y 2 C such that c T y D 1. It holds

ˇ
ˇ
ˇ
Th. 8.3.11 (x)
' y ˇ sup Lc .x; u /
sup ' y j Lc .x; u /
u 2Z u 2Z

sup ' y j Lc .x; t uN /
t >0

ˇ
D sup ' y ˇ F .x/ C t inf
hNu ; ui D C1
t >0 u2G.x/CK
Theorem 8.3.11 (ii) yields that sup Lc .x; u / D fC1g. t

u
u 2Z
The next example shows that the assumption G.x/CK is closed in the preceding
proposition cannot be omitted, not even in the scalar case. However, this assumption
is fulfilled in many important special cases, such as for vector–valued or compact–
valued functions G.
o F W R ! I , F .x/ Inf f0g ˚D bd C . We set A WD

2
nExample 8.3.17. Let
y 2 R2C j y2
y11 , G.x/ WD fxg C A and K WD u 2 R2 j u1
0 . Then
˚
K n f0g D u 2 R2 j u1 < 0; u2 D 0 . An easy computation shows

F .0/ if u 2 K C
Lc .0; u / WD :
f1g else.
It follows that sup Lc .0; u / D F .0/ ¤ fC1g, G.0/ C K is not closed but
u 2Z
G.0/ \ K D ;, i.e., 0 62 S .
Now we study another dual problem associated to (SP I ) where the dual
variables are operators instead of vectors as above. The following results are given
by Hernández, Löhne, Rodríguez-Marín, Tammer [247]. Exactly, we consider the
formulation of the Lagrangian map more used in the Literature (for instance, see
[112, 387, 402]) but in the framework of I -valued problems. We obtain the weak
duality from the completeness of the lattice and prove the strong duality applying
Theorem 8.3.15. S
Given T 2 L and ; ¤ A Z we write T .A/ D a2A T .a/.
Now the Lagrangian map L W X L ! I is defined by
L.x; T / WD F .x/ ˚ inf .T .u/ C bd C / ; (8.22)

u2G.x/CK
the dual objective function W L ! I is
.T / WD inf L.x; T /; (8.23)

x2X
and the dual problem is defined by
DQ WD sup .T /: (SDTL I )
T 2LC
Again we can similarly prove the following result like in Proposition 8.3.13 with
respect to the above Lagrangian.
Proposition 8.3.18. For all x 2 S it holds
sup L.x; T / D F .x/:

T 2LC
Theorem 8.3.19 (Weak Duality for (SP I ) and (SDTL I )). Prob-
lems (SP I ) and (SDTL I ) satisfy the weak duality inequality, that is,
DQ 4 PN .
Proof. Since I is a complete lattice (compare Proposition 2.6.40), we immediately
have
sup inf L.x; T / 4 inf sup L.x; T /; (8.24)

T 2LC x2X x2X T 2L
C
By Proposition 8.3.18 we know that inf sup L.x; T / 4 PN . t

u
x2X T 2L
C
To prove the strong duality from Theorem 8.3.15 we establish another formula-
tion for DN c (see (SD0c
L I ), c 2 int C ) in terms of linear operators as follows. Let
Lc be a subset of L defined by
˚
Lc WD T 2 L j T D hu ; i c for some u 2 K C :
Is is easy to check that Lc is a subset of LC , i.e.,
Lc LC ; (8.25)
and is isomorphic to K C Z . The dual problem (SD0c

L I ) can be equivalently
expressed by
DN c WD sup .T /; (8.26)
T 2Lc
where the duality function is defined in (8.23).

Remark 8.3.20. It is standard in the literature to consider the larger class LC
instead of Lc . Nevertheless, in [402] Lc has also been considered to establish
strong duality by using properly efficient solutions. On the other hand, we point out
that, in general, the Lagrangian multipliers are always given via elements T 2 Lc
338 8 Duality
for any c 2 int C see, for instance, [112, Theorem 4.1], [387, Theorem 3], [402].
Therefore, according to the obtained results the dual problem defined in such papers
could define by using Lc instead LC . So, we can consider vectors as dual variables
instead of operators by virtue of (8.26).
Strong duality between (SP I ) and (SDTL I ) can be easily derived from
previous results.
Theorem 8.3.21 (Strong Duality Between (SP I ) and (SDTL I )). Let F
be convex and G be K-convex. Suppose that
G.dom F / \ int K ¤ ;;
then we have strong duality, i.e., DQ D PN .

Proof. By Theorems 8.3.19 and 8.3.15 we obtain strong duality for (SP I )
and (SDTL I ) taking into account (8.25) and (8.26). u
t
The proof of the following result is analogous to that in Proposition 8.3.16 taking
into account that given u 2 K C nf0g and y 2 C there always exists T 2 LC
such that T y D u :
Proposition 8.3.22. If for some x 2 dom F \ dom G the set G.x/ C K is closed
and convex, then

F .x/ if x2S
sup L.x; T / D
T 2LC fC1g else.
8.4 Comparison of Different Approaches to Duality

in Set-Valued Optimization
In order to derive duality assertions in set-valued optimization one has the possibil-
ities to use an approach via conjugates, via Lagrangian technique or an axiomatic
approach. Conjugate duality statements, based on different types of perturbation of
the original problem, have been derived by Tanino and Sawaragi [569], Sawaragi,
Nakayama and Tanino [521], Löhne [394, 395], Boţ, Grad and Wanka [73], and
others. Lagrange duality for set-valued problems has been studied, for instance, by
Luc [402], Ha [226], Hernández and Rodríguez-Marín [249, 251], Li, Chen and
Wu [379], Löhne [395] and Hamel and Löhne [236]. An axiomatic approach was
given by Luc [402]. Furthermore, duality assertions can be developed for different
solution concepts, this means for the vector approach (cf. Luc [402], Boţ, Grad
and Wanka [73], Li, Chen and Wu [379]), for the set-approach (cf. Kuroiwa [351],
Hernández and Rodríguez-Marín [251]) and by using supremal and infimal sets (see
Nieuwenhuis [442], Tanino [563,566]) and/or infimum and supremum in a complete
lattice for the lattice approach (cf. Tanino [566], Song [535,536], Löhne [393–395],
Lalitha and Arora [360]).
8.4 Comparison of Different Approaches to Duality in Set-Valued Optimization 339
Another important difference in constructing the dual problems concerns the type
of dual variables, which can be vectors like in the scalar case (cf. Löhne [394, 395],
Boţ, Grad and Wanka [73], Section 7.1.3), extended vectors (cf. Hamel and Löhne
[236]) or operators (cf. Corley [112], Luc [402], Tanino [566], Boţ, Grad and Wanka
[73], Section 7.1.2, Hernández and Rodríguez-Marín [251], Lalitha and Arora [360],
Li, Chen and Wu [379]).
In the mentioned papers one can observe that it is very easy to derive weak
duality statements without additional assumptions. In order to get strong duality one
assumes additionally convexity and certain regularity assumptions. An important
approach for the formulation of regularity assumptions is the stability of the
primal set-valued problem. In the literature, stability of the primal set-valued
problem is formulated using the subdifferential of the minimal value map (cf.
Tanino and Sawaragi [569], Boţ, Grad and Wanka [73]). In this section we will
introduce a subdifferential notion based on infimal sets where subgradients are
vectors. A corresponding stability notion is used to prove strong duality statements.
Furthermore, we will discuss the space of dual variables and explain the relations
between some other approaches from the literature. Subdifferentials notions for
vector and set-valued maps have been investigated by many authors, see e.g., by
Tanino [566], Jahn [293], Boţ, Grad and Wanka [73].
Lagrange duality theory is an important tool in optimization and there are many
approaches to a corresponding theory for vector optimization problems, see e.g.
Corley [111,112,509], Boţ and Wanka [52], Li and Chen [387], Jahn [293], Göpfert,
Tammer, Riahi and Zălinescu [214], Boţ, Grad and Wanka [73] and references
therein.
The following results are given in the paper by Hernández, Löhne, Rodríguez-
Marín, Tammer [248].
In order to compare different approaches to duality in set-valued optimization we
discuss Lagrange duality assertions from Sect. 8.3.2 for optimization problems with
I -valued objective function and set-valued constraints using a notation adapted to
[566].
Assume that .Y; / is a partially ordered linear topological space, where the order
is induced by a proper pointed closed convex cone C satisfying ; ¤ int C ¤ Y .
Let X be a linear space and let Z be a separated locally convex space. Moreover,
let hZ; Z i be a dual pair. Let F W X Y and G W X Z be set-valued
maps. We set Dom F WD fx 2 X jF .x/ ¤ ;; F .x/ ¤ fC1gg and dom G WD
fx 2 X j G.x/ ¤ ;g. Let K Z be a proper closed convex cone with nonempty
interior. We denote by K the negative polar cone of K.
8.4.1 Lagrange Duality
In this section we use the notation introduced in Sect. 2.6.3, especially the infimal
set (Definition 2.6.31) and the hyperspace of self-infimal sets .I ; ˚I ; ˇI ; 4I /
340 8 Duality
(Definition 2.6.37). We now study a reformulation of problem (SP I ) from

Sect. 8.3.2 taking into account Theorem 2.6.41.
[
pN WD Inf F .x/; S WD fx 2 X j G.x/ \ K ¤ ;g : (SP)
x2S
Restrictions of this type have been considered in many papers (see Borwein [56];
Corley [112]; Jahn [290]; Luc [402]; Götz and Jahn [218]; Crespi, Ginchev and
Rocca [116]; Boţ and Wanka [52]; Hernández, Löhne, Rodríguez-Marín, Tammer
[248]).
For simplicity we put WDI .
In the sequel we consider the primal problem (SP). We assume throughout that a
fixed vector
c 2 int C (8.27)
is given. Several concepts, for instance, the Lagrangian, the dual objective function
and subgradients will depend on the choice of this vector c. Note that we do not
mention this dependance explicitly.
The Lagrangian map of problem (SP) is defined by (compare the I -valued
formulation in (8.18))
[
L W X Z Y ; L.x; u / D F .x/ C Inf hu ; ui fcg : (8.28)
u2G.x/CK
In the special case q D 1, C D RC , c D 1, the well-known Lagrangian coincides

with the Lagrangian of the scalar problem. For every choice of c 2 int C we have a
different Lagrangian map and a different corresponding dual problem, but we show
that weak duality and strong duality hold for all these problems.
With the notations of this section we get an equivalent formulation to Proposi-
tion 8.3.13).
Proposition 8.4.1. For all x 2 S it holds
[
Sup L.x; u / D F .x/:
u 2Z
We next define the dual problem. The dual objective function is defined by
[
W Z Y ; .u / WD Inf L.x; u /
x2X
and the dual problem (with respect to c 2 int C ) associated to (SP) is defined by
[
dN WD Sup .u /: (SDL )
u 2Z
As shown in Theorem 8.3.14, we have weak duality. Taking into account

Theorem 2.6.41, we get the following formulation.
Theorem 8.4.2 (Weak Duality Between (SP) and (SDL )). The problems (SP) and
(SDL ) satisfy the weak duality inequality, i.e.,
[ [
Sup .u / 4 Inf F .x/:
u 2Z x2S
The following strong duality theorem has been proven in Theorem 8.3.15 and is
reformulated now using Theorem 2.6.41.
Theorem 8.4.3 (Strong Duality Between (SP) and (SDL )). Let F be C -convex,
let G be K-convex, and let
G.Dom F / \ int K ¤ ;: (8.29)
Then, strong duality holds, that is,

[ [
Sup .u / D Inf F .x/:
u 2Z x2S
By strong duality and Theorem 2.6.41 we get

[ [ [ [
Sup Inf L.x; u / D Inf Sup L.x; u /:
u 2Z x2X x2X u 2Z
The next statement extends Proposition 8.4.1. Note that the additional assumption
of G.x/ C K being a closed convex set originates from the set-valued constraints
and could not be omitted if the objective function would be scalar-valued, see [395,
Example 3.21] for further details.
Proposition 8.4.4. Let F W X Y a set valued map with F .x/ D Inf F .x/ ¤
f1g for all x 2 X , Dom F ¤ ; and let the set G.x/ C K be closed and convex
for every x 2 X , then
[
F .x/ if x2S
Sup L.x; u / D
fC1g else.
u 2Z
8.4.2 Subdifferentials and Stability
In this section we will use the notations introduced in Sect. 8.4.1 and study
subdifferential notions and corresponding duality statements.
342 8 Duality
We start by introducing the notion of subgradient and stability for our framework.
These concepts are mainly motivated by Tanino [566] (see also Boţ, Grad and
Wanka [73] for numerous related results).
Let ' be a set-valued map from X Z to Y defined by

F .x/ if G.x/ \ .K u/ ¤ ;
'.x; u/ WD
; else.
Denote by W W Z Y the perturbation map defined by

[
W .u/ WD Inf '.x; u/:
x2X
Clearly, we have
[
W .0/ D Inf F .x/ D p:
N
x2S
This leads to the following definition of a subgradient.

Definition 8.4.5. A point u 2 K C is a called positive subgradient of W at
.Nu; y/ N written u 2 @C W .Nu; y/
N 2 Z Y with yN 2 W .x/, N for short, if
[
yN hu ; uN i c 2 Inf .W .u/ hu ; ui c/:
u2Z
Remark 8.4.6. Note that the subgradient in [379, Definition 2.5] (see Sect. 15.1.2,
especially Definition 15.1.10) is a stronger notion than the one in Definition 8.4.5.
In particular, in [379, Definition 2.5], subgradients are operators whereas in Def-
inition 8.4.5 subgradients are vectors. Furthermore, the definition of subgradients
in Definition 8.4.5 is closely related to that given by Boţ, Grad and Wanka [73,
Definition 7.1.9 (c)] (see Definition 8.1.3, (c)), where the subgradients are defined
using the vector approach (Pareto maximum) instead of the supremal set like in
Definition 8.4.5. However, in [73] (Definition 8.1.3, (c)) the so called c-subgradients
are vectors too.
Lemma 8.4.7. Consider the set-valued maps @C W .0; / W Y Z and W Z
Y . For u 2 Z with .u / Y , one has
u 2 @C W .0; y/ ” y 2 .u /:
Proof. By definition, u 2 @C W .0; y/ means

[
y 2 Inf .W .u/ hu ; ui c/:
u2Z
We have
[ [ [
Inf .W .u/ hu ; ui c/ D Inf Inf .'.x; u/ hu ; ui c/
u2Z u2Z x2X
[
D Inf .F .x/ hu ; ui c/
u2G.x/CK; x2X
[
D Inf .F .x/ C hu ; ui c/
u2G.x/CK; x2X
[ [
D Inf F .x/ C Inf hu ; ui c
x2X u2G.x/CK
[
D Inf L.x; u / D .u /;
x2X
which proves the claim. t

u
Definition 8.4.8. Problem (SP) is called stable if W .0/ ¤ fC1g, W .0/ ¤ f1g
and @C W .0; y/ ¤ ; for all y 2 W .0/.
Remark 8.4.9. In Boţ, Grad and Wanka [73] the definition of c-subgradients (see
Definition 8.1.3) is used in order to introduce the property that the primal set-valued
optimization problem is c-stable, compare Remark 8.1.5.
Theorem 8.4.10. If (SP) is stable, then strong duality holds for (SP) and (SDL ),
that is,
[ [
Sup .u / D Inf F .x/:
u 2Z x2S
S S
Proof. We set dN D Sup u 2Z .u / and pN D W .0/ D Inf x2S F .x/. By
assumption, we have W .0/ ¤ fC1g and W .0/ ¤ f1g, which implies ; ¨
Cl C W .0/ ¨ Y . Take some y 2 W .0/. Since (SP) is stable, there is some u 2 Z
with y 2 .u / (by Lemma 8.4.7). Using weak duality we get .u / 4 dN 4 p. N
Thus dN ¤ f1g and dN ¤ fC1g, which implies ; ¨ Cl C dN ¨ Y .
By weak duality, it remains to prove pN 4 dN . Taking into account Lemma 2.6.38,
we have to prove that .pN int C / \ dN D ;: On the contrary, suppose that there is
N Then there exists z 2 pN D W .0/ and c 2 int C such
y 2 Y with y 2 .pN int C /\ d.
that y D z c. On the other hand, there exists u 2 @C W .0; z/. By Lemma 8.4.7,
this means z 2 .u /. Hence y 2 ..u / int C / \ dN . By Lemma 2.6.38 this
contradicts .u / 4 dN . t
u
Theorem 8.4.11. If F is C -convex, G is K-convex,
G.Dom F / \ .int K/ ¤ ;; (8.30)
and W .0/ ¤ f1g, then (SP) is stable.

344 8 Duality
Proof. From (8.30), we get W .0/ ¤ fC1g.

Let yN 2 W .0/. By Lemma 8.4.7 we have to show that there exists u 2 Z
with yN 2 .u /. The map QW X Y Z defined by Q.x/ WD .F .x/; G.x// is
C K-convex. Thus, Q.X / C C K is a convex set. We next show that
.Q.S / C C K/ \ int .B .K// D ; (8.31)
N C . Indeed, if there exist x 0 2 S and .y; u/ such that

where B D fyg

.y; u/ 2 .F .x 0 /; G.x 0 // C C K \ int .B .K//;
then y 2 .F .x 0 / C C / \ .fyg N int C / and u 2 .G.x 0 / C K/ \ int K. Thus,

y D yN c where y 2 F .x / and c 0 2 int C S
0 0 0 0
and .G.x 0 / C K/ \ .K/ ¤ ; (that
0
is, x 2 S ) which contradicts yN 2 W .0/ D Inf x2S F .x/:
By (8.31), applying a separation theorem there exists a pair .y ; u / 2 Y Z n
f.0; 0/g such that
hy ; yi C hu ; ui hy ; bi C hu ; d i (8.32)
for all .y; u/ 2 Q.S / C C K, b 2 B and d 2 K. We deduce that .y ; u / 2

.C K / n f.0; 0/g: This implies
8.y; u/ 2 Q.S / C C K W hy ; yi C hu ; ui hy ; yi

N :
Since, by (2.62), Cl C W .0/ D Cl C F .S / D cl .F .S / C C /, we get
8y 2 Cl C F .S /; 8u 2 G.S / C K W hy ; yi C hu ; ui hy ; yi
N : (8.33)
We show that y ¤ 0. Assuming the contrary, we get u ¤ 0 and, by (8.33), we

have
8u 2 G.S / C K W hu ; ui 0:
On the other hand, by (8.30), there exists x 2 S and u0 2 G.x/ \ .int K/, i.e.,
hu ; u0 i > 0 as u 2 K n f0g. Since u0 2 G.x/ G.x/ C K this is a contradiction.
Since y 2 C n f0g, for the fixed vector c 2 int C according to (8.27), we have
hy ; ci < 0. Without loss of generality we can assume hy ; ci D 1.

Since yN 2 W .0/ Cl C F .S /, by (8.33), we have hu ; ui 0 for all u 2 G.x/

N
N
G.x/CK. Since u 2 K , we have hu ; uN i D 0 for all u 2 G.x/\K.
N Thus, (8.33)
can be written as
8y 2 Cl C F .S /; 8u 2 G.S / C K W hy ; y hu ; ui fcgi hy ; yi

N :
From weak duality, we know that yN 2 Cl C .u /. Assuming that yN 2 .u /Cint C ,
we obtain a contradiction to the latter inequality. Hence, by Proposition 2.6.33, we
have yN 2 .u /. t
u
8.4.3 Duality Statements with Operators as Dual Variables
We establish in this section another type of dual problem where the dual variables
are operators instead of vectors as in problem .SDcL I / (compare also problem
.SDTL I /, where the dual variables are operators too). The usage of operators is
more common in the literature (see, for instance [112, 251, 351, 387, 402]). We will
see, however, that a duality theory based on operators as dual variables is an easy
consequence of the above results.
Denote by L the set of all linear continuous operators from Z to Y and by LC
the subset of all positive operators, that is, LC WD fT 2 L W T .K/ C g. Given
T 2 L and A Z we write T .A/ D fT .a/ j a 2 Ag. Let Lc be a subset of
L defined by Lc WD fT 2 L j T D hu ; i c for some u 2 K C g: Obviously,
we have
Lc LC : (8.34)
Moreover, Lc is isomorphic to K C Z .
The Lagrangian map L W X L ! Y is defined by
L.x; T / WD F .x/ C T .G.x//: (8.35)
The dual objective function ˚ W L ! Y is defined by

[
˚.T / WD Inf L.x; T /;
x2X
and the associated dual problem is defined by

[
dQ WD Sup ˚.T /: (SDTL )
T 2LC
Comparing the two dual problems (SD L ) (with vectors as dual variables) and
(SDTL ) (with operators as dual variables), we observe that Lagrangian (8.28) for
problem (SDL ) involves the cone K but Lagrangian (8.35) for problem (SDTL )
does not. On the other hand, the supremum in (SDL ) is taken over the whole
linear space Z whereas in (SDTL ) only the subspace LC of the linear space L
is considered. A reformulation of problem (SDL ) clarifies the connection. Consider,
instead of (8.28), the Lagrangian
346 8 Duality
[
LO W X Z Y ; O
L.x; u / D F .x/ C Inf hu ; ui fcg : (8.36)
u2G.x/
and the corresponding dual objective function

[
O W Z Y ; O / WD Inf
.u O
L.x; u /:
x2X
Lemma 8.4.12. The dual objective function of problem (SDL ) can be expressed as

O / if u 2 K C
.u
.u / D (8.37)
f1g otherwise.
Proof. Since c 2 int C , we have
[
f0g if u 2 K C
Inf hd; u i c D
f1g otherwise.
d 2K
It follows
[
.u / D Inf L.x; u /
x2X
[ [
D Inf F .x/ C Inf hu; u i c
x2X u2G.x/CK
[ [

D Inf F .x/ C Inf hu; u i c C hk; u i c
x2X u2G.x/;k2K
[ [ [

D Inf F .x/ C Inf hu; u i fcg C Inf hk; u i fcg
x2X u2G.x/ k2K
[ [
D Inf O
L.x; u / C Inf hk; u i fcg
x2X k2K
[
O / C Inf
D .u hk; u i fcg :
k2K
Combining the two equations we obtain the result. t

u
As a consequence we can define a dual problem
[
dO WD Sup O /;
.u O
(D)
u 2K C
where we obviously have

[ [
dN D Sup .u / D Sup O / D dO :
.u (8.38)
u 2Z u 2K C
Since K C is isomorphic to Lc and Lc LC , we get (using Theorem 2.6.41)

[ [
dN D Sup .u / 4 Sup ˚.T / D dQ : (8.39)
u 2Z T 2LC
We next prove weak duality.

Theorem 8.4.13 (Weak Duality Between (SP) and (SDTL )). The problems (SP)
and (SDTL ) satisfy the weak duality inequality, i.e.,
[ [
Sup ˚.T / 4 Inf F .x/:
T 2LC x2S
Proof. By Theorem 2.6.41, we have

[ [ [ [
Sup Inf L.x; T / 4 Inf Sup L.x; T /:
T 2LC x2X x2X T 2LC
S
Since ˚.T / D Inf x2X L.x; T /, it remains to show
[ [ [
Inf Sup L.x; T / 4 Inf F .x/:
x2X T 2LC x2S
But this follows from Proposition 8.4.1 and

[ [ [ [
Inf Sup L.x; T / 4 Inf Sup L.x; T /;
x2X T 2LC x2S T 2LC
which is a consequence of Theorem 2.6.41. t

u
Finally we obtain strong duality as a conclusion of the Lagrange duality theorem
with vectors as variables.
Theorem 8.4.14 (Strong Duality Between (SP) and (SDTL )). Let F be C -convex,
let G be K-convex, and let
G.Dom F / \ .int K/ ¤ ;:
Then strong duality holds, that is,

[ [
Sup ˚.T / D Inf F .x/:
T 2LC x2S
348 8 Duality
Proof. From Theorem 8.4.3, inequality (8.39), and Theorem 8.4.13, we get
[ [ [ [
Inf F .x/ D Sup .u / 4 Sup ˚.T / 4 Inf F .x/;
x2S u 2Z T 2LC x2S
which yields the desired equation. t

u
Remark 8.4.15. Further relationships between approaches to duality are discussed
in Sect. 15.1.2, where the notations well-known in vector optimization are used
(compare Li, Chen and Wu [379]).
Chapter 9
Existence Results for Minimal Points
In this chapter we establish several existence results for minimal points with respect
to transitive relations; then we apply them in topological vector spaces for quasi-
orders generated by convex cones. We continue with the presentation of several
types of convex cones and compactness notions with respect to cones. We end the
chapter with existence results for vector and set optimization problems.
The presentation in Sects. 9.1, 9.2, 9.3 and 9.4 follows that in [214] or [539],
where one can find the proofs for the results which are only stated.
9.1 Preliminary Notions and Results Concerning Transitive

Relations
In the sequel Y is a nonempty set and t Y Y , that is, t is a relation on Y .

If ; ¤ A Y and t Y Y , the restriction of t to A is denoted by t A ; i.e.,
t A WD t \ .A A/.
With the relation t on Y we associate the following relations:
tR WD t [ Y ; tN WD t n t 1 D t n .t \ t 1 /; tNR D .tN /R :
Hazen and Morin [240] call tN the asymmetric part of t. Some properties of these
relations are given in the following proposition; the first three properties mentioned
in (ii) are stated by Dolecki and Malivert in [144].
Proposition 9.1.1. Let t be a transitive relation on Y and ; ¤ A Y .
(i) tR is reflexive and transitive; tN \ Y D ;;
(ii) t ı tN tN , tN ı t tN , tN ı tN tN , .tR /N D tN , tR ı tN D tN ı tR D tN ;
.tN /N D tN ;

DOI 10.1007/978-3-642-54265-7__9
350 9 Existence Results for Minimal Points
(iii) tNR is reflexive, antisymmetric,

and transitive; .tN /N D .tNR /N D tN ;
(iv) t A R D .tR /A , t A N D .tN /A .
A A A
Taking into
A
account
A proposition, we denote by tR , tN , and tNR the
the above
A
relations t R , t N , and t NR , respectively.
The preceding proposition shows that with every transitive relation one can
associate a partial order. It is useful to know whether they determine the same
maximal and minimal points. As noted in Sect. 2.1, Max.AI t/ D Min.A; t 1 / for
; ¤ A Y ; so it is sufficient to study only the problems related to minimal points.
Corollary 9.1.2. Let t be a transitive relation on Y and ; ¤ A Y ; then
Min.AI t/ D Min.AI tR / D Min.AI tN / D Min.AI tNR /:
The above corollary shows that the problem of existence (for example) of
minimal points w.r.t. to a transitive relation t reduces, theoretically, to the same
problem for the partial order tNR . Another way to reduce this problem to one for a
partial order is given by the following known result.
Proposition 9.1.3. Let t be a transitive relation on Y and take WD tR \ .tR /1 .
Then is an equivalence relation, and tO D f.x; O y/
O j .x; y/ 2 tg is a partial order
on YO WD Y =, where xO is the class of x 2 Y with respect to . Moreover, if x 2 Y ,
x 2 Min.Y I t/ if and only if xO 2 Min.YO I tO/.
In the sequel we shall also use the notation AC
t .x/ and At .x/ for the upper and
lower sections of A Y with respect to t and x 2 Y . So
AC
t .x/ WD fy 2 A j .x; y/ 2 tR g; A
t .x/ WD fy 2 A j .y; x/ 2 tR g;
respectively; the most common case is for x 2 A. Similarly, for E Y , we consider

[
AC
t .E/ WD fy 2 A j 9 z 2 E W .z; y/ 2 tR g D AC
t .z/;
z2E
[
A
t .E/ WD fy 2 A j 9 z 2 E W .y; z/ 2 tR g D A
t .z/I
z2E
C
clearly, E \ A A t .E/ \ At .E/:
In the rest of this
section weassume that t is a transitive relation on Y:
Note that A t A t .E/ D At .E/ (so one may suppose that E A) and
Min.A
t .E/I t/ D At .E/ \ Min.AI t/; Max.AC C
t .E/I t/ D At .E/ \ Max.AI t/I
(9.1)
moreover,
Min.AI t/ D A \ Min.YtC .A/I t/; Max.AI t/ D A \ Max.Yt .A/I t/; (9.2)

9.1 Preliminary Notions and Results Concerning Transitive Relations 351
and
Min.AI t/ D Min.YtC .A/I t/; Max.AI t/ D Max.Yt .A/I t/ (9.3)
provided that Yt .y/ \ YtC .y/ D fyg for every y 2 A (for example if tR is
antisymmetric).
Indeed, let us prove the first relation in (9.2). Take y 2 Min.AI t/. Then y 2
A YtC .A/. Take z 2 YtC .A/ with .z; y/ 2 tR ; then there exists y 2 A such that
.y; z/ 2 tR , and so .y; y/ 2 tR . It follows that .y; y/ 2 tR , whence .y; z/ 2 tR , and
so the inclusion holds. Conversely, consider y 2 A \ Min.YtC .A/I t/ and take
y 2 A . YtC .A// with .y; y/ 2 tR . Then .y; y/ 2 tR , and so y 2 Min.AI t/:
Assume now that Yt .y/ \ YtC .y/ D fyg for every y 2 A and take y 2
Min.YtC .A/I t/ . YtC .A//; there exists yQ 2 A . YtC .A// such that .y; Q y/ 2 tR ,
Q 2 tR . Thus y 2 Yt ./\YtC .y/
and so .y; y/ Q D fyg,
Q whence y D yQ 2 A. Therefore,
y 2 Min.A; y/ by (9.2), and so (9.3) holds.
Note that (9.2) makes more precise [189, Proposition 2.1 (a)].
Taking E the singleton fxg in (9.1), we get
Min.A
t .x/I t/ Min.AI t/; Max.AC
t .x/I t/ Max.AI t/: (9.4)
Relation (9.4) shows that in order to have that Min.AI t/ is nonempty it is sufficient
to get the nonemptiness of Min.A t .x/I t/ for some x 2 A:
We say that ; ¤ A Y has the domination property (DP) (w.r.t. t) if
Min.A t .y/I t/ ¤ ;, for every y 2 A (i.e., every element of A is dominated by
a minimal element of A).
A quite important problem is how to extend other notions related to partially
ordered sets, like chain or increasing net, to sets endowed with transitive relations.
Related to this problem we have the next result.
Proposition 9.1.4. Let A Y be nonempty.
(i) If tRA is a partial order on A, then tRA tNR .
(ii) If is a total order on A such that tNR , then D tNR
A
.
Note that in (ii) we cannot replace tNR by t or tR (take Y D R and t D R R).
Let .xi /i 2I Y be a net. We say that .xi / is t-increasing [t-decreasing] if
.xi ; xj / 2 t [.xj ; xi / 2 t] for all i; j 2 I with i j and i ¤ j ; .xi / is strictly
t-increasing [strictly t-decreasing] if .xi / is tN -increasing [tN -decreasing].
In the sequel we say that ; ¤ A .Y; t/ is a chain (w.r.t. to t) if tRA is a total
order on A, while A is well-ordered (w.r.t. to t) if A is well-ordered by tRA .
The following result is due to Gajek and Zagrodny (see [193]).
Proposition 9.1.5. Let t be a transitive relation on Y . Then there exists a nonempty
well-ordered subset W of Y such that for every x 2 Y n W there exists w 2 W with
.x; w/ 2 t or .w; x/ … t.
9.2 Existence of Minimal Elements with Respect

to Transitive Relations
We begin with the following result.

Proposition 9.2.1. Let A Y be nonempty. Suppose that one of the following
conditions holds:
(i) every nonempty set E A with E E t [ t 1 [ Y is lower bounded in A;
(ii) every chain in A is lower bounded in A;
(iii) every well-ordered subset of A w.r.t. t 1 is t-lower bounded in A.
Then A has the domination property. In particular, Min.AI t/ ¤ ;.
Corollary 9.2.2. Let A Y be nonempty. Suppose that one of the following
conditions holds:
(i) every t-decreasing net of A is lower bounded in A;
(ii) every strictly t-decreasing net of A is lower bounded in A;
(iii) every t-decreasing net .yi /i 2I A with I totally ordered is lower bounded
in A;
(iv) every strictly t-decreasing net .yi /i 2I A with I totally ordered is lower
bounded in A;
(v) every t-decreasing net .yi /i 2I A with I well-ordered is lower bounded in A;
(vi) every strictly t-decreasing net .yi /i 2I A with I well-ordered is lower
bounded in A.
Note that conditions (i), (ii), and (iii) of Corollary 9.2.2 are equivalent to
\
8 .yi /i 2I A; .yi / t–decreasing W A \ Yt .yi / ¤ ;; (9.5)
i 2I
\
8 .yi /i 2I A; .yi / strictly t–decreasing W A \ Yt
NR
.yi / ¤ ;; (9.6)
i 2I
\

8 D A; D totally ordered W A \ Yt .y/ ¤ ;; (9.7)
y2D
respectively.
Corollary 9.2.3. Let A Y be nonempty. Suppose that

[ [

8 E A; A Y n Yt .a/ ; 9 E0 2 F .E/ W A Y n Yt .a/ ;
a2E a2E0
(9.8)
9.2 Existence of Minimal Elements with Respect to Transitive Relations 353
where F .E/ is the class of finite subsets of E. Then A has the domination property.
In particular, Min.AI t/ ¤ ;.
Proof. Let us observe that (9.8) implies T (9.5).Indeed, take .yi /i 2I A a t-
decreasing net and assume S that A \ Y
i 2I t .y i / D ;. Then E WD fyi j
i 2 I g A and A S a2E Y n Yt .a/ . Using our hypothesis, there exists
E0 2 F .E/ with A a2E0 Y n Yt .a/ . Therefore, there exists I0 2 F .I /
T

such that E0 D fyi j i 2 I0 g, and so A \ i 2I0 Yt .yi / D ;. Because I is
directed, there exists i0 2 I such that i0
i for every i 2 I0 . It follows that .yi0 ; yi /
(and so yi0 2 Yt .yi /) for every i 2 I0 , getting so a contradiction. The conclusion
follows using Corollary 9.2.2 (i). t
u
An even stronger condition than (9.8) is

[ [

8 E Y; A Y n Yt .a/ ; 9 E0 2 F .E/ W A Y n Yt .a/ :
a2E a2E0
(9.9)
The following result is due, essentially, to Gajek and Zagrodny [193].

Proposition 9.2.4. Let A Y be nonempty. Suppose that the following two
conditions hold:
(i) every nonempty well-ordered subset W of A w.r.t. t 1 is at most countable;
(ii) every strictly t-decreasing sequence of A is lower bounded in A.
Observe that the set
[

E j E A; A Y n Yt .a/
a2E
is empty if and only if there exists y 2 A such that .y; y/ 2 t for every y 2 A, that
is, A has a smallest element w.r.t. t; similarly, the set
[

E j E Y; A Y n Yt .a/
a2E
is empty if and only if there exists y 2 A such that .y; y/ 2 t for every y 2 Y , that
is Y has a smallest element w.r.t. t which belongs to A. Of course, in these situations
conditions (9.8) and (9.9) are verified, respectively.
If is a topology on Y and the lower sections of Y are closed (for example,
if t is closed in Y Y and is Hausdorff), then the sets Y n YtC .a/ are open;
in this situation condition (9.8) is a kind of compactness of A. This is the reason
for which a set A Y verifying condition (9.8) is called order-semicompact
in [189, Definition 2.4 (a)]. Note also that condition (9.7) is equivalent to the fact
that A is order-totally-complete in the sense of [189, Definition 2.2 (a)]; so, [189,
Theorem 3.1] is practically Corollary 9.2.2 (iii).
The following result is related to this kind of conditions.
Proposition 9.2.5. Let A Y be nonempty. Assume that there exists a relation s
on Y such that
s ı tN tR (9.10)
and one of the following conditions holds:

\
8 .yi /i 2I A; .yi / strictly t-decreasing W A \ Ys .yi / ¤ ;; (9.11)
i 2I
\
8 D A; D totally order w.r.t. t W A \ Ys .y/ ¤ ;: (9.12)
y2D

In the sequel, in this section, we suppose that .Y; / is a topological space. In this
situation one can formulate other conditions for the existence of minimal points.
Having a net .xi /i 2I , by .x'.j / /j 2J , or simply .x'.j / /, we denote a subnet of .xi /;
this means that J is directed and ' W J ! I has the property that for every i 2 I
there exists ji 2 J such that '.j /
i for all j 2 J , j
ji :
Corollary 9.2.6. Let Y be endowed with the topology and ; ¤ A Y . Assume
that one of the following conditions holds:
8 .yi /i2I A t–decreasing; 9 .y'.j / /j 2J ! y 2 A; 8 j 2 J W .y; y'.j / / 2 tR ;

(9.13)
8 .yi /i2I A strictly t–decreasing; 9 .y'.j / /j 2J ! y 2 A; 8 j 2 J W .y; y'.j / / 2 tR :

(9.14)

When the upper or lower sections of Y are closed, we may consider other
conditions, too.
Proposition 9.2.7. Let Y be endowed with the topology and ; ¤ A Y .
Consider the following conditions:
8 .yi /i 2I A t–decreasing W 9 .y'.j //j 2J ! y 2 A; (9.15)
8 .yi /i 2I A strictly t–decreasing W 9 .y'.j / /j 2J ! y 2 A; (9.16)
8 .yi /i 2I A t–decreasing; 9 y 2 A W yi ! y; (9.17)
and
9.3 Existence of Minimal Points with Respect to Cones 355
8 .yi /i 2I A strictly t–decreasing; 9 y 2 A W yi ! y: (9.18)
If the lower sections of Y are closed, then (9.15) , (9.13) and (9.16) , (9.14),
while if t is a partial order and the upper and lower sections of Y (w.r.t. t) are
closed, then (9.15) , (9.17) and (9.16) , (9.18).
Corollary 9.2.8. Let Y be endowed with the topology and ; ¤ A Y . Assume
that the lower sections of Y are closed and A is compact. Then A has the domination
property. In particular, Min.AI t/ ¤ ;.
Hazen and Morin proved Corollary 9.2.8 in [240, Corollary 2.8]. The next result
is [189, Theorem 3.2].
Corollary 9.2.9. Let Y be endowed with the topology and ; ¤ A Y . Suppose
that

8x; y; z 2 Y W x 2 Yt .y/ n YtC .y/; z 2 cl Yt .y/ ) z 2 Yt .x/; (9.19)
and
\
8 D A; D totally ordered w.r.t. t W A \ Ys .y/ ¤ ; (9.20)
y2D
hold, where
s WD f.y; z/ 2 Y Y j y 2 cl Yt .z/g : (9.21)
Then A has the domination property; in particular, Min.AI t/ ¤ ;.

Proof. With our notation, condition (9.19) can be written as s ı tN tR , that
is (9.10), while condition (9.20) is just (9.12). Applying Proposition 9.2.5 we get
the conclusion. t
u
9.3 Existence of Minimal Points with Respect to Cones
To begin with, let Y be a real vector space and ; ¤ C Y a convex cone. As usual,
with C we associate the reflexive and transitive relation
C WD t WD f.x; y/ 2 Y Y j y x 2 C g: (9.22)
Taking
L WD C \ .C /;
the lineality space of C , the equivalence relation WD t \ t 1 is f.x; y/ 2 Y Y j

y x 2 Lg. So we get
tN D f.x; y/ 2 Y Y j y x 2 C n Lg: (9.23)
Using Proposition 9.1.1 we obtain that
C C .C n L/ D C n L; .C n L/ C .C n L/ C n LI
the above formulae were obtained by Luc in [401, 402]. So,
tNR D f.x; y/ 2 Y Y j y x 2 .C n L/ [ f0gg:
It follows that .C n L/ [ f0g is a pointed convex cone. Note that for x 2 Y and
A Y one has AC C
C .x/ WD AC .x/ D A \ .x C C / and AC .x/ D A \ .x C /.
Therefore the upper and lower sections of Y (w.r.t. C ) are closed if and only if C is
closed. Similarly, for E Y , AC
C .E/ WD A\.E CC / and AC .E/ WD A\.E C /.
In accordance with the notions introduced in Sect. 9.1, the net .xi /i 2I Y is
[strictly] C -increasing if xj xi 2 C [xj xi 2 .C n L/] for all i; j 2 I , i j ,
i ¤ j ; .xi /i 2I is [strictly] C -decreasing if .xi /i 2I is [strictly] C -increasing. Of
course, the set A Y is C -upper (lower) bounded if A x0 C (A x0 C C )
for some x0 2 Y . Moreover, the set Min.A; C / will be denoted by Min.AI C /. An
element y 2 Min.AI C / is also called an efficient point of A (w.r.t. C ).
It is obvious that for ; ¤ A Y and C D Y or C D f0g we have that
Min.AI C / D A. Taking into account this fact, in the sequel we shall suppose that
C is proper, i.e., f0g ¤ C ¤ Y .
To Proposition 9.2.1 and Corollary 9.2.2 corresponds the following result.
Proposition 9.3.1. Let ; ¤ A Y . Suppose that one of the following conditions
holds:
(i) every nonempty set E A such that E E C [ .C / is C -lower bounded
in A;
(ii) every chain in A w.r.t. C is C -lower bounded in A;
(iii) every well-ordered subset of A w.r.t. .C / is C -lower bounded in A;
(iv) every C –decreasing net of A is C -lower bounded in A.
Then A has the domination property. In particular, Min.AI C / ¤ ;.
In the following proposition we gather some properties of efficient sets.
Proposition 9.3.2. Let C; K Y be convex cones such that C K, x 2 Y , and
; ¤ A; E Y .
(i) Min.A
K .E/I C / D AK .E/ \ Min.AI C /. In particular, we have that

Min.AC .x/I C / D AC .x/ \ Min.AI C /.
(ii) If A E A C K and
K \ .C / C; (9.24)
then
Min.EI K/ Min.AI C / C .K \ .K// :
In particular, if Min.EI K/ ¤ ;, then Min.AI C / ¤ ;.

(iii) If (9.24) holds, then Min.AI K/ Min.AI C /.
(iv) If A E A C C , then Min.AI K/ Min.EI K/; moreover, if K \ .C / D
f0g, then Min.AI K/ D Min.EI K/.
(v) Suppose that (9.24) and
K C .C n L/ C (9.25)
hold. If A Min.AI K/ C K, then A Min.AI C / C C (i.e., if A has (DP)

w.r.t. K, then A has (DP) with respect to C ).
(vi) Suppose that K \ .C / D f0g and A Min.AI K/ C C . Then Min.AI C / D
Min.AI K/.
In the sequel we suppose that Y is a real Hausdorff topological vector space
(H.t.v.s. for short), its topology being denoted by .
Applying the results from Sect. 9.2 we obtain several existence theorems for
efficient points w.r.t. cones. Before stating them, let us recall or introduce some
notions or, more exactly, some possible properties of the cone C .
So, C is (sequentially Daniell) Daniell if every C -upper bounded and C -
increasing (sequence) net in Y has a least upper bound (or supremum) and converges
to it. Because we suppose that Y is Hausdorff, every (sequentially) Daniel cone is
pointed.
Other similar conditions are:
(P1) Every C -increasing and C -upper bounded net .xi /i 2I C is convergent to
an element of C .
(P2) Every C -increasing and C -upper bounded net .xi /i 2I Y is Cauchy.
(P3) Every C -increasing and -bounded net.xi /i 2I C is convergent to an
element of C .
(P4) Every C -increasing and -bounded net .xi /i 2I Y is Cauchy.
(P5) Every C -increasing and -bounded net .xi /i 2I Y which is contained in a
complete set is convergent.
The sequential variants of (P1)–(P4) are:
(SP1) Every C -increasing and C -upper bounded sequence .xn /n2N C is
convergent to an element of C .
(SP2) Every C -increasing and C -upper bounded sequence .xn /n2N Y is
Cauchy.
(SP3) Every C -increasing and -bounded sequence .xn /n2N C is convergent

to an element of C .
(SP4) Every C -increasing and -bounded sequence .xn /n2N Y is Cauchy.
If one of the above conditions holds, then, necessarily, C is pointed.
Note that an equivalent formulation for (P1) [and similarly for (P3), (SP1), and
(SP3)] is:
(P10 ) Every C -increasing and C -upper bounded net .xi /i 2I Y is convergent
to an element of Y which is a C -upper bound for fxi j i 2 I g.
Of course, (P10 ) ) (P1). For the converse implication consider a C -increasing
and C -upper bounded net .xi /i 2I Y . Fix k 2 I . Then the net .yi /k i , with
yi WD xi xk 2 C , is C -increasing and C -upper bounded. By (P1) there exists
y 2 C such that yi ! y. It follows that xi ! x D xk C y, and so xk C x. Since
k 2 I is arbitrary, (P10 ) holds.
Note that when Y is a normed space, Krasnosel’skij [339] says that a convex cone
C satisfying (SP2) is regular, while a convex cone C satisfying (SP4) is completely
regular. Isac [280] says that C is completely regular if every C -increasing and -
bounded net of C is convergent. Of course, if Y is quasi-complete (i.e. every closed
and bounded subset of Y is complete), then C is completely regular in Isac’s sense if
and only if C has property (P4). Németh [440] says that a convex cone C satisfying
(P1) [(SP1)] is regular [sequentially regular]; moreover, he says that C is fully
regular if C satisfies (P3). Condition (P5) was introduced by Ha [224] under the
name of property . /.
We have the following scheme of implications, where (D) means Daniell, while
(sD) means sequentially Daniell.
(
.D/ ) .P1/ ) .P2/ .P3/ ) .P4/ ) .P5/
+" +" m +" m #
(
.sD/ ) .SP1/ ) .SP2/ .SP3/ ) .SP4/ .P3/:
The implications marked by ), +, and m always hold; those marked by " are
valid when C is complete or when 0 2 Y has a countable basis of neighborhoods;
the implications # and are valid if C is complete; those marked ( are valid for
C closed. The nontrivial implications are stated in the next proposition.
Moreover, the implications ! below hold if C is normal, while the implications
hold if C is boundedly order complete, i.e., every C -increasing and -bounded
net of C has a supremum:
(P3) (P1); (P4) (P2); (SP3) (SP1) and (SP4) (SP2). (9.26)
The proof of the following result uses an idea from [440]. Note that Borwein (see
Proposition 2.7 (ii) of [61]) showed that C is Daniell iff C is sequentially Daniell
when “C admits a strict monotone functional.”
Proposition 9.3.3. Let C Y be a convex cone. Then

(i) .P 2/ , .SP2/ and .P 4/ , .SP4/.
(ii) If C is complete, or 0 2 Y has a countable basis of neighborhoods, then
.P1/ , .SP1/, .P 3/ , .SP3/, and C is Daniell if and only if C is
sequentially Daniell.
Before stating the next result we introduce other two notions. We say that the
subset A of Y is C -complete (resp. sequentially C -complete) if every Cauchy C -
increasing net (resp. Cauchy C -increasing sequence) is convergent to an element
of C .
Proposition 9.3.4. Let A Y be a nonempty set. Assume that one of the following
conditions holds:
(i) C satisfies (P1), while A is closed and C -lower bounded;
(ii) C is closed and satisfies (SP2), while A is C -complete and C -lower bounded;
(iii) C satisfies (P3), while A is closed and -bounded;
(iv) C is closed and satisfies (SP4), while A is C -complete and -bounded.
When every well-ordered subset W of Y w.r.t. .C / (see page 351) is at most
countable (i.e., .Y; .C / / is countable orderable in the sense of Gajek and Zagrodny
[193]) the closedness and completeness in the preceding result can be taken in the
weaker sequential sense.
Proposition 9.3.5. Let A Y be a nonempty set. Assume that every well-ordered
subset W of Y w.r.t. .C / is at most countable and one of the following conditions
holds:
(i) C satisfies (SP1), while A is sequentially closed and C -lower bounded;
(ii) C is closed and satisfies (SP2), while A is sequentially C -complete and C -
lower bounded;
(iii) C satisfies (SP3), while A is sequentially closed and -bounded;
(iv) C is closed and satisfies (SP4), while A is sequentially C -complete and -
bounded.
The next result is due to Ha [225].
Proposition 9.3.6. Assume that cl C satisfies (P5) and A Y is nonempty,
complete, and -bounded. Then Min.AI C / ¤ ;.
Proposition 9.3.7. Assume that C is closed and A Y is nonempty and compact.
Corollary 9.3.8. Let ; ¤ A Y and assume that
cl C \ .C / C: (9.27)
If there exists E Y such that A

C .E/ or Acl C .E/ is nonempty and compact (even
weakly compact if Y is locally convex), then Min.AI C / ¤ ;.
Proposition 9.3.9. Let C; K Y be convex cones and ; ¤ A Y . Assume that
K C .C n L/ C and
\
8 .yi /i 2I A strictly C -decreasing W A \ .yi K/ ¤ ; (9.28)
i 2I
holds. Then A has the domination property. In particular, Min.AI C / ¤ ;.

Note that in locally convex spaces, every result stated above has a weak version
when the topology is replaced by the weak topology on Y .
9.4 Types of Convex Cones and Compactness with Respect

to Cones
The following result is stated in Jameson [303, Theorem 3.8.7].

Theorem 9.4.1. Let Y be an H.t.v.s. and C be a nontrivial well-based convex cone.
Then condition (P4) holds.
Proposition 9.4.2. Let .Y; P/ be an H.l.c.s. If C is supernormal, then C satisfies
condition (P4), and therefore (P2).
Another type of cone is the ./-cone. Suppose that Y is a l.c.s.; the cone C is
a ./-cone if there exists x 2 C C such that fx 2 C j hx; x i 1g is relatively
weakly compact (see [542, Definition 2.1] and [87, Definition 4.1]). Sterna-Karwat
showed in [542, Proposition 2.1] that
C is ./-cone , cl C has a weakly compact base.
If the cone C has a bounded base , then C is supernormal; furthermore, if

is complete, it follows easily that C is complete, too. This discussion shows that
a cone with compact base is Daniell, supernormal, and complete. In particular, a
closed ./-cone is Daniell and complete w.r.t. the weak topology.
Luc in [401, Definition 2.3] and [402, Definition I.1.1] says that C is correct
if (9.25) holds for K D cl C . Every domination cone in the sense of Henig [241,
p. 112] is correct.
Concerning cone compactness, recall the following notions.
Hartley [239, p. 214] says that A is C -compact if Acl C .y/ is compact for every
y 2 A.
Corley in Definition 2.5 of [110] says that A Y is C -semicompact
if (9.8) holds for t equal to cl C . As a generalization of this notion, Luc ([401,
9.4 Types of Convex Cones and Compactness with Respect to Cones 361
Definition 2.1] and [402, Definition II.3.2]) says that A is C -complete ‘if (9.28)
holds for K D cl C .
Postolică [479] and Isac [281] say that A is C -bounded if there exists a -
bounded set A0 A such that A A0 C C , A is C -closed if A C C is closed, and
A is C -semicompact if A is C -bounded and C -closed; in fact, these notions are also
used in [280, Definition 3], but with A0 a singleton.
Dedieu [129] says that A is asymptotically compact (a.c. for short) if there
exist > 0 and U a neighborhood of 0 2 Y such that .Œ0; A/ \ U is relatively
compact; note that A is a.c. iff cl A is a.c. (see [612, Proposition 2.2(i)]). Of course,
every subset of Rm is asymptotically compact. Several properties of a.c. sets can be
found in [612, Proposition 2.2].
The asymptotic cone of the nonempty set A Y is
A1 WD fx 2 Y j 9 .ti /i 2I .0; 1/; ti ! 0; 9 .ai /i 2I A W ti ai ! xgI
if Y is a normed space, in particular, for Y D Rm , one can use sequences instead of

nets. Note that A1 D .cl A/1 . If A is closedT and convex, then A1 is given by the
known formula from convex analysis A1 D t >0 t.A a/ for some fixed a 2 A.
The importance of this notion in our context is shown by the following result.
Proposition 9.4.3. Let C Y be a closed convex cone and ; ¤ A Y closed.
(i) If A is asymptotically compact and A1 \ .C / D f0g, then A \ .x C / is
compact for every x 2 Y (hence A is C -compact).
(ii) Suppose that there exists a compact set Q Y such that 0 … Q, C D Œ0; 1/Q
and A1 \.C / D f0g. Then A\.xC / is compact for every x 2 Y . Moreover,
if C is pointed, then A C C is closed and .A C C / \ .x C / is compact for
every x 2 Y .
Conversely, if A is convex and A \ .x C / is nonempty and compact for some
x 2 Y , then A1 \ .C / D f0g.
Corollary 9.4.4. Let C Y be a closed convex cone and A a nonempty closed
subset of Y . Assume that A1 \ .C / D f0g. If A is asymptotically compact or
C D Œ0; 1/Q for some compact set Q Y with 0 … Q, then A has the domination
property.
Using the results in the previous section and the present one, one can deduce the
most part of existence theorems spread in the literature.
For example, using Proposition 9.3.1 one gets the results obtained in Corley [110,
Theorem 3.1] and Chew [95]; using Propositions 9.3.4 and 9.3.5 one gets the results
obtained in Jameson [303, Corollary 3.8.10], Penot [460, Theorem 3.3], Cesari and
Suryanarayana [87, Lemma 4.1], Borwein [68, Theorem 1], Isac [280, Theorem 2],
Postolică [480, Theorem 3.2, Corollary 3.2.1], Attouch and Riahi [13, Theorem 2.5],
Ha [224, 225], Ng and Zheng [441, Theorem 3.1]; using Proposition 9.3.7 one gets
the results obtained in Bitran and Magnanti [51, Proposition 3.1], Nieuwenhuis
[442, Theorem I-14], Henig [241, Theorem 2.1], Borwein [68, Theorem 1], Penot
and Sterna-Karwat [472, Remark 3.3], Jahn [291, Theorem 2.3(b), Theorem 2.6(b)],
Tanaka [561, Lemma 2.4]; using Proposition 9.3.9 one gets the results obtained
in Luc [401, Theorem 2.6, Corollary 2.12], [402, Theorem II.3.3], Malivert [411,
Theorem 3.5].
Note that there are other results concerning existence of efficient points w.r.t.
cones that can not be deduced from the existence results in Sect. 9.2.
9.5 Existence of Optimal Solutions for Vector and Set

Optimization Problems
In this section .X; / is a Hausdorff topological space and t is a reflexive transitive

relation on Y . Consider first the problem
Min f .x/ s.t. x 2 X; (P )
where f W X ! Y is a function. A solution of problem (P) is an element x0 2 X

such that f .x0 / 2 Min.f .X /; t/:
We say that f is t-monotonically lower semicontinuous (m.l.s.c. for short) at
x0 2 X if for every net .xi /i 2I X such that xi ! x0 and .f .xi //i 2I is t-
decreasing one has that .f .x0 /; f .xi // 2 t for every i 2 I ; f is t-monotonically
lower semicontinuous if f is t-m.l.s.c. at any x 2 X:
The next theorem subsumes the existence results established in Theorems 5.1, 5.2
and 5.8 of Flores-Bazán et al. [189].
Theorem 9.5.1. Let X be compact. Assume that one of the following conditions
holds:
(i) f is t-m.l.s.c.;

(ii) f 1 Yt .y/ is closed for every y 2 f .X /I
(iii) Y is a topological space, f 1 cl Yt .y/ is closed for every y 2 f .X /
and (9.19) holds.
Then f .X / has the domination property. In particular, problem .P / has optimal
solutions.
Proof. Set A WD f .X / Y:
(i) Let us apply Corollary 9.2.2 (i). For this take .yi /i 2I A be t-decreasing.
Hence, there exists .xi /i 2I X such that yi D f .xi / for every i 2 I . Because
X is compact, taking into account [417, Definition 2.1.26, Proposition 2.1.37], there
exists J a directed set and W J ! I an increasing
function
with .J / cofinal
such that .x .j / /j 2J converges to x 2 X . Since f .x .j / / j 2J is t-decreasing and
f is t-m.l.s.c., it follows that .f .x/; y .j / / 2 t for every j 2 J . Since .J / is
cofinal and .yi /i 2I is t-decreasing, we get .f .x/; yi / 2 t for every i 2 I . Using
Corollary 9.2.2 (i) we get the conclusion.
(ii) ) (i) (However, a direct proof of (ii), using Corollary 9.2.3, is possible.)
Indeed, take .xi /i 2I X such that xi ! x and .f .xi //
i 2I is t-decreasing. Setting
yi WD f .xi / 2 f .X /, we have that xj 2 f 1 Yt .yi / for all i; j 2 I with j
i .
9.5 Existence of Optimal Solutions for Vector and Set Optimization Problems 363

Since f 1 Yt .yi / .xj /j i ! x, it follows that x 2 cl f 1 Yt .yi / D
f 1 Yt .yi / , and so f .x/ 2 Yt .yi / for every i 2 I:
(iii) In order to apply Corollary
T
let us consider D A a totally ordered
9.2.9,
set and assume that A \ Y .y/ D ;, where s is defined by (9.21). Then
S
y2D s
A y2D Y n Ys .y/ . It follows that
S S
X f 1 y2D Y n Ys .y/ D y2D f
1
.Y n Ys .y// . X /:

Because X is compact and f 1 .YSn Ys .y// D X n f 1 Ys .y/ is open, there
exists D0 2 F .D/ such that X D y2D0 f 1 .Y n Ys .y//, and so
S 1 S
A D f .X / D y2D0 f f .Y n Ys .y// y2D0 Y n Ys .y/ :
The set D0 being finiteS and totally

w.r.t. t, there exists y0 2 D0 . A/ such
ordered
that Y n Ys .y0 / D y2D0 Y n Ys .y/ , and so A Y n Ys .y0 /. It follows that
y0 2 A \ cl Yt .y0 / D ;, a contradiction. Hence, by Corollary 9.2.9, f .X / has the
domination property. t
u
When Y is a topological vector space ordered by the proper convex cone C;

t WDC , as seen in Sect. 9.3, Yt .y/ D y C: In such a situation the fact that
taking
1
f Yt .y/ is closed for every y 2 Y means that f is C -lower semicontinuous.
So, in this special case Theorem 9.5.1 becomes the following result.
Corollary 9.5.2. Let X be compact. Assume that Y is a topological vector space
ordered by the proper closed convex cone C . If either (a) f is C -lower semi-
continuous or (b) f is cl C -lower semicontinuous and cl C C .C n .C // C ,
then f .X / has the domination property. In particular, problem .P / has optimal
solutions.
Consider now the set optimization problem
Min F .x/ s.t. x 2 X; (VP)
where F W X Y is a set-valued map; one says that x0 2 X is a solution of

problem (VP) if F .x0 / \ Min.F .X /; t/ ¤ ;:
The next result is obtained in Flores-Bazán et al. [189, Theorem 7.2].
Theorem 9.5.3. Assume that Y is a topological space and there exists y0 2 F .X /
such that F 1 .Yt .y// is compact for every y 2 Yt .y0 /. If F .x/ \ Yt .y0 / verifies
condition (9.9) for all x 2 X , then the problem (VP) has optimal solutions.
Proof. It is sufficient to show that Min .A; t/ is nonempty, where A WD F .X / \
Yt .y0 / D .F .X //
t .y0 / 3 y0 . For this let us prove that condition
T (9.5) holds.
So,

consider .yi /i 2I A a t-decreasing net and assume that A \ Y
i 2I t .yi / D ;.
For i 2 I set

Ui WD fx 2 X j F .x/ \ Yt .yi / D ;g D X n F 1 Yt .yi / :

Clearly, F 1 Yt .yi / ¤ ; and Ui Ui 0 for all i; i 0 2 I with i
i 0 . Because
yi 2 Yt .y0 /, we have that Ui is open for every i 2 I . Moreover,
S
F 1 Yt .y0 / i 2I Ui : (9.29)

Indeed, if x 2 F 1 Yt .y0 / then F .x/ \ Yt .y0 / ¤ ;. Because
S F .x/\ Yt .y0 /
F .X / \ Yt .y0 / D A we obtain that F .x/ \ Yt .y0 / i 2I S Y n Yt .yi / . Using
condition (9.9) we get I0 2 F .I / such that F .x/ \ Yt .y0 / i 2I0 Y n Yt .yi / .
The subset I0 of I being finite, there exists ix 2 I such that ix
i for every i 2 I0 .
It follows that Y n Yt .yi / Y n Yt .yix / for i 2 I0 , and so F .x/ \ Yt .y0 /
Y n Yt .yix /. Moreover, taking into account that Y n Yt .y0 / Y n Yt .yix /, we get

F .x/ D F .x/ \ Yt .y0 / [ F .x/ \ .Y n Yt .y0 // Y n Yt .yix /;

whence x 2 Uix . Therefore, (9.29) holds.Because
S F 1 Yt .y0 / is compact, there
exists I0 2 F .I / such that F 1 Yt .y0 / i 2I0 Ui . Taking i0 2 I such that i0

i for every i 2I0 we obtain
that F1 Yt .y0 / Ui0 , and so ; D F 1 Yt .y0 / \
F 1 Yt .yi0 / D F 1 Yt .yi0 / , a contradiction. Applying Corollary 9.2.2 we
obtain that Min .A; t/ ¤ ;. t
u
An immediate consequence of the preceding result is the following (see [189,
Corollary 7.4]).

Corollary 9.5.4. Assume that X is compact, F 1 Yt .y/ is closed for every y 2
Y and F .x/ \ Yt .y/ verifies condition (9.9) for all x 2 X and y 2 F .X /. Then
F .X / has the domination property.
In the case in which Y is a topological vector space ordered by a proper convex
cone C we have the following result.
Theorem 9.5.5. Let dom F be compact and Y be a topological space. Assume that
F is upper continuous and F .x/ verifies condition (9.9) with t replaced by s defined
by (9.21) for all x 2 dom F . If (9.19) holds, then F .X / has the domination property.
Proof. W.l.o.g. we suppose that X D dom F . In order to apply Proposition 9.2.5
let us show that A WD F .X / verifies condition (9.9) with t replaced by s defined
by (9.21). As observed in the proof of Corollary 9.2.9, conditionS(9.19) is (9.10) and

condition (9.20) is (9.12) in this case. Let E Y with A y2E Y n Ys .y/ .
Take x 2 X . Because F .x/S A and F .x/ verifies
condition
S Ex 2
(9.9), there exists
F .E/ such that F .x/ y2Ex Y n Ys .y/ D y2Ex Y n cl Yt .y/ DW Dx .
Since Dx is open and F is upper continuous at x, there exists an open neighborhood
Ux of x such that F .Ux / Dx . Since X is compact and X D [x2X Ux , there exists
X0 2 F .X / such that X D [x2X0 Ux . Taking E0 WD [x2X0 Ex 2 F .E/ we
have that

A D F .X / D F .[x2X0 Ux / D [x2X0 F .Ux / [x2X0 Dx D [y2E0 Y n Ys .y/ :
Because (9.9) with s instead of t implies (9.12), the conclusion follows. t

u
Note that in the case in which Y is a topological vector space ordered by a

(proper) convex cone C , condition (9.19) for t DC is equivalent to cl C C .C n
.C // C , and Ys .y/ D cl Yt .y/ D y cl C for every y 2 Y ; in this case,
instead of asking F to be upper continuous it is sufficient to ask that F is C -upper
continuous, getting so [189, Theorem 7.5, Corollary 7.6].
In the rest of this section we consider Y a topological vector space ordered
by a (proper) convex cone C . On the family P0 .Y / of nonempty sets of Y we
consider the relation l WDlC and u WDuC introduced by Kuroiwa [350] (see also
Definition 2.6.9) in the following way:
A l B if B A C C; A u B if A B C:
Note that
A uC B , .B/ lC .A/ , B lC A: (9.30)
It follows that for A; B 2 P0 .Y / one has
A 2 .P0 .Y //
l .B/ , Œ8b 2 B W A \ .b C / ¤ ; ; (9.31)
A 2 P0 .Y / n .P0 .Y //
l .B/ , Œ9b 2 B W A Y n .b C / (9.32)
and
A 2 P0 .Y / n .P0 .Y //
u .B/ , A \ ŒY n .B C / ¤ ;: (9.33)
It is clear that l and u are reflexive and transitive. So, all the results stated in
Sects. 9.1, 9.2 and the present one for a transitive relation t can be adapted for l
and u .
In order to translate Corollary 9.2.2 (i) to the present context for t Dl let us
introduce some more notation. Let us fix A P0 .Y / and take E A ; we set
CE WD CEA WD fH HE j 8A 2 A ; 9b 2 H W A Y n .b C /g ;
where
S
HE WD fE j E 2 E g : (9.34)
Clearly,
E 0 E A ) HE 0 HE HA ; CE 3 H H 0 HE ) H 0 2 CE :
Consequently, if E A , from (9.32) we have that

S
A B2E P0 .Y / n .P0 .Y //
l .B/ , HE 2 CE , CE ¤ ;: (9.35)
We have the following result stated in [250, Theorem 4.1].

Corollary 9.5.6. Assume that ; ¤ A P0 .Y / verifies the condition
8.Ai /i 2I A W .Ai /i 2I l -decreasing ) CEI D ;; (9.36)
where EI WD fAi j i 2 I g A . Then A has the domination property with respect

to l ; in particular Min.A ; l / ¤ ;:
Proof. Let us show that (9.5) holds for t WDl , that is

8.Ai /i 2I A ; .Ai /i 2I l -decreasing ; 9A 2 A ; 8i 2 I W A l Ai :
(9.37)
For this, take a l -decreasing net .Ai /i 2I A . By (9.35) we have that HEI 62
CEI , that is there exists A 2 A such that for all i 2 I and b 2 Ai we have that
A \ .b C / ¤ ;. This can be rewritten as Ai A C C , that is A l Ai , for
every i 2 I . Hence (9.37) holds. Using Corollary 9.2.2 (i) we obtain that A has the
domination property w.r.t. l . t
u
It is easy to prove that (9.37) ) (9.36), too.
The transcription of (9.8) for A P0 .Y / using (9.35) is
Œ8E A ; CE ¤ ; ; 9E 0 2 F .E / W CE 0 ¤ ;: (9.38)
A stronger condition is
8E A W CE ¤ ; ) CE \ F .HE / ¤ ;; (9.39)
and an even stronger condition is
8H 2 CA W CA \ F .H / ¤ ;: (9.40)
Of course, using Corollary 9.2.3, if A P0 .Y / verifies (9.40) then A has the

domination property w.r.t. l .
In the present context we consider the set optimization problem
Min F .x/ s.t. x 2 X; (SP)
where F W X Y is a set-valued map with nonempty domain; one says that

x0 2 dom F is a l (resp. u ) solution of problem (SP) if F .x0 / 2 Min.AF ; l /
(resp. F .x0 / 2 Min.AF ; u /), where
AF WD fF .x/ j x 2 dom F g: (9.41)

The next theorem is the first result established in the literature concerning the
existence of solutions for problem (SP) w.r.t. l , and is due to Kuroiwa [352,
Theorem 4.1] (see also [250, Theorem 5.8]). Before stating the result let us introduce
the limit superior of a net .Ei /i 2I of subsets of the topological space .X; / as being
the set of those x 2 X such that there exist X .xj /j 2J ! x and W J ! I
increasing and cofinal (that is .J / is cofinal in I ) such that xj 2 E .j / for every
j 2 J . Let us observe that
lim sup E .j / lim sup Ei (9.42)

j 2J i 2I
for each directed set J and each increasing cofinal mapping W J ! I;

moreover, one has equality in (9.42) if .Ei /i 2I is increasing. Observe also that
Ei0 lim supi 2I Ei for each i0 2 I when .Ei /i 2I is increasing.
9.5.7. Let X D dom F be compact. Assume that F .x/

l
Theorem
lim sup F .xi / C C whenever .xi /i 2I X is such that xi ! x 2 X and
.F .xi //i 2I is l -decreasing. Then AF WD fF .x/ j x 2 dom F g has the domination
property with respect to l . In particular, (SP) has optimal solutions with respect
to l .
Proof. Let us consider A WD fF .x/ C C j x 2 X g. Observe that A and
AF have the same minimal points w.r.t. l . We apply Corollary 9.5.6 to A . For
this take a l -decreasing net .Ai /i 2I A , that is Ai Ai 0 for all i; i 0 2 I
with i 0
i . Then Ai D F .xi / C C for some .xi /i 2I X . It follows that
.F .xi //i 2I is l -decreasing. Because X is compact, taking into account [417,
Definition 2.1.26, Proposition 2.1.37], there exists J a directed set and an increasing
cofinal mapping W J ! I such that .x .j / /j 2J converges to x 2 X . Since
j 2J is -decreasing. Using our assumption, F .x/
l l
is increasing,
.F .x .j / //
lim supj 2J F .x .j / /CC D lim supj 2J A .j / . Since .A .j / /j 2J is increasing, we
have that A .j / lim supj 2J A .j / A WD F .x/ C C for every j 2 J . Because
.Ai / is increasing and is cofinal we obtain that Ai A, that is A l Ai , for every
i 2 I . Therefore, (9.37) holds, and so (9.36) holds, too. Applying Corollary 9.5.6
we obtain that A has the domination property. t
u
Using the considerations above we get the next result established essentially by
Alonso and Rodríguez-Marín in [3] (see also [189, Theorem 6.4]).
Theorem 9.5.8. Assume that X D dom F is compact and F is C -upper continu-
ous. If C is closed, then AF WD fF .x/ j x 2 dom F g has the domination property
with respect to l . In particular, (SP) has optimal solutions with respect to l .
Proof. Let us show that condition (9.40) holds. Take H 2 CAF . For x 2 X there
exists bx 2 H such that F .x/ Y n .bx C /. Because C is closed and F is
C -u.c. at x, there exists an open neighborhood Ux of x 2 X such that F .Ux /
ŒY n .bx C / C C D Y n .bx C /. Because X D [x2X Ux , there exists X0 2
F .X / such that X D [x2X0 Ux . Set H 0 WD fbx j x 2 X0 g 2 F .H /. For x 0 2 X we
have that x 0 2 Ux for some x 2 X , and so F .x 0 / Y n .bx C /; hence H 0 2 CAF .

Therefore, condition (9.40) holds, and so (9.38) holds, too. The conclusion follows
using Corollary 9.2.3. t
u
We present now an existence result for problem (SP) w.r.t. u : First observe that
using (9.33) we have that
S
A B2E P0 .Y / n .P0 .Y //
u .B/ , 8A 2 A ; 9B 2 E W A \ ŒY n .B C / ¤ ;:
(9.43)
The next result is practically established in Alonso–Rodríguez-Marín [3, Propo-

sition 22].
Theorem 9.5.9. Assume that X D dom F is compact and F is C -lower continuous
with F .x/ C closed for every x 2 X . Then AF has the domination property with
respect to u . In particular, (SP) has optimal solutions with respect to u .
S
Proof. Let E A F be such that A B2E P0 .Y / n .P0 .Y // u .B/ ; hence
E D fF .x/ j x 2 Eg for some E X . Take x 2 X ; using (9.43), there exists
ux 2 E such that F .x/ \ ŒY n .F .ux / C / ¤ ;. Because F .ux / C is closed and
F is C -l.c., there exists an open neighborhood Ux of x such that

F .x 0 / \ Y n .F .ux / C / C C D F .x 0 / \ ŒY n .F .ux / C / ¤ ; 8x 0 2 Ux :
Since X is compact and X D [x2X Ux , there exists X0 2 F .X / such that

X D [x2X0 Ux . Taking E0 WD fux j x 2 X0 g .2 F .E// andSE0 WD fF .x/ j
x 2 E0 g .2 F .E //, and using again (9.43), we have that A B2E0 P0 .Y / n

.P0 .Y //
u .B/ . The conclusion follows using Corollary 9.2.3. t
u
Chapter 10
Ekeland Variational Principle
In this chapter we present existence results for minimal points of subsets of the
Cartesian product of a complete metric space and a topological vector space with
respect to order relations determined by generalized set-valued metrics; such results
are useful for deriving EVP type results for vector and scalar functions. Then we
derive EVP results of Ha’s type as well as an EVP result for bi-set-valued maps.
We end the chapter with an application to error bounds for set-valued optimization
problems.
The celebrated Ekeland variational principle (see [171]) has many equivalent
formulations and generalizations. Quite rapidly after the publication of the Ekeland
variational principle (EVP) in 1974 there were formulated extensions to functions
f W .X; d / ! Y , where Y is a real (topological) vector space. A systematization
of such results was done in [215] (see also [214]), where instead of a function f it
was considered a subset of X Y ; said differently, it was considered a set-valued
map from X to Y . The common feature of these results is the presence of a certain
term d.x; x 0 /k 0 in the perturbed objective function, where K is the convex ordering
cone in Y and k 0 2 K n f0g. Very recently d.x; x 0 /k 0 was replaced by d.x; x 0 /H
with H a bounded convex subset of K (see [40]) or by F .x; x 0 / K, F being a so
called K-metric (see [220]); in both papers one deals with functions f W X ! Y .
Our approach is that of [558], being similar to that in [215]. The results without
bibliographical references can be found in [558] or are new.
10.1 Preliminary Notions and Results
In the sequel .X; d / is a complete metric space, Y is a real topological vector space,
Y is its topological dual, and K Y is a proper convex cone; as usual, K C is the
positive dual cone of K and K # is the quasi interior of K C :

DOI 10.1007/978-3-642-54265-7__10
370 10 Ekeland Variational Principle
K C D fy 2 Y j y .y/
0 8 y 2 Kg;
K # D fy 2 Y j y .y/ > 0 8 y 2 K n f0gg:
If Y is just a real linear space we (can) endow it with the finest locally convex
topology, that is, the core convex topology (see Sect. 6.3).
We say that E Y is K C -bounded if y .E/ is bounded from below for every
y 2 K C . It is clear that any quasi (K-lower) bounded set is K C -bounded.

Let F W X X K satisfy the conditions:

(F1) 0 2 F .x; x/ for all x 2 X;
(F2) F .x1 ; x2 / C F .x2 ; x3 / F .x1 ; x3 / C K for all x1 ; x2 ; x3 2 X:
Using F we introduce a quasi-order on X Y , denoted by F , in the following
manner:
.x1 ; y1 / F .x2 ; y2 / ” y2 2 y1 C F .x1 ; x2 / C K: (10.1)
Indeed, F is reflexive by (F1). If .x1 ; y1 / F .x2 ; y2 / and .x2 ; y2 / F .x3 ; y3 /,

then
y2 D y1 C v1 C k1 ; y3 D y2 C v2 C k2 (10.2)
with v1 2 F .x1 ; x2 /, v2 2 F .x2 ; x3 / and k1 ; k2 2 K. By (F2) we have that v1 Cv2 D

v3 C k3 for some v3 2 F .x1 ; x3 / and k3 2 K, and so
y3 D y1 C v1 C k1 C v2 C k2 D y1 C v3 C k1 C k2 C k3 2 y1 C F .x1 ; x3 / C KI
hence .x1 ; y1 / F .x3 ; y3 /, and so F is transitive. Of course,
.x1 ; y1 / F .x2 ; y2 / ) y1 K y2 I (10.3)
moreover, by (F1), we have that
.x; y1 / F .x; y2 / ” y2 2 y1 C K ” y1 K y2 : (10.4)
Besides conditions (F1) and (F2) we shall consider also the condition
(F3) there exists z 2 K C such that
˚
8ı > 0 W .ı/ WD inf z .v/ j v 2 [d.x;x 0 /ı F .x; x 0 / > 0: (10.5)
Let us set
Fı WD [d.x;x 0 /ı F .x; x 0 / (10.6)

10.1 Preliminary Notions and Results 371
for ı
0; clearly, Fı Fı0 if 0 ı 0 < ı. So, condition (F3) can be rewritten as
9z 2 K C ; 8ı > 0 W inf z .Fı / > 0: (10.7)
A weaker condition is
8ı > 0; 9z 2 K C W inf z .Fı / > 0: (10.8)
An even weaker condition is the following
8ı > 0; 8.zn / Fı ; 9z 2 K C W lim sup z .zn / > 0: (10.9)
Clearly, if (F3) holds then 0 … convF .x; x 0 / for x ¤ x 0 :

Condition (F3) holds also if one has
8x; x 0 2 X W inf z .z/

d.x; x 0 /: (10.10)
z2F .x;x 0 /
In the next lemma we give an important example of set-valued map F satisfying

conditions (F1) and (F2).
Lemma 10.1.1. Let ; ¤ H K be a K-convex set, that is H C K is convex.
Consider
FH W X X K; FH .x; x 0 / WD d.x; x 0 /H: (10.11)
(i) FH verifies (F1) and (F2).

(ii) FH verifies condition (F3) iff FH verifies condition (10.8) iff there exists
z 2 K C such that inf z .H / > 0; if Y is a separated locally convex space,
then FH verifies condition (F3) iff 0 … cl.H C K/. Moreover, FH verifies
condition (10.9) iff
8.hn / H; 9z 2 K C W lim sup z .hn / > 0: (10.12)
Proof. (i) Clearly FH .x; x/ D f0g for every x 2 X . Take x1 ; x2 ; x3 2 X and

v1 2 FH .x1 ; x2 /, v2 2 FH .x2 ; x3 /; hence d.x1 ; x2 /Cd.x2 ; x3 / D d.x1 ; x3 /C˛
for some ˛ > 0 and v1 D d.x1 ; x2 /h1 , v2 D d.x2 ; x3 /h2 with h1 ; h2 2 H .
Using the convexity of H C K we get
v1 C v2 2 d.x1 ; x2 /H C d.x2 ; x3 /H d.x1 ; x2 /.H C K/ C d.x2 ; x3 /.H C K/

D Œd.x1 ; x2 / C d.x2 ; x3 / .H C K/ D Œd.x1 ; x3 / C ˛ .H C K/
d.x1 ; x3 /H C K D FH .x1 ; x3 / C K:
Hence (F2) is verified, too.

(ii) is obvious. t
u
If (10.10) holds then
.x1 ; y1 / F .x2 ; y2 / ) d.x1 ; x2 / z .y2 / z .y1 /: (10.13)
Indeed, since F .x1 ; x2 / K, from (10.3) we get first that y1 K y2 ; then

from (10.2)
z .y2 / D z .y1 / C z .v1 / C z .k1 /

z .y1 / C inf z .v/
z .y1 / C d.x1 ; x2 /;
v2F .x1 ;x2 /
and so (10.13) holds. Using (10.13) we obtain that

Œ.x1 ; y1 / F .x2 ; y2 /; .x2 ; y2 / F .x1 ; y1 / ) x1 D x2 ; z .y1 / D z .y2 / :
(10.14)
In fact (10.14) holds if F verifies conditions (F1)–(F3). Indeed for ı WD d.x1 ; x2 /,

from (10.2) we get
z .y1 /
z .y2 / C inf z .Fı /; z .y2 /
z .y1 / C inf z .Fı /;
and so jz .y1 / z .y2 /j

inf z .Fı /
0:
Equation (10.14) shows that F is anti-symmetric (and so F is a partial order)
when F verifies conditions (F1)–(F3) with z 2 K # .
For F satisfying conditions (F1) and (F2), and z 2 K C , we introduce the partial
order F;z on X Y by

.x1 ; y1 / D .x2 ; y2 / or
.x1 ; y1 / F;z .x2 ; y2 / ”
.x1 ; y1 / F .x2 ; y2 / and z .y1 / < z .y2 /:
It is easy to verify that F;z is reflexive, transitive, and antisymmetric. Of course,

for z D 0 the relation F;z coincides with the equality relation on X Y , while
for z 2 K # and 0 … F .x; x 0 / for x ¤ x 0 one has that F;z and F coincide.
The following result is established in [484].
Lemma 10.1.2. Let Y be a separated locally convex space. One has (10.8) )
(10.7) [that is, (10.8) ) (F3)] in each of the following situations:
(i) for every sequence .zn /n1 K C there exists a sequence .n /n1 P such

that convw fn zn j n
1g is w -compact;
(ii) K C has a weakly-star compact base (e.g. if int K ¤ ;);
(iii) for every sequence .zn /n1 Z there exists a sequence .n /n1 P such

that convw fn zn j n
1g is w -compact;
(iv) Y is a normed vector space.
10.2 Minimal Points in Product Spaces 373
Proof. Note first that (iv) ) (iii) ) (i). Indeed, the implication (iii) ) (i)
is obvious. For (iv) ) (iii) just note that having .zn /n1 Y , there exists
.n /n1 P such that n zn 2 UY WD fz j kz k 1g for every n
1, and

so convw fn zn j n
1g is w -compact.
(ii) ) (i) Let B be a weakly-star compact base of K C . It follows that B0 WD
Œ0; 1B is a w -compact convex subset of K C such that K C D RC B0 . This
shows that for every z 2 K C there exists > 0 such that z 2 B0 . So, for
.zn /n1 K C , there exists .n /n1 P such that fn zn j n
1g B0 , whence

convw fn zn j n
1g is w -compact.
(i) Assume that (10.8) holds. Then for every n 2 N there exists zn 2 K C such
that n WD inf zn .F1=n / > 0. By hypothesis, there exists a sequence .n /n1 P

such that B WD convw fn zn j n
1g is w -compact, and so

B0 WD Œ0; 1B D convw f0g [ fn zn j n
1g
P
is w -compact, too. Then z WD n
n1 2 n zn exists and belongs to B0
C
K
Pn . Indeed, because B0 is w -bounded, the sequence .z0 0
n /n1 with zn WD
k 0
kD1 2 k zk 2 B0 is w -Cauchy. Since .zn /n1 has a subnetP w -converging
w
to some z 2 B0 , we obtain that z0
n ! z . It follows that z D n1 2n n zn .
Clearly, for z 2 F1=n we have that z .z/
2n n zn .z/, and so inf z .F1=n /
2n n n > 0. Since for every ı > 0 we have that ı

1=n for some n 2 N , we
obtain that (10.5) holds. t
u
10.2 Minimal Points in Product Spaces
We take X; Y; K; F as above, that is, F verifies conditions (F1) and (F2).

Consider a nonempty set A X Y . In the sequel we shall use the following
condition on A:
(H1) for every F –decreasing sequence ..xn ; yn // A with xn ! x 2 X there
exists y 2 Y such that .x; y/ 2 A and .x; y/ F .xn ; yn / for every n 2 N.
The next theorem is the main result of this section.
Theorem 10.2.1. Let .X; d / be a complete metric space, Y a real topological
vector space and K Y a proper convex cone. Assume that:
(i) F W X X Y verifies conditions (F1), (F2) and (10.9), and
(ii) A X Y verifies (H1).
Then for every .x0 ; y0 / 2 A and every z 2 K C such that z .PrY .A// is bounded
from below, there exists a minimal element .x; N y/
N of A with respect to F;z such
N F;z .x0 ; y0 /. Moreover, if z .z/ > 0 for every z 2 [ı>0 Fı , then A 3
N y/
that .x;
.x; y/ F .x;
N y/
N implies x D x:
Proof. Let us fix .x0 ; y0 / 2 A and z 2 K C such that z .PrY .A// is bounded from
below. Then
˛ WD inf fz .y/ j 9x 2 X W .x; y/ 2 A; .x; y/ F;z .x0 ; y0 /g 2 R:
Let us set
˚
AF;z .x; y/ WD .x 0 ; y 0 / 2 A j .x 0 ; y 0 / F;z .x; y/ :
We construct a sequence ..xn ; yn //n0 A as follows: Having .xn ; yn / 2 A, we

take .xnC1 ; ynC1 / 2 AF;z .xn ; yn / such that
z .ynC1 / inffz .y/ j .x; y/ 2 AF;z .xn ; yn /g C 1=.n C 1/:
Of course, ..xn ; yn // is F;z –decreasing. It follows that .yn /n0 PrY .A/ is K –
decreasing, and so the sequence .z .yn //n0 is non increasing and bounded from
below; hence WD lim z .yn / 2 R.
If AF;z .xn0 ; yn0 / is a singleton (that is, AF;z .xn0 ; yn0 / D f.xn0 ; yn0 /g) for some
n0 2 N, then clearly .x; y/ WD .xn0 ; yn0 / is the desired element. In the contrary
case the sequence .z .yn // is (strictly) decreasing; moreover, < z .yn / for every
n 2 N.
Assume that .xn / is not a Cauchy sequence. Then there exist ı > 0 and the
sequences .nl /, .pl / from N such that nl ! 1 and d.xnl ; xnl Cpl /
ı for every
l. Since .xnl Cpl ; ynl Cpl / F;z .xnl ; ynl / we obtain that ynl ynl Cpl D vl C kl0
with vl 2 F .xnl Cpl ; xnl / Fı and kl0 2 K. By (10.9), there exists v 2 K C such
that v .vl / 6! 0. On the other hand, because .yn /n0 PrY .A/ and v 2 K C ,
.v .yn //n0 is bounded from below. The sequence .yn /n0 being K–decreasing we
obtain that lim v .yn / exists in R. Since v .ynl /v .ynl Cpl /
z .vl /
0 for every
l 2 N, we get the contradiction v .vl / ! 0. Therefore, .xn / is a Cauchy sequence
in the complete metric space .X; d /, and so .xn / converges to some xN 2 X .
Since ..xn ; yn // is F –decreasing, by (H1) there exists some yN 2 Y such that
N y/
.x; N 2 A and .x; N y/N F .xn ; yn / for every n 2 N. It follows that z .y/ N

lim z .yn /, and so z .y/ N < z .yn / for every n 2 N. Therefore .x; N y/
N F;z .xn ; yn /
for every n 2 N. Let .x 0 ; y 0 / 2 A be such that .x 0 ; y 0 / F;z .x; N y/.
N Since
N F;z .xn ; yn /, we have that .x 0 ; y 0 / F;z .xn ; yn / for every n 2 N. It
N y/
.x;
follows that
8 n
1 W 0 z .y/
N z .y 0 / z .yn / z .y 0 / 1=n;
whence z .y 0 / D z .y/.
N By the definition of F;z we obtain that .x 0 ; y 0 / D .x;
N y/.
N
Assume now that z .z/ > 0 for every z 2 [ı>0 Fı . Consider A 3 .x; y/ F
N y/
.x; N and assume that x ¤ x. Then y D y C v C k with v 2 F .x; x/ and k 2 K.
By our assumption we have that z .v/ > 0, and so z .y/ < z .y/. It follows that
.x; y/ F;z .x; N y/,
N contradicting the minimality of .x;
N y/
N with respect to F;z .
Hence x D x. t
u
Remark 10.2.2. As seen from the proof, to have the conclusion of Theorem 10.2.1
for a fixed .x0 ; y0 / 2 A, in its hypothesis, it is sufficient that AF;z .x0 ; y0 / verify
(H1) and z .PrY .AF;z .x0 ; y0 /// be bounded from below. For this just replace A by
AF;z .x0 ; y0 / and apply the preceding theorem.
Theorem 10.2.3. Assume that .X; d / is a complete metric space, Y is a real
topological vector space and K Y is a proper convex cone. Let F W X X Y
satisfy conditions (F1)–(F3) and A X Y satisfy (H1). Furthermore, suppose that
z (from (F3)) is bounded from below on PrY .A/. Then for every .x0 ; y0 / 2 A there
exists a minimal element .x;N y/
N of A with respect to F;z such that .x;N y/
N F;z
N y/
.x0 ; y0 /. Moreover, .x; N F .x0 ; y0 / and A 3 .x; y/ F .x;
N y/
N implies x D x:
Proof. The proof is the same as that of Theorem 10.2.1. One must observe that for
every ı > 0 we can take as z in (10.9) the element z provided by (F3). t
u
Remark 10.2.4. As in Remark 10.2.2, to have the conclusion of Theorem 10.2.3 for
a fixed .x0 ; y0 / 2 A, in its hypothesis, it is sufficient that AF;z .x0 ; y0 / verify (H1)
and z be bounded from below on the set AF;z .x0 ; y0 /.
As seen above, a possible choice for F is FH given in (10.11) with ; ¤ H K
a K-convex set; if z 2 K C is such that inf z .H / > 0, then conditions (F1)–(F3)
are verified (see Lemma 10.1.1). In the case H WD fk 0 g condition (H1) becomes
condition (H1) in [214, p. 199]. So Theorem 10.2.3 extends [214, Theorem 3.10.7]
to this framework, using practically the same proof.
For H as above, that is, H K n f0g is a nonempty set such that H C K is
convex, and FH defined in (10.11), we set H WDFH ; moreover, for z 2 K C we
set H;z WDFH ;z .
An immediate consequence of the preceding theorem is the next result.
Corollary 10.2.5. Assume that .X; d / is a complete metric space, Y is a real
topological vector space, K Y is a proper convex cone and ; ¤ H K
is a nonempty K-convex set. Let A X Y satisfy (H1) with respect to H .
If there exists z 2 K C such that inf z .H / > 0 and inf z .PrY .A// > 1,
then for every .x0 ; y0 / 2 A there exists .x;
N y/
N 2 A a minimal element of A with
respect to H;z such that .x;
N y/
N H;z .x0 ; y0 /. Moreover, .x;
N y/
N H .x0 ; y0 / and
A 3 .x; y/ H .x; N y/
N implies x D x:
Of course, a variant of Corollary 10.2.5 can be formulated corresponding to
condition (10.12).
A condition related to (H1) is the next one.

(H2) for every sequence .xn ; yn / A with xn ! x 2 X and .yn / K –
decreasing there exists y 2 Y such that .x; y/ 2 A and y K yn for every
n 2 N.
Remark 10.2.6. Note that (H2) holds if A is closed with PrY .A/ y0 C K for
some y0 2 Y and every K –decreasing sequence in K is convergent (i.e., K is
a sequentially Daniell cone). In fact, instead of asking that A is closed we may

assume that
8 ..xn ; yn //n1 A W Œxn ! x; yn ! y; .yn / is K –decreasing ) .x; y/ 2 A :
Remark 10.2.7. Note that (H1) is verified whenever A satisfies (H2) and
8u 2 X; 8X .xn / ! x 2 X W \n2N .F .xn ; u/ C K/ F .x; u/ C K:

Indeed, let .xn ; yn / A be F –decreasing with xn ! x. It is obvious that
.yn / is K –decreasing. By (H2), there exists y 2 Y such that .x; y/ 2 A and
y K yn for every n 2 N. It follows that
8 n; p 2 N W yn 2 ynCp C F .xnCp ; xn / C K y C F .xnCp ; xn / C K:
Fix n; then yn y 2 F .xnCp ; xn /CK for every p 2 N, and so, by our hypothesis,
yn y 2 F .x; xn / C K because limp!1 xnCp D x. Therefore, .x; y/ F .xn ; yn /:
Remark 10.2.8. When d generates the discrete topology on X , both conditions (H1)
and (H2) are equivalent to:
(H12) for every x 2 PrX .A/ and every K –decreasing sequence .yn /n1
A.x/ WD fy 2 Y j .x; y/ 2 Ag there exists y 2 A.x/ such that y K yn for
every n
1.
Indeed, it is clear that (H1) ^ (H2) ) (H12) (for any metric d ). Assume that
(H12) holds and take ..xn ; yn //n1 A a 4F -decreasing sequence with xn ! x;
in particular, .yn / is K –decreasing. Since xn ! x, there exists n0
1 such that
xn D x for n
n0 . It follows that .yn /nn0 A.x/. By (H12) we have that
y 2 A.x/, and so .x; y/ 4F .x; yn / D .xn ; yn / 4F .xn0 ; yn0 / 4F .xm ; ym / for
n n0 m. Hence (H1) holds. The proof of (H12) ) (H2) is similar.
Remark 10.2.9. In the case F D FH with ; ¤ H K a K-convex set, (H1) is
A satisfies
verified whenever (H2) and H C K is closed.
Indeed, let .xn ; yn / A be a H –decreasing sequence with xn ! x. It is
obvious that .yn / is K –decreasing. By (H2), there exists y 2 Y such that .x; y/ 2
A and y K yn for every n 2 N.
Fix n. If xn D x then clearly .x; y/ D .xn ; y/ H .xn ; yn /. Else, because
d.xnCp ; xn / ! d.x; xn / > 0 for p ! 1, we get d.xnCp ; xn / > 0 for sufficiently
large p, and so
yn 2 ynCp C d.xnCp ; xn /H C K y C d.xnCp ; xn /H C K D y C d.xnCp ; xn /.H C K/
for sufficiently large p. Since H C K is closed we obtain that
yn 2 y C d.xn ; x/.H C K/ D y C d.xn ; x/H C K;
that is, .x; y/ H .xn ; yn /:

Another condition to be added to (H2) in order to have (H1) was suggested by the
hypotheses of [40, Theorem 4.1]. Recall that a set E Y is cs-complete
P (see [614,
p. 9]) if for all sequences .
n /n1 RC , .yn /n1 E such that n1
n D 1
Pn P
and the sequence mD1
m ym n1 is Cauchy, the series n1
n yn is convergent
and itsPsum belongs to E. One says that P E Y is cs-closed if the sum of the
series n1
n yn belongs
P to E whenever n1
n yn is convergent and .yn / E,
.
n /n1 RC and n1
n D 1. Of course, any cs-complete set is cs-closed; if
Y is complete then the converse is true. Moreover, notice that any cs-closed set is
convex.
It is worth observing that the closed convex subsets as well as the open convex
subsets of a separated locally convex space are cs-closed; moreover, all the convex
subsets of finite dimensionalPnormed spaces
are cs-closed (hence cs-complete).
n
Note that the sequence
P mD1 m m n1 is Cauchy whenever .
n /n1 RC is

y
such that the series n1
n is convergent and .yn /n1 Y is such that convfyn j
n
1g is bounded; of course, if Y is a locally convex spacePthen B Y is bounded
iff conv B is bounded. Indeed, let .
n /n1 RC with n1
n convergent and
.yn /n1 Y with B WD convfyn j n
1g bounded. Fix V Y a balanced
neighborhood ofP 0. Because B is bounded, there exists ˛ > 0 such that PB ˛V .
nCp
Since the series n1
n is convergent there exists n0
1 such that kDn
k
˛ 1 for all n; p 2 N with n
n0 . Then for such n; p and some bn;p 2 B we have
X
nCp nCp
X

k yk D
k bn;p 2 Œ0; ˛ 1 B Œ0; ˛ 1 ˛V D V:
kDn kDn
Proposition 10.2.10. Assume that .X; d / is a complete metric space, Y is a real

topological vector space and K Y is a proper closed convex cone. Furthermore,
suppose that H K is a nonempty cs-complete bounded set with 0 … cl.H C K/.
If A satisfies (H2) then A satisfies (H1), too.

Proof. Let .xn ; yn / n1 A be a H –decreasing sequence with xn ! x.
It follows that .yn / is K –decreasing. By (H2), there exists y 2 Y such that
y K yn for every n 2 N.
.x; y/ 2 A and
Because .xn ; yn / n1 is H –decreasing we have that
yn D ynC1 C d.xn ; xnC1 /hn C kn (10.15)
with hn 2 H and kn 2 K for n

1. If xn D xn for n
n
1 we take x WD xn ; then
.x; y/ H .xn ; yn / for every n 2 N. Indeed, for n n we have that .xn ; yn / H
.xn ; yn /; because y K yn , by (10.4) we get .x; y/ H .x; yn / D .xn ; yn /, and
so .x; y/ .xn ; yn /. If n > n, using again (10.4), we have .x; y/ D .xn ; y/ H
.xn ; yn /:
Assume that .xn / is not constant for large n. Fix n
1. From (10.15), for p
0,
we have
X
nCp
X
nCp nCp
X
yn D ynCpC1 C d.xl ; xlC1 /hl C kl D ynCpC1 C d.xl ; xlC1 / hn;p
lDn lDn lDn
X
nCp nCp
X
0
C kl D y C kn;p C d.xl ; xlC1 / hn;p (10.16)
lDn lDn
0
P
for some hn;p 2 H and kn;p 2 K. Assuming that ln d.xl ; xlC1 / D 1, from
nCp
X 1 nCp
X 1
0
d.xl ; xlC1 / .yn y/ D hn;p C d.xl ; xlC1 / kn;p 2 H C K;
lDn lDn
we get the contradiction

P 0 2 cl.H C K/ taking the limit for p ! 1. Hence
1
0 < WD ln d.xl xlC1 / < 1. Set
l WD d.xl ; xlC1 / for l
n.
;
Since H isPcs-complete and convfhl j l
ng . H / is bounded we obtain that
the series ln
l hl is convergent and its sum hn belongs to H . It follows that
P
ln d.xl ; xlC1 /hl D hn , and so
k n WD lim kp0 D yn y hn 2 K
p!1
PnCp1
because K is closed. Since d.xn ; xnCp / lDn d.xl ; xlC1 /, we obtain that
d.xn ; x/ , and so
yn D y C d.xn ; x/hn C k n C . d.xn ; x// hn 2 y C d.xn ; x/H C K:
Hence .x; y/ H .xn ; yn / for every n 2 N. t

u
The most part of vector EVP type results are established for Y a separated locally
convex space. However, there are topological vector spaces Y whose topological
dual reduce to f0g. In such a case it is not possible to find z satisfying the conditions
of Corollary 10.2.5.
topological vector space. Let K Y be a proper closed convex cone and H K be
a nonempty cs-complete bounded set with 0 … cl.H C K/. Suppose that A X Y
satisfies (H2) and that PrY .A/ is quasi bounded. Then for every .x0 ; y0 / 2 A there
N y/
exists .x; N 2 A such that .x;N y/
N H .x0 ; y0 / and .x; y/ 2 A, .x; y/ H .x; N y/
N
imply x D x:N
Proof. First observe that A satisfies condition (H1) by Proposition 10.2.10. More-
over, because PrY .A/ is quasi bounded, there exists a bounded set B Y such that
PrY .A/ B C K:
Note that for .x; y/ 2 A the set PrX .AH .x; y// is bounded, where
AH .x; y/ WD f.x 0 ; y 0 / 2 A j .x 0 ; y 0 / H .x; y/g:
In the contrary case there exists a sequence ..xn ; yn //n1 AH .x; y/ with
d.xn ; x/ ! 1. Hence y D yn C d.xn ; x/hn C kn D bn C d.xn ; x/hn C kn0 with
hn 2 H , bn 2 B, kn ; kn0 2 K. It follows that d.xn ; x/1 .y bn / 2 H C K, whence
the contradiction 0 2 cl.H C K/:
Let us construct a sequence ..xn ; yn //n0 A in the following way: Having
.xn ; yn / 2 A, where n 2 N, because Dn WD PrX .AH .xn ; yn // is bounded, there
exists .xnC1 ; ynC1 / 2 AH .xn ; yn / such that
d.xnC1 ; xn /
1
2
supfd.x; xn / j x 2 Dn g
1
4
diam Dn :
We obtain in this way the sequence ..xn ; yn //n0 A, which is H –decreasing.

Since AH .xnC1 ; ynC1 / AH .xn ; yn /, we have that DnC1 Dn for every n 2 N.
Of course, xn 2 Dn . Let us show that diam Dn ! 0. In the contrary case there
exists ı > 0 such that diam Dn
4ı, and so d.xnC1 ; xn /
ı for every n 2 N.
As in the proof of Proposition 10.2.10, for every p 2 N, we obtain that
X
p X
p X
p
y0 D ypC1 C d.xl ; xlC1 / hp C kl D bp C d.xl ; xlC1 / hp C kp0
lD0 lD0 lD0
D bp C .p C 1/ıhp C kp00 ;
where hp 2 H , bp 2 B, kl ; kp0 ; kp00 2 K. It follows that Œ.p C 1/ı1 .y0 bp / 2

H C K for every p 2 N. Since .bp / is bounded we obtain the contradiction 0 2
cl.H C K/. Thus we have that the sequence .cl Dn / is a decreasing sequence of
nonempty closed subsets ofT the complete metric space .X; d /, whose diameters tend
to 0. By Cantor’s theorem, n2N cl Dn D fxg N for some xN 2 X . Of course, xn ! x. N
Since ..xn ; yn // A is a H –decreasing sequence, from (H1) we get an yN 2 Y
N y/
such that .x; N H .xn ; yn / for every n 2 N; .x; N y/N is the desired element. Indeed,
N H .x0 ; y0 /. Let .x 0 ; y 0 / 2 AH .x;
N y/
.x; N It follows that .x 0 ; y 0 / 2 AH .xn ; yn /,
N y/.
and so x 0 2 Dn cl Dn for every n. Thus x 0 D x. N t
u
If Y is a separated locally convex space, the preceding result follows immediately
from Corollary 10.2.5.
Of course, the set A X Y can be viewed as the graph of a set-valued map
W X Y ; then PrX .A/ D dom and PrY .A/ D Im .
For the nonempty set E Y let us set
BMMin E WD fy 2 E j E \ .y K/ D fygg
(see [26, (1.2)]); note that this set is different of the usual set
Min E WD fy 2 E j E \ .y K/ y C Kg;
but they coincide if K is pointed, that is, K \.K/ D f0g. As in [26, Definition 3.2],
we say that W X Y satisfies the limiting monotonicity condition at x 2
dom if for every sequence ..xn ; yn //n1 graph with xn ! x and .yn / K –
decreasing, there exists y 2 BMMin .x/ such that y yn for every n
1. As
observed in [26], if satisfies the limiting monotonicity condition at x 2 dom
N BMMin .x/ C K, that is, .x/
then .x/ N satisfies the domination property.
In [26, Proposition 3.3], in the case Y a Banach space, there are mentioned
sufficient conditions in order that satisfy the limiting monotonicity condition at
x 2 dom :
When X and Y are Banach spaces and H is a singleton the next result is
practically [26, Theorem 3.5].
topological vector space. Let K Y be a proper closed convex cone and H K
be a nonempty cs-complete bounded set with 0 … cl.H C K/. Suppose that
W X Y is K-l.s.c., satisfies the limiting monotonicity condition on dom and
Im is quasi-bounded. Then for every .x0 ; y0 / 2 graph there exist xN 2 dom
and yN 2 BMMin .x/ N such that .x; N y/
N H .x0 ; y0 / and .x; y/ 2 graph ,
.x; y/ H .x;
N y/
N imply x D x:N
Proof. In order to apply Theorem 10.2.11 for A WD graph we have only to show
that A verifies condition (H2). For this consider the sequence ..xn ; yn //n1 A
such that .yn / is K –decreasing and xn ! x. Clearly, xn 2 lev .y1 / for every
n; since is K-l.s.c., we have that x 2 lev .y1 / dom . Since satisfies the
limiting monotonicity condition at x, we find y 2 BMMin .x/ .x/ such that
y yn for every n. Hence (H2) holds. By Theorem 10.2.11 there exists .x; y/ 2 A
such that .x; y/ H .x0 ; y0 / and .x 0 ; y 0 / 2 graph , .x 0 ; y 0 / H .x; y/ imply x 0 D
x. Set xN WD x and take yN 2 BMMin .x/ N such that yN K y. By (10.4) we have that
.x; N H .x0 ; y0 /. Let now .x 0 ; y 0 / 2 graph D A with .x 0 ; y 0 / H .x;
N y/ N y/.
N Since
N y/
.x; N D .x; y/ N H .x; y/, we have that .x 0 ; y 0 / H .x; y/, and so x 0 D x D x. N
u
In the case when H is a singleton the next result is practically [25, Theorem 1]
under the supplementary hypothesis that Min .x/ is compact for every x 2 X .
topological vector space. Let K Y be a proper closed convex cone and H K
be a nonempty cs-complete bounded set with 0 … cl.H C K/. Suppose that
W X Y is K-l.s.c., Min .x/ is compact and .x/ K C Min .x/ for
every x 2 dom , and Im is quasi-bounded. Then for every .x0 ; y0 / 2 graph
there exist xN 2 dom and yN 2 Min .x/ N such that .x;
N y/
N H .x0 ; y0 / and
.x; y/ 2 graph , .x; y/ H .x;
N y/
N imply x D x:
N
Proof. In order to apply Theorem 10.2.11 for A WD graph we have only to show
that A verifies condition (H2). For this consider the sequence ..xn ; yn //n1 A
such that .yn / is K –decreasing and xn ! x. As in the proof of the preceding
corollary, x 2 lev .yn / for every n 2 N. Because .x/ K C Min .x/, for
10.3 Minimal Points in Product Spaces of Isac–Tammer’s Type 381
every n 2 N there exists yn0 2 Min .x/ such that yn0 yn . Because Min .x/
is compact, .yn0 / has a subnet .y 0 .i / /i 2I converging to some y 2 Min .x/; here
W .I; / ! N is such that for every n there exists in 2 I with .i /
n for
i in . Hence y 0 .i / y .i / yn for i in , whence y yn because K is closed.
Therefore, (H2) holds. By Theorem 10.2.11, for .x0 ; y0 / 2 graph , there exists
.x; y/ 2 A such that .x; y/ H .x0 ; y0 / and .x 0 ; y 0 / 2 graph , .x 0 ; y 0 / H .x; y/
imply x 0 D x. Set xN WD x and take yN 2 Min .x/ N such that yN K y. As in the
proof of Corollary 10.2.12 we find that .x; N y/
N is the desired element. The proof is
complete. t
u
10.3 Minimal Points in Product Spaces

of Isac–Tammer’s Type
Besides F W X X K considered in the preceding section we consider also

F 0 W Y Y K satisfying conditions (F1) and F(2), that is, 0 2 F 0 .y; y/ for
all y 2 Y and F 0 .y1 ; y2 / C F 0 .y2 ; y3 / F 0 .y1 ; y3 / C K for all y1 ; y2 ; y3 2 Y .
Then ˚ W Z Z K with Z WD X Y , defined by ˚..x1 ; y1 /; .x2 ; y2 // WD
F .x1 ; x2 / C F 0 .y1 ; y2 /, satisfies conditions (F1) and (F2), too. As in Sect. 10.2 we
obtain that the relation 4F;F 0 defined by
.x1 ; y1 / 4F;F 0 .x2 ; y2 / ” y2 2 y1 C F .x1 ; x2 / C F 0 .y1 ; y2 / C K
is reflexive and transitive. Moreover, for x; x1 ; x2 2 X and y1 ; y2 2 Y we have
.x1 ; y1 / 4F;F 0 .x2 ; y2 / H) .x1 ; y1 / 4F .x2 ; y2 / H) y1 K y2 ;

.x; y1 / 4F;F 0 .x; y2 / ” y2 2 y1 C F 0 .y1 ; y2 / C K ) y1 K y2 :
As in the preceding section, for F satisfying (F1)–(F3), F 0 satisfying (F1), (F2) and
z from (F3) we define the partial order F;F 0 ;z by

.x1 ; y1 / D .x2 ; y2 / or
.x1 ; y1 / F;F 0 ;z .x2 ; y2 / ”
.x1 ; y1 / F;F 0 .x2 ; y2 / and z .y1 / < z .y2 /:

topological vector space and K Y is a proper convex cone. Let F W X X K
satisfy conditions (F1)–(F3), let F 0 W Y Y K satisfy (F1) and (F2), and let
A X Y satisfy the condition
(H1b) for every 4F;F 0 –decreasing sequence ..xn ; yn // A with xn ! x 2 X
there exists y 2 Y such that .x; y/ 2 A and .x; y/ 4F;F 0 .xn ; yn / for every
n 2 N.
Suppose that z (from (F3)) is bounded from below on PrY .A/. Then for every
.x0 ; y0 / 2 A there exists a minimal element .x;
N y/
N of A with respect to F;F 0 ;z such
N y/
that .x; N F;F 0 ;z .x0 ; y0 /.
Proof. It is easy to verify that F;F 0 ;z is reflexive, transitive and antisymmetric.
To get the conclusion one follows the lines of the proof of Theorem 10.2.3. t
u
Clearly, taking F 0 D 0 in Theorem 10.3.1 we get Theorem 10.2.3. As mentioned
after the proof of Theorem 10.2.3, this extends significantly [214, Theorem 3.10.7],
keeping practically the same proof. One can ask if [214, Theorem 3.10.15] could be
extended to this framework, taking into account that the boundedness condition on
A in [214, Theorem 3.10.15] is much less restrictive. In [214, Theorem 3.10.15] it
is used a functional 'C;k 0 (defined in (10.19) below) in order to prove the minimal
point theorem. Because an element k 0 does not impose itself naturally, and we need
a stronger condition on the functional 'C;k 0 even if k 0 2 K int C , we consider an
abstract K–monotone functional ' to which we impose some conditions 'C;k 0 has
already.
satisfy conditions (F1)–(F3), let F 0 W Y Y K satisfy (F1) and (F2), and
let A X Y satisfy the condition (H1b). Assume that there exists a functional
' W Y ! R such that
(F4) .x1 ; y1 / F;F 0 .x2 ; y2 / H) '.y1 / C d.x1 ; x2 / '.y2 /.
If ' is bounded below on PrY .A/ then for every point .x0 ; y0 / 2 A with
'.y0 / 2 R, there exists .x;
N y/
N 2 A such that .x; N F;F 0 .x0 ; y0 /, and .x 0 ; y 0 / 2 A,
N y/
0 0 0
.x ; y / F;F 0 .x;
N y/N imply x D x: N Moreover, if ' is strictly K–monotone on
PrY .A/, that is, y1 ; y2 2 PrY .A/, y2 y1 2 K n f0g H) '.y1 / < '.y2 /, then
N y/
.x; N is a minimal point of A with respect to F;F 0 :
Proof. Let us construct a sequence ..xn ; yn //n0 A as follows: Having .xn ; yn / 2
A, we take .xnC1 ; ynC1 / 2 A, .xnC1 ; ynC1 / F;F 0 .xn ; yn /, such that
'.ynC1 / inff'.y/ j .x; y/ 2 A; .x; y/ F;F 0 .xn ; yn /g C 1=.n C 1/: (10.17)
Of course, the sequence ..xn ; yn // is F;F 0 –decreasing. From (F4) we obtain that
0 d.xnC1 ; xn / '.yn /'.ynC1 /, and so the sequence .'.yn // is non-increasing.
Because .yn / PrY .A/, .'.yn // is also bounded (from below), and so .'.yn // is
convergent in R. Because .xnCp ; ynCp / F;F 0 .xn ; yn / F;F 0 .xn1 ; yn1 /, using
again (F4) and (10.17) we get
8 n; p 2 N W d.xnCp ; xn / '.yn / '.ynCp / 1=n:

10.3 Minimal Points in Product Spaces of Isac–Tammer’s Type 383
It follows that .xn / is a Cauchy sequence in the complete metric space .X; d /, and
so .xn / is convergent to some xN 2 X .
By (H1b) there exists y 2 Y such that .x; y/ 2 A and .x;N y/
N F;F 0 .xn ; yn / for
every n 2 N, and so
8n
1 W '.y/ '.y/ C d.x; xn / '.yn / '.y/ C 1=n; (10.18)
whence '.yn / ! '.y/. Let us show that .x; y/ is the desired element. Indeed,
.x; N F;F 0 .x0 ; y0 /. Suppose that .x 0 ; y 0 / 2 A is such that .x 0 ; y 0 / F;F 0 .x;
N y/ N y/
N
(F;F 0 .xn ; yn / for every n 2 N). Thus '.y 0 / C d.x 0 ; x/ '.y/ by (F4);
moreover, (10.18) holds with y replaced by y 0 , and so '.y/ D '.y 0 / D lim '.yn /.
Hence d.x 0 ; x/
N D '.y/ N '.y 0 / D 0, whence x 0 D x.
Assuming that ' is strictly K–monotone, because y 0 K y and '.y/'.y N 0
/ D 0,
0
we have necessarily y D y. Hence .x; N y/N is a minimal point with respect to F;F 0 .u t
Note that if C Y is a proper convex cone such that K n f0g int C and
k 0 2 K n f0g, the functional 'C;k 0 W Y ! R defined by
˚
'C;k 0 .y/ WD inf t 2 R j y 2 tk 0 C (10.19)
is a strictly K–monotone continuous sublinear functional (see Corollary 5.2.8).

Moreover, if the condition
(B) PrY .A/ \ .yQ int C / D ; for some yQ 2 Y
holds, then ' WD 'C;k 0 is bounded from below on PrY .A/. Indeed, by
Corollary 5.2.8, we have that '.y/ C '.y/ Q
'.y y/
Q
0 for y 2 PrY .A/,
whence '.y/
'.y/ Q for y 2 PrY .A/:
Another example for a function ' is that defined by
'.y/ WD 'K;k 0 .y y/;

O (10.20)
where K is a proper convex cone, k 0 2 K n f0g, and yO 2 Y is such that

(B1) y0 yO 2 Rk 0 K; PrY .A/ \ .yO K/ D ;.
Then ' is K–monotone, '.y0 / < 1 and '.y/
0 for every y 2 PrY .A/:
For both of these functions in (10.19) and (10.20) we have to impose condition
(F4) in order to be used in Theorem 10.3.2.
Remark 10.3.3. Using the function ' D 'K;k 0 . y/ O (defined by (10.20)) in The-
orem 10.3.2 we can derive [237, Theorem 4.2] taking F .x1 ; x2 / WD fd.x1 ; x2 /k 0 g
and F 0 .y1 ; y2 / WD f" ky1 y2 k k 0 g when Y is a Banach space; note that, at its turn,
[237, Theorem 4.2] extends [285, Theorem 8].
10.4 Ekeland’s Variational Principles of Ha’s Type
The previous EVP type results correspond to Pareto optimality. Ha [227] established
an EVP type result which corresponds to Kuroiwa optimality. Theorem 10.4.4 below
is an extension of this type of result. For its proof we need the following result
established by Hamel and Tammer in [237, Theorem 2.2] which generalizes [575,
Theorem 3.1] (where S.x/ is assumed to be closed for every x 2 X ).
Theorem 10.4.1. Let .X; d / be a metric space, quasi-ordered by 4 such that
.X; d / is 4–complete, that is any 4–increasing Cauchy sequence .xn / X is
convergent. Assume that
(i) S.x/ WD fx 0 2 X j x 4 x 0 g is 4–upper closed for every x 2 X , that is for
every 4–increasing sequence .xn / S.x/ with xn ! u one has u 2 S.x/, and
(ii) any 4–increasing sequence .xn / X is asymptotic, that is d.xnC1 ; xn / ! 0:
Then for every x 2 X there exists x 2 X such that x 4 x and S.x/ D fxg:
Proof. Let x 2 X and set x0 WD x; fix also ˛ 2 0; 1=2Œ. There exists x1 2 S.x0 /
such that d.x1 ; x0 /
minf1; ˛ diam S.x0 /g. Then there exists x2 2 S.x1 / such
that d.x2 ; x1 /
minf1; ˛ diam S.x1 /g: Continuing in this way we get a sequence
.xn /n0 X such that xnC1 2 S.xn / and d.xnC1 ; xn /
minf1; ˛ diam S.xn /g
for every n. By (ii) we have that d.xnC1 ; xn / ! 0, whence diam S.xn / ! 0. For
" > 0 there exists n0 2 N such that diam S.xn / < " for n
n0 . Hence for n; m
n0 we have that xn ; xm 2 S.xn0 /, whence d.xn ; xm / < ". It follows that .xn / is
an increasing Cauchy sequence, and so .xn / is convergent to some x 2 X . Since
.xn /nk S.xk / is a 4–increasing sequence and S.xk / is 4–upper closed, we
obtain that x 2 S.xk / [and so S.x/ S.xk /] for every k
0. It follows that
0 diam S.x/ diam S.xk / ! 0, which shows that diam S.x/ D 0. Since
x 2 S.x/ we get S.x/ D fxg. As x 2 S.x/ D S.x0 /, the conclusion follows. t
u
Note that in Theorem 10.4.1 4 is in fact anti-symmetric. Indeed, take x; x 0 2 X
with x 4 x 0 and x 0 4 x. Then the sequence .xn /n0 defined by x2n WD x and
x2nC1 WD x 0 is 4-increasing; by (ii) we get d.x; x 0 / D d.xn ; xnC1 / ! 0, and so
x D x0 :
Remark 10.4.2. Conditions (i) and (ii) in Theorem 10.4.1 are equivalent to (i’) and
(ii’), respectively, where
(i’) for any 4–increasing sequence .xn / X with xn ! x 2 X one has xn 4 x
for every n
1,
(ii’) any 4–increasing sequence .xn / X is Cauchy.
Indeed, assume that (i) in Theorem 10.4.1 holds and take .xn / X a 4–
increasing sequence with xn ! x 2 X . Since .xn /nk S.xk / and S.xk / is
4–upper closed, we have that x 2 S.xk /, and so xk 4 x for every k. Conversely,
assume that (i’) holds and take .xn / S.x/ a 4–increasing sequence with xn ! u.
Then x 4 x1 4 u, whence u 2 S.x/:
10.4 Ekeland’s Variational Principles of Ha’s Type 385
Clearly, (ii’) ) (ii). Conversely, assume that (ii) holds, but the increasing
sequence .xn /n1 X is not Cauchy. Then there exists "0 > 0 such that for every
p
1 there exists n > p such that d.xn ; xp /
"0 . So, for p D 1 DW n1 , there exists
n2 > n1 such that d.xn2 ; xn1 /
"0 : For p WD n2 , there exists n3 > n2 such that
d.xn3 ; xn2 /
"0 . Continuing in this way we get the (strictly) increasing sequence
.nk /k1 N such that d.xnkC1 ; xnk /
"0 for all k
1. Setting xk0 WD xnk
for k
1 we have that .xk0 /k1 is 4–increasing, and so .xk0 / is asymptotic. This
0
contradicts the fact that d.xkC1 ; xk0 /
"0 for every k
1:
Remark 10.4.2 shows that Theorem 10.4.1 can be reformulated as in the
following result (see [391, Lemma 2.2]).
Theorem 10.4.3. Let .X; d / be a metric space, quasi-ordered by 4, such that any
4–increasing sequence .xn / X converges to some x 2 X with xn 4 x for every
n. Then for every x 2 X there exists x 2 X such that x 4 x and S.x/ D fxg:
Note also that Theorem 10.4.1 is slightly more general than the Dancs–Hegedüs–
Medvegyev Theorem (see [120]), in which .X; d / is assumed to be complete instead
of being 4–complete.
satisfy conditions (F1)–(F3) and W X Y be such that z (from (F3)) is
bounded below on .X /. If fx 2 X j .u/ .x/ C F .x; u/ C Kg is closed
for every u 2 X , then for every x0 2 dom there exists x 2 X such that
.x0 / .x/ C F .x; x0 / C K, and .x/ .x/ C F .x; x/ C K implies x D x:
Proof. Let us consider the relation 4 on X defined by
x 4 x 0 ” .x/ .x 0 / C F .x 0 ; x/ C K: (10.21)
By our hypotheses we have that S.x/ WD fx 0 2 X j x 4 x 0 g is closed for every x 2

X . Note that for x 2 X n dom we have that S.x/ D X , while for x 2 dom we
have that S.x/ dom . The relation 4 is reflexive and transitive. The reflexivity
of 4 is obvious. Let x 4 x 0 and x 0 4 x 00 . Then .x/ .x 0 / C F .x 0 ; x/ C K and
.x 0 / .x 00 / C F .x 00 ; x 0 / C K. Using (F2) we get
.x/ .x 00 / C F .x 00 ; x 0 / C K C F .x 0 ; x/ C K .x 00 / C F .x 00 ; x/ C K;
that is, x 4 x 00 . Consider
' W X ! R; '.x/ WD inf z . .x// ; (10.22)
with the usual convention inf ; WD C1. Clearly, '.x/

m WD inf z . .X // >
1. Moreover, if dom 3 x 4 x 0 then z . .x// z . .x 0 // C z .F .x 0 ; x// C
z .K/, whence '.x/
'.x 0 / C inf z .F .x 0 ; x//
'.x 0 /:
Fix x0 2 dom . The conclusion of the theorem asserts that there exists
x 2 X such that x 2 S.x0 / and S.x/ D fxg. To get this conclusion we apply
Theorem 10.4.1. Because .X; d / is complete and S.x/ is closed for every x 2 X ,
we may (and do) assume that dom D X (otherwise we replace X by S.x0 /).
In order to apply Theorem 10.4.1 let .xn /n1 X be 4–increasing and let us
show that .xn / is a Cauchy sequence. In the contrary case there exist ı > 0 and
.np /p1 N an increasing sequence such that d.xnp ; xnpC1 /
ı for every p
1.

Then, as seen above, '.xnp /
'.xnpC1 / C inf z F .xnpC1 ; xnp / , and so
X
p

'.xn1 /
' xnpC1 C inf z F .xnlC1 ; xnl /
m C p .ı/
lD1
with .ı/ > 0 from (F3). Letting p ! 1 we get a contradiction. Hence .xn / is
Cauchy, and so asymptotic. Applying Theorem 10.4.1 the conclusion follows. u
t
Note that instead of assuming S.u/ to be closed for every u 2 X it is sufficient
to have that S.u/ is 4–upper closed, where 4 is defined by (10.21).
Remark 10.4.5. Taking Y to be a separated locally convex space, K Y a pointed
closed convex cone and F .x; x 0 / WD fd.x; x 0 /k 0 g with k 0 2 K n f0g, we can deduce
[227, Theorem 3.1]. For this assume that .X / is quasi bounded, .x/CK is closed
for every x 2 X and is K-l.s.c. Since clearly z is bounded from below on Im ,
in order to apply the preceding theorem we need to have that S.u/ is closed for every
u 2 X ; this is done in [227, Lemma 3.2]. Below we provide another proof for the
closedness of S.u/:
First, if x … lev .b/ then there exists ı > 0 such that B.x; ı/ \ lev .b C ık 0 / D
;. Indeed, because x … lev .b/ we have that b … .x/ C K, and so b C ı 0 k 0 …
.x/ C K, that is, x … lev .b C ı 0 k 0 /, for some ı 0 > 0 (since .x/ C K is closed).
Because lev .bCı 0 k 0 / is closed, there exists ı 2 0; ı 0 such that B.x; ı/\lev .bC
ı 0 k 0 / D ;, and so B.x; ı/ \ lev .b C ık 0 / D ;:
Fix u 2 X and take x 2 X n S.u/, that is, .u/ 6 .x/ C d.x; u/k 0 C K. Then
there exists y 2 .u/ with b WD y d.x; u/k 0 … .x/ C K. By the argument above
there exists ı 0 > 0 such that B.x; ı 0 / \ lev .b C ı 0 k 0 / D ;, that is, y d.x; u/k 0 C
ı 0 k 0 … .x 0 / C K for every x 0 2 B.x; ı 0 /. Taking ı 2 0; ı 0 sufficiently small we
have that d.x 0 ; u/
d.x; u/ı 0 for x 0 2 B.x; ı/, and so y … .x 0 /Cd.x 0 ; u/k 0 CK
for every x 0 2 B.x; ı/, that is, B.x; ı/ \ S.u/ D ;:
If we assume that .x0 / 6 .x/ C k 0 C K for every x 2 X , then x provided by
the preceding theorem satisfies d.x; x0 / < 1. Indeed, in the contrary case, because
.x0 / .x/ C d.x; x0 /k 0 C K and d.x; x0 /k 0 C K k 0 C K, we get the
contradiction .x0 / .x/ C k 0 C K. Replacing k 0 by "k 0 and d by
1 d for
some ";
> 0 we obtain exactly the statement of [227, Theorem 3.1].
With a similar proof we get the next result; for Y a separated locally convex
space, K Y a closed convex cone, k 0 2 K n f0g; F .x; x 0 / WD d.x; x 0 /k 0 for
x; x 0 2 X , and .X / quasi-bounded, one gets [390, Corollary 3.1].

satisfy conditions (F1) and (F3), and W X Y be such that z (from (F3)) is
bounded below on .X /. Assume that (a) fx 2 X j .u/ .x/ C F .x; u/g is
closed for every u 2 X , and (b) .y/ .x/ C F .x; y/, .z/ .y/ C F .y; z/
imply .z/ .x/ C F .x; z/ for all x; y; z 2 X . Then for every x0 2 dom there
exists x 2 X such that .x0 / .x/CF .x; x0 /, and .x/ .x/CF .x; x/CK
implies x D x:
Proof. The proof is practically the same as that of Theorem 10.4.4. This time the
relation 4 on X is defined by x 4 x 0 if .x/ .x 0 / C F .x 0 ; x/. By (a) we
have that S.x/ WD fx 0 2 X j x 4 x 0 g is closed for every x 2 X . Note that
for x 2 X n dom we have that S.x/ D X , while for x 2 dom we have
that S.x/ dom . The reflexivity of 4 is obvious, while its transitivity is given
by (b). Consider ' given by (10.22). The rest of the proof is the same as that of
Theorem 10.4.4. t
u
In the case in which Y is just a topological vector space we have the following
version of Theorem 10.4.4.
topological vector space and K Y is a proper closed convex cone. Let H K
be nonempty convex set with 0 … cl.H C K/, and W X Y . If fx 2 X j .u/
.x/ C d.x; u/H C Kg is closed for every u 2 X and .X / is quasi bounded, then
for every x0 2 dom there exists x 2 X such that .x0 / .x/Cd.x; x0 /H CK
and .x/ .x/ C d.x; x/H C K implies x D x:
Proof. Let B Y be a bounded set such that .X / B C K:
Consider F WD FH with FH defined in (10.11). As seen before, F satisfies con-
ditions (F1) and (F2), and so the relation 4 defined in the proof of Theorem 10.4.4
is a quasi-order; moreover, by our hypotheses, S.x/ WD fx 0 2 X j x 4 x 0 g is
closed for every x 2 X . As in the proof of Theorem 10.4.4 we may (and do) assume
that X D dom and it is sufficient to show that every 4–increasing sequence
.xn /n1 X is Cauchy. In the contrary case there exist ı > 0 and .np /p1 N
an increasing sequence such that d.xnp ; xnpC1 /
ı for every p
1:
Fixing y1 2 .x1 /, inductively we find the sequences .yn /n0 Y , .hn /n0
H and .kn /n0 K such that yn D ynC1 C d.xn ; xnC1 /hn C kn for every n
1.
Using the convexity of H , and the facts that H K and .X / B C K, for
p 2 N we get h0p 2 H , bp 2 B and kp0 ; kp00 2 K such that
X
np
X
p
y1 D ynpC1 C d.xl ; xlC1 /hl C kl D bp C ı.hn1 C : : : C hnp / C kp0
lD1 lD0
D bp C pıh0p C kp00 :
It follows that .pı/1 .y1 bp / 2 H C K for every p

1. Since .bp / is bounded
we obtain the contradiction 0 2 cl.H C K/. The conclusion follows. t
u
Remark 10.4.8. In Theorem 10.4.7, instead of assuming that S.u/ is closed for
every u 2 X , it is sufficient to assume that S.u/ is 4–upper closed for u 2 X ,
where 4 is defined by (10.21) for F .x; x 0 / D d.x; x 0 /H .
Theorem 10.4.7 covers [227, Theorem 3.1], too.
topological vector space, K Y is a proper closed convex cone, and H
K is a nonempty convex set. Let W X Y be such that epi WD
f.x; y/ 2 X Y j y 2 .x/ C Kg is closed and .X / is quasi bounded. Assume
that one of the following conditions holds:
(i) H is bounded, cs-complete, and 0 … cl.H C K/;
(ii) Y is a n.v.s., and H is cs-complete with 0 … cl H and satisfying
9y 2 K C n f0g; 8h 2 H W hh; y i
khk I (10.23)
(iii) Y is a reflexive Banach space, and H is closed with 0 … H and satisfying
9 > 0; 8h 2 H; 8k 2 K W kh C kk
khk : (10.24)
Then for every x0 2 dom there exists x 2 X such that .x0 / .x/ C
d.x; x0 /H C K and .x/ .x/ C d.x; x/H C K implies x D x:
Proof. We apply Theorem 10.4.7, taking into account Remark 10.4.8. For this first
observe that 0 … cl.H C K/ in each case. Indeed, (10.23) implies (10.24) with
1
WD ky k > 0, and (10.24) together 0 … cl H imply kyk
d.0; H / > 0 for
every y 2 H C K:
We have to show that S.u/ is 4-upper closed for every u 2 X . Let B Y be a
bounded set such that .X / B C K:
Consider .xn /n1 S.u/ a 4–increasing sequence with xn ! x 2 X and
y 2 .u/. We have to show that y 2 .x/ C d.x; u/H C K. Setting x0 WD u and
y0 WD y we have that
yn D ynC1 C d.xn ; xnC1 /hn C kn

K ynC1 8n
0; (10.25)
where yn 2 .xn /, hn 2 H and kn 2 K for all n

0: It follows that
X
n X
n
y D ynC1 C d.xi ; xi C1 /hi C kn0 D bn C d.xi ; xi C1 /hi C kn00
i D0 i D0
D ynC1 C n h0n C kn0 D bn C n h0n C kn00 (10.26)

P P
for every n
0, where kn0 WD niD0 kn 2 K, kn00 2 K, n WD niD0 d.xi ; xi C1 /,
h0n 2 H and bn 2 B; we have taken into account the convexity of H and that
.X / B C K:
P
Set WD n0 d.xn ; xnC1 / D lim n 2 Œ0; 1: Clearly, d.xnC1 ; u/ n
nC1 for every n
0. If D 0 then x D u, and so y 2 .x/Cd.x; u/H CK.
Assume that D 1. From the fourth expression of y in (10.26) we have that
1
n .y bn / 2 H C K for n sufficiently large. Since .bn / is bounded we obtain the
contradiction 0 2 cl.H C K/:
Let 2 0; 1Œ. For n
0 let us set
X
n
yn0 WD n h0n D d.xi ; xi C1 /hi :
i D0
Assume first that (i) holds. Because H is bounded and convex we obtain
that the sequence .yn0 /n1 is Cauchy. Assume now that (ii) holds. From
the secondP expression of y in (10.26) and (10.23) we have that hy; y i

hbn ; y i C niD0 d.xi ; xi C1 / khi k. Because B is bounded we obtain that the series
P 0
n0 d.xn ; xnC1 /hn is absolutely convergent, and so the sequence .yn /n1 is
0
Cauchy. In both cases, because H is cs-complete, we have that yn ! h for some
h 2 H . From the first expression of y in (10.26) we have that yyn0 2 .xnC1 /CK,
and so .xnC1 ; y yn0 / 2 epi . Hence .xnC1 ; y yn0 / ! .x; y h/ 2 cl.epi / D
epi . It follows that y 2 .x/ C H C K .x/ C d.x; u/H C K (because
d.x; u/ ), and so x 2 S.u/:
Assume now that (iii) holds. Using again the fourth expression of y in (10.26),
from (10.24) we obtain that ky bn k
n kh0n k ; and so .h0n / is bounded. Because
Y is reflexive, .h0n /n1 has a subsequence converging weakly to h 2 H , and so h 2
Hp WD convfh0n j n
pg for every p
1. Fix some n1
1 such that d.xn ; x/ < 1
for every n
n1 . Because h 2Hn1 , there exists n2 > n1 and .
1n /n1 n<n2 RC
Pn2 1 1 Pn2 1 1 0
such that nDn1
n D 1 and h1 h < 1, where h1 WD nDn1
n hn 2 H .
Increasing if necessary n2 , we may (and do) assume that d.xn ; x/ < 1=2 for every
n
n2 . Continuing in this way, we find an increasing sequence .nk /k1 N
such that for k
1 we have d.x k n
nk and there exists
; x/ < 1=k for every
PnnkC1
.
kn /nk n<nkC1 RC such that nDn 1 k

D 1 and h h < 1=k, where hk WD
PnkC1 1 k 0 Pk n
nDnk
n hn 2 H . Because n D n0 d.xn ; xnC1 /
d.x; u/ d.xnC1 ; x/, we
have that n
d.x; u/ 1=k for every n
nk . Taking into account the third
expression of y in (10.26), then (10.24), it follows that
y
K ynC1 C n h0n
ynC1 C .d.x; u/ 1=k/h0n
K ynkC1 C .d.x; u/ 1=k/h0n
for every n 2 nk ; nkC1 1, whence y

ynkC1 C .d.x; u/ 1=k/hk . Hence
.xnkC1 ; y .d.x; u/ 1=k/hk / 2 epi for every k
1. Since hk ! h for
k ! 1, from the closedness of epi , we get .x; y d.x; u/h/ 2 epi , and
so y 2 .x/ C d.x; u/H C K. The proof is complete. t
u
Remark 10.4.10. Theorem 10.4.9 (iii) is exactly [391, Theorem 3.5 (iii)], Theo-
rem 10.4.9 (ii) is slightly more general than [391, Theorem 3.5 (i)] (in which,
moreover, Y is a Banach space, and H is closed and convex), and Theorem 10.4.9
(i) is much more general than [391, Theorem 3.5 (ii)] (in which, moreover, Y is a
Banach space, H is closed and convex, and (10.24) holds).
10.5 Ekeland’s Variational Principle for Bi-Set-Valued Maps
In [48] Bianchi, Kassay and Pini obtained an EVP type result for vector functions
of two variables; previously such results were obtained by Isac [282] and Li et al.
[384]. The next result extends [48, Theorem 1] in two directions: d.; /e is replaced
by F satisfying (F1)–(F3) and instead of a single-valued function f W X X ! Y
we take a multi-valued one.
satisfy conditions (F1)–(F3). Assume that G W X X Y has the properties:
(i) 0 2 G.x; x/ for every x 2 X;
(ii) G.x1 ; x2 / C G.x2 ; x3 / G.x1 ; x3 / C K for all x1 ; x2 ; x3 2 X;
(iii) z (from (F3)) is bounded below on the set Im G.x; / for every x 2 X:
If
(iv) fx 0 2 X j ŒG.x; x 0 / C F .x; x 0 / \ .K/ ¤ ;g is closed for every x 2 X;
then for every x0 2 X there exists x 2 X such that ŒG.x0 ; x/ C F .x0 ; x/\.K/ ¤
; and ŒG.x; x/ C F .x; x/ \ .K/ ¤ ; implies x D x:
Proof. Let us consider the relation 4 on X defined by

x 4 x 0 ” G.x; x 0 / C F .x; x 0 / \ .K/ ¤ ;:
Then 4 is reflexive and transitive. The reflexivity is immediate from (i) and (F1).
Assume that x 4 x 0 and x 0 4 x 00 . Then k 2 G.x; x 0 / C F .x; x 0 / and k 0 2
G.x 0 ; x 00 / C F .x 0 ; x 00 / with k; k 0 2 K. Hence, by (ii) and (F2),
k k 0 2 G.x; x 0 /CF .x; x 0 /CG.x 0 ; x 00 /CF .x 0 ; x 00 / G.x; x 00 /CF .x; x 00 /CK;
whence ŒG.x; x 00 / C F .x; x 00 / \ .K/ ¤ ;, that is, x 4 x 00 :

Setting S.x/ WD fx 0 2 X j x 4 x 0 g, by (iv) we have that S.x/ is closed for
every x 2 X . We have to show that for .xn /n1 X a 4–increasing sequence
one has d.xn ; xnC1 / ! 0. In the contrary case there exist an increasing sequence
.nl /l1 N and ı > 0 such that d.xnl ; xnl C1 /
ı for every l
1. Because .xn /
is 4–increasing, we have that kn 2 G.xn ; xnC1 / C F .xn ; xnC1 / for some kn 2 K
and every n
1. Then
k1 : : : kn 2 G.x1 ; xnC1 / C F .x1 ; x2 / C : : : C F .xn ; xnC1 / C K;

10.6 EVP Type Results 391
and so
8n
1 W inf z .Im G.x1 ; // C inf z .F .x1 ; x2 // C : : : C inf z .F .xn ; xnC1 // 0:

Since inf z .F .xn ; xnC1 //
0 for every n
1 and inf z F .xnl ; xnl C1 /
.ı/ > 0 for every l

1, taking n WD np with p
1, we obtain that p.ı/
inf z .Im G.x1 ; // for every p
1. This yields the contradiction .ı/ 0.
Hence d.xn ; xnC1 / ! 0. Applying Theorem 10.4.1 we get some x 2 S.x0 / with
S.x/ D fxg, that is, our conclusion holds. t
u
Remark 10.5.2. If we need the conclusion only for a fixed (given) point x0 2 X ,
we may replace condition (iii) by the fact that z (from (F3)) is bounded below on
the set Im G.x0 ; /:
Indeed, X0 WD S.x0 / is closed by (iv), and so .X0 ; d / is complete. If x 2 X0 then
k 2 G.x0 ; x/CF .x0 ; x/ G.x0 ; x/CK for some k 2 K, and so k 0 2 G.x0 ; x/
for some k 0 2 K: It follows that k 0 C G.x; u/ G.x0 ; x/ C G.x; u/ G.x0 ; u/ C
K, whence G.x; u/ G.x0 ; u/ C K for every u 2 X . Hence condition (iii) is
verified on X0 , and so the conclusion of the theorem holds for x0 :
Remark 10.5.3. For F .x; x 0 / WD fd.x; x 0 /k 0 g with k 0 2 K n f0g and G
single-valued, using Theorem 10.5.1 and the preceding remark one obtains [282,
Theorem 8] and [384, Theorem 3]; in [282] K is normal and closed, while in [384]
k 0 2 int K.
Note that condition (iv) in the preceding theorem holds when G is compact-
valued, G.u; / is K-l.s.c., K is closed and F .x; x 0 / WD fd.x; x 0 /k 0 g for some k 0 2
K. Indeed, assume that kn 2 G.u; xn / C d.xn ; u/k 0 for every n
1, where kn 2
K. Take " > 0. Then there exists n"
1 such that d.xn ;u/
d.x; u/ " DW " for
every n
n" . Then for such n we have that G.u; xn / \ " k 0 K ¤ ;, whence
G.u; x/ \ " k 0 K ¤ ;. Hence there exists y" 2 G.u; x/ such that y" C " k 0 2
K. Since G.u; x/ is compact, .y" /">0 has a subnet converging to y 2 G.x; u/.
Since lim"!0 " D d.x; u/ and K is closed, we obtain that y C d.x; u/k 0 2 K:
If Y is a separated locally convex space then we may assume that G is weakly
compact-valued instead of being compact-valued.
When G is single-valued and F .x; x 0 / WD fd.x; x 0 /k 0 g with k 0 2 K, where K
is closed and z .k 0 / D 1, the preceding theorem reduces to [48, Theorem 1].
10.6 EVP Type Results
The framework is the same as in the previous sections. We want to apply the
preceding results to obtain vectorial EVPs. To envisage functions defined on subsets
of X we add to Y the distinct elements 1, 1 not belonging to the space Y ,
obtaining thus the space Y WD Y [ f1; 1g. We consider that 1 K y K 1
for all y 2 Y . Consider now the function f W X ! Y . As usual, the domain of

f is dom f D fx 2 X j f .x/ ¤ 1g; the epigraph of f is epi f D f.x; y/ 2
X Y j f .x/ K yg; f is proper if dom f ¤ ; and f .x/ ¤ 1 for every x 2 X ;
the graph of the proper function f is graph f D f.x; f .x// j x 2 dom f g. For
y 2 K C we set .y ı f /.x/ WD C1 for x 2 X n dom f:
satisfy the conditions (F1), (F2) and (10.9), and let f W X ! Y be a proper
function such that the next condition holds:
(H3) for every sequence .xn / dom f with xn ! x 2 X and f .xn / 2 f .xnC1 / C
F .xnC1 ; xn / C K for every n 2 N one has f .xn / 2 f .x/ C F .x; xn / C K for
every n 2 N:
Then for every x0 2 dom f and every z 2 K C such that z ı f is bounded
from below on the set fx 2 dom f j f .x0 / 2 f .x/ C F .x; x0 / C Kg, there exists
x 2 dom f such that
f .x0 / 2 f .x/ C F .x; x0 / C K
and
8x 2 dom f W f .x/ 2 f .x/ C F .x; x/ C K ) x D x:
Proof. Consider A WD graph f WD f.x; f .x// j x 2 dom f g. Condition (H3)

says nothing else than that (H1) is verified. Applying Theorem 10.2.1 (taking into
account Remark 10.2.2) we get the conclusion. t
u
Note that the above result is equivalent to [484, Theorem 3.6], which, at its turn,
extends the main result of [220]. Moreover, the result above can be used when F
satisfies also (F3) and z ı f is bounded from below, where z 2 K C is the element
appearing in (F3).
In the case in which F .x; x 0 / D d.x; x 0 /H for some H K the condition (H3)
becomes
(H4) for every sequence .xn / dom f with xn ! x 2 X and f .xn / 2 f .xnC1 /C
d.xnC1 ; xn /H C K for every n 2 N one has f .xn / 2 f .x/ C d.x; xn /H C K
for every n 2 N:
In the case H WD fk 0 g condition (H4) is nothing else than condition (E1) in
[237].
Using Theorem 10.2.11 we have the following variant of the preceding result.
topological vector space and K Y is a proper closed convex cone. Let f W X !
Y be a proper function and H K be a nonempty cs-complete bounded set with
0 … cl.H C K/. Assume that f .dom f / is quasi bounded. If
10.6 EVP Type Results 393
(H5) for every sequence .xn / dom f such that xn ! x 2 X and .f .xn // is
K –decreasing one has f .x/ K f .xn / for every n 2 N
holds, then for every x0 2 dom f there exists x 2 dom f such that
.f .x0 / K/ \ .f .x/ C d.x; x0 /H / ¤ ;
and
8x 2 dom f n fxg W .f .x/ K/ \ .f .x/ C d.x; x/H / D ;:
Proof. Let us apply Theorem 10.2.11 for A WD graph f . To have the conclusion of
the theorem it is sufficient to observe that (H2) is verified. Indeed, let ..xn ; yn //
graph f be such that xn ! x 2 X and let .yn / be K –decreasing. Hence yn D
f .xn / for every n. By (H5) we have that y WD f .x/ K f .xn / D yn for every
n 2 N and, of course, .x; y/ 2 graph f . The proof is complete. t
u
Remark 10.6.3. According to L. Gajek and D. Zagrodny in [194], condition (H5) is
saying that f is monotonically semicontinuous with respect to K at any x 2 X .
Remark 10.6.4. Taking H to be complete, convex and bounded, then H is cs-
complete. In this case, Theorem 10.6.2 reduces to the main result in [40], that is,
[40, Theorem 4.1].
Remark 10.6.5. Taking H WD fk 0 g in the preceding theorem one obtains [214,
Corollary 3.10.6]; there K is assumed to be closed in the direction k 0 , the present
condition (H5) being condition (H4) in [214, Corollary 3.10.6].
Remark 10.6.6. Similar results can ˚ be stated using
Theorems 10.3.1 ˚ and 10.3.2.
When specializing to F .x1 ; x2 / D d.x1 ; x2 /k 0 and F 0 .y1 ; y2 / D " ky1 y2 k k 0
one recovers [237, Corollary 3.1] and [237, Theorem 4.2].
In the conclusions of [483, Lemma 2.2, Theorem 2.1] ı d appears instead of
d , as in the usual EVP type results, where is the class of those subadditive and
non decreasing functions W RC ! RC with .t/ D 0 only for t D 0. Based
on the next simple result we can show that Lemma 2.2 and Theorem 2.1 of [483]
can be deduced easily from the classical (scalar) EVP and [214, Corollary 3.10.14],
respectively.
Let us denote by 0 the set f 2 j is continuous at 0g; in [288, Theorem 6]
one observed that is continuous if 2 0 . Clearly, for 2 (resp. 2 0 )
and .tn / RC one has: .tn / ! 0 ) tn ! 0 (resp. .tn / ! 0 , tn ! 0).
The case 2 0 of the next result is practically stated in the proof of [288,
Theorem 6].
Proposition 10.6.7. Let 2 and let .M; d / be a metric space. Then d 0 WD ıd
is a metric on M . Moreover, if 2 0 then d and d 0 determine the same topology
and the same Cauchy sequences; in particular .M; d / is complete if and only if
.M; d 0 / is complete. If 2 n 0 then the topology determined by d 0 is the
discrete topology, and so, if .xn / M is asymptotic, that is d 0 .xn ; xnC1 / ! 0, then
.xn / is constant for large n:
Proof. The fact that d 0 is a metric is well known. Let 2 0 , .xn /n1 M and
x 2 M . As seen above, for .tn / RC one has .tn / ! 0 , tn ! 0, we get:
lim d 0 .xn ; x/ D 0 , lim d.xn ; x/ D 0;

n!1 n!1
0
lim d .xn ; xm / D 0 , lim d.xn ; xm / D 0:
n;m!1 n;m!1
The first equivalence implies that the closed sets with respect to d and d 0 are the
same, while the second equivalence shows that the Cauchy sequences with respect
to d and d 0 are the same.
Let 2 n 0 . Then necessarily ı WD inff .t/ j t > 0g > 0. It follows that
Bd 0 .x; r/ D fxg for every r 2 0; ıŒ, and so fxg is open for every x 2 X . Hence
the topology determined by d 0 is the discrete topology. Let .xn / M be such that
d 0 .xn ; xnC1 / ! 0, that is .d.xn ; xnC1 // ! 0. Then there exists n0
1 such that
.d.xn ; xnC1 // < ı, and so d.xn ; xnC1 / D 0, for n
n0 . t
u
Using the preceding proposition, if 2 , then f from [483, Lemma 2.2] is
ı d -lsc; the conclusion of [483, Lemma 2.2] follows from [614, Theorem 1.4.1]
just taking " D 1 and replacing d by ı d . A similar argument shows that [483,
Theorem 2.1] follows from [214, Corollary 3.10.14] by replacing d by ı d:
10.7 Error Bounds
Let us consider W X Y with graph ¤ ; ., dom ¤ ;/. We have the

following result of Hamel type.
vector space, K Y a proper closed convex cone, H K a nonempty convex
set with 0 … cl.H C K/, and W X Y a set-valued map with dom ¤ ;.
Assume that fx 2 X j .u/ .x/ C d.x; u/H C Kg is closed for every u 2 X ,
and .X / is quasi bounded. If S X is such that
8x 2 dom n S; 9u 2 X n fxg W .x/ .u/ C d.x; u/H C K; (10.27)
then S \ dom ¤ ; and
8x 2 X W .x/ .S / C d.x; S \ dom /H C K: (10.28)
Proof. Let x 2 dom . By Theorem 10.4.7, there exists x1 2 X such that
.x/ .x1 / C d.x1 ; x/H C K (10.29)

10.7 Error Bounds 395
and
.x1 / .x/ C d.x; x1 /H C K ) x D x1 I (10.30)
clearly, x1 2 dom . Assume that x1 … S . Using (10.27), there exists u 2 X n fx1 g

such that .x1 / .u/ C d.x1 ; u/H C K. From (10.30) we get the contradiction
u D x1 . In particular, because dom ¤ ;, we have that S 0 WD S \ dom ¤ ;,
too. From (10.29) we obtain that
.x/ .x1 / C d.x1 ; x/H C K .S / C d.x; S 0 /H C K; (10.31)
because d.x1 ; x/H D d.x; S 0 /H C Œd.x1 ; x/ d.x; S 0 / H d.x; S 0 /H C K.

Since (10.31) is obvious for x … dom , the conclusion holds. u
t
A slightly more general result is the following.
vector space, K Y a proper closed convex cone, H K a nonempty convex
set with 0 … cl.H C K/, and W X Y a set-valued map with dom ¤ ;.
Assume that fx 2 X j .u/ .x/ C d.x; u/H C Kg is closed for every u 2 X ,
and .X / is quasi bounded. Furthermore, assume that S; W dom with S ¤ ;
are such that for every x 2 dom n .S [ W / there exists u 2 X n fxg with
.x/ .u/ C d.x; u/H C K, and
8x 2 W W .x/ .S / C d.x; S /H C K (10.32)
for some > 0. Then, for WD min.; 1/;
8x 2 X W .x/ .S / C d.x; S /H C K: (10.33)
Proof. As in the proof of Theorem 10.7.1, consider x 2 dom . As there, there

exists x1 2 dom such that (10.29) and (10.30) hold. If x1 2 W then, by (10.32),
.x/ .S / C d.x; S /H C K, and so .x/ .S / C d.x; S /H C K
because 0 < . If x1 2 S then (10.31) holds, and so again .x/ .S / C
d.x; S /H C K because 1. In the contrary case, x1 2 X n .S [ W /, and so, by
our hypothesis, there exists u 2 X nfx1 g such that .x1 / .u/Cd.x1 ; u/H CK.
From (10.30) we get the contradiction u D x1 . t
u
Remark 10.7.3. The hypothesis “fx 2 X j .u/ .x/ C d.x; u/H C Kg is
closed for every u 2 X ” in Theorems 10.7.1 and 10.7.2 can be replaced by “epi is
closed and one of the conditions (i), (ii), (iii) of Theorem 10.4.9 holds” for getting
the same conclusions of Theorem 10.7.1 and Theorem 10.7.2.
A natural candidate for S in Theorems 10.7.1 and 10.7.2 is the set of solutions
(in a certain sense) of the set minimization problem
minK .x/; s.t. x 2 X: (P )

For example, x is a solution of (P ) in the sense of Kuroiwa if .x/ .x/ C K

implies .x/ .x/CK; the choice of solutions in the sense of Kuroiwa is natural
in the case of Theorem 10.7.1 because the conclusion of Theorem 10.4.7 is saying
that x is a strict solution (or strict minimizer, [227]) of (P ) in the sense of Kuroiwa
for replaced by 0 , where 0 .x/ WD .x/ C d.x; x/H .
A particular case of Theorem 10.7.1 is the following corollary which subsumes
results of Takahashi [555] and Hamel [233].
Corollary 10.7.4. Let .X; d / be a complete metric space and f W X ! R be a
proper lsc lower bounded function. If for a fixed 2 0; 1Œ and every x 2 X with
f .x/ > inf f there exists u 2 X n fxg such that f .u/ C d.x; u/ f .x/, then
S WD arg min f is nonempty and
8x 2 X W d.x; S / f .x/ inf f: (10.34)
Proof. Consider Y WD R, K WD RC , H WD f g K, and W X R defined

by .x/ WD ff .x/g for x 2 dom f and .x/ WD ; elsewhere. Clearly, and S
satisfy the conditions of Theorem 10.7.1. Hence S ¤ ; and .S / D finf f g, and
so the conclusion (10.28) of Theorem 10.7.1 becomes (10.34). t
u
proper lsc function. If for a fixed 2 0; 1Œ and every x 2 X there exists u 2 X nfxg
such that f .u/ C d.x; u/ f .x/, then inf f D 1:
Proof. Assume that inf f > 1 and take S D ;. Applying Corollary 10.7.4 we
obtain the contradiction S ¤ ;. t
u
Applying Corollary 10.7.4 for fC WD maxff; 0g we observe that [391,
Lemma 2.3] is valid without assuming that Œf 0 WD fx 2 X j f .x/ 0g
is nonempty.
A similar result to Corollary 10.7.4, corresponding to Theorem 10.7.2, is the
following.
proper lsc lower bounded function. Assume that S WD arg min f ¤ ;, W dom f ,
and ; 2 0; 1Œ are such that for every x 2 dom f n .S [ W / there exists u 2
X n fxg such that f .u/ C d.x; u/ f .x/, and d.x; S / f .x/ inf f for every
x 2 W . Then, for WD min. ; /;
8x 2 X W d.x; S / f .x/ inf f:
Let fi W X ! R be proper for every i 2 1; n and assume that S WD \niD1 Œfi

0 ¤ ; (that is, the solution set of the system fi .x/ 0; i 2 1; n, is nonempty). As
in [391], for x 2 X set I> .x/ WD fi 2 1; n j fi .x/ > 0g. For > 0, Liu and Ng
in [391] say that x 2 X n S has the -descent property if there exists xO 2 X n fxg
such that: (a)PI> .x/O I> .x/, (b) fi .x/ O fi .x/ for every i 2 I> .x/, and (c)
O
d.x; x/ i 2I> .x/ Œ.f /
i C .x/ .f O In fact, taking into account (a), for
i C .x/.
/
10.7 Error Bounds 397
i 2 1; n n I> .x/ we have that i 2 1; n n I> .x/,

O and so .fi /C .x/ D .fi /C .x/O D 0.
This shows that if x 2 X nS has the -descent property
Pn then there exists O
x 2 X nfxg
O C d.x; x/
such that f .x/ O f .x/, where f WD .f /
i D1 i C . It is obvious that
S D Œf 0 D arg min f . Using the preceding corollary we get the next result.
Corollary 10.7.7. Let .X; d / be a complete metric space and fi W X ! R, i 2 1; n,
be proper lsc functions with S WD \niD1 Œfi 0 nonempty. Assume that there exist
W D WD \niD1 dom fi and ; 2 0; 1Œ satisfying the following conditions:
(i) for each x 2 D nP.S [ W / there exists xO 2 X n fxg such that I> .x/
O I> .x/
O P i 2I> .x/ Œ.fi /C .x/ .fi /C .x/,
and d.x; x/ O and
(ii) d.x; S / i 2I> .x/ .fi /C .x/ for every x 2 W:
P
Then, d.x; S / niD1 .fi /C .x/ for every x 2 X , where WD min. ; /:
Pn
Proof. Consider f WD i D1 .fi /C . Clearly, D D dom f; inf f D 0, and S D
arg min f ¤ ;. Let x 2 dom f n .S [ W /. By (i), there exists xO 2 X n fxg
such that conditions (a) and (c) (above) are verified. The argument provided before
the statement of the present corollary shows that f .x/ O C d.x; x/
O f .x/. Since
d.x; S / f .x/ .D f .x/ inf f / for every x 2 W , the conclusion follows using
Corollary 10.7.6. t
u
Remark 10.7.8. Corollary 10.7.7 provides sufficient conditions of error bounds for
the system fi .x/ 0; i 2 1; n. This result is essentially [391, Theorem 4.1]; there
X is a Banach space and instead of (i) one asks that any x 2 X n .S [ W / has the -
descend property. For the proof of [391, Theorem 4.1] one uses [391, Theorem 3.5]
(see also Remark 10.4.10).
Chapter 11
Derivatives and Epiderivatives of Set-Valued
Maps
In set-valued optimization, derivatives, epiderivatives, and coderivatives of

set-valued maps play the most fundamental role. We give optimality conditions by
using derivatives, epiderivatives, and coderivatives. Sensitivity analysis is another
important area where these objects are the building blocks. Numerical solutions of
set-valued optimization problems are also computed by expressing optimality in
terms of derivatives, epiderivatives, and coderivatives.
This chapter presents a detailed study of derivatives and epiderivatives of
set-valued maps. In the literature, several approaches have been proposed to
define derivatives of set-valued maps. To show the wide spectrum of ideas
available to define the derivatives of set-valued maps, we refer the reader to
the interesting articles by Banks and Jacobs [23], Daniilidis and Pang [124],
Dentcheva [132], De Blasi [128], Gorokhovik and Zabreiko [216], Minchenko
and Volosevich [421], Nurminskiı̆ [444], Ovcharova and Gwinner [447, 448],
Plotnikov [477], Rolewicz [507], Sach [514], Schinas and Boudourides [523], Xia,
Wang, and Zhang [600], and the cited references therein. See also Chen, Nashed,
and Qi [94] for an elegant treatment of nonsmoothness.
Our focus in this chapter is on the derivatives of set-valued maps which are
motivated by the geometric interpretation of the classical notion of derivative for
single-valued maps as a local approximation of its graph. We study the contingent
derivative, the contingent epiderivative, and several analogues of these notions such
as the adjacent derivative, the Clarke derivative, the lower Dini derivative, and the
radial derivative. We also investigate their second-order analogues. In particular, we
study the second-order contingent derivative, the second-order adjacent derivative,
the second-order lower Dini derivative, and their epiderivatives counterparts. We
give several calculus rules for the first-order derivatives and epiderivatives and their
second-order extensions. The existence of contingent epiderivatives is an important
issue in set-valued optimization. We present a comprehensive study of the various
issues pertaining to the existence of contingent epiderivatives, their generalized
analogues, and their various characterizations. We also study important issues

DOI 10.1007/978-3-642-54265-7__11
400 11 Derivatives and Epiderivatives of Set-Valued Maps
such as proto-differentiability, semi-differentiability, and several related notions.

Numerous other concepts which play an important role in differentiability of
set-valued maps are given.
There are excellent monographs that address the notion of derivatives of
set-valued maps and related issues such as by Aubin and Ekeland [16], Aubin and
Frankowska [18], Borwein and Zou [66], Clarke [100], Hu and Papageorgiou [267],
Jahn [293], Mordukhovich [430, 431], Penot [470], Rockafellar and Wets [499],
and Schirotzek [524], among others. Our treatment of the subject has benefited
particularly from [16, 18, 267, 499].
11.1 Contingent Derivatives of Set-Valued Maps
We begin by introducing the notion of the contingent derivative which plays a

fundamental role in various aspects of set-valued optimization.
Definition 11.1.1. Let X and Y be normed spaces, let F W X Y be a set-valued
N y/
map, and let .x; N 2 graph .F /: A set-valued map DF.x;
N y/
N W X Y is called the
N y/;
contingent derivative of F at .x; N if
N y//
graph .DF.x; N D T .graph .F /; .x;
N y//;
N
N y//
where T .graph .F /; .x; N is the contingent cone of graph .F / at .x;
N y/:
N
A geometric interpretation of the contingent derivative is given in Fig. 11.1. The
graph of the contingent derivative is obtained by locally approximating the graph of
the underlying set-valued map by means of the contingent cone.
Fig. 11.1 Contingent

derivative of the set-valued
map F
11.1 Contingent Derivatives of Set-Valued Maps 401
In the subsequent development, for a single-valued map f W X ! Y; the

N f .x//
contingent derivative of f at .x; N 2 X Y will be denoted by Df .x/:
N That
N WD Df .x;
is, Df .x/ N f .x//:
N
Remark 11.1.2. The above notion of the contingent derivative was introduced
by Aubin [14] and has been used extensively in variational analysis, nonsmooth
analysis, optimization, viability theory, finance, among others. It has now become
a standard derivative notion for set-valued maps and its properties can be found
in most of the recent books on these subjects. In our treatment of the subject,
besides following the classical paper [14], we used the excellent books by Aubin
and Ekeland [16], Aubin and Frankowska [18], and Ha and Papageorgiou [267].
We recall that given normed spaces X and Y , a function f W X ! Y is said to
be Hadamard directionally differentiable at xN 2 X in the direction x 2 X; if the
Hadamard directional derivative
f .xN C tz/ f .x/

N
f 0 .x;
N x/ D lim ;
.t;z/!.0C ;x/ t
exists as an element of Y:
We have the following insightful result which is quite easy to prove (see
Ward [596]):
Proposition 11.1.3. Let X and Y be normed spaces and let f W X ! Y be
Hadamard directionally differentiable at xN 2 X in the direction x 2 X: Then
N f .x//.x/
Df .x; N D ff 0 .x;
N x/g:
The following simple examples illustrate the notion of the contingent derivative
(see [563]):
Example 11.1.4. Let F W R R be a set-valued map given by

fy 2 Rj y
0g; if x D 0;
F .x/ WD
fy 2 Rj y > jxjg; if x ¤ 0:
Then, for every x 2 R; the contingent derivatives of the set-valued maps F and
.F C RC / at .0; 0/ are given by D.F C RC /.0; 0/.x/ D DF.0; 0/.x/ D fyj y
jxjg:
Example 11.1.5. Let F W R R2 be a set-valued map given by

f.0; 0/g; if x 0;
F .x/ WD
f.y1 ; y2 / 2 R2 j y2 D y12 ; 0 y1 xg; if x > 0:
Then,

f.0; 0/g; if x 0;
DF.0; .0; 0//.x/ D
f.y1 ; y2 / 2 R2 j y2 D 0; 0 y1 xg; if x > 0;
whereas
D.F C R2C /.0; .0; 0//.x/ D f.y1 ; y2 /j y1

0; y2
0g;
for every x 2 R:

f0g; if x 0;
F .x/ WD p
f0; xg; if x > 0:
Then,

f0g; if x ¤ 0;
DF.0; 0/.x/ D
fyj y 0g; if x D 0;

fyj y
0g; if x < 0;
D.F C R/.0; 0/.x/ D
R; if x
0;
showing the difference in the derivative of a set-valued map and its profile map.

f0g; if x 0;
F .x/ WD p
f xg; if x > 0:
Then,
8
< f0g; if x < 0;
DF.0; 0/.x/ D fyj y 0g; if x D 0;
:
f;g; if x > 0;
depicting that the domain of the contingent derivative is not necessarily the whole
space.
In the following, we give various characterizations of the contingent derivative.
For this, we first recall that given normed spaces X and Y , a set-valued map F W
X Y is called Lipschitz around xN 2 X , if there exist a positive constant L
and a neighborhood U.x/ N dom .F / such that for every x; z 2 U.x/; N we have
F .x/ F .z/ C Lkx zkB:
Proposition 11.1.8. Let X and Y be normed spaces, let F W X Y be a

N y/
set-valued map, and let .x; N 2 graph .F /: Then:
1. A pair .x; y/ 2 graph .DF.x; N y//
N if and only if there are sequences ftn g P and
f.xn ; yn /g X Y with tn # 0 and .xn ; yn / ! .x; y/ such that yN C tn yn 2
F .xN C tn xn / for every n 2 N:
2. A pair .x; y/ 2 graph .DF.x; N y//
N if and only if

F .xN C t x/
O yN
lim inf d ;y D 0:
O /!.x;0C /
.x;t t
N y/
3. The contingent derivative of F at .x; N is the set-valued map DF.x;
N y/
N WX
Y given by
F .xN C t x/
O yN
N y/.x/
DF.x; N D lim sup :
O /!.x;0C /
.x;t t
4. If xN 2 int.dom .F // and F is Lipschitz around x;

N then .x; y/ 2 graph .DF.x;
N y//
N
if and only if

F .xN C tx/ yN
lim inf d ;y D 0:
t #0 t
Proof. The proof follows from the characterizations of the contingent cone (see
[18, 267]). t
u
It is at once evident from the first characterization given above that the contingent
derivative of the inverse of a set-valued map is the inverse of the contingent
derivative. That is,
DF1 .y;
N x/
N D DF.x; N 1 :
N y/
It is also clear that the domain of the contingent derivative does not need to be
the whole space X: The following result sheds some light on this issue. We recall
that a set-valued map F is pseudo-convex at .x; N y/
N 2 graph .F /, if its graph is a
pseudo-convex set at .x; N y/:
N That is,
graph .F / .x;
N y/
N T .graph .F /; .x;
N y//;
N
or equivalently, for every .x; y/ 2 graph .F /; we have
F .x/ yN DF.x;
N y/.x
N x/:
N

N y/
1. The following inclusion holds
N y///
cl .dom .DF.x; N T .dom .F /; x/:
N
N y/;
2. If the map F is pseudo-convex at .x; N then
N y///
cl .dom .DF.x; N D T .dom .F /; x/:
N
3. If Y is finite dimensional, xN 2 int.dom .F //; and F is Lipschitz around x;

N then
dom .DF.x; N y//
N D X: Moreover, the contingent derivative DF.x; N y/
N is Lipschitz
(with the same modulus).
Proof. The proof of the first statement being obvious, let us verify the second one.
By using the properties of the projection PX of X Y into X and the fact that F is
pseudo-convex, we have
N y///
cl .dom .DF.x; N D cl .PX .graph .DF.x;
N y////;
N
D cl .PX .T .graph .F /; .x;
N y////;
N
N
D T .PX .graph .F //; x/;
D T .dom .F /; x/:
N
For the final part, let x 2 X be arbitrary. Then for every sufficiently small t > 0;
we have
yN 2 F .x/
N F .xN C tx/ C LtkxkBY ;
implying that there exists yNt 2 F .xN C tx/ such that
yNt yN
yt D 2 LkxkBY ;
t
which, due to the compactness of the ball BY in the finite-dimensional space Y;
ensures that a subsequence fyNt g converges to some y: It follows at once that y 2
DF.x;N y/.x/;
N proving that dom .DF.x; N y//
N D X:
Finally, for the Lipschitz continuity of DF.x; N y/;
N let x; O xQ 2 X and let yO 2
DF.x;N y/.
N x/:
O We claim that there exists yQ 2 DF.x; N y/.
N x/Q such that kyO yk
Q
LkxO xk: Q Let ftn g P and fyOn g Y be such that tn ! 0; yOn ! yO and
yN C tn yOn 2 F .xN C tn xO n /: Then the Lipschitz continuity of F ensures that we can
find zn 2 F .xN C tn xQ n / such that kzn yN tn yOn k Ltn kxO xk:
Q
zn yN
By choosing a subsequence, if necessary, we may assume that ! yQ 2
tn
DF.x;N y/.
N x/:
Q Evidently, yQ satisfy the estimate kyO yk Q LkxO xk:Q The proof is
complete. t
u
The following simple result sheds some light on the kernel of the contingent
derivative:
N y/
T .F 1 .y/;
N x/N Ker.DF.x;
N y//
N WD DF.x; N 1 .0/:
N y/ (11.1)
Furthermore, the equality holds in the above formula, if F 1 is Lipschitz-like

N x/:
around .y; N
Proof. The proof of (11.1) follows from the definition of the contingent cone and
the converse inclusion is a consequence of the Lipschitz-like property as depicted in
many other results. t
u
For our next result, we recall the notion of calmness which is weaker than the
Aubin property (see [243]). Given normed spaces X and Y , a set-valued map
F W X Y is said to be calm at .x; N y/
N 2 graph .F /; if there exist c > 0
and neighborhoods U.x/ N and V .y/N such that for all x 2 U.x/N and for all y 2
F .x/ \ V .y/;
N we have
N c d.x; x/:
d.y; F .x// N (11.2)
We now give a useful property of the contingent derivatives of calm maps (see
[380]):
Proposition 11.1.11. Let X and Y be normed spaces, and let F W X Y be a
N y/
set-valued map that is calm at .x; N 2 graph .F /: Then
N y/.0/
DF.x; N D T .F .x/;
N y/:
N
Proof. Let y 2 DF.x; N y/.0/:

N Then there exist sequences ftn g P; f.xn ; yn /g
X Y such that tn # 0; .xn ; yn / ! .0; y/ and yN C tn yn 2 F .xN C tn xn /; for every
n 2 N: Let U.x/ N and V .y/ N be the neighborhoods appearing in the definition of
calmness. Then for sufficiently large n, we have xN C tn xn 2 U.x/ N and yN C tn yn 2
N such that there exists fyNn g F .x/
V .y/ N and kyN C tn yn yNn k ctn kxn k C tn2 :
Then, for zn WD tn1 .yNn y/;
N we have kzn yn k ckxn k C tn ; which confirms that
zn converges to y: Consequently, y 2 T .F .x/;
N y/:N The proof is complete. t
u
The following result gives useful information about the contingent derivative.
Theorem 11.1.12. Let X and Y be normed spaces, let L and S be nonempty sets
in X , where S is assumed to be open, let G W X Y be a set-valued map, and let
f W S ! Y be a single-valued map. We define a set-valued map F W X Y by

f .x/ G.x/ if x 2 L;
F .x/ D
; if x 62 L:
If the map f is Fréchet differentiable at xN 2 S \dom .F /; then for every .x;

N y/
N 2
graph .F /; we have

f 0 .x/.x/
N DG.x;
N f .x/
N y/.x/
N if x 2 T .L; x/;
N
N y/.x/
DF.x; N
; if x 62 T .L; x/:
N
Furthermore, the converse of the above inclusion holds if the set L is convex and
N In particular, if G.x/ D G; for every x 2 dom .G/,
the map G is Lipschitz at x:
then
N y/.x/
DF.x; N D f 0 .x/.x/
N T .G; f .x/
N y/
N for every x 2 T .L; x/:
N
Proof. Let .x; y/ 2 graph .DF.x; N y//

N be arbitrary. Then there are sequences
f.xn ; yn /g X Y and ftn g P such that .xn ; yn / ! .x; y/; tn # 0 and
yN C tn yn 2 F .xN C tn xn /; for every n 2 N; implying that
xN C tn xn 2 L;
yN C tn yn 2 f .xN C tn xn / G.xN C tn xn /:
Clearly, the first containment, in view of the properties of ftn g and fxn g; at once
confirms that x 2 T .L; x/:N Moreover, since the map f is Fréchet differentiable at
N we have
x;
N C tn f 0 .x/.x/
f .xN C tn xn / D f .x/ N C tn h.tn /;
with h.tn / ! 0 as tn # 0: Therefore,

N yN C tn f 0 .x/.x/
f .x/ N C h.tn / yn 2 G.xN C tn xn /;
which implies that
y 2 f 0 .x/.x/
N DG.x;
N f .x/
N y/.x/:
N
For the converse, let .x; y/ 2 X Y be such that
x 2 T .L; x/;
N
0
N
f .x/.x/ y 2 DG.x;
N f .x/
N y/.x/:
N
N D A.L; x/;
Since the set L is convex, we have T .L; x/ N and therefore there are
sequences ftn g P with tn # 0 and fxn g and fxN n g in X with xn ! x; xN n ! x and
fyn g Y with yn ! .f 0 .x/x
N y/ such that for every n 2 N; we have
xN C tn xn 2 L;
N yN C tn yn 2 G.xN C tn xn /:
f .x/
N we can find fyNn g Y with yNn !

Since the map G is locally Lipschitz at x;
f 0 .x/.x/
N y such that
N yN C tn yNn 2 G.xN C tn xn / for every n 2 N:

f .x/
N we have
Using the Fréchet differentiability of the map f at x;
N C tn f 0 .x/.x/
f .xN C tn xn / D f .x/ N C tn h.tn /;
with h.tn / ! 0 as tn # 0: Consequently, for every n 2 N, we have
f .xN C tn xn / tn .f 0 .x/.x/
N C h.tn // yN C tn yNn 2 G.xN C tn xn /;
implying
yN C tn .f 0 .x/.x/
N C h.tn / yNn / 2 f .xN C tn xn / G.xN C tn xn /;
where

f 0 .x/.x/
N C h.tn / yNn ! f 0 .x/.x/
N f 0 .x/.x/
N C y D y;
which confirms that y 2 DG.x; N y/.x/:

N
N f .x/
Finally, if G is constant, then DG.x; N y/
N D T .G; f .x/
N y/:
N The proof
is complete. t
u
11.1.1 Miscellaneous Graphical Derivatives of Set-valued

Maps
The following definition introduces some other notions of derivatives of set-valued

maps.
N y/
map, and let .x; N 2 graph .F /: Then:
N y/
1. A set-valued map DA F .x; N W X Y is called the adjacent derivative of F
N y/;
at .x; N if
N y//
graph .DA F .x; N D A.graph .F /; .x;
N y//;
N
N y//
where A.graph .F /; .x; N is the adjacent cone of graph .F / at .x;
N y/:
N
N y/
2. A set-valued map DC F .x; N W X Y is called the Clarke derivative of F at
N y/;
.x; N if
N y//
graph .DC F .x; N D C.graph .F /; .x;
N y//;
N
N y//
where C.graph .F /; .x; N is the Clarke tangent cone of graph .F / at .x;
N y/:
N
3. A set-valued map DR F W X Y is called the radial derivative of F at .x; N y/;
N
if
N y//
graph .DR F .x; N
N D R.graph N y//;
.F /; .x; N
N
where R.graph N y//
.F /; .x; N is the closed radial cone of graph .F / at .x;
N y/:
N
4. A set-valued map DL F W X Y is called the lower Dini derivative of F at
N y/;
.x; N if for every sequence ftn g P and fxn g with tn # 0 and xn ! x; there
exist a sequence fyn g Y and an index m 2 N such that yN Ctn yn 2 F .xN Ctn xn /
for every n
m:
5. A set-valued map DQ F W X Y is called the quasi-interiorly Clarke
derivative of F at .x; N y/;
N if for every sequence ftn g P and every sequence
f.xn ; yn /g graph .F / with tn # 0 and .xN n ; yNn / ! .x; N y/;
N and for every
fxn g ! x; there exist a sequence fyn g Y with yn ! y and an index m 2 N
such that yNn C tn yn 2 F .xN n C tn xn / for every n
m:
For a single-valued map f W X ! Y; the adjacent derivative, the Clarke
derivative, the radial derivative, the lower Dini derivative, and the quasi-interiorly
Clarke derivative are denoted by DA f .x/; N DC f .x/;
N DL f .x/, N DR f .x/;N and
N respectively.
DQ f .x/;
Remark 11.1.14. The lower Dini derivative was proposed in the celebrated paper
by Penot [462], where one also finds the definition of the contingent derivative
termed as the upper Dini derivative. This notion turns out to be quite useful in
giving optimality conditions and sensitivity analysis in set-valued optimization. The
notions of the adjacent derivative and the Clarke derivative are more akin to the
idea of the contingent derivative and have also been used very often in set-valued
optimization. The notion of the quasi-interiorly Clarke derivative was introduced by
Thibault [571]. However, the above sequential definition of this derivative was used
by Amahroq and Thibault [6]. The notion of the radial derivative was proposed by
Taa [548].
The following result gives characterizations of the adjacent derivative, the Clarke
derivative, and the lower Dini derivative.
Proposition 11.1.15. Let X and Y be normed spaces, let F W X Y be a set-
N y/
valued map, and let .x; N 2 graph .F /: Then:
1. A pair .x; y/ 2 graph .DA F .x; N y//
N if and only if for every ftn g P with tn # 0
there exists f.xn ; yn /g X Y with .xn ; yn / ! .x; y/ such that yN C tn yn 2
F .xN C tn xn / for every n 2 N:
2. A pair .x; y/ 2 graph .DC F .x; N y//
N if and only if for every ftn g P with
tn # 0 and for every fxN n ; yNn /g graph .F / with .xN n ; yNn / ! .x;
N y/;
N there exists
f.xn ; yn /g X Y with .xn ; yn / ! .x; y/ such that yNn C tn yn 2 F .xN n C tn xn /
for every n 2 N:
3. An element .x; y/ 2 graph .DA F .x;

N y//
N if and only if

F .xN C t x/
O yN
lim sup inf d ;y D 0:
O
t !0C x!x t

N then .x; y/ 2
N y//
graph .DA F .x; N if and only if

F .xN C tx/ yN
lim d ;y D 0:
t #0 t
5. An element .x; y/ 2 graph .DC F .x;

N y//
N if and only if

F .xO C t x/
Q yO
lim sup inf d ;y D 0:
graph .F / Q
x!x t
t !0C ; .x;
O y/
O ! .x;
N y/
N

N then .x; y/ 2
N y//
graph .DC F .x; N if and only if

F .xO C tx/ yO
lim d ;y D 0:
t !0C ; .x;
O y/
O
graph .F /
! .x;
N y/
N
t
7. An element .x; y/ 2 graph .DL F .x;

N y//
N if and only if

F .xN C t x/
O yN
lim d ;y D 0:
O /!.x;0C /
.x;t t
N y/
8. The lower Dini derivative of F at .x; N is a set-valued map DL F .x;
N y/
N WX
Y given by
F .xN C t x/
O yN
N y/.x/
DL F .x; N D lim inf :
O /!.x;0C /
.x;t t
9. An element .x; y/ 2 graph .DQ F .x;N y//

N if and only if for each neighborhood
U.y/; there exist a neighborhood U.x/;N a neighborhood U.y/;N a real number
s > 0; and a neighborhood U.x/ such that for every .u; v/ 2 .U.x/
N U.y//
N \
graph .F /; t 2 .0; s/; and w 2 U.x/; we have
Œ.u; v/ C t .fwg U.y// \ graph .F / ¤ ;:
The above result shows that for certain Lipschitz maps, the contingent derivative,
the Clarke derivative, and the adjacent derivative have a simpler structure. As it is
evident from the properties of the tangent cones studied earlier, when the graph
of the involved set-valued map is a convex set, all these derivatives coincide with
the radial derivative and they share the individual properties of these derivatives
listed above. Nonetheless, we have the following interesting property of the adjacent
derivative for set-valued maps taking convex values (see [267]):
Proposition 11.1.16. Let X and Y be normed spaces, let F W X Y be a convex
valued map, that is, for every x 2 dom .F /; the set F .x/ is convex. If F is Lipschitz
N then for every .x;
around x; N y/
N 2 graph .F /; the adjacent derivative DA F .x; N y/
N is
a convex valued map.
Proof. For x 2 dom .DA F .x; N y//;
N let y; z 2 DA F .x;
N y/.x/
N be arbitrary. Then
for any ftn g P with tn # 0; there are sequences f.xn ; yn /g and f.xQ n ; zQn /g with
.xn ; yn / ! .x; y/ and .xQ n ; zQn / ! .x; z/ such that yN C tn yn 2 F .xN C tn ; xn / and
yN C tn zQn 2 F .xN C tn ; xQ n /; for every n 2 N: By using the fact that F is Lipschitz
around x; N we can show that there exists a sequence fzn g such that zn ! z and
yN C tn zn 2 F .xN C tn xn /. Since F is convex valued, for every s 2 Œ0; 1; we have
yN C tn .syn C .1 s/zn / 2 F .xN C tn xn /;
and since .syn C .1 s/zn / ! .sy .1 s/z/, we deduce that .sy C .1

s/z/ 2 DA .x;N y/.x/;
N N y/.x/:
proving the convexity of the set DA F .x; N The proof
is complete. t
u
The following is an analogue of Theorem 11.1.12 for the adjacent and the Clarke
derivatives.
Theorem 11.1.17. Let X and Y be normed spaces, let L and S be nonempty sets
in X , where S is assumed to be open, let G W X Y be a set-valued map and let
f W S ! Y be a single-valued map. We define a set-valued map F W X Y by

f .x/ G.x/ if x 2 L;
F .x/ D
; if x 62 L:
Then:
1. If f is Fréchet differentiable at xN 2 S \ dom .F /; then for every .x;
N y/
N 2
graph .F /, we have

f 0 .x/.x/
N DA G.x;
N f .x/
N y/
N if x 2 A.L; x/;
N
N y/.x/
DA F .x; N D
; if x 62 A.L; x/:
N
2. If f is continuously differentiable at xN 2 S \ dom .F /; then for every .x;

N y/
N 2
graph .F /; we have

f 0 .x/.x/
N DC G.x;
N f .x/
N y/
N if x 2 C.L; x/;
N
N y/.x/
DC F .x; N D
; if x 62 C.L; x/:
N
As a direct consequence of the above result, we obtain the following useful

corollary. Here, by fjL ; we denote the restriction of the map f to the set L.
Corollary 11.1.18. Let X and Y be normed spaces, let L X be nonempty, and

let f W X ! Y be a single-valued map.
1. If f is Fréchet differentiable at xN 2 L; then
N D f 0 .x/jT .L;x/
DfjL .x/ N
N D f 0 .x/jA.L;x/
DA fjL .x/ N :
2. If f is continuously differentiable at xN 2 L, then
N D f 0 .x/jC.L;x/
DC fjL .x/ N :
In the following, we give a result concerning the computation of the contingent

derivative of a composition map. Let X , Y , and U be normed spaces, let F W U
X be a set-valued map, and let f W X U ! Y be a single-valued map. With these
we associate the set-valued map f F W U Y defined by
f F .u/ WD f .F .u/ fug/ :
N for some .Nu; y/

Our objective is to evaluate Df F .Nu; y/ N 2 graph .f F /. Recall
that a set-valued map G W X Y is upper Lipschitz at xN 2 dom.G/; if there exist
L; > 0 such that G.x/ G.x/ N C L kx xk N BY ; for every x 2 B.x;
N /.
We have the following:
Theorem 11.1.19. Suppose that xN 2 F .Nu/, yN D f .x; N uN / and that f is Fréchet
N uN /. Then the following inclusions hold:
differentiable at .x;
N uN / .DF.Nu; x/.u//
rx f .x; N C ru f .x;
N uN /.u/ Df F .Nu; y/.u/
N for every u 2 U;
(11.3)
N uN / .DL F .Nu; x/.u//
rx f .x; N C ru f .x;
N uN /.u/ DL f F .Nu; y/.u/
N for every u 2 U:
(11.4)
Furthermore, define a set-valued map FQ W U Y X by
FQ .u; y/ WD fx 2 F .u/ j f .x; u/ D yg;
and assume that X is finite-dimensional and one of the following two conditions
holds:
(a) F .Nu/ D fxg
N and F is upper Lipschitz at uN :
(b) FQ .Nu; y/ N and FQ is upper Lipschitz at .Nu; y/.
N D fxg N
Then
rx f .x;
N C ru f .x;
N uN /.u/ D Df F .Nu; y/.u/
N for every u 2 U:
(11.5)
Proof. The proof is adopted from [214]. Let x 2 DF.Nu; x/.u/ N and take y D
rx f .x;
N uN /.x/ C ru f .x;
N uN /.u/. Then there exist ftn g ! 0C , f.un ; xn /g ! .u; x/
such that xN C tn xn 2 F .Nu C tn un /; for every n 2 N. Since f is differentiable at
N uN /, we have that for some Y fvn g ! 0,
.x;
f .xN C tn xn ; uN C tn un / D yN C tn .rx f .x;

N uN /.xn / C ru f .x;
N uN /.un / C vn / ;
for every n 2 N. Taking
yn WD rx f .x;
N uN /.xn / C ru f .x;
N uN /.un / C vn ;
since rf .x; N uN / is continuous, we have that fyn g ! y. Since yN C tn yn 2 f F .Nu C

tn un /; for every n 2 N, y 2 Df F .Nu; y/.u/. N Hence (11.3) holds.
Consider x 2 DL F .Nu; x/.u/ N and y as above. Let f.tn ; un /g ! .0C ; u/ with uN C
tn un 2 domf F (D domF ), for every n. Then there exists fxn g ! x such that
xN C tn xn 2 F .Nu C tn un / for n
n0 . Continuing as above, we get fyn g ! y with
yN Ctn yn 2 f F .Nu Ctn un / for n
n0 . Therefore y 2 DL f F .Nu; y/.u/, N whence (11.4).
Suppose now that dim X < 1 and (a) or (b) holds. Consider y 2 Df F .Nu; y/.u/. N
Then there exist ftn g ! 0C , f.un ; yn /g ! .u; y/ such that yN C tn yn 2 f F .Nu C tn un /
for every n 2 N. Let xQ n 2 F .Nu C tn un / be such that yN C tn yn D f .xQ n ; uN C tn un /,
that is, xQ n 2 FQ .Nu C tn un ; yN C tn yn /, for every n. Since F is upper Lipschitz at uN and
F .Nu/ D fxg N in case (a) and FQ is upper Lipschitz at .Nu; y/ N and FQ .Nu; y/N D fxgN in case
(b), there exists a bounded sequence fxn g X such that xQ n D xN C tn xn ; for every
n 2 N. Since dim X < 1, there exists a subsequence fxnp gp2N converging to some
x 2 X . It follows that x 2 DF.Nu; x/.u/, N and so y 2 rx f .x; N uN / .DF.Nu; x/.u//
N C
ru f .x;N uN /.u/. Taking into account (11.3), we obtain that (11.5) holds. The proof is
complete. t
u
In the following example, which is taken from [214], one of the conditions in (a)
and (b) of Theorem 11.1.19 is not satisfied, and (11.5) does not hold.
Example 11.1.20. Let U D X D Y D R, F; F 0 W U X and f W X U ! Y
be defined by
F .u/ WD Œ0; max.1; 1 C u/; F 0 .u/ WD Œ0; max.1; 1 C u/Œ; f .x; u/ WD x 2 x;
uN D 0, xN D 0, yN D f .0; 0/ D 0. Then F , F 0 are upper Lipschitz at uN , but F .Nu/ ¤

N ¤ F 0 .Nu/ (so (a) does not hold for F and F 0 ) and
fxg

Œ 14 ; 0 if u 0; Œ 14 ; 0 if u 0;
f F .u/ D f 0 F .u/ D
Œ 14 ; u.1 C u/ if u > 0; Œ 14 ; u.1 C u/Œ if u > 0:
y y y y
Let fx1 ; x2 g D fx 2 R j x 2 x D yg, with x1 x2 , for y
14 . Then
8 y y
< fx1 ; x2 g if u 2 R; y 2 Œ 4 ; 0;
1
FQ .u; y/ D fx2 g y
if u > 0; y 2 0; u.1 C u/;
:
; otherwise,
8 y y
ˆ
ˆ fx ; x g if u 2 R; y 2 Œ 14 ; 0Œ;
< 1 2
f f0g if u 2 R; y D 0;
F 0 .u; y/ D
ˆ fx2y g if u > 0; y 2 0; u.1 C u/Œ;
:̂
; otherwise.
Now, FQ is upper Lipschitz at .Nu; y/

N but FQ .Nu; y/
N D f0; 1g ¤ fxg,
N while F f0 is not
0
N (so (b) does not hold for F and F ). One obtains that
upper Lipschitz at .Nu; y/
N
DF.Nu; x/.u/ D DF 0 .Nu; x/.u/
N D R RC ; for every u 2 R;
N
Df F .Nu; y/.u/ D Df F 0 .Nu; y/.u/
N D f.u; y/ 2 R2 j y max.0; u/g for every u 2 R:
Since rx f .x;
N uN / D 1 and ru f .x;
N uN / D 0,
rx f .x;
N C ru f .x;
N uN /.u/ D 1; 0 ¤ 1; u D Df F .Nu; x/.u/;
N
for every u > 0, and similarly for F 0 .

Remark 11.1.21. Formulae (11.3) and (11.5) are stated by Tanino [561] under the
assumption that U , X , and Y are finite-dimensional and f is of class C 1 and by
Klose [330, Theorem 4.1].
The following result highlights a useful feature of the radial derivative (see
Taa [548]):
Theorem 11.1.22. Let X and Y be normed spaces, let F be a set-valued map, and
N y/
let .x; N 2 graph .F /: Then, for every x 2 X; we have
N y/.x
F .x/ yN DR F .x; N x/:
N (11.6)
Furthermore, if graph .F / is a convex set, then
F .x/ yN DF.x;
N y/.x
N x/:
N (11.7)
Proof. For any y 2 F .x/ y; N we have yN C y 2 F .x/: This, by setting tn WD

1; yn WD y; and xn WD x x; N can be expressed as yN C tn yn 2 F .xN C tn xn /;
for every n 2 N: Since .xn ; yn / ! .x x; N y/; we deduce that .x x; N y/ 2
N
R.graph N y//
.F /; .x; N D graph .DR F .x; N y//:
N Consequently, y 2 DR F .x; N y/.x
N x/N
and the proof of (11.6) is complete. Inclusion (11.7) follows from the fact that, for
a convex set, the closed radial tangent cone and the contingent cone coincide. t
u
We have the following useful corollary of the above result.
Corollary 11.1.23. Let X and Y be normed spaces, let C Y be a pointed, closed

N y/
and convex cone, let F be a set-valued map, and let .x; N 2 graph .F /: If epi .F /
is convex, then for every x 2 X; we have
F .x/ yN D.F C C /.x;

N y/.x
N x/:
N (11.8)
Proof. This follows from using the fact that, for a convex set, the closed radial
tangent cone and the contingent cone coincide and the chain of inclusions
F .x/ yN F .x/ yN C C D.F C C /.x;

N y/.x
N x/;
N
where the first inclusion is obvious and the second inclusion is (11.7) written for
F C C. u
t
11.1.2 Convexity Characterization Using Contingent

Derivatives
Having shown a useful property of the convex set-valued maps, we take a pause to
give a characterization of a certain convexity notion for set-valued maps. Recall that
given a normed space X and a continuously differentiable map f W X ! R; the
following two characterizations of convexity are well-known:
1. The map f is convex if and only if its Fréchet derivative rf is monotone.
That is,
hrf .x/ rf .z/; x zi

0 for every x; z 2 X: (11.9)
2. The map f is convex if and only if
f .z/ f .x/
hrf .x/; z xi for every x; z 2 X: (11.10)
The above two characterizations of convexity have been extended in several

directions. For instance, generalized notions of convexity have been characterized
by using generalized notions of differentiability. See Clarke, Stern, and Wolenski
[101], Correa, Jofré, and Thibault [114, 115], Daniilidis and Hadjisavvas [123],
Ellaia and Hassouni [176], Luc and Swaminathan [406], Penot and Sach [471],
Sach [515, 516], Sach and Yen [518], Zagrodny [608] and the cited references
therein. The primary objective of this short section is to give an analogue of the
first statement given above by giving a characterization of certain convexity notion
for set-valued maps in terms of the contingent derivative. Here our intention is only
to briefly discuss the involvement of set-valued maps and contingent derivatives in
this interesting area. The interested reader can find more elaborate results in the
literature.
Let X and Y be normed spaces, let C Y be a pointed, closed, and convex

cone, and let F W X Y be a set-valued map. We recall that the map F is called
almost C -convex, if for all x; z 2 X and every t 2 Œ0; 1; we have
tF .x/ C .1 t/F .z/ cl ŒF .tx C .1 t/z/ C C : (11.11)
We have the following characterization of the almost C -convexity:

Theorem 11.1.24. Let X and Y be normed spaces, let C Y be a pointed, closed,
and convex cone, and let F W X Y be a set-valued map. We formulate the
following statements:
(a) The map F is almost C -convex.
(b) For every .x1 ; y1 /; .x2 ; y2 / 2 graph .F /; we have
y2 y1 2 DF.x1 ; y1 /.x2 x1 /: (11.12)
(c) The contingent derivative DF is monotone on graph .F / in the following sense:

for every .x1 ; y1 /; .x2 ; y2 / 2 graph .F /; we have
0 2 DF.x1 ; y1 /.x2 x1 / C DF.x2 ; y2 /.x1 x2 /: (11.13)
Then: (a) H) (b) H) (c).

Proof. We begin with a proof of the first implication, namely, (a) H) (b). Let F
be almost C -convex and let .x1 ; y1 /; .x2 ; y2 / 2 graph .F / be arbitrary. We claim
that since the map F is almost C -convex, the cl .epi .F // is a convex set. Indeed,
by the almost C -convexity of F , and by setting FN .x/ WD cl .F .x/ C C /; for every
x1 ; x2 2 X and every t 2 .0; 1; we have
t.F C C /.x1 / C .1 t/.F C C /.x2 / cl ..F C C /.tx1 C .1 t/x2 // C C;
which implies that

t FN .x1 / C .1 t/FN .x2 / cl FN .tx1 C .1 t/x2 / C C D FN .tx1 C .1 t/x2 /;
which shows that graph .FN / is a convex set. Note that the identity cl .graph .FN // D
cl .graph ..F C C /// also holds. Then, for every .x;
N y/
N 2 graph ..F C C //; we have
C.graph ..F C C //; .x;

N y//
N D T .graph ..F C C //; .x;
N y//
N
D cl cone .graph ..F C C // .x;
N y//;
N
and, by setting .x; N WD .x1 ; y1 /; we deduce .x2 x1 ; y2 y1 / 2 T .graph ..F C

N y/
C //; .x1 ; y1 //; which confirms that (b) holds.
For the second implication, namely, (b) H) (c), we note that (11.12) implies that
y2 y1 2 DF.x1 ; y1 /.x2 x1 /;
y1 y2 2 DF.x2 ; y2 /.x1 x2 /;
and by summing up the above inclusions, we get (11.13). The proof is complete. u
t
The above result is based on the work by Sach and Yen [518] where more
elaborate relationships among the statements of Theorem 11.1.24, under various
settings, can be found.
11.1.3 Proto-Differentiability, Semi-Differentiability,

and Related Concepts
In the previous sections, we witnessed various derivatives of set-valued maps. It

turns out that, in the setting when two or more derivative notions coincide, the
resulting derivative benefits from their individual features and is typically equipped
with nicer features. In his celebrated paper, Rockafellar [495] introduced the notion
of proto-differentiable set-valued maps to recapture the combined benefits of the
contingent and the adjacent derivative of set-valued maps and investigated related
concepts. He also studied the proto-differentiability of certain constraint set-valued
maps appearing in parametric optimization. In this section, we briefly discuss some
of these situations to which many authors have contributed. This section is partly
adopted from Amahroq and Thibault [6]. See also [19, 495].
We begin with the following definition:
N y/
map, and let .x; N 2 graph .F /:
1. The map F is said to be proto-differentiable or derivable at .x; N y/;
N if the
N y/
contingent derivative and the adjacent derivative of F at .x; N coincide. That is,
N y//
T .graph .F /; .x; N D A.graph .F /; .x;
N y//:
N
N y/;
2. The map F is said to be strictly proto-differentiable at .x; N if the contingent
N y/
derivative and the Clarke derivative of F at .x; N coincide. That is,
N y//
T .graph .F /; .x; N D C.graph .F /; .x;
N y//:
N
N y/;
3. The map F is said to be semi-differentiable at .x; N if the contingent derivative
N y/
DF.x; N y/
N and the lower Dini derivative DL F .x; N coincide. That is, if for every
x 2 X; we have
F .xN C t x/
O yN F .xN C t x/
O yN
lim sup D lim inf : (11.14)
O /!.x;0C /
.x;t t C
O /!.x;0 /
.x;t t
4. The map F is said to be strictly semi-differentiable at .x;N y/;

N if the contingent
derivative DF.x; N y/
N and the quasi-interiorly Clarke derivative DQ F .x;
N y/
N coin-
cide. That is, if for every x 2 X;
F .xN C tz/ yN F .x C tz/ y

lim sup D lim inf :
.t;z/!.0C ;x/ t graph .F /
.t;z/!.0C ;x/;.x;y/ ! .x; N y/
N
t
(11.15)
Remark 11.1.26. Clearly any set-valued map with a convex graph is strictly
proto-differentiable and hence derivable. However, the converse in not true in
general. For example, consider a set-valued map F W R R defined by F .x/ WD
fjxj; jxjg; for all x 2 R. Then F is not strictly proto-differentiable at the point
.0; 0/ 2 graph .F /, but only derivable at this point (see [18]).
We have the following result which was originally proved by Amahroq and
Thibault [6]:
Theorem 11.1.27. Let X and Y be normed spaces, let F W X ! Y be a set-valued
map which is Lipschitz-like around .x; N y/
N 2 graph .F /: Then F is strictly semi-
N y/
differentiable at .x; N if and only if F is strictly proto-differentiable at .x;
N y/:
N
Proof. In view of the definitions of the strictly semi-differentiability and the
strictly proto-differentiability, it is enough to show that for any .v; w/ 2
N y//;
C.graph .F /; .x; N we have
F .x C tz/ y
w2 lim inf : (11.16)
graph .F / t
.t;z/!.0C ;v/;.x;y/ ! .x;
N y/
N
Let ftn g P; f.xn ; yn /g graph .F /; and fvn g Y be sequences such that

.tn ; xn ; yn / ! .0; x;
N y/
N and vn ! v: Then there exists .Nvn ; wN n / ! .v; w/ such that
yn C tn wN n 2 F .xn C tn vN n / for every n 2 N:
Since F is Lipschitz-like around .x; N y/;
N for sufficiently large n 2 N; we have
yn C tn wN n 2 W \ F .xn C tn vN n / F .xn C tn vn / C L tn kNvn vn kBY ;
where L is a positive constant and W is a neighborhood of y: N Hence there exists

bn 2 BY and wn WD .w N n C L kNvn vn kbn / ! w such that for sufficiently large
n 2 N; we have yn C tn wn 2 G.xn C tn vn / and hence (11.16) holds. The proof is
complete. t
u
We have the following analogue of the above result.

map which is Lipschitz-like around .x; N y/
N 2 graph .F /: Then F is derivable at
N y/
.x; N if and only if F is semi-differentiable at .x;
N y/:
N
Proof. Evidently, if F is semi-differentiable at .x;N y/;
N then it is derivable at .x;
N y/:
N
For the converse, let F is derivable at .x; N y/N and let .x; y/ 2 graph .DF.x; N y//
N
be arbitrary. Then, for every ftn g P; there exists f.zn ; wn /g X Y such that
tn # 0, .zn ; wn / ! .x; y/ and yN Ctn wn 2 F .xN Ctn zn /: Let fxn g X be an arbitrary
sequence such that xn ! x: Using the fact that F is Lipschitz-like at .x; N y/,
N we can
show that there exist a sequence fyn g and an index m 2 N such that .xN C tn xn ; yN C
tn yn / 2 graph .F /; for every n
m: Consequently, y 2 DL F .x; N y/.x/.
N The proof
is complete. t
u
We will now investigate the proto-differentiability of a set-valued map associated
to a specific constraint set. For this we first give the following two results which are
of interest (see [6]):
Theorem 11.1.29. Let X; Y and Z be Banach spaces, let G W X Y Z be a
set-valued map, let S Y and Q Z be nonempty closed sets. Assume that the
N y;
following metric regularity condition holds: For .x; N zN/ 2 X S Q \ graph .G/;
there exist k
0 and neighborhoods U.x/; N V .y/;
N and W .Nz/ such that for every
x 2 U.x/;
N y 2 V .y/N \ S; and z 2 G.x; y/ \ W .Nz/; we have
d..x; y/; .X S / \ G 1 .Q// k d.z; Q/: (11.17)
Then for the set-valued map F W X Y given by
F .x/ D fy 2 S j G.x; y/ \ Q ¤ ;g; (11.18)
the following inclusion holds
ŒX A.S; y/ N y;
N \ DL G.x; N zN//1 ŒA.Q; zN/ A.graph .F /; .x;
N y//:
N (11.19)
Moreover, if G.; /\Q is lower semicontinuous at .x;

N y;
N zN/ relative to its domain,
then
ŒX C.S; y/ N y;
N \ DQ G.x; N zN//1 ŒC.Q; zN/ C.graph .F /; .x;
N y//:
N (11.20)
Proof. We will only prove (11.20) as (11.19) can be proved by similar arguments.
Let
.v; w/ 2 ŒX C.S; y/ N zN//1 ŒC.Q; zN/

N y;
N \ DQ G.x;
be arbitrary. Let ftn g P and f.xn ; yn /g graph .F / be arbitrary sequences such

that tn # 0; and .xn ; yn / ! .x;
N y/:
N To prove that .v; w/ 2 C.graph .F /; .x; N y//;
N
we have to show that there exists f.vn ; wn /g with .vn ; wn / ! .v; w/ such that .xn C
tn vn ; yn C tn wn / 2 graph .F /:
In view of the containment w 2 C.S; y/; N there exists a sequence fw O n g with
wO n ! w such that yn C tn wO n 2 S; for every n 2 N: Furthermore, due to
.v; w/ 2 DQ G.x; N y;N zN//1 ŒC.Q; zN/; there exists u 2 C.Q; zN/ such that .v; w; u/ 2
graph .DQ G.x; N y;
N zN//:
Moreover, since G.; / \ Q is lower semicontinuous relative to its domain, there
is a sequence zn 2 G.xn ; yn / \ Q such that .tn ; xn ; yn ; zn / ! .0; x; N y;
N zN/: Also,
from the containment u 2 C.Q; zN/; there exists a sequence fung with un ! u
such that zn C tn un 2 Q for every n 2 N: By the definition of DQ G.x; N y;N zN/ and
the containment .v; w; u/ 2 graph .DQ G.x; N y;
N zN//; there are a sequence fOun g with
uO n ! u and an index m 2 N such that for every n
m; we have zn C tn uO n 2
G.xn C tn v; yn C tn wO n /:
Since
d .zn C tn uO n ; Q/ kzn C tn uO n zn tn un k D tn kun uO n k;
we obtain from the metric regularity hypothesis that for sufficiently large n 2 N; we
have
d ..xn C tn v; yn C tn w
O n /; graph .F // k tn kun uO n k < tn2 C k tn kun uO n k;
and hence for large n there exists f.vn ; wn /g X Y with .xn C tn vn ; yn C tn wn / 2

graph .F /; and
k.xn C tn v; yn C tn wO n / .xn C tn vn ; yn C tn wn /k < tn2 C k tn kun uO n k;
implying that
k.v; wO n / .vn ; wn /k < tn C k kun uO n k;
and consequently .vn ; wn / ! .v; w/ as w

O n ! w: The proof is complete. t
u
For our next result, we recall the notion of compactly approximable set-valued
maps.
N y/
map, and let .x; N 2 graph .F /: The map F W X Y is called compactly
N y/
approximable at .x; N in a direction z 2 X; if there are a compact set L.z/ in
Y; a number ˛ > 0; and a function r W0; 1ŒX ! R such that
1. lim r.t; x/ D 0:
.t;x/!.0;z/
2. For each x 2 U˛ .z/ and t 20; ˛ with F .xN C tx/ ¤ ;; we have
F .xN C tx/ \ .yN C t.L.z/ C r.t; x/BY // ¤ ;:

Remark 11.1.31. The above notion of compactly approximable sets was originally
proposed by Amahroq and Thibault [6]. Numerous illustrative examples and useful
properties of the above class of set-valued maps can be found in [6]. See also
Taa [548].
The following result makes use of the notion of compactly approximiable
set-valued maps.
Theorem 11.1.32. Let X; Y and Z be Banach spaces, let G W X Y Z be
a set-valued map, let S Y and Q Z be closed. If for .x; N y;
N zN/ 2 X S
Q \ graph .G/; the map G.; / \ Q is compactly approximiable at .x; N y;
N zN/ in all
directions .v; w/ 2 T .graph .F /; .x;
N y//;
N then for F given in (11.18),
N y//
T .graph .F /; .x; N ŒX T .S; y/
N \ DG.x; N zN//1 ŒT .Q; zN/:
N y; (11.21)
Proof. Let .v; w/ 2 T .graph .F /; .x; N y//

N be arbitrary. Then there exist sequences
ftn g P and f.vn ; wn /g with tn # 0 and .vn ; wn / ! .v; w/ such that .xN C tn vn ; yN C
tn wn / 2 graph .F /; which in view of (11.18) implies that for every n 2 N; we have
yN C tn wn 2 S and G.xN C tn vn ; yN C tn wn / \ Q ¤ ;:
By the definition of compactly approximability, for sufficiently large n 2 N; we
have
G.xN C tn vn ; yN C tn wn / \ S \ ŒNz C tn L.v; w/ C tn rn BZ ¤ ;;
where rn ! 0: Therefore, for sufficiently large n 2 N; there exists `n 2 L.v; w/ and

bn 2 BZ such that
zN C tn .`n C rn bn / 2 G.xN C tn vn ; yN C tn wn / \ Q: (11.22)
By the compactness of L.v; w/; there exists a subsequence f`n gn2J ; where J is
a subset of N; converging to some ` 2 L.v; w/ Z: Therefore, using (11.22),
we have ` 2 T .Q; zN/ and .v; w; `/ 2 T .graph .G/; .x; N y;
N zN// which means that
.v; w/ 2 DG.x; N zN//1 ŒT .Q; zN/: However, we have also shown that w 2 T .S; y/:
N y; N
u
With the above preparation,we have the following main result (see [6]):
Theorem 11.1.33. Let X; Y and Z be Banach spaces, let G W X Y Z be
a set-valued map, let S Y and Q Z be closed. If G is Lipschitz-like and
N y;
derivable at .x; N zN/; the map G.; / \ Q is compactly approximable at .x;
N y;
N zN/
in all directions .v; w/ 2 T .graph F; .x;N y//
N where F is given by (11.18), (11.17)
N y;
holds at .x; N zN/; and S and Q are derivable at yN and zN , then F is derivable at
N y/
.x; N with
N y/.x/
DF.x; N D fy 2 A.S; y/j
N DG.x;
N y;
N zN/.x; y/ \ A.Q; zN/ ¤ ;g;
where G is derivable.
Moreover, if, in addition, the map G.; / \ Q is lower-semicontinuous at .x; N y;

N zN/
N y;
relative to its domain, G is strictly proto-differentiable at .x; N zN/; and S and Q are
tangentially regular at yN and zN, then F is strictly proto-differentiable at .x;
N y/
N with
N y/.x/
DF.x; N D fy 2 C.S; y/j
N DG.x;
N y;
N zN/.x; y/ \ C.Q; zN/ ¤ ;g;
where G is strictly proto-differentiable.

Proof. Since the set-valued map G is Lipschitz-like at .x; N y;
N zN//, we have
N y;
DA G.x; N zN// D DL G.x;
N y;
N zN//; and DC G.x;
N y;
N zN// D DQ G.x;
N y;
N zN//: The proof
now follows from Theorems 11.1.29 and 11.1.32. t
u
In the remaining part of this section, we give sufficient conditions for the semi-
differentiability of convex set-valued maps. For the proof of this theorem, we need
the following technical lemmas whose proofs can be found in [214]. We recall that
by icrA we mean the algebraic relative interior of A.
Lemma 11.1.34. Let ; ¤ A X be a convex set and T W X ! Y be a linear
operator.
(i) If intA ¤ ; and T is open, then int.0; 1ŒA/ D0; 1ŒintA and T .intA/ D
intT .A/.
(ii) If icrA ¤ ; (e.g., if X is finite-dimensional), then icr.0; 1ŒA/ D0; 1ŒicrA
and T .icrA/ D icrT .A/.
Lemma 11.1.35. Let A X Y be a nonempty convex set. Suppose that either
.x; y/ 2 intA or .x; y/ 2 icrA and dim X < 1. Then for every sequence fxn g
PrX .A/ with xn ! x there exists fyn g ! y such that .xn ; yn / 2 A for every n 2 N.
The following result, originally due to Penot [462], is adopted from [214].
Theorem 11.1.36. Let F W X Y be a convex set-valued map. Suppose that one
of the following conditions holds:
(i) int.graph F / ¤ ;, x 2 int.dom .F // and y 2 F .x/;
(ii) X and Y are finite-dimensional, x 2 icr.dom .F //, and y 2 F .x/.
Then F is semidifferentiable at .x; y/.
Proof. We give the proof of (ii), that of (i) being similar. We already know that for
every u 2 X , we have
DL F .x; y/.u/ DF.x; y/.u/ D fv 2 Y j .u; v/ 2 T .graph F; .x; y//g:
Let us show that v 2 DL F .x; y/.u/; if .u; v/ 2 icrT .graph F; .x; y//, even
without requiring that x 2 icr.dom .F //.
Without loss of generality, we consider that .x; y/ D .0; 0/. Take
.u; v/ 2 icrT .graph F; .0; 0// D icr.0; 1Œgraph F / D 0; 1Œicr.graph F /;

where we used the relation icrA D icr.clA/, valid if A is a convex subset of a

finite-dimensional space, for the first equality and Lemma 11.1.34 for the last one.
Then there exist
0 > 0 and .u0 ; v0 / 2 icr.graph F / such that .u; v/ D
0 .u0 ; v0 /.
Let ftn g ! 0C and fun g ! u such that tn un D x C tn un 2 dom .F /; for every
n 2 N. It follows that fun g aff.dom .F // D span.dom .F // and .un =
0 / !
u0 2 icr.dom .F //. So there exists n0 2 N such that un =
0 2 dom .F / for n
n0 .
Therefore, by using Lemma 11.1.35, there exists a sequence .v0n / Y such that
fv0n g ! v0 and .un =
0 ; v0n / 2 graph .F /; for n
n0 . Let vn WD
0 v0n for n
n0 and
be arbitrary for n < n0 . So vn ! v. Since tn ! 0C , there exists n1
n0 such that
tn 1=
0 for n
n1 . Since .0; 0/ 2 graph F , we obtain that .x; y/ C tn .un ; vn / D
tn
0 .un =
0 ; v0n / 2 graph F for n
n1 . Therefore v 2 DL F .x; y/.u/.
Consider now u 2 icrT .dom .F /I 0/ D 0; 1Œicr.dom .F // and
v 2 DF.x; y/.u/. Therefore, we have u 2 icr .0; 1Œdom .F // and .u; v/ 2
cl .0; 1Œgraph F /. By Lemma 11.1.34, there exists v0 2 Y such that .u; v0 / 2
icr .0; 1Œgraph F /. It follows that for every
2 0; 1Œ,
.u; .1
/v C
v0 / D .1
/.u; v/ C
.u; v0 / 2 icr .0; 1Œgraph F / :
By what precedes, we have that .1

/v C
v0 2 DL F .x; y/.u/ for every
2
0; 1Œ. Taking into account that DL F .x; y/.u/ is a closed set, we obtain that v 2
DL F .x; y/.u/ for
! 0. Therefore for every u 2 icrT .dom .F /; x/,
DL F .x; y/.u/ D fv 2 Y j .u; v/ 2 T .graph F; .x; y//g D DF.x; y/.u/;
that is, F is semidifferentiable at every u 2 icrT .dom .F /; x/. Under our initial
condition that x 2 icr.dom .F //; we have T .dom .F /; x/ D span.dom .F /
x/ D icrT .dom .F /; x/, and consequently F is semidifferentiable. The proof is
complete. t
u
11.1.4 Weak Contingent Derivatives of Set-Valued Maps
For optimization problems posed in infinite-dimensional spaces, a fruitful approach

is to modify the definition of tangent cones by replacing the strong convergence of
sequences by the weak-convergence. The usefulness of this approach was depicted
in a thorough study of this idea by Borwein [59]. On the other hand, while using
tangent cones in product spaces, for instance to define graphical or epigraphical
derivatives of set-valued maps, one possibility is to use a combination of the
strong and the weak-convergence. This idea was explored by Rodríguez-Marín and
Sama [504] on which the forgoing discussion is based.
In the following, the notion .un ; vn / ! .u; v/ indicates that fung converges to u
s;w
strongly and fvn g converges to v weakly. Equipped with this notion, we propose the
following:
Definition 11.1.37. Let X and Y be normed spaces, let A X Y; and let .x;N y/
N 2
N of A at .x;
A: The quasi-weak contingent cone TQW .A; .x; y// N y/
N consists of all
.x; y/ 2 X Y such that there are sequences ftn g P and f.xn ; yn /g A with
.xn ; yn / ! .x;
N y/
N and tn .xn x;
N yn y/
N ! .x; y/.
s;w s;w
N y//
The quasi-weak contingent cone TQW .A; .x; N is the sequential weak-closure of
a subset of sequences,

N y//
TQW .A; .x; N D . kk/ cl .tn ..xn ; yn /
w

.x;
N y///
N j ftn g P; A ..xn ; yn // ! .x;
N y/
N ;
s;w
with respect to the product topology defined by the norm topology in the first
space and the weak topology in the second space. Consequently, TW .A; .x; N y//
N
is sequentially closed with respect to this product topology. That is, for every
sequence f.un ; vn /g TQW .A; .x;
N y//
N such that .un ; vn / ! .u; v/ we have .u; v/ 2
s;w
N y//.
TQW .A; .x; N
We equip the notion of the quasi-weak contingent cone to propose the following:
N y/
Definition 11.1.38. A set-valued map DW F .x; N W X Y is called the weak
N y/;
contingent derivative of F at .x; N if
N y//
graph .DW F .x; N D TQW .graph .F /; .x;
N y//;
N
N y//
where TQW .graph .F /; .x; N is the quasi-weak contingent cone of graph .F / at
N y/:
.x; N
To establish some connection between the contingent and weak contingent
derivatives, we introduce the following notions.
N y/
N y/;
1. The map F is said to be stable at .x; N if there exist constants " > 0 and k > 0
such that
F .x/ fyg
N C k kx xk
N BY for every x 2 U" .x/
N \ dom .F /:
2. The map F is said to be directionally compact at .x; N y/

N in the direction
x 2 T .dom .F /; x/;N if for every sequence ftn g P with tn # 0 and for
every sequence fxn g X with xn ! x; any sequence fyn g Y satisfying
yN C tn yn 2 F .xN C tn xn /; for each n 2 N; contains a convergent subsequence. If
F is directionally compact at .x; N y/
N for every direction x 2 T .dom .F /; x/,
N then
F is said to be directionally compact at .x;N y/.
N
Remark 11.1.40. Clearly, if F is stable then F .x/ N D fyg.N Moreover, if Y is

finite-dimensional, then as a consequence of the compactness of the closed unit
N y/,
ball, it follows that, if the map F is stable at .x; N then F is directionally compact
N y/.
at .x; N
The following result gives some useful implications of the directional compact-
ness.
Theorem 11.1.41. Let X and Y be normed spaces, let F W X Y be a set-valued
N y/
map, and let .x; N 2 graph .F /: Assume that the map F is directionally compact
N y/.
at .x; N Then:
1. For every f.xn ; yn /g graph .F / with xn ! x,
N we have yn ! y.N
2. The map F is stable at .x;N y/
N if and only if DF.x;
N y/.0/
N D f0g.
Proof. For the first part, we begin by defining
N 1=2 ;
tn D kxn xk
.xn x/
N
xN n D :
tn
Clearly tn # 0 and xN n ! 0: For yNn WD .yn y/= N hn , by the directional

N y/,
compactness of F at .x; N there is a subsequence fyNnk g converging to an element
y. Therefore,
ynk yN D tnk yNnk ! 0:

k
Indeed, following this reasoning, every subsequence of fyn g has a convergent

subsequence converging to yN and this implies that yn ! y:N
N y/.0/
Next we show that the stability of the map F implies that DF.x; N D f0g. Let
y 2 DF.x;N y/.0/
N be arbitrary. Then there are sequences ftn g P and f.xn ; yn /g
graph .F / such that .xn ; yn / ! .x;
N y/;
N and tn .xn x;
N yn y/N ! .0; y/:
Since the map F is stable, we deduce that for sufficiently large n; we have
ktn .yn y/k

N k ktn .xn x/k
N ;
which, by passing to the limit, yields that kyk 0; and hence y D 0:

For the converse, assume that the map F is directionally compact but not stable
N y/.
at .x; N In this case, we can find a sequence f.xn ; yn /g graph .F / with xn ¤ 0
such that xn ! xN and
kyn yk
N
! 1: (11.23)
kxn xk
N
We define
tn D kyn yk
N ;
.xn x/
N
xN n D ;
hn
.yn y/
N
yNn D :
hn
Then tn # 0 and hence xN n ! 0 by (11.23). By the directional compactness

N y/;
of the map F at .x; N there exists a subsequence fyNnk g of fyNn g converging to an
element y 2 Y . Since kyNn k D 1, we get a contradiction to the hypothesis that
DF.x;N y/.0/
N D f0g. The proof is complete. t
u
With the above preparation, we give the following result:
and convex cone, let F W X Y be a set-valued map, and let .x;
N y/
N 2 graph .F /:
N y/.
Let F be directionally compact at .x; N Then the following equivalences hold:
1. D .F C C / .x;
N y/./
N D DW .F C C / .x;
N y/./:
N
N y/./
2. DF.x; N D DW F .x;
N y/./:
N
N y//
Proof. Since TQW .epi .F /; .x; N T .epi .F /; .x//
N is always true, it suffices to
show that
N y//
TQW .epi .F /; .x; N T .epi .F /; .x;
N y//:
N
Let .x; y/ 2 TQW .epi .F /; .x;

N y//
N be arbitrary. Then there are sequences ftn g
P, fcn g C and f.xn ; yn /g graph .F / such that
.xn ; yn C cn / ! .x;
N y/
N
s;w
tn .xn x;
N yn C cn y/
N ! .x; y/:
s;w
From the directional compactness of the map F at .x; N y/;

N Theorem 11.1.41
ensures that yn ! y: N Furthermore, by the same result, we can assume, taking
subsequences if necessary, that the sequence .tn .yn y//
N converges to an element
y1 2 Y . Consequently, by definition y1 2 DF.x;N y/.x/.
N In particular, we have
.tn .yn y//

N ! y1 ;
w
which, due to the convergence properties of the sequences, confirms that the
sequence ftn cn g converges weakly to an element c 2 C: Therefore y D y1 C c
and we deduce that
y D y1 C c 2 DF.x;
N y/.x/
N C C D .F C C / .x;
N y/.x/:
N
N y//
Since .x; y/ was arbitrary, we have shown that TQW .epi .F /; .x; N
N y//:
T .epi .F /; .x; N The proof of the second part can be obtained by similar
arguments. t
u
11.1.5 A Lyusternik-Type Theorem Using Contingent

Derivatives
In this section, our objective is to give an analogue of the well-known Lusternik

theorem for set-valued maps by employing a weaker notion of metric regularity
recently proposed by Pühl and Schirotzek [481]. We first recall the notion of metric
regularity and related notions:
Definition 11.1.43. Let X and Y be Banach spaces, let F W X Y be a closed
N y/
set-valued map, and let .x; N 2 graph .F /:
N y/;
1. The map F is said to be an open covering with linear rate r at .x; N if
N and V .y/
there exist neighborhoods of U.x/ N such that for any x 2 U.x/
N and
t sufficiently small, we have
F .x/ \ V .y/
N C t r BY F .xN C tBX /: (11.24)
N y/
2. The map F is said to be metrically regular at .x; N with modulus r; if there
N and V .y/
exist neighborhoods U.x/ N such that for any x 2 U.x/
N and y 2 V .y/;
N
we have
d.x; F 1 .y// r d.y; F .x//: (11.25)
We have the following result whose proof can be found in Penot [461] or Borwein
and Zhu [66]:
Theorem 11.1.44. Let X and Y be Banach spaces and let F W X Y be a closed
set-valued map. Then the following statements are equivalent:
N y/
1. F is an open covering with linear rate r at .x; N 2 graph .F /:
N y/
2. F is metrically regular with modulus 1=r at .x; N 2 graph .F /:
3. F 1 is pseudo-Lipschitz with modulus 1=r at .y; N 2 graph .F 1 /:
N x/
We now give an extension of the notions used in the above classical result:
Definition 11.1.45. Let X and Y be Banach spaces, let F W X Y be a set-valued
N y/
N y/;
1. The map F is said to be linearly semi-open around .x; N if there are numbers
c > 0 and s > 0 such that for every .x; y/ 2 graph .F / with x 2 Us .x/N and
y 2 Us .y/;
N and for every t 2 Œ0; s; we have
y C c tkx xkB
N F .x C tkx xkB/:
N (11.26)
2. The map F is said to be metrically semi-regular around .x; N y/;

N if there are
N y 2 Us .y/
numbers c > 0 and s > 0 such that for every .x; y/ with x 2 Us .x/; N
and d.y; F .x// skx xk;
N we have
d.x; F 1 .y// c d.y; F .x//: (11.27)
3. The map F is said to be semi-pseudo-Lipschitz around .x; N y/;

N if there are
numbers c > 0 and s > 0 such that for every .x; y/ 2 graph .F / with
x 2 Us .x/
N and y 2 Us .y/;
N there is a neighborhood V .y/ such that for every
xO 2 x C sky ykB;
N we have
F .x/ \ V .y/ F .x/

O C c kx xkB:
O (11.28)
We have the following result connecting the above three concepts:

Theorem 11.1.46. Let X and Y be Banach spaces, let F W X Y be a set-valued
N y/
map, and let .x; N 2 graph .F /: Let the map F be closed-valued. Then the
following properties are equivalent:
N y/:
1. The map F is linearly semi-open around .x; N
N y/:
2. The map F is metrically semi-regular around .x; N
3. The map F 1 is semi-pseudo-Lipschitz around .y;
N x/:
N
Proof. See Pühl and Schirotzek [481]. t
u
With the above preparation, we give the following Lyusternik-type theorem for
set-valued maps. This result was originally proved by Pühl and Schirotzek [481].
Theorem 11.1.47. Let X and Y be Banach spaces, let F W X Y be a set-valued
N y/
map, and let .x; N 2 graph .F /: Let the map F be metrically semi-regular around
N 0/: Then:
.x;
N D Ker.DF.x;
T .Ker.F /; x/ N 0//: (11.29)
Proof. Let x 2 T .Ker.F /; x/ N be arbitrary. Then there are sequence ftn g P and
fxn g X with tn # 0 and xn ! x such that xN C tn xn 2 Ker.F /; for every n 2 N:
Therefore, 0 2 F .xN C tn xn / for every n 2 N; and consequently,
F .xN C tn xn /
0 2 lim sup DF.x;
N 0/.x/;
n!1 tn
which implies that x 2 Ker.DF.x; N 0//; proving the inclusion T .Ker.F /; x/

N
Ker.DF.x;N 0//:
For the converse, let x 2 Ker.DF.x;N 0// be arbitrary. If x D 0; then x 2
N Therefore, we assume that x ¤ 0: Let k and s be the semi-regularity
T .Ker.F /; x/:
N 0/; there are sequences ftn g P and fxn g

parameters. By the definition of DF.x;
such that tn # 0; xn ! x, and yn 2 F .xN C tn vn / such that
yn
lim D 0: (11.30)
n!1 tn
Furthermore, for sufficiently large n; we have xN C tn xn 2 Us .x/:

N Since x ¤ 0;
for sufficiently large n, we have
kyn k 1
skxk skxn k;
tn 2
which implies that yn 2 s tn kxn kB and that yn 2 sB: The metric semi-regularity of
F then applied to .xN C tn xn ; 0/ gives
d.xN C tn xn ; Ker.F // k d.F .xN C tn xn /; 0/ k kyn k:
Therefore, there is a wn 2 Ker.F / such that
kxN C tn xn wn k 2k kyn k: (11.31)
We set un D tn1 .wn x/;

N and hence xN C tn un 2 Ker.F /. We claim that un ! x:
For this, we note that
kun xk kun xn k C kxn xk

wn xN
D t xn
C kxn xk

n
kwn .xN C tn xn / k
D C kxn xk
tn
yn
2k C kxn xk ! 0;
tn

u
Several particular cases of the above result, including the classical Lusternik
theorem, can be found in [481, 524].
11.2 Calculus Rules for Derivatives of Set-Valued Maps
In this section, we give calculus rules for graphical derivatives of set-valued maps
using two techniques. Namely, using a direct approach and by the calculus of the
involved tangent cones.
11.2 Calculus Rules for Derivatives of Set-Valued Maps 429
11.2.1 Calculus Rules by a Direct Approach
We begin with a scalar multiplication rule for the contingent derivative.

Theorem 11.2.1. Let X and Y be normed spaces and let F W X Y be a
set-valued map. For every
2 RC and for every .x;
N y/
N 2 graph .F /; we have
N
y/.x/
D.
F /.x; N D
DF.x;
N y/.x/:
N (11.32)
Proof. For .x; y/ 2 graph .

DF.x; N y//;
N there are sequences f.xn ; yn /g X Y
and ftn g P such that tn # 0; xn ! x; yn ! y
and yN Ctn yn 2 F .xN Ctn xn /; which
N
implies that
yC
t n yn 2
F .xCt
N n xn /; and hence .x; y/ 2 graph .D.
F /.x;
N
y//:
N
We have shown that
DF.x; N y/.x/
N D.
F .x; N t y/.x/:
N The converse inclusion can
also be proved by similar arguments. t
u
The next three results, whose proofs follow from the definitions of the involved
objects, give calculus rules which either involve the lower Dini derivative or a
combination of the lower Dini derivative and the contingent derivative ( see [462]).
Theorem 11.2.2. Let X; Y and Z be normed spaces, let F W X Y and G W
Y Z be set-valued maps, let .x; N y/
N 2 graph .F /; and let .y;
N zN/ 2 graph .G/:
Then the following chain rules hold:
N zN/ ı DL F .x;
DL G.y; N y/
N DL .G ı F /.x;
N zN/;
N zN/ ı DF.x;
DL G.y; N y/
N D.G ı F /.x;
N zN/:
Theorem 11.2.3. Let X and Y be normed spaces, let F1 ; F2 W X Y be

N yN1 / 2 graph .F1 /; and let .x;
set-valued maps, let .x; N yN2 / 2 graph .F2 /: Then the
following sum formulae hold:
N yN1 / C DL F2 .x;
DL F1 .x; N yN2 / DL .F1 C F2 /.x;
N yN1 C yN2 /;
N yN1 / C DL F2 .x;
DF1 .x; N yN2 / D.F1 C F2 /.x;
N yN1 C yN2 /:
Theorem 11.2.4. Let X; Y1 , and Y2 be normed spaces, let F1 W X Y1 and F2 W

X ! Y2 be set-valued maps, let .x;
N yN2 / 2 graph .F2 /:
Then:
DL .F1 F2 /.x;
N yN1 ; yN2 / D DL F1 .x;
N yN1 / DL F2 .x;
N yN2 /;
D.F1 F2 /.x;
N yN1 ; yN2 / DF 1 .x;
N yN1 / DL F2 .x;
N yN2 /;
D.F1 F2 /.x;
N yN1 ; yN2 / DF 1 .x;
N yN1 / DF 2 .x;
N yN2 /:
In the following, we give a strengthening of some of the above results by giving

calculus rules for the contingent derivatives by an approach proposed by Luc [400].
We need the following:
Definition 11.2.5. Let X and Y be Banach spaces and let F W X Y be a

set-valued map.
1. F is called lower semidifferentiable at .x; N y/
N 2 graph .F /; if for any sequence
fxn g dom .F / converging to xN and for any sequence ftn g P with ftn .xn
N converging to some x 2 X; there exists a subsequence fynm g such that the
x/g
sequence ftnm .ynm y/g
N is convergent.
2. F is called upper semidifferentiable at .x; N y/
N 2 graph .F /; if for any sequence
f.xn ; yn /g graph .F /; not identical to .x;
N y/;
N and converging to .x; N y/;
N there
exists a subsequence ftnm g P such that ftnm .xnm x;N ynm y/g N converges to
some nonzero .x; y/ 2 X Y:
3. F is called compact at xN 2 dom .F /; if any sequence f.xn ; yn /g graph .F /
possesses a convergent subsequence provided that fxn g converges to x: N
Many useful results concerning lower semidifferentiable and upper semidiffer-
entiable set-valued maps can be found in [400].
The following result gives a chain rule for the contingent derivatives.
Theorem 11.2.6. Let X , Y; and Z be Banach spaces and let F W X Y and
G W Y Z be set-valued maps. For .x;
N y/
N 2 graph .F / and .y;
N zN/ 2 graph .G/; let
F be compact at xN and upper semidifferentiable at .x;
N y/;
N let DF.x;N y/.0/
N D f0g;
N Then:
and let G be closed on F .x/:
D.G ı F /.x;
N zN/ DG.y;
N zN/ ı DF.x;
N y/:
N (11.33)
Proof. Let .x; z/ 2 graph .D.GıF /.x; N zN// be arbitrary. Then, .x; z/ 2 T .graph .Gı
N zN// and consequently there are sequences ftn g P and f.xn ; zn /g
F /; .x;
graph .G ı F / such that .xn ; zn / ! .x; N zN/ and tn .xn x; N zn zN/ ! .x; z/: Since
f.xn ; zn /g graph .G ı F /; we can find fyn g Y such that yn 2 F .xn / and
zn 2 G.yn /; for every n 2 N: Since the map F is compact at x; N we can assume
that fyn g converges to some yN 2 F .x/:
N Furthermore, since the map G is closed, we
deduce that zN 2 G.y/: N
At this juncture, we use the upper semidifferentiability of the map F to ensure
that there exists a sequence fsn g P such that a subsequence of fsn .xn x; N yn y/g
N
converges to a nonzero .u; v/: Wenretain o the same
n o notations for subsequences as
tn sn
well. We consider the sequences sn and tn ; and by taking subsequences if
necessary, we can assume that one of the two sequences converges to t: We first
assume that stnn converges to t. Then tn .yn y; N zn zN/ converges to .tv; z/ and
tn .xn x;N yn y/
N converges to .tu; tv/ D .x; tv/: Consequently, z 2 DG.y; N zN/.tv/;
where tv 2 DF.x; N y/.x/;
N confirming (11.33).
Next, we assume that stnn converges to t. Then sn .yn y; N zn zN/ converges to
.v; tz/ whereas sn .xn x;N yn y/
N converges to .tx; v/ D .u; v/: Assume that t ¤ 0:
Then z 2 DG.y; N zN/.v=t/ where v=t 2 DF.x; N y/.x/
N confirming (11.33). It remains
to show that t ¤ 0: Note that if t D 0; then u D 0; and due to the assumption that
DF.x; N y/.0/
N D f0g; we obtain v D 0: This, however, is a contraction. The proof is
complete. t
u
We give another result recently obtained by Ward [596] who noted the usefulness
of the notion of directional compactness for deriving calculus rules.
Theorem 11.2.7. Let X , Y; and Z be normed spaces, let F W X Y and G W
Y Z be set-valued maps, let .x;
N y/
N 2 graph .F /; and let .y;
N zN/ 2 graph .G/:
Define H W X Z be
H.x; z/ WD fy 2 Y j y 2 F .x/; z 2 G.y/g:
N zN/; y/
Let H be directionally compact at ..x; N in the direction .x; z/ 2
graph .D.G ı F /.x;
N zN//: Then
N zN/ ı DF.x;
.x; z/ 2 graph .DG.y; N y//:
N
Proof. Let .x; z/ 2 graph .D.G ı F /.x; N zN// with H directionally compact at
N zN/; y/
..x; N in the direction .x; z/: Then there exist sequences ftn g P and f.xn ; zn /g
with tn # 0; .xn ; zn / ! .x; z/ such that .Nz Ctn zn / 2 .G ıF /.xN Ctn xn /, for every n 2
N: Therefore, there exists fwn g such that wn 2 F .xN C tn xn / and zN C tn zn 2 G.wn /:
By setting yn WD tn1 .wn y/; N we have yN C tn yn 2 H.xN C tn xn ; zN C tn zn /: Now,
by invoking the directional compactness, we may assume, by taking subsequence if
necessary, that fyn g ! y for some y 2 Y: Then .x; y/ 2 T .graph .F /; .x; N y//
N and
.y; z/ 2 T .graph .G/; .y;N zN// implying that .x; z/ 2 graph .DG.y; N zN/ ı DF.x;N y//:
N
u
The following result gives a sum formula for the contingent derivatives.
Theorem 11.2.8. Let X and Y be Banach spaces and let F1 W X Y and F2 W
X Y be set-valued maps. Let F1 and F2 be closed at xN 2 X; let either F1 or F2
N let for each yN1 2 F1 .x/;
be compact at x, N F1 be upper semidifferentiable at .x;
N yN1 /
N yN1 /.0/ D f0g; and let .x;
with DF1 .x; N yN2 / 2 graph .F2 /: Then:
D.F1 C F2 /.x;
N yN1 C yN2 / DF 1 .x;
N yN1 / C DF 2 .x;
N yN2 /: (11.34)
Proof. The proof follows the same line of arguments as in Theorem 11.2.6 and is
omitted. t
u
Our objective now is to give some calculus rules for the Clarke derivative and the
quasi-interiorly Clarke derivative. For this, we recall the following notion introduced
by Thibault [571]:
Definition 11.2.9. Let X and Y be normed spaces, let F1 ; F2 W X Y be
N yN1 / 2 graph .F1 /, and let .x;
set-valued maps, let .x; N yN2 / 2 graph .F2 /. The maps
F1 and F2 are additively separate at .x;N yN1 ; yN2 /; if for every neighborhood U.yN1 /
and every neighborhood U.yN2 /; there exist a neighborhood U.yN1 C yN2 / and a
neighborhood U.x/N such that for every x 2 U.x/; N we have
.F1 .x/ C F2 .x// \ U.yN1 C yN2 / U.yN1 / \ F1 .x/ C U.yN2 / \ F2 .x/:

With the above notion, we are now in a position to announce the following sum
formula:
N yN1 / 2 graph .F1 /, and let .x;
set-valued maps, let .x; N yN2 / 2 graph .F2 /. Let F1
N yN1 ; yN2 /: Then:
and F2 be additively separate at .x;
N yN1 / C DC F2 .x;
DQ F1 .x; N yN2 / DC .F1 C F2 /.x;
N yN1 C yN2 /
N yN1 / C DQ F2 .x;
DQ F1 .x; N yN2 / DQ .F1 C F2 /.x;
N yN1 C yN2 /:
Proof. See Thibault [571]. t

u
We now recall the following notion of lower semicontinuity:
Definition 11.2.11. Let X and Y be normed spaces and let F W X Y be
a set-valued map. The map F is said to be lower-semicontinuous at .x; N y/
N 2
graph .F / relative to S X; if for each neighborhood U.y/;
N there is a neigh-
borhood U.x/N such that U.y/
N \ F .x/ ¤ ;; for every x 2 S \ U.x/:
N
We are now in a position to give the following formula for a composition:
Theorem 11.2.12. Let X , Y and Z be normed spaces, let F W X Y and G W
Y Z be two set-valued maps, let .x;N y/N 2 graph .F / and let .y;
N zN/ 2 graph .G/.
Let the set-valued map defined from X Z to Y by .x; z/ ! G.x/ \ F 1 .z/ be
N zN/; y/
lower-semicontinuous at ..x; N relative to graph .F ı G/: Then:
N zN/ ı DQ F .x;
DQ G.y; N y/
N DQ .G ı F /.x;
N zN/;
N zN/ ı DC F .x;
DQ G.y; N y/
N DC .G ı F /.x;
N zN/:
Furthermore, if G is Lipschitzian, then
N zN/ ı DC F .x;
DC G.y; N y/
N DC .G ı F /.x;
N zN/:
Proof. See Thibault [571]. t

u
Recently in an interesting work, Li, Meng, and Penot [380] derived numerous
calculus rules for derivatives of set-valued maps and gave numerous examples and
counterexamples.
11.2.2 Derivative Rules by Using Calculus of Tangent Cones
In set-valued optimization, an efficient approach for obtaining various calculus

rules can be devised by using certain properties of the tangent cones which are
being used to define the derivatives of the involved set-valued maps. This approach,
which seems to be devised independently, by Ward and Borwein [597] and Aubin
and Frankowska [17], has been extensively pursued ever since [17, 597]. Given
the generality of this approach and the fact that many derivatives of set-valued
maps defined via suitable tangent cones can benefit from this strategy, we would
give the calculus rules for abstract tangent cone approximations, relying only on
certain features of the underlying cones. Our approach is based on Aubin and
Frankowska [17, 18] and Ward [590, 596, 597].
In the following, we deal with abstract tangent cones in the sense of Sect. 4.5.
Furthermore, given an abstract tangent cone, the associated graphical derivative is
defined in a way analogous to the contingent derivative, that is, by taking the tangent
cone to the graph of the set-valued map.
We begin with the following general scheme for deriving an abstract chain rule:
Y Z be set-valued maps, let .x;N y/
N 2 graph .F /; and let .y; N zN/ 2 graph .G/:
Assume that P and Q are closed tangent cones such that for the maps ` W X Y
Y Z ! X Z and g W X Y Y Z ! Y; defined by
`.a; b; c; d / WD .a; d /; (11.35)

g.a; b; c; d / WD b c; (11.36)
and for S1 WD graph .F /; S2 WD graph .G/; S WD S1 S2 ; and x0 WD .x;

N y;
N y;
N zN/;
we have
`.P .S; x0 // P .`.S /; `.x0 //; (11.37)

N y//
P .S1 ; .x; N Q.S2 ; .y;
N zN// P .S1 S2 ; .x;
N y;
N y;
N zN//; (11.38)
P .S; x0 / \ rg.x0 /1 .0/ P .S \ g 1 .0/; x0 /: (11.39)
N zN/ and DQ F .y;

Then, for the graphical derivatives DP G.y; N zN/ defined via P and
Q, we have
N zN/ ı DP F .x;
DQ G.y; N y/
N DP .G ı F /.x;
N zN/: (11.40)
Proof. Following the reasoning used by Aubin and Frankowska [18] and by
Ward [590], we have
graph .DP .G ı F / .x;

N zN// D P .graph .G ı F /; .x;N zN//

D P .` Œgraph .F / graph .G/ \ g 1 .0/ ; .x;
N zN//

` P Œgraph .F / graph .G/ \ g1 .0/; .x;N y;
N y;
N zN/

` P .graph.F /graph.G/; .x; N y;
N y;
N zN//

\ rg.x;
N y; N zN/1 .0/
N y;

` P .graph .F /; .x;
N y//
N Q .graph.G/; .y;
N zN//

\ g 1 .0/

N y//
D ` graph .DP F .x; N zN// \ g 1 .0/
N graph.DQ G.y;
D graph .DQ G.y;
N zN// ı DP F .x;
N y//
N ;
and (11.40) is proved. The proof is complete. t

u
The above result conveys that to give a chain rule formula like (11.40), a
combination of cones can be used that satisfy (11.37)–(11.39). It has been shown by
Ward [590,593] that general calculus rules for epiderivative of nonsmooth functions
can be derived by the strategy depicted in the above result. We remark that, in set-
valued optimization, many cones discussed in Chap. 4 have not been employed
yet. However, based on their properties studied in Chap. 4, new derivatives can be
defined which will enjoy a rich calculus.
Remark 11.2.14. In Chap. 4, several inversion theorems have been given which
ensure (11.39). Inclusions (11.37) and (11.38) hold for a various combination of
cones. Note that (11.38) holds as an equality for the contingent cone for derivable
sets.
In the following, we collect a few results and identify the conditions where
(11.37)–(11.39) are met. We begin with the following result (see [18, Theo-
rem 5.3.3]):
Theorem 11.2.15. Let X , Y; and Z be Banach spaces where Y is finite-
dimensional, let F W X Y and G W Y Z be set-valued maps, let
N y/
.x; N 2 graph .F /; and let .y;
N zN/ 2 graph .G/: Assume that F and G are closed
and the following condition that involves the Clarke’s derivative, holds:
N y//
Im.DC F .x; N dom .DC G.y;
N zN// D Y: (11.41)
N zN/ ı DF.x;
DA G.y; N y/
N D.G ı F /.x;
N zN/; (11.42)
N zN/ ı DA F .x;
DA G.y; N y/
N DA .G ı F /.x;
N zN/; (11.43)
N zN/ ı DC F .x;
DC G.y; N y/
N DC .G ı F /.x;
N zN/: (11.44)
Proof. The proof follows the strategy of Theorem 11.2.13, and using Theo-
rem 4.7.7, which holds due to (11.41), ensures (11.37)–(11.39) for ` and g and
for T; C; and A. t
u
A useful extension, in finite dimensional spaces, is the following result by
Ward [596]:
Theorem 11.2.16. Let F W Rm Rn and G W Rn Rp be set-valued maps, let
N y/
graph .F / be locally closed at .x; N 2 graph .F /; and let graph .G/ locally closed
N zN/ 2 graph .G/: If

at .y;
Ker.D F .x; N \ D G.y;

N y// N zN/.0/ D f0g; (11.45)
where Ker.D F .x; N WD fy 2 Rn j 0 2 D F .x;

N y// N /g, then
N y/.y
N zN/ ı DA F .x;
DG.y; N y/
N D.G ı F /.x;
N zN/; (11.46)
N zN/ ı DA F .x;
DA G.y; N y/
N DA .G ı F /.x;
N zN/: (11.47)
Proof. For the maps ` W Rm Rn Rn Rp ! Rn Rp and g W Rm Rn Rn Rp !
Rn ; defined by (11.35) and (11.36), we assume that
0 2 y rg.x;
N y;
N y;
N zN/ C N.x;
N y;
N y;
N zN; S /:
Since the normal cone is product preserving, we deduce that .0; y / 2

N.graph .F /; .x; N and .y ; 0/ 2 N.graph .G/; .y;
N y// N zN//: Condition (11.45) then

implies that y D 0; which allows us to apply Theorem 4.7.10. Therefore, the
conditions (11.35) and (11.36) are satisfied and the arguments used in the proof of
Theorem 11.2.13 are applicable. t
u
We now give a general scheme for deriving an abstract sum formula.
set-valued maps, let f W X ! X be a single-valued map, let .x;
N yN1 / 2 graph .F1 /;
N yN2 / 2 graph .F2 ı f /: Assume that P and Q are closed tangent cones
and let .x;
such that for ` W .X Y /2 ! X Y and g W .X Y /2 ! X given by
`.a; b; c; d / WD .a; c C d /; (11.48)

g.a; b; c; d / WD f .a/ c; (11.49)
and for S1 WD graph .F1 /; S2 WD graph .F2 /; S WD S1 S2 ; and x0 WD

N yN1 ; f .x/;
.x; N yN2 /; we have
`.P .S; x0 // P .`.S /; `.x0 //;

N yN1 // Q.S2 ; .f .x/;
P .S1 ; .x; N yN2 // P .S1 S2 ; .x;
N yN1 ; f .x/;
N yN2 //;
P .S; x0 / \ rg.x0 /1 .0/ P .S \ g 1 .0/; x0 /:
Then the following sum formula, for the graphical derivatives given by P and
Q; holds:
N yN1 /./ C DQ F2 .f .x/;

DP F1 .x; N yN2 /.rf .x/.//
N DP .F1 C F2 ı f /.x;
N yN1 C yN2 /./:
(11.50)
Proof. Proceeding as in the proof of Theorem 11.2.13, we have
graph .DP .F1 C F2 ı f / .x;

N yN1 C yN2 //
D P .graph .F1 C F2 ı f /; .x;
N yN1 C yN2 //

D P .` Œgraph .F1 / graph .F2 / \ g 1 .0/ ; `.x;
N yN1 ; f .x/;
N yN2 //

` P .Œgraph .F1 / graph .F2 / \ g 1 .0/; .x;
N yN1 ; f .x/;
N yN2 /
` .P .graph .F1 /
\
graph .F2 /; .x;
N yN1 ; f .x/;
N yN2 // rg.x; N yN2 /1 .0/
N yN1 ; f .x/;
` .P .graph .F1 /; .x;

N yN1 //
\
Q .graph .F2 /; .f .x/;
N yN2 // rg.x; N yN2 /1 .0/
N yN1 ; f .x/;
D ` .graph .DP F1 .x;

N yN1 //
\
graph .DQ F2 .f .x/;
N yN2 // rg.x; N yN2 /1 .0/
N yN1 ; f .x/;
N yN1 //; .c; d /

D f.a; b C d /j .a; b/ 2 graph .DP F1 .x;
2 graph .DQ F2 .f .x/;
N yN2 /; rf .x/.a/
N D cg
D graph .DP F1 .x;
N yN1 / C DQ F2 .f .x/;
N yN2 /.rf .x///
N ;
and the inclusion (11.50) is established. The proof is complete. t

u
Remark 11.2.18. In Theorem 11.2.17, f should be sufficiently smooth to ensure
the smoothness of g. The above sum-chain type calculus rule for nonsmooth maps
was first devised by Ward [590].
We recall that given Banach spaces X and Y , a nonempty set S X; a map
f W X ! Y is called metrically subregular at .x;
N f .x//;
N where xN 2 S; with respect
to S if there exist real numbers k > 0 and > 0 such that for every x 2 U .x/
N \ S;
we have
d.x; f 1 .f .x/
N \ S / kkf .x/
N f .x/k:
Using the above notion, recently Durea and Strugariu [155] gave the following:
Theorem 11.2.19. Let X and Y be Banach spaces, let F1 ; F2 W X Y be
closed set-valued maps, let f W X ! X be a continuously differentiable map,
let .x; N yN2 / 2 graph .F2 ı f /: Assume that the map
g W X Y X Y ! X given by g.a; b; c; d / D f .a/ c is metrically subregular
N yN1 ; f .x/;
at .x; N yN2 ; 0/ with respect to graph .F1 / graph .F2 /: Assume that either
graph .F1 / is derivable at .x; N yN1 / or graph .F2 / is derivable at .f .x/;
N yN2 /: Then:
N yN1 /./ C DF 2 .f .x/;

DP F1 .x; N yN2 /.rf .x/.//
N D.F1 C F2 ı f /.x;
N yN1 C yN2 /./:
11.3 Contingently C -Absorbing Maps 437
Proof. The proof follows the scheme of Theorem 11.2.17 and employes Theo-
rem 4.7.6. t
u
We conclude this section by giving a chain rule for convex set-valued maps acting
between finite-dimensional spaces. In this case, it is natural to give corresponding
results for the associated coderivative. Recall that for a convex set-valued map F W
X Y; the coderivative D F .x;N y/
N W Y X is a set-valued map given by
D F .x; N / D fx j .x ; y / 2 N.graph .F /; .x;

N y/.y N y//g:
N
The following result is taken from [459]:

Theorem 11.2.20. Let X , Y; and Z be finite-dimensional spaces, let F W X Y
and G W Y Z be convex set-valued maps, let .x;N zN/ 2 graph .G ı F /; and let
N \ G 1 .Nz/: Then,
yN 2 F .x/
D.G ı F /.x;
N zN/ cl .DG.y;
N zN/ ı DF.x;
N y//:
N
D .G ı F /.x;
N zN/ D G.y;
N zN/ ı D F .x;
N y/;
N
and if ri range.F / \ ri dom.F / ¤ ;; then equalities hold.

Proof. The proof follows by the arguments used in Theorem 11.2.13 and makes use
of the intersection formula (4.31): t
u
Numerous particular cases of the above result can be found in [459]. Stronger
conditions can be imposed for obtaining analogues of the formulas given in the
above result for convex maps acting between general Banach spaces. Details can be
found in [16].
11.3 Contingently C -Absorbing Maps
As we have noted earlier in this chapter, under some nice properties of the graph
of a set-valued map, its contingent derivative enjoys useful properties. For instance,
when a set-valued map has a convex graph, then, under mild additional assumptions,
the contingent derivative of such a map is defined on the whole space. However,
the convexity of the graph is a very stringent assumption whereas the convexity of
its epigraph is a relatively milder assumption. Therefore, the question arises as to
whether the contingent derivative of a set-valued map can provide some insight in
looking at the contingent derivative of the profile map of the underlying set-valued
map. It turns out that the following notion plays an important role in studying this
relationship:
Definition 11.3.1. Let X and Y be normed spaces and let C Y be a suitable
cone. A set-valued F W X Y is called contingently C -absorbing at .x;
N y/
N 2
graph .F /; if the following identity holds:
N y/.x/
DF.x; N C C D D.F C C /.x;
N y/.x/
N for every x 2 dom .D.F C C /.x;
N y//:
N
(11.51)
The above notion plays an important role in several issues such as the existence
of contingent epiderivatives, generalized epiderivatives, and sensitivity analysis,
among others. In the following, we will give examples to show that (11.51) is in
general false. We will also give several conditions on the map F and on the cone C
which ensure that (11.51) remains valid.
Before any advancement, we recall the following notion proposed by Shi [530].
map and let .x; N 2 graph .F /. A set-valued map DS F W X Y is called the
N y/
N y/
S-derivative of F at .x; N if there are sequences ftn g P and f.xn ; yn /g X Y
with .xn ; yn / ! .x; y/ and tn xn ! 0 such that yN C tn yn 2 F .xN C tn xn / for every
n 2 N:
The following result, whose proof is adopted from [214, Theorem 3.6.1] will be
used shortly.
N y/
map, and let .x; N 2 graph .F /. Let C Y be a pointed, closed and convex cone.
Then
N y/.0/
DS F .x; N \ .C / D f0g; (11.52)
holds, provided that any of the following conditions is verified:

(i) yN 2 Bo-PMin .F .x/;
N C /, F is upper Lipschitz at x,
N that is, there exist L; > 0
such that
F .x/ F .x/
N C L kx xk
N BY for all x 2 B.x;
N /I
(ii) yN 2 Bo-PMin .F .x/I

N C /, X , Y are finite-dimensional, xN 2 icr.dom .F //, and
F is C -convex;
(iii) yN 2 GHe-PMin .F .x/I
N C /, X , Y are Banach spaces, xN 2 int.dom .F //, and
F is C -convex and C -closed.
Proof. (i) Let y 2 DS F .x;
N y/.0/\.C
N /: Then there exist ftn g P, f.xn ; yn /g
graph .F / such that xn ! 0 and tn ..xn ; yn / .x;N y//
N ! .0; y/. We may
assume that xn 2 B.x; N /; for every n 2 N. It follows that for every n there
exist yNn 2 F .x/
N and vn 2 BY such that yn D yNn C L kxn xk N vn . Hence
ktn .yNn y/
N tn .yn y/k
N Ltn kxn xk
N ;
for every n 2 N, whence tn .yNn y/

N ! y. Therefore y 2 cl cone.F .x/
N y/
N \
.C / D f0g, which is a contradiction.
(ii) Taking .x;N y/
N D .0; 0/ (replacing F by F .x; N y/)
N and replacing X
by span.dom .F //; if necessary, we may suppose that xN 2 int.dom .F //.
Since dim Y < 1 and yN 2 Bo-PMin .F .x/I N C /, we have that yN 2

N C /. Consequently, there exists a convex cone K such that
GHe-PMin .F .x/;
C n f0g intK ¤ Y and F .x/
N \ .yN K/ yN C K. Since
graph .F C K/ D graph .F / C f0g C C f0g K;
the set graph .F C K/ is convex. We have that .x; N y/

N … icr.graph .F C K//.
In the contrary case, for c 2 C n f0g; there exists some
> 0 such that
N y/
.x; N yN
c/ 2 graph .F / C f0g K; i.e., yN
c k 0 2 F .x/
N
.0; c/ D .x; N
for some k 2 K. This contradicts the hypothesis since
cCk 0 2 C nf0gCK
0
intK. Using a separation theorem, there exists .u ; y / 2 X Y n f.0; 0/g

such that
N u i C hy;
hx; N y i hx; u i C hy C k; y i 8 .x; y/ 2 graph .F /; 8 k 2 K:
(11.53)
If y D 0, then hx;
N u i hx; u i for every x 2 dom .F /, and so u D 0.
Therefore y ¤ 0. From (11.53) we obtain hk; y i
0 for k 2 K, whence

hk; y i > 0 for every k 2 intK C n f0g. Now assume that there exists
y 2 DS F .x;N y/.0/\.C
N /nf0g. Then there are ftn g .0; 1/ and f.xn ; yn /g
graph .F / such that xn ! 0 and tn ..xn ; yn / .x;
N y//
N ! .0; y/. From (11.53)
we obtain that
N u i C htn .yn y/;

htn .xn x/; N y i
0;
for every n, whence hy; y i 0, a contradiction.

(iii) Let K be a convex cone such that C nf0g intK ¤ Y and yN 2 Min.F .x/;
N K/.
If we show that int.graph .F C K/ ¤ ;, arguments like those in the proof of
(ii) show that (11.52) holds. For this aim, consider the relation
R WD f.y; x/ j .x; y/ 2 graph .F C K/g:
Clearly, R is closed, convex, and 0 2 int.ImR/. Applying the

Robinson–Ursescu theorem, we have that R.V / is a neighborhood of xN
for every neighborhood V of y. N Let k0 2 intK. There exists > 0 such
that B.k0 ; / K. It follows that for some ı > 0 we have B.x; N ı/
R .B.y;N =2//. Let us show that B.x;N ı/ B.k0 C y;N =2/ graph .F C K/.
Indeed, let .u; y/ belong to the first set. There exists y 0 2 R 1 .u/ such that
ky 0 yk
N =2. It follows that y D y 0 C .y y 0 / and

k0 .y y 0 / k.k0 C yN y/k C y 0 yN =2 C =2 D ;
which shows that y y 0 2 B.k0 ; / K. Therefore .x; y/ 2 R 1 Cf0gK D

graph .F C K/. The proof is complete. t
u
The following example, taken from [564], shows that the condition that F is
upper Lipschitz at xN in Theorem 11.3.3 (i) cannot be dropped.
Example 11.3.4. Let X D Y D R, C D RC , and let F W R R be defined by

f0g if x 0;
F .x/ WD p
f x; 0g if x > 0:
Then,

Œ0; 1Œ if x 0;
.F C C /.x/ D p
Œ x; 1Œ if x > 0:
Taking xN D 0 and yN D 0, we have

f0g if x ¤ 0;
N y/.x/
DS F .x; N D DF.x;
N y/.x/
N D
C if x D 0;

C if x < 0;
D.F C C /.x;
N y/.x/
N D
R if x
0;
showing that (11.52) is violated.

We remark that the condition yN 2 GHe-PMin .F .x/; N C / in the above theorem
cannot be replaced by yN 2 Min .F .x/I
N C /. In the next example F is upper Lipschitz
at xN 2 int.dom .F //, C -convex, and C -closed.
Example 11.3.5. Let X D R, Y D R2 , C D R2C , and F W X Y be defined by
˚
F .x/ WD .y1 ; y2 / j y2
y12 :
We take xN D 0 and yN D 0. Note that yN 2 Min.F .x/;

N C / n BMin.F .x/;
N C /. Then
N y/.x/
DS F .x; N D DF.x;
N y/.x/
N D R RC ;
and consequently,
N y/.x/
DS F .x; N \ .C / D 1; 0 f0g ¤ f0g:
The following example shows that Theorem 11.3.3 (ii) does not hold when the
space X is infinite dimensional.
Example 11.3.6. Consider X an infinite-dimensional normed vector space, Y WD
R; and W X ! R a noncontinuous linear functional. Let F W X Y be defined
by F .x/ WD f.x/g and C WD RC . Then it is easy to verify that
N y/.x/
DS F .x; N D DF.x;
N y/.x/
N D R;
N y/
for all .x; N 2 graph .F / and all x 2 X:
The following result gives conditions for set-valued maps to be contingently

C -absorbing. The original proofs of the second and the third parts were given by
Tanino [564, Proposition 2.1, Theorem 2.1] and Shi [530, Proposition 3.1], however,
our presentation is taken from [214, Theorem 3.6.6].
N y/
map and let .x; N 2 graph .F /. Let C Y be a pointed, closed and convex cone.
Then:
(i) For every x 2 X; we have
N y/.x/
DF.x; N C C D.F C C /.x;
N y/.x/:
N
(ii) If the set fy 2 C j kyk D 1g is compact, then for every u 2 X ,
Min .D.F C C /.x;

N y/.x/;
N C / Min .DF.x;
N y/.x/I
N C / DF.x;
N y/.x/:
N
(11.54)
(iii) If the set fy 2 C j kyk D 1g is compact and (11.52) holds, then F is

contingently C -absorbing, that is,
N y/.x/
DF.x; N C C D D.F C C /.x;
N y/.x/
N for every x 2 X: (11.55)
Proof. (i) Let y 2 DF.x; N y/.x/

N and c 2 C . Then there exist ftn g ! 0C and
f.xn ; yn /g ! .x; y/ such that .x;N y/Ct
N n .xn ; yn / 2 graph .F / for every n 2 N.
N y/
It follows that .x; N C tn .xn ; yn C c/ 2 graph .F C C / and .xn ; yn C c/ !
.x; y C c/. Therefore y C c 2 D.F C C /.0; 0/.x/.
In the rest of the proof we suppose that Q WD fy 2 C j kyk D 1g is
compact. Of course, C D coneQ.
(ii) Let y 2 Min .D.F C C /.x; N y/.x/;
N C / : Then there exists ftn g ! 0C , f
n g
Œ0; 1/, fqn g Q, f.xn ; yn /g X Y such that .xn ; yn C
n qn / ! .x; y/
and .x;N y/
N C tn .xn ; yn / 2 graph .F / for every n 2 N. Since Q is compact, we
may suppose that qn ! q 2 Q. Assume that there are a subsequence f
nk g
and > 0 such that
nk
for every k 2 N. Then .x; N y/
N C tnk .xnk ; ynk C
.
nk /qnk / 2 graph .F C C / and .xnk ; ynk C .
nk /qnk / ! .u; y
q/.
It follows that y q 2 D.F C C /.x; N y/.x/
N \ .y C / n fyg, a contradiction.
Therefore
n ! 0, and so .xn ; yn / ! .u; y/. It follows that y 2 DF.x; N y/.x/.
N
Using also (i), we have that for every x 2 X ,
Min .D.F C C /.x;

N y/.x/I
N C / DF.x;
N y/.x/
N D.F C C /.x;
N y/.x/:
N
From the above inclusions we obtain immediately that (11.54) holds.

(iii) The inclusion being proven in (i), let us prove the converse one. Of course,
we assume that (11.52) holds. Let y 2 D.F C C /.x; N y/.x/;
N there exist ftn g !
0C , f
n g Œ0; 1/, fqn g Q, f.xn ; yn /g X Y such that .xn ; yn C
n qn / !
N y/
.x; y/ and .x; N C tn .xn ; yn / 2 graph .F / for every n 2 N. We may suppose
that qn ! q 2 Q. Taking also a subsequence if necessary, we may assume that

n !
2 Œ0; 1. Suppose that
D 1. Since yn C
n qn ! y, we obtain that

1
n yn ! q. Of course, x
N C tn xn ! xN and

1 un yn
..xN C tn xn ; yN C tn yn / .x;
N y//
N D ; ! .0; q/:

n tn
n
n
It follows that q 2 DS F .x;

N y/.0/
N \ .C / n f0g, a contradiction. Therefore

< 1, whence .xn ; yn / ! .u; y
q/. Hence y 2 DF.x; N y/.x/
N C C . The
u
The following example, originally constructed in [214, Example 3.6.8], shows
that the hypothesis “fy 2 C j kyk D 1g is compact” in the above theorem is
essential for having (11.54) and (11.55).
Example 11.3.8. Let X D R, Y D `2 , C D `2C D f.xn /n2N 2 `2 j xn
0 8 n 2
Ng, and F W X Y be given by
(
f0g if x 2 1; 0 [ Œ1; 1Œ;
F .x/ WD 1
n Œa e n ; a if x 2 Œ nC1
1
; n1 Œ; n 2 N ;
where a 2 `2C n f0g and .en /n2N is the canonical base of `2 . Then
(
C if x 2 1; 0 [ Œ1; 1Œ;
.F C C /.x/ W D
n1 .a C en / C C if x 2 Œ nC1
1
; n1 Œ; n 2 N :
Take xN D 0 and yN D 0. Then

f0g if u 0;
N y/.u/
DF.x; N D
fuag if u > 0;

C if u 0;
D.F C C /.x;
N y/.u/
N D
ua C C if u > 0:
So (11.54) and (11.55) do not hold for u > 0.

Let us prove the formulae for DF.x;N y/.u/
N and D.F C C /.x; N y/.u/
N for u
0,
those for u < 0 being obvious. So, let u
0 and v 2 DF.x; N y/.u/.
N Then there
exist ftn g 0; 1Œ and graph F .xn ; yn / ! .0; 0/ such that .tn xn / ! u and
.tn yn / ! v. If P WD fn 2 N j xn 0g is infinite, then yn D 0 for every n 2 P ,
and so .u; v/ D .0; 0/. Suppose that P is finite (this is the case when u > 0). Since
.xn / ! 0, there exists n0 2 N such that xn 2 0; 1Œ for n
n0 . For every n
n0
there exists pn 2 N such that xn 2 Œ pn1C1 ; p1n Œ; of course, .pn / ! 1. Then for

every n
n0 there exists
n 2 Œ0; 1 such that yn D p1n .1 2
n /a
n en . It

follows that ptnn ! u and ptnn ..1 2
n /a
n epn / ! v. If u D 0, it is obvious
that v D 0. Suppose that u > 0. Taking eventually a subsequence, we may assume
that .
n / !
2 Œ0; 1. We obtain that .
n epn / ! .1 2
/a u1 v. If
¤ 0, we
get the contradiction that .en / contains a norm-convergent subsequence. Therefore

.
n / ! 0, whence v D ua. It is easy to get that ua 2 DF.x; N y/.u/
N for u
0 (in fact,
similar to the proof in the next paragraph). So the formula for DF.x; N y/.u/
N holds.
Consider first u > 0 and v D ua C k with k 2 C . Taking xn D nC1 1
, yn D
1 1
1
.a C u k/ D 1
.a e C u k C e / (2 .F C C /.x /), and t D .n C 1/u >0
n

n n n
n
n
for n 2 N , we have that .xn ; yn / ! .0; 0/ and tn .xn ; yn / ! .u; v/; if u D 0 and
v D k 2 k just take xn D 0, yn D n1 k, and tn D n. Therefore v 2 DFC .x; N y/.u/.
N
Let u
0 and v 2 D.F C C /. N
x; N
y/.u/. Then
there exist .t n / 0; 1Œ and
graph .F C C / .xn ; yn / ! .0; 0/ such that tn .xn ; yn / ! .u; v/. As in the first
part we may suppose that xn 2 0; 1Œ for n
n0 . So there exists N .pn / ! 1
with xn 2 Œ pn1C1 ; p1n Œ for n
n0 . Of course, yn D p1n .a C epn / C kn with kn 2 C .

It follows that ptnn ! u and ptnn .a C epn / C tn kn ! v. Since 0 D w-lim en ,
it follows that v C ua D w-lim.tn kn /. Because C is weakly closed, v C ua 2 C ,
whence v 2 ua C C .
In Example 11.3.4, (11.52) is not satisfied and F .x/ is not convex for x > 0
and (11.55) do not hold for u > 0.
For various purposes, a number of compactness assumptions have been intro-
duced in optimization. In the following, we recall two notions of compactness and
employ them to show that a given map set-valued map is contingently C -absorbing.
We recall that, given normed spaces X and Y , a set-valued map F W X Y
is called directionally compact at .x; N y/N 2 graph .F / in a direction x 2 X; if
for every sequence ftn g P with tn # 0 and every sequence fxn g X with
xn ! x; any sequence fyn g Y contains a convergent subsequence provided that
yN C tn yn 2 F .xN C tn xn /; for every n 2 N:
We have the following result which was originally given by Bednarczuk and
Song [39].
Theorem 11.3.9. Let X and Y be normed spaces and let C Y be a pointed,
closed, and convex cone. Let F W X Y be a set-valued map and let .x; N y/
N 2
N y/
graph .F /: Assume that F is directionally compact at .x; N in any direction x 2 X:
N y/;
Then, the map F is contingently C -absorbing at .x; N that is, (11.90) holds.
Proof. In view of (11.88), it suffices to show that
D.F C C /.x;
N y/.x/
N DF.x;
N y/.x/
N C C:
Let y 2 D.F C C /.x;N y/.x/:

N Then there exist sequences ftn g P; f.xn ; yn /g
X Y; and fcn g C such that tn # 0; .xn ; yn / ! .x; y/; and
yN C tn .yn cn =tn / D yN C tn yn cn 2 F .xN C tn xn /:
Since the map F is directionally compact, we may assume that yn cn =tn

converges to some yQ 2 Y: Hence yQ 2 DF.x;
N y/.x/
N and cn =tn ! .y y/
Q 2 C;
confirming that
y 2 DF.x;
N y/.x/
N C C;

u
The following result which was originally given by Taa [548, Proposition 2.2], is
obtained by using the concept of compactly approximable maps.
Theorem 11.3.10. Let X and Y be normed spaces, and let C Y be a pointed,
N 2
graph .F /: If the map F is compactly approximable at .x; N y/;
N then the map F is
N y/;
contingently C -absorbing at .x; N that is (11.90) holds.
Proof. For any x 2 dom.D.F CC /.x; N y//;
N let y 2 D.F CC /.x; N y/.x/
N be arbitrary.
Hence there exist sequences ftn g P and f.xn ; yn /g X Y such that tn #
0; .xn ; yn / ! .x; y/ and
yN C tn yn 2 F .xN C tn xn / C C:
Then,
yN C tn yn D zn C cn with zn 2 F .xN C tn xn / and cn 2 C:
Since F is compactly approximable, there exists fkn g R.x/; rn WD

r.tn ; xn /; bn 2 BY and n0 2 N such that rn ! 0; and
zn D yN C tn .kn C rn bn / for all n > n0 ;
which implies
yN C tn .kn C rn bn / 2 F .xN C tn xn /;
for all n > n0 : In view of the compactness of R.x/; we may assume that kn ! k 2
R.x/. Because of the fact that .kn C rn bn / ! k; we deduce that k 2 DF.x;
N y/.x/:
N
Since
yN C tn yn D yN C tn .kn C rn bn / C cn ;
for sufficiently large n; we have yn kn rn bn D tn1 cn 2 C: By passing to the

limit, we have y k 2 C and hence
y 2 k C C DF.x;
N y/.x/
N C C;

u
11.4 Epiderivatives of Set-Valued Maps 445
11.4 Epiderivatives of Set-Valued Maps
In this section, we introduce the notion of the contingent epiderivative and some
related epiderivatives which play a fundamental role in set-valued optimization. The
main motivation behind considering the epiderivatives is the fact that the epigraph of
a set-valued map has considerably better structure than the graph of the set-valued
map.
We begin with formulating the following notion of the contingent epiderivative
which was originally proposed by Jahn and Rauh [301].
Definition 11.4.1. Let X and Y be normed spaces and let C Y be a pointed,
closed, and convex cone. Let F W X Y be a set-valued map and let .x;N y/
N 2
graph .F /: A single-valued map DE F .x;N y/
N W X ! Y is called the contingent
N y/;
epiderivative of F at .x; N if the following identity holds
N y//
epi .DE F .x; N D T .epi .F /; .x;
N y//;
N
N y//
where T .epi .F /; .x; N is the contingent cone of epi .F / at .x;
N y/:
N The map F
N y/
is called contingently epidifferentiable, if the effective domain of DE F .x; N is
given by
L WD dom .D.F C C /.x;

N y//
N D fx 2 X j .x; y/ 2 T .epi .F /; .x;
N y//g:
N
A geometric interpretation of the contingent epiderivative is given in Fig. 11.2.

The epigraph of the single-valued contingent epiderivative is obtained by locally
approximating the epigraph of the underlying set-valued map by means of the
contingent cone.
In the following, for a single-valued map f W X ! Y; we denote the contingent
epiderivative of f at .x;N f .x//
N 2 graph .f / by DE f .x/:
N That is, DE f .x/ N WD
N f .x//:
DE f .x; N
Fig. 11.2 Contingent y
epiderivative of the set-valued

map F
epi (F )
graph (F )
x
T (epi(F ), (x̄, ȳ))
Remark 11.4.2. If .x;N y/

N belongs to the interior of epi .F /, then T .epi .F /; .x;
N y//
N
coincides with the product space X Y; and, in this case, the contingent epiderivative
N y/
DE F .x; N does not exist. It is clear that the domain of the contingent epiderivative
does not need to be the whole space X:
Although our focus will primarily be on the contingent epiderivative, the
following definition collects some of the related concepts which are based on the
same idea but take a different tangent cone for the local approximation.
closed, and convex cone. Let F W X Y be a set-valued map and let .x;
N y/
N 2
graph .F /:
1. A single-valued map DO E F .x;
N y/
N W X ! Y is called the adjacent epiderivative
N y/;
of F at .x; N if the following identity holds
epi .DO E F .x;

N y//
N D A.epi .F /; .x;
N y//;
N
N y//
where A.epi .F /; .x; N is the adjacent cone of epi .F / at .x;
N y/:
N The map F
is called adjacently epidifferentiable, if the effective domain of DO E F .x;
N y/
N is
given by
LO WD dom .DA .F C C /.x;

N y//
N D fx 2 X j .x; y/ 2 A.epi .F /; .x;
N y//g:
N
2. A single-valued map DQ E F .x;

N y/
N W X ! Y is called the Clarke epiderivative of
N y/;
F at .x; N if the following identity holds
epi .DQ E F .x;

N y//
N D C.epi .F /; .x;
N y//;
N
N y//
where C.epi .F /; .x; N is the Clarke cone of epi .F / at .x;
N y/:
N The map F is
called Clarke epidifferentiable, if the effective domain of DQ E F .x;
N y/
N is given
by
LQ WD dom .D.F
Q C C /.x;
N y//
N D fx 2 X j .x; y/ 2 C.epi .F /; .x;
N y//g:
N
11.4.1 Contingent Epiderivatives of Set-Valued Maps

with Images in R
To illustrate the underlying idea, we investigate the contingent epiderivative of set-

valued maps defined from a normed space to the real line and give some simple
illustrative examples. Contingent epiderivatives of set-valued maps acting between
general spaces will be studied in a later part of this section. Our exposition of this
subsection is based on [298].
Let X be a normed space and let F W X R be a given set-valued map. To the

map F , we associate another set-valued map FC W X R defined by
FC .x/ WD cl .F .x/ C RC /:
We define a single-valued map f W X ! R by
f .x/ D fy 2 FC .x/j y z; for every z 2 FC .x/g: (11.56)
We begin with the following intuitive result for the existence of contingent
epiderivatives.
Theorem 11.4.4. Let X be a normed space and let F W X R be a
set-valued map. Let the single-valued function f defined by (11.56) be convex and
lower-semicontinuous. Let xN 2 int.dom .F // and .x;
N f .x//
N 2 graph .F /. We set
yN D f .x/:
N Then the contingent epiderivative DE F .x;
N y/
N of F at .x;
N y/
N exists and
is given by
f .xN C
x/ f .x/
N
N y/.x/
DE F .x; N D lim ; for every x 2 X:

#0
Proof. The proof is based on straightforward arguments. It is evident from (11.56)

that the epigraph of FC is characterized by f: By using the lower-semicontinuity of
f , we obtain
epi .f / D cl .epi .f // D cl .epi .F //:
Due to the properties of the contingent cone, we have
N y//
T .epi .f /; .x; N D T .cl .epi .F //; .x;
N y//
N D T .epi .F /; .x;
N y//:
N
Since the function f is convex and lower-semicontinuous and xN 2 int.dom .f //,

N y/.x/
it follows that the contingent epiderivative DE f .x; N of the single-valued
functional f exists for all x 2 X (see [16, pp.198–199]). This amounts to say
that
N y/.x/
Df .x; N D inffyj .x; y/ 2 T ..f /; .x;
N y//g
N
is finite for all x 2 X: Since the contingent cone is closed, the above identity implies
that
N y//
epi .DE f .x; N D T .epi .f /; .x;
N y//:
N
N y/,
By virtue of the definition of the contingent epiderivative DE F .x; N we have
N y//
N y//
N D T .epi .f /; .x;
N y//
N D epi .DE f .x;
N y//;
N
and hence the existence of DE F .x; N y/.x/

N follows from the existence of
N y/.x/:
DE f .x; N The given characterization then follows from [16, pp.198–199].
The proof is complete. u
t
Next we slightly improve the above result by relaxing the assumption of
convexity to local convexity. To achieve this goal, the following result will be useful:
Theorem 11.4.5. Let X be a normed space, let F W X R be a set-valued map,
N y/
and let .x; N 2 graph .F /: Assume that there are functionals f1 ; f2 W dom .f1 / D
dom .f2 / D X ! R such that
epi .f1 / T .epi .F /; .x;

N y//
N epi .f2 /:
N y/
Then the contingent epiderivative DE F .x; N is given by:
N y/.x/
DE F .x; N D minfy 2 Rj .x; y/ 2 T .epi .F /; .x;
N y//g;
N for every x 2 X:
N y//
It is clear that the above theorem assures that dom .DE F .x; N D X . In fact,
it is a consequence of the assumptions that T .epi .F /; .x; N y//
N epi .f2 / and
dom .f2 / D X . Here the domain of F is of less relevance. The following simple
N y/
example shows that if dom .f2 / is a proper subset of X , then DE F .x; N might not
be well defined in some directions.
Example 11.4.6. Let F W R R be a set-valued map defined as follows (see
Fig. 11.3):
( ip h
jxj; 1 if x 6D 0;
F .x/ D
Œ0; 1Œ if x D 0:
N y/
We choose .x; N D .0; 0/. Then,
Fig. 11.3 A set-valued map y

and its contingent
epiderivative epi(F)
T (epi(F ), (x̄, ȳ))
x
˚
N y//
T .epi .F /; .x; N D .0; y/ 2 R2 j y
0 :
Now let us choose a function p1 W R ! R defined by p1 .x/ D jxj; for every

x 2 R; and another function p2 W f0g ! R defined by p2 .0/ D 0: It is easy to check
N y/
that if the contingent epiderivative of pi at .x; N is denoted by fi ; where i D 1; 2,
then the required condition
epi .f1 / T .epi .F /; .x;

N y//
N epi .f2 /;
holds. However, the contingent epiderivative DE F .x; N y/;

N whose effective domain
consists of the singleton f0g; is given by DE F .x;
N y/.0/
N D 0:
Let Z be a normed space and let ! 2 ˝ Z be arbitrary. We reserve
the notation U.˝; !/ to represent the set V .!/ \ ˝, where V .!/ is an arbitrary
neighborhood of !.
Our next result for the existence of contingent epiderivative is as follows:
Theorem 11.4.7. Let X be a normed space and let F W X R be a set-valued
map. Assume that xN 2 int.dom .F // and .x; N f .x//
N 2 graph .F /, where the
functional f is defined via (11.56). We set yN D f .x/: N Assume that there exist
functionals f1 ; f2 W X ! R; satisfying the following conditions:
N y/
1. .x; N 2 graph .f1 / \ graph .f2 / and xN 2 int.dom .f1 // \ int.dom .f2 //.
N y//
2. U.epi .f1 /; .x; N U.epi .f /; .x;
N y//
N U.epi .f2 /; .x; N y//:
N
3. The functionals f1 and f2 are convex and lower-semicontinuous.
N y/.x/
Then the contingent epiderivative DE F .x; N N y/
of F at .x; N exists for every
x 2 X.
Proof. In view of the first two assumptions and the isotony property of the
contingent cone, we have
N y//;
T .U.epi .f1 /; .x; N .x;
N y//
N T .U.epi .f /; .x;
N y//;
N .x;
N y//
N
T .U.epi .f2 /; .x;
N y//;
N .x;
N y//:
N
Moreover, from the local approximation nature of the contingent cone, it follows
that
N y//;
T .U.epi .f1 /; .x; N .x;
N y//
N D T .epi .f1 /; .x;
N y//;
N
N y//;
T .U.epi .f2 /; .x; N .x;
N y//
N D T .epi .f2 /; .x;
N y//:
N
By combining this with the preceding chain of inclusions, we obtain
N y//
T .epi .f1 /; .x; N T .epi .f /; .x;
N y//
N T .epi .f2 /; .x;
N y//:
N (11.57)
Since fi is convex and lower-semicontinuous, we have the existence of function

pi W X ! R; i D 1; 2; such that
epi .pi / D T .epi .fi /; .x;

N y//;
N i D 1; 2:
The above identity, when combined with (11.57)), yields
epi .p1 / T .epi .f /; .x;

N y//
N epi .p2 /:
N y/.x/;
The existence of DE F .x; N for every x 2 X; now follows from Theo-
rem 11.4.5. t
u
Remark 11.4.8. In the above result, if we use the so-called lower-semicontinuous
regularization (see Rockafellar and Wets [499]) of fi to generate pi ; then we can
dispense with the assumption that fi is lower-semicontinuous. Of course, it is also
possible to impose other conditions on fi which assure the existence of pi :
The following example highlights the use of the above result.
Example 11.4.9. Consider a set-valued map F W Œ1; 1 R given by (cf.
Fig. 11.4 (i)):
8
ˆ
8 x1 if 13 < x 1;
21 5
ˆ
ˆ 8
ˆ
ˆ if 0 x 13 ;
< 0; 4
F .x/ D 2x if 13 < x < 0;
ˆ
ˆ
ˆ
ˆ
3
x7 if 23 < x 13 ;
:̂ 2 x 16
2 2; 0 if 1 x 23 :
In the present case, the single-valued selection function f (cf. 11.56) of F is

given by:
8
ˆ
ˆ
21
x 58 if 13 < x 1;
ˆ
ˆ 8
ˆ
<0 if 0 x 13 ;
f .x/ D 2x if 13 < x < 0;
ˆ
ˆ
ˆ
ˆ2 x 6
3 7
if 23 < x 13 ;
:̂ x 1 if 1 x 23 :
2 2
The map f is discontinuous and dom .f / D Œ1; 1 (see Fig. 11.4 (ii)). Let
us consider the point .1=3; 2=3/ 2 graph .f /: We define two functions f1 ; f2 W
Œ1; 1 ! R as by
3 2 1
f1 .x/ D x Cx ; for every x 2 Œ1; 1;
2 2
and
A(1,2)
1.5
0.5
B(1/3,1/4)
G
0
D C
F(−2/3,−1/6)
−0.5
E(−1/3,−2/3)
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Fig. 11.4 (i) The set-valued map F. (ii) Functions fi ; i D 0; 1; 2
11
C 14
4 x if 13 < x 1;
f2 .x/ D
2x 43 if 1 x 13 :
The functions f1 and f2 are convex and satisfy the following conditions (see
Fig. 11.4 (ii)):
I. U.epi .f1 /; .1=3; 2=3// U.epi .f /; .1=3; 2=3// U.epi .f2 /;
.1=3; 2=3//:
II. .1=3; 2=3/ 2 graph .f1 / \ graph .f2 / and 1=3 2 int.dom .f1 // \
int.dom .f2 //.
Therefore, all the hypotheses of Theorem 11.4.7 are satisfied and hence the
existence of the contingent epiderivative is assured at . 13 ; 23 /. Note that the
epigraph of the set-valued map f is not convex, but only convex around . 13 ; 23 /.
Note that Theorem 11.4.7 does not provide any information for the existence of
the contingent epiderivative at . 23 ; 16 / as there does not exist any convex map
satisfying the hypotheses on the function f1 . Though the contingent epiderivative
exists at . 23 ; 16 /.
In order to illustrate the importance of the assumption xN 2 int.dom .f1 // \
int.dom .f2 //; let us consider the point . 13 ; 0/ 2 graph .f /. Since
T .epi .F /; .1=3; 0// D f.x; y/ 2 R2 j 0

x; y
0g;
the contingent epiderivative of F at . 13 ; 0/ with restricted domain is given by
DE F .1=3; 0/.x/ D 0 for every x 0;
and consequently,
dom .DE F .1=3; 0// D fx 2 Rj x 0g:

Remark 11.4.10. In the above discussion, our objective was to show the relevance
of the single-valued map marginal function (11.56) and, for clarity, we have imposed
the convexity assumption. Clearly, all the results can be modified to dispense with
the convexity, and analogous results can be given by assuming that the marginal
functional is differentiable in the Hadamard sense.
11.4.2 Contingent Epiderivatives in General Spaces
Having studied contingent epiderivatives of set-valued maps with images in R; we

now move to investigate this notion for maps acting between general spaces. For
the treatment of this subject in the remaining part of this section, besides relying on
the contribution by Jahn and Rauh [301], we use extensively the results based on a
thorough investigation of the subject by Rodríguez-Marín and Sama [500, 502] and
Sama [520]. See also [210, 246, 501, 503, 504, 519].
We begin with the following result that gives some insight on the existence of
contingent epiderivatives.
Theorem 11.4.11. Let X and Y be normed spaces and let C Y be a convex
cone. Let F W X Y be a set-valued map and let .x; N y/
N 2 graph .F /: If the
N y/
contingent epiderivative DE F .x; N W X Y of F at .x;
N y/
N exists, then for every
x 2 L WD dom .D.F C C /.x; N y//;
N we have
N y/.x/
DE F .x; N 2 SMin.D.F C C /.x;
N y/.x/;
N C /;
where SMin.D.F C C /.x; N y/.x/;

N C / is the set of all strongly minimal points
of D.F C C /.x;
N y/.x/
N with respect to the cone C .
Proof. By the definition of the contingent epiderivative, we have
N y//
N y//;
N
which implies that for every x 2 .D.F C C /.x;

N y//,
N the following identity holds
N y/.x/
DE F .x; N C C D D.F C C /.x;
N y/.x/;
N
N y/.x/
and consequently DE F .x; N 2 SMin .D.F C C /.x;
N y/.x/;
N C / : The proof is
complete. u
t
Remark 11.4.12. In the above result, we did not explicitly imposed the condition
that the cone C is closed. However, the existence of the contingent epiderivative
ensures that C is closed. In fact, if the contingent derivative exists, then we have
N y/.x/
DE F .x; N C C D D.F C C /.x;
N y/.x/
N for any x 2 dom .D.F C C /.x;
N y//;
N
which ensures that C D D.F C C /.x; N y/.x/

N DE F .x;
N y/.x/:
N Because the set
D.F C C /.x; N y/.x/
N is closed, we deduce the closedness of C from the fact that it
is a difference of a closed and a compact set.
The following result addresses the uniqueness issue for the contingent epideriva-
tive.
closed, and convex cone. Let F W X Y be a set-valued map and let
N y/
.x; N 2 graph .F /: If the set-valued map F is contingent epidifferentiable at
N y/;
.x; N then the contingent epiderivative DE F .x;
N y/
N is unique.
Proof. Suppose that DM E F .x;
N y/
N and DE F .x;
N y/
N are two contingent epiderivatives
such that
DM E F .x;
N y/.x/
N ¤ DE F .x;
N y/.x/
N for any x 2 dom .D.F C C /.x;
N y//:
N
Then, for such x 2 dom .D.F C C /.x;

N y//;
N we have
DM E F .x;
N y/.x/
N 2 SMin.D.F C C /.x;
N y/.x/;
N C /;
N y/.x/
DE F .x; N 2 SMin.D.F C C /.x;
N y/.x/;
N C /;
implying that
DM E F .x;
N y/.x/
N 2 DE F .x;
N y/.x/
N C C;
N y/.x/
DE F .x; N 2 DM E F .x;
N y/.x/
N C C;
and hence
DM E F .x;
N y/.x/
N DE F .x;
N y/.x/
N 2 C \ C;
which, by taking into account the pointedness of C; ensures that DM E F .x;

N y/.x/
N D
N y/.x/.
DE F .x; N t
u
The fact that the pointedness of the ordering cone C plays a vital role in the
uniqueness of the contingent epiderivative was singled out in [500]. It turns out that
if the ordering cone C is non-pointed, then there exists a family of maps such that
each member of the family satisfies the definition of the contingent epiderivative.
The following example illustrates this situation:
Example 11.4.14. We define a set-valued map F W R R2 by
F .x/ D f.0; 0/; .jxj ; j xj/g; for every x 2 R:
N y/
Then, for C D R RC and .x; N D .0; .0; 0//; we have
N y//
T .epi .F /; .x; N D R R RC
which means that for every x 2 dom .D.F C C /.x;

N y//
N D R; we have
D.F C C /.x;
N y/.x/
N D R RC :
Clearly, a map ' W R ! R2 is a contingent epiderivative of F at .x;

N y/;
N if for any
x 2 R, we have
'.x/ C R RC D D.F C C /.x;

N y/.x/:
N
Evidently, the maps
'1 .x/ D .x; 0/;

'2 .x/ D .x; 0/
are both contingent epiderivatives of F at .x;N y/;

N although '1 ¤ '2 : In fact,
there exists a family of epiderivatives given by D f' W R ! R2 j '.x/ D
.h.x/; 0/ where h W R ! R is any mapg:
We have the following useful property of the contingent epiderivative (see [301]).
N 2
graph .F /: Assume that the map F is C -convex. Then the contingent epiderivative
N y/
DE F .x; N is a sublinear map, that is, for every
2 RC and all x; y 2 dom .D.F C
N y//;
C /.x; N we have
N y/.
x/
DE F .x; N D
DE F .x;
N y/.x//;
N
N y/.x
DE F .x; N C y/ DE F .x;
N y/.x/
N C DE F .x;
N y/.y/:
N
t
Proof. The proof can be extracted from the proof of Theorem 11.4.16 given below.u
Note that if the ordering cone C is not pointed, then not necessarily every
member of the family of contingent epiderivatives is sublinear. In fact, this is evident
from Example 11.4.14. However, if epi .F / is locally convex around .x; N y/;
N then
there exists a sublinear map in the family of contingent epiderivatives of F at .x;N y/;
N
which in the following, is denoted by EF .x; N y/:
N This is shown in the following
result.
Theorem 11.4.16. Let X and Y be normed spaces and let C Y be a closed,
and convex cone. Let F W X Y be a set-valued map and let .x;
N y/
N 2 graph .F /:
N y/
Assume that EF .x; N ¤ ;: Then:
N y/:
1. If the map F is C -convex, then there exists a sublinear map ˚ 2 EF .x; N
N convex, then yN is a strong minimal point of F .x/:
2. If F .x/ N
N y/
Proof. By the proof Theorem 11.4.11, it follows that EF .x; N ¤ ; implies
SMin..D.F C C /.x;
N y/.x/;
N C/ ¤ ; for any x 2 L D dom .D.F C C /.x;
N y//:
N
Consider the set B D fe 2 Lj kek D 1g. For each e 2 B, we assign a fixed

element ye 2 SMin.D.F C C /.x;
N y/.e//.
N For the unique ˛ > 0 such that x D ˛e,
we define ˚ W L ! Y by
˚.0/ D 0
˚.e/ D ye
˚.x/ D ˛ye :
N y/.
We claim that ˚ 2 EF .x; N In fact, if x ¤ 0, then
˚.x/ D ˛ye 2 SMin.˛D.F C C /.x;

N y/.e/;
N C/
D SMin.D.F C C /.x;
N y/.˛e/;
N C / D SMin.D.F C C /.x;
N y/.x/;
N C /:
If x D 0, then we claim that
˚.0/ D 0 2 SMin.D.F C C /.x;

N y/.0//:
N
Indeed, if the above containment does not hold, then
0 … SMin.D.F C C /.x;
N y/.0/;
N C /: (11.58)
Therefore, by Theorem 11.4.11, there exists w ¤ 0 with w 2 SMin.D.F C

N y/.0/;
C /.x; N C /. Since 0 2 D.F C C /.x;
N y/.0/,
N by the minimality of w, we have
w 2 C . Furthermore, by (11.58), w … C . Since the contingent derivative is strictly
positive homogeneous,

w 2 D.F C C /.x;
N y/.0/
N for any
> 0:
This, in view of the minimality of w, confirms that w

w. Therefore w 2
C \ C; but this contradicts w … C:
By applying Theorem 11.4.11, we conclude ˚ 2 EF .x;N y/:
N It is easily seen that
˚ is positive homogeneous because ˚.0/ D 0 and for each x ¤ 0, we have
˚.
x/ D ˚.
.˛e// D ˚..
˛/e/ D
˛ye D
˚.x/:
N y//
We not proceed to show that ˚ is subadditive. Since T .epi .F /; .x; N is convex
and because ˚ 2 EF .x; N y/,
N we deduce that the cone epi .˚/ D T .epi .F /; .x; N y//
N
is convex as well. Therefore,
1
2
x C 12 y; 12 ˚.x/ C 12 ˚.y/ 2 epi .˚/; for any x; y 2 L;
implying that
1
1
2 ˚.x/ C 12 ˚.y/ 2 D.F C C /.x;
N y/
N 2x C 12 y ˚ 12 x C 12 y C C;
which, when multiplied by 2 on both sides, yields
˚.x/ C ˚.y/ ˚ .x C y/ C C:
For the second part, we suppose that the result is false. Then, there exists an
element y 2 F .x/N such that y yN … C: Since F .x/
N is a convex set, the sequence
y yN
fyn g; where yn WD yN C N and satisfies that
; belongs to F .x/
n
yn yN y yN
D ! v:
kyn yk
N ky yk
N
Obviously, v … C and v 2 D.F C C /.x; N y/.0/.

N From the first part, we note that
N y/
there is an element ' 2 EF .x; N such that '.0/ D 0. By the definition of ', we
have v
0 D '.0/ which contradicts that v … C and the proof is complete. t
u
As noted above, the notion of the contingent epiderivative is connected to
SMin.D.F C C /.x;
N y/;
N C /: The following result connects SMin.DF.x;N y/./;
N C/
and SMin.D.F C C /.x;
N y/./;
N C /:
Theorem 11.4.17. Assume that F is a contingently C -absorbing map at .x; N y/,
N
that is, (11.51) holds. Then, for every x 2 dom .D.F C C /.x; N y//;
N the set
SMin.DF.x; N y/.x/;
N N y/.x/
C / is nonempty, if and only if, DE F .x; N exists, and, in
that case, the two sets coincide.
Proof. Let any x 2 dom .D.F C C /.x;
N y//;
N be arbitrary. Using (11.51), we have
SMin .D.F C C /.x;

N y/.x/;
N C / D SMin .DF.x;
N y/.x/
N C C/;C/
D SMin.DF.x;
N y/.x/;
N C /:
N y/
By the proof of Theorem 11.4.11, if DE F .x; N exists then SMin .DF.x;
N y/.x/C
N
C / ; C / ¤ ;: The conclusion then follows from the above identity. The converse
implication is obvious. t
u
Clearly, the above result is a consequence of (11.51). Several examples have been
given in [500, 502] which show that the above result does not hold without (11.51),
and, in general, SMin.DF.x; N y/.x/;
N N y/.x/:
C / is different from DE F .x; N
11.4.3 Existence Theorems for Contingent Epiderivatives
Our objective in this section is to give some existence results for contingent
epiderivatives. Throughout this section, we assume that the ordering cone C is
strongly minihedral. That is, every subset which is C -bounded from below has a
supremum.
The following result gives some insight into the existence of contingent epi-
derivatives.
N 2
graph .F /: Assume that for every x 2 L WD dom .D.F C C /.x; N y//
N ; the set
D.F C C /.x;N y/.x/
N is C -lower bounded. Then for the map M W L ! Y defined by
M.x/ D SMax.fy 2 Y j D.F C C /.x;

N y/.x/
N y C C g; C /; (11.59)
the following properties hold:

1. M.
x/ D
M.x/; for any x 2 L and

0:
N y/
2. If DE F .x; N exists, then it coincides with M:
N y/
3. DE F .x; N exists, if and only if,
M.x/ 2 D.F C C /.x;

N y/.x/
N for every x 2 L:
Proof. We begin by proving the first part. By the strong minihedrality of the cone,
the map M is well-defined. To prove the first part, we fix x 2 L and
> 0. Since
D.F CC /.x;
N y/.x/
N is C -lower bounded and the map D.F CC /.x; N y/./
N is positive
homogenous, we have
D.F C C /.x;
N y/.
x/
N D
D.F C C /.x;
N y/.x/
N

M.x/;
and consequently, by the definition of M , we get
M.
x/

M.x/:
Analogously, because

D.F C C /.x;
N y/.x/
N D D.F C C /.x;
N y/.
x/
N
M.
x/;
we have
1
D.F C C /.x;
N y/.x/
N
M.
x/; (11.60)

implying
1
M.x/
M.
x/;

and it follows that

M.x/ D M.
x/.
If x D 0; then, using (11.60), we have
M.0/

M.0/ for every
2 RC ;
and hence M.0/ D 0:

The second part follows from the fact that for a C -lower bounded subset of X , it
holds that
SMin.A; C / D SMax.B; C /;
where B D fy 2 Y j A y C C g; provided that SMin.A; C / ¤ ;:

We now proceed to prove the third part. Let x 2 L be arbitrary. If DE F .x;
N y/
N
exists, then by Theorem 11.4.11, we have SMin.D.F C C /.x; y/.x/; C / ¤ ;.
Therefore, by Theorem 11.4.11 and the preceding arguments, we deduce that
M.x/ D SMin.D.F C C /.x; y/.x/; C /.
Conversely, if M.x/ 2 D.F C C /.x; y/.x/; then by D.F C C /.x; y/.x/
M.x/ C C , we have
M.x/ D SMin.D.F C C /.x; y/.x/; C /;
N y/
and hence Theorem 11.4.11 ensures that DE F .x; N exists. The proof is complete.u
t
In the above result, the assumption that D.F C C /.x;
N y/.x/
N remains C -lower
bounded for every x 2 L has played an important role. The following result gives a
necessary criteria to ensure this condition.
closed, and convex cone with int.C / ¤ ;: Let F W X Y be a set-valued map and
N y/
let .x; N 2 graph .F /: If F is upper locally Lipschitz at xN and yN 2 SMin.F .x/
N C
C; C /; then the set D.F C C /.x;N y/.x/
N is C -lower bounded for every x 2 L:
Proof. Let.x; y/ 2 T .epi .F /; .x; N y//
N be arbitrary. Then it follows from the
definition of the contingent cone that there exist sequences ftn g P; f.xn ; yn C
cn /g epi .F /; such that .xn ; yn C cn / ! .x;
N y/
N and tn .xn x;N yn C cn y/N !
N there exists a sequence fzn g F .x/
.x; y/: Since F is upper locally Lipschitz at x; N
such that
yn zn 2 l kxn xk
N BY for every n 2 N;
implying that yn zn ! 0; and therefore .zn C cn y/N ! 0. Since int.C / ¤ ;;

there exists a C -lower bound of BY , z 2 Y such that
tn .yn zn /
tn l kxn xk
N z:
Since yN is a local strongly minimal point of .F .x/

N C C /; we can assume without
loss of generality that .zn C cn y/
N 2 C , and, consequently
tn .yn C cn y/
N D tn .yn zn / C tn .zn C cn y/
N
l ktn .xn x/k
N z;
which, when passed to the limit as n ! 1; confirms that y

l kuk z. t
u
For a preparation for our next result, we need the following auxiliary result:
Proposition 11.4.20. Let X and Y be normed space and let C Y be a pointed,
closed, and convex cone. Let S X be nonempty, let F W S Y be a set-valued
N 2 graph .F /. Assume D.F C C /.0; y/.x/
map, and let .0; y/ N is C -lower bounded
for every x in its domain. If S L D dom .M / and M , defined in (11.59), is
continuous, then
D.F C M /.0; y/
N D M C DF.0; y/:
N
Proof. For any .x; y/ 2 graph .D.F C M /.0; y//; N we have .x; y/ 2 T .graph
.F C M /; .0; y//.
N Then there exist sequences ftn g P; f.xn ; yn /g X Y with
tn ! 0; .xn ; yn / ! .x; y/ such that tn xn 2 S and
yN C tn yn 2 F .tn xn / C M.tn yn /;
implying
yN C tn .yn M.xn // 2 F .tn xn /:
Since M is continuous, we have yn M.xn / ! y M.x/ and hence .x; y

M.x// 2 graph .DF.0; y//:N
For the converse, let .x; y/ 2 graph ..M C DF.0; y// N and y M.x/ 2
N
DF.0; y/.x/, then there exist sequences tn # 0; xn ! x; wn ! y M.x/ with
tn xn 2 S and yN C tn wn 2 F .tn xn / or equivalently,
yN C tn .wn C M.xn // 2 F .tn un / C M.tn xn /:
Since yn C M.xn / ! y, we obtain .x; y/ 2 graph .D.F C M /.0; y//:

N The proof
is complete. t
u
We have the following result regarding the existence of the contingent epideriva-
tive.
closed, and convex cone. Let S X be nonempty, let F W S Y be a set-valued
map, and let .x;N y/
N 2 graph .F /. Assume D.F C C /.x;
N y/.x/
N is C -lower bounded
for every x 2 L D dom .D.F C C /.x; N y//
N ; the map M; defined in (11.59), is
continuous and S xN C dom .M /. Then the contingent epiderivative DE F .x;
N y/
N
exists, if and only if,
0 2 D.F . C x/
N M C C /.0; y/.x/
N for any x 2 L:
Proof. Let x 2 L and H.x/ D F .x C x/; N then .0; y/

N 2 graph .H /: It is easy to
check that D.H C C /.0; y/N coincides with D.F C C /.x; N y/:
N By Theorem 11.4.18,
N y/
a necessary and sufficient condition for the existence of DE F .x; N is
M.x/ 2 D.H C C /.0; y/.x/

N for every x 2 L: (11.61)
Since dom .H C C / D S fxg

N dom .M / and M is continuous, by
Proposition 11.4.20, we get
D.H C C M /.0; y/.x/

N D D.F C C /.x;
N y/.x/
N M.x/ for any u 2 S fxg
N :
Therefore (11.61) is equivalent to
0 2 D.F . C x/
N M C C /.0; y/.x/;
N

u
In the following, we give another characterization for the existence of
N y/:
DE F .x; N
closed, and convex cone. Let S X , let xN 2 int.S /; let F W S Y be a set-valued
map, and let yN 2 SMin.F .x/;
N C /. Assume that L WD dom .D.F C C /.x; N y//
N DX
and there exists a neighborhood U.0/ with xN C U.0/ S satisfying that, for every
x 2 U.0/; the following conditions hold:
1. D.F C C /.x;
N y/.x/
N F .x C x/
N yN C C .
2. SMin.F .x C x/;
N C / ¤ ;.
N y/
Then, DE F .x; N exists provided that the function M is continuous.
Proof. We begin by claiming that D.F CC /.x; N y/.x/
N is C -lower bounded for every
x 2 L: In fact, by the imposed conditions, the set D.F C C /.x;N y/.0/
N is C -lower
bounded and D.F C C /.x; N y/.0/
N F .x/
N yN C C C .
For any 0 ¤ x 2 L, there exists ˛ 2 RC nf0g such that ˛ 1 x 2 U.0/. Therefore,
by the imposed conditions, we have
D.F C C /.x; N 1 x/ SMin.F .˛ 1 x C x/;

N y/.˛ N C / yN C C:
By the positive homogeneity of the contingent derivative, we deduce
D.F C C /.x;
N y/.x/
N D ˛D.F C C /.x; N 1 x/ ˛.SMin.F .˛ 1 x C x/;
N y/.˛ N C / y/
N C C;
ensuring that the element ˛.SMin.F .˛ 1 x C x/;

N C / y/
N is a C -lower bound of
D.F C C /.x;N y/.x/.
N
For x 2 dom .D.F . C x/

N M C C /.0; y//;
N there exist fxn g dom .F . C x/
N
M C C / \ U.0/; ftn g P such that xn ! 0 and tn xn ! x.
Using the imposed conditions once again, we have D.F C C /.x;N y/.x
N n / zn
yN C C; with zn D SMin.F .xn C x/;
N C /, and hence
M.xn /
zn y:
N
By choosing cn D .zn M.xn / y/

N 2 C; we get tn .xn ; zn M.xn / yN Ccn / !
.x; 0/ which confirms that
0 2 D.F . C x/
N M C C /.0; y/.x/;
N
and the proof finishes by Theorem 11.4.21. t

u
In the remaining part of this section, we focus on the set-valued maps taking
values in Rn :
We recall that, given a normed space X , if a map f W X ! R is upper locally
Lipschitz map at xN 2 X , then the contingent epiderivative DE f .x/
N exists and is
given by
N
DE f .x/.x/ D inffy 2 Rj y 2 Df .x;
N f .x//.x/g:
N
Based on this observation, we have the following:

Lemma 11.4.23. Let X be a normed space, let S X be nonempty, and let f W
X ! Rn be a given map. Assume that f D .f1 ; f2 ; : : : fn / is upper locally Lipschitz
at xN 2 S: For any i 2 f1; ::; ng and for given y 2 Dfi .x; N fi .x//.x/
N with x 2 L;
there exists .wj /j ¤i 2 Rn1 such that
.w1 ; w2 ; ::; y; : : : wn / 2 Df .x;

N f .x//.x/:
N
Proof. For x 2 L; let y 2 Dfi .x; N fi .x//.x/

N be arbitrary. Then there exist sequences
ftm g RC and fxm g S such that xm ! x, N tm .fi .xm / fi .x//
N ! y. By the
fact that the component functions ffi gi 2I are Lipschitz, it is possible to extract
subsequences ftmk g ftm g and fxmk g fxm g such that there exists .wj /j ¤i 2
Rn1 which verifies that tmk .fi .xmk / fi .x// N ! wj for any j 2 f1; ::; ngnfi g:
Consequently, .w1 ; w2 ; ::; y; : : : wn / 2 Df .x;
N f .x//.x/.
N t
u
In the following, by Mf we denote the map M associated with f (see (11.59)).
Theorem 11.4.24. Let X be a normed space, let S X be nonempty, and let Rn
be partially ordered by C D RnC : If f D .f1 ; f2 ; : : : fn / is upper locally Lipschitz
at xN 2 S , then
Mf D .DE f1 .x/;
N DE f2 .x/;
N : : : ; DE fn .x//:
N
Proof. By Theorem 11.4.19, we have
Mf .x/ D SMax .y 2 Y W D.f C C /.x;

N y/.x/
N y C C; C / ;
and since the upper locally Lipschitz maps are contingently C -absorbing, it is
equivalent to
Mf .x/ D SMax .fy 2 Y W D.f /.x;

N y/.x/
N y C Cg;C/:
Since
N
DE fi .x/.x/ D inffyj y 2 Dfi .x;
N y/.x/g
N for any i D 1; 2; ::; n;
we have
N y/.x/
Df .x; N N DE f2 .x/;
.DE f1 .x/; N ::; DE fn .x//
N C C;
which, by the definition of Mf , yields that
Mf .x/
.DE f1 .x/;
N DE f2 .x/;
N ::; DE fn .x//:
N
Suppose that Mf .x/ ¤ .DE f1 .x/;

N DE f2 .x/;
N ::; DE fn .x//:
N Therefore, there
exists an index i such that
N
Mfi .x/ > DE fi .x/.x/:
N
Since DE fi .x/.x/ 2 Dfi .x;
N y/.x/;
N it follows by Lemma 11.4.23 that there exists
.wj /j ¤i 2 Rn1 such that
N
.w1 ; w2 ; : : : ; DE fi .x/.x/; : : : ; wn / 2 Df .x;
N y/.x/
N Mf .x/ C C;
N
which contradicts Mfi .x/ > DE fi .x/.x/: The proof is complete. t
u
We intend to close this discussion by giving an existence result for the contingent
epiderivative for convex set-valued maps. For this we will first need the following
result.
Lemma 11.4.25. Let X be a normed space, let S X be nonempty, and let Rn
be partially ordered by C WD RnC : If f W S ! Rn is C -convex and continuous at
xN 2 int.S /, then the contingent epiderivative DE f .x;
N f .x//
N exists and is given by
N f .x//
DE f .x; N D .DE f1 .x/;
N DE f2 .x/;
N : : : ; DE fn .x//:
N
Proof. Because f is C -convex, each component fi is convex and locally lipschitz

at xN 2 int.S /. Therefore, it follows that (see [18]),
fi .xN C tx/ fi .x/

N
N
DE fi .x/.x/ D lim ; for any x 2 X; i 2 f1; ::; ng:
t !0C t
A direct computation shows that
N
.x; DE f1 .x/.x/; N
::; DE fn .x/.x// 2 T .epi .f /; .x;
N f .x///:
N
According to Theorem 11.4.24, we have
Mf .x/ D .DE f1 .x/.x/;

N N
::; DE fn .x/.x//;
asserting
.x; Mf .x// 2 T .epi .f /; .x;

N f .x///;
N
and, as a consequence, we get
Mf .x/ 2 D.f C C /.x;

N f .x//.x/;
N
N f .x//
and the existence of the contingent epiderivative Df .x; N D Mf follows from
Theorem 11.4.18. The proof is complete. t
u
We have the following existence result for the contingent epiderivative of set-
valued map convex maps.
Theorem 11.4.26. Let X be a normed space, let S X be convex, let xN 2 int.S /;
and let Rn be partially ordered by C WD RnC : Let F W S ! 2R be a C -
n
convex set-valued map such that SMin.F .x/; C / ¤ ;; for any x 2 S . If h.x/ D
N then the contingent epiderivatives DE h.x/
SMin.F .x/; C / is continuous at x; N and
N h.x//
DE F .x; N exists. Furthermore, we have
N D DE h.x; h.x//
DE F .x; h.x// N D .DE h1 .x/;
N DE h2 .x/;
N : : : ; DE hn .x//:
N
N h.x//
Proof. Since epi .h/ D epi .F /, the existence of DE F .x; N is equivalent to the
N h.x//:
existence of DE h.x; N Since F is C -convex, for any x1; x2 2 S and
2 Œ0; 1 ;
we have
h.
x1 C .1
/x2 / 2 F .
x1 C .1
/x2 /
F .x1 / C .1
/F .x2 / C C;
and by using the condition h.x/ D SMin.F .x/; C /, we deduce that
h.
x1 C .1
/x2 /
h.x1 / C .1
/h.x2 /
which confirms that the map h is C -convex. It follows from Lemma 11.4.25 that the
N exists and is given by
contingent epiderivative DE h.x/
N D .DE h1 .x/;
Dh.x; h.x// N DE h2 .x/;
N ::; DE hn .x//;
N

u
11.4.4 Variational Characterization of the Contingent

Epiderivatives
In this section, our objective is to give a variational characterization for the

contingent epiderivative by associating it to a variational system. This is done in
two ways. In the first approach, the variational system is obtained by scalarizing
the contingent derivative of the involved set-valued map, whereas in the second
approach, we employ the contingent epiderivative of the scalarization of the set-
valued map. Our exposition of this subsection is based on the investigations by
Rodríguez-Marín and Sama [502] and Sama [520].
We begin with the first approach by considering the following family of
variational systems: For a given x 2 L WD dom .D.F C C /.x; N y//
N ; find y 2 Y
such that
`.y/ D inff`.y/j y 2 D.F C C /.x;

N y/.x/g
N for any ` 2 C C : (11.62)
The following is the promised variational characterization of the contingent

epiderivative.
Theorem 11.4.27. Let X and Y be normed space and let C Y be a pointed,
N 2
graph .F /: Assume that for any x 2 L; there exists a weakly compact subset Bx
N such that D.F C C /.x; N y/.x/
N Bx C C . Then the contingent
N y/
epiderivative DE F .x; N exists, if and only if, for each x 2 L; the system (11.62) is
N y/
solvable. Moreover, if DE F .x; N exists, then DE F .x; N y/.x/
N is the unique solution
of (11.62) for x 2 L:
N y/
Proof. We recall that the contingent epiderivative DE F .x; N exists, if and only if,
for every x 2 L, we have SMin.D.F C C /.x; N y/.x/;
N C / ¤ ;: Furthermore, if
N y/
DE F .x; N exists, then we have
N y/.x/
DE F .x; N D SMin.D.F C C /.x;
N y/.x/;
N C /:
We will finish the proof in the following three steps:

1. For any weakly compact set A Y; if SMin.A; C / is nonempty, then it solves
the variational system: Find y 2 Y such that
`.y/ D inf `.a/; for any ` 2 C C : (11.63)

a2A
2. If the variational system (11.63) is solvable, then the solution is unique and is
given by SMin.A; C /.
3. If B Y be such that A B, B A C C , then SMin.B; C / ¤ ; provided
that (11.63) is solvable.
Assuming that A is weakly compact, we will first show that if (11.63) has solution
then it coincides with SMin.A; C /. Suppose that y0 2 Y is such that
`.y0 / D inf `.a/; for every ` 2 C C :

a2A
We define the set
S` WD fa 2 Aj `.a/ `.a/
Q for every aQ 2 Ag;
which, due to the weak-compactness of A, is nonempty and weakly compact. We

claim that
\`2C C S` ¤ ;:
In view of the weak-compactness of A; it suffices to show that any finite

subfamily .Sì /i 21;n of .S` /`2C C has a nonempty intersection. Note that,
S`1 C:::C`n \i 21;n Sì ; for all `1 ; : : : ; `n 2 C C :
In fact, by setting ` WD `1 C : : : C `n and taking a 2 S` ; we have
ì .y0 / ì .a/; for every i 2 1; n;
and by summing these inequalities term by term , we get
`.y0 / `.a/ D `.y0 /;
and consequently
ì .y0 / D ì .a/ D min ì .A/; for every i 2 1; n:
It follows that a 2 \i 21;n Sì which confirms the claim that \`2C C S` ¤ ;: For
aN 2 \`2C C S` ¤ ;; we have
N
0; for every ` 2 C C ; for every a 2 A;
`.a a/
confirming that a aN 2 C CC D C; for every a 2 A: Therefore, aN is the strong

minimum point of A: The last two parts follow immediately. t
u
We now proceed to give a variational characterization by taking the contingent
epiderivative of the scalarization of the involved set-valued map. For the sake of
a clear presentation, we only consider the case when Y D Rn and C D RnC : A
treatment of this issue in general spaces can be found in the useful contribution by
Sama [520].
Given a normed space X and a set-valued map F W X Rn ; we define the
following variational system associated with the scalar set-valued maps ` ı F for
` 2 C C : Find y 2 Y such that
`.y/ D D.` ı F /.x;

N `.y//.x/
N for any ` 2 C C : (11.64)
It follows that, for contingently C -absorbing maps, we have

Lemma 11.4.28. Let ` 2 C C , let SMin.F .x/;
N C / D fyg;
N and let F be stable at
N y/:
.x; N Then:
inff`.y/j y 2 D.F C C /.x;

N y/.x/g
N D inff`.y/j y 2 DF.x;
N y/.x/g
N for any
x 2 L WD dom .D.F C C /.x;
N y//:
N
We begin with exploring some properties of set-valued maps which get trans-
ferred to their scalarized analogues.
Lemma 11.4.29. Given a normed space X; let F W X Y , where Y D Rn ; be a
set-valued map. Let ` 2 Y . Assume that F .x/
N D fyg
N and that F is stable at .x;
N y/:
N
Then the following assertions hold:
1. N y/;
If F is stable at .x; N then ` ı F is stable at .x;
N `.y//.
N
2. N D dom .DF.x;
T .dom .F /; x/ N y//
N D dom .D.` ı F /.x; N `.y///:
N
3. ` ı DF.x;N y/.x//
N D D.` ı F /.x;
N `.y//.x/;
N for any x 2 T .dom .F /; x/.
N
4. For every x 2 T .dom .F /; x/;
N the set DF.x; N y/.x/
N is compact.
5. For any x 2 T .dom .F /; x/;
N we have
inff`.y/j y 2 DF.x;
N y/.x/g
N D inffyj y 2 D.` ı F /.x;
N `.y//.x/g
N 2 R:
6. For any x 2 L; there exists a (weakly) compact subset Bx D.F CC /.x;

N y/.x/
N
such that
D.F C C /.x;
N y/.x/
N Bx C C:
N y/;
Proof. We begin with a proof of the first claim. Since the map F is stable at .x; N
N and a constant c > 0 such that
there exists a neighborhood U.x/
y yN 2 B.0; c kx xk/
N for any y 2 F .x/; x 2 U nfxg:
N
The containment ` 2 Y ensures that there exists a constant k > 0 such that
j`.y y/j
N k kx xk
N ;
implying that
j`.y y/j
N ck kx xk
N ;
which completes the proof of the first part.

For the second part, we set dom .F / D S and note that the inclusion
dom .DF.x; N y//
N T .S; x/ N is always true. Therefore, it suffices to show that
N dom .DF.x;
T .S; x/ N y//.
N Let x 2 T .S; x/N be arbitrary. Then there are sequences
fxn g S , ftn g P such that tn ! 1; xn ! xN and tn .xn x/
N ! x. Since the map
F is stable at .x;N y/
N and F .x/N D fyg;
N there exists fyn g Y with yn 2 F .xn / such
that yn ! yN and for sufficiently large n 2 N; we have
tn .yn y/
N 2 B.0; ctn kxn xk/:
N (11.65)
Since Y is finite-dimensional, we can assume that there exists y 2 Y such

that tn .yn y/
N ! y. Consequently, .x; y/ 2 T .graph .F /; .x;N y//
N implying that
y 2 DF.x; N y/.x/
N and x 2 dom .DF.x;N y//.
N This proves that T .dom .F /; x/ N D
dom .DF.x; N y//:
N For the second identity, it suffices to note that from the first part,
the map ` ı F is stable at .x;
N `.y//:
N
We now proceed to prove the third part. We claim that for any x 2 T .S; x/,
N we
have
D.` ı F /.x;
N `.y//.x/
N `.DF.x;
N y/.x//:
N
Let y 2 D.` ı F /.x; N `.y//.y/

N be arbitrary. Then there exist ftn g P with
tn ! 1, f.xn ; yn /g graph .F / such that .xn ; `.yn // ! .x; N `.y//
N and tn .`.yn /
N ! y. Since the map F is stable at .x;
`.y// N y/N and SMin.F .x/;N C / D fyg,
N we have
yn ! yN and there exists z 2 Y such that a subsequence of ftn .yn y/g N converges
to z. This implies that z 2 DF.x; N y/.x/.
N It is clear that y D `.z/. For the converse,
let y 2 DF.x;N y/.x/
N be arbitrary. By the linearity and continuity of `; we can show
that `.y/ 2 D.` ı F /.x; N `.y//.x/:
N
The fourth part follows from the observation that, using (11.65), we get y 2
B.0; c kxk/ which means that DF.x; N y/.x/
N B.0; c kuk/.
For the fourth part, we begin by choosing x 2 T .S; x/ N arbitrarily. Since ` is
continuous, we deduce that the set of reals `.DF.x; N y/.x//
N is compact and therefore
inff`.y/j y 2 DF.x; N y/.x/g
N is finite. Then, from the third part, we have
inff`.y/j y 2 DF.x;
N y/.x/g
N D inffyj y 2 D.` ı F /.x;
N `.y//.x/g:
N
Finally, we proceed to prove the last part. For x 2 T .S; x/;

N we deduce from
Lemma 11.4.28 that
D.F C C /.x;
N y/.u/
N D DF.x;
N y/.x/
N C C;
and hence, by taking Bx D DF.x;

N y/.x/;
N we have that Bx D.F C C /.x;
N y/.x/
N
and D.F C C /.x;N y/.x/
N Bx C C . The proof is complete. t
u
We remark that the above results have been extended to Banach spaces by
Sama [520]).
The following is the variational characterization by using scalarization of the
involved map.
Theorem 11.4.30. Given a normed space X; let F W X Y D Rn be a set-valued
map. Let ` 2 Y 0 . Assume that F .x/N D fyg N and the map F is stable at .x; N y/:
N
Then DE F .x;N y/
N exists if and only if each system (11.64) has a solution for every
x 2 T .dom .F /; x/.
N Furthermore if it exists, DE F .x;
N y/
N is given by the unique
solution of (11.64) for any x 2 T .dom .F /; x/.
N
Proof. Let ` 2 C C . we set S D dom .F / and define a set-valued map FO W X Y
by setting FO .x/N D fyg N and FO .x/ D F .x/ for any x 2 S nfxg. N Clearly, L D
O
dom .D.F C C /.x; N y//
N D T .S; x/:N In fact, the inclusion dom .D FO .x;
N y//
N
dom .D.FO C C /.x; N y//
N T .S; x/,
N is always true and it follows from the above
lemma that L D T .S; x/. N
It is also evident that .F C C / D .FO C C / and .` ı F C RC / D .` ı FO C RC /:
Therefore, D.` ı FO C RC /.x; N `.y//
N coincides with DE .` ı F C RC /.x;N `.y//
N and
D.FO C C /.x; N y/
N coincides with D.F C C /.x; N y/:
N
Note that it suffices to prove that DE .` ı FO /.x;
N `.y//
N exists and
DE .` ı FO /.x;
N `.y//./
N D inff`.v/j v 2 D.FO C C /.x;
N y/./g;
N (11.66)
which would imply that DE .` ı F /.x;

N `.y//
N exists and
DE .` ı F /.x;
N `.y//./
N D inff`.y/j y 2 D.F C C /.x;
N y/./g:
N
As a consequence the family of variational systems (11.64) is equivalent

to (11.63).
From the last part of Lemma 11.4.29, the map FO verifies the compactness con-
dition of Theorem 11.4.27 and hence the existence of the contingent epiderivative
ensues.
Let us now prove (11.66). For x 2 L; it follows from Lemma 11.4.29 that the
map ` ı FO is stable at .x;
N `.y//
N , and hence by the fifth part of Lemma 11.4.29 and
by Lemma 11.4.28, we deduce that
inffyj 2 D.` ı FO /.x;

N `.y//.x/g
N D inff`.y/j D FO .x;
N y/.x/g
N
D inff`.y/j y 2 D.FO C C /.x;
N y/.x/g:
N
From the previous equalities, we conclude that DE .` ı FO /.x;

N `.x//
N exists and
verifies
DE .` ı FO /.x;
N `.x//.x/
N D inff`.y/j y 2 D.FO C C /.x;
N y/.u/g;
N for any x 2 L;

u
The following example illustrates the above approach for a single-valued map.
Example 11.4.31. Let X D R, Y D R2 , and
C D f.x; y/ 2 R2 j x y
0; x
0g:
We set xN D 0 and define f W R ! R2 by

8
ˆ
ˆ 1
< .e sin x; e sin x/ if x 2
x x
; n2N
f .x/ D n
ˆ 2 2 1
:̂ .x ; x / if x 62 ; n2N :
n
The map f is stable at xN D 0, and the dual cone C C is characterized by
C C D f.˛ ˇ; ˛/j ˛; ˇ 2 RC g:
For ` D .˛ ˇ; ˛/ 2 C C ; we have

8
ˆ
ˆ 1
< ˇe x sin x if x 2 ; n2N
` ı f .x/ D n
ˆ 1
:̂ ˇx 2
if x 62 ; n2N ;
n
and hence

ˇx if x > 0,
DE .` ı f /.0; `.0//.x/ D
0 if x 0.
N f .x//.x/
Therefore, for any x > 0; if the contingent epiderivative Df .x; N exists,
then it is the solution of the following variational system: Find .x; y/ 2 R2 such
that
.˛ ˇ/x ˛y D ˇx for any ˛; ˇ 2 RC :
It can easily be verified that .x; y/ D .u; u/ is a solution.

Analogously, for any x < 0; if the contingent epiderivative Df .x; N f .x//.x/
N
exists, then it is the solution of the following variational system: Find .x1 ; x2 / 2 R2
such that
.˛ ˇ/x1 ˛x2 D 0; for any ˛; ˇ 2 RC :
It is easily seen that .x; y/ D .0; 0/ is a solution. Therefore, Df .x;

N f .x//
N exists
and is given by

.x; x/ if x > 0,
N f .x//.x/
Df .x; N D
.0; 0/ if x 0.
11.5 Generalized Contingent Epiderivatives of Set-Valued

Maps
The existence theory for the contingent epiderivative of set-valued maps acting
between general spaces requires strong conditions on the data. This motivates the
introduction of the generalized contingent epiderivatives for which the existence
theorems can be given under more general conditions. This section is devoted to
the generalized contingent epiderivative and two other notions of epiderivatives
which are very closely related to it. Besides some simple examples, we give
existence theorems and characterizations of the generalized contingent epiderivative
and its variants. Our treatment of the subject is based on [297], however, we
have also benefited from the contributions by Chen and Jahn [92] and Bednarczuk
and Song [39], where the notion of the generalized contingent epiderivative was
originally introduced.
Definition 11.5.1. Let X and Y be normed spaces, let C Y be a pointed, closed
N y/
N 2 graph .F /.
Define a set-valued map G W X Y by
G.x/ WD D.F C C /.x;

N y/.x/
N for every x 2 L WD dom .D.F C C /.x;
N y//:
N
(11.67)
N y/
1. The generalized contingent epiderivative of F at .x; N is the set-valued map
N y/
DG F .x; N W X Y defined by
N y/.x/
DG F .x; N D Min.G.x/; C /:
N y/
2. The generalized proper contingent epiderivative of F at .x; N is the set-
N y/
valued map DP F .x; N W X Y defined by
N y/.x/
DP F .x; N D GHe-PMin.G.x/; C /;
where GHe-PMin.G.x/; C / represents the set of all Henig global properly

minimal points.
N y/
3. The generalized weak contingent epiderivative of F at .x; N is the set-valued
N y/
map DW F .x; N W X Y defined by
N y/.x/
DW F .x; N D WMin.G.x/; C /;
provided that the ordering cone C is solid.

In view of the relationship among GHe-PMin.; /, Min.; / and WMin.; /, the
following inclusions are immediate:
N y/./
DP F .x; N DG F .x;
N y/./
N DW F .x;
N y/./:
N (11.68)
11.5 Generalized Contingent Epiderivatives of Set-Valued Maps 471
Fig. 11.5 (i) A view of

epi .F /. (ii) A contradiction –30
to .1; 0/ 2
GHe-PMin.G.1/; R2C /
40 –20
20 –10
10 5 0
20 15
30 25 –20
10
20
30
40
In order to show that the above inclusion is strict in general and to illustrate the
above notions, we give some simple examples. We begin with an example showing
N y/./
that DP F .x; N N y/./.
is a proper subset of DG F .x; N
Example 11.5.2. Consider a set-valued map F W R R2 defined as follows:

; if x < 0;
F .x/ WD ˚
.y1 ; y2 / 2 R2 j y12 C y22 x 2 if x
0:
Let xN D 0 and let yN D .0; 0/. We note that (see Fig. 11.5 (i))
N y//
T .epi .F /; .x; N D epi .F /;
and hence the map G W RC R2 defined in (11.67) is given by:
G.x/ D f.y1 ; y2 / 2 R2 j .y1 ; y2 / 2 F .x/ C R2C g:
N y/
Consequently, the generalized contingent epiderivative DG F .x; N W RC R2
is defined by:
˚
N y/.x/
DG F .x; N D .y1 ; y2 / 2 R2 j y12 C y22 D x 2 ; y1 0; y2 0 :
Let x D 1. Then .1; 0/ 2 DG F .x; N y/.1/.

N We claim that .1; 0/ 62
N y/.1/.
DP F .x; N We have (see Fig. 11.5 (ii))
T .G.1/ C R2C ; .1; 0// D f.y1 ; y2 / 2 R2 j y1

0; y2 2 Rg;
and therefore
T .G.1/ C R2C ; .1; 0// \ .R2C / D f.y1 ; y2 / 2 R2 j y1 D 0; y2 0g 6D f0g;
confirming that .1; 0/ 62 Bo-PMin.G.1/; R2C / D GHe-PMin.G.1/; R2C /.

In fact, the generalized proper contingent epiderivative at x D 1 is given by:
˚
N y/.1/
DP F .x; N D .y1 ; y2 / 2 R2 j y12 C y22 D 1; y1 < 0; y2 < 0
N y/.1/:
which is a proper subset of DG F .x; N Indeed this is true for all x 2 RC =f0g:
However
N y/.0/
DP F .x; N D DG F .x;
N y/.0/
N D f0g:
Next we give an example showing that DG F .x; N y/./

N is a proper subset of
DW F .x;N y/./.
N N y/./nD
This example also makes it clear that the set DW F .x; N G
N y/./
F .x; N may be very large.

F .x/ D ˚; if x < 0
.y1 ; y2 / 2 R2 j y1 ; y2 2 Œ0; x if x
0:
Let xN D 0 and let yN D .0; 0/. We note that
N y//
T .epi .F /; .x; N D f.x; .y1 ; y2 // 2 R3 j x
0; y1
0; y2
0g;

˚
G.x/ D .y1 ; y2 / 2 R2 j y1
0; y2
0 :
N y/
is given by
N y/.x/
DG F .x; N WD f.0; 0/g;
N y/
and the weak contingent epiderivative DW F .x; N W RC R2 is given by (see
Fig. 11.6 (i)):
˚ [˚
N y/.x/
DW F .x; N D .0; y2 / 2 R2 j y2
0 .y1 ; 0/ 2 R2 j y1
0 ;
Weak minimal points

G(x)
of G (x)
G(x)
Weak minimal points of G(x)
F(x)
Minimal point Weak minimal points

of G(x) of G(x)
Intersection
(no minimal point)
2
-R + 2
-R +
Fig. 11.6 (i) Example 11.5.3: Minimal points. (ii) Example 11.5.4: Only weakly minimal points
N y/.x/
depicting the fact that the generalized weak contingent epiderivative DW F .x; N
is much larger than DG F .x; N y/.x/
N for all x 2 dom .DG F .x; N y//
N D
dom .DW F .x;N y//
N D RC :
By virtue of the connections among properly minimal, minimal and weakly
minimal points, it is clear that when DP F .x; N y/./
N N y/./
or DG F .x; N exists, then
DW F .x;N y/./
N also exists, provided that the ordering cone C has a nonempty interior.
However, it is possible that, out of these three epiderivatives, only DW F .x;N y/.x/
N
exists. Such a situation is furnished by the following example.
(
; if x < 0
F .x/ D ˚ S˚
.y1 ; y12 / 2 R2 j 0 y1 x .y1 ; y2 / 2 R2 j y1 0; y2
0 if x
0:
Let xN D 0 and let yN D .0; 0/. We note that
N y//
T .epi .F /; .x; N D f.x; .y1 ; y2 // 2 R3 j x
0; y2
0; y1 2 Rg;
G.x/ WD f.y1 ; y2 / 2 R2 j y2
0; y1 2 Rg:
N y/
is not defined as
N y/.x/
DG F .x; N D f;g:
However, the weak contingent epiderivative DW F .x; N y/

N W RC R2 is given by
(see Fig. 11.6 (ii)):
˚
N y/.x/
DW F .x; N WD .y1 ; 0/ 2 R2C j y1 2 R :
N y/.x/
Moreover, since DG F .x; N N y/.x/
is empty, DP F .x; N is empty as well.
11.5.1 Existence Theorems for Generalized Contingent

Epiderivatives
To give existence results for the generalized contingent epiderivatives, we need to

recall some notions and results (see Chap. 9). Let Y be a real Hausdorff topological
vector space. We formulate the following condition:
(P1) Every C -increasing and C -upper bounded net .xi /i 2I C is convergent to
an element of C .
Furthermore, following Németh [440], a convex cone C is called regular if it
satisfies (P1).
The following result is needed for our next result.
Proposition 11.5.5. Let D Y be nonempty. Assume that D is closed and
C -lower bounded, C satisfies the condition (P1). Then Min.D; C / 6D ; and the
domination property holds, that is, D Min.D; C / C C .
A proof of the above result as well as many other results concerning the existence
of minimal points, which could be useful in proving the existence of the generalized
contingent epiderivatives, can be found in Chap. 9.
We now give an existence theorem for the generalized contingent epiderivatives
and the generalized weak contingent epiderivatives.
Theorem 11.5.6. Let X and Y be normed spaces and let C Y be a closed,
convex, and regular cone. Let F W X Y be a set-valued map and let .x; N y/
N 2
graph .F /. Let, for every x 2 dom .G/, the set-valued map G given by (11.67),
be C -lower bounded. Then, for every x 2 dom .G/, the generalized contingent
N y/.x/
epiderivative DG F .x; N exists. Moreover, the following identity holds:
N y//
epi .DG F .x; N D T .epi .F /; .x;
N y//:
N (11.69)
If additionally int.C / 6D ;; then for every x 2 dom .G/, the generalized weak
N y/.x/
contingent epiderivative DW F .x; N also exists and the following identity holds:
N y//
epi .DW F .x; N D T .epi .F /; .x;
N y//:
N (11.70)
Proof. Since the contingent cone is always closed, the set G.x/ is closed for every
x 2 dom .G/. It has a C -lower bound by the assumption. Therefore, the existence
N y/.x/
of DG F .x; N now follows from Proposition 11.5.5 which ensures the existence
of minimal points for G.x/ with respect to the convex cone C . Moreover, it follows
from the same result that G.x/ has the domination property, that is
G.x/ DG F .x;
N y/.x/
N C C: (11.71)
We claim that equality holds in the above inclusion. For this, let y 2
N y/.x/
DG F .x; N C C be arbitrary. Then y 2 G.x/ C C: Now making use of the
fact that G.x/ D G.x/ C C , we obtain y 2 G.x/. This confirms the inclusion
N y/.x/
G.x/ DG F .x; N C C:
Therefore
N y/.x/
DG F .x; N C C D G.x/ D D.F C C /.x;
N y/.x/;
N
where D.F C C /.x; N y/.x/

N is the contingent derivative of the profile map .F C
C /: Furthermore, the above identity and the definition of the contingent derivative
implies
N y//
epi .DG F .x; N D graph .D.F C C /.x;
N y//
N
D T .graph .F C C /; .x;
N y//
N
D T .epi .F /; .x;
N y//;
N
which confirms the claim.

N y/.x/
The existence of DW F .x; N follows from the existence of DG F .x;N y/.x/
N and
the inclusion DG F .x;N y/.x/
N DW F .x; N y/.x/
N (see ((11.68)). It remains to prove
that the identity (11.70) holds. In fact, in view of the inclusions (11.68) and (11.70),
it follows that
G.x/ DW F .x;
N y/.x/
N C C:
Moreover, the reverse inclusion follows from arguments similar to those in the proof
of (11.69). The proof is complete. t
u
Remark 11.5.7. Note that if the generalized continent epiderivative DG F .x; N y/
N of
N y/,
F at .x; N then (11.69) holds if and only if the domination property holds for the
set D.F CC /.x; N y/.x/
N for any x 2 dom .G/: In fact, note that if epi .DG F .x;
N y//
N ¤
N y//
T .epi .F /; .x; N then there exists .x; y/ such that .x; y/ 2 T .epi .F /; .x; N y//
N
and .x; y/ … epi .DG F .x; N y//:
N This means that y 2 D.F C C /.x; N y/.x/
N but
y … DG F .x; N y/.x/
N C C; contradicting domination property holds for D.F C
N y/.x/:
C /.x; N See [500].
The relations (11.69) and (11.70) play a fundamental role in set-valued opti-
mization. The main reason for this is the fact that these identities enable the
use of several properties of the contingent cones to obtain various conclusions
about the corresponding epiderivatives. It should also be clear that in deriving the
identities (11.69) and (11.70), the domination property (11.70) has played a decisive
role.
We note that if the weak contingent epiderivative exists, then we can
prove (11.70) by assuming the following weak-domination property:
G.x/ WMin.G.x/; C / C C:
It is easy to check that the weak-domination property is much weaker than the
domination property.
The following existence result can be useful when less information is available
about the ordering cone C . Recall that given a set A Y and a point s 2 Y , the
C -lower section of A at s, denoted by As , is given by the set A \ .s C /.
N 2
graph .F /. Assume that for each x 2 dom .G/; there exists s 2 Y such that the
C -lower section Gs .x/ is nonempty and compact. Then for every x 2 dom .G/, the
N y/.x/
generalized contingent epiderivative DG F .x; N exists. Moreover, if int.C / 6D ;
N y/.x/
then the generalized weak contingent epiderivative DW F .x; N also exists.
N y/.x/
Proof. The existence of DG F .x; N N y/.x/)
(and hence of DW F .x; N follows from
[57, Theorem 1], which assures the existence of minimal points for G.x/ with
respect to the cone C . t
u
The following existence theorem in a finite-dimensional setting is also of
importance.
Theorem 11.5.9. Let X be a normed space and let C Rn be a pointed, closed,
and convex cone. Let F W X Rn be a set-valued map and let .x; N y/
N 2 graph .F /:
Let for each x 2 dom .G/, the set G.x/, given by (11.67), be convex and let s 2 Rn
be such that the C -lower section Gs .x/ is nonempty and compact. Then, for every
x 2 dom .G/, the epiderivatives DP F .x;
N y/.x/,
N N y/.x/
DG F .x; N N y/.x/
and DW F .x; N
exist. Moreover,(11.69) and (11.70) hold.
Proof. The existence of DG F .x; N y/.x/
N N y/.x/
and DW F .x; N follow from Theo-
rem 11.5.8 and the existence of DP F .x; N y/.x/
N follow from Luc [403, Corol-
lary 3.16]. Recall that in Rn every closed convex and pointed cone has a compact
base (see, for example, [126]). It follows from Lemma 11.5.5 (2) that the hypotheses
imposed on G.x/ together with its closedness assures the domination property. The
rest of the proof is similar to the proof of Theorem 11.5.6. t
u
In the following, we give some important properties of the generalized contingent
epiderivatives. In these results, the equations (11.69) and (11.70) play an important
role.
Theorem 11.5.10. If, besides the assumptions of Theorem 11.5.6 or Theo-

N y//
rem 11.5.9, we suppose that the contingent cone T .epi .F /; .x; N is convex, then
N y/./
the generalized contingent epiderivative DG F .x; N is a subadditive set-valued
map. That is
N y/.x/
DG F .x; N C DG F .x;
N y/.y/
N DG F .x;
N y/.x
N C y/ C C for all
N y//:
x; y 2 dom .DG F .x; N
Proof. It follows immediately from (11.69) and the convexity of T .epi .F /;

N y//.
.x; N t
u
Since
N y/.
x/
DG F .x; N D
DG F .x;
N y/.x/
N
for every x 2 dom .DG F .x; N y//

N and for every
> 0, the above result states that the
N y/./
generalized contingent epiderivative DG F .x; N is a strictly sublinear set-valued
map. That is, for all x; y 2 dom .DG F .x;
N y//;
N we have
N y/.x/
DG F .x; N C ˇDG F .x;

N y/.y/
N DG F .x;
N y/.
x
N C ˇy/ C C
> 0; ˇ > 0:
We have the following:

Corollary 11.5.11. Under the assumptions of Theorem 11.5.6 or Theorem 11.5.9,
the following holds:
N y//
epi .DW F .x; N D epi .DG F .x;
N y//:
N
Proof. This is a direct consequence of (11.69) and (11.70). t

u
It is well-known that single-valued functionals are identical if and only if
their epigraphs are identical, provided that the ordering cone is pointed. However,
contrary to this, two set-valued maps may not be identical even if their epigraphs are
N y/
identical. Clearly, Corollary 11.5.11 does not imply that DG F .x; N DW F .x;N y/.
N
Indeed the difference between DG F .x; N y/
N and DW F .x;N y/
N can be very large (see
Example 11.5.4).
Corollary 11.5.12. Under the assumptions of Theorem 11.5.6 or Theorem 11.5.9,
the following holds:
N y//
dom .DW F .x; N T .dom .F /; x/;
N
N y//
dom .DG F .x; N T .dom .F /; x/:
N
Proof. We only prove the second inclusion because the first follows directly
from Theroem 11.5.11. Let .u; v/ 2 epi .DG F .x; N y//:
N Therefore .u; v/ 2
N y//
T .epi .F /; .x; N and hence there are sequences ftn g P and f.un ; vn /g with
tn # 0 and .un ; vn / ! .u; v/ such that
yN C tn vn 2 F .xN C tn un / C C:
Therefore xN C tn un 2 dom .F / implying that .u; v/ 2 T .dom .F /; x/

N Y . The
u
The following result connects the contingent derivative and the generalized
contingent derivative.
closed, and convex cone. Let F W X Y be a set-valued map and let .x;
N y/
N 2
graph .F /. Let
SMin.G.x/; C / 6D ; for every x 2 dom .G/:
N y/./
Then the contingent epiderivative DE F .x; N and the generalized contingent
N y/./
epiderivative DG F .x; N coincide.
Proof. Since the set SMin.G.x/; C / is nonempty, it coincides with Min.G.x/; C /:
By the assumption, the convex cone C is pointed and therefore Min.G.x/; C / is a
singleton set. It is clear from the definition of a strongly minimal point that it satisfies
the domination property and hence we deduce that (11.69) holds. Consequently,
N y/./
DG F .x; N is a single-valued map satisfying (11.69). However, (11.69) is also
satisfied by the contingent epiderivative DE F .x; N y/./,
N which under the imposed
conditions is unique. This completes the proof. t
u
In view of the above theorems, we make a brief discussion of Example 11.4.6–
11.5.4. It is easy to verify that, for G.x/ given in Example 11.4.6, all the hypotheses
of Theorem 11.5.6-11.5.9 are satisfied. Hence the existence of DG F .x; N y/.x/
N is
assured by these theorems. The same is the case for Example 11.5.3. However, for
G.x/ defined in Example 11.5.4, there is no R2C -lower bound. Neither does it have
any compact section. Hence, the above theorems do not assure the existence of
N y/./.
DG F .x; N Indeed, we have seen that DG F .x; N y/./
N does not exist.
11.5.2 Characterizations of Generalized Contingent

Epiderivatives
We begin with the following simple result.

Theorem 11.5.14. Let X and Y be normed spaces, let C Y be pointed, closed,
N y/
N 2 graph .F /.
If for some x 2 X , the proper contingent epiderivative DP F .x; N y/.x/
N exists,
N y/.x/
then DP F .x; N and DG F .x;N y/.x/
N belong to the boundary of the contingent
derivative D.F C C /.x;N y/.x/:
N
Proof. Assume that for some x 2 X the proper contingent epiderivative

N y/.x/
DP F .x; N exists. Since D.F C C /.x;
N y/.x/
N is closed, it suffices to show
fDP F .x;
N y/.x/;
N N y/.x/g
DG F .x; N 6 int.D.F C C /.x;
N y/.x//:
N
Because DP F .x; N y/.x/

N DG F .x; N y/.x/,
N we will prove the statement for
N y/.x/
DG F .x; N only. Let y 2 DG F .x; N y/.x/
N be arbitrary. Clearly y 2 D.F C
N y/.x/.
C /.x; N Let y 2 int.D.F C C /.x; N y/.x//.
N Then there exists " > 0 so that
B" .y/.D y C B" .0// D.F C C /.x; N y/.x/,
N where B" .y/ is an open ball of radius
" at y. B" .0/ is absorbing, that is, for all z 2 Y there exists some `0 > 0 such that
j`j `0 implies `z 2 B" .0/. For an arbitrary c 2 C nf0g; there exists such that
c 2 B" .0/ and y c 2 D.F C C /.x; N y/.x/;
N implying y 62 DG F .x;
N y/.x/.
N This
contradiction finishes the proof. t
u
For the next result, we need to recall the following (see Borwein [57]):
Theorem 11.5.15. Let Z be a locally convex space, let P and Q be cones in Z
such that P \ Q D f0g. Assume that one of the following conditions hold:
1. P is weakly-closed and Q has a weakly-compact base.
2. P is closed and Q has a compact base.
Then there is a pointed, convex cone R such that Qnf0g int.R/ and P \ R D
f0g:
We begin with the following auxiliary result.
Theorem 11.5.16. Let Y be a normed space, let C Y be pointed, closed,
and convex cone, and let D Y be arbitrary. Then GHe-PMin.D; C /
Bo-PMin.D; C /: Furthermore, if either cl .D C C / is convex and C has a
weakly-compact base or C has a compact base, then the converse holds.
Proof. Let yN 2 GHe-PMin.D; C /, that is, there exists a pointed convex cone C1 6D
Y such that C nf0g int.C1 / and yN 2 Min.D; C1 /. Let yN 62 Bo-PMin.D; C / and
let y 2 T .D C C; y/
N \ .C nf0g/: There are f
n g # 0 and fyn g ! y such that
yN C
n yn 2 D C C: Since y 2 C nf0g int.C1 /, there exists N 2 N such that
yn 2 int.C1 /, 8n
N: Since
n > 0 we deduce that
n yn 2 int.C1 /, for every
n
N: Let wn 2 D be such that yN C
n yn 2 wn C C which implies
wn 2 yN C
n yn C yN int.C1 / C yN int.C1 / C1 nf0g C1 yN C1 nf0g:
Therefore yN 2 wn C C1 nf0g; which contradicts that yN 2 Min.D; C1 /.

Now, under the additional hypotheses, we show the converse. Let yN 2
Bo-PMin.D; C /; that is, T .D C C; y/ N \ .C / D f0g: If cl.D C C / is convex,
T .D C C; y/N D T .cl.D C C/; y/ N is convex as well, and hence weakly closed.
Therefore, by virtue of Proposition 11.5.15, there is a pointed convex cone C1 such
that C nf0g int.C1 / and T .D CC; y/\.C
N 1 / D f0g. Now taking into account
the convexity of the set cl .D C C /; we conclude that .D y/

N \ .C1 nf0g/ D ;:
This implies that yN 2 Min.D; C1 / and consequently yN 2 GHe-PMin.D; C /. u
t
Recall that for a convex cone C Y , the quasi-interior C # of the dual C C of C
is given by
C # D f` 2 Y j `.x/ > 0 8x 2 C nf0gg:
In the following, we give a characterization of the proper contingent epideriva-

tive.
and convex cone with a compact base, let F W X Y be a set-valued map, and let
N y/
.x; N 2 graph .F /. Let T .epi .F /; .x;
N y//
N be convex. Assume that for some x 2 X ,
the proper contingent epiderivative DP F .x;N y/.x/
N exists. Then
[
N y/.x/
DP F .x; N D `.G.x// ; (11.72)
`2C #
where the map G is given by (11.67) and `.G.x// WD fy 2 Y j `.y/ `.G.x//g:

Proof. Let y 2 [`2C # `.G.x// . Then there exists some `0 2 C # such that
`0 .y / `0 .y/; 8 y 2 G.x/: (11.73)
Let y 62 DP F .x;
N y/.x/.
N In view of Proposition 11.5.16 we have that y 62
Bo-PMin.G.x/; C /, where G.x/ is defined by (11.5.1). Let w 2 T .G.x/CC; y /\
.C nf0g/ be arbitrary. Since G.x/ C C D G.x/, we have w 2 T .G.x/; y /:
Therefore, there are a sequences f
n g P and fwn g G.x/ such that
n .wn
y / ! w and wn ! y . In view of (11.73), we have `0 .y / `0 .wn / for every
n 2 N and accordingly limn!1
n .`0 .wn / `0 .y //
0: Consequently `0 .w/
0
for each w 2 T .G.x/; y / \ .C nf0g/: This contradiction to `0 .w/ < 0 confirms
that y 2 DP F .x;
N y/.x/.
N Hence [`2C # `.G.x/ DP F .x;N y/.x/:
N
For the converse, let y 2 DP F .x;
N y/.x/:
N N y//
Since T .epi .F /; .x; N is convex,
G.x/ is convex for all x 2 dom .G/ and consequently T .G.x/ C C; y / is convex.
We have
T .G.x/ C C; y / D T .G.x/; y / G.x/ y :
From the definition of y and Proposition 11.5.16 we have
T .G.x/; y / \ .C nf0g/ D ;:
Therefore, by a cone separation theorem (see [426] [293]), there exists `0 2 C #

such that `0 2 ŒT .G.x/; y /C ; or, `0 .G.x/ y /
0; which ensures `0 .y /
`0 .G.x//: The proof is complete. t
u
In the following we characterize the weak contingent epiderivatives.

Theorem 11.5.18. Let X; Y; F; `.G.x// and .x; N y/
N be as in Theorem 11.5.17.
Let int.C / 6D ;. Assume that for some x 2 X , the weak contingent epiderivative
N y/.x/
DW F .x; N exists. Then:
[
N y/.x/
DW F .x; N `.G.x// : (11.74)
`2C C nf0g
N y//
Moreover, if T .epi .F /; .x; N is convex, then equality holds in (11.74).
N y/.x/
Proof. Since DW F .x; N D WMin.G.x/; C /, we have G.x/ \ .y int.C // D
; 8y 2 DW F .x; N y/.x/:
N Hence the standard separation theorems can be applied.
The arguments are quite similar to that of the previous theorems and hence
omitted. t
u
We now proceed to give some connections between the epiderivatives and
the contingent derivatives. For this we the recall the following result (see also
Jahn [293]).
Lemma 11.5.19. Let Y be a normed space, let C Y be pointed, closed, and
convex cone, and let A and B be nonempty subsets of Y: Assume that A C C D B:
Then the following relationships hold:
(a) GHe-PMin.A; C / D GHe-PMin.B; C /.
(b) Min.A; C / D Min.B; C /.
(c) WMin.A; C / WMin.B; C /:
We have the following result:
Theorem 11.5.20. Let X and Y be normed spaces and let C Y be pointed,
N 2
graph .F /. Assume that for x 2 X the epiderivatives DP F .x;N y/.x/,
N N y/.x/
DG F .x; N
and DW F .x;N y/.x/
N N y/;
exist. If the map F is contingently C -absorbing at .x; N then
the following relationships hold:
N y/.x/
(a) DP F .x; N D GHe-PMin.DF.x;N y/.x/;
N C /.
N y/.x/
(b) DG F .x; N D Min.DF.x;
N y/.x/;
N C /.
N y/.x/
(c) DW F .x; N WMin.DF.x;N y/.x/;
N C /.
Proof. (a) By virtue of
N y/.x/
DF.x; N C C D DF.x;
N y/.x/
N for all x 2 L;
we have
GHe-PMin.D.F C C /.x;
N y/.x/;
N C / D GHe-PMin.DF.x;
N y/.x/
N C C; C /;
which, in view of Lemma 11.5.19 (a), implies that DP F .x; N y/.x/

N D
GHe-PMin.DF.x; N y/.x/;
N C /:
The proof of (b) and (c) follows from the last two parts of Lemma 11.5.19. t
u
Finally, we relate the epiderivatives with a contingent cone in the image space.
convex, and solid cone, and let A X be nonempty. Let F W A W X Y be a
N y/
set-valued map and let .x; N 2 graph .F /: Let S WD dom .DW F .x;
N y//.
N Then:
[
N y/.x/
DW F .x; N T .F .A/ C C; y/:
N (11.75)
x2S
Proof. Let T .F .A/ C C; y/ N 6D Y otherwise there is nothing to prove. Let y 2

N y/.S
DW F .x; N / be arbitrary and let x 2 S be the corresponding element such that
y 2 DW F .x;
N y/.x/:
N Therefore
y 2 WMin.G.x/; C / G.x/;
where G is defined in (11.5.1). Hence .x; y/ 2 graph .G/ D T .epi .F /; .x; N y//:
N
By the definition of the contingent cone, there are sequences f
n g P; fxn g X
and fyn g Y such that
n ! 0; xn ! x; and yn ! y with yN C
n yn 2
N
F .xC
N n yn 2 F .xC
n xn /CC: This implies yC

N n xn /CC F .A/CC: Therefore,
in view of the fact that
n ! 0 and yn ! y; we deduce that y 2 T .F .A/ C C; y/: N
N y/.S
Since y is chosen arbitrarily, we have DW F .x; N / T .F .A/ C C; y/;
N and the
u
11.6 Calculus Rules for Contingent Epiderivatives
In this section we present some calculus rules for the contingent epiderivatives.
A formula for scalar multiplication, sum formulae, and chain rules are given.
This section is based on [299] and is strongly influenced by the general approach
of deriving calculus rules devised by Ward [590, 593, 597]). This section is an
extension of results given in Sect. 11.2.2 for the contingent derivatives to contingent
epiderivatives.
We begin with exploring the contingent epiderivative for the product of a
set-valued map and a strictly positive scalar. Let X and Y be normed spaces and let
C Y be a pointed, closed, and convex cone. Given a set-valued map F W X Y
and a scalar
> 0, we first define a set-valued map
˝ F W X Y with

˝ F .x/ D
F .
1 x/; for every x 2 X:
We show that
˝ F is a set-valued analogue of the epi-multiple function (see
[499]).
11.6 Calculus Rules for Contingent Epiderivatives 483
Lemma 11.6.1. Let X and Y be normed spaces, let C Y be a pointed, closed,

and convex cone C , let F W X Y be a set-valued map, and let
> 0: Then:
epi .
˝ F / D
epi .F /:
Proof. Clearly,
epi .
˝ F / D f.x; y/ 2 X Y j y 2
˝ F .x/ C C g
1
D f.x; y/ 2 X Y j y 2
F . x/ C C g

D f.
x;
y/ 2 X Y j y 2 F .x/ C C g
D
epi .F /;
which justifies the claim. The proof is complete. t

u
Proposition 11.6.2. Let X and Y be normed spaces, let C Y be a pointed,
closed, and convex cone C , let F W X Y be a set-valued map, and let .x; N y/
N 2
N y/
graph .F /: If the contingent epiderivative DE F .x; N exists, then for
> 0; the
contingent epiderivative DE .
˝ F /.
x;N
y/
N exists, and,
DE .
˝ F /.
x;
N
y/./
N D DE F .x;
N y/./:
N
Proof. From Lemma 11.6.1 and from the definition of the contingent epiderivative,
we have
epi .DE .
˝ F /.
x;
N
y//
N D T .epi .
˝ F /; .
x;
N
y//
N
D T .
epi .F /;
.x;
N y//
N
D
T .epi .F /; .x;
N y//
N
D T .epi .F /; .x;
N y//
N
D epi .DE F .x;
N y//
N
which implies the assertion. The proof is complete. t

u
We now present a formula for scalar multiplication.
Theorem 11.6.3. Let X and Y be normed spaces, let C Y be a pointed,
closed, and convex cone C , let F W X Y be a set-valued map, and let
N y/
.x; N y/
N 2 graph .F /: If the contingent epiderivative DE F .x; N exists, then for
> 0; the contingent epiderivative DE .

˝ F /.
x;
N
y/
N exists, and,
N
y/./
DE .
F /.x; N D
DE F .x;
N y/./:
N
Proof. For a homomorphism

W X Y ! X Y defined by
.x;
N y/
N WD .x;
N
y/,
N
it is obvious that
epi .
F / D f.x;
N y/
N 2 X Y; W yN 2 .
F /.x/
N C Cg
D f.x;
N
y/
N 2 X Y; W yN 2 F .x/
N C Cg
D
.epi .F //:
Then the definition of the contingent epiderivative implies that
N
y//
epi .DE .
F /.x; N D T .epi .
F /; .x;
N
y//
N
N y//
D T .
.epi .F //;
.x; N
D
.T .epi .F /; .x;
N y///
N
D
.epi .DE F .x;
N y///;
N
N
y/./
which confirms that DE .
F /.x; N D
DE F .x;
N y/./:
N t
The proof is complete.u
Our next result of this section is the following sum formula:
Theorem 11.6.4. Let X; Y; and Z be real Banach spaces, and let C Z be a
pointed, closed, and convex cone. Let F1 W X Z and F2 W Y Z be set-valued
maps, and let f W X ! Y be a single valued map being strictly differentiable
on Bı .x/ for ı > 0. Let .x; y1 / 2 graph .F1 / and .f .x/; y2 / 2 graph .F2 /.
Let the contingent epiderivatives DE .F1 C F2 ı f /.x; y1 C y2 /, DE F1 .x; y1 / and
DE F2 .f .x/; y2 / exist. Let epi .F1 / and epi .F2 / be epi-Lipschitz-like and derivable
at .x; y1 / and .f .x/; y2 /, respectively. Let
r .x; y1 ; f .x/; y2 /C.epi .F1 / epi .F2 /; .x; y1 ; f .x/; y2 // D Y; (11.76)
where W X Z Y Z ! Y is defined by .˛; ˇ; ; ı/ WD f .˛/ : Then
DE .F1 C F2 ı f /.x; y1 C y2 /./ DE F1 .x; y1 /./ C DE F2 .f .x/; y2 /.rf .x/.//:
Proof. Let us define a map ˚ W X Z Y Z ! X Z by
˚.˛; ˇ; ; ı/ D .˛; ˇ C ı/:
Clearly ˚ is linear and continuous. For simplicity we set F WD F1 C F2 ı f ,

˝ WD epi .F1 / epi .F2 / and K WD ˝ \ 1 .0/. From the definitions of the maps
and ˚ and of ˝, it is evident
˚.K/ D ˚..epi .F1 / epi .F2 // \ f.x; y; w; z/ W f .x/ D wg/

D ˚.f.x; y; w; z/ 2 X Z Y Z W y 2 F1 .x/ C C;
z 2 F2 .w/ C C; f .x/ D wg/

D f.x; y C z/ 2 X Z W y 2 F1 .x/ C C; z 2 F2 ı f .x/ C C g
D epi .F /: (11.77)
It follows from the definition of the contingent epiderivative, (11.77), and well-
known properties of the contingent cone that
epi .DE F .x; y1 C y2 // D T .epi .F /; .x; y1 C y2 //

D T .˚.K/; ˚.x0 // with x0 WD .x; y1 ; f .x/; y2 /
˚.T .K; x0 //: (11.78)
Since (11.76) holds and the set ˝ being a product of epi-Lipschitz-like sets is
epi-Lipschitz-like (see Definition 4.7.1), it follows that Theorem 4.7.5 is applicable
and consequently
T .K; x0 / D T .˝ \ 1 .0/; x0 / T .˝; x0 / \ r .x0 /1 .0/;
which, in view of (11.78), leads to
epi .DE F .x; y1 C y2 // ˚.T .˝; x0 / \ r .x0 /1 .0//

D ˚.T .epi .F1 / epi .F2 /; x0 / \ r .x0 /1 .0//:
(11.79)
Observe that derivability of epi .F1 / at .x; y1 / and of epi .F2 / at .f .x/; y2 /
implies derivability of ˝ at x0 . Consequently,
T .epi .F1 /; .x; y1 // T .epi .F2 /; .f .x/; y2 //

D T .epi .F1 / epi .F2 /; .x; y1 ; f .x/; y2 //: (11.80)
Therefore, with (11.79) and (11.80), we obtain
epi .DE F .x; y1 C y2 // ˚.T .epi .F1 /; .x; y1 //

T .epi .F2 /; .f .x/; y2 // \ r .x0 /1 .0//
D ˚..epi .DE F1 .x; y1 //
epi .DE F2 .f .x/; y2 /// \ r .x0 /1 .0//
D ˚.fu; v; w; r/ 2 X Z Y Z W
.u; v/ 2 epi .DE F1 .x; y1 //;
.w; r/ 2 epi .DE F2 .f .x/; y2 //; rf .x/.x/ D wg/
D f.u; v C r/ 2 X Z W .u; v/ 2 epi .DE F1 .x; y1 /;

.rf .x/.x/; r/ 2 epi .DE F2 .f .x/; y2 /g
D epi .DE F1 .x; y1 /./ C DE F2 .f .x/; y2 /.rf .x/.///;
which confirms that
epi .DE F .x; y1 C y2 /.// epi .DE F1 .x; y1 /./ C DE F2 .f .x/; y2 /.rf .x/.///;
implying
DE .F1 C F2 ı f /.x; y1 C y2 /./ DE F1 .x; y1 /./ C DE F2 .f .x/; y2 /.rf .x/.//;

u
By setting X D Y in the above theorem, we get the following result.
Corollary 11.6.5. Let the assumptions of Theorem 11.6.4 hold with f as the
identity map. Then:
DE .F1 C F2 /.x; y1 C y2 /./ DE F1 .x; y1 /./ C DE F2 .x; y2 /./:
We now investigate chain rules for the contingent epiderivatives. Throughout the
remaining part of this section, let X , Y and Z be Banach spaces. The space Z is
partially ordered by a pointed, closed, and convex cone C . Let F W X Y and
G W Y Z be given set-valued maps. In the following by PX Z we shall represent
the canonical projection from X Y Z onto X Z:
The following lemma will be used shortly ( see Pennanen [468]):
Lemma 11.6.6. For any F W X Y and G W Y Z the following holds:
.a/ graph .G ı F / D PX Z Œ.graph .F / Z/ \ .X graph .G//;
.b/ epi .G ı F / D PX Z Œ.graph .F / Z/ \ .X epi .G//:
Proof. We only prove part (b). Since
epi .G ı F / D f.x; z/ 2 X Z W 9 y 2 F .x/; z 2 G.y/ C C g

D f.x; z/ 2 X Z W 9 y W .x; y; z/ 2 .graph .F / Z/ \ .X epi .G//g;
the proof is complete. t

u
We are now ready to give the following chain rule:
Theorem 11.6.7. Let graph .F / and epi .G/ be epi-Lipschitz-like and derivable at
.x; y/ 2 graph .F / and .y; z/ 2 graph .G/, respectively. Assume that the contingent
epiderivative DE G.y; z/ exists. Moreover, assume that
r .x; y; y; z/C.graph .F / epi .G/; .x; y; y; z// D Y; (11.81)

where W X Y Y Z ! Y is defined by .˛; ˇ; ; ı/ WD ˇ . ThenW
epi .DE .G ı F /.x; z// epi .DE G.y; z/ ı DF.x; y//: (11.82)
Proof. Clearly ./, as defined above, is linear and continuous. Set ˙ WD

.graph .F / epi .G// \ 1 .0/ and ! WD .x; y; y; z/: In view of Lemma 11.6.6,
we can express epi .G ı F / as follows
epi .G ı F / D PX Z ..graph .F / epi .G// \ 1 .0//;
where PX Z is the canonical projection from X Y Y Z onto X Z:

From the definition of the contingent epiderivative, we obtain
epi .DE .G ı F /.x; z// D T .epi .G ı F /; .x; z// D T .PX Z .˙/; PX Z .!//;
and by using the fact that PX Z is linear and continuous, we also have
epi .DE .G ı F /.x; z// PX Z T .˙; !/: (11.83)
Since (11.81) holds and graph .F / epi .G/ is epi-Lipschitz-like, we get
T .˙; !/ D T ..graph .F / epi .G// \ 1 .0/; !/

T .graph .F / epi .G/; !/ \ r .!/1 .0/; (11.84)
which, in view of the derivability assumptions, yields
T .˙; !/ T .graph .F /; .x; y// T .epi .G/; .y; z// \ r .!/1 .0/;
D Œgraph .DF.x; y// epi .DE G.y; z// \ r .!/1 .0/:
Combining the above inclusion with (11.83) gives
epi .DE .G ı F /.x; z//

PX Z .graph .DF.x; y// epi .DE G.y; z// \ r .!/1 .0//
D epi .DE G.y; z/ ı DF.x; y//;
which completes the proof. t

u
Recall that f W X ! Y is said to be isotone, if f .y/ 2 f .x/ C C , provided that
y 2 x C C:
We now give a chain rule for the contingent epiderivatives.
Theorem 11.6.8. Besides the hypotheses of Theorem 11.6.7, assume that F is the
profile map of a set-valued map H , the contingent epiderivative DE G.y; z/ exists,
and the contingent epiderivative DE G.x; y/ is isotone. Then
DE .G ı F /.x; z/ DE G.y; z/ ı DE H.x; y/:
Proof. By virtue of the isotony property of DE G.x; y/; the following identity holds:
epi .DE G.y; z/ıDE H.x; y// D PX Z .epi .DE H.x; y// epi .DE G.y; z//\ 1 .0//:
Hence the above equation in view of (11.82) and the identity graph .DF.x; y// D
epi .DE H.x; y// gives
epi .DE .G ı F /.x; z// epi .DE G.y; z/ ı DE H.x; y//;
which implies that DE .G ı F /.x; z/ DE G.y; z/ ı DE H.x; y/: The proof is

complete. u
t
11.7 Second-Order Derivatives of Set-Valued Maps
In this section, we study second-order extensions of the graphical derivatives and

epiderivatives of set-valued maps studied earlier in this chapter. These notions are
defined by using second-order tangent cones and second-order tangent sets studied
in Chap. 4.
The following definition introduces some second-order derivatives of set-valued
maps.
N y/
map, let .x; N 2 graph .F /; and let .Nu; vN / 2 X Y: Then:
N y;
1. A set-valued map D 2 F .x; N uN ; vN / W X Y is called the second-order
N y/
contingent derivative of F at .x; N in the direction .Nu; vN /; if
N y;
graph .D 2 F .x; N uN ; vN // D T 2 .graph .F /; .x;
N y/;
N .Nu; vN //;
N y/;
where T 2 .graph .F /; .x; N .Nu; vN // is the second-order contingent set.
2. A set-valued map D 2 F .x; N y;
N uN ; vN / W X Y is called the second-order
compound contingent derivative of F at .x; N y/
N in the direction .Nu; vN /; if
N y;
graph .D 2 F .x; N uN ; vN // D T .T .graph .F /; .x;
N y//;
N .Nu; vN //;
N y//;
where T .T .graph .F /; .x; N .Nu; vN // is the contingent cone of T .graph .F /;
N y//
.x; N at .Nu; vN /:
N y;
3. A set-valued map DA2 F .x; N uN ; vN / W X Y is called the second-order adjacent
N y/
derivative of F at .x; N in the direction .Nu; vN /; if
N y;
graph .DA2 F .x; N uN ; vN // D A2 .graph .F /; .x;
N y/;
N .Nu; vN //;
N y/;
where A2 .graph .F /; .x; N .Nu; vN // is the second-order adjacent set.
11.7 Second-Order Derivatives of Set-Valued Maps 489
4. A set-valued map DA2 F .x;

N y;
compound adjacent derivative of F at .x; N y/
N y;
graph .DA2 F .x; N uN ; vN // D A.A.graph .F /; .x;
N y//;
N .Nu; vN //;
N y//;
where A.A.graph .F /; .x; N .Nu; vN // is the adjacent cone of A.graph .F /;
N y//
.x; N at .Nu; vN /:
5. A set-valued map DC2 F .x;
N y;
compound Clarke derivative of F at .x; N y/
N y;
graph .DC2 F .x; N uN ; vN // D C.C.graph .F /; .x;
N y//;
N .Nu; vN //;
N y//;
where C.C.graph .F /; .x; N .Nu; vN // is the adjacent cone of A.graph .F /;
N y//
.x; N at .Nu; vN /:
6. A set-valued map D2 F .x;
N y;
asymptotic contingent derivative of F at .x;N y/
N y;
graph .D2 F .x; N uN ; vN // D T2 .graph .F /; .x;
N y/;
N .Nu; vN //;
where T2 .graph .F /; .x;

N y/;
N .Nu; vN // is the second-order asymptotic contingent
cone.
7. A set-valued map D2A F .x;N y;
asymptotic adjacent derivative of F at .x; N y/
N y;
graph .D2A F .x; N uN ; vN // D A2 .graph .F /; .x;
N y/;
N .Nu; vN //;
where A2 .graph .F /; .x; N y/;

N .Nu; vN // is the second-order asymptotic adjacent cone.
8. A set-valued map DL2 F .x; N y;
lower Dini derivative of F at .x; N y/N in the direction .Nu; vN / provided that y 2
N y;
DL2 F .x; N uN ; vN /.x/; if and only if, for every .tn / P and for every fxn g X
with tn # 0 and xn ! x; there are a sequence fyn g Y with yn ! y and an
index m 2 N such that yN C tn vN C tn2 yn 2 F .xN C tn uN C tn2 xn /; for every n
m:
Remark 11.7.2. It is clear that if .Nu; vN / D .0; 0/ in the above definition, then the
second-order contingent derivative, the second-order compound contingent deriva-
tive, and the second-order asymptotic contingent derivative recover the contingent
derivative DF.x;N y/
N of F at .x; N y/;
N and the second-order adjacent derivative, the
second-order compound adjacent derivative, and second-order asymptotic adjacent
derivative recover the adjacent derivative DA F .x; N y/N of F at .x;N y/:
N The notion
of second-order lower Dini derivative is inspired by the lower Dini derivative
introduced by Penot [462]. In fact, if .Nu; vN / D .0; 0/, then the above notion recovers
the lower Dini derivative of [462]. The second-order analogue given above was first
introduced and used in [313] to give some necessary optimality conditions in set-
valued optimization.
It is natural to have a second-order version of derivability and related notions:

N y/
map, let .x; N 2 graph .F /; and let .Nu; vN / 2 X Y:
1. The map F is said to be second-order proto-differentiable or second-order
derivable at .x; N y/
N in the direction .Nu; vN / 2 X Y; if the second-order contingent
derivative and the second-order adjacent derivative of F at .x; N y/
N in direction
.Nu; vN / coincide, that is,
N y/;
T 2 .graph .F /; .x; N .Nu; vN // D A2 .graph .F /; .x;
N y/;
N .Nu; vN //:
2. The map F is said to be second-order compound proto-differentiable or

second-order compound derivable at .x; N y/
N in the direction .Nu; vN / 2 X
Y; if the second-order compound contingent derivative and the second-order
N y/
compound adjacent derivative of F at .x; N in direction .Nu; vN / coincide, that is,
N y//;
T .T .graph .F /; .x; N .Nu; vN // D A.A.graph .F /; .x;
N y//;
N .Nu; vN //:
3. The map F is said to be second-order semi-differentiable at .x; N y/ N in the

direction .Nu; vN / 2 X Y; if the second-order contingent derivative and the
N y/
second-order lower Dini derivative of F at .x; N in the direction .Nu; vN / coincide.
Note that if a set-valued map is strict proto-differentiable, then it is second-order
compound proto-differentiable.
The following is a second-order analogue of Theorem 11.1.28:
map which is Lipschitz-like around .x; N y/ N 2 graph .F /: Then F is second-order
N y/
derivable at .x; N in any direction .Nu; vN / 2 X Y if and only if F is second-order
N y/
semi-differentiable at .x; N direction .Nu; vN / 2 X Y .
Proof. In view of the chain of inclusions
N y;
DL2 F .x; N uN ; vN / DA2 F .x;
N y;
N uN ; vN / D 2 F .x;
N y;
N uN ; vN /;
it is evident that if F is second-order semi-differentiable at .x; N y/

N in any direction
.Nu; vN / 2 X Y then it is second-order derivable at .x; N y/
N in that direction. For the
converse, let F be second-order derivable at .x; N y/N in the direction .Nu; vN / 2 X Y
and let .x; y/ 2 graph .D 2 F .x; N y;
N uN ; vN // be arbitrary. Then, for every ftn g P; there
exists f.zn ; wn /g X Y such that tn # 0, .zn ; wn / ! .x; y/ and yN C tn vN C tn2 wn 2
F .xN C tn uN C tn2 zn /: Let fxn g X be an arbitrary sequence such that xn ! x: Using
the fact that F is Lipschitz-like at .x; N y/,N we can show that there exist a sequence
fyn g and an index m 2 N such that .xN C tn uN C tn2 xn ; yN C tn vN C tn2 yn / 2 graph .F /
for every n
m: Consequently, y 2 DL2 F .x; N y;
N uN ; vN /.x/. The proof is complete. u t
The following result connects the second-order contingent and the second-order
adjacent derivatives to the second-order Fréchet derivative of a single-valued smooth
map (see [594]).
Theorem 11.7.5. Let X and Y be normed spaces, and let F W X ! Y be twice

Fréchet differentiable at xN 2 X with derivative rF .x/
N and second-order derivative
r 2 F .x/:
N Then for every uN 2 X;
N F .x//;
T 2 .graph .F /; .x; N .Nu; rF .x/.N
N u/// D f.Nv; rF .x/N
N v C r 2 F .x/.N
N u; uN //j vN 2 X g;
N F .x//;
A2 .graph .F /; .x; N .Nv; rF .x/.N
N v/// D f.Nv; rF .x/N
N v C r 2 F .x/.N
N u; uN //j vN 2 X g:
Proof. Let t > 0; and let uN ; x 2 X be arbitrary. By Taylor’s theorem, we have
t2
F .xN Ct uN Ct 2 x=2/ D F .x/CtrF
N N u/C
.x/.N N
.rF .x/.x/Cr 2
N u; uN //Co.t 2 /;
F .x/.N
2
and the claim follows at once from the above equality. t
u
Remark 11.7.6. In fact, if F W X ! Y is a single valued map which is twice
continuously Fréchet differentiable around xN 2 K X; then the second-order
contingent derivative of the restriction FK of F to K at xN in a direction uN is given
by the formula (see [18, p. 215]):
(
2 rF .x/.x/
N C r 2 F .x/.N
N u; uN / if x 2 T 2 .K; x;
N uN /;
N F .x/;
D FK .x; N uN ; rF .x/.N
N u//.x/ D
; if x 62 T 2 .K; x;
N uN /:
(11.85)
The following result is a second-order analogue of Theorem 11.1.12 given for
the first-order contingent derivative (see [18, Theorem 5.6.2]).
Theorem 11.7.7. Let X and Y be normed spaces, let L and S be nonempty subsets
of X where S is assumed to be open, let G W X Y be a set-valued map and let
f W S ! Y be a twice continuously differentiable single-valued map. We define a
set-valued map F W X Y by

f .x/ G.x/; if x 2 L;
F .x/ WD
;; if x 62 L:
Then for any xN 2 S \ dom .F /; .x;

N y/
N 2 graph .F /; and .Nu; vN / 2
N y//;
graph .DF.x; N we have
N y;
D 2 F .x; N uN ; vN /.x/ rf .x/.x/
N
1
C r 2 f .x/.N
N u; uN /
2
N f .x/
D 2 G.x; N y;
N uN ; rf .x/.N
N u/ vN /.x/; (11.86)
for every x 2 T 2 .L; x;

N uN /: Furthermore, D 2 F .x;
N y;
N uN ; vN /.x/ is empty whenever x …
2
N uN /: Equality holds in (11.86) if either L is twice derivable or G is twice
T .L; x;
derivable and G is Lipschitz.
Proof. Let .x; y/ 2 graph .D 2 F .x; N y;

N uN ; vN // be arbitrary. Then there are sequences
f.xn ; yn /g X Y and ftn g P such that .xn ; yn / ! .x; y/; tn # 0 and yN C tn vN C
tn2 yn 2 F .xN C tn uN C tn2 xn /; implying that
xN C tn uN C tn2 xn 2 L;
yN C tn vN C tn2 yn 2 f .xN C tn uN C tn2 xn / G.xN C tn uN C tn2 xn /;
where the first containment confirms that x 2 T 2 .L; x;N uN /: Moreover, since the map
N we have
f is twice continuously differentiable at x;

1 2
N n uN Ctn2 xn / D f .x/Ct
f .xCt N n rf . N
x/.N
u /Ctn
2
rf . N
x/.x/ C r f . N
x/.N
u ; N
u / C h.tn /
2
with h.tn / ! 0 as tn # 0: Therefore,
N yN C tn .rf .x/.N
f .x/ N u/ vN /

1 2
Ctn rf .x/.x/
2
N C r f .x/.NN u; uN / y C h.tn / 2 G.xN C tn uN C tn2 xn /
2
which implies that
1
rf .x/.x/
N C r 2 f .x/.N
N u; uN / y 2 D 2 G.x;
N f .x/
N y;
N uN ; rf .x/.N
N u/ vN /.x/;
2
(11.87)
and the inclusion (11.86) follows at once.

For the converse, we work under the additional assumption that G is
second-order derivable and Lipschitz, x 2 T 2 .L; x; N uN /; and that (11.87) holds.
Then there are sequences ftn g P with tn # 0 and f.xn ; yn /g X Y such that
tn ! 0; xn ! x and yn ! rf .x/.x/ N C r 2 f .x/.N
N u; uN / y such that for every
n 2 N; we have xN C tn uN C tn xn 2 L and
2
N yN C tn .rf .x/.N
f .x/ N u/ vN / C tn2 yn 2 G.xN C tn uN C tn2 xn /;
where we used the Lipschitz continuity of F . Then, for some h.tn / ! 0; the
sequence
1
zn WD rf .x/.x/
N C r 2 f .x/.N
N u; uN / C h.tn / yn
2
converges to y and satisfies yN Ctn vN Ctn2 zn 2 f .xN Ctn uN Ctn2 xn /G.xN Ctn uN Ctn2 xn /;
u
In the following, we collect some of the useful features of second-order

derivatives of set-valued maps. We give these results in terms of the contingent
derivatives but most of the results can easily be extended to other variants.
Proposition 11.7.8. Let X and Y be normed spaces, and let C Y be a pointed,
closed, and convex cone. Let F W X Y be a set-valued map, let .x; N y/
N 2
graph .F /; and let .Nu; vN / 2 X Y: Then for every x 2 dom .D 2 F .x;
N y;
N uN ; vN //;
the following inclusion holds:
N y;
D 2 F .x; N uN ; vN /.x/ C C D 2 .F C C /.x;
N y;
N uN ; vN /.x/: (11.88)
Proof. For an arbitrary x 2 dom .D 2 F .x; N y; N uN ; vN //; let y 2 D 2 F .x;

N y;
N uN ; vN /.x/ be
arbitrary. We also choose an arbitrary element c 2 C: Then there exist sequences
ftn g P and f.xn ; yn /g X Y such that tn # 0; .xn ; yn / ! .x; y/; and
yN C tn vN C tn2 yn 2 F .xN C tn uN C tn2 xn /: Setting yNn WD yn C c; we note that yN C tn vN C
tn2 yNn 2 F .xN C tn uN C tn2 xn / C C: Since the sequence fyNn g converges to y C c; we
conclude that y C c 2 D 2 .F C C /.x; N y;
N uN ; vN /.x/: This establishes (11.88) and the
u
The following variant of the above result is useful when working with the
second-order contingent epiderivatives which will be defined shortly.
Proposition 11.7.9. Let X and Y be normed spaces, and let C Y be a pointed,
N 2
graph .F /; and let .Nu; vN / 2 X Y: Then for every x 2 dom .D 2 .F CC /.x;
N y;
N uN ; vN //;
the following identity holds:
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ C C D D 2 .F C C /.x;
N y;
N uN ; vN /.x/: (11.89)
Proof. The inclusion
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ C C D 2 .F C C /.x;
N y;
N uN ; vN /.x/
holds trivially whereas the inclusion
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ C C D 2 .F C C /.x;
N y;
N uN ; vN /.x/
follows from Proposition 11.7.8 by using the identity .F C C /./ C C D .F C

C /./. u
t
The converse inclusion of (11.88) does not hold in general. Moreover, the
effective domains of the two derivatives can be very different. We illustrate this
in the context of the first-order derivative (the case where .Nu; vN / D .0; 0//.
Example 11.7.10. We set X D Y D R and C D RC : Define a set-valued map
F W R R as follows:

0 if x ¤ 0
F .x/ D
Œ1; 2 if x D 0:
Then dom .DF.0; 1// D f0g and
DF.0; 1/.0/ D RC ;
D.F C RC /.0; 1/.x/ D R for every x 2 R:
Consequently,
D.F C RC /.0; 1/.0/ 6 DF.0; 1/.0/ C RC :
It is worth noting that for every x 2 R; we have
Min.D.F C RC /.0; 1/.x/; RC / D ;;
whereas
Min.DF.0; 1/.0/; RC/ D f0g:
Therefore,
Min.D.F C RC /.0; 1/.0/; RC/ ¤ Min.DF.0; 1/.0/; RC /:
We remark that the set-valued map F is convex-valued.

Given the importance of the converse of the inclusion given in (11.88), we
introduce the following concept:
Definition 11.7.11. Let X and Y be normed spaces and let C Y be a suitable
ordering cone. A set-valued map F W X Y is called second-order contingently
N y/
C -absorbing at .x; N 2 graph .F / in direction .Nu; vN / 2 X Y if for every x 2
dom .D 2 .F C C /.x;
N y;
N uN ; vN //, we have
N y;
D 2 F .x; N uN ; vN /.x/ C C D D 2 .F C C /.x;
N y;
N uN ; vN /.x/: (11.90)
To prove identity (11.90), we need to impose certain restrictions on the map

F and/or on the ordering cone C . In the following, we will give several such
conditions. We begin by proposing a second-order extension of the S -derivative
that has an important role to play.
Definition 11.7.12. Let X and Y be normed spaces, and let C Y be a pointed,
N 2
graph .F /; and let .Nu; vN / 2 X Y: The second-order S -derivative of F at .x;
N y/
N
N y;
in direction .Nu; vN / is the set-valued map DS2 F .x; N uN ; vN / W X Y such that
y 2 DS2 F .x;N y;N uN ; vN /.x/ if and only if there are sequences f˛n g P, fˇn g P;
and f.xn ; yn /g graph .F / satisfying that
xn ! x;
N
˛n Œ.xn ; yn / .x;
N y/
N ! .Nu; vN /;
ˇn Œ˛n Œ.xn ; yn / .x;
N y/
N .Nu; vN / ! .x; y/:
We are now ready to give a result that ensures that a set-valued map is
second-order contingently C -absorbing.
N 2
graph .F /; and let .Nu; vN / 2 X Y: Assume that the set B WD fz 2 C j kzk D 1g is
compact. Assume that
N y;
DS2 F .x; N uN ; vN /.0/ \ .C =f0g/ D ;: (11.91)
Then the map F is second-order contingently C -absorbing, that is, for every
x 2 dom.D 2 .F C C /.x;
N y;
N uN ; vN //; the following identity holds
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ D D 2 F .x;
N y;
N uN ; vN /.x/ C C:
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ D 2 F .x;
N y;
N uN ; vN /.x/ C C:
For an arbitrary x 2 dom.D 2 .F C C /.x; N y;

N uN ; vN //; let y 2 D 2 .F C
N y;
C /.x; N uN ; vN /.x/ be arbitrary. Therefore, there are sequences ftn g P; fcn g C;
and f.xn ; yn /g X Y with tn # 0; .xn ; yn C cn / ! .x; y/ such that
yN C tn vN C tn2 yn 2 F .xN C tn uN C tn2 xn /: Since the set B is compact and C D cone.B/;
for some rn > 0 and for some bn 2 B; we have cn D rn bn : Without any loss of
generality, we assume that rn ! r 2 Œ0; 1 and bn ! b; for some b 2 B.
We claim that r < 1: For the sake of argument, we assume that rn ! 1: This,
in view of the convergence .yn C rn bn / ! y; implies that rn1 yn ! b; and hence

1 1
.xN C tn uN C tn2 xn ; yN C tn vN C tn2 yn / .x;
N y/
N .Nu; vN /
rn tn tn

xn yn
D ; ! .0; b/:
rn rn
Therefore b 2 DS2 F .x;

N y;
N uN ; vN /.0/ \ .C =f0g/ which is a contradiction to (11.91).
Hence r < 1: This ensures that y 2 D 2 F .x; N y;
N uN ; vN /.x/ C C: The proof is
complete. t
u
The following result is then immediate.
Corollary 11.7.14. Let X and Y be normed spaces where the space Y is of finite
dimension, and let C Y be a pointed, closed, and convex cone. Let F W X Y be
N y/
a set-valued map, let .x; N 2 graph .F /; and let .Nu; vN / 2 X Y: Assume that (11.91)
remains valid. Then (11.90) holds.
Proof. The proof follows by noticing that, for finite-dimensional Y; the set B D
fz 2 C j kzk D 1g is a compact base for the cone C . u
t
By setting .Nu; vN / D .0; 0/ in the above result, we deduce the following corollary:
Corollary 11.7.15. Let X and Y be normed spaces, let C Y be a pointed, closed,
and convex cone. Let F W X Y be a set-valued map and let .x;N y/
N 2 graph .F /:
Assume that the set B D fz 2 C j kzk D 1g is compact. Assume that
N y/.0/
DS F .x; N \ .C =f0g/ D ;:
Then for every x 2 dom .D.F C C /.x;

N y//;
N the following identity holds:
D.F C C /.x;
N y/.x/
N D DF.x;
N y/.x/
N C C:
Remark 11.7.16. We note that in Example 11.7.10, we have y 2 DS F .0; 1/.0/

for all y > 0, consistent with the above corollary.
In the following we extend two notions of compactness and employ them to
prove (11.90).
map, let .x; N y/
N 2 graph .F /; and let .Nu; vN / 2 X Y: The map F is called
second-order directionally compact at .x; N y/
N with respect to .Nu; vN / in a direction
x 2 X; if for every sequence ftn g P with tn # 0 and every sequence fxn g X
with xn ! x; any sequence fyn g Y contains a convergent subsequence provided
that yN C tn vN C tn2 yn 2 F .xN C tn uN C tn2 xn /:
Remark 11.7.18. If .Nu; vN / D .0; 0/; then the second-order directionally compact
map F at .x;N y/
N with respect to .Nu; vN / in a direction x 2 X; reduces to the notion
N y/
of directional compactness at .x; N in the direction x: The notion of directional
compactness, introduced in [39] is very important in set-valued optimization. The
above notion of second-order directional compactness was introduced in [323].
Second-order directional compactness holds, in particular, in the case where F is
single-valued, the Hadamard directional derivative
F .xN C tu/ F .x/

N
vN WD F 0 .xI
N uN / WD lim
t #0;u!Nu t
exists, and the parabolic second-order directional derivative

N tF 0 .xI
F .xN C t uN C t 2 w/ F .x/ N uN /
F 00 .x;
N uN I x/ WD lim 2
t #0;w!x t
exists. To see this, suppose that tn # 0, xn ! x; and yN C tn vN C tn2 yn 2 F .xN C tn uN C

tn2 xn /: Then
N tn F 0 .xI
F .xN C tn uN C tn2 xn / F .x/ N uN /
yn D 2
;
tn
so that yn ! F 00 .x;
N uN I x/.
With the concept of second-order directional compactness, we give another proof
of (11.90).
N 2
graph .F /; and let .Nu; vN / 2 X Y: Assume that F is second-order directionally
N y/
compact at .x; N with respect to .Nu; vN / in any direction x 2 X: Then (11.90) holds.
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ D 2 F .x;
N y;
N uN ; vN /.x/ C C:
Let y 2 D 2 .F CC /.x;
N y;
N uN ; vN /.x/: Then there exist sequences .tn / P; .xn ; yn /
X Y; and cn 2 C such that tn # 0; .xn ; yn / ! .x; y/; and
yN C tn vN C tn2 .yn cn =tn2 / D yN C tn vN C tn2 yn cn 2 F .xN C tn uN C tn2 xn /:
Since F is second-order directionally compact, we may assume that yn cn =tn2

converges to some yQ 2 Y: Hence yQ 2 D 2 F .x; N y;
N uN ; vN /.x/ and cn =tn2 ! .y y/
Q 2 C;
confirming that y 2 D F .x;
2
N y;
N uN ; vN /.x/ C C: The proof is complete. t
u
In the following definition, we give a notion of second-order compactly approx-
imable set-valued maps.
Definition 11.7.20. Let X and Y be normed spaces, and let C Y be a pointed,
N 2
graph .F /; and let .Nu; vN / 2 X Y: The map F W X Y is called second-order
compactly approximable at .x; N with respect to .Nu; vN /; if for each x 0 2 X; there
N y/
exists a set-valued map R from X into the set of all nonempty compact subsets of Y;
a neighborhood N .x/ N of xN in X; and a function r W0; 1 X !0; 1 such that
1. lim r.t; x/ D 0:
.t;x/!.0;x 0 /
2. For each x 2 X and t 20; 1; we have
F .xN C t uN C t 2 x/ yN C t vN C t 2 .R.x 0 / C r.t; x/BY /:

Remark 11.7.21. If .Nu; vN / D .0; 0/; then the second-order compactly approximable
N y/
map F at .x; N with respect to .Nu; vN / in a direction x 2 X; reduces to the notion of
the compactly approximable map at .x; N y/
N in the direction x 2 X used in Taa [548].
We remark that set-valued compactly approximable map were originally introduced
by Amahroq and Thibault [6].
The following result is obtained by the above concept of second-order compactly
approximable maps.
N 2
graph .F /; and let .Nu; vN / 2 X Y: If the map F is second-order compactly
N y/
approximable at .x; N with respect to .Nu; vN /; then (11.90) holds.
Proof. For x 2 dom.D 2 .F C C /.x; N y;
N uN ; vN //; let y 2 D 2 .F C C /.x;
N y;
N uN ; vN /.x/:
Hence there exist sequences ftn g P and f.xn ; yn /g X Y such that tn #
0; .xn ; yn / ! .x; y/ and yN C tn vN C tn2 yn 2 F .xN C tn uN C tn2 xn / C C:
Then yN C tn vN C tn2 yn D zn C cn with zn 2 F .xN C tn uN C tn2 xn / and cn 2 C:
Since F is second-order compactly approximable, there exists fkn g R.x/;
rn WD r.tn ; xn /; bn 2 BY and n0 2 N such that rn ! 0; and zn D yN C tn vN C tn2 .kn C
rn bn / for all n > n0 : This implies yN C tn vN C tn2 .kn C rn bn / 2 F .xN C tn uN C tn2 xn /
for all n > n0 :
In view of the compactness of R.x/; we may assume that kn ! k 2 R.x/.
Because of the fact that .kn C rn bn / ! k; we deduce that k 2 D 2 F .x; N y;
N uN ; vN /.x/:
Since yN C tn vN C tn2 yn D yN C tn vN C tn2 .kn C rn bn / C cn for sufficiently large n; we
have yn kn rn bn D tn2 cn 2 C: By passing to the limit, we have y k 2 C and
hence y 2 k C C D 2 F .x; N y;
N uN ; vN /.x/ C C: The proof is complete. t
u
In view of Lemma 11.5.19, we obtain the following result.
N 2
graph .F /; and let .Nu; vN / 2 X Y: Assume that F is second-order contingently
N y/
absorbing at .x; N in direction .Nu; vN /: Then for every x 2 dom.D 2 F .x;
N y;
N uN ; vN //;
we have
1. GHe-PMin.D 2 .F C C /.x; N y; N uN ; vN /.x/; C / D GHe-PMin.D 2 F .x; N y;
N uN ; vN /
.x/; C /:
2. Min.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C / D Min.D 2 F .x; N y;
N uN ; vN /.x/; C /:
3. WMin.D .F C C /.x;
2
N y;
N uN ; vN /.x/; C / WMin.D 2 F .x; N y;N uN ; vN /.x/; C /.
Proof. Since for every x 2 dom .D 2 F .x;
N y;
N uN ; vN //; we have
N y;
D 2 F .x; N uN ; vN /.x/ C C D D 2 .F C C /.x;
N y;
N uN ; vN /.x/;
N y;
the proof follows directly from Lemma 11.5.19 by setting A D D 2 F .x; N uN ; vN /.x/;
and B D D 2 .F C C /.x; N y;
N uN ; vN /.x/. t
u
The following result is an analogue of Theorem 11.7.23 without (11.90):

N 2
graph .F /; and let .Nu; vN / 2 X Y: Assume that the set B D fz 2 C j kzk D 1g is
compact. Then for every x 2 dom .D 2 .F C C /.x; N y;
N uN ; vN //; the following inclusion
holds
Min.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C / D 2 F .x;
N y;
N uN ; vN /.x/: (11.92)
Proof. For x 2 dom.D 2 .F C C /.x; N y;

N uN ; vN //; let y 2 Min.D 2 .F C
N y;
C /.x; N uN ; vN /.x/; C / be arbitrary. Then y 2 D 2 .F C C /.x; N y;
N uN ; vN /.x/; and
hence there exist sequences .tn / P; .xn ; yn / X Y; and cn 2 C such
that tn # 0; .xn ; yn / ! .x; y/; and yN C tn vN C tn2 .yn cn / 2 F .xN C tn uN C tn2 xn /:
Since C D cone.B/; we have cn D rn bn for some rn > 0 and bn 2 B: Moreover,
because B is compact, we may assume that bn ! b 2 B: We will show that rn ! 0:
If not, we may assume, taking a subsequence if necessary, that there exists " > 0
such that rn > ": We set cNn D "cn =rn: Notice that cn cNn 2 C; from which it
follows that yN C tn vN C tn2 .yn cNn / 2 F .xN C tn uN C tn2 xn / C C:
Since cNn ! "b; we have that y "b 2 D 2 .F C C /.x; N y;
N uN ; vN /.x/ contradicting
the C -minimality of y: Therefore rn ! 0; and this implies that .yn cn / ! y:
Consequently, y 2 D 2 F .x; N y;
N uN ; vN /.x/; and the proof is complete. t
u
By setting .Nu; vN / D .0; 0/ in the above result, we recover the following result.
Corollary 11.7.25. Let X and Y be normed spaces, let C Y be a pointed, closed,
and convex cone, let F W X Y be a set-valued map, and let .x; N y/
N 2 graph .F /
be arbitrary. Assume that the set B D fz 2 C j kzk D 1g is compact. Then for every
x 2 dom .D.F C C /.x; N y//;
N the following inclusion holds
Min.D.F C C /.x;
N y/.x/;
N C / DF.x;
N y/.x/:
N
Example 11.7.26. As an illustration of Theorem 11.7.24 and Corollary 11.7.25, let

F W R R2 be defined by F .x/ WD f.y1 ; y2 / 2 R2C j y1 y2 D xg; and let C D R2C ,
xN D 1, and yN D .1; 1/. Then for all x 2 R,
N y/.x/
DF.x; N D f.y1 ; y2 / j x D y1 C y2 g;
D.F C C /.x;
N y/.x/
N D f.y1 ; y2 / j x y1 C y2 g;
so that MinC D.F C C /.x; N y/.x/

N D DF.x; N y/.x/;
N consistent with Corol-
lary 11.7.25.
For vN D .v1 ; v2 / and uN D v1 C v2 , one obtains (e.g., by Proposition 2.3 of [594])
N y;
D 2 F .x; N uN ; vN /.x/ D f.y1 ; y2 / j x D y1 C y2 C v1 v2 g;
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ D f.y1 ; y2 / j x y1 C y2 C v1 v2 g;
so that equality holds in (11.92).

11.8 Calculus Rules for Second-Order Contingent

Derivatives
We have seen in Sect. 11.2 of this chapter that a quite elegant approach for
various calculus rules for derivatives of set-valued maps can be developed by
exploiting certain properties of the involved tangent cones. In this section, we
propose a second-order extension of this approach and obtain calculus rules for
the second-order graphical derivatives. Following the scheme of Sect. 11.2, we
present abstract chain rules and sum formulas and then derive some particular
cases of interest. This section is influenced by Aubin and Frankowska [18] and
Ward [594, 596].
We begin with the following general scheme for an abstract chain rule:
Y Z be set-valued maps, let .x; N y/
N 2 graph .F /; let .y;
N zN/ 2 graph .G/; and let
.Nu; vN ; w/
N 2 X Y Z: Assume that P 2 and Q2 are second-order closed tangent sets
such that for the maps ` W X Y Y Z ! X Z and g W X Y Y Z ! Y;
given by
`.a; b; c; d / WD .a; d /; (11.93)

g.a; b; c; d / WD b c; (11.94)
and for S1 WD graph .F /; S2 WD graph .G/; S WD S1 S2 ; x0 WD .x;

N y;
N y;
N zN/; and
u0 WD .Nu; vN ; vN ; w/;
N we have
`.P 2 .S; x0 ; u0 // P 2 .`.S /; `.x0 /; `.u0 //; (11.95)

2
N y/;
P .S1 ; .x; N .Nu; vN // Q .S2 ; .y;
2
N zN/; .Nv; w//
N P .S1 S2 ; .x0 ; u0 //;
2
(11.96)
P 2 .S; x0 ; u0 / \ rg.x0 /1 .0/ P 2 .S \ g 1 .0/; x0 ; u0 /: (11.97)
N zN; vN ; w/
Then, for DP2 G.y; N and DQ2 F .x;
N y;
N uN ; vN /; we have
N zN; vN ; w/
DQ2 G.y; N ı DP2 F .x;
N y;
N uN ; vN / DP2 .G ı F / .x;
N zN; uN ; w/:
N (11.98)
Proof. Following the reasoning used in Theorem 11.2.13, we have

graph DP2 .G ı F / .x;
N zN; uN ; w/
N
D P 2 .graph .G ı F /; .x;
N zN/; .Nu; w//
N

D P .` Œgraph .F / graph .G/ \ g 1 .0/ ; .x;
2
N zN/; .Nu; w//N
2
` P .Œgraph .F / graph .G/ \ g 1 .0/; .x; N y;N y; N zN/ ; .Nu; vN ; vN ; w/
N

` P 2 .graph .F / graph .G/; .x; N y;
N y;
N zN/ ; .Nu; vN ; vN ; w//
N
11.8 Calculus Rules for Second-Order Contingent Derivatives 501

\ rg.x;
N y; N zN/1 .0/
N y;

N y/;
` P 2 .graph .F /; .x; N .Nu; vN //

Q2 .graph .G/; .y; N \ g 1 .0/
N zN/; .Nv; w//

D ` graph .DP2 F .x;
N y; N uN ; vN // graph .DQ G.y; N \ g 1 .0/
N zN; vN ; w//

D graph DQ2 G.y;N zN; vN ; w/
N ı DP2 F .x; N y;
N uN ; vN / ;

u
The above result depicts that a chain rule can be given by using a combination
of second-order tangent sets satisfying (11.95)–(11.97). In the following, we collect
a few results and identify the conditions where these requirements are met. The
following result was given by Ward [594]:
Theorem 11.8.2. Let X , Y; and Z be finite dimensional, let F W X Y and
G W Y Z be set-valued maps, let .x; N y/
N zN/ 2 graph .G/; and
let .Nu; vN ; w/
N 2 X Y Z: Assume that F is closed near .x; N y/;
N G is closed near
N zN/ and the following condition holds:
.y;
N y//
range.DC F .x; N dom .DC G.y;
N zN// D Y: (11.99)
N zN; vN ; w/
DA2 G.y; N ı D 2 F .x;
N y;
N uN ; vN / D 2 .G ı F /.x;
N zN; uN ; w/;
N (11.100)
N zN; vN ; w/
DA2 G.y; N ı DA2 F .x;
N y;
N uN ; vN / DA2 .G ı F /.x;
N zN; uN ; w/;
N (11.101)
2
N zN; vN ; w/
D G.y; N ı DA F .x;
2
N y;
N uN ; vN / D .G ı F /.x;
2
N zN; uN ; w/:
N (11.102)
Proof. The proof follows by applying Theorem 4.7.7 to f0g and graph .F /
graph .G/: We have
rg.x;
N y;
N y;
N zN/C.graph .F / graph .G/; .x;
N y;
N y;
N zN// C.0; 0/
D range.DC F .x;
N y//
N dom .DC G.y;
N zN//;
and hence the required properties hold. t

u
We now give a general scheme for an abstract sum formula.
set-valued maps, let f W X ! X be a linear map, let .x; N yN1 / 2 graph .F1 /; and
N yN2 / 2 graph .F2 ı f /: Assume that P 2 and Q2 are closed second-order
let .x;
tangent sets such that for the maps ` W .X Y /2 ! X Y and g W .X Y /2 ! X
defined by
`.a; b; c; d / WD .a; c C d /; (11.103)

g.a; b; c; d / WD f .a/ c; (11.104)
and for S1 WD graph .F1 /; S2 WD graph .F2 /; S WD S1 S2 ; x0 WD .x;

N yN1 ; f .x/;
N yN2 /;
and u0 WD .Nu; vN 1 ; f .Nu/; vN 2 /; we have
`.P 2 .S; x0 ; u0 // P 2 .`.S /; `.x0 /; `.u0 //;

N yN1 /; .Nu; vN 1 // Q2 .S2 ; .f .x/;
P 2 .S1 ; .x; N yN2 /; .f .Nu/; vN 2 // P 2 .S1 S2 ; .x0 ; u0 //;
P 2 .S; x0 ; uN 0 / \ rg.x0 /1 .0/ P 2 .S \ g 1 .0/; x0 ; u0 /:
Then the following inclusion holds:
N yN1 ; uN ; vN 1 /./ C DQ2 F2 .f .x/;

DP2 F1 .x; N yN2 ; f .Nu/; vN 2 /.f .//
DP2 .F1 C F2 ı f / .x;
N yN1 C yN2 ; f .Nu/; vN 1 C vN 2 /./:
Proof. Proceeding as in the proof of Theorem 11.8.1, we have

graph DP2 .F1 C F2 ı f / .x;
N yN1 C yN2 ; uN ; vN 1 C vN 2 /
D P 2 .graph .F1 C F2 ı f /; .x;

N yN1 C yN2 /; .Nu; vN 1 C vN 2 //
D P 2 .` .Œgraph .F1 /

graph .F2 / \ g 1 .0/ ; `.x; N yN1 ; f .x/;
N yN2 /; `.Nu; vN 1 ; f .Nu/; vN 2 //
2
` P .Œgraph .F1 /

graph .F2 / \ g 1 .0/; .x; N yN1 ; f .x/;
N yN2 / ; .Nu; vN 1 ; f .Nu/; vN 2 /

` P 2 .graph .F1 /
\
N yN1 ; f .x/;
graph .F2 /; .x; N yN2 /; .Nu; vN 1 ; f .Nu/; vN 2 // rg.x; N yN2 /1 .0/
N yN1 ; f .x/;

N yN1 /; .Nu; vN 1 //
` P 2 .graph .F1 /; .x;
\
Q2 .graph .F2 /; .f .x/;
N yN2 /; .f .Nu/; vN 2 // rg.x; N yN1 ; f .x/;N yN2 /1 .0/

N yN1 ; uN ; vN 1 //
D ` graph .DP2 F1 .x;
\
graph .DQ2 F2 .f .x/;
N yN2 ; f .Nu/; vN 2 // rg.x; N yN1 ; f .x/;N yN2 /1 .0/
N yN1 ; uN ; vN1 //
D `f.a; b; c; d /j .a; b/ 2 graph .DP2 F1 .x;
N yN2 ; f .Nu/; vN 2 //; rf .x/.a/
.c; d / 2 graph .DQ2 F2 .f .x/; N D cg;
N yN1 /; .Nu; vN 1 //;
D f.a; b C d /j .a; b/ 2 graph .DP2 F1 .x;
11.8 Calculus Rules for Second-Order Contingent Derivatives 503
.c; d / 2 graph .DP2 F2 .f .x/; N yN1 /; .f .Nu/; vN 1 /; f .a/ D cg;

2
D graph DP F1 .x; N yN1 ; uN ; vN 1 // C DQ2 F2 .f .x/; N yN2 ; f .Nu/; vN 2 /.f ./ ;
where we used the fact that rf .x/.x/

N D f .x/: The proof is complete. t
u
The following is a second-order sum formula for the second-order contingent
derivatives. This is an analogue of Theorem 11.2.19 and was recently given in [155].
Theorem 11.8.4. Let X and Y be Banach spaces, let F1 ; F2 W X Y be closed
set-valued maps, let f W X ! X be a linear, let .x; N yN1 / 2 graph .F1 /; and let
N yN2 / 2 graph .F2 ı f /: Assume that the map g W X Y X Y ! X given by
.x;
g.a; b; c; d / D f .a/ c is metrically subregular at .x;
N yN1 ; f .x/;
N yN2 ; 0/ with respect
to graph .F1 / graph .F2 /: Assume that either graph .F1 / is second-order derivable
N yN1 / in any direction or graph .F2 / is second-order derivable at .f .x/;
at .x; N yN2 / in
any direction.
Then the following sum formula, for the second-order contingent derivatives,
holds:
N yN1 ; uN ; vN 1 /./ C D 2 F2 .f .x/;

D 2 F1 .x; N yN2 ; f .Nu/; vN 2 /.f .//
D 2 .F1 C F2 ı f / .x;
N yN1 C yN2 ; f .Nu/; vN 1 C vN 2 /./:
We conclude this section by a second-order analogue of Theorem 11.2.7 given

by Ward [596]. This results employs the notion of the second-order directional
compact:
Y Z be set-valued maps, let .x;N y/
N zN/ 2 graph .G/; and let
.Nu; vN ; w/
N 2 X Y Z: Define H W X Z be
H.x; z/ WD fy 2 Y j y 2 F .x/; z 2 G.y/g:
Let H be second-order directionally compact at ..x; N zN/; y/

N with respect to
N vN / in the direction .x; z/ 2 T 2 .graph .G ı F //; .x;
..Nu; w/; N zN/; .Nu; w//:
N Then
.x; z/ 2 graph .D 2 G.y;

N zN; vN ; w/
N ı D 2 F .x;
N y;
N uN ; vN //:
Proof. Let .x; z/ 2 graph .D 2 .GıF /.x; N zN; uN ; w//

N with H second-order directionally
compact at ..x; N zN/; y/
N with respect to ..Nu; w/; N vN / in the direction .x; z/: Then there
exist sequences ftn g P and f.xn ; zn /g with tn # 0; .xn ; zn / ! .x; z/ such that
.Nz C tn wN C tn2 z2n / 2 .G ı F /.xN C tn uN C tn2 xn / for every n 2 N: Therefore, there exists
fwn g such that wn 2 F .xN C tn uN C tn2 xn / and zN C tn wN C tn2 z2n 2 G.wn /: By setting
yn WD tn1 .wn yN tn vN /; we have yN Ctn vN Ctn2 yn 2 H.xN Ctn uN Ctn2 xn ; zN Ctn wCt
N n2 z2n /:
Now, by invoking the second-order directional compactness, we may assume, by
taking subsequence if necessary, that fyn g ! y for some y 2 Y: Then .x; y/ 2
N y/;
T 2 .graph .F /; .x; N .Nu; vN // and .y; z/ 2 T 2 .graph .G/; .y; N zN/; .Nv; zN// implying that
.x; z/ 2 graph .D G.y;
2
N zN; vN ; w/
N ı D 2 F .x;
N y;
N uN ; vN //: The proof is complete. t
u
11.9 Second-Order Epiderivatives of Set-Valued Maps
In this section, we study second-order epiderivatives of set-valued maps. We begin

with the following:
closed and convex cone. Let F W X Y be a set-valued map, let .x;N y/
N 2
graph .F /; and let .Nu; vN / 2 X Y .
N y;
1. A single-valued map DE2 F .x; N uN ; vN / W X ! Y is called the second-order
N y/
contingent epiderivative of F at .x; N in the direction .Nu; vN /; if
N y;
epi .DE2 F .x; N uN ; vN // D T 2 .epi .F /; .x;
N y/;
N .Nu; vN //;
N y/;
where T 2 .epi .F /; .x; N .Nu; vN // is the second-order contingent set.
2. A set-valued map DG2 F .x; N y;N uN ; vN / W X Y is called the second-order
generalized contingent epiderivative of F at .x; N y/
N in direction .Nu; vN /; if
N y;
DG2 F .x; N uN ; vN /.x/ D Min.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C /
x 2 dom .D 2 .F C C /.x;
N y;
N uN ; vN //;
where D 2 .F C C /.x;
N y;
N uN ; vN // is the second-order contingent derivative.
3. A single-valued map D2E F .x; N y;
N uN ; vN / W X ! Y is called the second-order
asymptotic contingent epiderivative of F at .x; N y/
N y;
epi .D2E F .x; N uN ; vN // D T2 .epi .F /; .x;
N y/;
N .Nu; vN //;
where T2 .epi .F /; .x;

N y/;
N .Nu; vN // is the second-order asymptotic contingent cone.
4. A set-valued map D2G F .x; N y;N uN ; vN / W X Y is called the second-order
generalized asymptotic contingent epiderivative of F at .x; N y/
N in direction
.Nu; vN /; if
D2G F .x;
N y;
N uN ; vN /.x/ D Min.D2 .F C C /.x;
N y;
N uN ; vN /.x/; C /
x 2 dom .D2 .F C C /.x;
N y;
N uN ; vN //;
where D2 .F C C /.x;
N y;
N uN ; vN // is the second-order asymptotic contingent deriva-
tive.
5. A single-valued map DE2 F .x; N y;
N uN ; vN / W X ! Y is called the second-order
compound contingent epiderivative of F at .x; N y/
N in the direction .Nu; vN / if
11.9 Second-Order Epiderivatives of Set-Valued Maps 505
N y;
epi .DE2 F .x; N uN ; vN // D T .T ..epi .F /; .x;
N y//;
N .Nu; vN //;
N y//;
where T .T ..epi .F /; .x; N .Nu; vN // is the contingent cone of T .epi .F /; .x;
N y//
N
at .Nu; vN /:
6. A set-valued map DG2 F .x; N y;
compound generalized contingent epiderivative of F at .x; N y/N in direction
.Nu; vN / if
DG2 F .x;
N y;
N uN ; vN /.x/ D Min.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C /
x 2 dom .D 2 .F C C /.x;
N y;
N uN ; vN //;
where D 2 .F C C /.x;
N y;
N uN ; vN // is the second-order compound contingent deriva-
tive.
Remark 11.9.2. If, in the above .Nu; vN / D .0; 0/; then we recover the contingent
N y/
epiderivative DE F .x; N and the generalized contingent epiderivative DG F .x;N y/
N
N y/.
of F at .x; N Clearly, several analogues of the above second-order epiderivatives
can be defined by using other notions of second-order tangent cones/sets.
The following result clarifies the relationship between the second-order contin-
gent epiderivative, the second-order compound contingent epiderivative, and the
second-order asymptotic contingent epiderivative.
Proposition 11.9.3. Let X and Y be normed spaces and let C Y be a
pointed, closed and convex cone. Let F W X Y be a set-valued map, let
N y/
.x; N 2 graph .F /; and let .Nu; vN / 2 X Y . Assume that the map F is C -
convex. Then the second-order compound epiderivative DE2 .x; N y;
N uN ; vN / coincides
N y;
with the second-order contingent derivative D 2 F .x; N uN ; vN / of .F C C / provided
that 0 2 D 2 .F C C /.x;
N y;
N uN ; vN /.0/:
Proof. The proof follows from the fact that, for a convex subset S of a
normed space, the equality T 2 .S; x; y/ D T .T .S; x/; y/ holds provided that
0 2 T 2 .S; x; y/. t
u
The following example illustrates the above notions.
F .x/ WD fy 2 R j y
x 4 g for all x 2 R:
N y/
Let .x; N D .0; 0/ and let .Nu; vN / D .1; 0/. Then:

T 2 .epi .F /; .0; 0/; .1; 0// D z 2 R2 j9 .zn / ! z; .
n / # 0 W
n .1; 0/

2n
C zn 2 epi .F / :
2

2n
The condition
n .1; 0/ C 2 n
z 2 epi .F / is equivalent to

2n
2
n
zn2
.
n C n zn1 /4 D
4n .1 C zn1 /4 ;
2 2 2
and hence

n
zn2
2
2n .1 C zn /4 :
2 1
Since zn1 ! z1 ; zn2 ! z2 and

n # 0; we obtain that
T 2 .epi .F /; .0; 0/; .1; 0/ D R RC ;
implying that
D 2 FC .0; 0; 1; 0/.x/ D RC for every x 2 R:
On the other hand,
D 2 F .0; 0; 1; 0/.x/ D 0 for every x 2 R;

DG2 F .0; 0; 1; 0/.x/ D Min.D 2 FC .0; 0; 1; 0/.x/; RC/ D f0g for every x 2 R:
The following examples will further clarify the differences and the similarities
between two of the above notions.
Example 11.9.5. Let F W R R be a set-valued map given by:
F .x/ WD fy 2 R j y
x 3=2 g for all x 2 R:
N y/
Let .x; N D .0; 0/ and let .Nu; vN / D .1; 0/. Define RC D fr 2 Rj r
0g.
Then
D 2 .F C RC /.0; 0; 1; 0/.x/ D RC for every x 2 R;
whereas the second-order contingent derivative is not defined.

The following is an existence theorem for the second-order asymptotic contin-
gent epiderivatives. Analogous results for other derivatives can be given by repeating
the arguments.
Theorem 11.9.6. Let X and Y be normed spaces and let C Y be a proper,
closed, convex and regular cone. Let F W X Y be a set-valued map, let
N y/
.x; N 2 graph .F / and let .Nu; vN / 2 X Y: Let for every x 2 A WD dom .D2 .F C
N y;
C /.x; N uN ; vN //; the set D2 .F C C /.x;
N y;
N uN ; vN /.x/ have a C -lower bound. Then
11.9 Second-Order Epiderivatives of Set-Valued Maps 507
D2G F .x;
N y;
N uN ; vN /.x/ ¤ ;; for every x 2 A. Moreover the following relationships
hold:
D2 .F C C /.x;
N y;
N uN ; vN /.x/ D2G F .x;
N y;
N uN ; vN /.x/ C C for every x 2 A:
(11.105)
2
N y;
epi .DG F .x; N uN ; vN // D T .epi .F /; .x;
2
N y/;
N .Nu; vN //: (11.106)
N y;
Proof. For the existence of the generalized asymptotic epiderivative D2G F .x; N
uN ; vN /.x/ and for (11.105) we used arguments analogous to those used for the
first-order case. For this, we define a set-valued map H W X Y by
H.x/ WD D2 .F C C /.x;
N y;
N uN ; vN /.x/ for every x 2 A:
Since the second-order asymptotic contingent cone is always closed, the set H.x/
is closed for every x 2 dom .H / D A. It has a C -lower bound by the assumption.
N y;
Therefore, the existence of D2G F .x; N uN ; vN /.x/ now follows from Proposition 11.5.5,
which ensures the existence of minimal points for H.x/ with respect to the convex
cone C . Moreover, it follows from the same result that H.x/ has the domination
property, that is
H.x/ D2G F .x;

N y;
N uN ; vN /.x/ C C;
and hence (11.105) is proved.

We claim that equality holds in the above inclusion. For this, let y 2
D2G F .x;
N y;
N uN ; vN /.x/ C C be arbitrary. Then y 2 H.x/ C C: Now making use
of the fact that H.x/ D H.x/ C C , we obtain y 2 H.x/. This confirms the
inclusion
N y;
H.x/ D2G F .x; N uN ; vN /.x/ C C:
Therefore
D2G F .x;
N y;
N uN ; vN /.x/ C C D H.x/ D D2G F .x;
N y;
N uN ; vN /.x/:
Furthermore, the above identity and the definition of the second-order asymptotic
contingent derivative implies
N y;
epi .D2G F .x; N uN ; vN // D graph .D2 .F C C /.x;
N y;
N uN ; vN //
D T2 .graph .F C C /; .x;
N y/;
N .Nu; vN //
D T2 .epi .F /; .x;
N y/;
N .Nu; vN //;
which confirms the claim. t

u
Remark 11.9.7. Numerous results given for the first-order contingent epiderivatives
in this chapter have their second-order analogues which can be proved by employing
similar arguments. Some of such results are available in [300].
Chapter 12
Optimality Conditions in Set-Valued
Optimization
Let X and Y be normed spaces, let S X be a nonempty set, let C Y be a cone

inducing a partial ordering in Y , and let F W X Y be a set-valued map. We begin
this chapter with optimization problems that can be written in the following form
Minimize F .x/ subject to x 2 S; (12.1)
where the minimum is taken in the vector optimization sense. For instance, one
N y/
option is to seek an element .x; N 2 graph .F / such that F .S / \ .fyg
N C / D fyg;
N
where F .S / WD fF .x/j x 2 S g. Clearly, such a minimizer of (12.1) is a point
N y/
.x; N 2 X Y that satisfies yN 2 F .x/
N \ Min.F .S /; C /.
Note that if the map F is single-valued, then (12.1) reduces to a vector
optimization problem. Additionally, if Y WD R and C WD RC ; then we are in the
framework of classical optimization.
Numerous other possibilities for choosing a minimizer of (12.1) have been
thoroughly discussed in Chap. 2 and will be considered in this chapter. This includes
the proper minimizer, the weak-minimizer, and the strong-minimizer, among others.
In this chapter, unless stated otherwise, the proper minimizer refers to the Henig
global proper minimizer.
The problems of type (12.1) belong to the realm of set-valued optimization. This
chapter is devoted to necessary and sufficient optimality conditions of first-order
and second-order for set-valued optimization problems. We note that primarily there
are two approaches for giving optimality conditions in set-valued optimization. The
first approach uses suitable derivatives of the involved set-valued maps whereas
the second approach yields optimality conditions by means of alternative-type
theorems. Although interesting results have been obtained by means of alternative-
type theorems, the focus of this chapter is mainly on obtaining optimality conditions
using some suitable derivatives and epiderivatives of the involved set-valued maps.
Before describing the approach adopted in this chapter and giving the main
results, let us first recall some of the major developments in the context of

DOI 10.1007/978-3-642-54265-7__12
510 12 Optimality Conditions in Set-Valued Optimization
derivative-based optimality conditions in set-valued optimization. We recall that

set-valued optimization problems explicitly appear in the interesting paper by
Corley [109], where the contingent derivative and the circatangent derivative were
employed to give general optimality conditions. However, it should be pointed out
that already in 1982, W. Oettli in [445] studied optimization problems with set-
valued data. Moreover, S. Tagawa in [554] also investigated optimization problems
with set-valued data, and, in fact, introduced derivatives of set-valued maps.
Tagawa’s work was based on his Ph. D. thesis supervised by W. Oettli (see [553]).
Due to the fact that the theory of variational analysis has been enriched by various
notions of tangent cones, many extensions of Corley’s results have been given using
other approximating notions. For instance, Taa [548] replaced the contingent cone
by the closed radial cone (see also Flores-Bazan [188] and Kasimbeyli [314]).
As another direction of generalization and strengthening of the known results
in set-valued optimization, some authors considered more general optimization
problems, but still used the contingent derivative as the derivative notion to give
optimality conditions (see Luc [400]). Starting from Corley’s work, many subse-
quent contributions revolved around the graphs of the involved set-valued maps.
Although this idea gives suitable necessary optimality conditions, the sufficient
optimality conditions demand that graph of the involved set-valued map be convex.
This, however, is a quite stringent assumption. Of course, a quick remedy is to
work with the profile map of the involved set-valued maps. In fact, Corley [109]
himself used this idea. However, the profile map does not seem to be a natural
candidate for studying optimality conditions, particularly when the interest lies also
in analyzing particular cases of single-valued maps. Motivated by this unsatisfactory
situation, an attractive alternative approach was proposed by Jahn and Rauh [301]
where the optimality conditions were given using contingent epiderivatives. Here
the main idea is to use a certain part of the boundary of the epigraphs of the
involved set-valued maps (see also Ward [594]). Since [301], the epigraph-based
approach has been rigorously pursued by many authors ( see, for instance, Chen
and Jahn [92], Sama [519] and [296, 297]). Some of the recent developments in set-
valued optimization can be found in [8, 88, 116, 117, 318, 319, 325, 326, 361, 399],
and the cited references therein. See also [450–452].
During the last several decades, the so-called Dubovitskii-Milyutin approach has
been used to study various optimal control and optimization problems with ordinary
or partial differential equations as constraints. Several authors have also worked
on extending this approach to nonsmooth optimization problems. More recently, in
[284], an extension of the Dubovitskii-Milyutin approach to set-valued optimization
problems was proposed, where first-order necessary optimality conditions for the
notion of the proper-minimality, the weak-minimality, and the strong-minimality
were given.
Another fruitful approach in set-valued optimization is based on various notions
of coderivatives and has attracted a great deal of attention in recent years. Abdouni
and Thibault [2] were the ones who initiated this study by considering a set-valued
12 Optimality Conditions in Set-Valued Optimization 511
constrained optimization problem. They established optimality conditions with

Lagrange-Kuhn-Tucker and Lagrange-Fritz-John multipliers in terms of the
coderivatives of the involved set-valued mapping. Since the appearance of [2],
numerous authors have employed coderivatives to optimality conditions for various
set-valued optimization problems (see [27, 151, 620, 621]). In this chapter, such
optimality conditions will be discussed thoroughly.
All of the aforementioned results deal with first-order optimality conditions.
Although many new refinements and interesting ideas related to the first-order
optimality conditions in set-valued optimization are still in the making, recent
developments in non-smooth scalar and vector optimization have shown an acute
interest in the development of second-order or higher-order optimality condi-
tions. This remains particularly evident for vector optimization problems (see
the interesting work [304] and the cited references therein). Motivated by these
developments, in [300], the second-order contingent epiderivatives were intro-
duced and employed to give new second-order optimality conditions in set-valued
optimization. These results were further refined in [313], where the second-order
asymptotic epiderivatives were used. Some extensions of these and related results
for higher-order optimality conditions are given in [586, 587]. An extension of the
first-order Dubovitskii-Milyutin approach (see [148]) to set-valued optimization
with an emphasis on the case of multi-equality constraints, and new first-order
and second-order optimality conditions for several notions of optimality have been
recently given in [321, 322].
This chapter covers a wide variety of optimality conditions in set-valued opti-
mization. To be specific, the following directions of deriving optimality conditions
are investigated in sufficient details:
• Necessary first-order and second-order optimality conditions in set-valued opti-
mization by a direct approach using graphical derivatives and epiderivatives.
• Necessary first-order and second-order optimality conditions in set-valued opti-
mization by the Dubovitskii-Milyutin approach using graphical derivatives and
epiderivatives.
• Sufficient optimality conditions using graphical derivatives and epiderivatives.
• Optimality conditions for Q-minimizers using graphical derivatives and
coderivatives.
• Lagrange multiplier rules based on limiting subdifferentials.
• Necessary optimality conditions for the notions of approximate solutions for set-
valued optimization problems.
• Necessary and sufficient optimality conditions in set-valued optimization for
solution concepts based on the set approach.
• Necessary optimality conditions for solution concepts with respect to a general
preference relation.
• Set-valued optimization problems from the view of KKT-Points and correspond-
ing stability results.

by the Direct Approach
We begin with various necessary and sufficient optimality conditions of first-order.

In set-valued optimization, a common approach is to give necessary optimality
conditions by showing a disjunction of certain sets in the image space. We refer
to this approach as the direct approach. As a derivative notion, we choose the
contingent derivative of the profile maps of the involved set-valued maps. However,
we emphasize that the used arguments easily yield optimality conditions for several
other notions of derivatives and epiderivatives of set-valued maps.
We now formulate the set-valued optimization problems we intend to explore.
Let W; X; Y; and Z be normed spaces, and let C Y and D Z be pointed,
closed, and convex cones. Let S X be a nonempty set and let h 2 W be a given
element. Let F W X Y; G W X Z; and H W X W be given set-valued maps.
We are interested in studying the following three set-valued optimization prob-
lems:
.P0 / Minimize F .x/ subject to x 2 S:

.P1 / Minimize F .x/ subject to x 2 S1 WD fx 2 S j G.x/ \ D ¤ ;g:
.P2 / Minimize F .x/ subject to x 2 S2 WD fx 2 S j G.x/ \ D ¤ ;; h 2 H.x/g:
N y/
Recall that .x; N 2 graph .F / is a minimizer of set-valued optimization problem
(P2 ), if and only if, F .S2 / \ .yN C / D fyg.
N Analogous relationships are valid for
other kinds of minimizers and have been defined earlier in Chap. 2. Observe that
(P2 ) reduces to (P0 ), if G.x/ D 0 and H.x/ D h uniformly on S . In this case,
the set S defines abstract constraints. If, additionally, we have S D X , then (P2 ) is
an unconstrained set-valued optimization problem. Note that the optimality notions
given above are the global ones, that is, the whole set F .S2 / has been taken into
account. We recall that their local versions are naturally defined as follows: The
point .x;N y/
N 2 graph .F / is said to be a local minimizer of (P2 ) if there exists a
neighborhood U of xN such that yN 2 Min.F .S2 \ U /; C /. Local analogues of other
minimizers have been defined in Chap. 2.
We have the following necessary optimality condition for local proper optimality
for (P1 ).
Theorem 12.1.1. Let X; Y; and Z be normed spaces, let C Y and D Z
be pointed, closed, and convex cones, where D is also solid, and let S X . Let
F W S Y and G W S Z be set-valued maps. For xN 2 S; let .x; N y/
N 2 graph .F /
be a local proper minimizer of (P1 ) and let zN 2 G.x/ N \ .D/. Then, for every
x 2 ˝ WD dom .D.F C C; G C D/.x; N y;
N zN//; the following disjunction holds:
\
D.F C C; G C D/.x;
N y;
N zN/.x/ Œ.C nf0g/ .intD zN/ D ;: (12.2)
12.1 First-Order Optimality Conditions by the Direct Approach 513
N y/
Proof. Since the element .x; N is a local proper minimizer of (P1 ), there are a
neighborhood U of xN and a proper, pointed, closed, convex, and solid cone K with
C nf0g intK such that yN 2 Min.F .S1 \ U /; K/. We will show that for every
x 2 ˝; we have
\
ŒD.F C C; G C D/.x; N y;
N zN/.x/ C .0; zN/ Œ.intK/ .intD/ D ;; (12.3)
which, due the fact that C nf0g intK; at once ensures (12.2).
We will establish (12.3) by a contradiction. For this, we assume that (12.3) does
not hold, and hence, for some x 2 ˝; we have
\
ŒD.F C C; G C D/.x; N y;
N zN/.x/ C .0; zN/ Œ.intK/ .intD/ ¤ ;;
which implies that there exists .y; z/ 2 D.F C C; G C D/.x;

N y;
N zN/.x/ such that
\
.y; z C zN/ 2 ŒD.F C C; G C D/.x;
N y;
N zN/.x/ C .0; zN/ Œ.intK/ .intD/ ;
(12.4)
confirming that
.x; y; z/ 2 graph .D.F CC; GCD/.x;

N y;
N zN// D T .epi .F; G/; .x;
N y;
N zN//: (12.5)
Therefore, we may use the definition of the contingent cone T .epi .F; G/; .x; N y;
N
zN// which ensures that there are sequences f.xn ; yn ; zn /g epi .F; G/ and ftn g P
such that .xn ; yn ; zn / ! .x;
N y;
N zN/ and tn .xn x;
N yn y;
N zn zN/ ! .x; y; z/.
Since xn ! xN and fxn g dom .F / D S; there exists n1 2 N such that
xn 2 S \ U for every n
n1 ; (12.6)
N tied to the notion of local proper minimality.

where U is the neighborhood of x;
Moreover, since y 2 intK (see (12.4)) and tn .yn y/
N ! y; there exists n2 2 N
satisfying that tn .yn y/
N 2 intK; for every n
n2 . Now, using the fact that
tn > 0; we have
yn 2 yN intK for every n

n2 : (12.7)
Furthermore, because .xn ; yn / 2 epi .F /; there exists a sequence fwn g with wn 2

F .xn / such that yn 2 wn C C; for every n 2 N. Therefore, wn 2 yn C; and
by (12.7), we have
wn 2 yN intK C yN intK K D yN intK;
which means that wn 2 yN intK; for every n

n2 .
We claim that xn 2 S1 ; that is, G.xn / \ .D/ ¤ ;; for sufficiently large n. For
this, we note that since z C zN 2 intD (see (12.4)) and tn .zn zN/ ! z; there exists
n3 2 N such that
tn .zn zN/ C zN 2 intD for every n

n3 ;
and accordingly,

1
zn 1 zN 2 intD for every n
n3 :
tn
To make a good use of the above containment, we note that zn ! zN and

tn .zn zN/ ! z imply that there exists n4 2 N such that tn > 1; for every n
n4 .
Therefore, since we have zN 2 D; we also have that .1 t1n /Nz 2 D; for every
n
maxfn3 ; n4 g. Consequently, we deduce that

1
zn 2 1 zN intD D intD D intD for every n
maxfn3 ; n4 g:
tn
Since f.xn ; zn /g epi .G/; there exists un 2 G.xn / such that zn 2 un C D; for
every n 2 N. This implies that
un 2 D intD D intD for every n

maxfn3 ; n4 g;
confirming that
G.xn / \ .D/ ¤ ; for every n

maxfn3 ; n4 g:
Summarising the above observations, we have shown that, for sufficiently large
n 2 N (indeed for n
fn1 ; n2 ; n3 ; n4 g), there exists xn 2 S1 \ U such that
F .xn / \ .yN intK/ ¤ ;;
N y/
contradicting the assumption that .x; N is a local proper minimizer. The proof is
complete. u
t
From the above result, we recover the following necessary optimality condition
for (P0 ).
and convex cone, let S X; and let F W S Y be a set-valued map. If .x; N y/
N 2
graph .F / is a local proper minimizer of (P0 ), then for every x 2 dom .D.F C
N y//;
C /.x; N the following disjunction holds:
\
D.F C C /.x;
N y/.x/
N .C nf0g/ D ;: (12.8)
Remark 12.1.3. Disjunctions (12.2) and (12.8) remain valid as first-order neces-
sary optimality conditions for local proper optimality for (P1 ) and (P0 ) if the
involved contingent derivatives are replaced by the adjacent derivative, the lower
Dini derivative, the circatangent derivative, the radial derivative, the contingent
epiderivative, the adjacent epiderivative, the radial epiderivative, the circatangent
epiderivative, or by their generalized and weakly generalized counterparts. Note
that for epiderivatives, identities analogous to (12.5) will play an important role. We
emphasize that the disjunctions for the radial derivative and the radial epiderivative
do not directly follow from disjunctions (12.2) and (12.8) but can be proved using
similar arguments.
We have the following example to illustrate the above results:
Example 12.1.4. Consider a set-valued map F W Œ0; 1R2 defined by
˚
F .x/ D .y1 ; y2 / 2 R2 j y12 C y22 x 2 ;
and let C D f.x; x/ 2 R2 j x 2 RC g. It is easy to verify that .1; .0; 1// 2

graph .F / is a proper minimizer. For x 2 dom .D.F C C /.x;
N y//
N D R ; we have
D.F C C /.x;
N y/.x/
N D f.y1 ; y2 / 2 R2 j y1 2 R; y2
jxjg;
and consequently,
D.F C C /.x;
N y/.x/
N \ .C nf0g/ D ; for every x 2 dom .D.F C C /.x;
N y//;
N
as claimed by the result. t

u
The arguments used in Theorem 12.1.1 readily yield the following necessary
optimality conditions for local weak-optimality.
be pointed, closed, convex and solid cones, and let S X . Let F W S Y and
G W S Z be set-valued maps. For xN 2 S; let .x; N y/
N 2 graph .F / be a local weak
minimizer of (P1 ) and let zN 2 G.x/N \ .D/. Then, for every x 2 dom .D.F C
C; G C D/.x;N y;
N zN//; the following disjunction holds:
\
D.F C C; G C D/.x;
N y;
N zN/.x/ Œ.intC / .intD zN/ D ;: (12.9)

closed, convex, and solid cone, let S X; and let F W S Y be a set-valued
N y/
map. If .x; N 2 graph .F / is a local weak minimizer of (P0 ), then, for every
x 2 dom .D.F C C /.x; N y//;
N the following disjunction holds:
\
D.F C C /.x;
N y/.x/
N .intC / D ;: (12.10)
In the following result, we extend the well-known Lagrange multiplier rule for
the local proper optimality concept.
F W S Y and G W S Z be set-valued maps. For xN 2 S; let .x; N y/
N 2 graph .F /
be a local proper minimizer of (P1 ), and let zN 2 G.x/
N \ .D/. Assume that D.F C
C; G C D/.x; N y;
N zN/.˝/ is convex, where ˝ WD dom .D.F C C; G C D/.x; N y;
N zN//.
Then, there exists .; / 2 Y Z nf.0; 0/g with 2 C C and 2 D C such that
.Nz/ D 0 and
.y/ C .z/
0 for every .y; z/ 2 D.F C C; G C D/.x;
N y;
N zN/.˝/: (12.11)
Proof. We set M WD D.F C C; G C D/.x; N y;

N zN/.˝/. Then Theorem 12.1.1 ensures
that the convex set ŒM C .0; zN/ satisfies the following disjunction:
\
ŒM C .0; zN/ Œ .intK intD/ D ;:
By the Eidelheit separation theorem, there are .; / 2 Y Z nf.0; 0/g and
2 R such that
.y/ C .z C zN/

for every .y; z/ 2 M; (12.12)
.c/ C .d / > for every .c; d / 2 intK intD: (12.13)
It follows from the above inequalities that we can set D 0. Furthermore, in view
of the continuity of .; /, by taking c 2 intK arbitrary close to zero, we obtain that
2 D C and by taking d 2 D arbitrarily close to zero, we obtain 2 C C .
Since .0; 0/ 2 M; we get from (12.12) that .Nz/
0. However, the inequality
.Nz/ 0 is also true from 2 D C and zN 2 D. Finally, (12.11) follows
from (12.12). The proof is complete. t
u
The following result sheds some light on the convexity assumption used in the
above result:
Proposition 12.1.8. Under the setting of Theorem 12.1.7, the set D.F C C; G C
D/.x;N y;
N zN/.˝/ is convex if the contingent cone T .graph .F C C; G C D/; .x;
N y;
N zN//
is convex.
Proof. Let T .graph .F C C; G C D/; .x; N y;
N zN// be convex. We set M WD D.F C
C; G C D/.x; N y;
N zN/.˝/ and choose .y1 ; z1 /; .y2 ; z2 / 2 M arbitrarily. Then, there
exist x1 ; x2 2 ˝ such that .y1 ; z1 / 2 D.F C C; G C D/.x; N y; N zN/.x1 / and .y2 ; z2 / 2
D.F C C; G C D/.x; N y;N zN/.x2 /. Therefore, .x1 ; y1 ; z1 /; .x2 ; y2 ; z2 / 2 T .graph .F C
C; G CD/; .x; N y;
N zN//. Since T .graph .F CC; G CD/; .x; N y;N zN// is convex, for every
t 2 Œ0; 1; we have
t.x1 ; y1 ; z1 / C .1 t/.x2 ; y2 ; z2 / 2 T .graph .F C C; G C D/; .x;

N y;
N zN//;
confirming that .tx1 C .1 t/x2 / 2 ˝ and .t.y1 ; z1 / C .1 t/.y2 ; z2 // 2 M .

u
Remark 12.1.9. The above result shows that the convexity assumption can be
dropped by replacing contingent derivatives by Clarke derivatives.
Note that Theorem 12.1.7 does not exclude the possibility of the Lagrange
multiplier being zero, and if happens to be zero, then (12.11) does not provide
much information. Therefore, it is of interest to find conditions ensuring that
¤ 0. It is known from scalar and vector optimization theory that such conditions
include a derivative of the map defining the constraints and are referred to as
constraint qualifications. However, the existence of the Lagrange multipliers in
Theorem 12.1.7 is proved in a setting where the contingent derivative D.F C
C; G C D/.x; N y;
N zN/ of the composite map .F C C; G C D/ is being used. This,
in general, is different from the composition .D.F C C /.x; N y/;
N D.G C D/.x; N zN//
of the derivatives. Therefore, first an analogue of Theorem 12.1.1 should be given
that takes into account the derivatives D.F C C /.x;
N y/
N and D.G C D/.x; N zN/.
be pointed, closed, convex, and solid cones, and let S X . Let F W S Y and
G W S Z be set-valued maps. For xN 2 S; let .x; N y/N 2 graph .F / be a local
weak minimizer of (P1 ), and let zN 2 G.x/ N \ .D/. Assume that either .F C C / is
Lipschitz-like around .x; N y/
N or .G C D/ is Lipschitz-like around .x; N zN/. Assume that
N y/
either epi .F / is derivable at .x; N or epi .G/ is derivable at .x;N zN/. Then, for every
x 2 dom .D.F C C /.x; N y//
N \ dom .D.G C D/.x; N zN//; the following disjunction
holds:
\
.D.F C C /.x; N y/;
N D.G C D/.x; N zN// .x/ Œ.intC / .intD zN/ D ;:
(12.14)
Proof. The proof follows by combining the arguments used above and in Theo-
rem 12.2.1. t
u
We have the following Lagrange multiplier rule along with a constraint qualifi-
cation.
Theorem 12.1.11. Besides the hypotheses of Theorem 12.1.10, assume that
.D.F C C /.x;N y/;
N D.G C D/.x; N zN//.˝/ is convex, where ˝ WD dom .D.F C
N y//
C /.x; N \ dom .D.G C D/.x; N zN//. Then there exists .; / 2 Y Z nf.0; 0/g
with 2 C C and 2 D C such that .Nz/ D 0 and
.y/ C .z/
0 for every .y; z/ 2 .D.F C C /.x;
N y/;
N D.G C D/.x;
N zN// .˝/:
(12.15)
Furthermore, ¤ 0 if additionally we assume that
D.G C D/.x;
N y/.˝/
N C cone .D C zN/ D Z: (12.16)
Proof. The proof of (12.15) is identical to the proof of (12.11) and hence has been
omitted. It remains to show that ¤ 0. For this we assume that D 0 and choose
zO 2 Z arbitrarily. Then, for s > 0 and for some z 2 D.G C D/.x; N zN/.˝/, we
have zO D z C s.D C zN/; which implies that, for every d 2 D; we get .Oz/ D
.z/ C s ..d / C .Nz//
0; confirming that D 0. This, however, is a contradiction
to the fact that and are not both simultaneously zero. The proof is complete. u t
Note that the imposed convexity in Theorem 12.1.11 allowed the use of classical
separation arguments which ensured that the multipliers are linear and continuous.
It turns out that within the absence of convexity, it will still be possible to prove
the existence of multipliers, however, the linearity of the multipliers will be lost in
this procedure. We explain this by giving an optimality condition for local weak
minimality for (P0 ).
We recall that a function z W Y ! R is called strictly C -monotone, where Y is
a normed space and C Y is a cone, if y1 2 y2 C C nf0g implies that z.y1 / >
z.y2 /. We also need to recall the following result (see [198]), where the notion of a
C -monotone function is being used.
Lemma 12.1.12. Let W be a linear topological space, let C W be a convex and
solid cone, and let A W be nonempty. Then the following two statements are
equivalent:
1. A \ .intC / D ;.
2. There exists a continuous sublinear functional s W W ! R which is strictly
intC -monotone with the range .1; C1/ and possesses the following proper-
ties:
s.A/
0; s.int.A// > 0; s.C / 0; s.b.C // D 0;
s.intC / < 0; s.C /
0; s.intC / > 0:
The following result is a direct application of the above result and Theo-
rem 12.1.6:
closed, convex, and solid cone, let S X; and let F W S Y be a set-
N y/
valued map. Let .x; N 2 graph .F / be a local weak minimizer of (P0 ) and let
˝ WD dom .D.F C C /.x; N y//.
N Then, there exists a continuous, sublinear, strictly
intC -monotone functional s W W ! R[f˙1g possessing the following properties:
s.D.F C C /.x;
N y/.˝//
N
0; s.int.D.F C C /.x;
N y/.˝///
N > 0; s.C / 0;
s.b.C // D 0; s.intC / < 0; s.C /
0; s.intC / > 0:
In optimization theory, theorems of alternative-type play a crucial role (see

Gwinner and Oettli [222]). In the following, we give another such result:

solid, and convex cone C Y; let S X; and let F W S Y be a set-valued map.
For ˝ WD dom .D.F C C /.x; N y//;
N let D.F C C /.x;N y/.˝/
N be convex. Then, only
one of the following statements hold:
1. There exists x 2 ˝ such that
D.F C C /.x;
N y/.x/
N \ .intC / ¤ ;:
2. There exists ` 2 C C nf0g such that
`.y/
0 for every y 2 D.F C C /.x;
N y/.˝/:
N (12.17)
N y/
Consequently (12.17) is a necessary condition for .x; N to be a local weak-
minimizer of .P0 /.
Proof. The proof follows from standard separation arguments used above. t
u
For the next optimality condition, we recall the following result due to
Kurcyusz [346]:
Lemma 12.1.15. Let Y be a normed space and let K1 and K2 be two cones in Y
with int.K2 / ¤ ;. If K1 \ int.K2 / D ; then .K1 K2 / cannot be dense in Y .
Proof. We will show that .K1 K2 / \ int.K2 / D ; so .K1 K2 / cannot be dense.
Indeed, if the disjunction does not hold, then .k1 k2 / D k; for some k1 2 K1 ; k2 2
K2 and k 2 int.K2 /. Therefore, for k1 2 K1 ; we have k1 D k2 Ck 2 K2 Cint.K2 / D
int.K2 /. This, however, contradicts the assumption that K1 \int.K2 / D ;. The proof
is complete. t
u
We now give a necessary optimality condition as a non-denseness of certain sets:
solid and convex cone, let S X; and let F W S Y be a set-valued map. If
.x; N is a local weak minimizer of (P0 ), then, .D.F C C /.x;
N y/ N y/.˝/
N C C / ; where
˝ WD dom .D.F C C /.x; N y//;
N cannot be dense in Y .
Proof. Since D.F C C /.x;
N y/.˝/
N is a cone, and since D.F C C /.x;
N y/.˝/
N \
.intC / D ;; the claim follows directly from Lemma 12.1.15. The proof is
complete. t
u
In the following, we state a result for .P2 / when the map H is single-valued,
that is,
Minimize F .x/ subject to x 2 S2 WD fx 2 S j G.x/ \ D ¤ ;; H.x/ D 0g:

(12.18)
We have the following result that employs the lower Dini derivative:
Theorem 12.1.17. Let .x; N y;
N zN/ be local weak-minimizer of (12.18) and let intD
be nonempty. For any x 2 dom .DL .F; G/.x; N zN// with H 0 .x/.x/
N y; N D 0; if the
N y;
set .DL .F; G/.x; N zN/.x/ C C D/ is convex, then there exists nonzero vector
.; / 2 .C D/C such that
N y;
.y/ C .z C zN/
0 for every .y; z/ 2 DL .F; G/.x; N zN/.x/: (12.19)
Moreover, if the set .DL .F C C; G C D/.x; N y;

N zN/.X / C .C D// is convex,
then there exists nonzero vector .; ; / 2 .C D f0g/C such that
.y/ C .z/ C .w/

0 for every .y; z; w/ 2 DL .F C C; G C D; H /.X /:
The element is nonzero if the following condition holds:

\
DL .F C C; G C D; H /.x;
N y;
N zN; 0/.X / .Y fintDg f0g/ ¤ ;: (12.20)
Proof. The proof is based on arguments similar to those used above and can be
found in [405]. u
t
Finally, we give some necessary optimality conditions for the notion of strong
minimality.
F W S Y and G W S Z be set-valued maps. Let .x; N y/
N 2 graph .F / be a local
strong minimizer of (P1 ), and let zN 2 G.x/
N \ D. Then:
1. For every x 2 ˝ WD dom .D.F C C; G C D/.x;
N y;
N zN//, we have
ŒD.F C C; G C D/.x;
N y;
N zN/.x/ C .0; zN/ \ Œ.Y nC / .intD/ D ;:
(12.21)
2. Assume that either .F C C / is Lipschitz-like around .x; N y/

N or .G C D/ is
Lipschitz-like around .x;N zN/. Assume that either epi .F / is derivable at .x;
N y/
N or
N zN/. Then for every x 2 ˝ WD dom .D.F C C /.x;
epi .G/ is derivable at .x; N y/
N \
D.G C D/.x; N zN//, we have
.D.F C C /.x;
N y/;
N D.G C D/.x;
N zN// .x/ \ Œ.Y nC / .intD zN/ D ;:
(12.22)
Proof. We will prove (12.21) by contraposition. Assume that there exists x 2 ˝

such that
.y; z C zN/ 2 ŒD.F C C; G C D/.x;

N y;
N zN/.x/ C .0; zN/ \ Œ.Y nC / .intD/ :
Therefore, .x; y; z/ 2 T .epi .F; G/; .x; N y;

N zN//; and consequently, there are a
sequence ftn g P and a sequence f.xn ; yn ; zn /g X Y Z such that tn # 0;
.xn ; yn ; zn / ! .x; y; z/ and for every n 2 N; we have yN C tn yn 2 F .xN C tn xn / C C
and zN C tn zn 2 G.xN C tn xn / C D. Since un WD .xN C tn xn / ! x;
N there exists n1 2 N
such that un 2 S \ U for every n
n1 and for some neighborhood U of x. N Because
y 62 C; yn ! y; and the cone C is closed, there exists n2 2 N such that yn … C
for every n
n2 . Since tn > 0; we have tn yn … C . Let wn 2 F .un / be such that
for every n 2 N and for some cn 2 C; we have yN C tn yn D wn C cn and hence
cn … yN C C for n
maxfn1 ; n2 g. Since C C C D C; we have wn … yN C C for
n
maxfn1 ; n2 g. As in the arguments in the previous results, we can show that
there exists n3 2 N such that G.un / \ D ¤ ; for every n
n3 .
Therefore, we have shown that for sufficiently large n 2 N; there are wn 2
F .S1 \ U / such that wn … yN C C . This, however, is a contradiction to the
assumption that .x; N y/
N is a local strong minimizer. Hence (12.21) must hold. The
proof of (12.22) is analogous. t
u
As a direct consequence of the above result, we get the following:
and convex cone, and let S X . Let F W S Y be a set-valued map, and let
N y/
.x; N 2 graph .F / be a local strong minimizer of (P0 ). Then, for every x 2 ˝ WD
dom .D.F C C /.x;N y//,
N we have
D.F C C /.x;
N y/.˝/
N C: (12.23)
The following example illustrates the above results:

Example 12.1.20. Consider the set-valued map F W Œ0; 1R given by:
˚
F .x/ WD .y1 ; y2 / 2 R2 j y12 C y22 x 2 :
Let C D R2C be the ordering cone. We see that the point .1; .0; 1// 2
graph .F / is a weak-minimizer. We employ the weak contingent epiderivative of
F at .1; .0; 1//; which, for every x 2 R ; is given by
˚
DW F .1; .0; 1//.x/ D .y1 ; y2 / 2 R2 j y2
jxj y1 2 R :
The necessary optimality for weak-minimality holds because

DW F .1; .0; 1//.R / \ int R2C D ;:
On the other hand, the necessary optimality condition for proper minimality is
violated as
˚
DW F .1; .0; 1//.0/ \ R2C nf.0; 0/g D .y1 ; 0/ 2 R2 j y1 < 0 :

Furthermore, the point 1; p12 ; p12 is a proper minimizer of (P0 ), but it is
not a strong minimizer. The generalized contingent epiderivative is
n
1 1 p o
DG F 1; p ; p .x/ D .y1 ; y2 / 2 R2 j y1 C y2
2x :
2 2
The necessary optimality condition for proper optimality holds because

1 1
DG F 1; p ; p .R / \ R2C nf0g D ;:
2 2
Finally, since

1 1
DG F 1; p ; p .R / ª R2C ;
2 2
the necessary optimality condition for strong minimality is violated.

by the Dubovitskii-Milyutin Approach
Let be a normed space and let ˙i ; i D 1; : : : ; n C 1; be convex cones with

˙i ; i D 1; : : : ; n; open. The Dubovitskii-Milyutin-lemma asserts an equivalence
between the following two statements:
\
nC1
(˛) ˙i D ;.
i D1
(ˇ) There exist li 2 ˙i ; i D 1; : : : ; n C 1; not all simultaneously equal to zero,
such that
l1 C l2 C C lnC1 D 0:
In optimization theory, the relevance of the so-called Dubovitskii-Milyutin-

lemma (see Theorem 12.2.21) comes from the fact that the optimality of a point can
be conveniently expressed as the disjunction of certain sets. Now, if these sets are
locally approximated by using suitable tangent cones so that the disjunction is main-
tained, then, under suitable convexity assumptions, the functionals appearing in (ˇ)
give rise to the Lagrange multipliers. In applications of the Dubovitskii-Milyutin-
lemma, the cone ˙nC1 is often an approximation of the equality constraints and
the cones ˙i ; i D 1; : : : ; n; contain some information about the behavior of the
derivatives of the objective map and the inequality constraints. The approach based
on the Dubovitskii-Milyutin-lemma is very flexible for many applications and has
12.2 First-Order Optimality Conditions by the Dubovitskii-Milyutin Approach 523
proven to be of great use in dealing with scalar and vector optimization problems.
In this chapter, we present optimality conditions by presenting an extension of the
Dubovitskii-Milyutin approach to set-valued optimization.
Some of the works on the Dubovitskii-Milyutin approach are by Bazaraa
and Goode [35], Das [125], Ioffe and Tihomirov [279], Halkin [231, 232],
Kotarski [332–334], Kowalewski and Kotarski [337], Ledzewicz-Kowalewska [372–
374], Ledzewicz and Schättler [365–369], Ledzewicz and Walczak [370, 371],
Leung [375], Rigby [487], Studniarski [546], Virsan [583], and Watkins [598],
among others.
12.2.1 Necessary Optimality Conditions

We begin with necessary optimality conditions for the following set-valued opti-
mization problem (P1 ), which was introduced earlier in this chapter:
.P1 / Minimize F .x/ subject to x 2 S1 D fx 2 S j G.x/ \ D ¤ ;g:
Before giving the main result, we recall that, given normed spaces X and Y and
a set-valued map F W X Y; the weak-inverse image F ŒA1 of F with respect to
a set A Y is defined by:
F ŒA1 WD fx 2 X j F .x/ \ A ¤ ;g:
The following is a necessary optimality condition for the notion of weak

minimality:
Theorem 12.2.1. Let X; Y; and Z be normed spaces, C Y and D Z; be
pointed, closed, convex, and solid cones, and let S X . Let F W X Y and
G W X Z be set-valued maps. For xN 2 S; let .x; N y/
minimizer of (P1 ). Assume that .F C C / is Lipschitz-like around .x;
N y/.
N Assume
that either epi .F / is derivable at .x; N or GŒD1 is derivable at x.
N y/ N Then the
following disjunction holds:
\ \
D.F C C /.x;
N y/ŒintC
N 1 N
IT.S; x/ T .GŒD1 ; x/
N D ;: (12.24)
Proof. We will show that if (12.24) fails, then a feasible x can be obtained in a
vicinity of xN with F .x/\.yN intC / ¤ ;; hence violating the local weak minimality
N y/.
of .x; N Assume that
\ \
x 2 D.F C C /.x;
N y/ŒintC
N 1 N
IT.S; x/ T .GŒD1 ; x/:
N
From x 2 T .GŒD1 ; x/; N we deduce that there are sequences ftn g P and
fzn g X such that tn # 0; zn ! x and xN C tn zn 2 GŒD1 for every n 2 N.
Therefore, G.xN C tn zn / \ .D/ ¤ ; for all n 2 N.
Since x 2 D.F C C /.x; N y/ŒintC
N 1 ; there exists y 2 intC such that .x; y/ 2
graph .D.F C C /.x; N y//.
N This confirms that there are sequences fsn g P and
f.xn ; yn /g X Y such that sn # 0; .xn ; yn / ! .x; y/ and yN C sn yn 2 F .xN C
sn xn / C C; for every n 2 N. By the derivability hypothesis, we can set tn D sn ; and
hence, yN C tn yn 2 F .xN C tn xn / C C for every n 2 N. At this juncture, we intend
to use the Lipschitz-like property at .x;N y/.
N Because un WD .xN C tn xn / ! xN and
vn WD .xN C tn zn / ! x;N there exists n1 2 N such that un ; vn 2 U WD U1 \ U2 for
n
n1 ; where U1 is the neighborhood tied to the notion of local weak minimality
and U2 is a neighborhood of xN which exists, along with a neighborhood V of y; N as a
consequence of the Lipschitz-like property. Moreover, since .yN C tn yn / ! y; N there
exists n2 2 N such that yN C tn yn 2 V for all n
n2 . Since .F C C / is Lipschiz-like
N y/;
at .x; N we get
yN C tn yn 2 Œ.F C C /.xN C tn xn / \ V .for n

n2 / .F C C /.xN C tn zn /
CL k .xN C tn xn / .xN C tn zn / k BY .for n
maxfn1 ; n2 g/ D F .xN C tn zn /
CC C L tn k xn zn k BY .for n
maxfn1 ; n2 g/;
and hence we can choose bn 2 BY such that, for n

maxfn1 ; n2 g; we have yN C
tn wn 2 F .xN C tn zn / C C where wn D .yn Lbn k xn zn k/ ! y.
Since y 2 intC; tn > 0; and wn ! y; there exists n3 2 N such that tn wn 2
intC . We choose wQ n 2 F .vn / such that for every n
n3 ; we have yN C tn wn 2
wQ n C C for every n 2 N. We have
wQ n 2 yN C tn wn C yN intC C yN intC
and consequently
wQ n 2 F .vn / \ .yN intC / for n

maxfn1 ; n2 ; n3 g:
Finally, we note that because tn # 0 and zn ! x; the containment x 2 IT.S; x/ N

implies that there exists n4 2 N such that vn WD xN C tn zn 2 S for every n
n4 .
Summarizing, we have shown that for every n
maxfn1 ; n2 ; n3 ; n4 g; there are
vn 2 S1 \ U1 ; such that F .vn / \ .yN intC / ¤ ;; which, however, contradicts the
local weak optimality of .x;N y/.
u
In view of the proof of Theorem 12.2.1, the following results are immediate.
be pointed, closed, convex, and solid cones, and let S X . Let F W X Y and
G W X Z be set-valued maps. Let .x; N y/
N 2 graph .F / be a local weak minimizer
to .P1 /. Let either dom .F / D dom .G/ D S or xN 2 intS; and let .F C C / is
N y/.
Lipschitz-like around .x; N Assume that either epi .F / is derivable at .x;
N y/
N or
GŒD1 is derivable at x.N Then:
\
D.F C C /.x;
N y/ŒintC
N 1 T .GŒD1 ; x/
N D ;:

convex, and solid cone, let F W X Y be a set-valued map, and .x;
N y/
N 2 graph .F /
be a local weak minimizer to .P0 /. Then:
\
N
IT.S; x/ D.F C C /.x;
N y/ŒintC
N 1 D ;:
Moreover, if additionally either dom .F / D S or xN 2 intS . Then:
D.F C C /.x;
N y/ŒintC
N 1 D ;:
Remark 12.2.4. Note that all the above results remain valid if, instead of D.F C
N y/;
C /.x; N we take either the adjacent derivative, the lower Dini derivative, the
circatangent derivative, the radial derivative, the contingent epiderivative, the
adjacent epiderivative, the radial epiderivative, the circatangent epiderivative, or
their generalized and weakly generalized counterparts. We remark that all the
above results, in fact, will characterize the proper minimality if we replace intC
by intK; where K is a proper, pointed, closed, convex, and solid cone such that
C nf0g intK. Also notice that for weak minimality, Theorem 12.2.3 states that
N Y n intC for every x 2 dom .D.F C C /.x; N y//.
N In
fact, in most known results, this necessary optimality conditions is proved by direct
arguments, which, primarily because of the definition of the derivative, demands
that dom .F / D S .
Next we give an example to illustrate the above remark:
Example 12.2.5. Consider the set-valued map F W Œ0; 2R2 defined by:
˚
F .x/ WD .y1 ; y2 / 2 R2 j y12 C y22 x 2 :
˚
Let C WD .x; x/ 2 R2 j x 2 RC be the ordering cone and let S WD
Œ0; 1 be the set of implicit constraints. It is easy to verify that the point
.1; .0; 1// 2 graph .F / is a (global) proper minimizer to .P0 /. For x 2
dom .DG F .1; .0; 1/// D R; we have
˚
DG F .1; .0; 1//.x/ D .y1 ; y2 / 2 R2 j y1 2 R; y2 D x ;
and hence DG F .1; .0; 1//.x/ \ .C nf0g/ ¤ ; for every x 2 P. However, since
N D P and
I T .S; x/
DG F .1; .0; 1//ŒC nf0g1 D P;

we have
\
DG F .1; .0; 1//ŒC nf0g1 N D ;;
IT.S; x/
ensuring that the necessary optimality condition holds.

Another approach to prove (12.24), in fact more akin to the classical Dubovitskii-
Milyutin approach, is to express the optimality as a disjunction of certain sets
and then use some separation theorems. In the present setting, the following two
results generalize the Dubovitskii-Milyutin approach and give an alternative proof
of Theorem 12.2.1.
Proposition 12.2.6. Let X; Y; and Z be normed spaces, let C Y and D Z
G W X Z be set-valued maps. For xN 2 S; let .x;N y/
minimizer to .P1 /. Then:
\ \
N
IT.S; x/ IT.F ŒyN intC 1 ; x/
N T .GŒD1 ; x/
N D ;: (12.25)
Proof. We claim that

\ \ \
U S F ŒyN intC 1 GŒD1 D ;;
where U is a neighborhood of xN corresponding to the local weak minimality. In fact,

if there exists
\ \ \
x2U S F ŒyN intC 1 GŒD1 ;
T T
then from x 2 U S GŒD1 we notice that x is feasible and from x 2 F ŒyN
intC 1 we obtain
F .x/ \ .yN intC / ¤ ;
N y/.
which is a contradiction to the optimality of .x; N The assertion now follows from
the properties of the interiorly contingent cones and the contingent cones. The proof
is complete. t
u
Remark 12.2.7. We emphasize that in (12.25) and in all such results in this work, it
is implicitly assumed that all the involved sets are nonempty. For instance the use of
N demands that the set S has a nonempty interior.
IT.S; x/
Besides giving an alternative proof for (12.24), the following result has its own
significance.
Proposition 12.2.8. Let X and Y be normed spaces, let F W X Y be a set valued

map, let .F C C / is Lipschitz-like around .x;
N y/
N 2 graph .F / and derivable at this
point. Let C be a pointed, closed, convex, and solid cone. Then:
D.F C C /.x;
N y/ŒintC
N 1 IT.F ŒyN intC 1 ; x/:
N
Proof. The proof is based on the arguments given in the proof of Theorem 12.2.1.
t
u
Let dom .F / D dom .G/ D S , let .x;
N y/
N 2 graph .F / and let zN 2 G.x/
N \ .D/.
N y/
Then a necessary optimality condition for .x; N to be a weak minimizer to (P1 ), by
the direct approach, reads as follows
N y;
D.F; G/.x; N zN/.˝/ \ ..intC / .intD zN// D ;;
where D.F; G/.x; N y;

N zN/ is the contingent derivative of the map .F; G/ WD F G at
N y;
.x; N zN/ and ˝ is the domain of this derivative. It has been shown in the previous
section that, under suitable convexity assumptions, these disjoint sets given in the
image space can be separated and a multiplier rule can be obtained. However the
optimality condition (12.24) is of different nature. Firstly, the disjunction in (12.24)
is taking place in the objective space. Secondly, in (12.24) we have not taken into
account any derivative of G. To get a multiplier rule from Theorem 12.2.1, we need
to employ the Dubovitskii-Milyutin lemma to (12.24). For this, among other things,
we need to gather some information about the dual of D.F C C /.x; N y/ŒintC
N 1
1
and a suitable regularity condition relating the cone T .GŒD ; x/ N with some
derivative of G at a feasible point. All this is done in the next section.
12.2.2 Inverse Images and Subgradients of Set-Valued Maps
Recall that given a normed space X with X as its dual, the polar M of M X
is defined by
M D f` 2 X j `.x/ 0 for every x 2 M g;
and the positive dual M C is defined by M C D M . It is known that M1 M2

implies M2 M1 .
Given a set A R and b 2 R; by the inequality A
b we understand that a
b
for every a 2 A.
We formulate the following:
Definition 12.2.9. Let X and Y be normed spaces, let C Y be a pointed, closed
and convex cone, and let F W X Y be set-valued.
1. Given A Y ; the A-scalarized subgradient @A F .x;

N y/
N of F at .x;
N y/
N 2
graph .F / is defined by
@A F .x; N WD fL 2 X j 9 y 2 A such that L.x/ .y ı D.F C C /.x;

N y/ N y//.x/
N
8 x 2 dom .D.F C C /.x;
N y//g:
N
N y/
2. Given B Y; the B-subgradient @B F .x; N of F at .x;
N y/
N 2 graph .F / is
defined by
N y/
@B F .x; N WD fL 2 L .X; Y /j .D.F C C /.x;
N y/.x/
N L.x// \ B D ;
8 x 2 dom .D.F C C /.x;
N y//g:
N
Remark 12.2.10. We shall use the terms, generalized subgradient, proper sub-
gradient and weak subgradient, if B D C nf0g; intK; where K is a closed,
convex, pointed and solid cone with C nf0g intK; and intC; and denote these
N y/;
variants by @G F .x; N @P F .x;
N y/
N and @W F .x;
N y/;
N respectively. If, in the above
definition of the B-subgradient, we replaced D.F C C /.x; N y/
N by DG F .x;N y/;
N
then we shall replace the term subgradient by subdifferential. In the following,
the generalized subdifferential, generalized proper subdifferential, generalized
weak subdifferential of F at .x;N y/
N will be denoted by G F .x;
N y/;
N P F .x;
N y/
N and
N y/;
W F .x; N respectively. Sticking to our terminology, the notion of generalized
weak subdifferential was introduced and analyzed by Song [538]. All the other
notions have first appeared in [284].
The above notions are similar to a notion due to Bair-Jahn [22]. Recall that, given
a set-valued map F W X Y and .x; N y/
N 2 graph .F /; the contingent epiderivative
N y/
DE F .x; N W X ! Y is a single-valued map such that epi .DE F .x; N y//
N D
N y//.
T .epi .F /; .x; N By employing this epiderivative, a notion of subgradients for
set-valued maps, given in [22], is as follows:
O .x;
@F N y/
N WD fL 2 L .X; Y /j L.x/ DE F .x;
N y/.x/
N 8 x 2 X g:
Though our primary goal is to state some constraint qualifications in terms of

the above notions, we pause to give some of their basic properties. The following
N y/;
result focuses on the generalized weak subdifferential W F .x; N which can also
be written as:
N y/
W F .x; N D fT 2 L .X; Y /j T .x/ … DG F .x;
N y/.
N x;
N y/.x/
N C intC; for every x 2 X g :
The next two results are based on Song [538]. In these results, it is assumed that
the generalized contingent epiderivative exists.
Theorem 12.2.11. Let X and Y be normed spaces, let C Y be a proper, pointed,
closed, and convex cone, let F W X Y be a set-valued map, and let .x;N y/
N 2
graph .F /. Then:
N y/
(a) The set W F .x; N is closed in L .X; Y /.
(b) For any
> 0; the following formula holds:
N y/

W F .x; N D W .
F /.x;
N
y/:
N
(c) If f W X ! Y is Fréchet differentiable at xN 2 S \ dom.F /; where S X is

open, then
W .f C F /.x;
N f .x/ N D f 0 .x/
N C y/ N C W F .x;
N y/:
N
Proof. We begin with a proof of (a). Let fTn g W F .x; N y/

N be an arbitrary
sequence such that Tn ! T . We will prove the closedness of W F .x; N y/
N by
contradiction. Assume that T … W F .x; N y/.
N Then there exists x 2 X such that Tx 2
N y/.x/
DG F .x; N C intC . From the convergence Tn ! T; we deduce that Tn x ! Tx.
Therefore, there exists an index m 2 N such that Tn x 2 DG F .x; N y/.x/
N C intC for
every n
m. This contradicts the fact that fTn g W F .x;
N y/N and hence W F .x;
N y/
N
must be closed.
For (b), it suffices to note that, for every
> 0; we have
DG F .x; N y/.x/
N D
N
y/.x/.
DG .
F /.x; N
For (c), we note that, from the identity (which is easy to prove),
D.f C F C C /.x;
N f .x/
N C y/.x/
N D f 0 .x/.x/
N C D.F C C /.x;
N y/.x/;
N
it follows that DG .f C F /.x;

N f .x/
N C y/.x/
N D f 0 .x/.x/
N C DG F .x;
N y/.x/.
N
For T 2 W .f C F /.x;
N f .x/
N C y/;
N we have Tx … DG F .x; N y/.x/
N C f 0 .x/.x/
N C
intC for any x 2 X and hence Tx f 0 .x/.x/
N … DG F .x;N y/.x/
N C intC; which
confirms that T f 0 .x/
N 2 W F .x;
N y/.x/.
N This shows W .f C F /.x;N f .x/
N C y/N
f 0 .x/
N C W F .x;
N y/.
N The converse can be proved by similar arguments. t
u
Recall that given normed spaces X and Y; a solid cone C Y; a set-valued map
F W X Y; and .x; N y/
N 2 graph .F /; the weakly-subdifferential @F .x;
N y/;
N proposed
by Tanino [567], reads:
N y/
@F .x; N WD fT 2 L .X; Y /j yN T .x/
N 2 WMin .[x2X ŒF .x/ T .x/; C /g:
N y/
We recall the following result that gives some information on @F .x; N ¤ ; (see
Song [535, 538]):
Lemma 12.2.12. Let X and Y be normed spaces, and let C Y be a pointed,
closed, convex, and solid cone. Assume that F W X Y is a C -convex set-valued
map such that int.epi .F // ¤ ;. If xN 2 int.dom .F // and yN 2 WMin.F .x/;
N C /;
N y/
then @F .x; N ¤ ;.
The following result relates the weakly subdifferential and the weak generalized
gradient.

closed, convex, and solid cone, let F W X Y be a set-valued map, and let .x;
N y/
N 2
graph .F /. Then
N y/
@F .x; N W F .x;
N y/;
N
N y//
and the converse holds if F is C -convex and epi .DG F .x; N D T .epi .F /; .x;
N y//.
N
Proof. Let T 2 @F .x; N y/
N be arbitrary. Assume that T … W F .x; N y/.
N Then for some
x 2 X; we have DG F .x; N y/.x/
N \ .Tx intC / ¤ ;. Let y 2 DG F .x; N y/.x/
N \
.Tx intC / be arbitrary. Consequently, there are sequences f.xn ; yn /g X Y and
ftn g P such that tn # 0; xn ! x; yn ! y; and yN C tn yn 2 F .xN C tn xn / C C .
Therefore, there exists a sequence fkn g C such that yNn WD yN T .x/ N C tn .yn
T .xn // kn 2 F .xN C tn xn / T .xN C tn xn /.
Since y T .x/ 2 intC; we have tn .yn T .xn // 2 intC; for sufficiently large
n 2 N, confirming that yn .yN T .x// N 2 intC for such n 2 N. However, this
contradicts the assumption that T 2 @F .x; N y/N and hence T 2 W F .x;
N y/.
N
For the converse inclusion, let T 2 W F .x; N y/
N be arbitrary. Then Tx 62
N y/.x/
DG F .x; N C intC for every x 2 X . Since F is C -convex, we have F .x/ yN
N y/.x
DG F .x; N x/
N C C; and hence, for every x 2 X we have .F S .x/ T .x/ .yN
N
T .x/// \ .intC / D ; which implies yN T .x/ N 2 WMin. x2X .Fx Tx/; C /.
Therefore T 2 @F .x; N y/.
u
We next connect the scalarized subgradients and the generalized subgradients.
Proposition 12.2.14. Let X and Y be normed spaces and let F W X Y be set-
N y/
valued which is assumed to be locally C -convex at .x; N 2 graph .F /. Then
N y/
@P F .x; N ¤ ; implies that @C C nf0g F .x;
N y/
N ¤ ;.
N y/
Proof. Let @P F .x; N ¤ ; and let L 2 @P F .x; N y/
N be arbitrary. By the definition
of the proper subgradient, there is a closed, convex, pointed and solid cone K with
C nf0g intK; such that
.D.F C C /.x;
N y/
N L/.x/ \ .intK/ D ;
for every x 2 ˝ WD dom .D.F C C /.x;
N y//:
N (12.26)
We claim that .D.F C C /.x; N y/

N L/.˝/ is convex. For i 2 f1; 2g; we choose
yi 2 .D.F C C /.x; N y/
N L/.˝/; ensuring that there are xi 2 ˝ such that yi 2
D.F C C /.x; N y/.x
N i / L.xi / and hence yi C L.xi / 2 D.F C C /.x; N y/.x
N i /. This
implies that, for some
2 .0; 1; we have
y1 C .1
/y2 2 .D.F C C /.x; N y/
N
L/.
x1 C.1
/x2 /; and by using the convexity of the set ˝; we conclude that
y1 C
.1
/y2 2 .D.F C C /.x; N y/
N L/.˝/. Now, by employing a separation theorem
to the disjunction given in (12.26), we assure the existence of y 2 K C nf0g
C C nf0g such that
.y ı D.F C C /.x;
N y//.x/
N
.y ı L/.x/ for every x 2 ˝:
Since L 2 L .X; Y / and y 2 Y ; we have y ı L 2 X and hence y ı L 2

N y/.
@C C nf0g F .x; N t
u
The following technical result will play an important role in the subsequent
development of this chapter. This result, which is taken from [284], is an extension
of a result by Rigby [487, Proposition 5] given for single-valued maps.
Theorem 12.2.15. Let X and Y be normed spaces, let M X be convex set, and
let A Y be a closed, convex, and solid cone. Let T W M Y be a A-convex
set-valued map. Then:
1. If T Œint.A/1 ¤ ;; then for every ` 2 L ; where L WD T ŒA1 ; there exists
t 2 AC such that
t ı T .x/
`.x/ for every x 2 M:
2. If T Œint.A/1 D ;; then there exists t 2 AC nf0g such that
t ı T .x/
0 for every x 2 M:
Proof. Let us begin with the case when the set T Œint.A/1 is nonempty. Then
the (negative) dual L of L WD T ŒA1 is nonempty as well. We choose ` 2 L
arbitrarily and define a set
E WD f.y; `.x// 2 Y R j y 2 T .x/ C A; x 2 M g:
Using the facts that M is convex, T is A-convex and ` 2 Y ; we deduce that E

is a convex set. Indeed, let .y1 ; z1 /; .y2 ; z2 / 2 E be arbitrary. Then by the definition
of E; for i D 1; 2; there exists xi 2 X with zi D `.xi / and yi 2 T .xi / C A. For
2 .0; 1; we have

z1 C .1
/z2 D `.
x1 C .1
/x2 /; which, in view of the
A-convexity of T; ensures that

y1 C .1
/y2 2
T .x1 / C .1
/T .x2 / C A T .
x1 C .1
/x2 / C A:
Since the set M is convex, we at once obtain that

.y1 ; z1 / C .1
/.y2 ; z2 / 2 E.
Having established the convexity of the set E, we claim that
E \ .int.A/ P/ D ;: (12.27)
In fact, if this is not the case, then there exists .x; y/ 2 X Y such that y 2
.T .x/ C A/ \ .int.A// and `.x/ > 0. Let w 2 T .x/ be such that y 2 w C A. Then
w 2 y A int.A/ A D int.A/; which contradicts that ` 2 L .
Therefore (12.27) holds, and hence by a separation theorem, we get the existence
of .f; g/ 2 Y Rnf0; 0g and a real number ˛ such that we have
f .u/ C g.v/
˛ for every .u; v/ 2 E (12.28a)
f .c/ C g.d / < ˛ for every .c; d / 2 int.A/ P: (12.28b)
Since A is a cone, we can set ˛ D 0 in (12.28a) and (12.28b). By taking d 2 P

arbitrarily close to 0 and c 2 int.A/ arbitrarily close to 0; we obtain f 2 AC and
g 0; respectively. We claim that g < 0. Indeed, if g D 0; we get
f .c/ < 0 for every c 2 int.A/;

f .u/
0 for every u 2 T .M / C A;
which, however, is impossible because .T .M / C A/ \ .int.A// ¤ ; and hence

g < 0.
Moreover, from (12.28a), for every x 2 M; we have f ı .T C A/.x/
.g ı
`/.x/. Therefore, by setting t D .f =g/ 2 A and noticing that 0 2 A; we finish
the proof of the first part.
For the second part, we notice that if T Œint.A/1 D ;; we have T .M / \
int.A/ D ;; and hence using arguments similar to those given above, we can
prove the existence of t 2 A nf0g such that t ı T .x/
0 for every x 2 M . The
u
The following particular case is quite useful. Henceforth we set .T .; // D
N.; /.
Theorem 12.2.16. If in Theorem 12.2.15, we assume that A D T .B; w/; N where
B Y is a pointed, closed, convex, and solid cone with wN 2 B and T is B-convex,
then t 2 B C and t.w/
N D 0.
Proof. To apply Theorem 12.2.15, it suffices to show that the set-valued map T is
T .B; w/-convex.
N To show that B T .B; w/;
N it suffices to notice that BCwN B
is equivalent to B B C .w/.N For this, we notice that, due to wN 2 B; we have
B C .w/ N B C B D B. Now, the T .B; w/-convexity
N of the map T follows
from its B-convexity.
Furthermore, by invoking Theorem 12.2.15, for every ` 2 T ŒA1 ; with
A D T .B; w/, N there exists a functional t 2 .T .B; w//N C which satisfies the
assertion. Since B is convex, we have T .B; w/
N B Cw N and consequently
t.b C w/
N
0 for every b 2 B: (12.29)
By setting b D 0 in the above inequality, we obtain t.w/

N
0. Because B is a
cone and wN 2 B; we can substitute b D 2w N 2 B in (12.29) to obtain t.w/
N 0.
N D 0 and this, in
Therefore, combining the preceding two inequalities, we have t.w/
view of (12.29), yields t 2 B C . t
u
We have the following result which will play an important role in deriving
optimality conditions in set-valued optimization.
Theorem 12.2.17. Let X and Y be normed spaces and let F W X Y be a set-
N y/
valued map which we assume to be locally C -convex at .x; N 2 graph .F /. That is,
D.F C C /.x;N y/
N is a convex set-valued map. Let K be a solid, closed, and convex
cone such that C K. Then:
1. If
1
D.F C C /.x;
N y/Œint.K/
N ¤ ;;
then
1 1
.D.F C C /.x;
N y/ŒK
N / .D.F C C /.x;
N y/Œint.K/
N / @C C F .x;
N y/:
N
(12.30)
2. If
1
D.F C C /.x;
N y/Œint.K/
N D ;;
then
N y/:
0 2 @C C nf0g F .x; N
Proof. The first inclusion is a consequence of

1 1
D.F C C /.x;
N y/Œint.K/
N D.F C C /.x;
N y/ŒK
N ;
and the second inclusion is essentially a restatement of Theorem 12.2.15 when we

notice that the map D.F CC /.x;N y/
N is K-convex and set T D.F CC /.x; N y/.
N t
u
We conclude this section by using the scalarized subgradient to express a neces-
sary optimality condition for the local proper minimality in set-valued optimization.
N y/
Theorem 12.2.18. Let .x; N 2 graph .F / be a local proper minimizer to .P0 /. Let
the map F W X Y be locally C -convex at .x;
N y/
N with dom .F / D S . Then,
N y/:
0 2 @C C nf0g F .x; N
N y/
Proof. Since .x; N is a local proper minimizer, by the similar arguments as in
Theorem 12.2.1 we can show that
1
D.F C C /.x;
N y/ŒintK
N D;
for some pointed, closed, convex, and solid cone K satisfying C nf0g intK.
(In fact this link has lead to the proof of Theorem 12.2.2.) Now, by applying
N y/.
Theorem 12.2.17, we obtain that 0 2 @C C nf0g F .x; N The proof is complete. t
u
12.2.3 Separation Theorems and the Dubovitskii-Milyutin

Lemma
In this section, we give a separation theorem and from that we deduce the so-called
Dubovitskii-Milyutin lemma. This is a classical result and can be found in many
books. Our proofs are based on the excellent paper by Ben-Tal and Zowe [45].
We recall that, for a subset S of a normed space X , its support functional
.; S /; defined on the topological dual X of X with values in the extended real
line R [ f˙1g; is given by:
.x ; S / D suphx ; xi for x 2 X : (12.31)

x2S
The effective domain of .; S / is denoted and defined as .S / WD fx 2

X j .x ; S / < 1g. It is easy to verify that .; S / is a positively homogeneous,

closed, and convex function, and that .S / is a convex cone. Furthermore, whenever
S is a cone, we have
.S / D S C (12.32)

0 if x 2 .S /;
.x ; S / D (12.33)
1 otherwise.
We begin with the first main result of this section:

Theorem 12.2.19. TLet X be a normed space, let S1 ; : : : ; Sn be convex subsets of
X; and let x 2 n
i D1 Si . If
!
\
n1
int.Si / \ Sn ¤ ;; (12.34)
i D1
then
! ( )
\
n X
n

x ; Si D min .xi jSi /j
x D x1 C xn ; xi 2 .Si / :
i D1 i D1
Proof. The proof is by induction over n. For n D 1; there is nothing to prove. Now
assume that the assertion holds for n 1 and proceed to n sets. We set
\
n1
K1 WD Si ;
i D1
K2 WD Sn :
Tn T
With the given x 2 i D1 Si D .K1 K2 / ; we define the following two
disjoint sets in X R:

x
A WD such that x 2 K1 ; " > 0 ;
hx ; xi "

x
B WD j x 2 K 2 :
.x ; K1 \ K2 /
The sets A and B are nonempty, convex, and because of (12.34), A has a
nonempty interior. Hence, there exists a closed hyperplane in X R separating
A and B. That is, we can choose y 2 X ;
2 R; not both simultaneously zero,
and ˇ 2 R such that, for every x 2 K1 and every z 2 K2 ; we have

hy ; xi C
hx ; xi "
ˇ
hy ; zi C
x ; K1 \ K2 ; (12.35)
from which it follows that

0. Without any loss of generality, we set
D 1.
Indeed, if
D 0; then hy ; x zi
0 for every x z 2 K1 K2 ; and because
int .K1 / \ K2 ¤ ;; we deduce that y D 0; which is a contradiction. Therefore, we
get from (12.35) that

z WD x y 2 .K1 / and z ; K1 " ˇ;

y 2 .K2 / and y ; K2 ˇ C .x ; K1 \ K2 /;
and hence x D z C y with z 2 .K1 / and y 2 .K2 / with
.z ; K1 / C .y ; K2 / " C .x ; K1 \ K2 /: (12.36)
However, the following inequality also holds trivially,
.x ; K1 \K2 / "C.z ; K1 \K2 /C.y ; K1 \K2 / "C.z ; K1 /C.y ; K2 /;
and consequently, we have .z ; K1 / C .y ; K2 / D .x ; K1 \ K2 /. This forces

" D 0; and this proves the assertion for n. The proof is complete. t
u
Our second result is the following separation theorem:
Theorem 12.2.20. Let X be a normed space, let S1 ; : : : ; Sn ; SnC1 ; be nonempty
convex subsets of X; where S1 ; : : : ; Sn are open. Then
\
nC1
Si D ; (12.37)
i D1
if and only if there are xi 2 .Si /; i D 1; : : : ; n C 1; such that
x1 C x2 C C xnC1

D 0; (12.38)
.x1 ; S1 / C .x2 ; S2 / CC
.xnC1 ; SnC1 / 0: (12.39)
Proof.T Assume that (12.38) and (12.39) hold but (12.37) fails. That is, there exists
xN 2 nC1i D1 Si . We set Ki WD Si x;
N for i D 1; 2; : : : ; n C 1. Then 0 2 Ki and hence
.xi ; Ki /
0; for every i . Note that xj ¤ 0 for at least one j n; as otherwise

xnC1 D x1 C : : : C xn D 0; implying that x1 D D xnC1

D 0. However, since

0 2 int.Kj /; we have .xj ; Kj / > 0 for this special j . Therefore,
X
nC1 X
nC1 X
nC1 X
nC1
.xi ; Si / D .xi ; Ki / C hxi ; xi
N D .xi ; Ki / > 0;
i D1 i D1 i D1 i D1
which is a contradiction to (12.39).

T
For the converse, we set S WD niD1 Si and assume thatTK ¤ ;. Then S and SnC1
are nonempty convex sets, S is open, and by (12.37), S SnC1 D ;. Therefore, by
a standard separation theorem, we have a suitable x 2 X ; x ¤ 0 such that
hx ; xi hx ; zi for all x 2 S; z 2 SnC1 ;
implying that x 2 .S /; x 2 .SnC1 /; and

!
\
n

x ; Si C x ; SnC1 0:
i D1
We now apply Theorem 12.2.19 to ensure that, with suitable xi 2 .Si /; i D
1; 2; : : : ; n and
x D x1 C x2 C xn ;

!
\
n Xn

x ; Si D xi ; Si ;
i D1 i D1
equations (12.38) and (12.39) follow if we put xnC1 WDTx . Note that, ifTS D
T n m mC1
i D1 Si D ;; then there exists 1 m < n such that i D1 Si ¤ ; and i D1

Si D ;. We apply the result just proved above with n replaced by m and set xmC2 D

D xnC1 D 0. The proof is complete. t
u
We have the following well-known separation theorem for convex cones, also
known as the Dubovitskii-Milyutin lemma. An alternative proof of this is given in
Girsanov [211, pp 37].
Theorem 12.2.21. Let X be a normed space, let S1 ; : : : ; Sn ; SnC1 be nonempty

convex cones in X , and let S1 ; : : : ; Sn be open. Then,
\
nC1
Si D ;
i D1
if and only if there are xi 2 SiC ; i D 1; 2; : : : ; n C 1; not all zero, such that
x1 C C xnC1

D 0:
Proof. The proof follows from Theorem 12.2.20 and the observation that (12.39)
becomes the trivial inequality 0 0. t
u
12.2.4 Lagrange Multiplier Rules by the Dubovitskii-Milyutin

Approach
In our next result we give the promised Lagrange multiplier rule.

Theorem 12.2.22. Let X , Y , and Z be normed spaces, let C Y; and D Z be
pointed, closed, and convex cones, where C is also solid, and let F W X Y and
G W X Z be set-valued maps. For xN 2 S X; let .x; N y/
N 2 graph .F / and let
zN 2 G.x/
N \ .D/. Let .x;N y/
N be a local weak minimizer of (P1 ). Assume that the
following conditions hold:
1. The map .F CC / is Lipschitz-like around .x;N y/
N and locally convex and derivable
at this point.
2. There exists an open convex cone M IT.S; x/. N
3. Either IT.F ŒyN intC 1 ; x/
N is convex or D.F C C /.x;N y/ŒintC
N 1 is open.
4. dom .D.F C C /.x; N y//
N D dom .D.G C D/.x; N zN// D X; and the following
regularity condition .RC / holds: If
N zN/ŒT .D; Nz/nf0g1 D ;;

D.G C D/.x;
N zN/, otherwise T .GŒD1 ; x/

then 0 2 @N.D;Nz/nf0g G.x; N is convex and we have:
N.GŒD1 ; x/
N @N.D;Nz/ G.x;
N zN/: (12.40)
Then there exists .s; t; u/ 2 X Y Z ; not all zero, such that s 2 M C ;

t 2 C C and u 2 D C . Moreover the complementary slackness condition u.Nz/ D 0
and the following inequality hold:
t ı D.F C C /.x;
N y/.x/
N C u ı D.G C D/.x;
N zN/.x/
s.x/ for every x 2 X:
(12.41)
If either xN 2 intS or dom .F / D dom .G/ D S then s 2 Z can be set to zero.

In this particular case, we have t ¤ 0 if the following regularity condition holds:
D.G C D/.x;
N zN/.X / C cone.D C zN/ D Z: (12.42)
Proof. We begin by showing that the assertions would hold trivially if either of the
following disjunctions hold:
D.F C C /.x;
N y/ŒintC
N 1 D ;;
N zN/ŒT .D; Nz/nf0g1 D ;:
D.G C D/.x;
Indeed, if D.F C C /.x;N y/ŒintC

N 1 D ;; then, in view of Theorem 12.2.17,
C
there exists t 2 C nf0g such that for every x 2 X; we have
t ı D.F C C /.x;
N y/.x/
N
0;
and hence by choosing s D 0 and u D 0; we get the desired result.

N zN/ŒT .D; Nz/nf0g1 D ;; then, in view of
On the other hand, if D.G C D/.x;
the condition
0 2 @N.D;Nz/nf0g G.x;
N zN/;
we assure the existence of
u 2 N.D; Nz/nf0g D T .D; Nz/ nf0g D T .D; Nz/C nf0g;
such that for every x 2 X; we have
u ı D.G C D/.x;
N zN/.x/
0;
and by choosing s D 0 and t D 0; we obtain (12.41). The proof for u 2 D C and the
complementary slackness condition follows from Theorem 12.2.16.
Therefore, without any loss of generality, we can assume that
D.F C C /.x;
N y/ŒintC
N 1 ¤ ;;
N zN/ŒT .D; Nz/nf0g1 ¤ ;:
D.G C D/.x;
We first consider the case when IT.F ŒyN intC 1 ; x/

N is convex. Now, by using
Proposition 12.2.6 and the containment M IT.S; x/,N we have
IT.F ŒyN intC 1 ; x/

N \ M \ T .GŒD1 ; x/
N D ;:
The above disjunction, in view of the Dubovitskii-Milyutin lemma, ensures the

existence of functionals
l1 2 M ;
l2 2 IT.F ŒyN intC 1 ; x/
N ;
l3 2 N.GŒD1 ; x/;
N
such that
l1 C l2 C l3 D 0: (12.43)
At this juncture, we recall the following inclusion (see Proposition 12.2.8),
D.F C C /.x;
N y/ŒintC
N 1 IT.F ŒyN intC 1 ; x/;
N
and consequently, we have
l1 2 M ;
l2 2 .D.F C C /.x;
N y/ŒintC
N 1 / ;
l3 2 N.GŒD1 ; x/:
N
Note that, if D.F C C /.x; N y/ŒintC

N 1 is open then since it is a convex
cone, we can directly obtain the above conclusion. Now, because D.G C
N zN/ŒT .D; Nz/nf0g1 ¤ ;; we have
D/.x;
N.GŒD1 ; x/
N @N.D;Nz/ G.x;
N zN/;
and consequently l3 2 @N.D;Nz/ G.x;

N zN/.
This observation and Theorem 12.2.17, with A D C; ensure the existence of
functionals t 2 C C and u 2 N.D; Nz/ D T .D; Nz/C such that
t ı D.F C C /.x;
N y/.x/
N
l2 .x/ for every x 2 X
u ı D.G C D/.x;
N zN/.x/
l3 .x/ for every x 2 X:
Combining the above two inequalities with (12.43) yields
t ı D.F C C /.x;
N y/.x/
N C u ı D.G C D/.x;
N zN/.x/
l1 .x/ for every x 2 X:
By setting s D l1 2 M C we finish the proof of (12.40). The remaining proof

for u 2 D C and for the complementary slackness condition u.Nz/ D 0 is same as in
Theorem 12.2.16.
Finally, if either xN 2 intS (that is, IT.S; x/

N D X ) or dom .F / D dom .G/ D
S; then we can apply the above arguments to the following disjunction given in
Theorem 12.2.2:
\
IT.F ŒyN intC 1 ; x/
N T .GŒD1 ; x/
N D ;:
It remains to show that under the regularity condition (12.42), we have t ¤ 0.

For this, assume that t D 0 and choose zQ 2 Z arbitrarily. Since
zQ D z C ˇ.d C zN/ where z 2 D.G C D/.x;

N zN/.X / d 2 D; ˇ > 0;
we deduce that for an arbitrary zQ 2 Z; we have
u.Qz/ D u.z/ C ˇ.u.d / C u.Nz//

0:
This, however, implies that u D 0; and we have a contradiction to the fact that
.t; u/ ¤ .0; 0/. The proof is complete. t
u
As evident from the above proof, the assumption (RC ) is vital to ensure the
existence of the multipliers. A detailed discussion of this condition, which is
partly motivated by the Guignard constraint qualification, can be found in [284].
Furthermore, this condition can be bypassed when the cone D is solid. This can be
seen in Sect. 12.5 where an extension of the Dubovitskii-Milyutin is presented. We
also remark that a treatment of (P2 ) for optimality condition by the Dubovitskii-
Milyutin approach is given in [284]. To avoid unnecessary repetition, we treat an
analogue of (P2 ) in Sect. 12.5.
We now focus on the strong minimizers. Recall that a point .x; N y/
N 2 graph .F /
is called a local strong minimizer of (P2 ) if there exists a neighborhood U of xN such
that F .S2 \ U / yN C C . In the following result, we give necessary optimality
conditions for the local strong minimality. Let the cone C be closed and convex and
the cone D be solid.
Theorem 12.2.23. Let X and Y be normed spaces, let F W X Y be set-valued
N y/
and let .x; N 2 graph .F /. Let the map .F C C / possess the Aubin property around
N y/.
.x; N
N y/
.a/ If .x; N is a strong minimizer to .P1 / then the following holds:
\ \
D.F C C /.x; N nC 1
N y/ŒY N
IT.S; x/ T .GŒD1 ; x/
N D ;:
.b/ If .x; N is a strong minimizer to .P1 / then for zN 2 G.x/

N y/ N \ D; the following
holds:
\ \
D.F C C /.x;
N y/ŒY
N nC 1 N
IT.S; x/ N zN /ŒIT.D; zN /1 / D ;:
cl .D.G C D/.x;
N y/
.c/ If .x; N is a strong minimizer to .P2 / then the following holds:
\ \ \
D.F C C /.x; N nC 1
N y/ŒY N
IT.S; x/ IT.GŒD1 ; x/
N T .H Œh1 ; x/
N D ;:
.d / Let the map .G C D/ possess the Aubin property around .x; N zN/ where zN 2
N \ D. If .x;
G.x/ N y/
N is a strong minimizer to .P2 / then the following holds:
\ \
D.F C C /.x;
N y/ŒY
N nC 1 N
IT.S; x/ D.G C D/ŒIT.D/; zN/1
\
T .H Œh1 ; x/
N D ;:
Proof. (a) Assume that there exists an element x 2 X such that

\ \
x 2 D.F C C /.x; N nC 1
N y/ŒY N
IT.S; x/ T .GŒD1 ; x/:
N
In view of the containment x 2 D.F C C /.x; N nC 1 ; there exists y 2 Y nC

N y/ŒY
such that .x; y/ 2 graph .D.F C C /.x; N y//.
N This implies that there are sequences
.
n / P and ..xn ; yn // X Y such that
n ! 0; .xn ; yn / ! .x; y/ and
yN C
n yn 2 F .xN C
n xn / C C for all n 2 N. Also, since x 2 I T .S; x/;N there
exists n1 2 N such that xN C
n xn 2 S for n
n1 . Since y … C; yn ! y and the
cone C is closed, there exists n2 2 N such that yn … C for all n
n2 . Because
n > 0; we have
n yn … C . For un WD xN C
n xn ; let wn 2 F .un / be such that for
n 2 N; we have yN C
n yn D wn C cn ; where cn 2 C . Hence wn C cn … yN C C
for n
n2 . Since C C C D C; we have wn … yN C C; for every n
n2 . As
in the proof of Theorem 12.2.1, we can show that there exists n3 2 N such that
G.un / \ .D/ ¤ ; for all n
n3 . Therefore, we have shown that, for sufficiently
large n 2 N; there are wn 2 F .S1 / such that wn … yN C C; which, however, is a
N y/
contradiction to the assumption that .x; N is a strong minimizer. Hence part (a) is
true. The parts (b)–(d) can now be shown by unifying the above arguments with
those given in the preceding section. t
u
We conclude this section by giving a simple example which shows that the
various optimality conditions given above are not, in general, sufficient optimality
conditions:
Example 12.2.24. Consider a set-valued map F W RR given by:
F .x/ D Œx 2 ; 1/ for every x 2 R:
Let RC be the ordering cone. We choose a point .x; N y/

N D .0; 0/ 2 graph .F /.
N y/
Then the generalized contingent epiderivative DG F .x; N is given by:
N y/.x/
DG F .x; N D 0 for every x 2 dom .DG F .x;
N y//
N D R:
N y/.R/
Since DG F .x; N RC ; the optimality condition for strong minimality
N y/
holds, however, the point .x; N is not a strong minimizer.
12.3 Sufficient Optimality Conditions in Set-Valued

Optimization
It is evident from scalar and vector optimization that, for sufficient optimality
conditions, some additional assumptions must be imposed on the involved data.
Most naturally, this remains valid in set-valued optimization as well (see Exam-
ple 12.2.24). In this section, our objective is to give some sufficient optimality
conditions for set-valued optimization problems.
12.3.1 Sufficient Optimality Conditions Under Convexity

and Quasi-Convexity
Before giving more general sufficient optimality conditions for constrained set-
valued optimization problems, we give sufficient optimality conditions for the
unconstrained problem .P0 /. The following result illustrates an important aspect of
the behaviour of the involved derivatives that is instrumental in deriving sufficient
optimality conditions.
Theorem 12.3.1. Let X and Y be normed spaces, let S X be convex, let C Y
be a pointed, closed, and convex cone, and let F W S Y be a C -convex set-valued
map.
N y/
1. An element .x; N 2 graph .F / is a weak-minimizer, if for every x 2 S; we have
D.F C C /.x;
N y/.x
N x/
N \ .intC / D ;: (12.44)
N y/
2. An element .x; N 2 graph .F / is a minimizer, if for every x 2 S; we have
D.F C C /.x;
N y/.x
N x/
N \ .C nf0g/ D ;: (12.45)
Proof. Since the map F is C -convex, for every x 2 S; we have
F .x/ yN .F .x/ C C y/
N D.F C C /.x;
N y/.x
N x/;
N (12.46)
and consequently, it follows from (12.44) that .F .S / y/

N \ intC D ;; which
N y/
confirms that .x; N is a weak minimizer of .P0 /. The proof of the second part
follows from similar arguments. t
u
12.3 Sufficient Optimality Conditions in Set-Valued Optimization 543
Remark 12.3.2. The key ingredient in the above proof is the inclusion (12.46).
Since this inclusion always holds for the radial derivatives, the convexity hypothesis
is not needed when using the radial derivatives. Although we have used the
contingent derivative of the profile map, the same arguments are used when
employing contingent epiderivatives or their generalized analogues.
We now formulate a notion of convexity taken from [297]. This notion was
inspired by a related notion proposed by Jahn and Sachs [302] for single-valued
maps. See also [218].
Definition 12.3.3. Let X and Y be normed spaces, let S X; and let Y .
Let F W S X Y be a set-valued map, and let .x; N y/
N 2 graph .F /: The map F
N y/
is called -contingently-quasi-convex at .x; N if for every x 2 S; the condition
.F .x/ y/
N \ ¤ ; ensures that .DF .x;
N y/.x
N x//
N \ ¤ ;.
Recall that given a convex cone in a normed space Y , the quasi-interior C # of
C
C is given by
C # D f` 2 Y j `.x/ > 0 8x 2 C nf0gg:
In the following result, we give a sufficient optimality condition for minimality.

be pointed, closed, and convex cones, let S X be convex, let F W S Y and
G W S Z be set-valued maps, let .x; N y;
N zN/ 2 graph .F; G/; and let S xN
dom .D.F; G/.x; N zN//. Assume that there are t 2 C # and u 2 D C such that
N y;
u.Nz/ D 0 and
t.y/ C u.z/
0 for every .y; z/ 2 D.F; G/.x;
N y;
N zN/.x x/
N for all 2 S:
(12.47)
N y/
Then the element .x; N is a minimizer of F on
SO WD fx 2 S j G.x/ \ .D C cone .Nz/ cone .Nz// ¤ ;g (12.48)
if and only if .F C C; G C D/ W SO Y Z is -contingently quasi convex at

N y;
.x; N zN/ with D .C nf0g/ .D C cone .Nz/ cone .Nz//.
Proof. We claim that for every x 2 SO , the following disjunction holds:
ŒD.F C C; G C D/.x;
N y;
N zN/.x x/
N \ D ;: (12.49)
Indeed, if (12.49) is violated, then there exists xQ 2 SO such that .y;

Q zQ/ 2 D.F C
C; G C D/.xQ x/; yQ 2 C nf0g; and zQ 2 .D C cone .Nz/ cone .Nz//. Since
t 2 C # and u.Nz/ D 0; we have t.y/ Q C u.Qz/ < 0 which is a contradiction to (12.47).
Therefore, (12.49) holds. Since .F C C; G C D/ is -contingently-quasi convex at

N zN/; (12.49) ensures that there is no x 2 SO with
N y;
.x;
.F .x/ C C y/
N .G.x/ C D zN/ \ ¤ ;;
or, there is no x 2 SO with .F .x/ C C y/\.C

N nf0g/ ¤ ; and .G.x/ C D zN/\
.D C cone .Nz/ cone .Nz// ¤ ;. This means that there is no x 2 SO with
.F .x/ y/\.C
N nf0g/ ¤ ; and G.x/\.DCcone .Nz/cone .Nz// ¤ ;. Therefore,
there is no feasible point x such that .F .x/ y/
N \ .C nf0g/ ¤ ;. Consequently,
N is a minimizer of F on SO .
N y/
.x;
N is a minimizer of F on SO . Therefore, there
N y/
For the converse, assume that .x;
O
is no x 2 S with .F .x/ y/ N \ .C nf0g/ ¤ ; and .G.x// \ .D C cone .Nz/
cone .Nz// ¤ ;. Consequently, for every x 2 SO ;
..F .x/ C C y/
N .G.x/ C D fNz// \ D ;;
which, using (12.49), confirms that .F CC; G CD/ is -contingently-quasi convex

N y;
at .x; N zN/. t
u
We have the following particular case of interest:
Corollary 12.3.5. Under the setting of Theorem 12.3.4, if the map .F CC; GCD/ W
SO Y Z is -contingently quasi convex at a minimizer of .P1 /.
Proof. The proof follows from the fact that S1 SO and hence .x;
N y/
N is a minimizer
of .P1 /. t
u
We have the following analogue of Theorem 12.3.4 for the weak-minimality.
Theorem 12.3.6. Let X; Y; and Z be normed spaces, let C Y and D Z be
pointed, closed, and convex cones where C is also solid, let S X be convex, let
F W S Y and G W S Z be set-valued maps, let .x; N y;
N zN/ 2 graph .F; G/; and let
S xN 2 dom .D.F; G/.x; N zN//. Assume that there are t 2 C C nf0g and u 2 D C
N y;
with u.Nz/ D 0 satisfying (12.47).Then .x; N is a weak-minimizer of F on SO given
N y/
in (12.48) if and only if .F C C; G C D/ W SO Y Z is -contingently quasi
N y;
convex at .x; N zN/ with D .intC / .D C cone .Nz/ cone .Nz//.
Proof. The proof is analogous to that of Theorem 12.3.4 and hence omitted. t
u
The following result is immediate:
Corollary 12.3.7. Under the setting of Theorem 12.3.6, if the map .F CC; GCD/ W
SO Y Z is -contingently quasi convex at a weak-minimizer of .P1 /.
Proof. The proof follows from the fact that S1 SO and hence .x;
N y/
N is a weak-
minimizer of .P1 /. t
u
Some results related to this subsection can be found in [218, 297].
12.3.2 Sufficient Optimality Conditions Under Paraconvexity
In the following, we explore the notion of -paraconvex and compactly -

paraconvex set-valued maps and use these notions to give some first-order
sufficient optimality conditions for set-valued optimization problems. The notion
of paraconvex functions was proposed by S. Rolewicz (see [505, 506]) and used by
many authors (see [270, 311, 552] and the cited references therein). Our treatment
of this section is based on Amahroq and Taa [5].
Definition 12.3.8. Let X and Y be Banach spaces and let F W X Y be a set-
valued map.
1. The map F is called -paraconvex, where > 0; if there exists a constant c > 0
such that for all x; z 2 X and ˛ 2 .0; 1/; the following inclusion holds:
˛F .x/ C .1 ˛/F .z/ F .˛x C .1 ˛/z/ C ckx zk BY : (12.50)
2. The map F is called compactly -paraconvex with > 1; if there exists a

convex and compact subset ˘ of Y with 0 2 ˘ such that
˛F .x/C.1˛/F .z/ F .˛x C.1˛/z/Cmin.˛; 1˛/kx zk ˘: (12.51)
Evidently, any convex set-valued map is compactly -paraconvex and hence

-paraconvex but the converse may be false. The following example, given by
Jourani [310], shows that -paraconvex set-valued maps need not to be convex:
Example 12.3.9. Let f W R ! R be given by f .x/ D jjxj 1j. Then the set-valued
map F W RR defined by F .x/ D f .x/ C RC is 1-paraconvex but it fails to be
convex.
The following result is a useful implication of the notion of compact -
paraconvexity.
Proposition 12.3.10. Let X and Y be Banach spaces, let F W X Y be a set-
N y/
valued map, let .x; N 2 graph .F /; and let > 1. If F is compactly -paraconvex,
then for every x 2 X we have
F .x/ yN DF .x;
N y/.x
N x/
N C kx xk
N ˘; (12.52)
where ˘ is the convex and compact set appearing in (12.51).

Proof. Let ˛ 2 .0; 1/. Since the map F is compactly -paraconvex, we have
˛F .x/ C .1 ˛/yN ˛F .x/ C .1 ˛/F .x/

N F .˛x C .1 ˛/x/
N C ˛kx xk
N ˘;
which can be rearranged as
F .x/ yN ˛ 1 ŒF .xN C ˛.x x//

N y
N C kx xk
N ˘;
and hence
F .x/ yN lim sup ˛ 1 ŒF .xN C ˛.x x//

N y
N C kx xk
N ˘:
Let y 2 lim sup ˛ 1 ŒF .xN C ˛.x x// N y N C kx xk

N ˘ be arbitrary. Then
there exist sequences f.˛n ; yn /g P Y with .˛n ; yn / ! .0C ; y/; fpn g ˘ and
yN C ˛n .yn kx xk N pn / 2 F .xN C ˛n .x x//.
N
Since ˘ is compact, there exists an infinite subset I N and p 2 ˘ with
.yn kx xk N pn /n2I ! .y kx xk N p/ ; confirming that y 2 DF .x; N y/.x
N
N C kx xk
x/ N ˘ . The proof is complete.

t
u
We have the following special case of the above result:
Corollary 12.3.11. Let X and Y be Banach spaces, where Y is finite-dimensional,
let F W X Y be a set-valued map, and let .x;
N y/
N 2 graph .F /. If F is -paraconvex
with > 1, then for any x 2 X;
F .x/ yN DF.x;
N y/.x
N x/
N C ckx xk
N BY :
In the following, we summarize some other useful features of paraconvex set-

valued maps:
Proposition 12.3.12. Let X , Y and Z be Banach spaces, let F W X Y and G W
X Z be set-valued maps, and let .x;
N y/
N 2 graph .F / and .x;
N zN/ 2 graph .G/.
1. If F and G are -paraconvex, then .F; G/ is -paraconvex.
2. If F and G are compactly -paraconvex, then .F; G/ is compactly -paraconvex.
3. Let > 1. If F and G are -paraconvex, then for every x 2 X;
.F; G/.x/ .y;

N zN/ C .DF.x;
N y/;
N DG.x;
N zN//.x x/
N C kx xk
N .˘1 ˘2 /;
where ˘1 and ˘2 are compact, convex subsets of Y and Z, respectively.

Proof. The proof follows from Proposition 12.3.10 and the involved definitions.
t
u
We are now ready to give the promised sufficient optimality conditions for .P1 /
(with S D X ).
Theorem 12.3.13. Let X; Y; and Z be Banach spaces with X finite-dimensional,
let C Y and D Z be pointed, closed, and convex cones where C is also
solid and C D compactly based, let F W X Y and G W X Z be compactly
-paraconvex set-valued maps with > 1, let .x;N y/

N 2 graph .F / and let zN 2
G.x/N \ .D/. Assume that the following conditions hold:
N y/;
1. .DF.x; N DG.x;N zN// .0/ \ .C; T .D; Nz// D f.0; 0/g.
N y/;
2. .DF.x; N DG.x;N zN// .x/ \ .C; T .D; Nz// D ; for every 0 ¤ x 2
N y//
dom .DF.x; N \ dom .DG.x; N zN//.
N y/
Then .x; N is a local minimizer of .P1 / (with S D X ).
N y/
Proof. Assume that .x; N is not a local minimizer of .P1 /. There there are
sequences fxn g X; fyn g Y; and fzn g Z such that xn ! xN and for every
n 2 N; we have yn 2 F .xn /; zn 2 G.xn / and
.yN yn ; zN zn / 2 .C .D C zN// nf.0; 0/g: (12.53)
For n 2 N; we set ˛n .pn ; qn / WD .yN yn ; zn / ; with
˛n D k.yN yn ; zn /k;

pn D ˛n1 .yN yn /;
qn D ˛n1 .zn /:
Since D is convex and since Nz 2 D; we have D Nz C T .D; Nz/. Moreover,
it follows from (12.53) that for every n 2 N; we have
zN C ˛n qn 2 T .D; Nz/: (12.54)
By using the assumption that the maps F and G are compactly -paraconvex, it
follows from Proposition 12.3.12 that there are sequences fan g ˘1 and fbn g
˘2 so that for any n 2 N; we get
˛n pn kxn xk
N an 2 DF.x;
N y/.x
N n x/;
N (12.55)
Nz ˛n qn kxn xk
N bn 2 DG.x;
N zN/.xn x/:
N (12.56)
Since C D has a compact base, by extracting subsequences if necessary, we

can assume that there are p 2 C and q 2 D with k.p; q/k D 1 such that pn ! p
and qn ! q.
We shall analyze various possibilities. First assume that f˛n g has no convergent
subsequence. Then f˛n g ! 1. It follows from (12.55) and (12.56) that
.p; q/ 2 .DF.x;
N y/;
N DG.x;
N zN// .0/; (12.57)
.p; q/ 2 .C; T .D; Nz// nf.0; 0/g; (12.58)
which is a contradiction to the hypothesis.

Next we assume that f˛n g has a convergent subsequence. By using the same
notation for the subsequences as well, we assume that f˛n g converges to some ˛ 2
Œ0; 1/. We will analyze the cases when ˛ D 0 and ˛ ¤ 0 separately. If ˛ ¤ 0; then
by dividing (12.55) and (12.56) by ˛n and taking the limit as n ! 1; we obtain
.p; ˛ 1 zN q/ 2 .DF.x;
N y/;
N DG.x;
N zN// .0/;
.p; ˛ 1 zN q/ 2 .C; T .D; Nz// nf.0; 0/g;
and hence the first hypothesis is once again contradicted.

For the case when ˛ D 0; we need to analyze the two cases, namely, zN D 0 and
zN ¤ 0. First assume that zN ¤ 0. Then by taking the limit n ! 1 in (12.56), we
obtain z 2 DG.x; N zN/ \ T .D; Nz/ which is a contradiction. Let’s now focus on
the case when zN D 0. In this setting, we first analyze the case when ˛n1 .xn x/
N has
a subsequence converging to some x0 2 X . Dividing (12.55) and (12.56) by ˛n ; we
deduce that
N 1 ˛n1 kxn xka

pn kxn xk N n 2 DF.x; N n1 .xn x//
N y/.˛ N ;
N 1 ˛n1 kxn xkb
qn kxn xk N zN/.˛n1 .xn x//
N n 2 DG.x; N ;
and by taking the limit n ! 1; we obtain
.p; q/ 2 .DF.x;
N y/;
N DG.x;
N zN// .x0 /
.p; q/ 2 .C; T .D; Nz// ;
which is a contradiction to the hypothesis for both cases x0 D 0 and x0 ¤ 0.

Finally, we analyze the case when ˛n1 .xn x/ N has no convergent sequences.
In this case, we have ˛n .kxn xk N 1 / ! 0. Since X is finite-dimensional, by
extracting subsequences if necessary, we can assume that Œ.xn x/kx
N N 1 ! y
n xk
with kyk D 1. Then by dividing (12.55) by kxn xk N and taking the limit n ! 1;
we deduce that 0 2 DF .x; N y/.y/;
N which is again a contradiction. The proof is
complete. t
u
From the above result, we obtain the following sufficient optimality condition
for .P0 / W
Theorem 12.3.14. Let X and Y be Banach spaces where X finite-dimensional,
let C Y be a pointed, closed, convex and compactly based, let F W X Y be
N y/
compactly -paraconvex set-valued map with > 1, and let .x; N 2 graph .F /.
Assume that the following conditions hold:
N y/.0/
1. DF.x; N \ .C / D f0g.
N y/.x/
2. DF.x; N \ .C / D ;; for every 0 ¤ x 2 dom .DF.x;
N y//.
N
N y/
Then .x; N is a local minimizer of .P0 / (with S D X ).
12.4 Second-Order Optimality Conditions in Set-Valued Optimization 549

Semidifferentiability
In this section, to show yet another interesting direction of deriving sufficient

optimality conditions in set-valued optimization, we present a brief discussion of
results obtained by Luc [400]. In the following by using the notion of upper semi-
differentiability, we state sufficient optimality conditions for .P0 / (with S D X )
and (P1 ).
and convex cone, let F W X Y be set-valued, let xN 2 dom .F /; and let yN 2
N C /. Assume that:
Min.F .x/;
\
1. DF.x;N y/.x/
N .C / D ; for every 0 ¤ x 2 dom .DF.x; N y//.
N
\
2. DF.x;N y/.0/
N .C / D f0g.
3. F is compact at xN and upper semidifferentiable at .x;
N y/.
N
N y/
Then .x; N is a local minimizer of .P0 / (with S=X).
Proof. See Luc [400]. t
u
be pointed, closed, and convex cones, let F W X Y and G W X Z be set-valued
maps, let xN 2 dom .F / with G.x/
N \ D ¤ ;; and let yN 2 Min.F .x/;
N C /. Assume
that the following conditions hold:
N upper semidifferentiable at .x;
1. F and G are compact at x; N y/
N and .x;
N zN/ for every
zN 2 G.x/
N \ .D/.
2. .DF.x;N y/.x/;
N N zN/.x// \ .C; T .D; Nz// D ; for every 0 ¤ x 2
DG.x;
dom .DF.x; N y//
N \ dom .DG.x; N zN// and zN 2 G.x/
N \ .D/.
3. .DF.x;N y/.0/;
N N zN/.0// \ .C; T .D; Nz// D f0g for every zN 2 G.x/
DG.x; N \
.D/.
N y;
Then .x; N zN/ is a local minimizer of .P1 / for every zN 2 G.x/
N \ .D/.
Proof. See Luc [400]. t
u
12.4 Second-Order Optimality Conditions in Set-Valued

Optimization
In this section, we employ various second-order derivatives of set-valued maps

to give second-order optimality conditions in set-valued optimization. We note
that given a set-valued map F , we shall use the second-order contingent
N y;
derivative D 2 F .x; N uN ; vN /, the second-order lower Dini derivative DL2 F .x;
N y;
N uN ; vN /,
the second-order asymptotic derivative D2 F .x; N y;
N uN ; vN /; and the second-order
compound derivative D 2 F .x; N y;

N uN ; vN /. Here .Nu; vN / is the direction in which the
second-order derivative is being computed. In view of the properties of the
second-order tangent sets and the second-order tangent cones, it is evident that
the second-order objects can only be nonempty if the directions are included in
the first-order tangent cones. This, and the first-order optimality conditions, impose
restriction on .Nu; vN /. Such conditions are always evident from the context. Therefore,
for notational simplicity, throughout this work we will work under the assumption
that the directions are suitably chosen and the second-order objects are non-trivial.
This section is based on [283, 300, 313, 321, 322]. See also [381, 385, 626]

In this subsection, we present a second-order extension of the Dubovitski-Milyutin

approach explored in Sect. 12.2. We begin with the following necessary optimality
condition for (P1 ):
Theorem 12.4.1. Let X; Y; and Z be normed spaces and let C Y and D Z
be pointed, closed, convex, and solid cones. Let F W X Y and G W X Z be set-
valued maps. For xN 2 S X; let .x; N y/
N 2 graph .F / be a local weak-minimizer
of (P1 ). For .Nu; vN / 2 X Y; let .F C C / be second-order asymptotic derivable at
N y;
.x; N uN ; vN / and Lipschitz-like around .x;
N y/.
N Then
D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1 \ IT2 .GŒD1 ; x;

N y; N uN / \ T2 .S; x;
N uN / D ;:
(12.59)
Proof. Since .x;N y/

N is a local weak-minimizer of .P1 /, there exists a neighborhood
U1 of xN such that yN 2 WMin.F .S1 \ U1 /; C /. Therefore, F .S1 \ U1 / \ .fyg N
intC / D ;. We will show that if (12.59) fails, then a feasible u can be constructed
in a suitable vicinity of xN such that F .u/ \ .yN intC / ¤ ;; and hence violating the
local weak minimality of .x; N y/.
N
For the sake of argument, we assume that there exists an x 2 X that
violates (12.59), that is
x 2 D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1 \ IT2 .GŒD1 ; x;

N y; N uN / \ T2 .S; x;
N uN /:
(12.60)
Since x 2 T2 .S; x; N uN / from (12.60), the definition of the second-order asymptotic

contingent cone ensures that there are a sequence fxn g X with xn ! x and
a sequence f.sn ; tn /g P P with .sn ; tn / # .0; 0/ and sn =tn ! 0 such that
From x 2 IT2 .GŒD1 ; x; N uN / and the facts that xn ! x; .sn ; tn / # .0; 0/ and
sn =tn ! 0; there exists n1 2 N such that xN C sn uN C sn tn xn 2 GŒD1 for every
n
n1 ; or G.xN Csn uN Csn tn xn /\D ¤ ; for every n
n1 . Due to the containment
x 2 D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1 ; there exists y 2 IT.C; vN / with .x; y/ 2
N y;
T .epi .F /; .x;
2
N y/;
N .Nu; vN //. Hence, there are a sequence f.xO n ; yOn /g X Y with
.xO n ; yOn / ! .x; y/ and a sequence f.pn ; qn /g P P with .pn ; qn / # .0; 0/ and
pn =qn ! 0 such that yN C pn vN C pn qn yOn 2 F .xN C pn uN C pn qn xO n / C C; for every
n 2 N.
By using the second-order asymptotic derivability of epi .F /, we can set pn D sn
and qn D tn . Next, we define un WD .xCs N n uN Csn tn xn /; and uO n WD .xCs
N n uN Csn tn xO n /.
Since both fun g and fOung converge to x; N there exists n2 2 N such that un ; uO n 2
U WD U1 \ U2 for n
n2 where U1 is the neighborhood of xN described above
and U2 is a neighborhood of xN which exists, along with a neighborhood V of y; N as a
consequence of the Lipschitz-like property. Moreover, since .yCs N n vN Csn tn yOn / ! y;
N
there exists n3 2 N such that yN C sn vN C sn tn yOn 2 V for all n
n3 .
Since .F C C / is Lipschitz-like at .x; N y/;
N we get
yN C sn vN C sn tn yOn 2 ŒF .xN C sn uN C sn tn xO n / C C \ V .for n

n3 /
F .xN C sn uN C sn tn xn / C C
CL sn tn kxn xO n kBY .for n
maxfn1 ; n2 ; n3 g/:
In view of the above inclusion, there exists bn 2 BY such that for n
maxfn1 ; n2 g; we have yN C sn vN C sn tn yn 2 F .xN C sn uN C sn tn xn / C C; where

yn D .yOn Lbn kxn xO n k/ ! y.
Since y 2 IT.intC; vN /; tn # 0 and yn ! y; it follows from the definition of the
interiorly contingent cone that there exists n4 2 N such that for n
n4 ; we have
vN Ctn yn 2 intC . Using the fact that sn > 0; we obtain yN Csn vN Csn tn yn 2 yN intC .
Next we choose wn 2 F .un / such that yN C sn vN C sn tn yn 2 wn C C for every n 2 N.
We have wn 2 yN C sn vN C sn tn yn C yN intC C yN intC; and consequently
wn 2 F .un / \ .yN intC /; for n
maxfn1 ; n2 ; n3 ; n4 g.
Therefore, we have shown that for every n
maxfn1 ; n2 ; n3 ; n4 g there are un 2
S1 \ U1 such that F .un / \ .yN intC / ¤ ;; contradicting the weak optimality of
.x;N y/.
u
Let us now deduce the following necessary optimality condition for (P0 ):
convex, and solid cone, and let S X . Let F W X Y be a set-valued map and
N y/
let .x; N 2 graph .F / be a local weak minimizer of (P0 ). For .Nu; vN / 2 X Y ,
let .F C C / be second-order asymptotic derivable at .x;
N y;
N uN ; vN / and Lipschitz-like
N y/.
around .x; N Then
D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1 \ T2 .S; x;

N y; N uN / D ;: (12.61)
Clearly, Theorem 12.4.1 and Theorem 12.4.2 remain valid if the asymptotic con-
tingent derivatives and cones are replaced by second-order contingent derivatives.
Note that Theorem 12.4.1 does not involve any derivative of the constraint map G.
Therefore, in order to give multiplier rules, it becomes necessary to either impose
some constraint qualifications that connect the asymptotic contingent derivative of
the map G to the cone IT2 .GŒD1 ; x; N uN /. Another possibility that serves this
purpose is furnished in the following:
Theorem 12.4.3. Let X and Z be normed spaces and let D Z be a pointed,
closed, convex, and solid cone. Let G W X Z be a set-valued map, let .x; N zN/ 2
graph .G/; and let .Nu; w/ N 2 X Z. Assume that epi .G/ is second-order asymptotic
N zN; uN ; w/
derivable at .x; N and .G C D/ is Lipschitz-like around .x; N zN/. Then the
following inclusion holds:
D2 .G C D/.x;
N zN; uN ; w/ŒIT
N 2
N 1 IT2 .GŒD1 ; x;
.D; zN; w/ N uN / (12.62)
Proof. Let x 2 D2 .G C D/.x; N zN; uN ; w/ŒIT

N 2
N 1 be arbitrary. Then there
.D; zN; w/
exists an element z 2 Z such that z 2 D .G C D/.x; 2
N y;
N uN ; w/.x/
N \ IT2 .D; zN; uN /.
Since .x; z/ 2 T .epi .G/; .x;
2
N zN/; .Nu; w//;
N there are a sequence f.xO n ; zOn /g X Z
with .xO n ; zOn / ! .x; z/ and a sequence f.pn ; qn /g P P with .pn ; qn / # .0; 0/ and
pn =qn ! 0 so that for every n 2 N; we have zN C pn wN C pn qn zOn 2 G.xN C pn uN C
pn qn xO n / C D.
Choose arbitrary sequences fxn g X with xn ! x and f.sn ; tn /g P P with
.sn ; tn / # .0; 0/ and sn =tn ! 0. To show that x 2 IT2 .GŒD1 ; x; N uN /; it suffices
to show that there exists m 2 N such that xN C sn uN C sn tn xn 2 GŒD1 for every
n
m; or G.xN C sn uN C sn tn xn / \ D ¤ ; for every n
m.
In view of the derivability, we can set pn D sn and qn D tn . Furthermore, by
following the same lines of argument as in the proof of Theorem 12.4.1, by using
the Lipschitz-like property, we can show that there exists n1 2 N such that for every
n
n1 ; we have zN C sn wN C sn tn zn 2 G.xN C sn uN C sn tn xn / C D; where for some
bn 2 BY ; we have zn D .Ozn Lbn kxn xO n k/ ! z. Because z 2 IT2 .D; zN; w/; N we
also have zN C sn w N C sn tn zn 2 D for sufficiently large n 2 N.
We now choose wn 2 G.xN C sn uN C sn tn xn / so that zN C sn wN C sn tn zn 2 wn C D
implying that wn 2 .Nz C sn w N C sn tn zn / D and hence for sufficiently large n 2 N,
we have wn 2 D. Summarizing, for sufficiently large n; we have xN C sn uN C
sn tn xn 2 GŒD1 . This, however, confirms that x 2 IT2 .GŒD1 ; x; N uN /. The
u
A direct consequence of the above result is the following analogue of Theo-
rem 12.4.1:
G W X Z be set-valued maps. Let .x; N y/
of (P1 ) and let zN 2 G.x/
N \ .D/. For .Nu; vN ; w/
N 2 X Y Z, let .F C C / be
N y;
second-order asymptotic derivable at .x; N uN ; vN / and Lipschitz-like around .x;
N y/.
N
Let .G C D/ be second-order asymptotic derivable at .x;

N zN; uN ; w/
N and Lipschitz-like
N zN/. Then
around .x;
D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1 \ D2 .G C D/.x;

N y; N y;
N uN ; w/
N
N 1 \ T2 .S; x;
ŒIT2 .D; zN; w/ N uN / D ;: (12.63)
For notational simplicity, for the following Lagrange multiplier rule, we set
L WD IT.C; Nv/I
N WD IT2 .D; Nz; w/I
N
P WD D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1 I
N y;
Q WD D2 .G C D/.x;
N y;
N uN ; w/ŒIT
N 2
N 1 :
.D; zN; w/
Theorem 12.4.5. Let X; Y; and Z be normed spaces, let C Y and D Z be

pointed, closed, convex, and solid cones, and let S X . Let .x; N y/ N 2 graph .F / be
a local weak minimizer of (P1 ) and let z 2 G.x/\.D/.
N For .Nu; vN ; w/
N 2 X Y Z,
let .F C C / be second-order asymptotic derivable at .x; N y;
N uN ; vN / and Lipschitz-like
around .x;N y/.
N Let .G C D/ be second-order asymptotic derivable at .x; N zN; uN ; w/
N
and Lipschitz-like around .x; N zN/. Assume that P and Q are open convex cones and
assume that there exists a nonempty, closed, and convex cone M T2 .Q; x; N uN /. Let
D2 .F C C /.x;N y;
N uN ; vN / W X Y be L-convex and D2 .G C D/.x; N y;
N uN ; w/
N W X Z
be N -convex. Then there exist functionals p 2 M C ; t 2 LC ; s 2 N C ; not all zero,
such that for every x 2 X; the following inequality holds
t ı D2 .F C C /.x;
N y;
N uN ; vN /.x/ C s ı D2 .G C D/.x;
N zN; uN ; w/.x/
N
p.x/: (12.64)
Proof. We will prove the theorem by analyzing the following three possibilities:
(i) P D ;;
(ii) Q D ;;
(iii) P ¤ ; and Q ¤ ;.
Let us begin with case (i) where P D ;; we have
D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1 D ;:

N y;
Then it follows from Theorem 12.2.15 that there exists t 2 LC nf0g such that
t ı D2 .F C C /.x;
N y;
N uN ; vN /
0 for every x 2 X:
We choose s D 0 and p D 0 and obtain the desired result.

For case (ii) where Q D ;; we have
D2 .G C D/.x;
N y;
N uN ; w/ŒIT
N 2
N 1 D ;:
.D; zN; w/
By invoking Theorem 12.2.15 once again, we obtain s 2 N C nf0g such that
s ı D2 .G C D/.x;
N y;
N uN ; w/.x/
N
0 for every x 2 X:
By setting t D 0; and p D 0; we obtain the desired result.

N y/
We now consider case (iii). Since .x; N is a weak minimizer of .P1 /; Theo-
rem 12.4.4 yields:
P \ Q \ M D ;:
Since P; Q; and M are all nonempty, we can apply the Dubovitskii-Milyutin

lemma to assure the existence of
l 2 M ;

l0 2 D2 .F C C /.x; N uN ; vN /ŒIT.C; vN /1
N y;

l1 2 D2 G.x;
N y;
N uN ; w/ŒIT
N 2
N 1 ;
.D; zN; w/
such that
l C l0 C l1 D 0: (12.65)
Now, in view of Theorem 12.2.15, we get the existence of functionals t 2 LC

and s 2 N C such that for all x 2 X; the following inequalities hold
t ı D2 .F C C /.x;
N y;
N uN ; vN /.x/
l0 .x/I
s ı D2 .G C D/.x;
N y;
N uN ; w/.x/
N
l1 .x/:
Combining the above inequalities with (12.65) and setting p D l

yields (12.64). t
u
12.4.2 Second-Order Optimality Conditions by the Direct

Approach
In this section, we give some second-order optimality conditions by using the direct
approach. The main objective here is that we want to shed some light on the
approach which is more commonly used in set-valued optimization and contrast
it with the Dubovitskii-Milyutin approach. We begin with the following:
G W S Z be set-valued maps, let .x;N y/ N 2 graph .F / be a local weak minimizer

of .P1 /; and let zN 2 G.x/
N \ .D/. For .Nu; vN ; w/ N 2 X Y Z, let .F C C / be
second-order asymptotic derivable at .x; N y; N uN ; vN / and Lipschitz-like around .x;
N y/.
N
Then for every x 2 dom ..D2 F .x;
N y;
N uN ; vN /; D2 G.x; N zN; uN ; w//;
N we have
N y;
.D2 F .x; N uN ; vN /; D2 G.x;
N zN; uN ; w/.x/\I
N T .C; vN /IT2 .D; zN; w/
N D ;: (12.66)
Proof. Assume that (12.66) is false. Then there exists x 2 dom ..D2 F .x;
N y;
N uN ; vN /; D2
N zN; uN ; w//
G.x; N with
N y;
.y; z/ 2 .D2 F .x; N uN ; vN /; D2 G.x;
N zN; uN ; w//.x/
N \ I T .C; vN / IT2 .D; zN; w/:
N
Consequently, we have .x; y/ 2 T2 .graph .F /; .x; N y/;

N .Nu; vN // and .x; z/ 2
T .graph .G/; .x;
2
N zN/; .Nu; w//.
N Since .x; y/ 2 T2 .graph .F /; .x; N y/;
N .Nu; vN //; there are
a sequence f.xO n ; yOn /g X Y with .xO n ; yOn / ! .x; y/ and a sequence f.pn ; qn /g
P P with .pn ; qn / # .0; 0/ and pn =qn ! 0 so that for every n 2 N; we have
yN C pn vN C pn qn yOn 2 F .xN C pn uN C pn qn xO n /.
From .x; z/ 2 T2 .graph .G/; .x; N zN/; .Nu; w//,
N we deduce that there are a sequence
f.xN n ; zNn /g X Y with .xN n ; zNn / ! .x; z/ and a sequence f.sn ; tn /g P P
with .sn ; tn / # .0; 0/ and sn =tn ! 0 so that for every n 2 N; we have zN C sn wN C
sn tn zNn 2 G.xN C sn uN C sn tn xN n /. By using the second-order asymptotic derivability of
graph .G/; we can set sn D pn and tn D qn . Furthermore, by the arguments given in
the proof of Theorem 12.4.1, we can show that for sufficiently large n there exists a
feasible sequence fan g that violates the weak minimality of .x; N y/.
N t
u
The following second-order necessary optimality for .P0 /, which employs the
notion of the contingent derivative, can be proved by arguments used above.
convex, and solid cone, let S X; and let F W S Y be a set-valued map. Let
N y/
.x; N 2 graph .F / be a local weak minimizer of .P0 /. ThenT for every uN 2 ˝0 WD
dom .D.F C C /.x; N y//;
N for every vN 2 D.F C C /.x;N y/.N
N u/ .@C / (@ denotes the
boundary ) and every x 2 ˝1 WD dom .D 2 .F C C /.x;N y;
N uN ; vN //; we have
\
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ .intC fNvg/ D ;: (12.67)
Remark 12.4.8. Note that by setting .Nu; vN / D .0; 0/ in (12.67), we obtain the
following first order optimality condition: For every x 2 dom .D.F C C /.x;N y//;
N
we have D.F C C /.x; N y/.x/
N Y nfintC /g.
Theorem 12.2.1 extends the second-order theory known from nonlinear program-
ming to the general set-valued optimization (and nonsmooth vector optimization)
case. For instance, the following result, which generalizes [86, Theorem 5.3], is a
direct consequence of Theorem 12.2.1.
Corollary 12.4.9. Let Y D R and let C D RC WD ft 2 Rj t

0g in
Theorem 12.2.1. Then:
D.F C C /.x;
N y/.x/
N RC for every x 2 ˝0 WD dom .D.F C C /.x;
N y//:
N
Furthermore, for uN 2 ˝0 with 0 2 D.F C C /.x;

N y/.N
N u/; we have
D 2 .F C C /.x;
N y;
N uN ; 0/.x/; RC for every x 2 dom .D 2 .F C C /.x;
N y;
N uN ; 0//:
The following example illustrates the above result.

Example 12.4.10. Consider the following set optimization problem
min F .x/
x2R
with F W RR given by F .x/ WD fy 2 R j y

x 2 g for everyx 2 R.
N y/
Clearly, .x; N WD .0; 0/ is a local weak minimizer. For every x 2 R; we have
D.F C C /.0; 0/.x/ D fy 2 R j .x; y/ 2 T .epi .F /; .0; 0//g D RC ;
and for every uN 2 R and for every x 2 R; we have

˚
D 2 .F C C /.x;
N y;
N uN ; 0/.x/ D D 2 .F C C /.0; 0; uN ; 0/.x/ D y 2 R j y
2Nu2 :
Summarizing, for the local minimizer .0; 0/; for uN 2 R; and for every x 2 R;
we have
D.F C C /.0; 0/.x/ D RC ;

D .F C C /.0; 0; uN ; 0/.x/ RC ;
2
that is, the necessary optimality conditions of Corollary 12.4.9 are fulfilled.
We next give optimality conditions by using the second-order compound contin-
gent derivative defined by taking the contingent cone of the contingent cone. Since
this approach allows to combine two different tangent cone of first-order, it seems
to have some potential ( see [283]).
The following is yet another necessary optimality conditions for (P1 ).
G W S Z be set-valued maps, let .x; N y/
of .P1 /; and let zN 2 G.x/N \ .D/. Let .Nu; vN ; w/N 2 X Y Z. Then, for every
x 2 dom .D 2 .FG/.x; N y;
N zN; uN ; vN ; w//;
N we have
D 2 .FG/.x;
N y;
N zN; uN ; vN ; w/.x/
N \ IT.C; vN / IT.IT.D; zN/; w/
N D ;: (12.68)
12.5 Generalized Dubovitskii-Milyutin Approach in Set-Valued Optimization 557
Proof. Assume that (12.68) is violated. Then there exists x 2 dom .D 2 .FG/.x;
N y;
N zN;
uN ; vN ; w//
N such that
.y; z/ 2 D 2 .F; G/.x;

N y;
N zN; uN ; vN ; w/.x/
N \ IT.C; vN / IT.IT.intD; zN/; w/;
N
which implies that .x; y; z/ 2 T .T .graph .F; G/; .x; N y;N zN//; .Nu; vN ; w//.
N Therefore,
there are exists ftn g P; f.xn ; yn ; zn /g X Y Z such that tn # 0; .xn ; yn ; zn / !
.x; y; z/ and .Nu C tn xn ; vN C tn yn ; wN C tn zn / 2 T .graph .F; G/; .x; N y;
N zN//. Since
y 2 IT.intC; vN /; there exists n1 2 N such that vN C tn yn 2 intC for n >
n1 . Analogously, since z 2 IT.IT.intD; zN/; w/; N there exists n2 2 N such that
wN C tn zn 2 IT.intD; zN/ for n > n2 . For n
maxfn1 ; n2 g; we fix elements
un WD uN C tn xn ; vn WD vN C tn yn ; and wn WD wN C tn zn . Note that we have
.un ; vn ; wn / 2 T .graph .F G/; .x; N y;N zN// and .vn ; wn / 2 intC IT.intD; zN/.
In view of the definition of the contingent cone, there are sequences ftm g P;
f.xm ; ym ; zm /g X Y Z; such that tm # 0; .xm ; ym ; zm / ! .un ; vn ; wn /
and .yN C tm ym ; zN C tm zm / 2 .F; G/.xN C tm xm / which means that yN C tm ym 2
F .xN C tm xm / C C and zN C tm zm 2 G.xN C tm xm / C D. Since ym ! vn and
vn 2 intC; there exists m1 > 0 such that tm ym 2 intC for every m > m1 .
Let am 2 F .xN C tm xm / be such that yN intC 2 am C C; and consequently we have
am 2 yN intC . Similarly, because wn 2 IT.intD; zN/; there exists m2 > 0 such
that zN C tm zm 2 intD for every m > m4 . Let bm 2 G.xN C tm xm / be such that
zN C tm zm 2 bm C D; and consequently we have bm 2 intD.
Therefore, we have shown that for sufficiently large m; we have cm WD xN C
tm xm 2 U.x/; N G.cm / \ D ¤ ; and F .cm / \ .yN intC / ¤ ;: This contradicts the
optimality of .x;N y/.
N t
u
The following necessary optimality conditions for (P0 ) following at once.
convex, and solid cone, and let S X . Let F W S Y be a set-valued map and
N y/
let .x; N 2 graph .F / be a local weak minimizer of .P0 /. Then, for every x 2
N y;
dom .D 2 .F /.x; N uN ; vN //; we have
D 2 F .x;
N y;
N uN ; vN /.x/ \ IT.int.C /; vN / D ;: (12.69)
Remark 12.4.13. Lagrange multiplier rules can be obtained by combining the

contingent cone and the Clarke tangent cone to define new convex derivatives.
12.5 Generalized Dubovitskii-Milyutin Approach

in Set-Valued Optimization
In this section, we consider a more general set-valued optimization problem and give
necessary optimality conditions by giving an extension of the Dubovitskii-Milyutin
approach.
Let I WD f1; 2; : : : ; mg and J WD f1; 2; : : : ; ng be two index sets. Let X and

Y be Banach spaces, let C Y be a pointed, closed, convex, and solid cone, let
S X; and let F W X Y be a given set-valued map. For i 2 I; let Zi be a normed
space, let Ci Zi be a pointed, closed, convex, and solid cone, and let Gi W X Zi
be a given set-valued map. Analogously, for j 2 J; let Wj be a Banach space, let
Dj Wj be closed and convex set, and let Hj W X Wj be a given set-valued
map.
In this section, we focus on the following set-valued optimization problem
.GP1 /:
Minimize F .x/ subject to x 2 S1 WD fx 2 S j Gi .x/ \ Ci ¤ ;;

8 i 2 I; Hj .x/ \ Dj ¤ ;; 8 j 2 J g:
Here we only study the local weak-optimality, however, other optimality notions
can be treated analogously. We will also focus on the following particular case:
Minimize F .x/ subject to x 2 S1 WD fx 2 S j Gi .x/ \ Ci ¤ ;; 8 i 2 I g:
By choosing the maps Gi and Hj to be single valued, from (P1 ) we recover the
following set-valued optimization problem with single-valued constraints:
Minimize F .x/ subject to x 2 fx 2 S j Gi .x/

0; 8 i 2 I; Hj .x/ D 0; 8 j 2 J g:
We remark that the above optimization problem with multi-equality constraints

can not be tackled by the classical Dubovitskii-Milyutin approach studied in the
previous sections. This is mainly due to the fact that the separation arguments used
in the classical Dubovitskii-Milyutin approach are applicable to an intersection
of cones in which at most one cone can be closed. On the other hand, a proper
formulation of multi-equality constraints leads to an intersection with several closed
cones. To handle optimization problems with multi-equality constraints, a gener-
alized Dubovitskii-Milyutin theory has been developed for smooth optimization
problems (see [334]). This theory is enriched by the notion of cones of the same
sense and cones of the opposite sense introduced by Walczak [585]. Although
many interesting contributions have appeared in this direction, the generalized
Dubovitskii-Milyutin theory so far has only been used to study vector optimization
problems with differentiable data. The primary objective of this section is to
present an extension of the generalized Dubovitskii-Milyutin theory to set-valued
optimization problems. In this process, we also present new extensions of this theory
to nonsmooth optimization and to more general vector optimization problems.
12.5.1 A Separation Theorem for Multiple Closed and Open

Cones
This section is devoted to some specific properties of cones and their use in giving
a separation theorem for a disjunction involving multiple closed and multiple open
cones.
We begin with the following notion (see Walczak [585]):
Definition 12.5.1. Let X be a normed space, let fKi gkiD1 be a system of cones in X .
Then:
1. The cones fKi gkiD1 are of the same sense , if there exists "1 ; : : : ; "k > 0 such that
for any x 2 B" \ .K1 C C Kk / where x D x1 C C xk with xi 2 Ki ; we
have
xi 2 B"i \ Ki for i D 1; : : : ; k:
2. The cones fKi gkiD1 are of the opposite sense, if there exists nontrivial vectors
fxi gni , that is, .x1 ; : : : ; xk / ¤ .0; : : : ; 0/ with xi 2 Ki ; such that x1 C x2 C C
xk D 0.
Remark 12.5.2. 1. It follows from the definition that the cones fKi gkiD1 are of the
same sense if for any " > 0 there exists "1 ; : : : ; "k > 0 such that for each x D
x1 C x2 C : : : C xk ; with xi 2 Ki ; the inequality kxk " implies the inequalities
kxi k "i i D 1; 2; : : : :n.
2. We also note that the system of same sense is disjoint with the system of opposite
sense. Furthermore, any subsystem of a system of the same sense is of the same
sense. If any subsystem is of opposite sense, then the whole system is of the
opposite sense.
3. If two cones K1 and K2 are subspaces and if K1 \K2 contains a nonzero element
then these cones are of the opposite sense.
We following result can be found in Walczak [585, Theorem 3.5].
Theorem 12.5.3. Let X be a Banach space and let X be the topological dual of
X . Then the algebraic sum of a finite number of weak* closed convex cones of the
same sense is weak* closed.
With the above preparation, we are ready to prove the following result given by
Kotarski [333]:
˚
Theorem 12.5.4. Let X be a Banach space and let K1 ; K2 ; : : : ; Kp ; : : : ; Kn be
a collection of cones in X . Assume that the following conditions hold:
1. The cones K1 ; K2 ; : : : ; Kp are open and convex with vertices at 0.
2. The cones KpC1 ; : : : ; Kn are closed and convex with vertices at 0.
C
3. The dual cones KpC1 ; : : : ; KnC are either of the same sense or the opposite sense.
Then the following two statements are equivalent:

\
n
I. Ki D ;.
i D1
II. There exist functions ì 2 KiC ; i D 1; : : : ; n; not all simultaneously zero,
such that
`1 C `2 C C `n D 0:
C
Proof. Let us consider the case when the cones KpC1 ; : : : ; KnC are of the same
sense because the case when the cones are of the opposite sense follows trivially.
We set
\
p
K WD Ki ;
i D1
\
n
C WD Ki :
i DpC1
There are the following four possibilities:

1. K ¤; C ¤ ;;
2. K D; C ¤ ;;
3. K ¤; C D ;;
4. K D; C D ;.
We begin by analyzing the first scenario. Clearly K is an open convex cone with
vertex at the origin and C is a convex cone with vertex at the origin. Therefore,
by using the classical separation theorem, there exists a nonzero linear continuous
functional ` such that
!C
\
p
`2 Ki
i D1
0 1C
\
n
` 2 @ Ki A :
i DpC1
It is known that ` has the following representation:
X
p
`D ì where ì 2 KiC i D 1; : : : ; p: (12.70)
i D1
Furthermore, from Theorem 12.5.3, we have

0 1C 0 1 0 1
\
n X
n X
n X
n
@ Ki A D wcl conv @ KiC A D wcl @ KiC A D KiC ;
i DpC1 i DpC1 i DpC1 i DpC1
where wcl denotes the closure in the weak-star topology. Consequently, ` admits
the representation
X
n
`D ì where ì 2 KiC i D p C 1; : : : ; n: (12.71)
i DpC1
P
Using (12.70) and (12.71), we get niD1 ì D 0 and the proof is complete for
this case.
\p
For the second case, we have Ki D ;; which implies that there exists 1
i D1
s < p such that
\
s
Ki ¤ ;;
i D1
\
sC1
Ki D ;;
i D1
and the conclusion follows by using a separation theorem and setting `sC2 D D
`n D 0.
The remaining two cases can be proved in an analogous manner.
For the converse, we assume that
X
n
ì D 0;
i D1
and, if possible, assume that
\
n
Ki ¤ ;;
i D1
Let x 2 Ki for i D 1; : : : ; n be arbitrary. For a certain 1 j p; we have

`j ¤ 0. However, since Kj is an open cone, we also have `j .x/ > 0; implying
X
n
0D ì .x/
`j .x/ > 0
i D1
which is impossible. The proof is complete. t

u
We conclude this section by formulating the following notion:

Definition 12.5.5. A system of sets f˝1 ; ˝2 ; : : : ; ˝n g is called optimally posi-
tioned, if for any z 2 ˝1 \ ˝2 \ \ ˝n ; we have
T .˝1 \ ˝2 \ \ ˝n ; z/ D T .˝1 ; z/ \ T .˝2 ; z/ \ \ T .˝n; z/: (12.72)
Numerous conditions are given in Chap. 4 which ensure (12.72).
12.5.2 First-Order Generalized Dubovitskii-Milyutin Approach
In the following, we give some technical results to be used later.

N y/
Proposition 12.5.6. Let .x; N 2 graph .F / be a local weak-minimizer to .GP/.
Then:
\ \ \
m \
n
U S F ŒyN int.C /1 Gi ŒCi 1 Hj ŒDj 1 D ;; (12.73)
i D1 j D1
where U is a neighborhood of xN corresponding to the definition of the local weak-

minimality.
Proof. Assume that there exists
\ \ \
m \
n
x2U S F ŒyN int.C /g1 Gi ŒCi 1 Hj ŒDj 1 :
i D1 j D1
From the containment
\
m \
n
x2S Gi ŒCi 1 Hj ŒDj 1 ;
i D1 j D1
we notice that x 2 S and for all i 2 I and all j 2 J; we have
Gi .x/ \ Ci ¤ ;;
Hj .x/ \ Dj ¤ ;;
ensuring that x is a feasible point. From x 2 U \ F ŒyN int.C /g1 ; we obtain that
there exists x 2 U; a neighborhood of x;N such that F .x/ \ .yN int.C // ¤ ;. This,
however, in view of the feasibility of x, contradicts the local weak-minimality of
N y/.
.x; N t
u
The above result then leads to the following new optimality condition:
N y/
Proposition 12.5.7. Let .x; N 2 graph .F / be a local weak minimizer to (P ).
Assume that the sets Hj ŒDj 1 for j 2 J; are optimally positioned. Then:
\ \
m \
n
N
IT.S; x/ IT.F ŒyN int.C /1 ; x/
N IT.Gi ŒCi 1 ; x/
N T .Hj ŒDj 1 ; x/
N D ;:
iD1 j D1
(12.74)
Proof. In view of Proposition 12.5.6 and the properties of the interiorly contingent
cones and the contingent cones, we obtain
\ \
m
N
N IT.Gi ŒCi 1 ; x/
N
i D1
\ \
n
T. Hj ŒDj 1 ; x//
N D ;:
j D1
The assertion then follows by the assumption that the sets Hj ŒDj 1 for j 2 J
are optimally positioned and hence
\
n \
n
T. Hj ŒDj 1 ; x//
N D T .Hj ŒDj 1 ; x/;
N
j D1 j D1

u
We will also need the following result where the lower Dini derivative is being
used:
Proposition 12.5.8. Let X and Y be normed spaces, let F W X Y be a set valued
N y/
map and let .x; N 2 graph .F /. Let C be a proper, closed, convex, and solid cone.
Then:
N y/ŒintC
DL F .x; N 1 IT.F ŒyN intC 1 ; x/:
N
Proof. Let x 2 DL F .x; N y/ŒintC

N 1 be arbitrary. Therefore, there exists y 2
N y/.x/
DL F .x; N \ intC . Let fxn g X and f
n g P be arbitrary sequences
such that xn ! x and
n # 0. It suffices to show that there exists m 2 N
such that xN C
n xn 2 F ŒyN intC 1 for every n
m. By the definition
N y/./;
of DL F .x; N there exist .yn / Y with yn ! y and n1 2 N such that
yN C
n yn 2 F .xN C
n xn / for every n
n1 . Since y 2 intC and yn ! y;
there exists n2 2 N such that
n yn 2 intC for every n
n2 . This implies that
yN C
n yn 2 F .xN C
n xn / \ .yN intC / for n
m WD maxfn1 ; n2 g. Hence for the
sequences .xn / and .
n /; we have xN C
n xn 2 F ŒyN intC 1 for n
m. This is
equivalent to saying that x 2 I T .F ŒyN intC 1 ; x/.
u
Proposition 12.5.9. Let X and Z be normed spaces, let G W X Z be a set valued

N zN/ 2 graph .G/. Let A Z with int.A/ ¤ ;. Then the following
map and let .x;
holds:
N zN/ŒIT.A; zN/1 IT.GŒA1 ; x/:

DL G.x; N
Proof. Let u 2 DL G.x; N zN/ŒIT.A; zN/1 be arbitrary. Let fung X and f

n g
P be arbitrary sequences such that un ! u and
n # 0. It suffices to show that
there exists m 2 N such that xN C
n un 2 GŒA1 for every n
m. Since u 2
DL G.x;N zN/ŒIT.A; zN/1 ; there exists v 2 DL G.x; N zN/.u/ \ IT.A; zN/. Therefore,
there are a sequence fvn g Z and an integer n1 2 N such that vn ! v and
zN C
n vn 2 G.xN C
n un / for every n
n1 . Because of the containment v 2
IT.A; zN/; there exists n2 2 N such that zN C
n vn 2 A for every n
n2 . Therefore
we have zN C
n vn 2 G.xN C
n un / \ .A/ for every n
m WD maxfn1 ; n2 g.
Consequently, u 2 IT.GŒA1 ; x/. N t
u
We recall the following notion.
Definition 12.5.10. Let X and Y be normed spaces and let R W X Y be a set-
valued map. The map R is called Dini locally convex at .x;N y/
N 2 graph .R/; if
N y/
the lower Dini derivative DL R.x; N of R at .x;
N y/
N is a convex set-valued map.
N y/
The map R is called Dini locally regular at .x; N 2 graph .R/; if additionally
dom .DL R.x;N y//
N D X and the weak-inverse image DL R.x; N y/
N with respect to an
open cone is an open cone.
Given A R and b 2 R; by the inequality A
b we understand that a
b
for every a 2 A. With this convention in mind we are ready to give the promised
multiplier rule.
Theorem 12.5.11. Let .x; N y/
N 2 graph .F / be a local weak-minimizer of (GP1 )
and let zNi 2 Gi .x/
N \ .Ci / where i 2 I WD f1; 2; : : : ; mg. Let there exist an open
convex cone L IT.S; x/. N Assume that the sets fHj ŒDj 1 gnj D1 are optimally
positioned. Assume that Mj WD T .Hj ŒDj 1 ; x/ N for j 2 J are nontrivial and
convex and fMjCgnj D1 are either of the same sense or the opposite sense. Let F be
Dini locally regular at .x; N y/
N and let Gi be Dini locally regular at .x;
N zNi / for i 2 I .
Then there exist functionals s 2 LC ; t 2 C C ; ui 2 .T .Ci ; Nzi //C , vj 2 MjC ; not
all zero, such that ui .Nzi / D 0. Moreover, the following inequality holds for every
x2X W
N y/.x/
t ı DL F .x; N C u1 ı DL G1 .x;
N zN1 /.x/ C C um ı DL Gm .x;
N zNm /.x/

s.x/ C v1 .x/ C C vn .x/: (12.75)
Proof. Note that the identity

\ \
m \
n
N
N IT.Gi ŒCi 1 ; x/
N T .Hj ŒDj 1 ; x/
N D ;;
i D1 j D1
and the following inclusions
N y/ŒintC
DL F .x; N 1 IT.F ŒyN intC 1 ; x/;
N
N zN/ŒIT.Ci ; zNi /1 IT.GŒCi 1 ; x/;
DL Gi .x; N
which follow from Propositions 12.5.8 and 12.5.9 (by taking A D Ci and G D Gi )
ensures that
\ \
m
N
IT.S; x/ N y/ŒintC
DL F .x; N 1 N zNi /ŒIT.Ci ; zNi /1
DL Gi .x;
i D1
\
n
T .Hj ŒDj 1 ; x/
N D ;:
j D1
We define
˚ WD DL F .x;
N y/ŒintC
N 1
N zNi /ŒIT.Ci ; zNi /1
i WD DL Gi .x; .i 2 I /:
We shall prove the theorem by analyzing the three possibilities, namely:

(i) ˚ D ;;
(ii) i D ; for some i 2 I ;
(iii) ˚ ¤ ; and i ¤ ; for every i 2 I .
We begin with case (i). Since ˚ D ;; we can apply Theorem 12.2.15 for
T WD DL F .x;
N y/
N W X Y; M WD X; A WD C
such that there exists t 2 C C nf0g with
N y/.x/
t ı DL F .x; N
0 for every x 2 X:
Setting s D 0; ui D 0 for every i 2 I; and vj D 0 for every j 2 J; we get the

desired result.
For case (ii), let there exist i 2 I such that i D ;. Then again by invoking
Theorem 12.2.15 with
T WD DL Gi .x;
N zNi / W X Zi ; D WD X; A WD T .Ci ; Nzi /; .i 2 I /
we obtain ui 2 .T .Ci ; Nzi //C nf0g such that
.ui ı DGi .x;

N zNi //.x/
0 for every x 2 X:
By setting s D 0; vj D 0W for every j 2 J; and uj D 0Zj ; i ¤ j 2 I; we

obtain (12.75). For ui .Ci CNzi /
0; it suffices to notice that, in view of the convexity
of Ci ; we have T .Ci ; Nzi / Ci C zNi and hence ui .z C zNi /
0 for every z 2 Ci .
Finally, we consider the case (iii). Since .x; N y/
N is a local-minimizer of .P /; it
follows from Proposition 12.5.7 and the imposed conditions that we have
\ \
m \
n
L ˚ i Mj D ;:
i D1 j D1
Since L; ˚; i ; (i 2 I ) and Mj ; (j 2 J ) are all nonempty, we can apply

Theorem 12.5.4 to assure the existence of
` 2 LC
`0 2 .DL F .x;
N y/ŒintC
N 1 /C
N zNi /ŒIT.Ci ; zNi /1 /C ;
ì 2 .DL Gi .x; i 2I
`mCj 2 MjC j 2J
such that
` `0 `1 `2 `m `mC1 `mCn D 0: (12.76)
Now, in the case ˚ ¤ ;; in view of Theorem 12.2.15 for
T WD DL F .x;
N y/
N W X Y; M WD X; A WD C;
we get the existence of functionals t 2 C C such that for all x 2 X and `0 2
N y/ŒC
.DL F .x; N 1 / the following inequality holds
.t ı DL F .x;
N y//.x/
N
`0 .x/: (12.77)
Analogously, in the case ¤ ; with Theorem 12.2.15 for
N y/
T WD DL Gi .x; N W X Zi ; D WD X; A WD T .Ci ; Nzi /; .i 2 I /
we get the existence of functionals ui 2 .T .Ci ; Nzi //C such that for all x 2 X and
N zNi /ŒT .Ci ; zNi /1 / the following inequality holds
ì 2 .DL Gi .x;
.ui ı DL Gi .x;
N zNi //.x/
ì .x/ i 2 I: (12.78)
Combining the above inequalities (12.77) and (12.78) with (12.65) and setting
s D ` and vj D `mCj with j 2 J yield (12.75). This completes the proof. t
u
12.5.3 Second-Order Generalized Dubovitskii-Milyutin

Approach
We begin with the following necessary optimality condition for Problem (GP1 ):
N 2 graph .F / be a local weak minimizer of (GP1 ). Let
.Nu; vN / 2 X Y . Then:
\ \
m
N uN /
IT2 .S; x; D2L F .x; N uN ; vN /ŒIT.C; vN /1
N y; IT2 .Gi ŒCi 1 ; x;
N uN /
i D1
\ \
n
T2 . Hj ŒDj 1 ; x;
N uN / D ;: (12.79)
j D1
Proof. Since .x; N y/

N is a local weak-minimizer of (P ), there exists a neighborhood
U of xN such that F .S \ U / \ .fygN intC / D ;. We will show that if (12.79) fails
then this criteria for the weak-minimality will be violated.
For the sake of argument, we assume that there exists an x 2 X such that
\ \
m
x 2 IT2 .S; x;
N uN / D2L F .x; N uN ; vN /ŒI T .C; vN /1
N y; IT2 .Gi ŒCi 1 ; x;
N uN /
i D1
\ \
n
T2 . Hj ŒDj 1 ; x;
N uN /:
j D1
T
In view of the containment x 2 T2 . nj D1 Hi ŒDi 1 ; x; N uN / there are a sequence
.xn / X with xn ! x and a sequence .sn ; tn / P P with .s Tn ; tn / # .0; 0/ and
sn =tn ! 0 so that for every n 2 N; we have xN C sn uN C sn tn xn 2 nj D1 Hj ŒDj 1 ;
implying that Hj .xN C sn uN C sn tn xn / \ Dj ¤ ; for every j 2 J .
Furthermore, due to the containment x 2 IT2 .Gi ŒCi 1 ; x; N uN /; and the facts
that xn ! x; .sn ; tn / # .0; 0/ and sn =tn ! 0; we ensure the existence of
n1 2 N such that xN C sn uN C sn tn xn 2 Gi ŒCi 1 for every n
n1 and
for everyi 2 I; or equivalently Gi .xN C sn uN C sn tn xn / \ Ci ¤ ; for every
n
n1 and every i 2 I . Moreover, since x 2 D2L F .x; N uN ; vN /ŒI T .C; vN /1 ; there
N y;
exists y 2 I T .C; vN / such that .x; y/ 2 graph.DL F .x; 2
N y; N uN ; vN //. This, in view
of the definition of D2L F .x; N y;
N uN ; vN / and the sequences .sn ; tn /; ensure that there
are a sequence .yn / Y and an index n2 2 N such that yN C sn vN C sn tn yn 2
F .xN C sn uN C sn tn xn /. Since y 2 I T .intC; vN /; tn # 0 and yn ! y; it follows
from the definition of the interior tangent cones that there exists n3 2 N such that
for n
n3 ; we have vN C tn yn 2 intC . Using the fact that sn > 0; we obtain that
yN C sn vN C sn tn yn 2 yN intC .
Finally, from the containment x 2 IT2 .S; x; N uN / and the definition of sequences
.xn / and .sn ; tn /; we infer that there exists n4 2 N such that for every n > n4 ; we
have xN C sn uN C sn tn xn 2 U \ S .
Therefore, we have shown that for every n

maxfn1 ; n2 ; n3 ; n4 g there are un WD
N n uN Csn tn xn 2 S \U; satisfying the constraints such that F .un /\.yintC
xCs N / ¤ ;.
This, however, contradicts the weak optimality of .x;N y/.
N The proof is complete. u t
To obtain a variant of the above optimality condition that involves the derivative
of the maps Gi for i 2 I as well, we need the following:
Theorem 12.5.13. Let X and Z be normed spaces, let G W X ! Z be a set-valued
N zN/ 2 graph .G/ and let .Nu; w/
map, let .x; N 2 graph .DL G.x;
N zN//. Let A Z with
int.A/ ¤ ;. Then,
D2L G.x;
N zN; uN ; w/ŒIT
N 2
N 1 IT2 .GŒA1 ; x;
.A; zN; w/ N uN / (12.80)
Proof. Let x 2 D2L G.x; N zN; uN ; w/ŒIT

N 2
N 1 be arbitrary. Choose arbitrary
.A; zN; w/
sequences .xn / X with xn ! x; and .sn ; tn / P P with .sn ; tn / # .0; 0/ and
sn =tn ! 0. It suffices to show that there exists m 2 N such that xN C sn uN C sn tn xn 2
GŒA1 for every n
m. Since x 2 D2L G.x; N zN; uN ; w/ŒIT
N 2
N 1 ; there exists
.A; zN; w/
z 2 DL G.x;
2
N zN; uN ; w/.x/
N \ IT .A; zN; w/.
2
N Therefore, there are a sequence fzn g Z
and an integer n1 2 N such that zn ! z and zN C sn wN C sn tn zn 2 G.xN C sn uN C sn tn xn /
for every n
n1 . Because of the containment z 2 IT2 .A; zN; w/ N there exists
n2 2 N such that zN C sn wN C sn tn zn 2 A for every n
n2 . Therefore we have
zN C sn wN C sn tn zn 2 G.xN C sn uN C sn tn xn / \ .A/ for every n
m WD maxfn1 ; n2 g.
Consequently x 2 IT2 .GŒA1 ; x; N uN /. t
u
Combining the above two results, we obtain the following optimality condition:
of (P ). Let .Nu; vN / 2 graph .DL F .x; N y//.
N Let .x; N zNi / 2 graph .Gi /; let uN 2
\mi D1 dom .DL Gi . N
x; N
z i //; and let N
w i 2 D L Gi . N
x; N
z i /.N
u /. Then
\ \
m
N uN /
IT2 .S; x; D2L F .x; N uN ; vN /ŒIT.C; vN /1
N y; D2L Gi .x;
N zNi ; uN ; wN i /
i D1
\ \
n
ŒIT2 .Ci ; zN; wN i /1 T2 . Hj ŒDj 1 ; x;
N uN / D ;:
j D1
12.6 Set-Valued Optimization Problems with a Variable

Order Structure
As has already been discussed in Chap. 2, in recent years, optimality conditions

for vector optimization and set-valued optimization problems with a variable order
structure have attracted a great deal of attention. The primary objective of this
section is to give first-order and second-order necessary and sufficient optimality
12.6 Set-Valued Optimization Problems with a Variable Order Structure 569
conditions for a weak v-minimizer of set-valued optimization problems by using

the lower Dini derivative and the contingent derivative.
Let X and Y be normed spaces, ; ¤ S X , let C W X Y be a set-valued map
such that for each x 2 X , C.x/ is a pointed, closed, convex, and solid cone, and let
F W X Y be a set-valued map. We consider the following set-valued optimization
problem with a variable order structure
.SP/v v-minimize F .x/ subject to x 2 S;
where “v-minimize” stands for problems with variable ordering structure with
respect to a cone-valued mapping C W X Y . Here our interest is in the weak
N y/
v-minimizer, that is, we seek a pair .x; N 2 graph .F / such that
.F .S / y/ \ .intC.x//
N D ;:
We begin with the following result using the lower Dini derivatives:
Theorem 12.6.1. Let X and Y be normed spaces and let C W X Y be a set-
valued map such that for each x 2 X , C.x/ is a pointed, closed, convex, and solid
cone. For S X; let F W S Y be a set-valued map and let .x;N y/
N 2 graph .F / be
a local weak v-minimizer of .SP/v .
1. For every x 2 T .S; x/;
N we have
N y/.x/
DL F .x; N \ .int.C.x///
N D ;: (12.81)
T
N y//
2. Let uN 2 dom .DL F .x; N and vN 2 DL F .x;N y/.N
N u/ .@C.x//; N where @ denotes
the boundary, be arbitrary. Then, for every x 2 T 2 .S; x;
N uN /; we have
\
N y;
DL2 F .x; N uN ; vN /.x/ N vN / D ;:
IT.C.x/; (12.82)
Proof. Since (12.81) follows from (12.82), we will prove (12.82). Let x 2
N uN / be such that
T 2 .S; x;
\
N y;
DL2 F .x; N uN ; vN /.x/ N vN / ¤ ;:
IT.C.x/;
Then, there exists y 2 DL2 F .x; N y;N uN ; vN /.x/ such that y 2 IT.C.x/;
N vN /. Since
x 2 T 2 .S; x;N uN /; there are sequences ftn g P and fxn g X with tn # 0; xn ! x;
and xN C tn uN C tn2 xn 2 S . Therefore, it follows from the definition of the second-
order lower Dini derivative that there are a sequence fyn g Y and an index n 2 N
such that yN C tn vN C tn2 yn 2 F .xN C tn uN C tn2 xn / for every n
n1 . Since y 2
IT.C.x/; N vN / D IT.int.C.x//; N vN /; we have vN C tn yn 2 int.C.x// N and hence
yN C tn vN C tn2 yn 2 yN int.C.x//
N for sufficiently large n. Therefore, for such n 2 N,
un WD xN Ctn uN Ctn2 xn 2 S satisfies that F .un /\.yN int.C.x/// N ¤ ;. This, however,
contradicts that .x; N y/
N 2 graph .F / is a weak v-minimizer of .SP /v . The proof is
complete. t
u
Remark 12.6.2. If in the above result we use the first-order and the second-order
contingent derivatives, then (12.81) and (12.82) would only hold on the effective
domains of the derivatives.
In any case, the following analogue remains valid for the contingent derivative:
Theorem 12.6.3. Let X and Y be normed spaces and let C W X Y be a cone-
valued map such that int C.x/ ¤ ; for all x 2 X . For S X; let F W S Y be
N y/
a set-valued map and let .x; N 2 graph .F / be a weak v-minimizer of .SP/v . Then,
for every x 2 S; we have
N y/.x
DF.x; N x/
N \ .int.C.x///
N D ;: (12.83)
Proof. We show the result by contradiction. Let .x; N y/

N 2 S F .x/ N be a weak
v-minimizer of .SP/v . Suppose that there is an element x0 2 S such that
DF.x; N y/.x
N 0 x/ N 6 Y nfint.C.x//g.
N Therefore, there exists an element z such
that z 2 DF.x;N y/.x
N 0 x/N and z 2 int.C.x//.
N By the definition of the contingent
derivative, we get .x0 x; z/ 2 T .graph .F /; .x;
N y//.
N This means that there are
sequences f.xk ; yk /g graph.F / and f
k g P with .xk ; yk / ! .x; N y/
N and
k .xk x;N yk y/N ! .x0 x; N z/. We assert that there is a positive integer
such that for every k
we have yk 2 yN int C.x/.N Since yk 2 F .xk / F .S /,
N y/
it follows that .x; N cannot be a weak v-minimizer of .SP/v , a contradiction. t
u
For sufficient optimality conditions for a weak v-minimizer of .SP/v ; we need
certain convexity assumptions which are proposed in the following:
Definition 12.6.4. Let X and Y be normed spaces, let C W X Y be a cone-valued
map, and let F W X Y be a set-valued map. The map F is said to be C.x/-pseudo-
N
N y/
convex at .x; N 2 graph.F / if and only if for every x 2 dom .F /; we have
F .x/ yN DF.x;
N y/.x
N x/
N C C.x/:
N
Definition 12.6.5. Let X and Y be normed spaces, let C W X Y be a cone-valued

map and let F W X Y be a set-valued map. For a convex set S X , let xN 2 S and
let yN 2 F .S /. The map F is said to be C.x/-convex
N N y/
at .x; N 2 graph.F /; if for
any x; z 2 S and
2 .0; 1/,

F .x/ C .1
/F .z/ F .
x C .1
/z/ C C.x/:
N
N
Remark 12.6.6. The C.x/-pseudo-convexity reduces to the usual C -convexity
of F .
The next proposition is important for the proof of the sufficient optimality
condition in Theorem 12.6.8. We show the sufficient optimality condition under
the assumption that F is compactly approximable at .x; N y/
N 2 graph F (see
Definition 11.1.30).
12.6 Set-Valued Optimization Problems with a Variable Order Structure 571
Proposition 12.6.7. Let X and Y be normed spaces, let C W X Y be a cone-

valued map, and let F W X Y be a set-valued map with a convex domain. Let the
N
map F be compactly approximable and C.x/-convex N y/
at .x; N 2 graph .F /. Then,
N
F is C.x/-pseudo-convex N y/.
at .x; N
Proof. Consider an element .x; y/ 2 graph.F / and define a sequence f.xk ; yk /g in
X Y by
1 1
xk WD xN C .x x/;
N yk WD yN C .y y/:
N
k k
N
Since dom.F / is convex and since F is C.x/-convex N y/,
at .x; N it follows that for
all k 2 N ,

1 1
xk D 1 xN C x 2 dom F
k k
and

1 1 1 1
yk D 1 yN C y 2 F 1 xN C x C C.x/
N D F .xk / C C.x/:
N
k k k k
Thus, .xk ; yk / 2 graph.F C C.x//

N for every k 2 N and .xk ; yk / ! .x;
N y/
N for
k ! 1. Moreover, we have k.xk x; N yk y/
N ! .x x;N y y/.
N This implies
.x x;
N y y/N 2 T .graph.F C C.x//;
N .x;N y/.
N So we get, for every x 2 dom .F /
and for every y 2 F .x/ that y yN 2 D.F C C.x//.
N x; N y/.x
N x/.
N Since F is
N y/,
supposed to be compactly approximable at .x; N we get (see Theorem 11.3.10 or
[91, Proposition 2.72])
D.F C C.x//.
N x; N y/.x
N x/
N D DF.x;
N y/.x
N x/
N C C.x/:
N
Thus, for every x 2 dom .F /; we have y yN 2 DF.x;

N y/.x
N x/
N C C.x/.
N This
N
means that F is C.x/-pseudo-convex N y/.
at .x; N The proof is complete. t
u
Now, we will show the announced sufficient condition for a weak v-minimizer
of .SP /v under certain convexity assumptions (see [91, Theorem 2.73]).
Theorem 12.6.8. Let X and Y be normed spaces, let C W X Y be a cone-valued
function such that int.C.x// ¤ ; for all x 2 X , let S X be nonempty convex set,
let F W X Y be compactly approximable and C.x/-convex
N N y/
at .x; N 2 graph .F /,
N be proper, pointed, closed, and convex cone. If for every x 2 S; we
and let C.x/
have
N y/.x
DF.x; N x/
N \ . int.C.x///
N D ;;
N y/
then .x; N is a weak v-minimizer of .SP/v .
Proof. Because of our assumptions, we have
N y/.x
DF.x; N x/
N \ . int C.x//
N D ;:
N is pointed and convex, for every x 2 S; we have

Furthermore, because C.x/
N y/.x
.DF.x; N x/
N C C.x//
N \ . int C.x//
N D ;:
N
Since F is C.x/-pseudo-convex N y/
at .x; N and because of Proposition 12.6.7,
we get
.F .x/ y/
N \ . int C.x//
N D ;;
N y/
which means that .x; N is a weak v-minimizer of .SP/v . t
u
12.7 Optimality Conditions for Q-Minimizers in Set-Valued

Optimization
We have noted that the optimality conditions for the proper minimality and the
weak minimality are quite similar and largely depend on using the interior of the
cones involved. In fact, as seen in Chap. 2, the notion of Q minimizers also builds
on exploiting the fact that numerous other notions of minimality in set-valued
optimization can be stated by using an open cone Q. The objective of this section
is to give some optimality conditions for the abstract notion of a Q-minimizer and
then derive optimality conditions for many other optimality notions. This section is
based on [223].
12.7.1 Optimality Conditions for Q-Minimizers Using Radial

Derivatives
In the following, we consider the following unconstrained set-valued optimization

problem:
minimize F .x/ subject to x 2 X; (P )
where X and Y are normed or Banach spaces and F W X Y is a set-valued map.

In the following result, we use the radial derivative (see Definition 11.1.13).
Proposition 12.7.1. Let X and Y be normed spaces, let Q Y be an open cone,
let F W X Y be a set-valued map, and let .x;
N y/
N 2 graph .F /. Then .x;
N y/
N is a
Q-minimizer of (P ) if and only if
12.7 Optimality Conditions for Q-Minimizers in Set-Valued Optimization 573
N y/.x/
DR F .x; N \ .Q/ D ; for every x 2 X: (12.84)
N y/
Proof. We recall that .x; N is a Q-minimizer of (SP ) if and only if
.F .X / y/
N \ .Q/ D ;: (12.85)
Assume that .x;N y/

N is a Q-minimizer of (P ) but (12.84) does not hold. Then,
there exists x 2 X such that y 2 Q and y 2 DR F .x; N y/.x/.
N By the definition
of the radial derivative, there exist sequences ftn g P and f.xn ; yn /g such that
xn ! x and yn ! y and yN C tn yn 2 F .xN C tn xn / for every n 2 N. Since
yn ! y 2 Q and since Q is open, we have tn yn 2 Q; for sufficiently large n.
Therefore, tn yn 2 .F .xN C tn xn / y/
N \ .Q/ for sufficiently large n. This, however,
is a contradiction to (12.85).
For the converse, assume that (12.84) holds. Then, for any x 2 X , we have
.F .x/ y/
N \ .Q/ DR F .x;
N y/.x
N x/
N \ .Q/ D ;:
N y/
Therefore, (12.85) holds, confirming that .x; N is a Q-minimizer of (P ). The proof
is complete. t
u
The following result shows the usefulness of the abstract notion of Q-minimality:
Theorem 12.7.2. Let X and Y be Banach spaces, let Q Y be an open cone, let
F W X Y be a set-valued map, and let .x;
N y/
N 2 graph .F /.
1. The element .x;N y/
N is a weak minimizer of (P ) if and only if (12.84) holds with
Q D intC .
N is a S -proper minimizer of (P ) if and only if (12.84) holds
with Q D fy 2 Y j '.y/ < 0g for some ' 2 C # .
N is a Hartley proper minimizer of (P ) if and only if (12.84)
hold with Q D C."/ for some scalar " > 0.
N is a Borwein proper minimizer of (P ) if and only if (12.84)
holds with Q being some dilation of C .
5. The element .x; N y/
N is a Henig global proper minimizer of (P ) if and only
if (12.84) holds with Q D int.K/ where K is a proper, pointed, convex, and
solid cone in Y such that C f0g intK.
N is a Henig proper minimizer of (P ) if and only if (12.84)
holds with Q D V for some scalar satisfying 0 < < ı.
7. Assuming that C has a bounded base, the element .x; N y/
N is a super proper
minimizer of (P ) if and only if (12.84) holds with Q D V for some scalar
satisfying 0 < < ı, provided that C has a compact base.
Proof. The proof is a direct consequence of the fact that a certain notion of mini-
mality satisfies (12.85) for the associated value of the cone Q (see Definition 2.6.3
and Theorem 2.4.11). t
u
Remark 12.7.3. Note that in Theorem 12.7.2, we only included cases where either
Q is expressed directly in terms of C; or Q is a dilation of C .
12.7.2 Optimality Conditions for Q-Minimizers Using

Coderivatives
In the following, our objective is to give optimality conditions for the notion of
Q-minimality by using the Ioffe approximate coderivative (introduced in Defi-
nition 3.6.2 and the Clarke coderivative (see Definition 3.4.5). These optimality
conditions are given by using a scalarization by means of the functional Q
introduced and discussed in Sect. 5.3.
Proposition 12.7.4. Let X and Y be Banach spaces, let Q Y be an open cone,
let F W X Y be a set-valued map with a closed graph, and let .x;
N y/
N 2 graph .F /.
N is a Q-minimizer of (P ) then there exists y 2 @A Q .0/ such that the
N y/
If .x;
following inclusions hold:
0 2 DA F .x; N /;
N y/.y (12.86)
02 DC F .x; N /:
N y/.y (12.87)
N y/
Proof. By using Proposition 5.3.3, we note that .x; N is a solution of (P ) if an only
N y/
if .x; N is a solution of the following problem
Minimize Q .y y/
N C ..x; y/I graph .F //:
Therefore, by Proposition 3.6.3, we have
.0; 0/ 2 @A .Q .0/ C graph .F / .x;

N y//
N
f0g @A Q .0/ C @A graph .F / .x;
N y/
N
D f0g @A Q .0/ C NA ..x;
N y/I
N graph .F //;
and hence there exists y 2 @A Q .0/ with .0; y / 2 NA ..x;N y/I
N graph .F //;
which further implies that 0 2 DA F .x; N /; showing that (12.86) holds.
N y/.y
Finally, (12.87) is a direct consequence of (12.86) and Proposition 3.6.3 (iv). The
u
The following is an analogue of the above result for other notions of minimality.
Theorem 12.7.5. Let X and Y be Banach spaces, let Q Y be an open cone, let
F W X Y be a set-valued map with a closed graph, and let .x;
N y/
N 2 graph .F /.
N y/
1. If .x; N is a weak minimizer of (P ) or .x; N y/
N is a Hurwicz proper minimizer
of (P ) and the image F .X / has a nonempty interior, then there exists y 2
C C n f0g such that (12.86)–(12.87) hold.
N y/
2. If .x; N is a global Henig proper minimizer, (in particular, if .x;N y/
N is either a
positively proper minimizer, or a Henig proper minimizer, or a super minimizer)
of (P ), then there exists y 2 C # such that (12.86)–(12.87) hold.
N y/
3. If .x; N is a super minimizer of (P ) and C has a bounded base, then there exists
y 2 int.C C / such that (12.86)–(12.87) hold.
Proof. In view of Definition 2.6.3 and Theorem 2.4.11, the element .x; N y/
N is a
minimizer of (P ) in some sense if an only if it is a Q-minimizer of (P ) with Q
being some open set which is chosen in correspondence with Theorem 2.4.11. The
proof then follows from Propositions 5.3.4 and 12.7.4. t
u
In the following, we establish the Lagrange multiplier rule for the constrained
set-valued optimization problem (SP ). We consider a set-valued optimization
problem
minimize F .x/ subject to x 2 S1 ; (SP )
where X , Y and Z are Banach spaces and F W X Y and G W X Z are

set-valued maps, S X , D Z and the feasible set S1 of (SP ) is given by
S1 WD fx 2 S j G.x/ \ D ¤ ;g: (12.88)
Remark 12.7.6. Note that the only difference between (SP ) and .P1 / is that in the
former we have more general constraint set. To emphasize this difference, we have
preferred renaming this set-valued optimization problem.
In order to prove the following assertion we assume that the map F is Lipschitz-
like around .x;N y/
N 2 graph .F / (see Definition 3.3.9) and the map G is metrically
regular around .x;N zN/ 2 graph .G/ relative to S D, (see Remark 12.8.10 and [2])).
That is, there exist scalars r > 0 and > 0 such that, for all .x; z/ 2 Œ.xN C rBX /
.Nz C rBZ / \ .S D/;
d..x; z/; S D/ \ graph G/ d.z; G.x//:
We give the following Lagrange multiplier rule for the Q-minimizers of (SP ).
Proposition 12.7.7. Let X , Y; and Z be Banach spaces, let F W X Y and
G W X Z be set-valued maps, and let S X and D Z be closed. Let
F and G be closed and Lipschitz-like around .x; N y/
N 2 graph .F / and .x; N zN/ 2
graph .G/ respectively, where zN 2 G.x/\D,
N and let G be metrically regular around
N zN/ relative to S D. If .x;
.x; N is a Q-minimizer of (SP), then there exists y 2
N y/
@A Q .0/ such that
0 2 DA F .x; N / C DA G.x;

N y/.y N zN/.z / C NA .xI
N S/ (12.89)
and
0 2 DC F .x; N / C DC G.x;

N y/.y N zN/.z / C NC .xI
N S /; (12.90)
where z 2 NA .NzI D/ and z 2 NC .NzI D/, respectively.

N y/
Proof. The proof is similar to that of [2, Theorem 3.7]. By Definition 2.6.3, .x; N
is a Q-minimizer of (SP ) if an only if
.F .S1 / y/
N \ .Q/ D ;:
N y;
Proposition 5.3.3 yields that .x; N zN/ is a minimal solution of the scalar problem
minimize q.x; y; z/ subject to .x; y/ 2 graph F and .x; z/ 2 .S D/ \ graph G;
where q.x; y; z/ WD Q .y y/. N Then by Clarke’s penalization ( see [100,

Proposition 2.4.3]), the metric regularity assumption and [2, Proposition 2.6 ], for
N y;
some integer l > 0 large enough, .x; N zN/ is an unconstrained local minimizer of
.x; y; z/ ! q.x; y; z/Cld..x; y/; graph F /Cld..x; z/; graph G/Cld..x; z/; S D/:
Taking into account the assertions (ii) and (iv) of Proposition 3.6.3, zero is in the
sum of the subdifferentials. That is, there exist
y1 2 @A q.x;
N y;
N zN/ D @A Q .0/ @C Q .0/;
.x2 ; y2 / 2 l@A d..x;
N y/;
N graph F / NA ..x;
N y/I
N graph F /;
.x3 ; z3 / 2 l@A d..x;
N zN/; graph G/ NA ..x;
N zN/I graph G/;
.x4 ; z4 / 2 l@A d..x;
N zN/; S D/ NA ..x;
N zN/I S D/;
such that
0 D x2 C x3 C x4 ; 0 D y1 C y2 ; and 0 D z3 C z4 :
Putting y D y1 D y2 and z D z4 D z3 , we obtain
0 2 DA F .x; N / C DA G.x;

N y/.y N zN/.z / C NA .xI
N S/
and (12.89) holds. Finally, (12.90) follows from (12.89) and Proposition 3.6.3 (iv).
t
u
The following Lagrange multiplier rule for several types of minimizers of (SP )
are derived in [228, Theorem 21.20].
Theorem 12.7.8. Let X , Y; and Z be Banach spaces, let F W X Y and G W

X Z be set-valued maps, and let S X and D Z be closed. Let F and
N y/
G be closed and Lipschitz-like around .x; N 2 graph .F / and .x;
N zN/ 2 graph .G/
respectively, where zN 2 G.x/
N \ D, and let G be metrically regular around .x;
N zN/
relative to S D.
N y/
1. If .x; N is a weak minimizer of (SP ) or .x;N y/
N is a Hurwicz proper minimizer
of (SP ) and the image F .S1 / has a nonempty interior, then there exists y 2
C C n f0g such that (12.89)–(12.90) hold.
N y/
2. If .x; N is a global Henig proper minimizer (in particular, a positive proper
minimizer, or a Henig proper minimizer) of (SP ), then there exists y 2 C #
such that (12.89)–(12.90) hold.
N y/
3. If .x; N is a super minimizer of (SP ) and C has a bounded base, then there
exists y 2 intC C such that (12.89)–(12.90) hold.
Proof. The proof can devised in the same way as Theorem 12.7.5 (but applying
Proposition 12.7.7 instead of Proposition 12.7.4) and hence is omitted. t
u
Remark 12.7.9. The above version of the Lagrange multiplier rule for weak
minimizers were derived in [2], for strong minimizers or positive proper minimizers
in [228], and are given in [228] for the case of Hurwicz proper minimizer, global
Henig proper minimizer, Henig proper minimizer and super minimizers. We note
that for weak minimizers to (SP ), there have also been obtained Lagrange-Kuhn-
Tucker multipliers in terms of the subdifferential of set-valued maps introduced by
Sawaragi and Tanino [568, 569]. Recently, the Lagrange multiplier rule for super
minimizers to a constrained set-valued optimization problem has been obtained by
Bao and Mordukhovich in Asplund spaces [26] and by Huang in Banach spaces
[269]. Bao and Mordukhovich used the extremal principle of variational analysis in
their proof while Huang’s technique is based on exploiting a property of the normal
cone due to Clarke. In contrast to Ha’s work, they considered super minimizers of
the set-valued optimization problem with the geometric constraint x 2 S1 .
Using a characterization of a Benson proper efficient point in a separable or
reflexive Banach space obtained in [482], it is possible to derive the Lagrange claim,
the Fermat rule and the Lagrange multiplier rule for Benson proper minimizers from
Theorems 12.7.2, 12.7.5 and 12.7.8.
Theorem 12.7.10. Assume that Y is a separable Banach space, or that Y is a
reflexive Banach space and C has a base.Consider the problem (SP ) where the
feasible set is given by (12.88).
1. Let xN 2 X and .x;
N y/
N 2 graph F . Assume that F has a closed graph and F yN
is nearly C -subconvexlike on X; that is, cl cone.F .X / yN C C / is convex. Then
N is a Benson proper minimizer of (SP ) if and only if there exists y 2 C #
N y/
.x;
such that (12.84) hold with Q D fy W '.y/ < 0g and only if there exists y 2 C #
such that (12.86)–(12.87) hold.
2. Let xN 2 S1 and .x;N y/

N 2 graph F . let S X and D Z be closed. Let F and G
be closed and Lipschitz-like around .x; N y/
N 2 graph .F / and .x;N zN/ 2 graph .G/
respectively, where zN 2 G.x/ N \ D, and let G be metrically regular around .x;N zN/
relative to S D. Assume in addition that F yN is nearly C -subconvexlike on
S1 ; that is, cl cone.F .S1 / yN C C / is convex. If .x;
N y/
N is a Benson minimizer
of (SP ), then there exists y 2 C # such that (12.89)–(12.90) hold.
Proof. It has been established (see [482, Corollaries 4.2 and 4.5 ]) that if Y is
a separable Banach space, or Y is a reflexive Banach space and C has a base
and if F yN is nearly C -subconvexlike on X (on S1 , respectively), then .x; N y/
N
being a Benson proper minimizer of the unconstrained problem (SP ) (of the
constraint problem (SP ) with the feasible set S1 , respectively) is also a positive
proper minimizer. The assertion follows from this fact and Theorems 12.7.2, 12.7.5
and 12.7.8. t
u
Remark 12.7.11. Recently, Benson proper minimizers to set-valued optimization
problems have received more attention, and several optimality conditions for them
were obtained under assumptions on generalized convexity of set-valued data such
as convexlikeness, subconvexlikeness, near convexlikeness and near subconvexlike-
ness (see [482] for references). Near subconvexlikeness, first presented in [602]
and also used in [513] is the weakest convexity among the above four kinds of
generalized convexity. Under the assumptions on the near subconvexlikeness of
the objective and constraints, it has been established in [513] that a Benson proper
minimizer of (SP ) can be expressed in terms of saddle points defined in a suitable
sense.
12.8 Lagrange Multiplier Rules Based on Limiting

Subdifferential
The aim of this section is to show necessary optimality conditions using the limiting
subdifferential by Mordukhovich ([342, 426, 430], see Sect. 3.5) for the set-valued
optimization problem (SP ) introduced in Sect. 2.6. The corresponding calculus
rules are working in Asplund spaces such that we assume in this section that X and
Y are Asplund spaces. We will use the solution concept based in vector approach
(see Sect. 2.6.1) introduced in Definitions 2.6.1. This means that we will derive our
results for minimizers of the following set-valued optimization problem (SP):
minimize F .x/ subject to x 2 S; (12.91)
where the cost mapping F W X Y is a set-valued mapping, Y and X are Asplund

spaces, C Y is a proper pointed convex closed cone, S is a subset of X , and we
are looking for minimizers of (SP ) in the sense of Definition 2.6.1.
12.8 Lagrange Multiplier Rules Based on Limiting Subdifferential 579
In many papers Lagrange multiplier rules are shown for weak minimizers of
the set-valued optimization problem (SP ) (see Definition 2.6.2 and Sect. 12.7.2).
The most important characteristic of the weakly minimal points of a subset A of Y
with respect to C (cf. Definition 2.4.2 and the corresponding solution concept for
the set-valued problem (SP ) given in Definition 2.6.3) is the possibility to study
its optimality conditions via numerous powerful scalarization techniques under the
additional condition
int C ¤ ;: (12.92)
However, the condition int C ¤ ; is a very strong restriction in infinite dimensional

spaces since “the class of ordered topological vector spaces possessing cones with
nonempty interiors is not very broad” ([475, page 183]); in particular, the natural
ordering cones in the Lebesgue spaces l p and Lp for 1 p < C1 have empty
interior.
So the aim of this section is to present optimality conditions without assum-
ing (12.92) derived by Bao and Tammer [29].
Obviously, every Pareto minimal point of a vector optimization problem is
weakly minimal provided that the nonempty interior condition (12.92) holds;
otherwise, the weak concept is no longer defined. In the latter case, the Henig global
proper minimality (see Definition 2.4.4 (f)) takes the center and plays the role of the
weak optimality since it is defined via the Pareto optimality with respect to a bigger
convex cone with a nonempty interior (compare Definition 2.4.4 (f)).
Recently, a few works have challenged the lack of the nonempty interior
condition (12.92) imposed on ordering cones. In [151] some necessary conditions
were established for the so-called approximate Pareto minimizers via a suitable
scalarization scheme. In [149] some Lagrange multiplier rules for Pareto minimal
solutions were obtained by using the separation theorem for nonconvex sets
(see Lemma 12.8.1) provided that the ordering cone enjoys the asymptotical
compactness (AC) condition, i.e., the intersection of it and the closed unit ball
is compact. This requirement is strict in infinite-dimensional settings. Moreover,
the necessary results therein are in fact applied to properly minimal points since
Min.A; C / D GHe-PMin.A; C / under the AC condition and other standard
assumptions and since the first set is bigger than the second in general. In difference
to this scalarization approach, the variational approach is mainly based on the
extremal principle (Theorem 5.5.9), a variational counterpart of local separation
for nonconvex sets. It is initiated by Kruger and Mordukhovich in [342, 426], and
recently applied to multiobjective optimization in [24, 25, 28, 430]. The advantage
of the latter approach is the possibility of deriving necessary conditions to Pareto
optimal points of a multiobjective optimization problem directly without converting
it to a scalar problem in both finite- and infinite-dimensional settings. However, to
implement it, the sequential normal compactness (SNC) condition (see Sect. 5.5) of
ordering cones is unavoidable due to the natural lack of the compactness property
in infinite dimensional spaces which is automatic in finite dimensions. It is known

from [431, Theorem 1.21] that a convex cone fulfills the SNC condition if it has a
nonempty relative interior r-int C ¤ ; and cl .C C / has a finite-codimensions.
In summary, both approaches can not be applied to the positive ordering cones of
the Lebesgue spaces l p and Lp for 1 < p < C1 since they are neither AC nor the
SNC at the origin.
The main aim of this section is to establish subdifferential necessary optimality
conditions for Pareto minimal points of sets and minimizers of constrained set-
valued optimization problems where the ordering cone of a infinite-dimensional
Banach space does not necessarily have an empty interior (even relative interior),
or a finite-dimensional span following the paper by Bao and Tammer [29]. Our
approach combines the scalarization scheme in [149, 151] and several advanced
tools of variational analysis used in [24, 25, 28]. Precisely, we proceed as follows:
Given a Pareto minimal point y to a set A Y with respect to an ordering cone C
with int C D ;. In contrast to [149] in which the existence of a bigger cone D Y
satisfying the condition C n f0g int D is required (so that, y is proper efficient,)
in this section we will choose an appropriate nonempty interior cone that might not
contain in the given cone or be contained in it. Such a cone automatically enjoys the
SNC condition, but it might not have the AC property. Then utilizing the strategy
in [24, 28] which deals with epigraphical multifunctions, we derive subdifferential
necessary conditions thanks to the full calculus for generalized differentiation in
[430, Chapter 3] (see Sect. 3.5).
In order to derive multiplier rules we use the fundamentals from the book by
Mordukhovich [431] (see Sect. 3.5), especially subdifferential constructions for
set-valued mappings with values in partially ordered spaces first introduced in
[25]. We derive necessary optimality conditions for Pareto minimal points of sets,
and minimizers of set-valued optimization problems without the nonempty interior
requirement imposed on ordering cones.
We use the standard notation of variational analysis presented in Sect. 3.5, espe-
cially the limiting normal cone (see Definition 3.5.6) and the limiting subdifferential
(see Definition 3.5.11); cf. the books [431,499], and assume that all the spaces under
consideration are Asplund (see Definition 3.5.3) unless otherwise stated.
In the proofs of our assertions we apply the results concerning the scalarizing
functional 'A WD 'A;k 0 W Y ! R given in (5.5) with A Y and k 0 .¤ 0/ 2 Y .
The most important properties (from Theorem 5.2.3, Proposition 5.2.23 and Corol-
lary 5.2.24) for deriving multiplier rules are collected in the following lemma:
Lemma 12.8.1 (Scalarization Functions for Nonconvex Sets). Let D Y be
a proper, closed and convex cone with a nonempty interior in a Banach space Y .
Given k 0 2 int D, the functional 'D;k 0 W Y ! R defined by
'D;k 0 .y/ D inff

2 R j
k 0 2 y C Dg (12.93)
is continuous, sublinear, strictly-int D-monotone. Moreover, the following relations

hold:
(i) 'D;k 0 .y C tk 0 / D 'D;k 0 .y/ C t 8y 2 Y; 8t 2 R.

(ii) @'D;k 0 .0/ D fy 2 D j hy ; k 0 i D 1g with D WD fy 2 Y j hy ; yi
0; 8y 2 Dg D N.0I D/.
(iii) @'D;k 0 .y/ D fy 2 D j hy ; k 0 i D 1; hy ; yi D 'D;k 0 .y/g for any y 2 Y .
(iv) Given a nonempty set A Y , if y 2 A is a weakly minimal point of A with
respect to D (see Definition 2.4.2), then one has
'D;k 0 .y y/
0 for all y 2 A:
This real-valued function plays the role of an utility function. In this section
we present necessary conditions for Pareto minimal points of sets, for Pareto
minimizers of vector-valued optimization problems and for Pareto solutions of set-
valued optimization problems in the Asplund setting due to the full calculus for
(Mordukhovich) generalized differentiation. It is worth to stressing that our results
can be extended to the general Banach setting by using the Ioffe approximate
differentiation by imposing stronger assumptions on the given data such that the
corresponding calculus rules are applied (compare Sect. 12.7.2).
Let Y be an Asplund space, and let C Y be a proper, convex and pointed cone
which generates a partial order “C ” defined by
y1 C y2 if and only if y2 y1 2 C for all y1 ; y2 2 Y: (12.94)
Given a nonempty subset A in Y and y 2 Min.A; C / (see Definition 2.4.1). For

each k 0 2 C n f0g and for each " 2 .0; kk 0 k/ we consider the following cone
˚
Ck 0 ;" WD cone B.k 0 ; "/ D t z j t
0; z 2 B.k 0 ; "/ ; (12.95)
where B.k 0 ; "/ is a closed ball with the center k 0 and the radius ". Obviously, it
is a proper, closed, convex and pointed cone and its interior is nonempty since
k 0 2 int Ck 0 ;" . This cone will take the center in our procedure in formulating
necessary conditions for Pareto points with respect to an empty interior ordering
cone (Fig. 12.1).
It is worth emphasizing that the approach in [149] involved the Henig global
proper minimality (Definition 2.4.4 (f)). Recall that Henig minimality is defined in
the same way of proper minimality by specifying the proper cone D Y with C n
f0g int D (see Theorem 2.4.11) in form of a dilation of C (see Definition 2.4.9)
given by
[
CH;" WD Ck 0 ;" ;
k 0 2B
where B is a base of the cone C with B is a convex set such that 0 62 cl B and
C D cone B, i.e., C is well-based (see Definition 2.2.6). In our approach, we just
use one component of the Henig cone.
Fig. 12.1 The cone C with int C D ; and the corresponding cone Ck0 ;" with int Ck0 ;" ¤ ;
Theorem 12.8.2 (Necessary Conditions for Pareto Minimal Points of Sets). Let
y 2 Min.A; C /. Assume that the epigraphical set of A with respect to C and A C C
are locally closed at y. Then for every k 0 2 C n f0g satisfying
k 0 62 cl cone .A C C y/; (12.96)
there are a positive number " > 0 and y 2 Y such that
y 2 N.yI A C C /; y 2 N.0I C / \ N.0I Ck 0;" / and hy ; k 0 i D 1;

(12.97)
where Ck 0 ;" is given in (12.95).

Proof. First employing [293, Lemma 4.7 (a)], we have the implication
y 2 Min.A; C / D
) y 2 Min.A C C; C / (12.98)
for any proper, convex and pointed ordering cone C . Taking into account Defini-
tion 2.4.1, we get from (2.19) that
.A C C y/ \ .C / D f0g; and thus cone .A C C y/ \ .C / D f0g:

(12.99)
Since k 0 62 cl cone .ACC y/ by (12.96), there is a positive number " 2 .0; kk 0 k/
such that

cl cone .A C C y/ \ B.k 0 ; "/ D ;; and thus cl cone .A C C y/ \ .Ck 0 ;" / D f0g;
which justifies y 2 Min.A C C; Ck 0 ;" /, where Ck 0 ;" is given in (12.95).

Now we can apply Lemma 12.8.1 for D WD Ck 0 ;" and the element k 0 2 int Ck 0 ;"
to get that y is a minimum of the scalarization function 'D;k 0 defined by (12.93)
over A C C :
minimize 'D;k 0 .y y/ subject to y 2 A C C:
Since the functional 'D;k 0 ./ is Lipschitz continuous (in fact, sublinear and continu-
ous) by Lemma 12.8.1, and since the constraint set A C C is assumed to be locally
closed at y, all the assumptions of the lower-subdifferential necessary condition for
local minima in [430, Proposition 5.3] are satisfied. Employing it to the minimum y
of the above problem, we have
0 2 @'D;k 0 . y/.y/ C N.yI A C C / D @'D;k 0 .0/ C N.yI A C C /: (12.100)
Taking into account the description of the subdifferential of 'D;k 0 at the origin in
Lemma 12.8.1 (ii)
@'D;k 0 .0/ D fy 2 CkC0 ;" j hy ; k 0 i D 1g with C

Ce;" D N.0I Ck 0 ;" /;
we find from the inclusion (12.100) a subgradient y 2 @'D;k 0 .0/ with y 2

N.yI A C C / and y 2 N.0I Ck 0;" /, which justifies the necessary conditions
in (12.97) provided that y 2 N.0I C /. To show this, we further elaborate
the definition of the normal cone (3.25) to the set A C C at y. We get from
y 2 N.yI A C C / a sequence .yk ; yk / 2 Y Y satisfying
ACC w
yk ! y; yk ! y with yk 2 NO .yk I A C C /:
The last inclusion gives
hyk ; yi hyk ; y yk i
lim sup D lim sup
C kyk C ky yk k
y !0 yyk !0
hyk ; y yk i
D lim sup
C Cyk ky yk k
y ! yk
hyk ; y yk i
lim sup 0;
ACC ky yk k
y ! yk
which verifies yk 2 NO .0I C /, and thus y 2 N.0I C / as k ! C1. The proof
is completed. t
u
Taking into account the definition of the normal cones N.0I C /, N.0I Ck 0;" / and
condition (12.97) in Theorem 12.8.2 we get the following corollary:
Corollary 12.8.3 (Necessary Conditions for Pareto Minimal Points of Sets). Let
y 2 Min.A; C /. Assume that the epigraphical set of A with respect to C and A C C
are locally closed at y. Then for every k 0 2 C n f0g satisfying (12.96) there are a
positive number " > 0 and y 2 C C such that
y 2 N.yI A C C /; hy ; k 0 i
"jjy jj and hy ; k 0 i D 1:
(12.101)
Proof. Because of y 2 N.0I C / \ N.0I Ck 0;" / (with Ck 0 ;" is given by (12.95))

in (12.97) in Theorem 12.8.2 and the properties of the included normal cones we get
y 2 N.0I C / D C C (cf. Lemma 12.8.1 (ii)). Furthermore, y 2 N.0I Ck 0;" /
implies y 2 CkC0 ;" and so we have
8
> 0; 8y 2 Y with jjyjj D 1 W hy ;
.k 0 C "y/i
0:
This yields for y ¤ 0 and for all y 2 Y with jjyjj D 1
y y
h ; k 0
i
"h ; yi;
jjy jj jjy jj
and so
y y
h ; k 0
i
sup "h ; yi D ";
jjy jj jjyjjD1 jjy jj
i.e., hy ; k 0 i
"jjy jj . t
u
Remark 12.8.4. Note that the condition (12.96) is fulfilled if the cone cone .A C
C y/ happens to be closed. Furthermore, (12.96) is satisfied for any k 0 2 int C
provided that int C ¤ ;.
Note that the condition (12.96) is essential since we do not assume that the
ordering cone has a nonempty interior. Analyzing the proof of Theorem 12.8.2, the
existence of a vector k 0 satisfying (12.96) is ensured provided that
C \ bd cone .A C C y/ ¤ ;:
Remark 12.8.5. It is worth emphasizing that if the cone cone .A C C y/ is closed

at the origin, then every Pareto minimal point y of A with respect to C is a Benson
minimal point (see Definition 2.4.4 (d)) in the sense that
cl cone .A C C y/ \ .C / D f0g: (12.102)
In the book chapter by Ha [229] necessary conditions for Benson proper

minimality are (indirectly) established under the condition that the cone D WD
Y n cl cone .A C C y/ has a nonempty interior. Ha [229] used the concept
of Benson proper minimality to obtain sufficient optimality conditions.
Remark 12.8.6. Theorem 12.8.2 can be applied to Pareto minimality and not
Benson proper minimality. Take A be a subset of R3 defined by
A WD f.x; y; z/ j x 2 C y 2 C .z 1/2 D 1g;
where R3 is partially ordered by the non-solid cone C WD f0g R2C , i.e., int C D ;.
Obviously, the origin is a Pareto point of A with respect to C , but it is not a Benson
properly minimal point since cl cone .A C C 0/ D f.x; y; z/ j z
0g is a closed
half-space of R3 and f.0; 0/g R D cl cone .A C C 0/ \ .C /. Moreover,
every vector k 0 D .0; y; z/ with z > 0 satisfies the condition (12.96). Therefore, the
necessary conditions in Theorem 12.8.2 can be applied to this example.
The necessary conditions in Theorem 12.8.2 do not hold true p without the
condition (12.96). Consider a closed set in R2 given by A WD epi . x/, where
R2 is partially ordered by the cone C D f0g RC . Obviously, the origin is a Pareto
minimal point but not a Benson properly minimal point. It is easy to check that
N.0I A C C / D N.0I A/ D R 0:
There is only the unique choice of y D .1; 0/ 2 N.0I A C C / and k 0 D .0; 1/ 2

C . We can easily check that y 62 N.0I Ck 0;" / for every " > 0, and thus the
necessary conditions in Theorem 12.8.2 is not applicable to this example.
Next, we will establish a revised version of Theorem 12.8.2
Theorem 12.8.7 (Refined Necessary Conditions for Pareto Minimal Points of
Sets). Let y 2 Min.A; C /. Assume that A C C is locally closed at y and cone .A C
C y/ is closed. Then for every k 0 2 C n f0g, there are a positive number " > 0
and y 2 Y such that
y 2 N.yI A C C /; y 2 N.0I C / \ N.0I Ck 0;" / and hy ; k 0 i D 1;

(12.103)
where Ck 0 ;" is given in (12.95).

Proof. The assertion follows by the proof of Theorem 12.8.2. Since the cone
cone .A C C y/ is closed, we get from (12.99) that for every k 0 2 C n f0g,
the condition (12.96) is fulfilled, and thus the necessary condition is proved. t
u
Note that the third condition in (12.103) implies the nontriviality condition
y ¤ 0 in the Lagrange-type necessary conditions while the second provides the
range of multipliers. Note also that the necessary conditions in Theorem 12.8.7 are
new when the ordering cone is not SNC at the origin; otherwise, they are weaker
than the existing necessary results due to the additional closedness assumption
imposed on cone .A C C y/.
Remark 12.8.8 (Alternative Necessary Conditions).
(1) It is important to note that the closedness of the cone cone .A C C y/ and the
closedness of the set itself are different conditions. Indeed, in R2 equipped with
the usual Pareto order C D R2C , the set A WD graph.x 2 / has its epigraphical set
.A C C 0/ which is closed, but its cone is not closed at the origin, while the
set A WD f.x; y/ j x > .y 1/2 1 and y
xg [ f0g has the epigraphical
set which is not closed, but its cone is closed at the origin.
(2) Observe that the closedness assumption of the set .ACC y/ in Theorem 12.8.7
can be dropped. In such a case, i.e., A C C is not locally closed at y, the first
condition in (12.103) will be formulated by

y 2 N yI cl .A C C / :
Observe also that the closedness assumption of the cone cone .A C C y/ is

essential. Indeed, if A is the graph of the cardioid r.˛/ WD 1 C sin.˛/ in R2 ,
C WD f.0; t/ 2 R2 j t
0g, and y D 0, then 0 is a Pareto minimal point to
A. Since cone .A C C y/ D R2 n .C /, and since the only vector k 0 2 C
with kk 0 k D 1 is k 0 D .0; 1/, there is no " > 0 such that 0 2 Min.A; Ck 0 ;" /.
Therefore, the arguments in the proof are no longer valid.
(3) The first condition in (12.103) formulated with the epigraphical set .A C C / is
much better than the conventional condition with the set itself
y 2 N.yI A/
provided that A C C is order continuous at y since we have the following

inclusion
N.yI A C C / N.yI A/: (12.104)
To justify this, let us first recall that A is order continuous at y if for any
sequence fyk g A C C converging to y, there is a sequence fvk g A with
vk C yk and vk ! y; cf. [26, Definition 4.1] for the order continuity of set-
valued mappings. In fact, a set A is order continuous if the constant mapping
F W R Y with F .x/ A is order continuous at .0; y/. The reader is referred
to [26] for several efficient conditions ensuring this property.
To justify inclusion (12.104), we mimic the proof of [26, Proposition 4.3].

Take y 2 N.yI A C C /, then find a sequence .yk ; yk / 2 Y Y such that
ACC w
yk ! y yk ! y with yk 2 NO .yk I A C C /:
Fix k and write yk in the form yk D vk C tk with vk 2 A and tk 2 C . For every

> 0 we get from the definition of Fréchet normal xk 2 NO .yk I A C C / that
hyk ; y yk i ky yk k
for all y 2 .A C C / \ .yk C BY / for sufficiently small > 0. Taking now a
neighborhood of vk in form vk C BY , for every yQ 2 A \ .vk C BY / we have
y D yQ C tk 2 .A C C / \ .yk C BY /, and thus
hyk ; yQ vk i D hyk ; .y C tk / .vk C tk /i D hyk ; y yk i ky yk k D kyQ vk k;
which justifies that yk 2 NO .vk I A/. This together with the convergent sequence
vk ! y as k ! C1 derived from the order continuity assumption implies
y 2 N.yI A/. We have proved the inclusion (12.104) under the order
continuity assumption.
Note that the normal cones to a set and its epigraphical set are incomparable
in general. To illustrate this, consider the set
A WD f.x; y/ 2 R2 j x 2 C y 2 D 2 and x C y
0g
in R2 equipped with the usual ordering cone C WD R2C , then it is easy to check
that
N..1; 1/I A/ WD span f.1; 1/g and N..1; 1/I A C C / WD bd R2 :
Now we consider a set-valued optimization problem with a general geometric

constraint:
minimize F .x/ subject to x 2 S; (SP )
where the cost mapping F W X Y is a set-valued mapping, Y and X are

Asplund spaces, S is a subset of X , and “minimization” is understood in the sense
of Definition 2.6.1. In the following theorem we assume that F is epigraphically
Lipschitz-like (ELL) around .x; y/ (see Definition 3.3.19) and we use the limiting
subdifferential introduced in Definition 3.5.11.
Theorem 12.8.9 (Necessary Conditions for Minimizers in Set-Valued Optimiza-
tion). Let .x; y/ 2 graph F with x 2 S be a minimizer of the problem (SP ).
Assume that F is ELL around .x; y/, epi F is closed around this point, and S is

closed around x. Assume furthermore that cone F .S / C C y is closed. Then for
every k 0 2 C n f0g, there are a positive number " > 0 and a dual element y 2 Y
such that
0 2 @F .x; y/.y / C N.xI S /; y 2 N.0I C / \ N.0I Ck 0 ;" / and hy ; k 0 i D 1:

(12.105)

Proof. Since cone F .S / C C y is assumed to be closed and epi F is closed
around .x; y/ (this yields that F .S / C C y is locally closed at the origin), we
can show that for every
k 0 2 C nf0g, there exists a positive number " 2 .0; kk 0 k/
such that y 2 Min F .S / C C; Ck 0 ;" . Then, we get from Lemma 12.8.1 (iv) with
D D Ck 0 ;" that .x; y/ is a minimum of the problem

minimize '.x; y/ C A .x; y/ ;
where the cost ' W X Y ! R is given by
'.x; y/ WD 'D;k 0 .y y/;

and the indicator function A has the value 0 for all elements of A and the value
C1 for all elements of X Y not in A, and A WD epi FS . Recall that FS is the
restriction of the mapping F on S given by
(
f0g. Y / if x 2 S;
FS .x/ WD F .x/ C 4.xI S / with 4.xI S / WD
; otherwise:

Since A D epi F \ .S Y / is locally closed at .x; y/, the indicator function A is
lower semicontinuous at that point. Since the function 'D;k 0 is Lipschitz continuous,
so is the cost function '. Employing first the Fermat rule to this auxiliary scalar
optimization problem and then the sum rule to the sum of one Lipschitz continuous
function and one lower semicontinuous one, we have

0 2 @ ' C N.I A/ .x; y/

@'.x; y/ C N .x; y/I A

D f0g @'D;k 0 .0/ C N .x; y/I epi F .S Y / ;
where the equality holds due to the definitions of the functions.

Taking into account now the subdifferential of 'D;k 0 (see Lemma 12.8.1 (ii)) and
the fact that A D epi FS D graph EFS we have the existence of y 2 N.0I Ck 0;" /
with hy ; k 0 i D 1 such that

.0; y / 2 N .x; y/I A D N .x; y/I epi FS ;
which is equivalent (see Definition 3.5.8) to the following inclusion

0 2 D EFS .x; y/.y / D D EF C 4.I S / .x; y/.y /: (12.106)
Note that the graphs of both set-valued mappings EF and 4.I S / are locally closed
at .x; y/ under the closedness assumptions imposed on F and S . Note also that the
imposed ELL property of F at .x; y/ is nothing but the Lipschitz-like property of
the epigraphical multifunction EF at the same point. Hence, the pair of mappings
EF and 4.I S / meet all the assumptions of the coderivative sum rule from [431,
Proposition 3.12], and thus we get from this rule that

D EF C 4.I S / .x; y/.y / D EF .x; y/.y / C N.xI S /: (12.107)
Substituting (12.107) into (12.106) and taking into account the definition of
subdifferential
(3.29) and the implication D EF .x; y/.y / ¤ ; H) y 2
N.0I C / in (3.30), we arrive at the necessary condition (12.105)
0 2 @F .x; y/.y / C N.xI S /
and complete the proof. t

u
Note that if the cost mapping F D f W X ! Y in Theorem 12.8.9 happens to
be single-valued, the necessary conditions in Theorem 12.8.9 reduce to those in [29,
Theorem 3.8] while the reduced assumption that epi f is closed around .x; f .x// is
abundant since it is automatic from the Lipschitz continuity of f . However, the ELL
property of a set-valued mapping does not imply the closedness of the epigraph of
F in general. Consider the set-valued mapping F W R R defined by
8
ˆ
ˆ .x; C1/ if x < 0;
ˆ
<
F .x/ WD Œ0; C1/ if x D 0;
ˆ
ˆ
:̂
.x; C1/ if x > 0;
where the usual order on real numbers is used, i.e., C WD Œ0; C1/. It is easy to
check that F is ELL at .0; 0/, but its epigraph epi F D graph F is not locally closed
at this point.
In comparison with necessary conditions in [24, 25, 27, 149, 151, 157, 431, 620,
621] and the references therein, the results in Theorem 12.8.7 and Theorem 12.8.9
hold even when the ordering cone has an empty interior, and enjoys neither AC nor
SNC conditions.
Remark 12.8.10 (Necessary Conditions Under Regularity Assumptions). Mimick-
ing the penalization scheme used in [149, Theorem 4.4], we are able to obtain
a refined necessary condition for the problem (SP ) under the metric regularity
assumption, i.e., there is > 0 and a neighborhood U of .x; y/ 2 epi F \ .S Y /
such that

d .x; y/; epi F \ .S Y / d.x; S / C d..x; y/; epi F / ; for all .x; y/ 2 U:
(12.108)
Necessary conditions with regularity properties. Let .x; y/ 2 graph F with x 2

S be a Pareto minimizer of the problem (SP ), where the image space Y is partially
ordered by a proper closed, convex and pointed cone C whose interior might be
empty. Assume that S is closed around x, F .S /CC y and cone F .S /CC y
are closed around the origin. Assume also that the regularity condition (12.108) is
fulfilled. Then for every k 0 2 C n f0g, there is a positive number " > 0 and y 2 Y
with y 2 N.0I C / \ N.0I Ck 0;" / and hy ; k 0 i D 1 satisfying

.@d.x; S / fy g/ \ @d .x; y/; epi F ¤ ;; (12.109)
which surely implies the relationships in (12.105).

Remark 12.8.11 (Variants of Necessary Optimality Conditions in Banach Spaces).
Although the results in this section were established in Asplund spaces, they are
easily to extend to general Banach spaces by using other types of generalized
differentiation which have the calculus rules used in the proofs of Theorem 12.8.7
and Theorem 12.8.9; for example, Clarke’s generalized differentiations in [100], and
the Ioffe’s approximate constructions in [275–277]. In Sect. 12.7.2 we presented
corresponding results in general Banach spaces for Q-minimizers of set-valued
optimization problems (see Definition 2.6.3). In this section we used the concept
of generalized differentiability by Kruger/Mordukhovich (see Sect. 3.5) since it has
a full calculus for normal cones to sets and coderivatives for set-valued mappings
under the SNC conditions and the Mordukhovich qualification conditions; they are
satisfied under the Lipschitz-like properties, and since it is smaller or at least not
bigger than the other generalized differentiation constructions.
If we want to have necessary conditions in terms of Ioffe’s normal cones and
coderivatives, we need the corresponding conditions for the fulfillment of the
calculus rules such as the sum rule [277, Theorem 5.6 ], the chain rule [277,
Theorem 6.2 ], the chain rule for coderivatives [312, Theorem 3.1 ], and the sum rule
for coderivatives [312, Theorem 3.4]. Loosely speaking, instead of the Lipchitz-like
assumptions needed for the Mordukhovich’s calculus, the Ioffe’s calculus requires
the strongly compactness Lipschitz continuity property whose definition is quite
complicated even for vector-valued functions (compare Definition 3.3.20). We were
dealing with necessary conditions based on Ioffes’s subdifferential calculus in
Sect. 12.7.
12.9 Necessary Conditions for Approximate Solutions of Set-Valued. . . 591
12.9 Necessary Conditions for Approximate Solutions

of Set-Valued Optimization Problems
In order to show necessary conditions for approximate solutions of a set-valued

optimization problem we will use the extremal principle of variational analysis
[430, Theorem 2.20] for the case of two closed sets ˝1 and ˝2 in the Asplund
space X (see Sect. 5.5).
In this section we present a subdifferential variational principle for set-valued
mappings established in [27, Theorem 3.9] as an extension of the corresponding
scalar result by Mordukhovich and Wang. This version of the subdifferential
variational principle is derived under the limiting monotonicity condition (compare
Sect. 10.2) imposed on the set-valued mapping F W X Y with no pointedness
assumption supposed on the proper closed convex ordering cone C Y . Further-
more, we assume that F is epiclosed, i.e., its epigraph is closed in X Y . It is
easy to see that every epiclosed mapping is level-closed, but the opposite may not
be true in the case of set-valued mappings. Similarly to [25, 430, 434], the proof
of the subdifferential variational principle given below is based on the extremal
principle (Theorem 5.5.9) and the variational principle given in Corollary 10.2.12
(see [27, Theorem 3.5]), and hence it requires the Asplund property of the Banach
spaces in question. The necessary condition in the following theorem is formulated
for approximate solutions of set-valued optimization problems in the sense of
Definition 2.6.61 using the Fréchet subdifferential of F at .x; y/ 2 graph F
introduced in (3.32). The following result is shown by Bao and Mordukhovich [27,
Theorem 3.9].
Theorem 12.9.1 (Subdifferential Variational Principle for Set-Valued Map-
pings). Let X and Y be Asplund spaces, C Y a proper closed convex cone
and F W X Y . Suppose that the assumptions of Corollary 10.2.12 are fulfilled.
Furthermore, assume additionally that F is epiclosed with respect to the ordering
cone C Y .
Then for any " > 0,
> 0, k 0 2 C n .C / with jjk 0 jj D 1, and a strict
approximate "k 0 -minimizer .x 0 ; y 0 / 2 graph F for the mapping F , there exists
.x; y/ 2 graph F such that jjx x 0 jj
and
O .x; y/ \ " BX ¤ ;:
@F (12.110)

Proof. We consider the strict approximate "k 0 -minimizer .x 0 ; y 0 / 2 graph F for

the mapping F . Taking into account Definition 2.6.61 there is a positive number
"Q < " such that .x 0 ; y 0 / is an approximate "Qk 0 -minimizer for this mapping. Set
" C "Q

Q WD
with 0 <
Q <
(12.111)
2"
and apply the variational principle in Corollary 10.2.12 to the mapping F and its
Q
approximate "Qk 0 -minimizer .x 0 ; y 0 / with the chosen parameters "Q and
.
The variational principle in Corollary 10.2.12 yields the existence of .u; v/ 2
graph F with
v 2 BMMin F .u/; Q
jjx 0 ujj
; (12.112)
and
"Q
y vC jjx ujjk 0 … C for all .x; y/ 2 graph F with .x; y/ ¤ .u; v/:
Q
(12.113)
We introduce a single-valued Lipschitz continuous mapping g W X ! Y

defined by
"Q
g.x/ WD v jjx ujjk 0 (12.114)

Q
and consider the following two closed subsets of the Asplund product space X Y
given by
˝1 WD epi F and ˝2 WD graph g: (12.115)
We will show that .u; v/ is an extremal point of the set system f˝1 ; ˝2 g
from (12.115) in the sense of Definition 5.5.1. Indeed, the condition .u; v/ 2 ˝1 \˝2
is obvious. Furthermore, by (5.37) we have to check the existence of a sequence
fai g X Y such that ai ! 0 as i ! C1 and
˝1 \ .˝2 C ai / D ; for all i 2 N:
In order to do this, we show
˝1 \ .˝2 C .0; i 1 k 0 // D ; for all i 2 N ; (12.116)
i.e., (5.37) holds with ai D .0; i 1 k 0 /. By the contrary, suppose that (12.116) does
not hold for some fixed i 2 N . Then by the constructions of ˝1 and ˝2 in (12.115)
there is .x; v/ satisfying the conditions
v D g.x/ i 1 k 0 and .x; v/ 2 epi F: (12.117)
Taking into account the definition of the epigraph of the set-valued map F W
X Y , we find y 2 F .x/ and c 2 C with v D y C c. We get with v D y C c
and (12.117)
y D v c D g.x/ i 1 k 0 c 2 g.x/ C;
12.9 Necessary Conditions for Approximate Solutions of Set-Valued. . . 593
because of k 0 2 C and c 2 C . This, together with condition (12.113)

and the definition of g in (12.114), implies that .x; y/ D .u; v/. From .x; y/ D
.u; v/, (12.114), (12.117), and v D y C c we get that
y D v D g.u/ D g.x/ D v C i 1 k 0 D y C c C i 1 k 0 ;
and so c C i 1 k 0 D 0. This yields k 0 D i c 2 C in contradiction to the

choice of k 0 2 C n .C /. This contradiction implies that (12.116) holds and thus
f˝1 ; ˝2 ; .u; v/g is an extremal system in the sense of Definition 5.5.1.
Now, we can apply the extremal principle (Theorem 5.5.9) to the system
f˝1 ; ˝2 ; .u; v/g imposing the sum norm jj.x; y/jj WD jjxjj C jjyjj on X Y that
generates the dual norm
jj.x ; y /jj D maxfjjx jj ; jjy jj g for .x ; y / 2 X Y
on X Y . Proceeding in this way for any > 0, we find elements

.xj ; yj ; xj ; yj / 2 X Y X Y with j D 1; 2 satisfying the relations
.xj ; yj / 2 ˝j ; jjxj ujj C jjyj vjj ;

.xj ; yj / 2 NO ..xj ; yj /I ˝j /; j D 1; 2; (12.118)
1 1
maxfjjxj jj ; jjyj jj g C ; j D 1; 2;
2 2

maxfjjx1 C x2 jj ; jjy1 C y2 jj g : (12.119)
From the Lipschitz continuity of the mapping g in (12.114) with constant L D

"Q=
Q and the coderivative estimate from [430, Theorem 1.43] it follows that
"Q
jjx2 jj jjy2 jj and hence y2 ¤ 0

Q
by (12.118), (12.119) with > 0 sufficiently small. The latter yields by Theo-
rem 5.5.9 that
jjx1 jj "

jjy1 jj ¤ 0 and < ; (12.120)
jjy1 jj
see more details in the similar setting of [25, Theorem 3.5]. Moreover, from the third
condition in (12.118) with j D 1 we find yQ1 2 F .x1 / with
.x1 ; yQ1 / 2 graph F; .x1 ; y1 / 2 NO ..x1 ; yQ1 /I epi F /; and y1 2 NO .0I C /:
(12.121)
Finally, denoting .x; y/ WD .x1 ; yQ1 /, x WD x1 =jjy1 jj , and y WD y1 =jjy1 jj
and taking into account the definition of the Fréchet subdifferential (see (3.32)),
we get the desired subdifferential condition (12.110) from the relations in (12.120)
and (12.121). The condition jjx x 0 jj <
in the assertions of the theorem follows
from the second condition in (12.112), condition (12.118) for j D 1, and the choice
of
Q in (12.111). The proof is completed. t
u
Remark 12.9.2. In Theorem 12.9.1 it is supposed that jjk 0 jj D 1. If we don’t
assume jjk 0 jj D 1, then condition (12.110) can be replaced by the modified
subdifferential condition
O .x; y/ \ " jjk 0 jjBX ¤ ;;

@F

if jjk 0 jj is selected arbitrarily from C n .C /, with no change in the proof.

Remark 12.9.3. The proof of this theorem, based on the extremal principle (Theo-
rem 5.5.9) and the variational principle in [27, Theorem 3.5] (see Corollary 10.2.12)
is similar to that one given in [25, Theorem 2]. In [25] the reader can find more
details.
12.10 Necessary and Sufficient Conditions for Solution

Concepts Based on Set Approach
In this section we derive necessary and sufficient conditions for solutions described
by the set approach using a directional derivative of the set-valued map. We assume
that Y is a n.v.s., C is a proper closed convex pointed cone in Y with int C ¤ ;,
int C C ¤ ; and we use the notations introduced in Sect. 2.6.4. Especially, let G
be the family of all nonempty C -convex and C C -bounded subsets of Y , V D
fŒA; Bj.A; B/ 2 G 2 g.
Moreover, we assume that S is a convex subset of a normed space .X; k k/.
Consider c 2 int C and a weak* compact base W WD fy 2 C C j hy ; ci D 1g of
C C , V D V .W / , that is jŒA; Bj < C1 when ŒA; B 2 V , and F : S G .
In this section we are dealing with weakly minimal solutions defined using the
following binary relation lC on G : For A; B 2 G ,
A lC B W” 9V Y such that A C C B C V; (12.122)
where V is a neighborhood of 0. Furthermore, we discuss minimal solutions based

on the set relation introduced in (2.64)
A cl
C B W” cl.A C C / B; (12.123)
where A; B 2 G .
12.10 Necessary and Sufficient Conditions for Solution Concepts Based on. . . 595
We consider the set-valued problem
C minimize F .x/; subject to x 2 S; (SP C )
where the set order relation C is given by (12.122) or by (12.123).

The results presented in the following are shown by Kuroiwa [354]. An important
and interesting discussion of this approach is given in a recent paper by J. Jahn:
Directional derivatives in set optimization with the set less order relation (2014),
page 8.
The derivatives introduced in the following definition are used for deriving the
optimality conditions for solutions of (SP C ).
Definition 12.10.1. Let x 2 S and d 2 X .
1
CF.x; d / WD fŒA; B 2 V j 9f
k g ! C0 s.t. f ŒF .x C
k d /; F .x/g ! ŒA; Bg

k
is said to be V -directional derivative cluster of F at x in the direction d .

If CF.x; d / is a singleton, then the element is written by DF.x; d / and called
V -directional derivative of F at x in the direction d , and F is said to be V -
directional differentiable at x in the direction d .
Example 12.10.2. Let F W RR2 be a set-valued map defined for all x 2 R by
F .x/ D convf.jxj; jxj C 1/; .jxj C 1; jxj/g;
and let C D R2C D f.x1 ; x2 /jx1 ; x2

0g . Then F is V -directional differentiable
in any direction d 2 R, for example, if x0 D 0:
DF.x0 ; d / D Œf.0; 0/g; jd j convf.1; 1/; .1; 1/g .d 2 R/;
if 0 < x0 < 12 :

Œf.0; 0/g; jd j convf.1; 1/; .1; 1/g if d
0;
DF.x0 ; d / D
Œjd j convf.1; 1/; .1; 1/g; f.0; 0/g if d < 0;
and if x0 D 12 :
DF.x0 ; d / D Œjd j convf.1; 1/; .1; 1/g; f.0; 0/g .d 2 R/:
Example 12.10.3. Consider a set-valued map F W S Y defined by

X
F .x/ WD g.x/ C ri .x/Ai ; x 2 S;
i 2I
where g is a function from S to Y which is directional differentiable at x0 2 S , I

is a nonempty finite set, and for each i 2 I; ri W S ! .0; C1/ is a function which
is directional differentiable at x0 2 S , and Ai 2 G . Then F is V -directional

differentiable at x0 for each direction d 2 X , and it holds
X 0
DF.x0 ; d / D g 0 .x0 ; d /; f0g C ri .x0 ; d / ŒAi ; f0g
i 2I
2 3
X X
D 4g 0 .x0 ; d / C ri0 .x0 ; d /Ai ; ri0 .x0 ; d /Ai 5
i 2IC .d / i 2I .d /
where IC .d / D fi 2 I j ri0 .x0 ; d / > 0g and I .d / D fi 2 I j ri0 .x0 ; d / < 0g .

We present necessary and sufficient optimality conditions for solutions of
(SP C ) using the derivatives introduced in Definition 12.10.1 and the cone
.C / D fŒA; B 2 V jB clC Ag (see (2.65)), where C is defined by (12.123).
cl
The following implication is given in [354, Proposition 1].

Proposition 12.10.4. For any A; B 2 G it holds
ŒA; B 2 Int .C / H) B lC A;
where Int .C / is the set of all interior points w.r.t. .V ; j j/.

Definition 12.10.5. An element x0 2 S is said to be a weakly minimal solution
of (SP C ) if
6 9x 2 S s.t. F .x/ lC F .x0 /:
Furthermore, we define local solutions of (SP C ) using the order relation cl
C
given by (12.123).
Definition 12.10.6. An element x0 2 S is said to be a local minimal solution
of (SP C ) if there exists a neighborhood U of x0 such that
x 2 U \ S; F .x/ cl
C F .x0 / H) F .x0 / C F .x/I
cl
and x0 2 S is called a local weakly minimal solution of (SP C ) if there exists

a neighborhood U of x0 such that
6 9x 2 U \ S s.t. F .x/ lC F .x0 /:
The following necessary condition of local weakly minimal solutions of

(SP C ) is given in Kuroiwa [354, Theorem 1].
Theorem 12.10.7. If x0 is a local weakly minimal solution of (SP C ), then for
all x 2 S it holds
CF .x0 ; x x0 / \ . Int .C // D ;:
12.10 Necessary and Sufficient Conditions for Solution Concepts Based on. . . 597
Furthermore, the following sufficient condition for local minimality concern-

ing (SP C ) is presented in Kuroiwa [354, Theorem 2].
Theorem 12.10.8. Assume that X is a finite dimensional space, and F is V -
directional differentiable at x0 2 S in each direction d 2 X . Moreover, suppose
that DF .x0 ; d / D lim
!C0
1 ŒF .x0 2 S C
d /; F .x0 / converges uniformly and
continuous with respect to d 2 X on the unit ball. If for all d 2 T .S; x0 /nf0g
DF .x0 ; d / 62 .C /;
then x0 is a local minimal solution of (SP C ).

Example 12.10.9. We discuss Example 12.10.2, in view of the optimality condi-
tions given above.
At first, it is easy to check that there are no (global) minimal, and no (global)
weakly minimal solutions of (SP C ), 0 is the only local minimal solution of
(SP C ) , and for each x 2 R; x is a local weakly minimal solution of (SP C ).
If x0 D 0; DF .x0 ; d / 2 .C / holds if and only if
R2C jd j convf.1; 1/; .1; 1/g;
but it does not hold when d ¤ 0 . From Theorem 12.10.8, we get that x0 is a local
minimal solution of (SP C ) .
Furthermore, if 0 < x0 < 12 ; DF .x0 ; d / 2 .C / holds if and only if

R2C jd j convf.1; 1/; .1; 1/g if d
0;
R2C C jd j convf.1; 1/; .1; 1/g 3 .0; 0/ if d <0
and
R2C C jd j convf.1; 1/; .1; 1/g 3 .0; 0/
is always true when d < 0. This is consistent with Theorem 12.10.8, since x0 is not
a local minimal solution.
Moreover, DF .x0 ; d / 2 Int.C / holds if and only if

.0; C1/2 jd j convf.1; 1/; .1; 1/g if d
0
.0; C1/2 C jd j convf.1; 1/; .1; 1/g 3 .0; 0/ if d < 0
does not hold for each d 2 R . This is also consistent with Theorem 12.10.7, since
x0 is a local weakly minimal solution of (SP C ) .
12.11 Necessary Conditions for Solution Concepts

with Respect to a General Preference Relation
In this section we present necessary conditions for solutions introduced in

Sect. 2.6.5 that are related to applications in welfare economics (see Sect. 15.3).
The main tool for the proof of the necessary conditions is the extended extremal
principle in Theorem 5.5.10.
We derive the necessary optimality conditions for set-valued optimization prob-
lems with geometric constraints
where X and Y are Banach spaces, S is a subset of X and the cost mapping
F W S Y is a set-valued mapping.
The “minimization” in (SP) is understood with respect to some preference
relation on Y (see Sect. 2.6.5). We consider a general (abstract) preference relation
on Y given by a nonempty subset R Y Y (y 1 is preferred to y 2 (y 1 y 2 ) if
.y 1 ; y 2 / 2 R) and the level-set mapping L W Y Y associated with the preference
relation :
L.y/ WD fu 2 Y j u yg: (12.124)
To derive necessary conditions for fully localized minimizers of the set-valued

optimization problem (SP) (Definition 2.6.49), Bao and Mordukhovich [28] devel-
oped a variational approach based on the extremal principle (Theorem 5.5.10) and
generalized differential calculus (see Sect. 3.5, compare [430]).
In order to prove first order necessary conditions for localized minimizers
in Theorem 12.11.3 the following asymptotic closedness property is supposed
concerning the level-set mapping L.
Definition 12.11.1 (Asymptotic Closedness Property). Let A Y be a subset in
a Banach space Y , and let y 2 cl A. We say that A is asymptotically closed at y if
there is a neighborhood V of y and for any " > 0 there is c 2 "BY such that
.cl A C c/ \ V A n fyg:
This asymptotic closedness property is not very restrictive and fulfilled for many
models in welfare economics (compare [28, Section 4]). This is illustrated by the
following examples given by Bao and Mordukhovich [28].
Example 12.11.2. • Each proper closed convex pointed cone C Y with C n
.C / ¤ ; has the asymptotic closedness property at the origin.
• The asymptotic closedness property at the origin is satisfied for each proper
convex cone C Y with nonempty interior and for its nonconvex complement
Y n C.
12.11 Necessary Conditions for Solution Concepts with Respect to a General. . . 599
• The epigraph of a functional ' W X ! R [ fC1g that is lower semicontinuous

around x 2 X has the asymptotic closedness property at .x; '.x//.
The following necessary optimality condition for all the types of fully localized
minimizers of the set-valued optimization problem (SP) in the sense of Defini-
tion 2.6.49 is shown in [28, Theorem 3.4] using the extended extremal principle
in Theorem 5.5.10 and will be applied in Sect. 15.3 to set-valued problems in
welfare economics. The SNC assumptions in the following theorem are introduced
in Sect. 5.5.
Theorem 12.11.3 (First Order Necessary Conditions for Localized Minimizers
of (SP)). Let X and Y be Asplund spaces and F W X Y be a set-valued mapping.
Assume that graph F is closed around some point .x; y/ 2 graph F , let S X be
closed around x, and let L W Y Y be a preference mapping on Y locally satiated
at y. Furthermore, suppose that one of the following SNC assumptions on the initial
data .F; S; L/ is fulfilled:
(a) graph F is SNC at .x; y/;
(b) S is SNC at x and cl L.y/ is SNC at y;
(c) F is PSNC at .x; y/ and cl L.y/ is SNC at y;
(d) S is SNC at x and F 1 is PSNC at .y; x/.
Then there exists a pair .0; 0/ ¤ .x ; y / 2 X Y satisfying the necessary
optimality conditions
x 2 D F .x; y/.y / \ .N.xI S // and y 2 N.yI cl L.y// (12.125)
in each of the following cases:

• .x; y/ is a fully localized weak minimizer for (SP) provided that the set L.y/ is
asymptotically closed at y 2 cl L.y/;
• .x; y/ is a fully localized minimizer for (SP) provided that the set cl L.y/ is
asymptotically closed at y;
• .x; y/ is a fully localized strong minimizer for (SP) provided that either cl L.y/
or S is asymptotically closed at y or x, respectively.
Proof. The main tool for the proof is the application of the extended extremal
principle in Theorem 5.5.10. Therefore, it is to check that a fully localized minimizer
.x; y/ is a local extremal point of the set system f˝1 ; ˝2 g (in the Asplund space
X Y ) with
˝1 WD graph F and ˝2 WD S cl L.y/
(see [28, Theorem 3.4]). t

u
12.12 KKT-Points and Corresponding Stability Results
The aim of this section is to study some set-valued optimization problems from the
point of view of stability of KKT points. All of the results presented in this section
are derived by Durea, Dutta, Tammer in [152].
Let us consider the set-valued map F W X Y , G W X Z, where X , Y and
Z are finite dimensional normed vector spaces. Furthermore, C Y and K Z
are proper closed convex pointed cones.
In this section, we study the following set-valued optimization problems where
we use the vector approach as solution concept, this means that the “minimization”
is to understand in the sense of Definition 2.6.1. Consider the problem
minimize F .x/; subject to x2S (SP )
where the feasible set S is given by
S WD fx 2 X j 0 2 G.x/ C Kg: (12.126)
In fact, we shall also consider a simplified version of problem (SP ) where G is not
a set-valued map but a single-valued one denoted by g. Then the problem above is
represented as
minimize F .x/; subject to x 2 S 0 ; (SP 0 )
where the feasible set is given by
S 0 WD fx 2 X j g.x/ 2 Kg: (12.127)
The main tools we use in this section are the generalized differentiability
concepts introduced by Mordukhovich (see [430, Chapters 1, 3] and Defini-
tions 3.5.6, 3.5.8). It is well known that these objects enjoy rich calculus in Asplund
spaces. However, as we already said, we work here on finite dimensional vector
spaces.
A set-valued mapping H acting between two finite dimensional normed vector
spaces X and Y is said to be inner-semicompact at x 2 X if for every sequence
xk ! x there is a sequence yk 2 H.xk / which has a convergent subsequence. For
example if H is uniformly bounded at x then H is inner-semicompact at x.
We start by considering the problem (SP ) assuming that F is Lipschitz-like at
.x; y/ in the sense of Definition 3.3.9. The following theorem is shown in [150,
Theorem 4.5].
12.12 KKT-Points and Corresponding Stability Results 601
Theorem 12.12.1. Consider the problem (SP ) and let .x; y/ 2 graph F be a
minimizer of (SP ) in the sense of Definition 2.6.1. Assume that F is Lipschitz-
like around .x; y/ and G.x/ is compact. Further assume that the mapping x 7!
G.x/ \ .K/ is inner-semicompact at x. Moreover, we assume to be true the
following qualification condition: For any z 2 G.x/ \ .K/ one has
N.zI K/ \ ker D G.x; z/ D f0g: (12.128)
Then there exist y 2 C C with jjy jj D 1 such that the set
Lıy D fz 2 Z j 9z 2 G.x/ \ .K/; with z 2 N.zI K/ such that

0 2 D F .x; y/.y / C D G.x; z/.z /g: (12.129)
is nonempty and bounded.

Now, we introduce a notion of approximate KKT points in this set-valued setting.
We emphasize the fact that, as always in the case of generalized differentiation
objects, the main motivation for their study is related to the nice theoretical
properties. For example, in many cases the set-valued maps can actually be single-
valued in the neighborhood of a point where it is single-valued. This is indeed
true for the subdifferential of a cone-convex vector-valued function. So, from the
computational point of view, in the minimizing sequence of " KKT points, we
may just get points where the subdifferential is actually single-valued.
Definition 12.12.2. Let "
0. A feasible point .x; y/ 2 graph F of problem (SP )
is said to be an "KKT point for (SP ) if there exist y 2 C C ; jjy jj D 1; z 2 Z;
z 2 G.x/ \ K with z 2 N.z; K/ and x 2 X ; kx k " s.t.
x 2 D F .x; y/.y / C D G.x; z/.z /:
Denote by T ı the set-valued map which assigns to every " 2 RC the subset of
X Y of all " KKT points of (SP ). In this nonsmooth set-valued case the stability
result we have is the following. When " D 0 in Definition 12.12.2, .x; y/ is called a
KKT point for (SP ).
Theorem 12.12.3. Suppose that F is Lipschitz-like around .x; y/ 2 graph F; G
has closed graph, there exists r > 0 such that G.B.x; r// is bounded and for any
z 2 G.x/ \ K
N.zI K/ \ ker D G.x; z/ D f0g:
If "
0 and .x; y/ 2 Limsup!" T ı ./, then .x; y/ 2 T ı ."/.
Proof. Let " 2 RC and take .x; y/ 2 Limsup!" T ı ./. Thus, there exist some
sequences fn g ! "; f.xn ; yn /g graph F; f.xn ; yn /g ! .x; y/; fyn g
C C ; kyn k D 1; fzn g Z ; zn 2 G.xn / \ K with zn 2 N.zn I K/; fxn g X ;
kxn k n s.t.
xn 2 D F .xn ; yn /.yn / C D G.xn ; zn /.zn /: (12.130)
First, we show that fzn g is bounded. Suppose that this is not the case. There exists
a subsequence (denoted fzn g as well) s.t. kzn k ! C1. It is clear that zn ¤ 0 for
n large enough, whence, by division with kzn k ; one gets
1 1 1
kzn k xn 2 D F .xn ; yn /.kzn k yn / C D G.xn ; zn /.kzn k zn /;
i.e., there exist fun g X s.t.
1
un 2 D F .xn ; yn /.kzn k yn /
1 1
un C kzn k xn 2 D G.xn ; zn /.kzn k zn /:
Note that in finite dimensions the graph of the set-valued mapping N.I K/ is
closed (see [499]). Now we employ the hypotheses in order to deduce, by moving
to subsequences if needed, that kzn k1 1
yn ! 0; kzn k xn ! 0; un ! u 2 X

(since F is Lipschitz-like around .x; y/; fun g is bounded, see [430, Theorem 1.43])
and zn ! z 2 G.x/ \ K (since xn B.x; r/ for n large enough), kzn k1
zn !

z 2 N.zI K/ n f0g, and
u 2 D F .x; y/.0/
u 2 D G.x; z/.z /:
Therefore,
0 2 D F .x; y/.0/ C D G.x; z/.z /:
One uses again the fact that F is Lipschitz-like around .x; y/ to observe that
D F .x; y/.0/ D f0g; whence
0 2 D G.x; z/.z /
in contradiction with the qualification condition we have imposed. Therefore, for

the second step of the proof, we know that in relation (12.130) the sequence fzn g
is bounded so it is convergent (on a subsequence, if necessary) towards an element
z 2 N.zI K/ for an z 2 G.x/ \ K. Similarly, fyn g is convergent towards
12.12 KKT-Points and Corresponding Stability Results 603
an element y of norm 1 from C C . Moreover fxn g is bounded too, whence it is

convergent to some x 2 X with kx k ". Passing to the limit (as above)
in (12.130) we get
x 2 D F .x; y/.y / C D G.x; z/.z /
i.e., .x; y/ 2 T ı ."/. The proof is complete. t

u
We have the following consequence.
Corollary 12.12.4. Suppose that f.xn ; yn /g graph F are "n KKT points
for (SP ) where f"n g ! 0 s.t. f"n g RC and .x; y/ 2 graph F s.t. f.xn ; yn /g !
.x; y/. Suppose that F is Lipschitz-like around .x; y/ 2 graph F; G has closed
graph, fG.xn /g is uniformly bounded and for any z 2 G.x/ \ K
N.zI K/ \ ker D G.x; z/ D f0g:
Then .x; y/ is a KKT point for (SP ).

We turn now our attention to the problem (SP 0 ). If we adapt [150, Theorem 4.4]
to the case of finite dimensional spaces we get the following result.
Theorem 12.12.5. Let .x; y/ be a minimizer of (SP 0 ) in the sense of Defini-
tion 2.6.1. Suppose that F is Lipschitz-like around .x; y/ and g is locally Lipschitz
at x. If
z 2 K C ; 0 2 D g.x/.z / ) z D 0; (12.131)
then there exists y 2 C C ; ky k D 1 s. t.
Lı0 C
y WD fz 2 K j 0 2 D F .x; y/.y / C D g.x/.z /g: (12.132)
is non-empty and bounded.

Definition 12.12.6. Let "
0. A feasible point .x; y/ 2 graph F of problem .SP 0 /
is said to be an " KKT point for (SP 0 ) if there exists y 2 C C ; jjy jj D 1;
z 2 K C and x 2 X ; kx k " s.t.
x 2 D F .x; y/.y / C D g.x/.z /:
Denote by T ı0 the set-valued map which assigns to every " 2 RC the subset of
X Y of all " KKT points of (SP 0 ). If " D 0 in Definition 12.12.6, then .x; y/
is called a KKT point for (SP 0 ). Our stability results for this problem uses the same
hypotheses as above. The proof is also on similar lines and therefore we omit it.
Theorem 12.12.7. Suppose that F is Lipschitz-like around .x; y/ 2 graph F; and

g is locally Lipschitz at x and the following implication holds:
z 2 K C ; 0 2 D g.x/.z / ) z D 0:
If "
0 and .x; y/ 2 Limsup!" T ı0 ./, then .x; y/ 2 T ı0 ."/.
Again, we record a corollary.
Corollary 12.12.8. Suppose that .xn ; yn / graph F are "n KKT points for (SP 0 )
where ."n / ! 0 s.t. ."n / RC and .x; y/ 2 graph F s.t. .xn ; yn / ! .x; y/.
Suppose that F is Lipschitz-like around .x; y/ 2 graph F; and g is locally Lipschitz
at x and the following implication holds:
z 2 K C ; 0 2 D g.x/.z / ) z D 0:
Then .x; y/ is a KKT point for (SP 0 ).

Chapter 13
Sensitivity Analysis in Set-Valued Optimization
and Vector Variational Inequalities
Sensitivity analysis, the quantitative analysis of the perturbation map, is of

paramount interest in optimization theory and has applications in several branches of
pure and applied mathematics. During the last five decades, substantial progress has
been made in sensitivity analysis for optimization problems with scalar objectives.
On the other hand, the differentiability issues of the perturbation map for vector
optimization problems and set optimization problems are rather involved and they
require modern tools from variational analysis. The main difficulty here stems from
the fact that the perturbation map for such problems is, in general, set-valued.
To fix these ideas, we begin by describing the general framework for such
problems. Let X , Y , and Z be normed spaces, and let C Y be a nonempty
cone in Y . Let H W X Z and G W X ZY be set-valued maps. We consider the
following parametrized family of set-valued optimization problems defined by
P .x/ D Min.F .x/; C /; (13.1)
where
F .x/ WD fy 2 G.x; z/j z 2 H.x/g: (13.2)
Here we interpret X as the parameter space, Z as the decision space, and Y

as the objective space, with the cone C inducing a partial ordering on objective
values in Y . For each value of the parameter x, the set H.x/ defines a feasible
set in Z, whereas the set G.x; z/ gives a set of objectives in the space Y . The set
F .x/ comprises the graph of the objective values over the entire feasible region of
the optimization problem. In this setting, P .x/ singles out those objective values
that cannot be improved relative to the ordering induced by C . Evidently, when
G is single-valued, the above model then recovers the setting of parametric vector
optimization. If additionally, Y C R and C D RC , then we are in the framework of
scalar parametric optimization. The objective of this chapter is to employ derivatives

DOI 10.1007/978-3-642-54265-7__13
606 13 Sensitivity Analysis in Set-Valued Optimization and Vector Variational Inequalities
of set-valued maps to estimate how sets F .x/ and P .x/ vary with changes in the
parameter x. For this purpose, we will use graphical derivatives as well as the
coderivatives of these set-valued maps. For this study, a crucial role is played by
the derivatives of F , F C C , P , P C C and their minimal points. In this chapter, we
will also study differentiability properties of set-valued gap functions associated
to vector variational inequalities and sensitivity analysis for vector variational
inequalities. For set-valued optimization and for vector variational inequalities, both
the first-order as well as the second-order results are given.
13.1 First Order Sensitivity Analysis in Set-Valued

Optimization
Many authors have focused on sensitivity analysis using first-order contingent

derivatives. For instance, Tanino [564, 565] derived many interesting relationships
among the first-order contingent derivative of the perturbation maps associated
with some set-valued maps. Since then, several authors have investigated the
differentiability of set-valued perturbation maps (see [99, 344, 345, 530, 531, 547]
and the cited references therein). Interesting results on sensitivity analysis for non-
smooth optimization problems are available in [462, 594], among others. A very
comprehensive treatment of sensitivity analysis based on graphical derivatives can
be found in [214, Section 3.6] and which is the main source for our presentation in
this section.
Throughout this section, X and Y are normed spaces, and f0g ¤ C Y is a
pointed, closed, and convex cone. When dealing with weakly minimal points of a
set, we also assume that int.C / ¤ ;. Let F W X Y be a given set-valued map. We
are interested in the assessment of the differentiability properties of the set-valued
perturbation maps P; Q, and R, defined for x 2 X , by
P .x/ D GHe-PMin.F .x/; C /; (13.3a)

Q.x/ D Min.F .x/; C /; (13.3b)
R.x/ D WMin.F .x/; C /; (13.3c)
where GHe-PMin.F .x/; C / is the set of global Henig-proper minimal points of

F .x/ with respect to C , Min.F .x/; C / is the set of minimal points of F .x/ with
respect to C , and WMin.F .x/; C / is the set of weakly minimal points of F .x/ with
respect to C .
The perturbation maps P; Q, and R will be termed as the proper perturbation
map, the perturbation map, and the weak perturbation map, respectively. Our goal
here is to investigate the relationships among the first-order contingent derivatives
of the perturbation maps P , Q, and R, the first-order contingent derivative of the
map F , and various minimal points of these derivatives.
13.1 First Order Sensitivity Analysis in Set-Valued Optimization 607
We begin by recalling the following result, which is a restatement of The-

orem 11.5.20 and which was proved under the assumption that the map F is
contingently C -absorbing. In the following, the imposed conditions ensure that F
is contingently C -absorbing.
Theorem 13.1.1. Assume that fy 2 C j kyk D 1g is compact and
N y/.0/
DS F .x; N \ .C / D f0g; (13.4)
N y/
where DS F .x; N is the S -derivative of F at .x;
N y/.
N Then, for every x 2 X , we have
N y/.x/;
Min .DF.x; N C / D Min .D.F C C /.x;
N y/.x/;
N C/: (13.5)
N y/.x/;
GHe-PMin .DF.x; N C / D GHe-PMin .D.F C C /.x;
N y/.x/;
N C/:
(13.6)
Moreover, if intC ¤ ;, then for every x 2 X , we have
N y/.x/;
WMin .DF.x; N C / WMin .D.F C C /.x;
N y/.x/;
N C/; (13.7)

N y/.x/;
WMin .DF.x; N C / D WMin D.F C CQ /.x;
N y/.x/;
N C ; (13.8)
where CQ is a closed convex cone such that CQ f0g [ int.C /.

Proof. See [214]. t
u
Remark 13.1.2. When X , Y are finite-dimensional, formulae (13.5), (13.6),
and (13.8) are established in [345, Theorem 2.1], while formulae (13.6) and (13.8)
are established in [344, Theorem 2.1] for xN 2 int.domF / and F locally C -convex
N
near x.
Remark 13.1.3. It is easy to verify that in Example 11.3.4, (13.4) is not satisfied
and (13.6), (13.7), (13.8) (for CQ D C ) do not hold for x > 0.
We have the following result concerning the differentiability of P and Q.
Theorem 13.1.4. Let .x;N y/
N 2 graph F . Assume that fy 2 C j kyk D 1g is
compact and F is C -dominated by Q near x.N That is, F .x/ Q.x/ C C for
every x 2 V for some neighborhood V of x.
N
(i) If yN 2 Q.x/,
N then for every x 2 X ,
Min .D.F C C /.x;

N y/.x/;
N C / Min .DQ.x;
N y/.x/;
N C / DQ.x;
N y/.x/:
N
(13.9)
Moreover, if F .x/ is convex for x 2 V , with V a neighborhood of x,

N then for
every x 2 X ,
N y/.x/;
Min .DF.x; N C / Min .DQ.x;
N y/.x/;
N C/: (13.10)
(ii) If (13.4) holds, then yN 2 Q.x/

N and for every x 2 X ,
N y/.x/;
Min .DF.x; N C / D Min .DQ.x;
N y/.x/;
N C / DQ.x;
N y/.x/;
N
(13.11)
N y/.x/;
GHe-PMin .DF.x; N C / D GHe-PMin .DQ.x;
N y/.x/;
N C / DQ.x;
N y/.x/:
N
(13.12)
Proof. See [214]. t

u
Remark 13.1.5. Formula (13.9) of Theorem 13.1.4 is proved by Tanino [561,
Theorem 3.1] for finite-dimensional Y and by Klose [330, Theorem 3.4] for
arbitrary Y ; for finite-dimensional spaces X and Y , formula (13.10) is proved by
Shi [531, Theorem 4.1], Theorem 13.1.4 (ii) is proved by Shi [530, Theorem 4.1].
In the next example, which is taken from [345], F .x/ is not C -dominated
for x < 0, but the other conditions of Theorem 13.1.4 hold. As we show
below, (13.10), (13.11), and (13.12) do not hold for x < 0.
Example 13.1.6. Let X D R, Y D R2 , C D R2C , and F W R R2 be defined by

f.y1 ; y2 / j y1
0; y2
y12 g if x
0;
F .x/ WD
f.y1 ; y2 / j y1 > 0; y2
y12 g if x < 0:
Then

f.0; 0/g if x
0;
Q.x/ WD
; if x < 0:
Take xN D 0, yN D .0; 0/. Then
N y/.x/
DF.x; N D D.F C C /.x;
N y/.x/
N D C;
N y/.x/
DQ.x; N D Q.x/;
for every x 2 X . t
u
In the following corollary Bo-PMin stands for the Borwein minimal points.
Corollary 13.1.7. Suppose that yN 2 Bo-PMin .F .x/;
N C / and F is C -dominated
N Then (13.11) and (13.12) hold if one of the following conditions is
by Q near x.
satisfied:
(i) F is upper Lipschitz at xN and fy 2 C j kyk D 1g is compact;
(ii) X , Y are finite-dimensional, xN 2 int.domF /, and F is C -convex near x.
N
Remark 13.1.8. Corollary 13.1.7 (i) was obtained by Tanino [561, Theorem 3.2] for
Y finite-dimensional, Shi [530, Corollary 4.1], and Klose [330, Theorem 3.5], while
part (ii) by Shi [531, Theorem 4.2].
The following is an analogue of Theorem 13.1.4 for the weak perturbation map:
N 2 graph .F /. Assume that (13.4) holds, dim Y < 1,
intC ¤ ;, and F is CQ -dominated by R near xN for some closed convex cone CQ with
CQ f0g [ intC . Then for every x 2 X ,
N y/.x/;
WMin .DF.x; N C / D WMin .DR.x;
N y/.x/;
N C / DR.x;
N y/.x/:
N (13.13)
Proof. Since (13.4) holds, we have that yN 2 Q.x/ N R.x/.N Since the conditions
of Corollary 13.1.1 hold, equality (13.8) holds, too. Since graph .R/ graph .F /,
condition (13.4) is satisfied with F replaced by R, hence (13.8) holds for the same
substitution. Because F is CQ -dominated by R near x,N we have that .F C CQ /.x/ D
.R C CQ /.x/ for every x in a neighborhood of x,
N hence
D.F C CQ /.x; N D D.R C CQ /.x;

N y/ N y/:
N
The conclusion is obtained immediately by using the properties of minimal

points. t
u
Remark 13.1.10. Theorem 13.1.9 was proved by Kuk, Tanino, and Tanaka in [345,
Theorem 3.1] in the same conditions.
Remark 13.1.11. In Examples 11.3.4 and 13.1.6 relation (13.13) does not hold.
Note that in Example 11.3.4, (13.4) is violated and in Example 13.1.6, F .x/
is not CQ -dominated by R.x/ for x < 0 for some closed convex cone CQ with
CQ f0g [ intC .
The following result is immediate:
Corollary 13.1.12. Suppose that X , Y are finite-dimensional, xN 2 int.domF /, yN 2
N If F is CQ -dominated by R near xN
N C /, and F is C -convex near x.
Bo-PMin .F .x/;
for some closed convex cone CQ with CQ f0g [ intC , then (13.13) holds.
N
In the following example, which is taken from [531], F is not C -convex (near x)
and relation (13.13) does not hold for every x 2 X .
Example 13.1.13. Let X D Y D R, C D RC , and let F W R R be defined by
p
jxj ; 1Œ [f jxjg if jxj 1;
Œp
F .x/ WD
Œ jxj; 1Œ if jxj > 1:
Then, for every x 2 X ,

p
.F C C /.x/ D Œ jxj; 1Œ;
p
Q.x/ D f jxjg:
Take xN D 0, yN D 0. Then,
N y/.x/
DF.x; N D Œ jxj ; 1Œ; D.F C C /.x;
N y/.x/
N D X; 8 x 2 X;

1; 0 if x D 0;
N y/.x/
DQ.x; N D
; if x ¤ 0:
Corollary 13.1.14. Suppose that Y is finite-dimensional, yN 2 Bo-PMin .F .x/;

N C /,
N and one of the following two
F .x/ is C -closed for all x in a neighborhood of x,
conditions is satisfied:
(i) (13.4) holds and F .x/ is C -bounded (i.e., .F .x//1 \ .C / D f0g) for all x
N
in a neighborhood of x;
(ii) X is finite-dimensional, xN 2 int.domF /, and F is C -convex and C -dominated
N
by Q near x.
Then for every x 2 X ,
N y/.x/;
Min .DF.x; N C / D Min .DP.x;
N y/.x/;
N C / DP.x;
N y/.x/:
N
N y/.0/
Remark 13.1.15. In Example 13.1.13, DS F .x; N D X . So (13.4) does not hold
N the conclusion of the preceding corollary does not
and F is not C -convex near xI
hold.
We have the following analogue for the weak perturbation map:
Theorem 13.1.16. Suppose that intC ¤ ; and consider yN 2 R.x/.
N If one of the
following conditions holds,
N y/I
(i) F is semidifferentiable at .x; N
(ii) F is C -convex and xN 2 int.domF /;
(iii) X , Y are finite-dimensional, F is C -convex and xN 2 raint.domF /,
then for every x 2 X , we have
N y/.x/
DR.x; N WMin .DF.x;
N y/.x/;
N C/: (13.14)
Suppose that either (a) (13.4) and (i) hold or (b) yN 2 Bo-PMin .F .x/;
N C / and
(iii) hold. If fy 2 C j kyk D 1g is compact and F is CQ -dominated by R or Q near
N where CQ f0g [ intC is a closed convex cone, then for every x 2 X ,
x,
N y/.x/
DR.x; N D WMin .DF.x;
N y/.x/;
N C/; (13.15)
or
N y/.x/
DQ.x; N D WMin .DF.x;
N y/.x/;
N C/; (13.16)
respectively.
Proof. Assume that condition (i) holds and consider y 2 DR.x; N y/.x/.
N If y does not
belong to WMin .DF.x; N y/.x/;
N C /, then there exists k 2 int.C / such that y k 2
DR.x; N y/.x/.
N Therefore there are ftn g ! 0C and f.xn ; yn /g ! .x; y k/ such
that .x;N y/N C tn .xn ; yn / 2 graph .R/ graph .F /. Since F is semidifferentiable
N y/
at .x; N and ftn g ! 0C , xn ! u, there exists fzn g ! y k such that .x; N y/
N C
tn .xn ; zn / 2 graph F for every n
n0 . Therefore yN C tn zn 2 F .xN C tn xn / and
yN C tn yn 2 R.xN C tn xn / F .xN C tn xn /. Since

.yN C tn zn / .yN C tn yn / =tn D zn yn ! y k y D k 2 int.C /;
there exists some n1

n0 such that .yN C tn zn / .yN C tn yn / 2 int.C / for every
n
n1 , contradicting the fact that yN C tn yn 2 R.xN C tn xn /.
Assume now that (ii) holds; since intC ¤ ;, .x; N y;
N k/ 2 int.graph .F C C // for
k 2 intC . By Theorem 11.1.36 (i) .F C C / is semidifferentiable at .x; N y/.
N Taking
y 2 DR.x; N y/.x/,
N with the same argument as in the proof of (i), we obtain the same
sequences, the sole difference being that yN C tn zn 2 F .xN C tn xn / C C for n
n0 .
The same contradiction is obtained.
If (iii) holds, by Theorem 11.1.36 (ii), .F C C / is semidifferentiable at .x;N y/.
N
The proof is the same as for (ii).
(b) ) (a). Indeed, if (iii) holds, using Theorem 11.3.3 (ii) we get that (13.4) is
satisfied.
Assume that (a) is satisfied and F .x/ is CQ -dominated by R.x/ for all x 2 V ,
for some neighborhood V of x. N It follows that F .x/ C CQ D R.x/ C CQ for x 2 V .
Q
So, D.F C C /.x; N D D.R C CQ /.x;
N y/ N Applying Theorem 11.3.7 for .F; CQ / and
N y/.
Q
.R; C /, we obtain that DF.x; N y/.x/
N C CQ D DR.x; N y/.u/
N C CQ for every x 2 X .
Q
Applying now Proposition p12 (iv) for the pair .C ; C0 /, we get for every x 2 X ,
N y/.x/;
WMin .DF.x; N C / D WMin .DR.x;
N y/.x/;
N / DR.x;
N y/.x/:
N
From (13.14) and the above relation (13.15) follows.

If F is CQ -dominated by Q near x,N all we said above is true if we replace R by
Q. Therefore (13.16) holds, too. Note that yN 2 Q.x/
N because (13.4) holds. t
u
Remark 13.1.17. In Example 11.3.4 F is neither semidifferentiable at x, N nor C -
convex; relation (13.14) does not hold for x D 0. In Example 13.1.6 (a) and (b) are
satisfied but F is not CQ -dominated by R or Q near x,N for some closed convex cone
CQ f0g [ intC I relations (13.15) and (13.16) do not hold for x < 0.
In the next example, which is taken from [561], neither (a) nor (b) is satisfied;
relations (13.15) and (13.16) do not hold for x > 0.
Example 13.1.18. Let X D R, Y D R2 , C D R2C , and let F W R R2 be
defined by
F .x/ WD f.y1 ; y1 / j y1 xg [ f.y1 ; 1 y1 / j y1 > 0g:

Then
Q.x/ D R.x/ D f.y1 ; y1 / j y1 min.0; x/g [ f.y1 ; 1 y1 / j y1 > 0g:
Let xN D 0 and yN D .0; 0/. Then
N y/.x/
DF.x; N D Min .DF.x;
N y/.x/;
N C / D f.v1 ; v1 / j v1 xg;
N y/.x/
DQ.x; N D Min .DF.x;
N y/.x/;
N C / D f.v1 ; v1 / j v1 min.0; x/g:
Remark 13.1.19. For R replaced by Q and X , Y finite-dimensional, (13.14) was

obtained by Shi [531, Theorem 5.1] under condition (iii) and by Kuk, Tanino, and
Tanaka [345, Theorem 3.3] under (i). Kuk, Tanino, and Tanaka obtained (13.14)
under condition (iii) and (13.15) under condition (b) in [344, Theorem 3.2] and
(13.16) under condition (b) in [344, Theorem 3.3].
We now consider another normed vector space U , a set-valued map G W U X ,
a single-valued map f W X U ! Y , and define a set-valued F W U Y by
F .x/ WD f .G.x/ fxg/.
With this preparation, we have the following:
Theorem 13.1.20. Suppose that the following conditions hold:
(i) G is upper Lipschitz at uN 2 dom.G/ and G.u/ is compact for u 2 U , with U a
neighborhood of uN ;
(ii) xN 2 G.Nu/ and yN D f .x;
N uN / 2 BMin.F .Nu/I C /;
(iii) X is finite-dimensional, C has a compact base, f is Lipschitz on bounded sets
and Fréchet differentiable at .x;N uN /;
(iv) G.Nu/ D fxg N or the set-valued map
GQ W U Y X; Q
G.u; y/ WD fx 2 G.u/ j f .x; u/ D yg;
is upper Lipschitz at .Nu; y/ Q u; y/

N and G.N N D fxg.
N
Then for every u 2 X ,
N uN / .DG.Nu; x/.u//
Min .rx f .x; N C ru f .x;
N uN /.u/; C / D Min .DQ.Nu; y/.u/;
N C/:
In order to prove this result we need the following lemma.

Lemma 13.1.21. Let G, f , F be as above and uN 2 domG. If f is Lipschitz on
bounded sets, G.Nu/ is bounded, and G is upper Lipschitz at uN , then F is upper
Lipschitz at uN .
Proof. See [214, Lemma 3.6.23]. t
u
Remark 13.1.22. Evidently, a sufficient condition for f to be Lipschitz on bounded
sets is that f be of class C 1 and U and X have finite dimension.
13.2 Second Order Sensitivity Analysis in Set-Valued Optimization 613
Proof of Theorem 13.1.20. From (i) and (iii), using Lemma 13.1.21, we obtain that
F is upper Lipschitz at uN . Since yN 2 Bo-PMin.F .Nu/; C /, from Theorem 11.3.3 (i),
N Since G.u/ is compact for u 2 U and f
we obtain that (13.4) holds for F at .Nu; y/.
is continuous, F .u/ is compact for u 2 U . Therefore F .u/ Q.u/ C C for u 2 U .
Now using Theorem 13.1.4(ii), for every u 2 U , we have
N
Min .DF.Nu; y/.u/I C / D Min .DQ.Nu; y/.u/I
N C/:
However, from the second part of Theorem 11.1.19, for every u 2 U , we have
N
DF.Nu; y/.u/ D rx f .x;
N uN / .DG.Nu; x/.u//
N C ru f .x;
N uN /.u/:
The conclusion follows. t

u
Remark 13.1.23. The above result is taken from [214, Theorem 3.6.22]. A slightly
weaker version of this result was originally given by Tanino in [561, Theorem 4.1]
for U , X , Y finite-dimensional and f of class C 1 and by Klose [330, Theorem 4.4].
Variants of Lemma 13.1.21 are stated and proved in [561, Lemma 4.1] and [330,
Lemmas 4.2, 4.3].
13.2 Second Order Sensitivity Analysis in Set-Valued

Optimization
This section is devoted to the second-order sensitivity analysis for set-valued

perturbation maps. To obtain second-order analogues of the known first-order
sensitivity analysis results presented in the previous section, we introduce and
employ several concepts. This includes a second-order analogue of the S -derivative
and second-order directional compactness, among others.
Let X and Y be normed spaces, and let C Y be a pointed, closed, and convex
cone. When dealing with weak minimality, we will assume that int.C / is non-empty.
Let F W X Y be a given set-valued map. We are interested in the assessment of
differentiability properties of the set-valued perturbation maps P; Q, and R, defined
in (13.3). That is,
P .x/ D GHe-PMin.F .x/; C /;

Q.x/ D Min.F .x/; C /;
R.x/ D WMin.F .x/; C /:
Our goal here is to investigate the relationships among the second-order contin-
gent derivatives of the perturbation maps P , Q, and R, the second-order contingent
derivative of the map F , and various minimal points of these derivatives.
We begin with the following result.

closed, and convex cone, and let F W X Y be a set-valued map. Let .x; N y/
N 2
graph .F / and let .Nu; vN / 2 X Y be arbitrary. Assume that the set B D fz 2
C j kzk D 1g is compact. Assume that F is C -dominated by Q near x. N Then for
every x 2 U.x/,
N where U.x/ N is a neighbourhood of x,
N we have
Min.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C / D 2 Q.x;
N y;
N uN ; vN /.x/: (13.17)
Proof. Since Q.x/ F .x/ and F is C -dominated by Q near x, N there exists a

N of xN such that F .x/ C C D Q.x/ C C for every x 2 U.x/.
neighbourhood U.x/ N
Consequently
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ D D 2 .Q C C /.x;
N y;
N uN ; vN /.x/; for every x 2 U.x/;
N
which implies that
Min.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C /
D Min.D 2 .Q C C /.x;
N y;
N uN ; vN /.x/; C /; for every x 2 U.x/:
N (13.18)
Because B is compact, we have
Min.D 2 .Q C C /.x;
N y;
N uN ; vN /.x/; C / D 2Q.x;
N y;
N uN ; vN /.x/; for every x 2 U.x/;
N
and hence (13.17) follows from the above inclusion and (13.18). The proof is
complete. u
t
Example 13.2.2. Consider the map F defined in Example 11.7.26. Let C D R2C ,
xN D 1, and yN D .1; 1/. Since Q.x/ D F .x/ for x > 0, F is C -dominated by Q
N As in Example 11.7.26, for vN D .v1 ; v2 / and uN D v1 C v2 , we have
near x.
N y;
D 2 Q.x; N uN ; vN /.x/ D f.y1 ; y2 / j x D y1 C y2 C v1 v2 g;
D 2 .F C C /.x;
N y;
N uN ; vN /.x/ D f.y1 ; y2 / j x y1 C y2 C v1 v2 g;
so that both sides of (13.17) are equal. t

u
In our next result we derive analogues of (13.17). We recall that N y;
DS2 F .x; N uN ; vN /
stands for the second-order S derivative of F .
closed, and convex cone, and let F W X Y be a set-valued map. Let .x; N y/
N 2
graph .F / and let .Nu; vN / 2 X Y be arbitrary. Assume that one of the following
conditions holds.
N y;
(a) DS2 F .x; N uN ; vN /.0/ \ .C =f0g/ D ;.
N y/
(b) Q is second-order directionally compact at .x; N with respect to .Nu; vN / in any
direction x 2 X .
(c) Q is second-order compactly approximable at .x;N y/
N with respect to .Nu; vN /.
Then the following are valid:
N then there exists a neighborhood U.x/
(1) If F is dominated by Q near x, N of xN such
that for every x 2 U.x/,
N we have
N y;
Min.D 2F .x; N uN ; vN /.x/; C / D 2 Q.x;
N y;
N uN ; vN /.x/: (13.19)
(2) If F is CQ -dominated by R near x,

N then there exists a neighborhood U.x/
N of xN
such that for every x 2 U.x/,
N we have
N y;
WMin.D 2F .x; N uN ; vN /.x/; C / D 2 R.x;
N y;
N uN ; vN /.x/: (13.20)
(3) If F is C -bounded and C -closed, Y is finite dimensional, and F is dominated

by Q near x,N then there exists a neighborhood U.x/N of xN such that for every
x 2 U.x/,
N we have
N y;
GHe-PMin.D 2F .x; N uN ; vN /.x/; C / D 2 P .x;
N y;
N uN ; vN /.x/: (13.21)
Proof. We begin by deducing some simple implications of (a)-(c). Due to the chain
of inclusions
N y;
DS2 Q.x; N uN ; vN /.0/ DS2 R.x;
N y;
N uN ; vN /.0/ DS2 F .x;
N y;
N uN ; vN /.0/;
we obtain
N y;
DS2 F .x; N uN ; vN /.0/ \ .C =f0g/ D ; ) DS2 R.x;
N y;
N uN ; vN /.0/ \ .C =f0g/ D ;
) DS2 Q.x;
N y;
N uN ; vN /.0/ \ .C =f0g/ D ;:
Therefore, the following formula holds for D F; Q; R W
D 2 . C C /.x;
N y;
N uN ; vN /.x/
N y;
D D 2 .x; N uN ; vN /.x/ C C; for all x 2 dom.D 2 . C C /.x;
N y;
N uN ; vN //:
In fact, the above formula also remains valid under (b) and (c), because when Q is
either second-order directionally compact or second-order compactly approximable,
then so are the maps Q and R.
Since F is C -dominated by Q near x,N for each x 2 U.x/,
N we have
N y;
MinC D 2F .x; N uN ; vN /.x/ D Min.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C /
D Min.D 2 .Q C C /.x;
N y;
N uN ; vN /.x/; C /
N y;
D Min.D 2Q.x; N uN ; vN /.x/; C /
D 2Q.x;
N y;
N uN ; vN /.x/;
establishing (13.19).
Analogously, since F is CQ -dominated by R near x,
N for each x 2 U.x/,
N we have
WMin.D 2F .x; N uN ; vN /.x/; C / D WMin.D 2 .F C CQ /.x;

N y; N y;
N uN ; vN /.x/; C /
D WMin.D 2 .R C CQ /.x;
N y;
N uN ; vN /.x/; C /
N y;
D WMin.D 2 R.x; N uN ; vN /.x/; C /
D 2 R.x;
N y;
N uN ; vN /.x/;
proving (13.20).
Finally, for (13.21), we begin by noticing that D 2P .x; N y;
N uN ; vN /.x/
2
D Q.x;N y;
N uN ; vN /.x/. Moreover, due to a known relationship between minimal and
properly minimal points we also have that
Q.x/ cl.P .x//; for every x 2 U.x/;

N
which implies that D 2Q.x; N y;

N uN ; vN /.x/ D 2P .x;
N y;
N uN ; vN /.x/, or, D 2Q.x;
N y;
N uN ; vN /
.x/ D D P .x;
2
N y;
N uN ; vN /.x/.
Consequently
N y;
PMinC D 2F .x; N uN ; vN /.x/ D GHe-PMin.D 2 .F C C /.x;
N y;
N uN ; vN /.x/; C /
D GHe-PMin.D 2 .Q C C /.x;
N y;
N uN ; vN /.x/; C /
N y;
D GHe-PMin.D 2Q.x; N uN ; vN /.x/; C /
D GHe-PMin.D 2P .x;
N y;
N uN ; vN /.x/; C /
D 2P .x;
N y;
N uN ; vN /.x/;
confirming (13.21). The proof is complete. t

u
The following notion of normally minimal points will be used in the sequel.
Definition 13.2.4. Let Y be a normed space, let C Y be a proper, closed, and
convex cone, and let S be a nonempty subset of Y such that S C C is convex. A
point z 2 Min.S; C / is called a normally C -minimal point of S if
T .S C C; z/ int.C / [ f0g;
where C is the negative dual of C and T .S C C; z/ is the negative dual of T .S C

C; z/.
For our next result, we also need to recall that, given a normed space Z, a set S
Z is second-order derivable at zN in direction w 2 Z, if A2 .S; zN; w/ D T 2 .S; zN; w/.
Example 13.2.5. To illustrate these notions, let
S D f.x1 ; x2 / 2 R2C j x1 x2 D 1g:
Then, for zN D .1; 1/, we have
T .S; zN/ D f.z1 ; z2 / 2 R2 j z2 D z1 g;
and for w D .w1 ; w1 /,
T 2 .S; zN; w/ D A2 .S; zN; w/ D f.z1 ; z2 / 2 R2 j z2 D z1 C w1 2 g;
which means that S is second-order derivable at zN in direction w. For C D R2C ,
T .S C C; zN/ D f.z1 ; z2 / 2 R2 j z2
z1 g;
so that
T .S C C; zN/ D f.z1 ; z1 / j z1 0g;
and hence zN is a normally C -minimal point of S .

Theorem 13.2.6. Let X and Y be finite-dimensional normed spaces, let C Y be
a proper, pointed, closed, and convex cone, and let F W X Y be a set-valued map.
N y/
Let .x; N 2 graph .F / and let .Nu; vN / 2 X Y be arbitrary. We make the following
assumptions:
1. xN 2 int.dom.F // and graph .F C C / is convex.
N y/
2. graph .F / is second-order derivable at .x; N in direction .Nu; vN /.
3. yN is a normally C -minimal point of F .x/.
N
Then for every x 2 dom .D 2Q.x;
N y;
N uN ; vN //, the following inclusion holds:
N y;
D 2Q.x; N uN ; vN /.x/ Min.D 2F .x;
N y;
N uN ; vN /.x/; C /: (13.22)
Proof. Assume that y 2 D 2Q.x; N y; N uN ; vN /.x/ which implies that y 2 D 2F .x;N y;

N uN ; vN /
.x/. Now suppose that y 62 Min.D 2F .x; N y;
N uN ; vN /.x/; C /, and consequently
there exists yQ 2 D 2F .x; N y;
N uN ; vN /.x/ such that y yQ 2 C nf0g. Since yQ 2
N y;
D 2F .x; N uN ; vN /.x/, there are sequences fsn g P and f.xQ n ; yQn /g X Y such that
sn # 0, .xQ n ; yQn / ! .x; y/,
Q and
yN C sn vN C sn2 yQn 2 F .xN C sn uN C sn2 xQ n /: (13.23)

Analogously, because y 2 D 2Q.x; N y;

N uN ; vN /.x/, there are sequences ftn g P and
f.xn ; yn /g X Y such that tn # 0, .xn ; yn / ! .x; y/, and
yN C tn vN C tn2 yn 2 Q.xN C tn uN C tn2 xn / D Min.F .xN C tn uN C tn2 xn /; C /:
Therefore, .xN C tn uN C tn2 xn ; yN C tn vN C tn2 yn / is a boundary point of the convex

set graph .F C C /. By a standard separation argument, there are nonzero vectors
.
n ; n / 2 X Y such that
h
n ; xN C tn uN C tn2 xn i C hn ; yN C tn vN C tn2 yn i
O C hn ; yi

h
n ; xi O for all .x;
O y/
O 2 graph .F C C /: (13.24)
We normalize these vectors so that k.

n ; n /k D 1, and assume that .
n ; n / !
.
; /. Passing to the limit as n ! 1 in (13.24), we obtain
h
; xi
N C h; yi
N
h
; xi
O C h; yi
O O y/
for all .x; O 2 graph .F C C /: (13.25)
Let z 2 F .x/N C C be arbitrary. The map F C C , being convex, is lower-

semicontinuous at xN (see [565]). Therefore, there exists a sequence fzn g Y such
that zn ! z and zn 2 F .xN Ctn uN Ctn2 xn /CC . By setting .x;
O y/
O D .xN Ctn uN Ctn2 xn ; zn /
in (13.24), we obtain
h
h
n ; xN C tn uN C tn2 xn i C hn ; zn i:
Passing to the limit as n ! 1 in the above inequality, we obtain h; yi N

h; zi,
and because z 2 F .x/ N C C was chosen arbitrarily, the above inequality confirms
that 2 T .F .x/
N C C; y/N . Since yN is a normally C -minimal point, we deduce that

2 .int.C / [ f0g/. However, due to the assumption that xN 2 int.dom.F //, we
have ¤ 0, ensuring that 2 int.C /. This, in view of the fact that y yQ 2 C nf0g,
confirms that
h; yi < h; yi:
Q (13.26)
Due to the assumption that F is derivable, we can set tn D sn . Using (13.23)

in (13.24), we have
h
h
n ; xN C tn uN C tn2 xQ n i C hn ; yN C tn vN C tn2 yQn i;
which after simplifying yields
h
n ; xn i C hn ; yn i
h
n ; xQ n i C hn ; yQn i:
Passing to the limit as n ! 1 in the above inequality, we obtain

h; yi
h; yi,
Q which, however, is a contradiction to (13.26). Consequently
y 2 Min.D 2F .x;
N y;
N uN ; vN /.x/; C /. The proof is complete. t
u
By setting .Nu; vN / D .0; 0/, we obtain the following result.
Corollary 13.2.7. Let X and Y be finite-dimensional normed spaces, let C Y be
a proper, pointed, closed, and convex cone, let F W X Y be a set-valued map, and
N y/
let .x; N 2 graph .F /. Assume that xN 2 int.dom.F // and graph .F C C / is convex.
If yN is a normally C -minimal point of F .x/,
N then for every x 2 dom.DQ.x; N y//,
N
N y/.x/
DQ.x; N N y/.x/:
MinCDF .x; N
The following results show that, when dealing with weak-perturbation maps, we
may dispense with the requirement that yN is a normally C -minimal point of F .x/.
N
Theorem 13.2.8. Let X and Y be finite-dimensional normed spaces, let C Y be
a proper, pointed, closed, and convex cone, and let F W X Y be a set-valued map.
N y/
Let .x; N 2 graph .F / and let .Nu; vN / 2 X Y be arbitrary. We make the following
assumptions:
1. xN 2 int.dom .F // and graph .F C C / is convex.
N y/
2. graph .F / is derivable at .x; N in direction .Nu; vN /.
3. yN is a normally C -minimal point of F .x/.
N
Then
N y;
D 2 R.x; N uN ; vN /.x/ WMin.D 2F .x;
N y;
N uN ; vN /.x/; C /: (13.27)
Proof. Assume that y 2 D 2 R.x; N y;

N uN ; vN /.x/ but y 62 WMin.D 2F .x; N y;
N uN ; vN /.x/; C /.
Therefore, there exists yQ 2 D F .x;
2
N y;N uN ; vN /.x/ such that y yQ 2 int.C /. As in the
proof of (13.25), we prove that
h
; xi
N C h; yi
N
h
; xi
O C h; yi
O for all .x;
O y/
O 2 graph .F C C /;
which, as above, will confirm that 2 C nf0g. However, now we have y yQ 2

int.C /, which implies h; y yi Q < 0. (In fact, the normality assumption was
required for this inequality). By the same approach used in the proof of (13.22),
we obtain a contradiction to h; y yi
Q < 0. The proof is complete. t
u
The following result uses the second-order lower Dini derivative and the notion
of semi-differentiability:
closed, and convex cone, and let F W X Y be a set-valued map. Let .x;N y/
N 2
graph .F / and let .Nu; vN / 2 X Y be arbitrary. Assume that F is second-order

N y/
semidifferentiable at .x; N in direction .Nu; vN /. Then (13.27) holds.
Proof. Let y 2 D 2 R.x; N y;
N uN ; vN /.x/. Then there exist sequences ftn g P,
f.xn ; yn /g X Y such that tn # 0; .xn ; yn / ! .x; y/ and
an WD yN C tn vN C tn2 yn 2 R.xN C tn uN C tn2 xn /: (13.28)
Assume that y 62 WMinC D 2F .x; N y;

N uN ; vN /.x/. Then there exists yQ 2 D 2 F .x;
N y;
N uN ; vN /
.x/ D D` F .x;
2
N y;
N uN ; vN /.x/ such that y yQ 2 int.C /. Since tn # 0 and xn ! x, there
exist n0 2 N and yQn ! yQ such that
bn WD yN C tn vN C tn2 yQn 2 F .xN C tn uN C tn2 xn / for n > n0 :
Since y yQ 2 int.C /, we have yn yen 2 int.C / for sufficiently large n. Moreover,

because C is a cone, we obtain that
an bn D .yN C tn vN C tn2 yn / .yN C tn vN C tn2 yQn / D tn2 .yn yen / 2 int.C /;
which, however, is a contradiction to (13.28). The proof is complete. t

u
Remark 13.2.10. The mapping defined in Example 3.2 is second-order semi-
differentiable at .1; .1; 1// in the direction .v1 C v2 ; .v1 ; v2 //.
For our next result, we consider the following parameter dependent vector-
optimization problem
Min.f .w; x/; C / subject to w 2 H.x/; (13.29)
where H is a given set-valued map, and confine our discussion to the case when the
set-valued map F is given by
F .x/ WD fy 2 Y j y D f .w; x/ for some w 2 H.x/g; (13.30)
where X is the parameter space, f W W X ! Y is a single-valued map, and

H W X W is a set-valued map.
For this situation, we have the following result.
Theorem 13.2.11. Let X and Y be finite-dimensional normed spaces, let C Y
be a proper, pointed, closed, and convex cone, and let F W X Y be a set-valued
N y/
map. Let .x; N 2 graph .F / and let .Nu; vN / 2 X Y be arbitrary. We make the
following assumptions:
1. f is continuous and C -convex.
2. H is convex.
N
3. F is C -dominated by Q near x.
Then
N y;
Min.D 2F .x; N uN ; vN /.x/; C / D 2Q.x;
N y;
N uN ; vN /.x/:
Proof. For any y 2 Min.D 2F .x; N y;

N uN ; vN /.x/; C /, we have y 2 D 2F .x;
N y;
N uN ; vN /.x/,
and hence there are sequences ftn g P and f.xn ; yn /g X Y such that tn # 0,
.xn ; yn / ! .x; y/, and yN Ctn vN Ctn2 yn 2 F .xN Ctn uN Ctn2 xn /. Since F is C -dominated
by Q near x, N for sufficiently large n, we have yCt N n vN Ctn2 yn 2 Q.xCt
N n uN Ctn2 xn /CC .
Consequently, there exists a sequence fyQn g Y such that yN C tn vN C tn2 yQn 2 Q.xN C
tn uN C tn2 xn / and
yn yQn 2 C: (13.31)
Assume that the sequence fyQn g is bounded. Since Y is finite dimensional, we may
assume that yQn ! yQ which implies yQ 2 D 2Q.x; N y;
N uN ; vN /.x/ D 2F .x;
N y;
N uN ; vN /.x/.
In view of (13.31), we have yyQ 2 C , and because y 2 MinC D 2F .x; N y;
N uN ; vN /.x/,
we must have y D y. Q This implies that y 2 D 2Q.x; N y;N uN ; vN /.x/.
Therefore, to complete the proof it suffices to show that fyQn g remains bounded.
If possible, assume that this is not the case and kyQn k ! 1. In view of (13.30), there
are sequences fwn g and fw Q n g such that
wN C tn uN C tn2 wn 2 H.xN C tn uN C tn2 xn /

yN C tn vN C tn2 yn D f .wN C tn uN C tn2 wn ; xN C tn uN C tn2 xn /
wN C tn uN C tn2 wN n 2 H.xN C tn uN C tn2 xn /
yN C tn vN C tn2 yQn D f .wN C tn uN C tn2 wQ n ; xN C tn uN C tn2 xn /:
Due to the convexity of H , for some 0 ˛ 1, we have
˛.wN C tn uN C tn2 wn / C .1 ˛/.wN C tn uN C tn2 wQ n / D wN C tn uN C tn2 .˛wn C .1˛/wQ n /

2 H.xN C tn uN C tn2 xn /:
Moreover, due to the assumption that f is C -convex, we obtain
yN C tn vN C tn2 yn .˛/ WD f .wN C tn uN C tn2 .˛wn C .1 ˛/wQ n /; xN C tn uN C tn2 xn /

C yN C tn vN C tn2 .˛yn C .1 ˛/yQn /
C yN C tn vN C tn2 yn ;
where we used the fact that
.yN C tn vN C tn2 yn / .yN C tn vN C tn2 .˛yn C .1 ˛/yQn // D tn2 .1 ˛/.yn yQn / 2 C:

Since f is continuous, we have
yN C tn vN C tn2 yn .˛/ ! yN C tn vN C tn2 yQn as ˛ ! 0

yN C tn vN C tn2 yn .˛/ ! yN C tn vN C tn2 yn as ˛ ! 1:
Since kyQn k ! 1 and yn ! y, by taking ˛n appropriately close to 1, we have
"tn2 kyN C tn vN C tn2 yn .yN C tn vN C tn2 yn .˛n //k tn2 ;
where " 2 .0; 1/. Setting yn .˛n / D zn we see that " kyn zn k 1.
Because yn ! y, we deduce that .zn / is bounded. Assume that zn ! z, which
implies z 2 D 2F .x;
N y;
N uN ; vN /.x/. Since " kyn zn k, we have " ky zk, and
hence y ¤ z.
From the inequality
yN C tn vN C tn2 zn C yN C tn vN C tn2 yn ;
we infer that yn zn 2 C , which implies y z 2 C . This, however, is a contradiction

to the fact that y 2 MinC D 2F .x;
N y;
N uN ; vN /.x/. Therefore, fyQn g has to be a bounded.
u
Example 13.2.12. As an illustration of Theorem 13.2.11, let X D W D .0; C1/,
Y D R2 , C D R2C , and define H W X W by H.x/ D Œx; C1/ and f W
W X ! Y by f .w; x/ D .w; 1=w C x 2 /. In this example, the hypotheses of
Theorem 13.2.11 are satisfied for any xN > 0. One can also see directly that the
conclusion of Theorem 4.7 holds, since for each x > 0, we have
Q.x/ D F .x/ D f.w; 1=w C x 2 / j w

xg:
To see what DF and D 2 F look like in this example, consider the case where xN D 1
and yN D .2; 3=2/. Then
DF.1; .2; 3=2//.x/ D f.y; 2x y=4/ j y 2 Rg
and
D 2 F .1; .2; 3=2/; v1; .v2 ; 2v1 v2 =4//.x/ D f.y; 2x y=4 C v12 C v2 2 =8/ j y 2 Rg:
t
u
13.3 Sensitivity Analysis in Set-Valued Optimization Using Coderivatives 623
13.3 Sensitivity Analysis in Set-Valued Optimization Using

Coderivatives
In the previous sections, the notion of the contingent derivative was used to study
the sensitivity properties of set-valued optimization problems. In this section, our
objective is show that various coderivatives can also be used as the derivative
concept to investigate the sensitivity properties of set-valued optimization problems.
Numerous authors have used coderivatives in parametric optimization. One of
the earlier works is the useful contribution by Levy and Mordukhovich [376],
where the authors considered parametric families of constrained problems in
mathematical programming and conducted a local sensitivity analysis for set-valued
solution maps. More akin to this section is the work by Huy, Mordukhovich,
and Yao [272] where the authors obtained formulas for estimating and computing
the normal and mixed coderivatives of frontier and solution maps in parametric
multiobjective optimization. Yao and Yen [604] employed coderivatives for some
calculations related to a parametric affine variational inequality. Recently, Chung
and Yao [98] used coderivatives of minimal point set-valued map in parametric
vector optimization. In an interesting paper, Nam [438] provided precise formulae
in terms of the initial data for the computation of the Fréchet and limiting normal
cones to the normal cone mapping to a given convex polyhedron when the set of
generating elements corresponding to the active constraints is linearly independent.
Based on those results, he gave an explicit representation of the coderivatives of the
normal cone mapping and computed the coderivative of solution maps to parametric
variational inequalities. Our treatment of the subject in the following is based on the
results derived by Chuong [97] where the author provides estimates for the Clarke
coderivative of the efficient set map associated with a parametric vector optimization
problem. These results are analogous to some of the results earlier obtained in [272].
We recall the setting of the model problem. Let P , X , and Y be Banach spaces,
let f W P X ! Y be a given single-valued map, and let M W P X be a given
set-valued map. Let C Y be a pointed, closed, and convex cone inducing a partial
ordering in Y . We consider the following parametric vector optimization problem
min ff .p; x/j x 2 M.p/g ; (13.32)

C
depending on the parameter p 2 P .

We define the associated set-valued map F W P Y by
F .p/ D .f ı M /.p/ WD ff .p; M.p//g D ff .p; x/j x 2 M.p/g :
Our objective is to investigate the differentiability properties of the following

set-valued perturbation map Q W P Y associated to the vector optimization
problem (13.32):
Q.p/ D Min.F .p/; C /; p 2 P: (13.33)

We begin by formulating and recalling a few concepts.

Definition 13.3.1. Let Y be a Banach space and Y be the topological dual of Y .
Let C be a closed, convex, and pointed cone. The uniformly positive polar to
the cone C is defined by
C
CUP WD fy 2 Y j 9 ˇ > 0; hy ; ki
ˇkkk; 8 k 2 C g:
We recall that given Banach spaces X and Y , a single-valued map g W X ! Y is

called strictly differentiable at xN if there exists a linear, continuous map Dg.x/
N W
X ! Y such that
g.x/ g.z/ hDg.x/;

N x zi
lim D 0:
x;z!xN kx zk
We also recall that given Banach spaces X and Y and S X , a single-valued

map L W S ! Y is said to be local upper-Lipschitzian at xN 2 S if there are
numbers > 0 and `
0 such that
kL.x/ L.x/k
N `kx xk
N for every x 2 B .x/
N \ S:
Definition 13.3.2. Let X and Y be Banach spaces, let L W D Y with D D

dom .L/ be a set-valued map, and let .x;N y/
N 2 graph .L/. The map L is said to
N y/
admit a local upper-Lipschitzian selection at .x; N if there exists a single-valued
mapping L W D ! Y which is local upper-Lipschitzian at xN satisfying L.x/ N D yN
and L.x/ 2 L.x/, for every x 2 D.
We recall that given a Banach space, a set S X , and xN 2 S , the Clarke normal
cone, the set of Fréchet "-normals, the Fréchet normal cone, the Mordukhovich
normal cone, are given by:
NC .S; x/ N D fx 2 X j hx ; xi 0; x 2 C.S; x/g:

N WD C.S; x/ N
8 9
< hx ; x xi
N =
NO " .S; x/
N WD x 2 X j lim sup " :
: S kx xk N ;
x !xN
8 9
< hx
; x N
xi =
NO .S; x/
N WD x 2 X j lim sup 0 :
: S kx xk N ;
x !xN
N WD lim sup NO " .S; x/:

N.S; x/ N
S
x !x;
N "#0
We also recall that the following chain of inclusions is known to hold:
NO .S; y/
N N.S; y/
N NC .S; y/:
N (13.34)
The set S is called Fréchet normally regular at xN if
NO .S; x/
N D NC .S; x/:
N
Finally, given Banach spaces X and Y with X and Y their duals, a set-valued
map F W X Y with .x; N y/
N 2 graph .F /, the Clarke coderivative and the
Mordukhovich coderivative are given by:
DC F .x; N / WD fx 2 X j .x ; y / 2 NC .graph .F /; .x;

N y/.y N y/g;
N
D F .x; N / WD fx 2 X j .x ; y / 2 N.graph .F /; .x;
N y/.y N y/g:
N
It follows from above definition that for every y 2 Y , the following inclusion
holds:
D F .p; N / DC F .p;

N y/.y N /:
N y/.y (13.35)
We begin with the following auxiliary result:

Theorem 13.3.3. Let P and Y be Banach spaces with P and Y their topological
duals, let G W P Y be a set-valued, and let .p;
N y/
N 2 graph .G/. Let C Y be a
pointed, closed, and convex cone.
1. Assume that .G C C / is tangentially regular at .p; N Then for every y 2 Y ,
N y/.
DC .G C C /.p; N / DC G.p;

N y/.y N /:
N y/.y (13.36)
2. Assume that the following conditions hold:

a) The map G is Fréchet normally regular at .p; N y/.
N
b) The map H W P Y Y given by H.p; y/ WD G.p/ \ .y C / admits a
N y;
local upper-Lipschitz selection at .p; N y/.
N
Then for every y 2 CUP
C
, the following inclusion holds:
DC .G C C /.p; N / DC G.p;

N y/.y N /:
N y/.y (13.37)
Proof. By using the tangential regularity of G C C , we have
N y//
C.graph .G/; .p; N T .graph .G/; .p;
N y//
N T .graph .G C C /; .p;
N y//
N
D C.graph .G C C /; .p;
N y//;
N
which implies NC .graph .G C C /; .p; N y//

N NC .graph .G/; .p; N y//,
N and conse-
quently, for every y 2 Y , we have DC .G C C /.p; N / DC G.p;
N y/.y N /,
N y/.y
which completes the proof of the first part.
For the second part, we fix an element y 2 CUP C
and choose an arbitrary

p 2 DC G.p; N y/.y
N /. To obtain a contradiction, we begin by assuming that p …
DC .G C C /.p; N /. By (13.35), we deduce that p … D .G C C /.p;

N y/.y N /,
N y/.y
which implies that .p ; y / … NO .graph .G C C /; .p;

N y//.
N
Therefore, by using the definition of the Fréchet normal cone, there exists a
sequence f.pn ; yn /g graph .G C C / with .pn ; yn / ! .p;
N y/
N such that
h.p ; y /; .pn ; yn / .p;

N y/i
N
lim sup > 0: (13.38)
n!1 k.pn ; yn / .p;N y/k
N
From the assumptions that the map H admits a local upper-Lipschitzian selection
N y;
at .p; N y/
N and that dom .H / D graph .G C C /, there are a number ` > 0 and
neighborhoods U V 2 U.p; N y/N such that for each .p; y/ 2 .U V / \ graph .G C
C /, we can find yO 2 H.p; y/ such that
kyO yk
N `k.p; y/ .y;
N y/k:
N
Due to the convergence .pn ; yn / ! .p; N y/,

N there exists m 2 N such that
.pn ; yn / 2 U V , for every n
m, and hence, for every n
m, there exists
yOn 2 H.pn ; yn / D G.pn / \ .yn C / such that
kyOn yk
N `k.pn ; yn / .p;
N y/k:
N
Using the definition of the map H , for each n

m, there exists yOn 2 G.pn / and
graph .G/
kn 2 C such that yOn D yn kn which implies that there exists .pn ; yOn / !
N y/
.p; N satisfying that
k.pn ; yOn / .p;

N y/k
N .` C 1/k.pn ; yn / .p;
N y/k
N for every n
m:
Since y 2 CUP
C
, there exists ˇ > 0 such that hy ; ki
ˇkkk
0 for every
k 2 C , and consequently, for each n
m, we have
h.p ; y /; .pn ; yOn / .p; N D h.p ; y /; .pn p;

N y/i N yn kn y/i
N
D h.p ; y /; .pn p; N C hy ; kn i
N yn y/i

h.p ; y /; .pn p;
N yn y/i;
N
which, by using (13.38), yields
h.p ; y /; .pn ; yOn / .p;

N y/i
N h.p ; y /; .pn p;N yn y/i
N
lim sup
lim sup
n!1 k.pn ; yOn / .p;N y/k
N n!1 k.pn ; yOn / .p;
N y/k
N
h.p ; y /; .pn p;
N yn y/i
N

lim sup
n!1 .` C 1/.k.pn ; yn / .p;
N y/k/
N
> 0:
Therefore,
h.p ; y /; .p; y/ .p;

N y/i
N
lim sup > 0;
graph .G/ k.p; y/ .p; N y/k
N
.p;y/ ! .p;
N y/
N
which implies that .p ; y / … NO .graph .G/; .p;

N y//,
N and by the fact that G
is Fréchet normally regular, we deduce that p … DC G.p; N /, which is a
N y/.y
contradiction. This completes the proof. t
u
Remark 13.3.4. Bao and Mordukhovich [26] proved an analogue of (13.36) for the
Mordukhovich coderivative and an analogue of (13.37) for the mixed coderivative.
The following example shows that (13.37) does not hold without additional
assumptions.
Example 13.3.5. For P D Y D R and C D RC , we define a set-valued map
G W RR by

f0; 2g if p
0;
G.p/ D
; otherwise.
We choose .p; N D .0; 2/ 2 graph .G/ and y D 2. By a simple calculation,

N y/
we obtain that
NO .graph .G/; .0; 2// D NC .graph .G/; .0; 2// D f.x; y/ 2 R2 j x 0; y 2 Rg;
and
DC .G C C /.0; 2/.2/ D ;;

DC G.0; 2/.2/ D RC ;
showing that DC .G C C /.p; N / 6 DC G.p;

N y/.y N /.
N y/.y
Remark 13.3.6. We note that for the data defined in Example 13.3.5, the map H
does not admit a local upper Lipschitzian selection at .0; 2; 2/. Also note that in this
example, the map G is Fréchet normally regular at .0; 2/ and y 22 CUP C
.
The following example shows that the Fréchet normal regularity is essential
for (13.37).
Example 13.3.7. We set P D Y D R and C D RC . We define a map G W RR
by G.p/ WD fjpjg.
We choose .p; N D .0; 0/ 2 graph .G/ and y D 1 2 CUP
N y/ C
. By a simple
calculation, we obtain that
NO .graph .G/; .0; 0// D f.x; y/ 2 R2 j y x y; y 0g;

NC .graph .G/; .0; 0// D R2 ;
and
DC .G C C /.0; 0/.1/ D Œ1; 1;

DC G.0; 0/.1/ D R;
showing that DC .G C C /.p; N / 6 DC G.p;

N y/.y N /.
N y/.y
In the following result, we give an estimate for the Clarke coderivative of the
composition of a single-valued map and a set-valued map.
Theorem 13.3.8. Let P , X and Y be Banach spaces, let f W P X ! Y be
a single-valued map, and let G W P X be a set-valued map. Let pN 2 P and
yN 2 .f ıG/.p/.
N For xN 2 G.p/
N with .p; N 2 f 1 .y/,
N x/ N let f be Fréchet differentiable
N x/
at .p; N with rf .p;
N x/N WD .rp f .p;
N x/;
N rx f .p;
N x//
N as the derivative.
1. Assume that the composite map f ı G is tangentially regular at .p;
N y/.
N Then for
every y 2 Y ,
DC .f ı G/.p; N / rp f .p;

N y/.y N .y / C DC G.p;
N x/ N x/.r
N N .y //:
N x/
x f .p;
(13.39)
2. Assume that the following conditions hold:

a) The map G is Fréchet normally regular at .p; N x/.
N
b) The map GQ W P Y X given by G.p; Q y/ WD fx 2 G.p/j y D f .p; x/g
admits a local upper Lipschitzian selection at .p;N y;
N x/.
N
Then for every y 2 Y , the following inclusion remains valid:
DC .f ı G/.p; N / rp f .p;

N y/.y N .y / C DC G.p;
N x/ N x/.r
N N .y //:
N x/
x f .p;
(13.40)
Proof. We claim that for every p 2 P , the following inclusion holds:
f.p; rf .p;
N x//.p;
N x/j.p; x/ 2 C.graph .G/; .p;
N x//g
N C.graph .f ıG/; .p;
N y//:
N
(13.41)
To verify our claim, for p 2 P , we pick .p; x/ 2 C.graph .G/; .p; N x//.
N Then
.p; x/ 2 T .graph .G/; .p; N x//,
N and hence there are sequences ftn g P and
f.pn ; xn /g P X with tn # 0 and .pn ; xn / ! .p; x/ such that xN C tn xn 2
G.pN Ctn pn / for every n 2 N. Therefore, f .pN Ctn pn ; xN Ctn xn / 2 .f ıG/.pN Ctn pn /
for every n 2 N, and by rearranging the above containment, for every n 2 N, we get
f .pN C tn pn ; xN C tn xn / f .p;
N x/
N
yN C tn 2 .f ı G/.pN C tn pn /:
tn
N x/,
In view of the Fréchet differentiability of the map f at .p; N we get
f .pN C tn pn ; xN C tn xn / f .p;
N x/
N
lim D rf .p;
N x/.p;
N x/;
n!1 tn
and hence
.p; rf .p;
N x/.p;
N x// 2 T .graph .f ı G/; .p;
N y//
N D C.graph .f ı G/; .p;
N y//;
N
(13.42)
where the last equality is due to the tangential regularity of f ı G at .p;

N y/.
N This
proves (13.41).
Since for each .p; x/ 2 C.graph .G/; .p; N and .u ; rx f .p;
N x// N y/ 2
N x/
N x//,
NC .graph .G/; .p; N
N x/
h.rp f .p; N y C u ; y /; .p; rf .p;
N x/.p;
N x//i D u .p/ rx f .p; N y .x/;
N x/
we deduce that for all .p; x/ 2 C.graph .G/; .p; N and .u ; rx f .p;
N x// N y/ 2
N x/
N x//,
NC .graph .G/; .p; N
N y C u ; y /; .p; rf .p;
N x//
h.rx f .p; N x/.p;
N x//i 0: (13.43)
Consequently, by using (13.41), we get
NC .graph .f ı G/; .p;

N y//
N f.rp f .p; N y C u ; y /j .u ; rx f .p;
N x/ N y/
N x/
2 NC .graph .G/; .p;
N x//g:
N
For y 2 Y , let p 2 DC .f ı G/.p; N / be arbitrary. Then,

N y/.y

.p ; y / 2 NC .graph .f ı G/; .x; N y//,
N and in view of the above inclusion,
there exists .u ; rx f .p; N y / 2 NC .graph .G/; .p;
N x/ N such that .p ; y / D
N x//

N x/
.rp f .p; N y C u ; y /, which gives
.p rp f .p; N y ; rx f .p;

N x/ N y / 2 NC .graph .G/; .p;
N x/ N x//
N
and consequently, we have p 2 rp f .p; N y C DC G.p;

N x/ N x/.r
N N y /.
N x/
x f .p;

Therefore, we have shown that p 2 rp f .p;N x/
N y CDC G.p; N x/.r
N N y /,
N x/
x f .p;
and the proof of (13.39) is complete.
For the converse, let u be a fixed element such that u … DC .f ı G/.p; N /.
N y/.y
We will show that
u … rp f .p; N y C DC G.p;
N x/ N x/.r
N N y /:
N x/
x f .p; (13.44)
First note that u … DC .f ı G/.p; N / implies .u ; y / … NO .graph .f ı

N y/.y
N y//.
G/; .p; N Thus,
h.u ; y /; .p; y/ .p;

N y/i
N
lim sup > 0;
graph .f ıG/ k.p; y/ .p; N y/k
N
.p;y/ ! N y/
.p; N
implying that there exist f.pn ; yn /g graph .f ıG/ and ˛ > 0 such that .pn ; yn / !
N y/
.p; N and
hu ; pn pi
N
hy ; yn yi
N C ˛.kpn pk
N C kyn yk/;
N for sufficiently large n 2 N:
(13.45)
Since the map GQ admits a local upper Lipschitzian selection at .p;N y;

N x/,
N there
exists a function ` W dom .G/O ! X which is local upper Lipschitzian at .p; N y/
N
and satisfies that `.p;N y/
N D xN and `.p; y/ 2 G.p;O O
y/ for all .p; y/ 2 dom .G/
N y/.
sufficiently close to .p; N Therefore, there exists L > 0 such that
kxn xk
N L.kpn pk
N C kyn yk/
N (13.46)
for sufficiently large n, where xn WD `.pn ; yn / 2 G.p Q n ; yn /. For such n, since

Q
xn 2 G.pn ; yn /, xn 2 G.pn /, and yn D f .pn ; xn /, it follows from (13.45) that
hu ; pn pi
N
hy ; f .pn ; xn /f .p;
N x/iC˛.kp
N n pk
N C kf .pn ; xn /f .p;
N x/k/
N
D hy ; rf .p;
N x/.p
N n p;
N xn x/i
N C o.kpn pk
N C kxn xk/
N C ˛.kpn pk
N
C kf .pn ; xn / f .p;
N x/k/
N
D hrf .p; N y ; .pn p;
N x/ N xn x/i
N C o.kpn pk
N C kxn xk/
N
C ˛.kpn pk
N C kf .pn ; xn / f .p;
N x/k/;
N
and by using (13.46), we get that
hu ; pn pi
N
hrf .p; N y ; .pn p;
N x/ N xn x/i
N C o.kpn pk
N C kxn xk/
N
˛ ˛
C kpn pk N C kxn xk
N
2 2L

hrf .p; N y ; .pn p;
N x/ N xn x/i
N C o.kpn pk
N C kxn xk/
N
C ˛O .kpn pk
N C kxn xk/
N ;
where ˛O D minf˛=2; ˛=.2L/g. Therefore,
hu rx f .p; N y ; p pi
N x/ N hrx f .p; N y ; x xi
N x/ N
lim sup
˛;
O
graph .G/ kp pk N C kx xk
N
.p;x/ ! .p;
N x/
N
which implies that .u rp f .p; N .y /; rx f .p;

N x/ N y /…NO .graph .G/; .p;
N x/ N x//.
N
Then, by using the Fréchet normal regularity of G at .p; N we have u
N x/,
rp f .p; N .y / … DC G.p;
N x/ N x/.r
N x f .p; N y /, and hence (13.44) is verified.
N x/
u
The following example shows that the imposed conditions cannot be removed:
Example 13.3.9. We set P D Y D R, C D RC , and f .p; x/ WD x 2 . We define a
set-valued map:
p p
f p; pg if p
0;
G.p/ WD
; otherwise:
Then,
˚
graph .G/ D .p; x/ 2 R2 j x 2 D p ;
graph .f ı G/ D f.p; y/ 2 R2 j p
0; y D pg:
We choose pN D xN D yN D 0 and y D 0. By a simple calculation, we obtain that
NO .graph .G/; .0; 0// D NC .graph .G/; .0; 0// D f.x; 0/ 2 R2 j x 2 Rg;
and
DC .f ı G/.0; 0/.0/ D RC ;

DC G.0; 0/.0/ D R;
proving that the inclusion (13.40) fails to hold. Note that the set-valued map GQ does
not admit any local upper Lipschitzian selection at .0; 0; 0/.
We will now give some coderivative estimates for the perturbation map Q defined
in (13.33). For this, we define two set-valued maps T W P Y Y and MQ W
P Y X as follows:
T .p; y/ WD Q.p/ \ .y C /
MQ .p; y/ WD fx 2 M.p/j y D f .p; x/g:
Theorem 13.3.10. Let pN 2 P and xN 2 M.p/ N be such that yN WD f .p;N x/

N 2
N Let f be Fréchet differentiable at .p;
Q.p/. N x/
N with the derivative rf .p;
N x/
N WD
N x/;
.rp f .p; N rx f .p;
N x//.
N Assume that the domination property holds for F around
pN and T admits a local upper Lipschitzian selection at .p;
N y;
N y/.
N
1. If F and F C C are tangentially regular at .p;
N y/
N and Q is Fréchet normally
N then for every y 2 CUP
N y/,
regular at .p; C
,
DC Q.p; N / rp f .p;

N y/.y N y C DC M.p;
N x/ N x/.r
N N y /:
N x/
x f .p; (13.47)
2. If Q C C is tangentially regular at .p;

N y/,
N F is Fréchet regular at .p;
N y/,
N and
N x/
M is Frećhet normal at .p; N and admits a local upper Lipschitzian selection
N y;
.p; N then for every y 2 CUP
N x/, C
,
DC Q.p; N / rp f .p;

N x/ N x/.r
N N y /:
N x/
x f .p; (13.48)
Proof. Since for every p 2 P , the inclusion Q.p/ F .p/ holds and since the
domination property holds for F around p, N there exists a neighborhood U of pN
such that
Q.p/ C C D F .p/ C C; for every p 2 U;
and consequently, for every y 2 Y , we have
DC .Q C C /.p; N / D DC .F C C /.p;

N y/.y N /:
N y/.y (13.49)
Since the map Q is Fréchet normally regular at .p; N x/

N and the map T admits a
N y;
local upper Lipschitzian selection at .p; N y/,
N it follows from (13.37) that for every
y 2 CUP
C
, we have
DC Q.p; N / DC .Q C C /.p;

N y/.y N /;
N y/.y (13.50)
and, in view of the tangential regularity, we also have for every y 2 Y ,
DC .F C C /.p; N / DC F .p;

N y/.y N /:
N y/.y (13.51)
Furthermore, using (13.39), for every y 2 Y , we get
DC F .p; N / rp f .p;

N x/ N x/.r
N N y /:
N x/
x f .p; (13.52)
By combining (13.49) thorough (13.52) we finish the proof of the inclu-

sion (13.47).
For (13.48), we begin by noticing that since the map QCC is tangentially regular
N it follows from Theorem 13.3.3 that for every y 2 Y , we have
N y/,
at .p;
DC .Q C C /.p; N / DC Q.p;

N y/.y N /:
N y/.y (13.53)
Define the set-valued map H W P Y by
H.p; y/ WD F .p/ \ .y C /;
and note that T .p; y/ H.p; y/ for every .p; y/ 2 P Y , dom .T / D

graph .Q C C /, and dom .H / D graph .F C C /. Therefore, the existence of a
local upper Lipschitzian selection of T at .p;N y;

N y/
N implies the existence of local
upper Lipschitzian selection of H at this point. Therefore, by the Fréchet normal
N for every y 2 CUP
N y/,
regularity of F at .p; C
, we get
DC .F C C /.p; N / DC F .p;

N y/.y N /:
N y/.y (13.54)
N and because MQ admits a local

N x/
Because M is Fréchet normally regular at .p;
N y;
upper Lipschitzian selection at .p; N x/,
N it follows from Theorem 13.3.3 that for
every y 2 Y , we have
DC F .p; N / rp f .p;

N x/ N x/.r
N N y /:
N x/
x f .p; (13.55)
Combining (13.49), (13.53), (13.54), and (13.55) completes the proof. t

u
The following examples depicts the importance of the domination property.
Example 13.3.11. We set P D Y D R and C D RC . We define f .p; x/ WD x and

x 2 Rj x
0; if p D 0;
M.p/ WD
x 2 Rj x > jpj; otherwise.
Then,
[
graph .M / D f.p; x/ 2 R2 j x > 0; 0 x < p < xg f.0; 0/g;
F .p/ D F .p/ C C D M.p/ for every p 2 P:
We choose pN D xN D 0 and y D 1 2 CUPC

. By a simple calculation, we obtain
that

f0g; if p D 0;
Q.p/ WD
;; otherwise,
confirming that yN WD f .p; x/

N D 0 2 Q.p/.
N We compute
DC Q.0; 0/.1/ D .1; 1/

rp f .0; 0/ .1/ C DC M.0; 0/.1/ D Œ1; 1;
which shows that (13.47) does not hold.

Notice that the set-valued maps F and .F C C / are tangentially regular at .0; 0/
and the map Q is Fréchet normally regular at .0; 0/. However, the domination
property does not hold at F around pN D 0. t
u
Analogues of the above formulas for special constraints can be found in [98, 99,
272] and the cited references therein.
13.4 Sensitivity Analysis for Vector Variational Inequalities
Vector variational inequalities have been of great interest in the academic and
professional communities ever since the path-breaking paper by Giannessi [199].
The interested reader can find some of the recent developments on this useful field
in [4, 96, 118, 122, 177, 200, 203, 204, 377, 416, 424, 478] and the cited references
therein. Vector variational inequalities provide an elegant and useful extension of
variational inequalities. We remark that in recent years the theory of variational
inequalities emerged as one of the most promising branches of pure, applied, and
industrial mathematics, and provides us with a convenient mathematical apparatus
for studying a wide range of problems arising in diverse fields such as structural
mechanics, elasticity, economics, optimization, optimal control, inverse problems,
financial mathematics, and others (see [30, 33, 121, 139, 201, 202, 221, 238, 289, 320,
327, 331, 508]).
This section is devoted to differentiability and sensitivity properties of gap
functions associated with vector-variational inequalities. Many authors have focused
on studying sensitivity analysis of vector variational inequalities. For instance,
Li, Yan, Chen [383] conducted a detailed study of differential and sensitivity
properties of gap functions associated with vector variational inequalities and also
gave optimality conditions for solutions of vector variational inequalities. This study
was further enhanced by Meng and Li [418] who investigated the differentiability
and sensitivity properties of gap functions associated with Minty vector-variational
inequalities. Recently, Li and Li [378] presented the second-order differentiability
and sensitivity properties of the gap functions for vector variational inequalities. Our
treatment of the subject in this section follows the approach of [378, 383].
We begin by recalling the following notions of vector variational inequalities:
Definition 13.4.1. Let C Rm be a pointed, closed, and convex cone, and let
K Rn be nonempty closed set. Let F W K ! L.Rn ; Rm / be a single-valued
map.
1. The vector variational inequality seeks xN 2 K such that
hF .x/;
N x xi
N … C nf0g for every x 2 K: (13.56)
2. Given that C is solid, the weak vector variational inequality seeks xN 2 K such
that
hF .x/;
N x xi
N … int.C / for every x 2 K: (13.57)
The following definition recalls the notion of the gap functions associated to the
above vector variational inequalities. See Chen, Goh, and Yang [90] and Meng and
Li [418] for details.
13.4 Sensitivity Analysis for Vector Variational Inequalities 635
Definition 13.4.2. Let C Rm be a pointed, closed, and convex cone, and let
K Rn be nonempty closed set. Let F W K ! L.Rn ; Rm / be a given single-valued
map.
1. A set-valued map N W Rn Rm is said to be a gap function for vector variational
inequality (13.56) if and only if the following conditions hold:
a) 0 2 N.x/
N if and only if xN solves vector variational inequality (13.56).
b) N.x/ \ .C / D f0g for every x 2 K.
2. Given that the cone C is also solid, a set-valued map W W Rn Rm is said to be
a gap function for weak vector variational inequality (13.57) if and only if the
following conditions hold:
a) 0 2 W .x/
N if and only if xN solves weak vector variational inequality (13.57).
b) W .x/ \ .int.C // D f;g for every x 2 K.
Given a compact set K Rn and a single-valued map F W K ! L.Rn ; Rm /, we
introduce
[
G.x/ WD fhF .x/; x zig D hF .x/; x Ki: (13.58)
z2K
Furthermore, using the above map G, we introduce the following set-valued

maps:
N.x/ D Max .G.x/; C / ; (13.59)

W .x/ D WMax .G.x/; C / : (13.60)
We have the following result regarding the above set-valued maps:

Theorem 13.4.3. Let C Rm be a pointed, closed, and convex cone, and let K
Rn be nonempty closed set. Let F W K ! L.Rn ; Rm / be a single-valued map.
Then:
1. The set-valued map N is a gap function for vector variational inequality (13.56).
2. The set-valued map W is a gap function for weak vector variational inequal-
ity (13.57).
Proof. Follows directly from the definition of the gap functions (see also [90]). u
t
In the following result, we study the differentiability properties of the set-valued
map G:
Theorem 13.4.4. Let C Rm be a pointed, closed, and convex cone, let K Rn
be a compact set, and let F W K ! L.Rn ; Rm / be a continuously differentiable
N u 2 K, let yN D hF .x/;
map. For x; N xN ui 2 G.x/,
N and
lim khF .x/;

N zik D 1: (13.61)
kzk!1
Then, for any x 2 dom .DG.x;

N y//,
N the following identity holds:
N y/.x/
DG.x; N D hrF .x/.x/;
N xN ui C hF .x/;
N x T .K; u/i;
where DG.x; N y/
N is the contingent derivative of G at .x;
N y/
N and rF .x/
N is the
Fréchet derivative of F .
Proof. For any x 2 dom .DG.x; N y//,
N let y 2 DG.x; N y/.x/.
N Then there are
sequences f.xn ; yn /g Rn Rm and ftn g P such that .xn ; yn / ! .x; y/, tn # 0,
and yN C tn yn 2 G.xN C tn xn / D hF .xN C tn xn /; xN C tn xn Ki for every n 2 N.
Therefore, for every n 2 N, there exists fzn g K such that
yN C tn yn D hF .xN C tn xn /; xN C tn xn zn i;
which, after a rearrangement of terms, reads
N xN ui hF .xN C tn xn /; xN zn i D hF .xN C tn xn /; tn xn i tn yn :
hF .x/; (13.62)
Using the Taylor expansion, we have
F .xN C tn xn / D F .x/
N C tn rF .x/.x
N n / C o.tn xn /: (13.63)
By substituting (13.63) into (13.62), and by dividing the resulting identity by tn ,

we get

zn u o.tn xn /
N
F .x/; D rF .x/.x
N n/ C ; xN zn C hF .xN C tn xn /; xn i yn :
tn tn
(13.64)
Since the set K is compact, we can assume that zn ! xQ 2 K. Since

o.tn xn /
N n/ C
rF .x/.x ; xN zn C hF .xN C tn xn /; xn i yn
tn
! hrF .x/.x/;
N xN xi
Q C hF .x/;
N xi y; (13.65)
nD Eo
we deduce from (13.64) that F .x/; N zntu must be a convergent sequence.
n o n
We claim that zntu remains bounded. Indeed, if this is not the case, then
n
n o

there is a subsequence zmtmu such that zmtmu ! 1. Then (13.61) yields
D E

N zmtmu ! 1, which is a contradiction of (13.65).
F .x/;

Therefore, we can assume that there exists ˇ > 0 such that zntu
n
ˇ, which,
due to the fact that tn # 0, implies that xQ D u. Since we can assume that zntu
n
! x,
M
we get xM 2 T .K; u/. From (13.64) and the fact that xQ D u, we have hF .x/;
N xiM D
N
hrF .x/.x/; xN ui C hF .x/;
N xi y, implying
[
y D hrF .x/.x/;
N N
xuiChF N x xi
.x/; M 2 hrF .x/.x/;
N N
xuiC hF .x/;
N xzi;
z2T .K;u/
N y/.x/
and since y was arbitrary, we have DG.x; N hrF .x/.x/;
N xN uiChF .x/;
N x
T .K; u/i.
For the converse, let z 2 T .K; u/ be such that y D hrF .x/.x/;N xN ui C
hF .x/;
N x zi. Due to the containment z 2 T .K; u/, there are sequences fzn g K,
zn u
ftn g P such that zn ! u, tn # 0, and ! z. We choose fxn g Rn such
tn
that xn ! x and define fyn g Rm by

F .xN C tn xn / F .x/
N zn u
yn D ; xN u C F .xN C tn xn /; xn :
tn tn
Clearly, we have yn ! y and yN C tn yn D hF .xN C tn xn /; xN C tn xn zn i 2

G.xN C tn xn /, which confirms that y 2 DG.x;
N y/.x/.
N This completes the proof. u t
In the following result, we study the sensitivity properties of the gap function W :
Theorem 13.4.5. Let C Rm be a pointed, closed, convex and solid cone, let K
Rn be a compact set, and let F W K ! L.Rn ; Rm / be a continuously differentiable
map. For x;N u 2 K, we define yN D hF .x/; N xN ui 2 W .x/.N Assume that G is
N y/
derivable at .x; N and (13.61) holds. Then for any x 2 dom .DW.x;
N y//,
N
N y/.x/
DW.x; N WMax .DG.x;
N y/.x/;
N C/:
Proof. For any x 2 dom .DW.x; N y//,

N let y 2 DW.x;N y/.x/.
N By the definition of
W , we have y 2 DG.x; N y/.x/.
N Assume that y … WMax .DG.x;N y/.x/;
N C /. Then,
there exists v 2 DG.x;
N y/.x/
N such that
v y 2 int.C /: (13.66)
The containment y 2 DW.x; N y/.x/

N ensures that there exist sequences
f.xn ; yn /g Rn Rm and ftn g P with .xn ; yn / ! .x; y/, tn # 0 such that
yN C tn yn 2 W .xN C tn xn / for every n 2 N.
Furthermore, from v 2 DG.x; N y/.x/
N and the fact that the map G is derivable
N y/,
at .x; N there exists .un ; vn / with .un ; vn / ! .x; v/ such that yN C tn vn 2 G.xN C
tn un / D hF .xN C tn un /; xN C tn un Ki, and consequently, there exists fzn g K such
that
yN C tn vn D hF .xN C tn un /; xN C tn un zn i: (13.67)
N we also have
Since the map F is continuously differentiable at x,
F .xN C tn un / D F .x/
N C tn rF .x/.u
N n / C o.tn un /; (13.68)
F .xN C tn xn / D F .x/
N C tn rF .x/.x
N n / C o.tn xn /: (13.69)
By rearranging the above equations, and using the notion o.tn / WD

o.tn un / o.tn xn /
, we get
tn
hF .xN C tn un /; xN C tn un zn i D hF .xN C tn xn /; xN C tn un zn i
C htn rF .x/.u
N n xn / C o.tn /; xN C tn un zn i
D hF .xN C tn xn /; xN C tn xn zn i
C hF .xN C tn xn /; tn .un xn /i
C htn rF .x/.u
N n xn / C o.tn /; xN C tn un zn i:
(13.70)
We set

o.tn /
˛.n/ D hF .xN C tn xn /; un xn i C rF .x/.u
N n xn / C ; xN C tn un zn :
tn
Clearly ˛.n/ ! 0, and from (13.67) and (13.70), we get yN C tn vn D hF .xN C

tn xn /; xN C tn xn zn i C tn ˛.n/ and hence yN C tn .vn ˛.n// 2 G.xN C tn xn /, which
in view of the definition of W implies that .yN C tn .vn ˛.n// .yN C tn yn / … int.C /
implying vn ˛.n/ yn … int.C /, and hence v y … int.C /, a contradiction
of (13.66). The proof is complete. t
u
Remark 13.4.6. Several related results can be found in [384]. Furthermore, dif-
ferentiability and sensitivity of gap functions for the Minty vector variational
inequalities can be found in [418].
In the remaining part of this section, we derive second-order analogues of the
derivative formulas obtained above. We begin with the following result which was
originally given in [378]:
Theorem 13.4.7. Let C Rm be a pointed, closed, and convex cone, and let K
Rn be a compact set. Let F W K ! L.Rn ; Rm / be a twice Fréchet differentiable
N u 2 K be given elements and let yN D hF .x/;
map. Let x; N xN ui 2 G.x/. N Assume
that (13.61) holds. Then, for any x 2 dom .D 2 G.x;
N y;
N uN ; vN //, we have
[
N y;
D 2 G.x; N uN ; vN /.x/ D N x zi C hrF .x/.N
ŒhF .x/; N u/; 2.Nu w/i
Q
z2T 2 .K;x;w/
Chr 2 F .x/.N
N u; uN /; xN ui C hrF .x/.x/;
N xN ui;
with w satisfying hF .x/;

N wi D Nv C hrF .x/N
N u; xN ui C hF .x/;
N uN i.
Proof. For x 2 dom .D 2 G.x; N y;
N uN ; vN //, let y 2 D 2 G.x;
N y;
N uN ; vN /.x/ be arbitrary.
Then, there are sequences f.xn ; yn /g Rn Rm , ftn g P with .xn ; yn / ! .x; y/
and tn # 0 such that for every n 2 N,

1 2 1 2
yN C tn vN C tn yn 2 G xN C tn uN C tn xn
2 2

1 1
D F xN C tn uN C tn2 xn ; xN C tn uN C tn2 xn K :
2 2
Therefore, there exists zn 2 K such that

1 2 1 2 1 2
yN C tn vN C tn yn D F xN C tn uN C tn xn ; xN C tn uN C tn xn zn ; (13.71)
2 2 2
and since K is compact, we have zn ! xM 2 K. The above equation then yields

yN D hF .x/;
N xN xi.
M
Since F is twice differentiable, we have

1 1 1
F xN C tn uN C tn2 xn D F .x/ N C tn rF .x/.N
N u/ C tn2 rF .x/x
N n C tn2 r 2 F .x/
N
2 2 2
2
1 1 1 2
uN C tn xn ; uN C tn xn C o tn uN C tn xn

:
2 2 2
(13.72)
Using (13.71) and (13.72), we have

1 1
yN C tn vN C tn2 yn D hF .x/;
N xN zn i C tn hF .x/;
N uN i C tn hrF .x/NN u; xN zn i C tn2 hF .x/;
N xn i
2 2

1 1 1
C tn2 hrF .x/N
N u; 2Nu C tn xn iC tn2 rF .x/.x N n uN C tn2 xn zn
N n /; xCt
2 2 2

1 1 1 1
C tn2 r 2 F .x/
N uN C tn xn ; uN C tn xn ; xN C tn uN C tn2 xn zn
2 2 2 2
2
* 1 2 +
1 2 o tn uN C 2 tn xn 1 2
C tn 1 2
; xN C tn uN C tn xn zn ; (13.73)
2 2 tn
2
which yields

zn u
N
F .x/; ! Nv C hF .x/;
N uN i C hrF .x/N
N u; xN xi:
M (13.74)
tn
n o
zn u
We claim that the sequence is bounded. In fact, if this is not the case,
tn

then there exists a subsequence such that zntu ! 1. By (13.61), we have
D E n
zn u
N
F .x/; ! 1, which contradicts (13.74). Therefore, we can assume that
tn
zn u
is bounded. This implies that u D x.M Moreover, we can assume that
tn
zn u
! w which implies that w 2 T .K; u/. Furthermore,
tn
hF .x/;
N wi D Nv C hF .x/;
N uN i C hrF .x/N
N u; xN ui:
Using (13.73) and (13.74), we deduce that

* zn u
+ * +
tn
w zn u
N
F .x/; 1
D yn rF .x/.N
N u/; 1 C hF .x/;
N xn i
t
2 n
t
2 n
C hrF .x/.N
N u/; 2Nu C tn xn i

1
N n ; xN C tn uN C tn2 xn zn
C rF .x/x
2

1 1 1 2
C r F .x/
2
N uN C tn xn ; uN C tn xn ; xN C tn uN C tn xn zn
2 2 2
* +
o tn uN C 12 tn2 xn
2
1
C 1 2
; xN C tn uN C tn2 xn zn ;
t
2 n
2
and hence
* zn u
+
tn w
N
F .x/; 1 2
! hF .x/;
N xi C hrF .x/N
N u; 2.Nu w/i C hrF .x/x;
N xN ui
2 tn
˝ ˛
C r 2 F .x/.N
N u; uN /; xN u y:
zn u
w
By using the same arguments, we obtain tn1 t 2 ! z with z 2 T 2 .K; u; w/.
2 n
Therefore,
[
y2 N x zi C hrF .x/.N
ŒhF .x/; N u/; 2.Nu w/i C r 2 F .x/.N
N u; uN /; xN ui
N
z2T 2 .K;x;w/
ChrF .x/.x/;
N xN ui:
For the converse, assume that
y D hF .x/;
N xziChrF .x/N
N u; 2.Nuw/iChrF .x/.x/;
N N
xuiChr 2
N u; uN /; xui
F .x/.N N
with z 2 T 2 .K; u; w/. Then there is a sequence fzn g K with zn ! u and tn # 0

zn u
w
such that zntu
n
! w and tn1 t ! z.
2 n
Since F is twice differentiable, we can take sequence xn and yn such that xn ! x
and
* zn u
+
1 2 tn w
yn D F .xN C tn uN C tn xn /; xn 1
2 2 tn
* +
F .xN C tn uN C 12 tn2 xn / F .x/
N
C ; 2.Nu w/ C hrF .x/x N n ; xN ui
tn
* +
F xN C tn uN C 12 tn2 xn F .x/ N rF .x/.t N n uN C 12 tn2 xn /
C 1 2
; xN u :
t
2 n
Therefore yn ! y and
1
yN C tn .hrF .x/.N
N u/; xN ui C hF .x/;
N uN wi/ C tn2 yn
2

1 2 1 2 1
D F xN C tn uN C tn xn ; xN C tn uN C tn xn zn 2 G xN C tn uN C tn2 xn :
2 2 2
Since vN D hrF .x/.N

N u/; xN ui C hF .x/;
N uN wi, we get y 2 D 2 G.x;
N y;
N uN ; vN /.x/.
u
We now give a second-order sensitivity analysis for the weak vector variational
inequalities. In the following result, we continue with the setting of Theorem 13.4.7.
Theorem 13.4.8. Let xN 2 K, yN 2 G.x/ N and .Nu; vN / 2 T .graph .G/; .x;
N y//.
N Assume
that G is second-order derivable .x; N y/
N in the direction .Nu; vN /. Then for any x 2
N y;
dom .D 2 W .x; N uN ; vN //, we have

N y;
D 2 W .x; N uN ; vN /.x/ WMax D 2 G.x;
N y;
N uN ; vN /.x/; C : (13.75)
Proof. Let y 2 D 2 W .x; N y;

N uN ;vN /.x/. Therefore, y 2 D 2 G.x;
N y;
N uN ; vN /.x/. If y …
2
WMax D W .x; N y;
N uN ; vN /.x/; C , then there exists v 2 D 2 G.x;
N y;
N uN ; vN /.x/ such that
v y 2 int.C /: (13.76)
N y;
Since y 2 D 2 W .x; N uN ; vN /.x/, there are sequences .xn ; yn / ! .x; y/ and tn !
0 such that

1 2 1 2
yN C tn vN C tn yn 2 W xN C tn uN C tn xn : (13.77)
2 2
Analogously, since v 2 D 2 G.x; N y;

N uN ; vN /.x/, there exists .un ; vn / ! .x; v/ such
that

1 1
yN C tn vN C tn2 vn 2 G xN C tn uN C tn2 un :
2 2
Therefore, there exists fzn g K such that

1 2 1 2 1 2
yN C tn vN C tn vn D F xN C tn uN tn un ; xN C tn uN C tn un zn : (13.78)
2 2 2
Since F is twice differentiable, we have

1 1
F xN C tn uN C tn2 un D F .x/ N C tn rF .x/ N uN C tn un
2 2

1 2 2 1 1
C tn r F .x/ N uN C tn un ; uN C tn un
2 2 2
2
1 2
Co
tn uN C 2 tn un : (13.79)
Using (13.72) and (13.79), we have

1 2 1 2
F xN C tn uN C tn un ; xN C tn uN C tn un zn
2 2

1 1
D F xN C tn uN C tn2 xn ; xN C tn uN C tn2 xn zn
2 2

1 1
C tn2 F xN C tn uN C tn2 xn ; un xn
2 2

1 2 1 2
C tn rF .x/.u N n xn /; xN C tn uN C tn zn
2 2

1 1
C tn3 r 2 F .x/.N N u; un xn /; xN C tn uN C tn2 un zn
2 2

1 1
C tn4 r 2 F .x/.u N n ; un / r 2 F .x/.x
N n ; xn /; xN C tn uN C tn2 un zn
8 2
* +
1 o.t 2 / 1
C tn2 1 n2 ; xN C tn uN C tn2 un zn ;
2 t
2 n
2
2 2
where o.tn2 / D o tn uN C 12 tn2 un o tn uN C 12 tn2 xn .
We define

1 2
˛.n/ D F xN C tn uN C tn xn ; un xn
2

1
C rF .x/.uN n xn /; xN C tn uN C tn2 un zn
2

1
Ctn r 2 F .x/.N
N u; un xn /; xN C tn uN C tn2 un zn
2
2 2
tn 2 tn 2 1 2
C r F .x/.u
N n ; un / r F .x/.x N n ; xn /; xN C tn uN C tn un zn
4 4 2
* +
o.t 2 / 1
C 1 n2 ; xN C tn uN C tn2 un zn ;
t
2 n
2
and note that ˛.n/ ! 0.

Furthermore

1 1 1
yN C tn vN C tn2 Œvn ˛.n/ D F xN C tn uN C tn2 xn ; xN C tn uN C tn2 xn zn
2 2 2

1
2 G xN C tn uN C tn2 xn ;
2
and it follows from the definition of the map W and (13.79) that

1 2 1 2
yN C tn vN C tn Œvn ˛.n/ yN C tn vN C tn yn … int.C /
2 2
which implies that vn ˛.n/ yn … int.C /. Therefore, v y … int.C / which is a

contradiction to (13.76). The proof is complete. t
u
Remark 13.4.9. We remark that numerous variants of the above results and illustra-
tive examples can be found in [378]. All the results stated in this section are given
in a finite-dimensional setting and it is natural to expect their extensions in general
setting.
Chapter 14
Numerical Methods for Solving Set-Valued
In this chapter we present solution procedures for solving set-valued optimization

problems. In Sect. 14.1 a Newton method for solving general set-valued optimiza-
tion problems is shown. For a special class of set-valued optimization problems
where the objective map is polyhedral, convex and set-valued, we present an
algorithm in Sect. 14.2.
Furthermore, in Sect. 15.4 we discuss the relationship between multiobjec-
tive optimization problems with uncertainties and set-valued optimization and
present solution procedures for these special set-valued optimization problems (see
Sect. 15.4.4).
14.1 A Newton Method for Set-Valued Maps
In Chap. 12 we have shown necessary optimality conditions for solutions of set-

valued optimization problems. In algorithms for solving set-valued problems one
has to study these optimality conditions, especially given in Theorem 12.8.9. This
means that we have to solve certain inclusions involving set-valued maps.
It is well known from scalar optimization that the Newton method is one
of the most powerful tools for solving systems of nonlinear equations. In the
paper by Hoheisel, Kanzow, Mordukhovich, Phan [266] a new generalized Newton
algorithm based on graphical derivatives is proposed and justified. The method in
[266] reduces to the classical Newton method in the smooth case, being different
from previously known versions of Newton’s method in the case of Lipschitz
continuous mappings. Based on advanced techniques of variational analysis and
generalized differentiation, the well-posedness of the algorithm, its local superlinear
convergence, and its global convergence of the Kantorovich type is derived in [266].
Set-valued generalizations of the Newton method are described in the paper by
Dias and Smirnov [137]. In this section we follow the approach given in [137].

DOI 10.1007/978-3-642-54265-7__14
646 14 Numerical Methods for Solving Set-Valued Optimization Problems
The Newton method developed in [137] is a general tool to solve systems of

nonlinear equations with geometric constraints and systems of nonlinear inequalities
in a unified manner.
In the classical case these nonlinear equations are given for continuously
differentiable functions f W Rn ! Rq (compare [446]) by
f .x/ D 0 (14.1)
and the Newton method generates the following sequence of iterates
x kC1 WD x k C x k ; k D 0; 1; ; (14.2)
where x k is a solution of the system of linear equations
rf .x k /x k D f .x k /: (14.3)
Dias and Smirnov [137] introduced a Newton-type method suitable to solve

inclusions given for a set-valued map F W Rn Rq by
0 2 F .x/: (14.4)
These Newton-type methods generate sequences of iterates (compare (14.2))
x kC1 WD x k C x k ; k D 0; 1; ;
where x k is now a solution of an inclusion that is a generalization of the linear

system (14.3). This generalized system involves set-valued map derivatives (see
[16]). Namely, instead of (14.3) one solves an inclusion
vk 2 .x k ; vk /.x k /; (14.5)
where vk is a nearest to zero point belonging to the set F .x k /, and the graph of the
set-valued map .x k ; vk /./ is a tangent cone in some sense to the graph of F at the
point .x k ; vk /. In order to solve the linear system (14.3) and to prove convergence
results it is essential to have the non-singularity of the matrix rf .x k /. This
condition will be replaced by the condition of metric regularity (cf. [430, 499] and
a survey in [21]). In the case of linear operators this method goes back to the open
mapping theorem and in the case of smooth maps to the Lyusternik theorem. It was
successfully applied to justify the well-posedness and convergence of the Newton
method for nonsmooth equations by Hoheisel, Kanzow, Mordukhovich, Phan in
[266]. Dias and Smirnov [137] show that the metric regularity of Lipschitzian set-
valued maps is equivalent to the well-posedness of the Newton method for perturbed
maps.
14.1 A Newton Method for Set-Valued Maps 647
In the following we use the distance between a point x 2 Rn and a set S Rn

denoted by
d.x; S / D inffjjx sjj j s 2 S g:
The projection of a vector x onto S is defined by
PS .x/ WD fs 2 S j jjx sjj D d.x; S /g:
In the formulation of the Newton method for set-valued maps we use the
contingent derivative of F W Rn Rq at the point .x; y/ 2 graph F , i.e., the set-
valued map DF.x; y/ W X Rq defined by
graph DF.x; y/ D T .graph F; .x; y//;
where T .graph F; .x; y// is the contigent cone to the set graph F at the point
.x; y/ 2 graph F (see Sect. 4.1.2). In other words, y 2 DF.x; y/.x/ if and only
if .x; y/ 2 T .graph F; .x; y//. The contingent cone to a set S X (X a real
normed space) at a point xN 2 S is introduced in Sect. 4.1.2, Definition 4.1.11. We
will use in the following the cone T13 studied in Theorem 4.1.12:

d.xN C tx; S /
T .S; x/ N D x 2 X j lim inf
N WD T13 .S; x/ D0 :
t !C0 t
Furthermore, we use the coderivative by Mordukhovich D F .x; y/W Y X of

a set-valued map F at .x; y/ 2 graph F introduced in Definition 3.5.8.
Next, we recall the notion of metric regularity for set-valued maps and its
coderivative characterization (see Mordukhovich [430], Rockafellar, Wets [499]),
Durea, Strugariu [153, 154], Apetrii, Durea, Strugariu [9]). A set-valued map F W
Rn Rq is called metrically regular around .x; y/ 2 graph F if there exist " > 0
and > 0 such that
d.x; F 1 .y// d.y; F .x//; x 2 x C "BRn ; y 2 y C "BRq :
The following coderivative characterization of the metric regularity for set-

valued maps is given in [430] and [499].
Theorem 14.1.1. A set-valued map F W Rn Rq with closed graph is metrically
regular around .x; y/ 2 graph F if and only if the inclusion 0 2 D F .x; y/.y /
implies that y D 0.
The well-posedness of the Newton method for perturbed set-valued maps
F .x/ y,
O where F W Rn Rq is a Lipschitz continuous map in the sense of
Definition 3.3.11 and the vectors yO 2 Rq have a sufficiently small norm, can be
used in order to characterize the metric regularity. Using the contingent derivative
we say that the Newton method for perturbed set-valued maps F is well-posed
around .x; y/ 2 graph F with modulus if there exists > 0 such that for all
x 2 x C BRn and yO 2 y C BRq there is y 2 PF .x/yO .0/ satisfying the condition
DF 1 .y C y;
O x/.y/ \ jjyjjBRn ¤ ;: (14.6)
Condition (14.6) implies that the largest possible Newton inclusion (14.5) for the
perturbed map x ! F .x/ yO (the generalization of the Newton equation (14.3))
has at least one solution x 0 2 BRn .
The following theorem given by Dias and Smirnov [137, Theorem 3.1] explains
under which conditions the generalized Newton method is well defined.
Theorem 14.1.2. Assume that the set-valued map F W Rn Rq with closed values
is Lipschitz continuous with a constant LF > 0 in a neighbourhood of a point x in
the sense of Definition 3.3.11. Then the following conditions are equivalent:
1. The map F is metrically regular around .x; y/ 2 graph F .
2. The Newton method for perturbed set-valued maps F is well posed around
.x; y/ 2 graph F .
In the following we present a convergence analysis concerning the set-valued
version of the Newton method. Consider a set-valued map F W Rn Rq with closed
values and .x k ; y k / 2 graph F . The step-length in the algorithm is given by t k 2
.0; 1. If t k D 1 this method is called Newton method and if t k 2 .0; 1, the
method is called the damped Newton method. For example, the step-length t k can
be chosen from the condition
d.0; F .x k C t k x k // D min d.0; F .x k C tx k //:

t 2Œ0;1
For a given point .x k ; y k / 2 graph F we define the next iterate by
x kC1 WD x k C t k x k ; y kC1 2 PF .x kC1 / .0/; (14.7)
where x k is a solution of the inclusion
y k 2 .x k ; y k /.x k /;
and .x; y/ W Rn ! Rq is a positively homogeneous set-valued mapping satisfying

the inclusion .x; y/.x/ DF.x; y/.x/ for all .x; y/ 2 graph F and x. The vector
x k is chosen from the condition
x k 2 P..x k ;y k //1 .y k / .0/: (14.8)
The condition (14.8) uniquely defines x k if ..x k ; y k //1 .y k / is convex.

14.1 A Newton Method for Set-Valued Maps 649
Suppose x … F 1 .0/, y 2 PF .x/ .0/ and x 2 ..x; y//1 .y/. Then there exists
p.t/ 2 Rq such that
y C ty C p.t/ 2 F .x C tx/
and
jjp.t/jj D tı.x; y; x; y; t/;
where
ı.x; y; x; y; t/ WD t 1 d.y C ty; F .x C tx//: (14.9)
If F is Lipschitz continuous in the sense of Definition 3.3.11, then
lim inf ı.x; y; x; y; t/ D 0:

t !C0
Furthermore, if y D y, then we have
d.0; F .x C tx// jj.1 t/y C p.t/jj .1 t/jjyjj C tı.x; y; x; y; t/:
This means that the rate of convergence of the Newton method depends on the
properties of the function ı. The following result concerning the rate of convergence
is shown in [137, Theorem 4.1].
Theorem 14.1.3. Suppose that the following conditions are satisfied:
(i) There exists > 0 such that for all x 2 Rn , y 2 PF .x/ .0/ and y 2 Rq the set
..x; y//1 .y/ \ jjyjjBRn is non-empty.
(ii) There exists a monotone increasing function ! W Œ0; C1/ ! Œ0; C1/ such
that lim˛!C0 !.˛/ D 0 and ı.x; y; x; y; t/ jjyjj!.tjjyjj/ for all x 2 Rn n
F 1 .0/, y 2 PF .x/ .0/ and x 2 P..x;y//1 .y/ .0/.
Then for any initial point x 0 there exists a monotone non-decreasing sequence
t > 0, such that t k D 1 for large k, and the corresponding damped Newton
k
method / Newton method, starting at x 0 , converges to a point x 1 WD limk!1 x k

with x 1 2 x 0 C 2d.0; F .x 0 //BRn . If graph F is closed, then x 1 2 F 1 .0/ and
the convergence is R-superlinear. If !.˛/ D O.˛/, ˛ ! C0, then the convergence
is R-quadratic. Furthermore, if F is additionally locally Lipschitz continuous, then
the convergence is Q-quadratic.
In the following, we discuss the set-valued version of the Newton method for
systems of nonlinear equations with geometric constraints.
Suppose that f W Rn Rl ! Rq is a continuously differentiable function.
Its derivative is supposed to be Lipschitz continuous with a constant Lrf . Let us
consider the nonlinear equation
f .x; u/ D 0: (14.10)
We are looking for a solution .x; u/ such that the variable u satisfies the geometric
constraints
u 2 U; (14.11)
where U Rl is a closed convex set. Let U1 be the recession cone of U , i.e.,
U1 WD fu 2 Rl j u C U U g:
Consider, for example, a system of equations and inequalities given for functions
g W Rn ! Rql and h W Rn ! Rl by
g.x/ D 0; h.x/ 0:
This system can be rewritten in the form (14.10) and (14.11) in the following way:
g.x/ D 0; h.x/ C u D 0; u
0:
In this case it holds
U D U1 D fu 2 Rl j u
0g:
The system (14.10) and (14.11) is equivalent to the inclusion
0 2 F .x/ WD f .x; U /:
Let y D f .x; u/ 2 PF .x/ .0/. In order to apply the generalized Newton method
to this inclusion, put for x 2 Rn
.x; u/.x/ WD rx f .x; u/x C ru f .x; u/U1 : (14.12)
O uO / 2 Rn U such that f .x;

Let .x; O uO / D 0. Moreover, assume that
O uO //T /1 U1
..ru f .x; C
\ ker rx f .x;
O uO / D f0g:
Now, from Theorem 14.1.1 it follows that the set-valued map F is metrically regular
O 0/. Furthermore, because of Theorem 14.1.2 the Newton method is well
around .x;
O 0/ and using the Lipschitz constant L we get
posed around .x;
d.0; ..x; u//1 .y// L jjyjj;
whenever .x; u/ 2 .x;

O uO / C BRnCl and > 0 is sufficiently small.
Let y 2 PF .x/ .0/. Because of the well posedness of the Newton method, (14.5)
and (14.12) there exist x 2 Rn and w 2 U1 such that
y D rx f .x; u/x C ru f .x; u/w

14.2 An Algorithm to Solve Polyhedral Convex Set-Valued Optimization. . . 651
and
1
.jjxjj2 C jjwjj2 / 2 L jjyjj:
Because of u C tw 2 U , t > 0, we get (see (14.9))
ı.x; y; x; y; t/ D t 1 d.y ty; F .x C tx//

t 1 jjy ty f .x C tx; u C tw/jj
tLrf .jjxjj2 C jjwjj2 /
tLrf .L jjyjj/2 :
Applying Theorem 14.1.3 one gets assertions concerning the convergence of the
generalized Newton method for solving the system (14.10) and (14.11). Especially,
we can see that the Newton method, starting in a sufficiently small neighbourhood
O converges Q-quadratically to the set F 1 .0/.
of x,
14.2 An Algorithm to Solve Polyhedral Convex Set-Valued

Kuroiwa [354] introduced an embedding approach for set-valued optimization

problems (see Sect. 2.6.4) such that it is possible to use methods of vector
optimization in order to solve set-valued problems. Furthermore, Löhne, Schrage
[396] derived an algorithm for solving polyhedral convex set-valued optimization
problems using a vectorial relaxation of the set-valued optimization problem. Very
recently, Jahn [294] developed a derivative-free descent method for the solution of
set optimization problems. In this section we present the results shown by Löhne,
Schrage [396] for set-valued problems where the graph of the set-valued objective
map is supposed to be given by finitely many linear inequalities. A minimal solution
of the set-valued optimization problem is understood to be a set of elements in the
domain satisfying two conditions: The attainment of the infimum and minimality
with respect to a set order relation (see Sect. 2.6.2). The algorithm presented in this
section is divided into two phases: In the first phase, a linear vector optimization
problem, called the vectorial relaxation, is solved. The generated pre-solution yields
the attainment of the infimum but, in general, not minimality. In the second phase
of the algorithm, minimality is established by solving certain linear programs in
combination with vertex enumeration of some values of the set-valued objective
map.
Throughout this section we suppose that C Rq is a proper pointed polyhedral
convex cone with int C ¤ ;. We present a solution procedure for set-valued
optimization problems where the objective map F W Rn Rq has a polyhedral

convex graph. We will use a solution concept given by the order relation defined
by (compare Sect. 2.6.2)
F .x/ F .u/ ” F .x/ C C F .u/:
In the following G denotes the family of all closed convex subsets of Rq . The
objective map F W Rn Rq can be considered as a function from Rn into G .
Furthermore, by GC we denote the subfamily of those elements P of G having
the additional property
P D P C C:
For P; Q 2 G we use the lower set less order relation lC (see Definition 2.6.9,
for simplicity we denote this relation by in this section):
P Q ” P C C Q C C; (14.13)
which can be equivalently written as (compare Sect. 2.6.2)
P C C Q:
The order relation is reflexive and transitive in G (preorder) and, additionally,

antisymmetric in GC (partial order). An equivalence relation is defined for P; Q 2
G by
P Š Q W” P C C D Q C C: (14.14)
Clearly, the quotient space G = Š is isomorphic to GC and thus is a partial ordering

in G = Š.
The partially ordered set .G = Š; / provides a complete lattice, i.e., for every
subset of G = Š there exist the infimum and supremum (see Sect. 2.6.3) taking
into account Proposition 2.6.25. To simplify the notation we express the infimum
and supremum in terms of the (preordered) space .G ; /, where we have in mind
that we are dealing with representatives of equivalence classes. Thus, for nonempty
sets P G we have (compare Sect. 2.6.3, Proposition 2.6.48 and Löhne [395,
Proposition 1.62])
[ \
inf P D cl conv .P C C /; sup P D .P C C /:
P 2P P 2P
As usual, we set inf ; D ; and sup ; D Rq .

14.2.1 Formulation of the Polyhedral Convex Set-Valued

Optimization Problem
We consider an set-valued objective map F W Rn Rq and the following set

optimization problem (compare problem (SP ) in Sect. 2.6.2):
minimize F .x/; subject to x 2 Rn : (SP )
Problem (SP ) is called feasible if dom F ¤ ;.

The objective map F can be considered as a map from Rn into the space G of all
closed convex subsets of Rq . With the above ordering relation (14.13), one obtains
a complete lattice, i.e., the infimum inffF .x/ j x 2 Rn g in the sense of a greatest
lower bound with respect to always exists. Solution concepts for complete-lattice-
valued problems are discussed in [257].
In this section we present an algorithm for solving (SP) under the assumption
that F W Rn Rq is a polyhedral convex set-valued map, this means that its graph
graph F is a polyhedral convex set.
Furthermore, we assume throughout that
(A1) There is a matrix Z 2 Rqp such that the cone C has the representation
C D fy 2 Rq j Z T y
0g; (14.15)
(A2) There are A 2 Rmn , B 2 Rmq and b 2 Rm such that the graph of F has
the representation:
graph F D f.x; y/ 2 Rn Rq j Ax C By
bg: (14.16)
(A3) (SP ) is bounded in the sense that
9v 2 Rq W fvg infn F .x/:

x2R
It is sufficient for the boundedness of (SP ) in the sense of (A3) that dom F
is a bounded set and
8x 2 Rn ; 9v 2 Rq W fvg F .x/:
Obviously, the latter condition is fulfilled for a map F with bounded values F .x/.
Remark 14.2.1. The algorithm presented in Sect. 14.2.2 can verify whether the
problem is bounded or not.
Remark 14.2.2. Because of assumption (A2) we consider problems with a polyhe-
dral convex set-valued objective map.
The following solution concept is based on a combination of minimality and

infimum attainment as these notions do no longer coincide in vector and set
optimization. It is an adaptation of the concepts introduced in Sect. 2.6.3 and [257]
to the present setting.
In order to express minimality in the following definition, we introduce for
P; Q 2 G :
P Q W” P Q and not P Š Q:
Definition 14.2.3. A point x 2 dom F is said to be a minimal solution for (SP)

if there is no x 2 Rn with F .x/ F .x/. A finite set S dom F is called a finite
infimizer for (SP ) if the infimum is attained in S , that is,
inf F .x/ D infn F .x/:

x2S x2R
A finite infimizer S of (SP ) is called a solution set of (SP ) if it consists of

only minimal solutions.
For deriving the algorithm we study the vectorial relaxation of the set-valued
optimization problem (SP ). Consider the linear function f W Rn Rq ! Rq
defined by f .x; y/ WD y. Because of formal reasons we understand f as a set-
valued map whose values are singleton sets, i.e.,
f W Rn Rq Rq ; f .x; y/ D fyg:
The vectorial relaxation of the set-valued optimization problem (SP ) is

defined with f W Rn Rq Rq as:
minimize f .x; y/ subject to y 2 F .x/: (VecRel)
Obviously, (VecRel) can be seen as a special case of a set-valued optimiza-

tion problem, whence the above definitions apply also to (VecRel). Further-
more, (VecRel) is feasible if and only if (SP ) is feasible. Since f is
single-valued and the constraint y 2 F .x/ can be expressed by finitely many linear
inequalities, (VecRel) is (equivalent to) a linear vector optimization problem.
It holds
inf F .x/ D infn inf fyg D inf f .x; y/; (14.17)

x2Rn x2R y2F .x/ x2Rn ;y2F .x/
i.e., (SP ) and (VecRel) have the same infima. This implies that (SP ) is
bounded if and only if (VecRel) is bounded. Equation (14.17) is a motivation for the
following solution concept.
Definition 14.2.4. A finite set fx i 2 Rn j i D 1; ; kg is called a pre-solution set

of (SP ) if there exist y i 2 Rq , i D 1; ; k, such that f.x i ; y i / 2 Rn Rq j i D
1; ; kg is a solution set of the vectorial relaxation (VecRel) of (SP ).
It is important to mention that a pre-solution set of (SP ) is in general not a
solution set of (SP ) (see Example 3.2 in [396]).
This discussion leads us to the algorithm described in the following section.
14.2.2 An Algorithm for Solving Polyhedral Convex Set-Valued

Taking into account the assumptions (A1) and (A2), let a problem of type (SP )
be given by A 2 Rmn , B 2 Rmq , b 2 Rm , Z 2 Rqp . The algorithm generates a
solution set of (SP ) if the problem is feasible and bounded. Otherwise it detects
whether (SP ) is infeasible or unbounded.
In the first phase of the algorithm, the vectorial relaxation (VecRel), that is
(equivalent to) a linear vector optimization problem, is solved. A solution set
of (VecRel) can be obtained, for instance, with Benson’s algorithm, see e.g.
[46,160,395,528,529]. We know that (SP ) is bounded if and only if (VecRel) is
bounded. However, Benson’s algorithm is able to detect if (VecRel) is unbounded. It
is important to mention that Benson’s algorithm was extended for unbounded linear
vector optimization problems in the book by Löhne in [395].
In the second phase, for every element x 0 belonging to the pre-solution set
obtained in the first phase, one constructs a sequence .x 0 ; x 1 ; x 2 ; ; x l / with
F .x 0 / F .x 1 / F .x 2 / F .x l / until a minimal solution x l for (SP )
in the sense of Definition 14.2.3 is obtained after finitely many steps.
For parameters w 2 C and x 2 dom F , we consider the following scalar
problem:
maximize wT y subject to y 2 F .x/; F .x/ F .x/: (P .w; x/)
As the y D .y1 ; ; yq /T 2 R (yi 2 R for all i D 1; ; q) are considered to

be auxiliary variables, we use the following convention: xO is said to be a solution
of (P .w; x/) if there exists yO 2 Rq such that .x;
O y/
O is a solution of (P .w; x/) in the
O y/
ordinary sense. In practice this means that .x; O is generated but only xO is used. Of
course, there is the following lower bound ˇ for the optimal value ˛ of (P .w; x/):
˛.w; x/ WD supfwT y j y 2 F .x/; F .x/ F .x/g

supfwT y j y 2 F .x/g DW ˇ.w; x/: (14.18)
In the following we will use certain representations of F .x/ C C . The first

representation of F .x/ C C is given for some x 2 dom F , wj 2 Rq , j 2 R
(j D 1; ; r) by
F .x/ C C D fy 2 Rq j .wj /T y j ; j D 1; ; rg: (14.19)
Furthermore, we use a second representation of F .x/ C C given for y 1 ; ; y s 2

Rq by
F .x/ C C D convfy 1 ; ; y s g C C: (14.20)
It is well known that every non-empty polyhedral convex subset of Rq can be

expressed as a (generalized) convex hull of finitely many points and finitely many
directions belonging to Rq . Note that the assumption that (SP ) is bounded was
used in (14.20), otherwise the cone on the right hand side can be a superset of C.
Remark 14.2.5. Numerical methods to compute a representation given by (14.20)
from a representation given by (14.19) and vice versa are called vertex enumeration,
see e.g. [31] and [80].
(14.18) yields the following sufficient optimality condition for minimal solutions
of the set-valued optimization problem (SP ) in the sense of Definition 14.2.3.
Lemma 14.2.6. Let (SP ) be feasible and bounded. Furthermore, for some
x 2 dom F let the representation (14.19) of the set F .x/ C C be given with the
corresponding wj 2 Rq , j 2 R (j D 1; ; r). If ˛.wj ; x/ D ˇ.wj ; x/ for all
j 2 f1; ; rg, then x is a minimal solution for (SP ).
Proof. Suppose that x is not a minimal solution of (SP), i.e., there exists x 2 Rn
with F .x/ F .x/. Hence there is some y 2 F .x/ and some c 2 C such that
y C c … F .x/ C C . Because of C C C D C , we conclude that y … F .x/ C C .
T
Thus there is some j 2 f1; ; rg such that wj y > j which implies ˛.wj ; x/ >
j
ˇ.w ; x/, a contradiction to ˛.w ; x/ D ˇ.wj ; x/ for all j 2 f1; ; rg. u
j j
t
Using the representation (14.20) of F .x/ C C , problem (P .w; x/) can be
transformed into a linear program.
By (14.16), the first constraint y 2 F .x/ in (P .w; x/) can be expressed as Ax C
By
b. The second constraint in (P .w; x/) can be transformed as follows:
F .x/ F .x/ ” F .x/ C C F .x/ C C

” 8i D 1; ; s W y i 2 F .x/ C C
” 8i D 1; ; s; 9c i 2 C W y i c i 2 F .x/ (14.21)
” 8i D 1; ; s; 9c 2 C W Ax Bc
b By :
i i i
Thus, (P .w; x/) is equivalent to the linear program:
max wT y subject to .x; y; c 1 ; ; c s / 2 S; (P .wI y 1 ; : : : ; y s /)
where
S W D f.x; y; c 1 ; ; c s / 2 Rn Rq Rq Rq j Ax C By
b;
Ax Bci
b By i ; Z T c i
0 .i D 1; ; s/g
which has n C q.s C 1/ variables and m C ms C ps constraints. According to the

above convention we will speak about a solution x 2 Rn and do not mention the
q.s C 1/ auxiliary variables y; c 1 ; ; c s 2 Rq .
The following algorithm for solving the set-valued problem (SP) is developed
by Löhne, Schrage [396] under the assumptions (A1), (A2) and (A3).
Algorithm SetOpt
Input:
Representation of graph F according to (14.16): A 2 Rmn , B 2 Rmq , b 2 Rm ;
Representation of the ordering cone C according to (14.15): Z 2 Rqp ;
Output:
A solution set X of (SP ) in the case that (SP ) is feasible and bounded,
X D ; otherwise;
The solution status for (SP );
Phase 1: Solve the linear vector optimization problem (VecRel)

X ;;
solve (VecRel);
if (VecRel) is infeasible then status 00 .SP / is infeasible:00 ; stop; end;
if (VecRel) is unbounded then status 00 .SP / is unbounded:00 ; stop; end;
store a pre-solution set fx 1 ; ; x k g of (SP );
Phase 2: Computation of minimal solutions of (SP )

for i 1 to k do
flag 1;
K ;;
while flag D 1 do
compute the representations (14.20) and (14.19) of F .x i / C C :
F .x i / C C D convfy 1 ; ; y s g C C
F .x i / C C D fy 2 Rq j .wj /T y j ; j D 1; ; rgI
for j 1 to r do
if wj =jjwj jj … K then
K K [ fwj =jjwj jjg;
solve .P .wj I y 1 ; y s //; xi solution; ˛ optimal value;
ˇ maxf.wj /T y 1 ; ; .wj /T y s gI
if ˛ > ˇ then break (i.e., exit the inner-most loop);

end;
if j D r then flag 0;
end;
end;
X X [ fx i g;
end;
status 00 .SP / has been solved:00 ;
14.2.3 Properties of the Algorithm
In [396, Lemmata 4.2 and 4.3] it is shown that the algorithm works correctly and is
finite. This assertion is prepared by two lemmata.
Lemma 14.2.7. Let (SP ) be feasible and bounded and w 2 Rq be fixed.
Consider some x 2 dom F , a halfspace H D fy 2 Rq j wT y g containing
the set F .x/ C C and finitely many points y 1 ; ; y s 2 Rq such that (14.20) holds.
Then,
(a) The linear program (P .wI y 1 ; : : : ; y s /) has a minimal solution;
(b) The lower bound ˇ defined in (14.18) can be expressed as
ˇ.w; x/ D maxfwT y 1 ; ; wT y s g:
Proof. Let y 2 F .x/, such that Ax C By

b because of (A2). By (14.20), for all
i 2 f1; ; sg, we have y i 2 F .x/CC , i.e., there is some c i 2 C such that y i c i 2
F .x/, or equivalently, Ax Bci
b Byi . Hence, the point .x; y; c 1 ; ; c s / is
feasible for (P .wI y 1 ; : : : ; y s /).
There exists v 2 Rq such that fvg C C F .x/ for all x 2 Rn since (SP ) is
supposed to be bounded. Moreover, it holds w 2 C because H contains F .x/CC .
It follows that
supfwT y j x 2 Rn ; y 2 F .x/g wT v C supfwT c j c 2 C g D wT v;
which implies that (P .w; x/) and hence (P .wI y 1 ; : : : ; y s /) is bounded. So (a) is
proven.
Statement (b) follows from (14.20) taking into account that supc2C wT c D 0 for
w 2 C . t
u
Remark 14.2.8. The assumption that the halfspace H D fy 2 Rq j wT y g
contains the set F .x/ C C implies w 2 C such that it is not necessary to suppose
additionally w 2 C .
Lemma 14.2.9. Let (SP) be feasible and bounded and let x 2 dom F . If xO 2 Rn
is a solution of (P .w; x/) for some w 2 C , then ˛.w; u/ D ˇ.w; u/ for every
u 2 Rn with F .u/ F .x/.
O
Proof. It holds ˇ.w; u/ ˛.w; u/ (see (14.18). For u 2 Rn with F .u/ F .x/,
O we
obtain
˛.w; u/ D supfwT y j y 2 F .x/; F .x/ F .u/g

supfwT y j y 2 F .x/; F .x/ F .x/g
O D ˛.w; x/:
O
Taking into account that F .u/ F .x/

O can be written as F .u/ F .x/
O C C and, as
w 2 C , we get
ˇ.w; u/ D supfwT y j y 2 F .u/g

supfwT y j y 2 F .x/g
O D ˇ.w; x/:
O
The point xO is supposed to be a solution of (P .w; x/), this implies that there
exists yO 2 F .x/
O such that ˇ.w; x/
O
wT yO D ˛.w; x/
˛.w; x/. O Altogether we
get ˛.w; u/ ˛.w; x/O ˇ.w; x/
O ˇ.w; u/ ˛.w; u/, which yields the desired
result. t
u
In [396, Theorem 4.4] Löhne and Schrage show that the algorithm SetOpt
terminates after finitely many steps:
Theorem 14.2.10. The algorithm SetOpt generates a solution set of (SP )
whenever (SP ) is feasible and bounded. Otherwise SetOpt states whether
(SP ) is infeasible or unbounded. If the representation
F .x/ C C D fy 2 Rq j .wj /T y j ; j D 1; ; rg
(see (14.19)) computed in Phase 2 contains no redundant inequalities, SetOpt

terminates after finitely many steps. To be more precise, let the pre-solution set
computed in the Phase 1 consists of k points and let l be the number of linear
inequalities necessary to describe the polyhedral convex set graph F C .0Rn C /.
In Phase 2 of SetOpt, at most l k linear programs have to be solved.
Proof. If (SP ) is infeasible or unbounded, so is (VecRel) and the algorithm
terminates with the corresponding status. If (SP ) is feasible and bounded, a
pre-solution set fx 1 ; ; x k g, k
1 of (SP ) is obtained by solving (VecRel).
For fixed i 2 f1; ; kg, denote by x i the value of the variable x i before the while
loop has been entered and let xO i be the value after the while loop has been left.
Then, X WD fx 1 ; ; x k g is the pre-solution set computed in Phase 1 and XO WD

fxO 1 ; ; xO k g is the result of the algorithm. We will show that for all i 2 f1; ; kg,
F .xO i / F .x i / and 6 9x 2 Rn W F .x/ F .xO i /: (14.22)
In order to show (14.22), let i 2 f1; ; kg be fixed. The constraint

F .x/ F .x i / of the equivalent formulation .P .wj ; x i // of the linear program
.P .wj I y 1 ; ; y s // directly implies the first condition of (14.22). In the case
where (SP ) is feasible and bounded there always exists a minimal solution
of .P .wj I y 1 ; ; y s // by Lemma 14.2.7. The while loop is left only after flag
has been set to zero. This requires r iterations in the inner for loop, which occurs
only if for every j 2 f1; ; rg, ˛ D ˛.wj ; xO i / equals ˇ D ˇ.wj ; xO i /. In case of
wj =jjwj jj 2 K, this is known by Lemma 14.2.9, i.e., .P .wj I y 1 ; ; y s // does not
need to be solved. But ˛.wj ; xO i / D ˇ.wj ; xO i / for all j 2 f1; ; rg implies that xO i
is a minimal solution, compare Lemma 14.2.6. So the second condition in (14.22)
is shown.
The first condition in (14.22) implies
inf F .x/ inf F .x/:

x2XO x2X
Since X is a pre-solution set of (SP ), the infimum is attained in X , that is,
inf F .x/ D infn F .x/:

x2X x2R
It follows that the infimum is also attained in XO . The second condition in (14.22)
states that XO consists of only minimal solutions, whence XO is a solution set of
(SP ).
To show finiteness, note first that there is a finite algorithm (such as Benson’s
algorithm) to solve (VecRel) in Phase 1. Consider the set-valued map FQ W Rn Rq ,
FQ .x/ WD F .x/ C C . Of course, graph FQ D graph F C .0Rn C / is a polyhedral
convex set. Thus it can be expressed as (compare (14.16))
Q C By
graph FQ D f.x; y/ 2 Rn Rq j Ax Q
Q
bg;
for some AQ 2 Rln , BQ 2 Rlq , bQ 2 Rl . For every x 2 dom F , we have
Q
bQ Axg:
F .x/ C C D fy 2 Rq j By Q
Consequently, every representation of F .x/ C C given by (14.19) that contains

no redundant inequalities consists of at most l inequalities. With other words, there
are at most l different vectors wj =jjwj jj for j 2 f1; ; kg in the algorithm, and
the proof is completed. t
u
Theorem 14.2.10 immediately implies the following existence result (see [396,
Corllary 4.5]).
Corollary 14.2.11. If (SP ) is feasible and bounded, a solution set of (SP )
exists.
In order to reduce the computational effort of the algorithm for specific problems,
an additional rule is suggested in [396]. Let f.x1 ; y1 /; ; .xk ; yk /g denote the
solution set of (VecRel) obtained in Phase 1 and consider iteration i 2 f1; ; kg of
the outer for loop in Phase 2:
If yj 2 F .xi / for j with i < j k then skip all commands in iteration j :
Obviously, this rule maintains the attainment of the infimum and thus the
algorithm still works correctly.
Finally, we discuss the special case where F W Rn ! GC , i.e., we consider
F .x/ D F .x/ C C for all x 2 Rn :
In this case, (14.21) can be replaced by
F .x/ F .x/ ” F .x/ F .x/

” 8i D 1; ; s W y i 2 F .x/
” 8i D 1; ; s W Ax
b By i :
” Ax
maxfb By i j i D 1; ; sg:
This means that the linear program (P .wI y 1 ; : : : ; y s /) has only n C q variables
and 2m constraints. The problem to get a representation of graph.F ./ C C / given
by (14.19 ) from a representation of graph F given by (14.16) seems to be difficult
in practice where it is typical that n is much more larger than q. One way to obtain
it is vertex enumeration of a polyhedral convex set in Rn Rq . In contrast, SetOpt
involves vertex enumeration only in Rq .
Chapter 15
Applications
15.1 Set-Valued Approaches to Duality in Vector

Optimization
Applying the duality assertions for set-valued problems given in Sect. 8.3 we derive
duality statements for vector optimization problems in this section.
In the literature there are several approaches to construct a dual problem for a
given vector optimization problem. For instance, Luc [402] distinguishes between
conjugation, Lagrangian and axiomatic Duality. However, there seems to be no
unified approach to dualization in vector optimization. One of the difficulties is in
the fact that the minimal solution in multi-objective optimization is not necessarily
a single element, but in general becomes a subset of the image space. The definition
of infimum (or supremum) of a set with partial order plays a key role in development
of duality theory in multi-objective optimization. An interesting discussion of these
aspects is given in the book by Pallaschke, Rolewicz [453] and in the paper by
Nakayama [437]. There are at least three main ideas which are used for overcoming
the difficulties that arise when generalizing well-known duality assertions from the
scalar optimization theory to the vector-valued case.
The first one is the usage of scalarization in the formulation of the dual problem
(see Schönfeld [525], Breckner [78], Jahn [290,292]). In this approach, scalarization
concepts and corresponding duality assertions from real-valued optimization are
often used in order to derive useful dual problems, to prove duality assertions or in
order to solve the dual problem. In a lot of papers this procedure is used in the proofs
too (see e.g. [52]). However, as shown in [234], this approach has the disadvantage
that even in the case of linear vector optimization a duality gap may occur, although
the usual assumptions are fulfilled.
A second category of dual problems is based on the observation that a dual
vector optimization problem is naturally set-valued (see Tanino, Sawaragi [568],
Kawasaki [316], Corley [111], Tanino [563, 566], Luc [402], Nakayama [437],

DOI 10.1007/978-3-642-54265-7__15
664 15 Applications
Tammer [557], Dolecki, Malivert [145], Pallaschke, Rolewicz [453], Song [536],
Hamel, Heyde, Löhne, Tammer, Winkler [234], Löhne, Tammer [397]). Duality
assertions for vector optimization problems are shown without a scalarization “from
the beginning”. Instead, the dual problem becomes set-valued. For instance, in
the paper by Tanino [566] the set-valued structure of the primal and dual vector
optimization problem is taken into account: Embedding the primal problem into
a family (depending from perturbation parameters) of set-valued optimization
problems and applying an extension of Fenchel’s inequality, Tanino derives a weak
duality assertion and using the relationship between a map and its biconjugate
he shows a strong duality statement. Furthermore, in the paper by Dolecki,
Malivert [145] a set-valued approach in combination with Lagrangian techniques
and perturbations of marginal relations is used in order to show duality assertions
for general vector optimization problems where the solution concept is described by
a transitive, translation-invariant relation.
A third type of dual problems is based on solution concepts with respect to the
supremum and infimum in the sense of a vector lattice. In the book by Pallaschke
and Rolewicz [453], special conditions concerning the order in the image space
are supposed in order to prove duality assertions. A duality theory for objective
functions with values in vector lattices is developed in [453] using corresponding
notions of infimum and supremum in the sense of utopia minimum (maximum).
However, these infima and suprema may not be solutions in the sense of vector
optimization (Pareto) because they may not belong to the set of the objective
function’s values. In [453] notions of vector-convexity, vector-subgradients and
vector duality are introduced. These notions are analogous to the scalar case taking
into account the order in the vector lattice.
In the paper by Nieuwenhuis [442] solution concepts on the basis of infimal
(supremal) sets are introduced because the assumption that the objective func-
tion has its values in a vector lattice is too restrictive for vector optimization.
Nieuwenhuis [442] and Taninio [563, 566] derived duality assertions to these
solution concepts. These infimal sets are closely related to weakly minimal points
(see Definition 2.4.2). Dolecki and Malivert [145] extended these concepts to infimal
sets being closely related to other kinds of efficiency, too.
In contrast to the these investigations the approach in this section (see Löhne,
Tammer [397]) is characterized by an embedding of the image space of the
vector-valued problem into a complete lattice without linear structure, namely a
sublattice of the power set of the image space (compare Sect. 2.6.3). On the one
hand, our primal and dual problems are set-valued and therefore related to the
problems of the second type. The infimal (supremal) sets of Nieuwenhuis [442]
are involved into the definition of the infimum (supremum) in the lattice which
yields a relationship to the papers by Nieuwenhuis [442] and Taninio [563,566]. On
the other hand, a consequent usage of the lattice structure yields that we can carry
over the formulations, statements and proof techniques from the scalar optimization
theory, even though our assertions are for solution concepts in the sense of vector
15.1 Set-Valued Approaches to Duality in Vector Optimization 665
optimization (weakly minimal points, see Definition 2.4.2). In [442, 563, 566] the
construction of the dual problem is not completely analogous to the scalar case
because the lattice structure is not taken into account.
Furthermore, we point out the relationship between duality in vector optimization
and duality for set-valued problems based on the ordering relation “set inclusion”,
in the sequel called set inclusion problems (this approach is also based on the lattice
structure of the image space), which were investigated in [393,394] (see Sect. 2.6.3).
In this section we discuss the relationship between a vector optimization problem
and set-valued optimization problems with F -valued and I -valued objective maps
(see Definitions 2.6.36 and 2.6.37). Since F and I are complete lattices (see
Proposition 2.6.40) we can assign to a given problem a dual problem following
the lines of scalar duality theory. In Sect. 8.3 we have shown that under the usual
assumptions weak as well as strong duality assertions for F -valued and I -valued
problems can be obtained. By the fact F and I being isomorphic and isotone
(Proposition 2.6.39) it is clear that it is sufficient to prove the duality assertions just
for one case, either for the F -valued case or for the I -valued case.
The advantage of the F -valued case is that the operations, the ordering and the
infimum and supremum have an easier structure, which is beneficial for proofs. In
this case we speak about set inclusion problems.
The advantage of the I -valued case is that it is closely related to vector
optimization problems, therefore let us speak about vector optimization problems
in this case.
This relationship can be seen as follows. Assume that .Y; / is a partially ordered
linear topological space, where the order is induced by a proper pointed closed
convex cone C with int C ¤ ;. For an arbitrary set X , consider a set-valued
objective map F W X Y . Using the notations introduced in Sect. 2.6.3, especially
Definition 2.6.43, we study the vector optimization problem
[
PN WD Inf F .x/: (SVP)
x2X
This means (compare Definition

S 2.6.31) that we are looking for the set of weakly
minimal points of ClC S x2X F .x/ with respect to C (see Definition 2.4.2). As
usual, if putting F .X / D x2X F .x/ we have the following relationships.
Because of Proposition 2.7.5 we have
Inf F .X / D WMin.F .X / C C; C /
under the assumption that F .X / C C is closed.

Furthermore, under the assumption that F .X / D cl.F .X / C C / it holds (see
Proposition 2.7.4)
WMin.F .X /; C / D WMin.cl.F .X / C C /; C / D Inf F .X /:

666 15 Applications
S S
Since Inf x2X F .x/ D Inf x2X Inf F .x/, (SVP) can be expressed as an I -
valued problem (see Sect. 2.6.3); without loss of generality we can assume that
the sets F .x/ are self-infimal, i.e., F W X ! I . Thus we consider the following
problem (compare Definition 2.6.43)
[
PN D Inf F .x/ D inf F .x/: (SP I )
x2X
x2X
We assign to (SP I ) a dual problem (SD I ): Let V be a set and D W

V I,
[
DN WD Sup D.v/ D sup D.v/: (SD I )
v2V v2V
As usual (see Sect. 8.3.2), we speak about weak duality between (SP I )
and (SD I ) when DN 4I PN and we speak about strong duality when DN D
PN . In contrast to the primal problem the self-infimality of the values of the dual
O
objective function D./ plays an important role. If we replace a value D.v/ by D.v/
O O N
with Inf D.v/ D Inf D.v/ but D.v/ being not self-infimal, D might be changed.
Therefore we shall understand (SD I ) itself as the dual problem to (SVP).
Remark 15.1.1. The vector optimization problem (SVP) means that we compute
weakly minimal elements of the upper closure (elements belonging to the infimal
set) of the image set with respect to the proper pointed closed convex cone C
with nonempty interior (compare Sect. 2.6.3). We express the vector optimization
problem by the I -valued problem (SP I ) in order to use a complete lattice
structure and to derive assertions analogously to the scalar optimization theory.
In the dual problem (SD I ) we study the problem to determine weakly
maximal elements of the lower closure (elements belonging to the supremal set)
of the dual image set with respect to the cone C .
Also in the case that the primal problem (SVP) is single-valued, the dual
problem (SD I ) is always a set-valued problem. Indeed, by many authors (see
Corley [111], Tanino [563,566], Luc [402], Dolecki, Malivert [145], Tammer [557],
Pallaschke, Rolewicz [453]) it was observed that the dual of a vector optimization
problem is “naturally” set-valued. Taking into account the space I and its lattice
structure we give an explanation for this by our approach.
Remark 15.1.2. In case of D; N PN Y , the weak duality inequality DN 4I PN can
be equivalently expressed by .DN int C / \ PN D ;, what is a well-known relation
in vector optimization. Indeed, DN 4I PN is equivalent to PN Cl C DN and by
Proposition 2.6.33 (iv), (v), (vi) this is equivalent to .DN int C / \ PN D ;.
The reformulation of the vector optimization problem (SVP) as an I -valued
problem and the relationship to a set inclusion problem, in our setting an F -valued
problem (see Definition 2.6.36), is discussed and used for deriving Fenchel duality
assertions in Sect. 15.1.1 (see Löhne, Tammer [397]). Furthermore, considering the
reformulation of the given primal problem (SVP) as an I -valued problem (see

Definition 2.6.37) we derive corresponding Lagrange dual problems in Sect. 15.1.2.
Finally, in Sect. 15.1.3 we study the special case of linear vector optimization
problems and apply our duality statements in order to show corresponding results for
the linear case. So it is possible to understand better the structure of the dual linear
vector optimization problem in comparison with results in the paper by Isermann
[287].
15.1.1 Fenchel Duality for Vector Optimization Problems

Using Corresponding Results for F -Valued Problems
In this section we consider the vector optimization problem (SVP) for the special
case Y D Rq , partially ordered by a proper pointed closed convex cone C Rq
with int C ¤ ; and X D Rn . We derive duality assertions for (SVP) using the
duality statements that we have shown in Theorem 8.3.4 for F -valued (compare
Definitions 2.6.36) problems following the paper by Löhne and Tammer[397]. Since
F is a complete lattice (Proposition 2.6.40) we can assign to a given problem a
dual problem following the lines of scalar duality theory. We know that F and I
being isomorphic and isotone (Proposition 2.6.39) and so it is sufficient to prove
the duality assertions just for one case, either for the F -valued case or for the
I -valued case. So we proceed as follows: We take the reformulation of (SVP)
as I -valued problem (SP I ), use the relationships between I -valued and
F -valued problems and apply duality assertions for F -valued problems given in
Theorem 8.3.4. As already mentioned, it is beneficial for the proofs of the duality
assertions that in the F -valued case the operations, the ordering and the infimum
and supremum have an easier structure.
So it is useful to consider a dual pair of F -valued problems simultaneously. For
p W X ! F ; p.x/ WD Cl C F .x/ and d W V ! F ; d.u/ WD Cl C D.u/ consider
the problems
[
pN WD cl p.x/ D inf p.x/; (SP F )
x2X
x2X
\
dN WD d.u/ D sup d.u/: (SD F )
u2V u2V
We define weak and strong duality for the F -valued problems (SP F ) and
(SD F ) like in Sect. 8.3.1, i.e., by dN pN (that is dN 4F p)
N and dN D p, N
respectively.
Of course, by Proposition 2.6.39 we have PN D Inf p,N DN D Inf dN , Cl C PN D pN
N
and Cl C DN D d and we have weak duality between (SP I ) and (SD I ) if
and only if we have weak duality between (SP F ) and (SD F ), and we have
668 15 Applications
strong duality between (SP I ) and (SD I ) if and only if we have strong
duality between (SP F ) and (SD F ), i.e., we have
DN 4I PN ” dN p;
N
DN D PN ” dN D p:
N
In Sect. 8.3.1 we have shown weak and strong duality assertions for F -valued
problems (see Theorem 8.3.4). Now, we will apply these results for deriving duality
assertions for vector optimization problems.
Let F W X Rq and G W Z Rq be two set–valued maps, A W X ! Z is a
linear map, where X D Rn and Z D Rm (as in Sect. 8.3.1). We now consider the
following vector optimization problem
[
PN WD Inf F .x/ C G.Ax/ : (SVPF )
x2X
In the same manner as at the beginning of this section, we do not loose generality
if F and G are considered to be functions F W X ! I , G W Z ! I and the
vector optimization problem (SVPF ) can be equivalently expressed as an I -valued
problem
[
PN D Inf F .x/ ˚I G.Ax/ : (SPF I )
x2X
We define the conjugate (compare [566]) of a function F W X ! I (with respect

to c 2 Rq ) by
[
Fc W X ! S ; Fc .x / WD Sup hx ; xi fcg F .x/ ;
x2X
where S is the space of self-supremal sets with respect to C , which is defined

analogously to I . Following the lines of scalar optimization we consider the
following dual problem:
[
DN c WD Sup Fc .AT u / ˚I Gc .u / : (SDcF I )
u 2Z
In view of Proposition 2.6.39 this is just a transformation of the dual pair of

problems considered in Sect. 8.3.1 into the framework of functions with self-infimal
values. Therefore we immediately obtain the same duality assertions.
Remark 15.1.3. Note that convexity of a function F W X ! I (see Defini-
tion 8.3.8) is equivalent to the so-called C -convexity (see e.g. Jahn [292, 293],
Luc [402] and Sect. 8.3.2) of the corresponding set-valued map FN W X Rq ,
FN .x/ WD F .x/ if F .x/ Rq , FN .x/ WD ; if F .x/ D fC1g, F N
.x/ WD R if
q
N
F .x/ D f1g, where C -convexity means that graph cl F ./ C C is convex.
In the following theorem we will show duality results for the I -valued problem
(SPF I ) (that is equivalent to the vector optimization problem (SVPF ) and the
I -valued dual problem (SDcF I ) using the results given in Theorem 8.3.4
for problems with F -valued objective functions taking into account the relations
between F -valued and I -valued problems (see Proposition 2.6.39).
Theorem 15.1.4 (Weak and Strong Duality Between (SPF I ) and
(SDcF I )). For all c 2 Rq it holds weak duality between (SPF I ) and
(SDcF I ), i.e., DN c 4I PN . If F and G are convex, 0 2 r-int.dom G A dom F /
and c 2 int C , then we have strong duality between (SPF I ) and (SDcF I ),
i.e., DN c D PN ; if additionally PN ¤ f1g, we have
[
DN c D WMax. Fc .AT u / ˚I Gc .u / ; C /:
u 2Z
Proof. Consider the F -valued dual pair (SP F ), (SDF F ) from Section
8.3.1 and denote the corresponding F -valued objective functions by p and dc .
The first part (weak and strong duality) follows from Theorem 8.3.4 and the
considerations in Sect. 2.6.3, in particular Proposition S2.6.39. For the second part let
PN ¤ f1g, hence ; ¤ Cl C PN ¤ Rq . Let yN 2 Sup u 2Z Dc .u / D DN c D PN D
Inf p.N Then yN 62 pCint
N C . By a separation theorem, there exists some yN 2 C nf0g

such that hyN ; yi N
pCint
N C .yN / D pN .yN / C int C .yN / D pN .yN /. By (8.13) there

exists some uN 2 Z such that hyN ; yi N
dc .Nu / .yN /. Assuming that yN 2 dc .Nu / C

int C we obtain hy ; yi N < dc .Nu / .yN / for all y 2 C n f0g, a contradiction. Hence

yN 62 dc .Nu / C int C . On the other hand, yN C int C 2 Inf pN C int CSD pN C int C
dc .Nu / C int C . This yields that
S yN 2 Inf dc.Nu / D Dc .Nu / u 2Z Dc .u /.

Together we have yN 2 WMax. u 2Z Dc .u /; C /. t

u
Remark 15.1.5. Our dual problem (SDcF I ) differs from the dual problem given
by Tanino [566] because of the formulation taking into account the complete lattice
structure of the space I of self-infimal sets. So our formulation is completely
analogous to that of scalar Fenchel duality theorems, that means our dual objective
function is expressed in terms of the conjugates. Note further that the origin space
of the dual problem is just the space Z instead of the space of linear continuous
operators L .Z ; Rq /.
Remark 15.1.6. In scalar optimization the primal and dual attainment of the solu-
tion are of interest in many cases. Even though we cannot expect that the infimum
and supremum are attained by a single x 2 X and u 2 Z , respectively, we speak
about primal and dual attainment if we can replace the infimal set by the set of
minimal (weakly minimal) elements in (SPF I ) and the supremal set by the set
of maximal (weakly maximal) elements in (SDcF I ), respectively. Concerning
the primal attainment let us mention that even in scalar convex programming the
infimum is not attained without additional assumptions. Therefore, also in the case
of vector optimization one needs additional assumptions. In the paper by Dolecki,
Malivert [145] and in the book by Pallaschke, Rolewicz [453] the submission
(domination) property is supposed. This property guarantees the existence of
670 15 Applications
minimal points. Using the infimal set in the primal problem, we do not need such a
domination property.
As in the scalar theory the dual attainment in Theorem 15.1.4 follows without
any additional assumption.
Remark 15.1.7. In many papers on duality for vector optimization problems
(compare Jahn [292]) a strong direct and a strong inverse duality assertion is
shown. In the formulation of the strong direct duality assertion the existence of a
weakly minimal solution of the primal vector optimization problem is supposed
and the existence of a weakly maximal solution of the dual problem with the
same objective function value is shown. Conversely, in the strong inverse duality
assertion a closedness condition concerning the primal problem and the existence
of a weakly maximal solution of the dual problem are supposed in order to show
the existence of a weakly minimal solution of the primal problem with the same
objective function value. Such formulations like strong direct and inverse duality
assertions are discussed in Sect. 15.1.3 (Corollaries 15.1.20, 15.1.21) for the case
of linear vector optimization problems, but could be also obtained for the convex
setting, as a conclusion of Theorem 15.1.4.
15.1.2 Lagrange Duality for Vector Optimization Problems

Based on Results for I -Valued Problems
In this section we apply the duality statements shown in Sect. 8.3.2 for deriving
Lagrange duality assertion for set-valued vector optimization problems following
Hernández, Löhne, Rodríguez-Marín, Tammer [247]. We study the primal and dual
problem introduced by Li, Chen and Wu [379] and derive relationships to the dual
problems discussed in Sect. 8.3.2 using the solution concepts of weakly minimal /
maximal elements for vector optimization problems (see Definition 2.4.2).
Let X be a linear space, Y be a separated locally convex space partially ordered
by a proper pointed closed convex cone C such that ; ¤ int C ¤ Y and let Z be a
separated locally convex space. Moreover, let hZ; Z i be a dual pair. Let F W X
Y be a set-valued map, let G W X Z be a set-valued map with dom G D X
and let K Z be a proper closed convex cone. We use the notations introduced in
Sect. 2.6.3, especially in Definition 2.6.31.
We consider a set-valued vector optimization problem
[
PN WD Inf F .x/ (SVPS )
x2S
where
S D fx 2 X j G.x/ \ K ¤ ;g
(compare problem (SP) in Sect. 8.4.1).

From the relation
[ [
Inf F .x/ D Inf Inf F .x/;
x2S x2S
and setting
FI W X ! I ; FI .x/ WD Inf F .x/;
we can reformulate (SVPS ) as an I -valued problem
PN W D inf FI .x/: (P)

x2S
Denote by L WD L .Z; Y / the set of all linear continuous operators from

Z to Y and by LC the subset of all positive operators, that is, LC W D fT 2
L j T .K/ C g.
Thus, according to Sect. 8.3.2 we can associate to problem (P) two dual problems
as follows: The first one is formulated as
DN c WD sup c .u /; (SVDCL )

u 2Z
where
c .u / WD inf Lc .x; u /

x2X
and (compare (8.18))
Lc .x; u / WD FI .x/ ˚ inf fhu ; uifcg C bd C g:

u2G.x/CK
Furthermore, the second one is given by
DQ WD sup .T /: (SVDTL )
T 2LC
where
.T / WD inf L.x; T /
x2X
and (compare (8.22))
L.x; T / D FI .x/ ˚ inf fT .u/ C bd C g:

u2G.x/CK
From now on without loss of generality we can assume that X D dom FI .

672 15 Applications
Remark 15.1.8. We emphasize that problem (P) has been studied in several papers,
for instance, Tanino [566], Song [535], Li et al [379] and Chen et al [89] to give
conjugate duality. The authors consider an objective map F from X to Y [ fC1g
but they work on Dom F D fx 2 X j F .x/ ¤ C1 and F .x/ ¤ ;g. Note that the
corresponding I -valued map FI defined in problem (P) satisfies that Dom FI D
fx 2 X j F .x/ ¤ C1 and F .x/ ¤ ;g: For a comparison of different approaches
to duality in set-valued optimization see Sect. 8.4.
Li et al in [379] study problem (P) and the following dual problem
[
dNL W D WMax.fInf .F .x/ C T .G.x/// j T 2 LC g; C / (DL )
x2X
(denoting Max instead of WMax).

By the previous notations, the problem (DL ) can be understood as the following
I -valued problem
[
dL W D sup Inf .F .x/ C T .G.x/// :
T 2LC x2X
Now we show the relationships between problem (DL ) and the dual problems
studied in Sect. 8.3.2, especially the construction based on the Lagrangian map
given in (8.22). The next result is easy to check taking account some properties
established in Sect. 2.6.3.
Lemma 15.1.9. For any T 2 LC , the following conditions hold
inf fT .u/ C bd C g D inf fT .u/ C bd C g D Inf T .G.x//:

u2G.x/CK u2G.x/
Therefore,
inf fFI .x/ ˚ inf fT .u/ C bd C gg D inf fF .x/ C T .G.x//g

x2X u2G.x/CK x2X
[
D Inf .F .x/ C T .G.x/// :
x2X
From Lemma 15.1.9 we obtain the following relationship between (SVDTL ) and (DL )
dNL dL D D:
Q (15.1)
Consequently, the weak duality result presented in [379, Proposition 3.2 ] can be
deduced from Theorem 8.3.19 taking into account Lemmata 8.3.7 and 15.1.9.
In order to establish relationships between the strong result given in [379] and
the obtained results in Sect. 8.3.2 we need to consider the following stability notions
(compare Sect. 8.4).
As in Sect. 8.4.2 let ' be a set-valued map from X Z to Y defined by

F .x/ if G.x/ \ .K u/ ¤ ;
'.x; u/ WD
; else.
Again,
[ we use the perturbation map W W Z Y given by W .u/ WD
S
Inf '.x; u/ with W .0/ D Inf x2S F .x/ D Inf F .S /.
x2X
In [379, Definition 2.5] the following definition of positive subgradients is
introduced in order to formulate the stability of the problem (SVPS ), compare
Definition 8.4.5, where a corresponding formulation based on infimal sets is given.
Furthermore, the positive subgradients in Definition 8.4.5 are vectors, where as the
positive subgradients in the following definition are operators.
Definition 15.1.10. T 2 LC is a positive subgradient of W at .Nu; y/
N 2 graph W
(denoted by T 2 @C W .Nu; y/),
N if
[
T .Nu/ yN 2 WMax. .T .u/ W .u//; C /:
u2Z
The following definition for stable problems (SVP) is given in [379, Definition
5.2] (compare Definition 8.4.8).
Definition 15.1.11. The problem (SVPS ) is stable if W .0/ ¤ fC1g, W .0/ ¤
f1g, @C W .0; y/ ¤ ; for all y 2 W .0/.
Concerning the existence of positive subgradients in the sense of Defini-
tion 15.1.10 we get the following assertion (compare Lemma 8.4.7 for positive
subgradients in the sense of Definition 8.4.5).
Lemma 15.1.12. Let y 2 W .0/ D Inf F .S / and T 2 LC . Then T 2 @C W .0; y/ if
and only if
[
y 2 Inf .F .x/ C T .G.x///:
x2X
[
Proof. Since W .u/ D Inf '.x; u/ we have that T 2 @C W .0; y/ if and only if
x2S
[ [ [
y 2 WMax. .T .u/ W .u//; C / D WMax. .T .u/ Inf '.x; u//; C /
u2Z u2Z x2X
or equivalently
!
[ [
y 2 WMin. .Inf '.x; u/ T .u//; C /:
u2Z x2X
674 15 Applications
By definition of '.x; u/ we have

!
[ [
WMin. .Inf '.x; u/ T .u//; C /
u2Z x2X
!
[ [
D WMin. Inf F .x/ T .u/ ; C /
u2G.x/CK x2X
!
[ [
D WMin. Inf F .x/ C T .v/ ; C /
uDv2G.x/CK x2X
!
[ [
D WMin. Inf F .x/ C T .v/ ; C /
v2G.x/CK x2X
0 1
[ [
D WMin.Inf @F .x/ C T .v/A; C /
x2X v2G.x/
[
D WMin.Inf .F .x/ C T .G.x///; C //
x2X
[
D Inf .F .x/ C T .G.x/// :
x2X
t
u
Under certain convexity assumptions it is possible to show that the problem
(SVPS ) is stable (compare Theorem 8.4.11).
Theorem 15.1.13. If F G is .C K/-convex, G.dom F / \ int K ¤ ; and
W .0/ ¤ f1g, then (SVPS ) is stable.
Proof. Because of G.dom F / \ int K ¤ ; we have W .0/ ¤ fC1g. By
Lemma 15.1.12 we have to show that for each yN 2 W .0/ D Inf F .S / there exits
T 2 LC such that
[
yN 2 Inf .F .x/ C T .G.x///:
x2X
If yN D 1 then 0 2 @C W .0/.
Suppose that yN 2 Y . Then yN 2 WMin.Cl C .F .S //; C / Cl C .F .S //. By
hypothesis the map QW X Y Z defined by Q.x/ WD .F .x/; G.x// is .C K/-
convex. Thus, Q.X / C .C K/ is a convex set. Moreover, it is easy to check that
(compare the proof of Theorem 8.4.11)
.Q.X / C .C K// \ int .B .K// D ; (15.2)
N C . Indeed, if there exist x 0 2 X and .y; u/ with

where B D fyg

.y; u/ 2 .F .x 0 /; G.x 0 // C .C K/ \ int .B .K//
then y 2 .F .x 0 /CC /\.yN int C / and u 2 .G.x 0 /CK/\int K. Thus, y 0 D yN c 0

where y 0 2 F .x 0 / and c 0 2 int C and .G.x 0 / C K/ \ .K/ ¤ ; (that is, x 0 2 S ), a
contradiction to yN 2 Inf F .S /:
By (15.2), applying the Hahn-Banach theorem there exists a pair .
; / 2 Y

Z n f.0; 0/g such that
h
; yiCh; ui h
; biCh; d i for all .y; u/ 2 Q.x/C.C K/; x 2 X; b 2 B and d 2 K:
(15.3)
From this we deduce that .

; / 2 C K n f.0; 0/g: In particular, by (15.3)
we have
h
; yi C h; ui h
; yi
N for all .y; u/ 2 Q.x/ C .C K/; x 2 X: (15.4)
We show
¤ 0. On the contrary, ¤ 0 and by (15.4), we have
h; ui 0 for all u 2 G.x/; x 2 X:
On the other hand, by hypothesis there exists u0 2 G.x/ \ .int K/ then h; u0 i > 0
since 2 K n f0g. Hence,
¤ 0:
Consider T 2 L defined by
T .u/ WD h; ui c0 for u 2 Z
where c0 2 int C and h

; c0 i D 1 (note that c0 there exists since
2 C n f0g).
It is clear that T 2 LC . Moreover, we have
u 2 G.x/ \ K ) T .u/ D kc0 being k

0: (15.5)
Case I. yN 2 F .S /. Then there exists xN 2 S such that yN 2 F .x/.

N By (15.4) we
have h; ui 0 for all u 2 G.x/.
N In addition, h; uN i D 0 for uN 2 G.x/
N \ K
since 2 K :
Therefore, if uN 2 G.x/
N \ K, by (15.4) and taking into account that T .Nu/ D 0
we have
h
; y T .u/i h
; yN T .Nu/i for all .y; u/ 2 Q.x/ C .C K/; x 2 X:
Since T .Nu/ D 0, applying [566, Proposition 2.7] we deduce

[
yN 2 Inf .F .x/ C T .G.x//
x2X
and we conclude the proof.

676 15 Applications
Case II. yN 62 F .S /. Since yN 2 Inf F .S / we have yN 62 F .S / C int C and yN C

int C F .S / C int C . Thus, for each ˛ > 0 there exist x 0 2 S , y 0 2 F .x/ and
c 0 2 int C such that yN c 0 D y 0 ˛c0 .
By (15.3) we obtain
h
; yi C h; ui h
; yN c 0 i for all .y; u/ 2 Q.x/ C .C K/; x 2 X
or equivalently for any ˛ > 0
h
; yi C h; ui h
; y 0 ˛c0 i D h
; y 0 i C h
; ˛c0 i
for all .y; u/ 2 Q.x/ C .C K/; x 2 X: (15.6)
Let u0 2 G.x 0 / \ K (note that x 0 2 S ). Taking into account (15.5) there exists
˛0 > 0 such that T .u0 / D ˛0 c0 : From this and (15.6) we have
h
; yiCh
; T .u/i h
; y 0 iCh
; T .u0 /i for all .y; u/ 2 Q.x/C.C K/; x 2 X:
Since h
; T .u0 /i D h; u0 i, by (15.4), we obtain h
; y 0 i C h
; T .u0 /i h
; yi.
N
Hence,
h
; yi C h
; T .u/i h
; yi
N for all .y; u/ 2 Q.x/ C .C K/; x 2 X:
[
To conclude that yN 2 Inf .F .x/ C T .G.x// applying [566, Proposition 2.7]
x2X
it is sufficient to show that
[
yN 2 Cl C .F .x/ C T .G.x//
x2X
[
which follows easily from Cl C F .S / Cl C .F .x/ C T .G.x// (taking into
x2X
account that F .x/ F .x/ C T .u/ C int C for x 2 S and u 2 G.x/ \ K) and
yN 2 Cl C F .S /. t
u
Note that the assumptions given in [379, Proposition 5.3 ] are more stronger that
those presented in the Theorem 15.1.13.
Consequently, we obtain the following result which improves [379, Theorem 5.2]
and [535, Theorem 3.1]. Furthermore, we get relationships to the dual problems
constructed in Sect. 8.3.2, especially to the dual problem (SDTL I ) with DQ WD
supT 2LC .T / and the dual problem (SD0c N
L I ) with Dc WD supu 2K C c .u /

T C
(see problems (SVDL ) and (SVDL ) in this section).
Theorem 15.1.14. Let c 2 int C . If F is C -convex, G is K-convex, G.dom F / \
int K ¤ ;, then
PN D DQ D DN c D dNL D dL :
Proof. By Theorem 15.1.13 applying [379, Theorem 5.1] we have PN D dNL D dL .

On the other hand, by Lemma 8.3.9 and Theorems 8.3.15 and 8.3.21 we have PN D
DN c D D:
Q t
u
As conclusion, it is possible to simplify or to reduce the feasible set of several
dual problems studied in the literature by considering problem (P) where the
objective function is an I -valued function.
15.1.3 Duality Assertions for Linear Vector Optimization

Based on Lattice Approach
In this section we investigate the special case of linear vector optimization problems
(see Löhne, Tammer [397], Heyde, Löhne, Tammer [258, 259]). We show that
we can maintain a large part of the structures of scalar linear programming. In
particular, we have no duality gap in the case that the right-hand side of the
inequality constraints is zero, compare the discussion in [234]. On the other
hand, our dual objective maps take their values in the space of self-infimal sets
(compare Definition 2.6.37), in particular, we cannot expect a single-valued dual
objective map. Nevertheless the values of the dual objective map have a simple
structure, namely they are boundary points of translated cones,
S in some cases even
hyperplanes. As usual, we use the abbreviation f .S / D x2S f .x/. Suppose that
C Rq is a proper closed convex cone with nonempty interior.
Consider the following linear vector optimization problems using the notation
given in Sect. 2.6.3, especially the infimal set Inf A of a set A Rq (see
Definition 2.6.31):
Inf M.S /; S WD fx 2 Rn j Ax
bg ; (LP1 )
Inf M.S /; S WD fx 2 Rn j x
0; Ax
bg ; (LP2 )
Inf M.S /; S WD fx 2 Rn j x
0; Ax D bg ; (LP3 )
where M 2 Rnq ; A 2 Rnm ; b 2 Rm . We calculate the corresponding

dual problems using the approach via conjugates presented in Sect. 15.1.1 and
apply Theorem 15.1.4 following Löhne, Tammer [397]. Exemplarily we show the
calculations starting with (LP2 ). Starting with (LP1 ) or (LP3 ) we obtain the result
similarly and even easier. We set

fMxg C bd C if x
0 bd C if u
b
F .x/ WD and G.u/ WD
fC1g else fC1g else.
678 15 Applications
For the choice c 2 int C , an easy calculation shows that

T
Fc .x / D Inf.M cx / RnC ;

hu ; bi fcg C bd C if u
0
Gc .u / D
f1g else.
In order to obtain dual side conditions of a simple structure, it is useful to

characterize the condition Inf.M c x T /.RnC / D f1g. For this purpose consider
the next two assertions.
Lemma 15.1.15. Let C1 ; C2 Rq be nonempty closed convex cones with int C2 ¤
;. Then
C1 \ int C2 D ; ” C1 \ .C2 n f0g/ ¤ ;:
Proof. Note that y 2 int C2 , z 2 .C2 n f0g/ implies that hy; zi > 0. On the other
hand, y 2 C1 , z 2 C1 implies that hy; zi 0. This proves the implication “(”.
If C1 \ int C2 D ;, by a separation theorem (e.g. [489, Theorem 11.3]),
we obtain that there exists y 2 Rq n f0g such that infy2C2 hy ; yi D
supy2C1 hy ; yi D 0. Hence we have y 2 C1 \ .C2 n f0g/. t
u
Proposition 15.1.16. Let K Rn be a nonempty closed convex cone and let H 2
Rqn . Then, the following statements are equivalent:
(i) Inf H.K/ ¤ f1g;
(ii) 9c 2 C n f0g W H T c 2 K .
Proof. Set L WD H.K/. Since L Rq is a nonempty cone, we have
Inf L ¤ f1g ” 9y2Rq ; 8˛ > 0 W ˛y … LCint C ^ ˛yCint C LCint C

” 0 2 Inf L:
By Lemma 15.1.15 we have
0 2 Inf L ” 0 62 LCint C ” L\int C D ; ” L \.C nf0g/ ¤ ;:
Using the bipolar theorem [489, Theorem 14.1] we can conclude
L \ .C n f0g/ ¤ ; ” 9c 2 C n f0g W 8l 2 L W hc ; li 0
˝ ˛
” 9c 2 C n f0g W 8x2K W hc ; H xi D H T c ; x 0
” 9c 2 C n f0g W H T c 2 K :
Together we obtain the desired assertion. t

u
C −C−
Bc C = −C− = Rq+
c Bc
c
Fig. 15.1 Two examples for the set Bc
Applying the previous result for c 2 int C we can write

(
Inf.M cx T / RnC Rq if 9c 2 C n f0g W x M T c Tc c
Fc .x / D
f1g else.
˚
In the following, the set Bc WD c 2 C C j hc; c i D 1 is used to express the dual
side conditions. In Fig. 15.1 we illustrate this set by two examples (with kck D 1):
Proposition 15.1.17. For c 2 int C , the set Bc is a compact convex subset of Rq .
Proof. Obviously, Bc is closed and convex. Thus, it remains to show that Bc is
bounded. Assuming the contrary, we obtain a sequence cn 2 Bc with kcn k ! 1.
Without loss of generality we can assume that cn = kcn k ! cN 2 C C n f0g. Since
hcn = kcn k ; ci D 1= kcn k ! 0, it follows that hcN ; ci D 0. But cN 2 C C n f0g and
c 2 int C implies that hcN ; ci > 0, a contradiction. t
u
According to problem (SDcF I ) in Sect. 15.1.1 we obtain the dual problem to
(LP2 ) with c 2 int C as
[
DN c D Sup c u b C Inf.M c u A/.RnC / ;
T T
(LD2c )
u 2Tc
with
˚
Tc WD u 2 Rm j u
0; 9c 2 Bc W AT u M T c :
By a similar calculation we obtain the dual problems to (LP1 ) and (LP3 ) as

[
DN c D Sup c u b C Inf.M c u A/.Rn / ;
T T
(LD1c )
u 2Tc
680 15 Applications
with
˚
Tc WD u 2 Rm j u
0; 9c 2 Bc W AT u D M T c
and
[
DN c D Sup c u b C Inf.M c u A/.RnC / ;
T T
(LDc3 )
u 2Tc
with
˚
Tc WD u 2 Rm j 9c 2 Bc W AT u M T c :
In the next proposition we collect some properties of the dual feasible set Tc .
Proposition 15.1.18. The set Tc in .LD1c /-.LD3c / (where c 2 int C ) is always a
closed convex subset of Rm . Moreover, if C is polyhedral, then Tc is polyhedral,
too.
Proof. Exemplarily
˚ we give the proof for Tc in .LD3c /. The set Tc can be expressed
as Tc D u 2 R j AT u 2 M.Bc / RnC . Of course, M.Bc / RnC is a convex
m
set. By [489, Theorem 3.4], it follows that Tc is convex. Since Bc is compact

(Proposition 15.1.17), the set M.Bc / RnC is closed. As the map AT W Rm ! Rn
is continuous, Tc is closed. If C is polyhedral, we conclude that C , Bc , M.Bc /
and M.Bc / RnC are polyhedral, too. By [489, Theorem 19.3], it follows that Tc is
polyhedral. t
u
From the Fenchel duality result in Theorem 15.1.4 and some additional consid-
erations we obtain the following duality assertion.
Theorem 15.1.19 (Weak and Strong Duality Between .LPi / and .LDic / (i D
1; 2; 3)). For c 2 int C it holds weak and strong duality between .LPi / and .LDic /
(i D 1; 2; 3). More precisely we have
(i) DN c D PN Rq if S ¤ ; and Tc ¤ ;, where “Sup” can be replaced by “Max”
in this case,
(ii) DN c D PN D f1g if S ¤ ; and Tc D ;,
(iii) DN c D PN D fC1g if S D ; and Tc ¤ ;.
Proof. In the case where S ¤ ; we obtain the strong duality by Theorem 15.1.4
taking into account Remark 8.3.6. In case (i), we have PN Rq [ f1g and DN c
Rq [fC1g, hence DN c D PN Rq . In case (ii) we have PN D DN c D Sup ; D f1g.
In case (iii) we need some additional considerations. Let uN 2 Tc and S D ;.
Exemplarily we show the assertion for .LP2 / and .LD2c /. By the Farkas Lemma
there exists some uO 2 Rm such that uO
0, AT uO 0 and b T uO > 0. Hence, for
all ˛ > 0 we have uN C ˛ uO 2 Tc . For ˛ ! C1 we have .Nu C ˛ uO /T b ! C1.
Hence DN c D fC1g D PN . t
u
Next we relate our result to classical formulations in vector optimization (see

Remark 15.1.6), i.e., we formulate the duality assertions for weakly minimal/weakly
maximal points introduced in Definition 2.4.2. Denoting the dual objective functions
in .LD1c /-.LD3c / by Dc W Rm ! I , .i D 1; 2; 3/, for all three pairs of dual problems
we easily obtain the following conclusions from Theorem 15.1.19.
Corollary 15.1.20 (Direct Strong Duality). For every c 2 int C it holds
WMin.M.S /; C / WMax.Dc .Tc /; C /:
Corollary 15.1.21 (Inverse Strong Duality). If M.S / is upper closed, then for all
c 2 int C it holds
WMax.Dc .Tc /; C / WMin.M.S /; C /:
Finally, we give an example.

Example 15.1.22. (see Fig. 15.2) Let q D m D n D 2, C D R2C and consider the
problem (LP2 ) with the data

10 12 2
M D ; AD ; bD :
01 21 2
Let us calculate the dual

˚ problem for c D .1; 1/T 2 int R2C . For this choice, we set
Bc is given by Bc D c1 ; c2
0 j c1 C c2 D 1 . The dual side conditions are as

follows
u1 ; u2
0; 9c1
0 W u1 C 2u2 c1 ; 2u1 C u2 1 c1 :
M(S) = M·x
x∈S Sup Dc (Tc )
Inf M(S)
Dc (v1 )
Dc (v2 )
Dc (v3 )
Fig. 15.2 The primal and dual values in Example 15.1.22

682 15 Applications
˚
This can be equivalently expressed by Tc D u1 ; u2
0 j u1 C u2 1=3 . The
vertices of Tc are the points v1 D .0; 0/T , v2 D .1=3; 0/T and v3 D .0; 1=3/T . The
matrices Hi WD .M c vi TA/ can be easily computed as
2 2 1 1
10 3
H1 D ; H2 D 1 1 ; H3 D 32 32 :
3
01 3 3 3 3
Hence we obtain
Inf H1 .R2C / D bd R2C ;

˚
Inf H2 .R2C / D y 2 R2 j y1 C 2y2 D 0 ;
˚
Inf H3 .R2C / D y 2 R2 j 2y1 C y2 D 0 :
Consequently, the values of the ˚ dual objective function at v1 ; v2 ; v3 are

D
˚ c .v 1 / D bd R 2
C , D c .v 2 / D y 2 R 2
j y1 C 2y2 D 2 and Dc .v3 / D
y 2 R2 j 2y1 C y2 D 2 . We see that the three dual feasible points v1 ; v2 ; v3 2 Tc
are already sufficient for strong duality.
15.1.4 Further Set-Valued Approaches to Duality in Linear

Vector Optimization
In this section we will derive further duality statements for linear vector optimization
problems where the solution concept is based on the concept of weakly minimal
points (see Definition 2.4.2) using the lattice approach. We now discuss a set-valued
dual problem to a linear vector optimization problem where also the scalarization
parameter c 2 Rq is considered as dual variable in difference to the construction of
the dual problem in Sect. 15.1.3. We will show duality assertions for a primal linear
vector optimization problem and a corresponding dual problem using a reformula-
tion of the primal problem based on self-infimal sets (see Definition 2.6.37). The
results presented in this section are derived in the paper by Heyde, Löhne, Tammer
[258].
Let m; n; q 2 IN and A 2 Rmn ; M 2 Rqn ; b 2 Rm be given. We consider the
following vector optimization problem
q q
WMin..M.S / C RC /; RC /; S WD fx 2 Rn j Ax
bg ; (LVP)
where M.S / D [x2S fMxg. This means that we are looking for weakly minimal
q q
points of the set .M.S / C RC / with respect to RC in the sense of Definition 2.4.2.
Furthermore, a point x 2 S is called a weakly efficient solution of (LVP) iff
0
q q q
Mx0 2 WMin.M.S /; RC / C RC or equivalently Mx0 2 WMin.M.S /; RC /:
Note the a point x 0 is a weakly efficient solution of (LVP) if and only if it is a weakly
efficient solution of the more common problem
q
WMin.M.S /; RC /; S WD fx 2 Rn j Ax
bg ;
q q q
even though the sets WMin.M.S /; RC / and WMin..M.S / C RC /; RC / can be
q q
different. The set WMin..M.S / C RC /; RC / is closely related to the infimal set of
M.S / (compare Sect. 2.6.3, Definition 2.6.31).
Put k WD .1; 1; : : : ; 1/T 2 Rq . Consider the following set-valued dual objective
map
˚
H W Rm Rq Rq ; H .u; c/ WD y 2 Rq j c T y D b T u :
The dual feasible set is given by

˚
U WD .u; c/ 2 Rm Rq j .u; c/
0; k T c D 1; AT u D M T c :
We use the following notation

[
H .U / WD H .u; c/:
.u;c/2U
Since this dual objective map is set-valued we use a corresponding solution concept,
namely, the concept introduced in Definition 2.6.5. As the dual problem to (LVP)
we consider the problem
q
WMax.H .U /; RC /: (LVD)
This means the dual problem consists in determining weakly maximal points of the
union of the hyperplanes H .u; c/ defined by the points .u; c/ 2 U in the sense
Definition 2.6.5.
Remark 15.1.23. The new idea in this approach compared with the results given in
Sect. 15.1.3 (compare also [397] and [259]) consists in having a pair .u; c/ of dual
variables and having hyperplanes as values of the dual objective without making
any assumptions on the rank of M .
A point .u0 ; c 0 / 2 U is called a weakly efficient solution of (LVD) iff
H .u0 ; c 0 / \ WMax.H .U /; C / ¤ ;;
or equivalently,
˚ 0 q
9y 0 2 H .u0 ; c 0 / W 8.u; c/ 2 U W y C int RC \ H .u; c/ D ;: (15.7)
684 15 Applications
This means that ..u0 ; c 0 /; y 0 / 2 U H.u0 ; c 0 / is a weak maximizer of (LVD)

(compare Sect. 2.6.1, Definition 2.6.5). Subsequently, weakly efficient solutions of
(LVP) or (LVD) are referred for simplicity as solutions of (LVP) or (LVD). The
notion of a solution of problem (LVP) as a feasible point whose image is weakly
minimal is common in vector optimization. We adapt this concept for the set-valued
dual problem by defining solutions of (LVD) as feasible points whose image, which
is a hyperplane, contains weakly maximal points (see Definition 2.6.5). Thus the
solution concept for the dual problem (LVD) coincides with the concept based on
vector approach introduced in Sect. 2.6.1.
In the following we prove weak and strong duality between the two problems
directly. In the proofs the following pairs of dual scalar linear optimization problems
depending on parameters c; y 2 Rq play an important role.
c T Mx ! min s.t. Ax
b; (P1 .c/)
b T u ! max s.t. u
0; AT u D M T c; (D1 .c/)
z ! min s.t. Ax
b; Mx kz y; (P2 .y/)
b T u y T c ! max s.t. u; c
0; AT u M T c D 0; k T c D 1: (D2 .y/)
The first pair of problems comes from classical linear scalarization and is mainly
used for characterizing solutions of (LVD). The second pair of problems is very
useful for characterizing weakly minimal and weakly maximal points in the image
space Rq . Similar problems also occur, for instance, in [286]. Note that the problem
(P2 (y)) also provide a very common scalarization method in vector optimization,
see e.g. [196, 454] and Chap. 5.
The following notion might also be useful for characterizing solutions of (LVP)
and (LVD). A pair of points .x; z/ 2 Rn R and .u; c/ 2 Rm Rq is called
complementary for the problems (P2 (y)) and (D2 (y)) if uT .Ax b/ D 0 and
c T .Mx kz y/ D 0.
Lemma 15.1.24. If .x; z/ 2 Rn R and .u; c/ 2 U are complementary points for
(P2 (y)) and (D2 (y)) then z D b T u y T c.
Proof. If .u; c/ 2 U we have k T c D 1 and AT u D M T c. Hence c T .Mxkzy/ D
0 and uT .Ax b/ D 0 imply z D c T Mx c T y D uT Ax c T y D uT b c T y. u
t
Subsequently, we use the following notation
q
M WD M.S /CRC D fy 2 Rq j 9x 2 S W M x yg ; F .u; c/ WD H .u; c/\M :
The following lemma can be interpreted as evidence of weak duality.

Lemma 15.1.25. If .u; c/ 2 U and y 2 M then c T y
b T u.
Proof. Since y 2 M there is some x 2 S such that y
Mx. Hence .x; 0/ is feasible
for (P2 (y)). Duality between (P2 (y)) and (D2 (y)) implies b T u y T c 0. t
u
Remark 15.1.26. The preceding lemma implies weak duality in classical sense,
i.e.,
x 2 S; .u; c/ 2 U ; y 2 H .U / H) Mx 6< y:
Indeed, if x 2 S; .u; c/ 2 U ; y 2 H .U / and Mx < y would be true, then c T Mx <

c T y D b T u but Lemma 15.1.25 implies c T Mx
b T u, a contradiction.
An interpretation of weak duality with the help of set relations is given later.
The next lemma states a sufficient optimality condition for (LVD), which is based
on weak duality.
q
Lemma 15.1.27. If .u0 ; c 0 /2U and y 0 2 F .u0 ; c 0 / then y 0 2 WMax.H .U /; RC /.
Proof. Let .u0 ; c 0 / 2 U and y 0 2 F .u0 ; c 0 /. Therefore we have y 0 2 H .U /. We
show that
˚ 0 q
y C int RC \ H .u; c/ D ; for all .u; c/ 2 U :
Assume on the contrary that there are .u; c/ 2 U and y 2 H .u; c/ with y > y 0 .
Since c
0; c ¤ 0 this implies c T y > c T y 0
b T u D c T y, a contradiction. u
t
Remark 15.1.28. The assertion of Lemma 15.1.27 says that under the given condi-
tions .u0 ; c 0 ; y 0 / is a weak maximizer of (LVD) in the sense of Definition 2.6.5.
The following theorem provides different characterizations of (weakly efficient)
solutions of (LVD).
Theorem 15.1.29. Let .u0 ; c 0 / 2 U . Then the following statements are equiva-
lent.
(i) .u0 ; c 0 / is a solution of (LVD),
(ii) u0 solves (D1 (c 0 )),
T
(iii) there exists some x 0 2 S with c 0 Mx0 D b T u0 ,
(iv) F .u ; c / is nonempty.
0 0
Proof. (i))(ii). Assume u0 does not solve (D1 (c 0 )). Then there is some u 2 Rm
such that .u; c 0 / 2 U and b T u > b T u0 . But for each y 2 H .u0 ; c 0 / we get
q
y C k.b T u b T u0 / 2 fyg C int RC \ H .u; c 0 /
contradicting (15.7), i.e., .u0 ; c 0 / being a solution of (LVD).

686 15 Applications
(ii))(iii). If u0 solves (D1 (c 0 )) then by duality between the problems (P1 (c 0 ))

T
and (D1 (c 0 )) there is some x 0 2 S such that c 0 Mx0 D b T u0 .
(iii))(iv). If (iii) holds then Mx0 2 H .u0 ; c 0 /. Since Mx0 2 M we have Mx0 2
F .u0 ; c 0 /.
(iv))(i). By Lemma 15.1.27. t
u
Our aim is to show a strong duality theorem in the sense that the set of weakly
minimal points for (LVP) and the set of weakly maximal points for (LVD) coincide.
In the proof of the strong duality assertion in Theorem 15.1.46 for I -valued
problems we will essentially use the following first strong duality result.
Theorem 15.1.30. The following four statements are equivalent.
q
(i) y 0 2 WMin.M ; RC /,
(ii) there is some x 0 2 Rn such that .x 0 ; 0/ solves (P2 (y 0 )),
T
(iii) there is some .u0 ; c 0 / 2 U with b T u0 D y 0 c 0 solving (D2 (y 0 )),
q
(iv) y 2 WMax.H .U /; RC /.
0
Proof. (ii))(i). If .x 0 ; 0/ solves (P2 (y 0 )), then x 0 2 S and Mx0 y 0 , hence y 0 2

M . Assume that there is some y 2 M (i.e., there is some x 2 S with Mx y)
with y < y 0 . Then there is some z < 0 such that y y 0 C kz. This implies
Mx kz y kz y 0 , i.e., .x; z/ is feasible for (P2 (y 0 )) and z < 0 contradicts the
optimality of .x 0 ; 0/.
q
(i))(ii). If y 0 2 WMin.M ; RC / then there exists some x 0 2 S with Mx0 y 0 ,
i.e., .x ; 0/ is feasible for (P2 (y )). Assume that there is some .x; z/ 2 RnC1 with
0 0
z < 0 being feasible for (P2 (y 0 )). Let y WD y 0 C zk then y < y 0 and Mx
y 0 C kz D y, i.e., y 2 M contradicting the weak minimality of y 0 .
(ii),(iii). By duality of (P2 (y 0 )) and (D2 (y 0 )).
q
(iii),(iv). We have y 0 2 WMax.H .U /; RC / iff
y 0 2 H .U / (15.8)
and
q
y 0 62 H .U / int RC : (15.9)
Condition (15.8) is equivalent to

T
9.u0 ; c 0 / 2 U W y 0 c 0 D b T u0 ; (15.10)
and (15.9) is equivalent to

T
8.u; c/ 2 U W y 0 c
b T u: (15.11)
Since (iii) is equivalent to (15.10) together with (15.11), the statement follows. t
u
Now we are able to prove the following theorem which provides sufficient condi-
tions for solutions of (LVP) and (LVD).
Theorem 15.1.31. Let .u0 ; c 0 / 2 U and x 0 2 S be given. Then x 0 is a solution
of (LVP) and .u0 ; c 0 / is a solution of (LVD) if one of the following equivalent
conditions is satisfied.
T
(i) b T u0 D c 0 Mx0 ,
(ii) u0 solves (D1 (c 0 )) and x 0 solves (P1 (c 0 )),
(iii) .x 0 ; 0/ solves (P2 (Mx0 )) and .u0 ; c 0 / solves (D2 (Mx0 )),
(iv) for all y 2 Rq there is some z0 2 R such that .x 0 ; z0 / and .u0 ; c 0 / are
complementary points for (P2 (y)) and (D2 (y)).
Proof. First we show the equivalence of the four conditions.
(i),(ii). By duality between (P1 (c 0 )) and (D1 (c 0 )).
(i),(iii). By duality between (P2 (Mx0 )) and (D2 (Mx0 )).
(i),(iv). If .u0 ; c 0 / 2 U , x 0 2 S and
z0 D b T u0 y T c 0 (15.12)
then we have
T T T
u0 .Ax0 b/ D c 0 .Mx0 kz0 y/ D c 0 Mx0 b T u0 : (15.13)
If (i) holds we define z0 by (15.12) and then (i) and (15.13) imply (iv). If (iv) holds
then (15.12) holds by Lemma 15.1.24 and then (iv) and (15.13) imply (i).
Now, sufficiency of these equivalent conditions for x 0 and .u0 ; c 0 / being
solutions of (LVP) and (LVD) follows from Theorem 15.1.29 and Theorem 15.1.30.
t
u
In the following we prove some statements showing the relationship between proper
faces (in particular facets) of M and solutions of (LVD). Let us recall some facts
concerning the facial structure of polyhedral sets. Let A Rq be a convex set. A
convex subset F A is called a face of A iff
y1; y2 2 A ;
2 .0; 1/;
y 1 C .1
/y 2 2 F ) y1; y2 2 F :
A face F of A is called proper iff ; ¤ F ¤ A . A set E ˚ A is called an exposed

face of˚ A iff there are c 2 Rq and 2 R such that A y 2 Rq j c T y
and
E D y 2 Rq j c T y D \ A . The proper .r 1/-dimensional faces of an r-
dimensional polyhedral set A are called facets of A . A point y 2 A is called a
vertex of A iff fyg is a face of A .
Theorem 15.1.32 ([599], Theorem 3.2.2). Let A be a polyhedral set in Rq . Then
A has a finite number of faces, each of which is exposed and a polyhedral set. Every
proper face of A is the intersection of those facets of A that contain it, and r-bd A
688 15 Applications
(the relative boundary of A ) is the union of all the facets of A . If A has a nonempty
face of dimension s, then A has faces of all dimensions from s to dim A .
Remark 15.1.33. If M ¤ ; then M is a q-dimensional polyhedral set, hence
the facets of M are the .q 1/-dimensional faces of M , i.e., the maximal (w.r.t.
inclusion) proper faces. A subset F M is a proper face iff it is a proper exposed
face, i.e., iff there is a supporting
˚ hyperplane H to M such that F D H \ M .
We call a hyperplane H WD y 2 Rq j c T y D (i.e., c ¤ 0) supporting to M iff
c T y
for all y 2 M and there is some y 0 2 M such that c T y 0 D .
˚
Lemma 15.1.34. If H D y 2 Rq j c T y D is a supporting hyperplane to M
then c
0.
Proof. If H is a supporting hyperplane to M then there is some y 0 2 M with
c T y 0 D and c T y
for all y 2 M . By definition of M we have y 0 C w 2 M
q q
for all w 2 RC , hence c T w
0 for all w 2 RC . This implies c
0. t
u
Lemma 15.1.35. A set F M is a proper face of M if and only if there is a
solution .u; c/ 2 U of (LVD) such that F D F .u; c/.
Proof. ‘(H”: If .u; c/ 2 U is a solution of (LVD) then there is some x 0 2 S such
that Mx0 2 H .u; c/, hence Mx0 2 F .u; c/. Moreover, if y 2 M then c T y
b T u
by Lemma 15.1.25. Consequently, H .u; c/ is a supporting hyperplane to M and
F .u; c/ is a proper face of M .
‘H)”: If F is˚ a proper face of M then there is some c 2 R n f0g; 2 R
q
such that H WD y 2 R j c y D is a supporting hyperplane to M and F D

q T
H \ M . By Lemma 15.1.34 we have c

0. Since c ¤ 0 we obtain k T c > 0.
Without loss of generality we can assume that k T c D 1. Since H is a supporting
hyperplane, we have c T y
for all y 2 M and c T y 0 D for some y 0 2 M .
Hence there is some x 0 2 S such that c T Mx0 D c T y 0 D , i.e.,
˚
D c T Mx0 D min c T Mx j x 2 S :
By duality between (P1 (c)) and (D1 (c)), problem (D1 (c)) has a solution u with
b T u D D c T Mx0 . Thus .u; c/ 2 U is a solution of (LVD) by Theorem 15.1.29,
and H .u; c/ D H .
Hence F D F .u; c/. t
u
q
Corollary 15.1.36. Each proper face of M is a subset of WMin.M ; RC /.
Proof. Let F be a proper face of M . By Lemma 15.1.35 there is a solution .u; c/ 2
U of (LVD) such that F D F .u; c/. Let y 2 F D F .u; c/, then y 2 M
(implying the existence of x 2 S such that Mx y, i.e., .x; 0/ is feasible for
(P2 (y))) and b T u D c T y. Duality between (P2 (y)) and (D2 (y)) implies that .u; c/ is
optimal in (D2 (y)) and .x; 0/ is optimal in (P2 (y)), hence y is weakly minimal by
Theorem 15.1.30. t
u
q
Corollary 15.1.37. WMin.M ; RC / ¤ ; if and only if ; ¤ M ¤ Rq .
Proof. This is a direct consequence of Corollary 15.1.36, Theorem 15.1.32 and the
fact that a nonempty set in A Rq has a nonempty boundary iff A ¤ Rq . t
u
The following lemma shows that facets of M may be described by extreme
solutions of (LVD) (i.e., solutions of (LVD) being a vertex of the feasible set U ).
Lemma 15.1.38. If F is a facet of M then there is an extreme solution .u0 ; c 0 / of
(LVD) such that F D F .u0 ; c 0 /.
Proof. Let
UN WD f.u; c/ 2 U j F .u; c/ D F g :
By Theorem 15.1.29, all points of UN are solutions of (LVD) because F is nonempty

as a facet of M . Let y 2 r-int F be arbitrary. Since F is a .q 1/-dimensional face
we have .u; c/ 2 UN if and only if .u; c/ 2 U and y 2 H .u; c/, i.e., b T u D y T c.
Hence UN D U \ Hy with
˚
Hy WD .u; c/ 2 Rm Rq j y T c b T u D 0 :
q
Since y 2 WMin.M ; RC / by Corollary 15.1.36, Theorem 15.1.30 implies that
Hy is a supporting hyperplane to U , hence UN is a nonempty face of U . Since
UN U RC contains no lines there is a vertex .u0 ; c 0 / of UN (see [489,
mCq
Corollary 18.5.3]). Hence .u0 ; c 0 / is also a vertex of U , i.e., an extreme solution of

(LVD). t
u
We define the following sets.
pFaces.M / WD fF M j F is a proper face of M g ;

Facets.M / WD fF M j F is a facet of M g ;
Sol.D/ WD f.u; c/ 2 U j .u; c/ is a solution of (LVD)g ;
ExtrSol.D/ WD f.u; c/ 2 Sol.D/ j .u; c/ is a vertex of U g :
Now we can extend the strong duality result in Theorem 15.1.30. Later we interpret
the following result as the attainment of the supremum in the dual problem in
extreme solutions.
Theorem 15.1.39. We have the following chain of equalities
q
[
WMin.M ; RC / D bd M D F .u; c/
.u;c/2ExtrSol.D/
q q
D WMax.H .ExtrSol.D//; RC / D WMax.H .U /; RC /:
690 15 Applications
Proof. Theorem 15.1.32, Lemma 15.1.38, Lemma 15.1.35 and Corollary 15.1.36
imply the following chain of inclusions
[ [ [
bd M D F F .u; c/ F .u; c/
F 2Facets.M / .u;c/2ExtrSol.D/ .u;c/2Sol.D/
[ q
D F WMin.M ; RC / bd M :
F 2pFaces.M /
Hence the first two equalities hold.

q q
The equality WMin.M ; RC / D WMax.H .U /;SRC / was already shown
in Theorem 15.1.30. Thus it remains to show that .u;c/2ExtrSol.D/ F .u; c/ D
q
WMax.HS .ExtrSol.D//; RC /.
If y 2 .u;c/2ExtrSol.D/ F .u; c/ then there exists some .u; c/ 2 ExtrSol.D/ such
that y 2 F .u; c/ D H .u; c/ \ M , i.e., y 2 H .ExtrSol.D//. Since .u; c/ is a
q q
solution of (LVD) we have .y C int RC / \ H .U / D ;, hence .y C int RC / \
q
H .ExtrSol.D// D ; implying y 2 WMax.H .ExtrSol.D//; RC /.
q
On the other hand, if y 2 WMax.H .ExtrSol.D//; RC / then y 2
H .ExtrSol.D// and
q
y 62 H .ExtrSol.D// int RC :
This is equivalent to
N 2 ExtrSol.D/ W y T cN D b T uN
9.Nu; c/ (15.14)
and
8.u; c/ 2 ExtrSol.D/ W y T c
b T u: (15.15)
By Theorem 15.1.29, uN solves (D1 (c)) N hence S ¤ ; by duality of (P1 (c)) N and
N
(D1 (c)).Thus the feasible set for (P2 (y)) is nonempty as well. Since .Nu; c/ N 2 U,
i.e., U ¤ ;, problem (D2 (y)) has an optimal solution .u0 ; c 0 / being a vertex
of U . Optimality of .u0 ; c 0 / for (D2 (y)) implies optimality of u0 for (D1 (c 0 ))
hence .u0 ; c 0 / 2 ExtrSol.D/ by Theorem 15.1.29. Now, (15.15) implies that
y T c 0
b T u0 . Moreover, optimality of .u0 ; c 0 / for (D2 (y)) implies b T u0 y T c 0
b T uN y T cN D 0, i.e., y T c 0 D b T u0 . Consequently we haveS y 2 H .u ; c / and

0 0
q
y 2 WMin.M ; RC / M by Theorem 15.1.30, i.e., y 2 .u;c/2ExtrSol.D/ F .u; c/.
t
u
In order to show weak and strong duality statements we give now a reformulation
of the problems using the space I of self-infimal sets (compare Definition 2.6.37)
and corresponding notations. Sequentially, it is important that .I ; I / is a
complete lattice (see Proposition 2.6.40).
We considered the linear vector optimization problem (LVP). It is easy to see that
q q q
WMin..M.S / C RC /; RC / D Inf M.S / ” S ¤ ; and M.S / C RC ¤ Rq :
Our aim is to reformulate problem (LVP) and its dual problem (LVD) as
optimization problems with I -valued objective function (compare Sect. 15.1.2).
Consider the function
q
P W Rn ! I ; P .x/ WD Inf fMxg D fMxg C bd RC :
It holds
[ [
Inf M.S / D Inf fMxg D Inf Inf fMxg D inf P .x/:
x2S
x2S x2S
Hence, we have
8
< fC1g if S D ;
inf P .x/ D f1g if M D Rq
x2S : q
WMin.M ; RC / otherwise.
q
Note that, by Corollary 15.1.37, WMin.M ; RC / ¤ ; iff ; ¤ M ¤ Rq . This
q
means, if the set WMin.M ; RC / is nonempty, it coincides with infx2S P .x/, other-
q
wise if WMin.M ; RC / is empty, we distinguish between two cases: infx2S P .x/ D
fC1g if S D ; and infx2S P .x/ D f1g otherwise. Thus, (LVP) is essentially
equivalent to
inf P .x/; S WD fx 2 Rn j Ax
bg : (LVP0 )
x2S
Moreover, using the order relation 4WD4I in I introduced in Sect. 2.6.3 it is easy
to see that x 2 S is a (weakly efficient) solution of (LVP) if and only if

x 2 S; P .x/ 4 P .x /
0
) P .x/ D P .x 0 / (15.16)
The above considerations show the relationships between the solution concepts for
(LVP) and (LVD) and lattice theoretical solution concepts for the primal problem.
We next want to reformulate the dual problem (LVD) using the supremum in I .
We first consider two auxiliary assertions.
q
Lemma 15.1.40. The set H .U / RC is closed.
q
Proof. Let fyi gi 2IN be a sequence in H .U / RC converging to yN 2 Rq , thus for
q
each i there is some .ui ; ci / 2 U with yi 2 H .ui ; ci / RC , i.e., yiT ci b T ui . We
N 2 U with yN cN b uN .
have to show that there is some .Nu; c/ T T
692 15 Applications
Assume on the contrary that yN T c b T u > 0 for all .u; c/ 2 U . Since U is

polyhedral there is some > 0 with yN T c b T u
for all .u; c/ 2 U . Take
i0 2 IN such that kyi0 yk
N 1 < , then
.yN yi0 /T ci0 kyi0 yk

N 1 kci0 k1 <
hence
yN T ci0 b T ui0 < yiT0 ci0 C b T ui0 ;
a contradiction. t
u
q q q
Lemma 15.1.41. It holds WMax. H .U / RC ; RC / D WMax.H .U /; RC /.
Proof. We have
q q
y 2 WMax.H .U /; RC / ” y 2 H .U / and y 62 H .U / int RC
and

q q q q
y 2 WMax. H .U / RC ; RC / ” y 2 H .U / RC and y 62 H .U / int RC :
Thus it remains to show that

q q
y 2 H .U / RC and y 62 H .U / int RC H) y 2 H .U /:
q
Indeed, y 62 H .U / int RC implies y T c
b T u for all .u; c/ 2 U and y 2
q
H .U / RC implies the existence of some .Nu; c/
N 2 U with y T cN b T uN . Thus we
obtain y cN D b uN , i.e., y 2 H .U /.
T T
t
u
Note that the hyperplane H .u; c/ Rq is a self-infimal set, whenever .u; c/ 2 U .
Therefore the term sup.u;c/2U H .u; c/ is well defined. The next lemma clarifies the
relationship between this supremum and the solution concept of problem (LVD).
Lemma 15.1.42. It holds
8
< f1g if U D ;
q
sup H .u; c/ D fC1g if H .U / RC D Rq
: q
.u;c/2U
WMax.H .U /; RC / otherwise.
Proof. (i) If U D ;, we have sup.u;c/2U H .u; c/ D Sup H .U / D Sup ; D

f1g, by definition.
q
(ii) The case H .U / RC D Rq follows from the definition of the supremal set.
(iii) Since H .U / Rq , we have

q q
sup H .u; c/ D Sup H .U / D WMax .cl .H .U / RC /; RC /;
.u;c/2U
by the definition of the supremal set. Lemma 15.1.40 and Lemma 15.1.41 yield
that
q q q
WMax.cl .H .U / RC /; RC / D WMax.H .U /; RC /:
The proof is completed. t

u
Remark 15.1.43. The preceding three lemmata remain valid if the set U is replaced
by any finite or polyhedral subset.
Lemma 15.1.42 shows in fact the relationship between problem (LVD) and the
following problem,
sup H .u; c/; (LVD 0 )

.u;c/2U
where
˚
U WD .u; c/ 2 Rm Rq j .u; c/
0; k T c D 1; AT u D M T c :
q
Indeed, if WMax.H .U /; RC / is nonempty, it coincides with sup.u;c/2U H .u; c/
in problem (LVD0 ). Otherwise, if WMax.H .U /; RC / is empty, we distinguish
q
between the following two cases:

f1g when U D ;
sup H .u; c/ D
.u;c/2U fC1g otherwise.
The solution concept for (LVD) as introduced in Sect. 2.6.3 can be expressed in
terms of the ordering relation 4WD4I in the complete lattice I . This characteri-
zation is completely analogous to (15.16). So we obtain yet another motivation for
this solution concept.
Lemma 15.1.44. A point .u0 ; c 0 / 2 U is a (weakly efficient) solution of (LVD) if
and only if

.u; c/ 2 U ; H .u0 ; c 0 / 4 H .u; c/ ) H .u0 ; c 0 / D H .u; c/:
(15.17)
694 15 Applications
Proof. Let .u0 ; c 0 / 2 U be a solution of (LVD). Hence u0 solves (D1 (c 0 )) by

Theorem 15.1.29. Consider .u; c/ 2 U with H .u0 ; c 0 / 4 H .u; c/. Then we have
c 0 D c and b T u0 b T u. Since c 0 D c, u is feasible for (D1 (c 0 )) hence b T u b T u0
and consequently b T u0 D b T u. This means we have H .u0 ; c 0 / D H .u; c/.
Let .u0 ; c 0 / 2 U be no solution of (LVD). By Theorem 15.1.29 there exists some
uN
0 with AT uN D M T c0 and b T uN > b T u0 . Hence, we have H .u0 ; c 0 / 4 H .Nu; c 0 /
but H .u0 ; c 0 / ¤ H .Nu; c 0 /, i.e., (15.17) is not satisfied. t
u
As a consequence of the weak and strong duality assertion given in
Lemma 15.1.25 and Theorem 15.1.30 and the above considerations, we present
here duality assertions for vector optimization problems, formulated along the lines
of the classical scalar duality theory. The complete lattice .I ; 4/ of self-infimal
q
subsets of R plays a key role in these results.
The first result shows that we have weak duality between the I -valued problems
(LVP0 ) and (LVD 0 ).
Theorem 15.1.45 (Weak Duality). Let x 2 S and .u; c/ 2 U . Then it holds
H .u; c/ 4 P .x/:
q
Proof. For all y 2 P .x/ D fMxg C bd RC M , Lemma 15.1.25 yields y T c
q
b T u, hence P .x/ H .u; c/ C RC . This implies H .u; c/ 4 P .x/. t
u
The next result shows strong duality between (LVP0 ) and (LVD 0 ). The following
distinction between the three cases is well-known from scalar linear programming.
Theorem 15.1.46 (Strong Duality). Let at least one of the sets S and U be
nonempty. Then it holds strong duality between (LVP0 ) and (LVD 0 ), i.e.,
V WD sup H .u; c/ D inf P .x/:

.u;c/2U x2S
Moreover, the following statements are true.

(i) If S ¤ ; and U ¤ ;, then f1g ¤ V ¤ fC1g and
q q
V D WMax.H .U /; RC / D WMin.P .S /; RC / ¤ ;:
(ii) If S D ; and U ¤ ;, then V D fC1g.

(iii) If S ¤ ; and U D ;, then V D f1g.
Proof. By the weak duality we have
sup H .u; c/ 4 inf P .x/:

.u;c/2U x2S
(i) If S ¤ ; and U ¤ ;, this implies that neither sup.u;c/2U H .u; c/ nor

infx2S P .x/ can be f1g or fC1g. Hence, Theorem 15.1.30 implies
q q
sup H .u; c/ D WMax.H .U /; RC / D WMin.M ; RC / D inf P .x/:
.u;c/2U x2S
(ii) If S D ; and U ¤ ;, we have infx2S P .x/ D fC1g. Theorem 15.1.30

implies that
q q
WMax.H .U /; RC / D WMin.M ; RC / D ;:
q
Since U ¤ ;, we conclude H .U / RC D Rq and Lemma 15.1.42 yields
sup H .u; c/ D fC1g :

.u;c/2U
(iii) If S ¤ ; and U D ;, we have sup.u;c/2U H .u; c/ D f1g. Theorem 15.1.30

implies that
q q
WMin.M ; RC / D WMax.H .U /; RC / D ;:
Since S ¤ ;, we obtain M D Rq , hence infx2S P .x/ D f1g. t

u
In scalar linear programming, the attainment of the supremum of the problem in a
vertex of the feasible set plays a key role in the simplex algorithm. It is therefore
sufficient to search for a solution on a finite subset of the feasible set. The next result
shows that we have a corresponding result for our dual problem. Typically, in our
case, the supremum in (D0 ) is not attained in a single vertex, but in a finite number
of vertices, namely, in the set of those vertices of U being solutions of (LVD), i.e.,
the set ExtrSol.D/ of extreme solutions of (LVD).
Theorem 15.1.47 (Dual Attainment in Vertices). Let S ¤ ; and U ¤ ;. Then
the supremum in the dual problem (LVD 0 ) is attained in extreme solutions of
(LVD), i.e.,
sup H .u; c/ D sup H .u; c/:

.u;c/2U .u;c/2ExtrSol.D/
Proof. Since U ¤ ; and S ¤ ; we have

q
sup H .u; c/ D WMax.H .U /; RC /
.u;c/2U
q q
by Theorem 15.1.46. WMax.H .U /; RC / D WMax.H .ExtrSol.D//; RC / fol-
lows from Theorem 15.1.39.
696 15 Applications
It remains to show that

q
WMax.H .ExtrSol.D//; RC / D sup H .u; c/:
.u;c/2ExtrSol.D/
If S ¤ ; and U ¤ ; then we conclude from Theorem 15.1.46 and Corol-

lary 15.1.37 that ; ¤ M ¤ Rq . Thus M has a facet and consequently
q
ExtrSol.D/ ¤ ; by Lemma 15.1.38. Moreover, H .ExtrSol.D// RC H .U /
q
RC ¤ Rq . Hence the desired statement follows from Remark 15.1.43. t
u
15.2 Applications in Mathematical Finance
In this section we apply the duality statements for linear vector optimization
problems presented in Sect. 15.1.4 for deriving corresponding duality assertions in
portfolio management.
We consider a Markowitz-type bicriterial portfolio optimization problem (see
Heyde, Löhne, Tammer [258]), where the expected return of the portfolio should be
maximized and the risk of the portfolio, measured by the Conditional Value at Risk
(see Example 5.1.4 and [191, Section 4.4, Definition 4.43]), should be minimized.
For details about the Conditional Value at Risk (sometimes also called Average
Value at Risk) see e.g. [497] or [191, Section 4.4].
We consider a market with n different financial instruments with returns rj ; j D
1; : : : ; n being random variables combined in a random vector r D .r1 ; : : : ; rn /T
with a given probability distribution P. The decision vector x 2 Rn represents a
portfolio of these instruments, where the components xj denote the fraction of the
capital invested in instrument j . This yields the constraints
X
n
x
0; xj D 1:
j D1
The return of a portfolio x is given by r T x so the bicriterial optimization

problem consists in minimizing the negative expected return, i.e., E.r T x/ and
the Conditional Value at Risk of the return (see Example 5.1.4 and [191, Section
4.4, Definition 4.43]), i.e., CVaRˇ .r T x/, for a given risk level ˇ 2 Œ0; 1/. We can
approximate this problem by a linear one by sampling the probability distribution
of r like it is done in [496]. If r 1 ; : : : ; r m denotes a sample of size m then
1 X kT
m
E.r T x/ r x
m
kD1
15.2 Applications in Mathematical Finance 697
and
(
1 X m
CVaRˇ .r x/ inf ˛ C
T
zk j ˛ 2 R; 8k 2 f1; : : : ; mg W
.1 ˇ/m
kD1
)
T
zk 2 RC ; r k x C ˛ C zk
0 :
Then the given problem accords essentially with the following linear vector
optimization problem:
WMin.f .S / C R2C ; R2C /; (PM )
where
( )
X
n
kT
S WD .x; z; ˛/ 2 RnC Rm
C R j xj D 1; 8k 2 f1; : : : ; mg W r xC˛Czk
0
j D1
and
0 1
1 X kT
m
B r x C
B m C
B kD1 C
f .x; z; ˛/ D B C:
B X m C
@ 1 A
˛C zk
.1 ˇ/m
kD1
As already noted in Sect. 15.1.4, finding solutions of .PM / is equivalent to finding

weakly minimal solutions of the problem WMin.f .S /; R2C /.
We put
0 1 0 1
In 0 0 0
B C B C
B C B C
0 1 1 B 0 Im 0 C B0C
B C B C
m 1Tm RT 0 0 B C B C
M WD @ A ; A WD B 1T 0 0 C ; b WD B
B C C
B 1 C;
1
1T 1 B n C B C
0 .1ˇ/m m B C B C
B1T 0 0 C B1C
B n C B C
@ A @ A
T
R Im 1 m 0
where
0 1 1
r1 r1m
B :: : : :: C ;
R WD @ : : : A
rn1 rnm
698 15 Applications
I` is the `-dimensional identity matrix and 1` is the `-dimensional vector with all
components being 1. Then the problem .PM / is equivalent to
q q ˚
WMin.M.S / C RC ; RC /; S WD xN 2 RnCmC1 j AxN
b ;
a problem of type (LVP) in Sect. 15.1.4.

As the corresponding dual problem to .PM / we derive the following problem with
a set-valued objective map as a special case of problem (LVD) in Sect. 15.1.4:
q
WMax.H .U /; RC /; (DM )
where
˚
U D .Nu; c/ 2 RnC2mC2
C R2C j c1 C c2 D 1; AT uN D M T c :
In fact, we have
(
U D .w; p; v1 ; v2 ; u; c/ 2 RnC Rm
C RC RC RC RC j c1 C c2 D 1;
m 2
)
c1 c2
w C 1n .v1 v2 / C Ru D R1m ; p C u D 1m ; 1Tm u D c2
m .1 ˇ/m
and the set-valued objective map is given by

˚
H .w; p; v1 ; v2 ; u; c/ D y 2 R2 j c1 y1 C c2 y2 D v1 v2 :
Interpreting w and p as slack variables and defining v WD v1 v2 we arrive at

(
U D .v; u; c/ 2 R Rm
C RC j c1 C c2 D 1;
2
1Tm u D c2 ;
)
c1 c2
1n v C Ru R1m ; u 1m
m .1 ˇ/m
and
˚
H .v; u; c/ D y 2 R2 j c1 y1 C c2 y2 D v :
The following transformation of the dual variables results in dual variables being
interpretable as probabilities. Note that H does not depend on u and for each
.v; u; c/ 2 U there is
15.2 Applications in Mathematical Finance 699
(
.v; q; c/ 2 UN WD .v; q; c/ 2 R Rm
C RC j c1 C c2 D 1;
2
1Tm q D 1;
)
c1 1
1n v C Rqc2 R1m ; q 1m ;
m .1 ˇ/m
where q is given by c12 u if c2 ¤ 0 and can be chosen as qk D m1 for all k if c2 D 0.

On the other hand for each .v; q; c/ 2 UN we have .v; c2 q; c/ 2 U , hence U can be
replaced by UN and problem .DM / is equivalent to
WMax.H .UN /; RC /; (DN M )

q
where
(
X
m
UN D .v; q; c/ 2 R Rm R2 j c
0; c1 C c2 D 1; qk D 1;
kD1
1
8k D 1; : : : ; m W 0 qk ;
.1 ˇ/m
m
X )
1 k
8j D 1; : : : ; n W v r c1 C rj qk c2
k
:
m j
kD1
N M /. A triple
Applying Theorem 15.1.29 we can characterize the solutions of .D
.v ; q ; c / 2 UN is a solution of .D
N M / if and only if
˚ ˚
v D max v j .v; q ; c / 2 UN D max v j .v; q; c / 2 UN ;
i.e., if and only if
m
X m
X
1 1
v D min rjk c1 C rjk qk c2 D max min rjk c1 C rjk qk c2
j D1;:::;n m q2Q j D1;:::;n m
kD1 kD1
with
( )
X
m
1
Q WD q2R j m
qk D 1; 8k D 1; : : : ; m W 0 qk
.1 ˇ/m
kD1
Pm of ˇ
0.
being a nonempty set because
Since q
0 and kD1 qk D 1, the numbers qk may be interpreted as
probabilities
Pm describing an alternative probability distribution Pq for the samples r k .
Then kD1 rj qk D EPq .rj /, the expectation of rj under the alternative distribution
k
700 15 Applications
P
Pq , and m kD1 m rj D EP .rj /, the expectation of rj under the given distribution P.
1 k
The numbers qk are related to the dual description of the coherent risk measure
Conditional Value at Risk. This dual description signifies that the Conditional
Value at Risk of some financial position equals the worst case expected loss of this
position under a certain set of alternative probability distributions (for details see e.g.
[191, Theorem 4.47]). Moreover, the scalarization weights c1 and c2 describe the
model uncertainty, i.e., c1 can be interpreted as the probability for P being the right
probability distribution and c2 as the probability that Pq provides the appropriate
distribution. Then P.c;q/ WD c1 P C c2 Pq describes a probability distribution being a
mixture of P and Pq and EP.c;q/ .rj / D c1 E.rj / C c2 EPq .rj /. Hence, a solution for
the dual problem consists of some .c ; q / determining an alternative probability
distribution P.c ;q / and a number v D minj D1;:::;n EP.c ;q / .rj / where the
vector q 2 Q must be chosen such that it maximizes minj D1;:::;n EP.c ;q/ .rj /
or minimizes maxj D1;:::;n EP.c ;q/ .rj /, i.e., the largest expected return of the n given
financial instruments, given the value of c . That means, .c ; q / provides the worst
case for the expected return of the “best” of the given financial instruments under
the considered alternative probabilities P.c ;q/ .
Using the results of Sect. 15.1.4 we can see that a point .x ; z ; ˛ / 2 S is a
N M / such that
solution of .PM / if and only if there is a solution .v ; q ; c / of .D
!
c X k T X
m m
1
1 r x C c2 ˛ C zk D v D min EP.c ;q / .rj /
m .1 ˇ/m j D1;:::;n
kD1 kD1
or equivalently if
c1 Eappr .r T x / C c2 CVaRˇ .r T x / D v D min EP.c ;q / .rj /;

appr
(15.18)
j D1;:::;n
appr
where Eappr and CVaRˇ are the approximations of the expectation and the
Conditional Value at Risk with the help of the samples. Thus one can find a
solution of the portfolio optimization problem by first determining some “worst
N M/
case” alternative probability P.c ;q / belonging to a solution .v ; q ; c / of .D
and then searching for a portfolio x such that (15.18) is satisfied.
For vector optimization problems one often does not want to choose a scalar-
ization in advance and prefers computing the whole set of minimal solutions.
Concerning the dual problem, it might be also useful to compute all solutions of
.DN M / together with the corresponding efficient portfolios and to provide the decision
maker (the investor) with this information because from solving the dual problem
the investor gets an information about the relationship between the scalarization
weights c and the “worst case” alternative probability scenario P.c ;q / taken into
account under this scalarization.
15.3 Set-Valued Optimization in Welfare Economics 701
15.3 Set-Valued Optimization in Welfare Economics
This section concerns applications of methods from variational analysis and general-
ized differentiation (especially Theorems 5.5.10 and 12.11.3) to nonconvex models
of welfare economics (see [28] and [278]) with infinite-dimensional commodity
spaces. Bao and Mordukhovich [28] established relationships between necessary
conditions in set-valued optimization and appropriate extensions of the second
fundamental theorem of welfare economics to nonconvex economies with general
preference relations. The variational approach developed in the paper by Bao and
Mordukhovich [28] (see Sect. 12.11) yields necessary conditions for various types
of local optimal solutions of constrained set-valued problems such that it is possible
to derive new versions of the second welfare theorem applied to Pareto as well as
weak, strict, and strong Pareto optimal allocations of nonconvex economies (see
Definition 15.3.2) under certain regularity conditions.
It is well known that vector optimization has its roots in economic modeling
and general equilibrium theory. In our book we take attention to recent mathe-
matical areas studying minimal points of sets and minimal solutions of set-valued
optimization problems. We present some applications of set-valued optimization in
welfare economics in this section following the presentation in the paper by Bao
and Mordukhovich [28].
First, we formulate a conventional model of welfare economics that we will
study in the following. For a given normed commodity space E, we consider an
economy
Eco D .CS1 ; ; CSn ; PS1 ; ; PSm ; W / .m 2 N ; n 2 N / (15.19)
involving m firms with production sets PSj E (j D 1; ; m), n customers

with consumption sets CSi E (i D 1; ; n), and a net demand constraint set
W representing constraints related to the initial inventory of commodities in the
economy Eco. We denote production strategies by v D .v1 ; ; vm / 2 PS1
PSm and consumption plans by y D .y1 ; ; yn / 2 CS1 CSn . In the sequel
we suppose that all the consumption sets CSi (i D 1; ; n) are closed. Then the
pair .v; y/ is an admissible state of the economy Eco.
Furthermore, associate with each consumer a preference / level set Li .y/ that
consists of elements in CSi preferred to yi by this consumer at the consumption
plan y. The corresponding preference mappings Li W Y E are set-valued with
Y WD E n . By definition we have yi … Li .y/ for every i D 1; ; n and naturally
suppose that Li .y/ ¤ ; at least for some i 2 f1; ; ng. Put for convenience
cl Li .y/ WD fyi g if Li .y/ D ;.
Definition 15.3.1. An admissible state .v; y/ of the economy Eco in (15.19) is
called feasible allocation of Eco if
X
n X
m
w WD yi vj 2 W; (15.20)
i D1 j D1
702 15 Applications
where W is the net demand constraint set representing constraints related to the
initial inventory of commodities in the economy Eco.
The decision makers in welfare economics are looking for certain types of Pareto-
type solutions (allocations) of the economy Eco in the following sense:
Definition 15.3.2 (Pareto-Type Optimal Allocations). Let .v; y/ 2 E m E n be
a feasible allocation of the economy Eco.
(i) .v; y/ is called a local weak Pareto optimal allocation of Eco if there is a
neighborhood U E m E n of .v; y/ such that for every feasible allocation
.v; y/ 2 U we have yi … Li .y/ for some i 2 f1; ; ng.
(ii) .v; y/ is called a local Pareto optimal allocation of Eco if there is a
.v; y/ 2 U either yi … cl Li .y/ for some i 2 f1; ; ng, or yi … Li .y/ for all
i 2 f1; ; ng.
(iii) .v; y/ is called a local strict Pareto optimal allocation of Eco if there is a
.v; y/ 2 U with y ¤ y we have yi … cl Li .y/ for some i 2 f1; ; ng.
(iv) .v; y/ is called a local strong Pareto optimal allocation of Eco if there is a
.v; y/ 2 U with .v; y/ ¤ .v; y/ we have yi … cl Li .y/ for some i 2 f1; ; ng.
(v) We omit the adjective “local” in (i) - (iv) or replace it by “global” if U D
E m E n above.
Remark 15.3.3. Of course, from the definitions given above it follows that (iv) )
(iii) ) (ii) ) (i) but not vice versa; the same implications hold with respect to the
global version in (v). The notions of (both local and global) weak Pareto and Pareto
optimal allocations are quite natural and often used in welfare economics. These
notions correspond to similar Pareto-type concepts (weakly minimal and minimal
solutions) for vector optimization problems based on preferences given by utility
functions (see [214, 293, 402, 415]). The notion of strong Pareto optimal allocations
(Definition 15.3.2 (iv)) was introduced by Khan [324] for models of welfare
economics and later on intensively studied by Mordukhovich in [410,428,429,431].
Bao and Mordukhovich [28] established such relationships in a more general
context. The notion of strict Pareto optimal allocations (Definition 15.3.2 (iii))
was first appeared in [431, Remark 8.15] as a modified version of the strong
Pareto concept. Bao and Mordukhovich [28] show that local strict Pareto optimal
allocations in welfare economics can be reduced in fact to the so-called fully
localized minimizers (see Definition 2.6.49) in associated problems of constrained
set-valued optimization.
In order apply the necessary optimality conditions derived in Theorem 12.11.3 to
the general nonconvex model of welfare economics we formulate a constrained set-
valued optimization problem which we associate with the welfare economic model
introduced above.
We will formulate the welfare economic model as a set-valued optimization

problem of the following type:
where F W X Y is a set-valued cost mapping between Banach spaces, S X is

a subset in X , and “ -minimization” in (15.21) is understood with respect to some
preference relation on Y (compare Sect. 2.6.5 and [431, Subsection 5.3.1]): Given
a subset R Y Y , we say that y1 is preferred to y2 (y1 y2 ) if .y1 ; y2 / 2 R.
We consider the level-set mapping L W Y Y associated with the preference
by
L.y/ WD fu 2 Y j u yg (15.22)
introduced in Sect. 2.6.5.

In the following we study a set-valued optimization problem in form (15.21)
with X D E mC1 , x D .v; w/, and Y WD E n constructed upon the initial data of the
welfare economy Eco:
where
X
n X
m
F .x/ WD fy 2 Y j w D yi vj g;
i D1 j D1
and
Y
m
S WD . PSj / W X:
j D1
The “minimization” in (15.23) is understood with respect to the level-set mapping

L W Y Y (compare (15.22)) defined in the product form
Y
n
L.y/ WD Li .y/; y 2 YI (15.24)
i D1
via the given preference mappings Li W Y E of the welfare economy Eco.

For deriving the extended second welfare theorem for local Pareto-type optimal
allocations (see Definition 15.3.2) we recall at first that the notions of local
weak Pareto, strict Pareto, and strong Pareto optimal allocations of the welfare
economy (15.19) are equivalent to, respectively, the notions of fully localized weak
704 15 Applications
minimizers, minimizers, and strong minimizers in the sense of Definition 2.6.49

for the set-valued optimization problem (15.23). So it is possible to derive a second
welfare theorem for (15.19) from the corresponding necessary optimality conditions
in Theorem 12.11.3 (see [28, Theorem 3.4]) for fully localized weak minimizers,
minimizers, and strong minimizers.
The following result is shown in ([28, Theorem 4.1]).
Theorem 15.3.4 (Equivalence Between Local Pareto-Type Optimal Allocations
in Welfare Economics and Fully Localized Minimizers in Set-Valued Optimiza-
tion). Let .v; y/ be a feasible allocation of the welfare economy
P Eco P
in (15.19) with
the preference sets Li .y/, and let x WD .v; w/ with w WD niD1 y i m j D1 vj .
Then the following equivalence relationships hold:
(i) .v; y/ is a local weak Pareto optimal allocation of Eco if and only if .x; y/ is a
fully localized weak minimizer of the set-valued optimization problem (15.23)
with respect to the preference mapping L defined in (15.24).
(ii) .v; y/ is a local strict Pareto optimal allocation of Eco if and only if .x; y/ is
a fully localized minimizer of the set-valued optimization problem (15.23) with
respect to the preference mapping L given in (15.24).
(iii) .v; y/ is a local strong Pareto optimal allocation of Eco if and only if it is a
fully localized strong minimizer of the set-valued optimization problem (15.23)
with respect to the preference mapping L defined in (15.24).
In Sect. 12.11 necessary optimality conditions for set-valued optimization
problems of type (15.21) are shown. Now, we will use Theorem 12.11.3 to establish
an extended version of the second fundamental theorem of welfare economics
for local weak Pareto, strict Pareto, and strong Pareto optimal allocations of the
economy Eco under the asymptotic closedness condition (see Definition 12.11.1)
supposed at the corresponding optimal allocation on certain sets involved in the
economy. The following theorem is shown in [28, Theorem 4.3] and ensures the
existence of a marginal price supporting the local Pareto-type optimal allocations
(see Definition 15.3.2) that are formalized via the limiting normal cone (3.25)
computed at the optimal allocation in consideration. This theorem is derived
directly from the corresponding first-order necessary conditions for fully localized
minimizers of constraint set-valued problems presented in Theorem 12.11.3 (see
[28, Theorem 3.4]) and based on the extended extremal principle in Theorem 5.5.10.
Theorem 15.3.5 (Extended Second Welfare Theorem for Local Pareto-Type
Optimal Allocations). Let .v; y/ be a local optimal allocation of economy (15.19)
in the senses listed below with respect to the preference sets Li .y/ under the local
satiation requirement:
y i 2 cl Li .y/ for all i D 1; ; n: (15.25)
Suppose that the commodity space E is Asplund, that the sets PS1 ; ; PSm , and W
are closed around the points in consideration, and that one of the sets
cl Li .y/; i D 1; ; nI PSj ; j D 1; ; mI and W (15.26)

P P
is SNC at y i , vj , and w D niD1 y i m j D1 vj , respectively. Then there exists a
marginal price 0 ¤ p 2 E satisfying the conditions
p 2 N.x i I cl Li .y//; i D 1; ; n;
p 2 N.vj I PSj /; j D 1; ; m;
p 2 N.wI W /
in each of the following cases of local optimal allocations of the economy Eco:
• .v; y/ is a local weak Pareto optimal allocation of Eco provided that all the sets
Li .y/, i D 1; ; n, are asymptotically closed at the corresponding points y i ;
• .v; y/ is a local strict Pareto optimal allocation of Eco provided that there is an
index i 2 f1; ; ng such that the set cl Li .y/ is asymptotically closed at y i ;
• .v; y/ is a local strong Pareto optimal allocation of Eco provided that one of the
sets in (15.26) is asymptotically closed at the corresponding point.
Remark 15.3.6. In order to prove Theorem 15.3.5 one needs a specification of the
asymptotic closedness property (compare Definition 12.11.1) for the product sets
involved into the economy Eco (see [28, Proposition 4.2]).
Now, we consider economies with commodity spaces E ordered by proper
closed cones C E. We recall some conditions, well-recognized in welfare
economics, that imply the asymptotic closedness property of the corresponding sets.
The following result shows that the asymptotic closedness property holds in ordered
Banach spaces E with generating proper closed cones C , i.e., C C D E.
The next proposition is proven in [28, Proposition 4.4].
Proposition 15.3.7 (Asymptotic Closedness Property in Ordered Banach
Spaces). Let E be a Banach space with the generating proper closed ordering
cone C E. Assume that A is a closed subset of E satisfying the condition
A C A; (15.27)
and consider y 2 bd A. Then the set A is asymptotically closed at y.

According to the conventional terminology in welfare economics (compare
Sect. 5.2.2), one says that the economy Eco (15.19) with the commodity space E,
the generating proper closed ordering cone C E and the preference sets Li .y/
satisfies:
• the implicit free disposal of commodities if
cl W C cl W (15.28)
706 15 Applications
• the free disposal of production if
cl PSj C cl PSj for some j 2 f1; ; mgI (15.29)
• the desirability condition if
cl Li .y/ C C cl Li .y/ for some i 2 f1; ; ng: (15.30)
The following corollary (see [28, Corollary 4.5]) is a consequence of Theo-

rem 15.3.5 and employs the result of Proposition 15.3.7 providing in this way
efficient implementations of the established extended version of the second welfare
theorem for the cases of local strict Pareto and strong Pareto optimal allocations of
nonconvex economies with ordered commodity spaces.
Corollary 15.3.8 (Extended Second Welfare Theorem for Local Strict Pareto
and Strong Pareto Optimal Allocations with Ordered Commodities). In addi-
tion to the assumptions of Theorem 15.3.5, assume that the commodity space E
is generated by its proper closed ordering cone C E. Then there is a positive
marginal price p 2 C C n f0g satisfying the conditions of the second welfare
theorem (Theorem 15.3.5) in each of the following cases:
• .v; y/ is a local strict Pareto optimal allocation of the economy Eco where the
desirability condition (15.30) with respect to its preferences is fulfilled;
• .v; y/ is a local strong Pareto optimal allocation of the economy Eco where
either the implicit free disposal of commodities (15.28), or the free disposal of
production (15.29), or the desirability condition (15.30) is fulfilled.
15.4 Robustness for Vector-Valued Optimization Problems
In this section we apply solution concepts for set-valued optimization problems

based on set approach, especially we use the set order relations introduced by
Kuroiwa [348,349,351] (see Definition 2.6.9) in order to study robustness for vector-
valued optimization problems.
Dealing with uncertainty in vector optimization is very important in many
applications. On the one hand side, most real world optimization problems are
contaminated with uncertain data, especially traffic optimization, scheduling prob-
lems, portfolio optimization, network flow and network design problems. On the
other hand side, many real world optimization problems require the minimization
of multiple conflicting objectives (see [543]), e.g. the maximization of the expected
return versus the minimization of risk in portfolio optimization, the minimization of
production time versus the minimization of the cost of manufacturing equipment,
or the maximization of tumor control versus the minimization of normal tissue
complication in radiotherapy treatment design.
15.4 Robustness for Vector-Valued Optimization Problems 707
This leads us to a concept of robustness for vector-valued optimization problems

introduced and studied in a paper by Ehrgott, Ide and Schöbel [159]. In this section
we follow the explanations in [159]. In [159] the authors extend the well-known
concept of minmax robustness from single objective to multiobjective optimization.
In single objective uncertain optimization this concept has been extensively studied,
see e.g. [44, 173, 540] for older contributions and [43] for a recent survey. Detailed
overviews of the existing robustness concepts for vector-valued optimization prob-
lems can be found in [541] and [130]. Furthermore, it is interesting to mention
that another line of research investigates the use of multiobjective optimization
techniques, especially scalarization methods, in order to model uncertain single
objective optimization problems, see, e.g., [556] and [328].
For optimization problems with uncertain data it is typically that at the time when
it is solved these data are not completely known. It is very important to estimate the
effects of this uncertainty and so it is necessary to evaluate how sensitive an optimal
solution is to perturbations of the input data. One way to deal with this question is
sensitivity analysis (for an overview see [526]). Sensitivity analysis is a posteriori
approach and provides ranges for input data within which a solution remains
feasible or optimal. It does not, however, provide a course of action for changing
a solution should the perturbation be outside this range. In contrast, stochastic
programming (see Birge and Louveaux [50] for an introduction) and robust
optimization techniques take the uncertainty into account during the optimization
process. Depending from the concrete application one can decide whether robust or
stochastic optimization is the more appropriate way of dealing with uncertainty.
For example, the emergency department with landing place for rescue helicopters
in a ski resort should be chosen in such a way that the flight time to all ski slopes in
the resort that are to be protected is minimized in the worst case, even though flight
times are uncertain due to unknown weather conditions in the future. Similarly, if
a aircraft-schedule of an airline is to be determined, one would want to be able to
treat as many passengers as possible in a cost-effective manner, even though the
exact number of passengers is not known at the time the schedule is fixed.
An important approach of dealing with such optimization problems is described
in the concept of robustness: Instead of assuming that all data are known, one
allows different scenarios for the input parameters and looks for a solution that
works well in every uncertain scenario. Robust optimization is a growing up field
of research, we refer to Ben-Tal, L. El Ghaoui, Nemirovski [43], Kouvelis and Yu
[336] for an overview of results and applications for the most prominent concepts.
Several other concepts of robustness were introduced more recently, e.g. the concept
of light robustness by Fischetti and Monaci [186] or the concept of recovery-
robustness in Liebchen et al. [388]. A scenario-based approach is suggested in
Goerigk and Schöbel [212]. In all these approaches, the uncertain optimization
problem is replaced by a deterministic version, called the robust counterpart of the
uncertain problem.
For specific robustness concepts, the connection between uncertain scalar opti-
mization problems and an associated (deterministic) multiple objective counterpart
were observed by several authors. Kouvelis and Sayin [335, 522] use this relation to
708 15 Applications
develop efficient solution methods for bi- and multiple objective discrete optimiza-
tion problems based on algorithms that were originally developed to solve uncertain
scalar optimization problems. They focus on two classical robustness concepts, see
also [336].
Perny et.al. [476] use a multiple objective counterpart to introduce a robustness
measure based on the Lorenz dominance rule in the context of minimum spanning
tree and shortest path problems. From the stochastic programming perspective, a
multiple objective counterpart for a two-stage stochastic programming problem was
introduced in Gast [195] and used to interrelate stochastic programming models
with the concept of recoverable robustness, see Stiller [544]. A critical analysis is
given in Hites et.al. [264] who give a qualitative description of the similarities and
differences between the two modeling paradigms.
In the literature, two main ways of dealing with optimization problems with
uncertainty are suggested. In stochastic optimization, knowledge about the proba-
bility distribution of the uncertain data is assumed. Some authors use a probabilistic
approach, for example, to minimize the probability that a solution is - in some sense
- not good or not feasible. Other authors use the expected value of some cost function
depending on the realization of the uncertain data as the objective function.
While in stochastic optimization a solution is desired which is good on average,
a robust solution hedges against the worst case that may happen. Hence, different
from stochastic optimization, no probability distribution is needed in robust
optimization. There are different ways of defining robustness in the literature.
Here, we focus on several models where uncertainties are assumed to occur in the
objective function.
First, we formulate a scalar optimization problem with uncertainties.
Throughout this section, let U Rm be the uncertainty set reflecting the potential
scenarios that may occur. One could think of 2 U being real numbers or real
vectors. Furthermore, let f W Rn U ! R. Then an uncertain scalar optimization
problem (uncertain OP) P .U / is defined as a family of optimization problems
.P ./; 2 U /; (15.31)
where for a given 2 U the optimization problem (P ./) is given by
f .x; / ! min
(P ./)
s.t. x 2 S Rn :
At the time the uncertain OP (15.31) has to be solved, it is not known which value
2 U is going to be realized.
An uncertain scalar optimization problem P .U / is defined as a family of
optimization problems with different objective functions defined by each scenario
2 U . In order to clarify how it is possible to evaluate a feasible solution we
are looking for feasible solutions x 2 S that are robust optimal solutions of the
uncertain optimization problem. For scalar optimization problems, many different
approaches to model robustness have been suggested in the literature. Minmax

robustness (also called strict robustness) has been first mentioned by Soyster
[540] and then formalized and analyzed by Ben-Tal, El Ghaoui, and Nemirovski
in numerous publications, see e.g. [44] for early contributions and [43] for an
extensive collection of results. The idea is that the worst possible objective function
value is minimized in order to get a solution that is “good enough” even in the
worst case scenario. Furthermore, constraints have to be satisfied for every scenario
2 U . Then the strictly robust counterpart of the uncertain optimization problem
.P ./; 2 U / is defined by
RC .x/ D max f .x; / ! min

2U
(RC )
s.t. x 2 S Rn :
In Ehrgott, Ide, Schöbel [159] the concept of minmax robustness is extended to

vector-valued optimization problems. We derive corresponding results in more
general settings (compare Ide, Köbis, Kuroiwa, Schöbel, Tammer [273]).
Unless otherwise mentioned, in the following we suppose that X is a linear space,
Y a linear topological space and C Y is a proper closed convex and pointed cone.
In deterministic vector optimization one studies the problem (see Sect. 2.4)
Min.f .S /; C / (VOP)
with a vector-valued objective function f W X ! Y , and a feasible set S X .

In the following we assume that the objective function f may depend on
scenarios which are unknown or uncertain. As in uncertain single objective
optimization, for an uncertainty set U Rm , an uncertain vector-valued
optimization problem P .U / is given as the family
.P ./; 2 U /
of vector-valued optimization problems
Min.f .S; /; C / (P ./)
with the objective function f W X U ! Y , a feasible set S X and the notation

(for 2 U )
f .S; / WD ff .x; / j x 2 S g:
We call 2 U a scenario and .P .// an instance of P .U /.

Naturally, the issue of uncertain data affects these vector-valued optimization
problems as it does in scalar optimization. Therefore, being able to find robust
solutions would be very valuable in many applications. Research in this area
710 15 Applications
deals with Branke’s concept of robustness for single objective optimization [77].
Furthermore, Deb and Gupta [133] introduced certain concepts of robustness for
vector-valued optimization problems.
Given an uncertain vector-valued optimization problem P .U /, the same ques-
tion as in single objective optimization arises, namely, how to evaluate feasible
solutions x 2 S . In uncertain vector-valued optimization problems, we cannot
evaluate solutions by just taking the worst case over all scenarios because we obtain
a vector of objective values for each scenario. For each x 2 X the set of objective
values of it is given by
F .x/ WD ff .x; / j 2 U g Y: (15.32)
Dealing with an uncertain vector-valued optimization problem P .U / leads us to

the following set-valued optimization problem with an objective map F W X Y
given in (15.32) and an order relation (see Sect. 2.6.2):
minimize F .x/; subject to x 2 S: (SP )
Let C Y be a proper closed convex and pointed cone. Furthermore, we

assume that F is compact-valued. In the following we will discuss several concepts
of robustness for vector optimization problems based on the set order relations
introduced in Sect. 2.6.2 (see Ide, Köbis, Kuroiwa, Schöbel, Tammer [273]).
15.4.1 uC -Robustness
With these notions we are able to apply the solution concept based on the upper set
less order relation uC introduced in Definition 2.6.9 for A; B Y :
A uC B ” A B C:
Ehrgott, Ide, Schöbel [159, Definition 3.1] introduced robust efficient elements
q
(minmax robust efficiency) for the case X D Rn , Y D Rq , C D RC : A feasible
q
solution x 2 S robust efficient, if F .x/ RC does not contain any other set F .x/
with x ¤ x 2 S . Furthermore, Ehrgott, Ide, Schöbel formulated in [159, Definition
3.1] robust weakly efficient and robust strictly efficient elements. We will extend the
concepts given in [159] to more general cones C Y using the upper set less order
relation uC .
Definition 15.4.1 (uC -Robustness). Given an uncertain vector-valued optimiza-
tion problem P .U /, a solution x 0 2 S is called uC -robust for P .U /, if there
is no x 2 S n fx 0 g such that
F .x/ uC F .x 0 /:
The concepts of robust weak efficiency and robust efficiency introduced in

Ehrgott, Ide, Schöbel [159, Definition 3.1] inspire us to extent the concept of upper
set less order relation uC introduced in Definition 2.6.9 for a proper closed convex
and pointed cone C Y and A; B Y to the case of general nonempty sets
Q Y . Suppose that Q Y is a proper set with C cl Q and cl Q \ . cl Q/ D
f0g. Under these assumptions we introduce an order relation with respect to Q
analogously to the upper set less order relation uC introduced in Definition 2.6.9:
A uQ B WD” A B Q: (15.33)
If we are dealing with Q D int C we suppose int C ¤ ;.

Definition 15.4.2 (uint C -Robustness and uC nf0g -Robustness). Given an uncer-
tain vector-valued optimization problem P .U / we call a solution x 0 2 S
• uint C -robust , if int C ¤ ;, Q D int C in (15.33) and there is no x 2 S n fx 0 g
such that
F .x/ uQ F .x 0 /I
• uC nf0g -robust , if Q D C n f0g in (15.33) and there is no x 2 S n fx 0 g such that
F .x/ uQ F .x 0 /I
Remark 15.4.3. In the case of uC -robust solutions in the sense of Definition 15.4.1
we have Q D C in (15.33).
Remark 15.4.4. The motivation behind the concept of uQ -robustness is the follow-
ing: When comparing sets with the upper set less order relation uQ (with Q D C ,
Q D int C or Q D C n f0g), the upper bounds of these sets, i.e., the “worst cases”,
are considered. Minimizing these worst cases is closely connected to the concept of
minmax robust efficiency (see [159, Definition 3.1]) where one wants to minimize
the objective function in the worst case. This risk averse approach would reflect a
decision makers strategy to hedge against a worst case and is rather pessimistic.
Remark 15.4.5. The first scenario-based concept to uncertain multi-objective opti-
mization, or minmax-robustness adapted to multi-objective optimization, has been
introduced by Kuroiwa and Lee [355] and studied in [356]. In [355, 356] robust
solutions of multi-objective optimization problems are introduced in the following
way. The authors propose to consider the robust counterpart to P.U /
U q
Min.fRC .S /; RC /; (15.34)
0 1
max1 2U1 f1 .x; 1 /
U
fRC .x/ WD @ ::: A; (15.35)
maxq 2Uq fq .x; q /
712 15 Applications
with functionals fi W Rn Ui ! R for i D 1; : : : ; q and the convex and compact

uncertainty sets U WD .U1 ; : : : ; Uq / .Ui Rm ; i D 1; : : : ; q/. In [355], solutions
to (15.34) are called robust.
Note that in [159] the authors pointed out that this concept differs from the
concept of minmax robust efficiency.
In deterministic vector-valued optimization there is a relationship between weak
minimality, minimality and strict minimality. This relationship also holds for robust
minimality.
Lemma 15.4.6. Let P .U / be an uncertain vector-valued optimization problem,
where C Y is a proper closed convex and pointed cone with int C ¤ ;. Then we
have:
x is uC -robust ) x is uC nf0g -robust ) x is uint C -robust :
Proof. The assertion holds because of int C C n f0g C . t

u
The robust counterpart of an uncertain vector-valued optimization problem is the
problem of identifying all x 2 S which are uint C -robust , uC nf0g -robust , or uC -
robust for P .U /.
The next lemma is shown by Ehrgott, Ide, Schöbel [159] for the case X D Rn ,
q
Y D Rq , C D RC and gives important properties of the order relation that are used
in order to derive solution procedures.
Lemma 15.4.7. Consider the uncertain vector-valued optimization problem
P .U /. Then it holds for all x 0 , x 2 S and for Q D int C (Q D C n f0g,
Q D C , respectively),
F .x 0 / F .x/ Q ” F .x 0 / C F .x/ Q (15.36)

F .x 0 / F .x/ Q ” 8 2 U 9 2 U W f .x 0 ; / 2 f .x; / Q:
(15.37)
Proof. (15.36): (i) “H)”: Suppose F .x 0 / F .x/ Q. Then f .x 0 ; / 2 F .x/ Q

for every 2 U . This implies
f .x 0 ; / c 2 F .x/ Q
for every 2 U and for all c 2 C because of C cl Q. So we can conclude
F .x 0 / C F .x/ Q:
(ii) “(H”: Suppose F .x 0 / C F .x/ Q. Because of 0 2 C we have
F .x 0 / F .x/ Q:
(15.37): (i) “H)”: Suppose the contrary. Then
9 2 U 8 2 U W f .x 0 ; / … f .x; / Q H) 9 2 U W f .x 0 ; / … F .x/ Q
H) F .x 0 / 6 F .x/ Q:
(ii) “(H”: Suppose the contrary. Then
9 2 U W f .x 0 ; / … F .x/ Q H) 9 2 U 8 2 U W f .x 0 ; / … f .x; / Q:
t
u
q
Under the additional assumption that X D Rn , Y D Rq , C D RC in Ehrgott,
Ide, Schöbel [159, Lemma 3.4] the following result is derived:
Lemma 15.4.8. Consider the uncertain vector-valued optimization problem P .U /
with X D Rn , Y D Rq , C D RC . Then it holds for all x 0 , x 2 S and for Q D int C
q
(Q D C n f0g, Q D C , respectively),
F .x 0 / F .x/ Q H) sup fi .x 0 ; / sup fi .x; / 8i 2 f1; kg: (15.38)

2U 2U
If max2U fi .x; / exists for all x 2 S , then
F .x 0 / F .x/ int C H) max2U fi .x 0 ; / < max2U fi .x; / 8i 2 f1; kg:

(15.39)
Proof. (15.38): Let F .x 0 / F .x/ Q. Suppose the contrary. Then there exists an
i 2 f1; qg such that
sup fi .x 0 ; / > sup fi .x; /:

2U 2U
Since the inequality is strict there exists a scenario 2 U such that fi .x 0 ; / >
sup2U fi .x; /. This means that for all 2 U it holds f .x 0 ; / … f .x; / Q.
This yields
F .x 0 / 6 F .x/ Q;
because of (15.37), a contradiction.

(15.39): Assume max2U fi .x; / exists for all x 2 S and let F .x 0 / F .x/
int C . Suppose the contrary. Then there exists an i 2 f1; qg such that
max fi .x 0 ; /
max fi .x; /:
2U 2U
714 15 Applications
This inequality implies that there exists a scenario 2 U such that fi .x 0 ; /
max2U fi .x; / and so we get for all 2 U that f .x 0 ; / … f .x; / int C . This
yields
F .x 0 / 6 F .x/ int C;
because of (15.37), a contradiction. t

u
In the following we will see that for special structures of set-valued optimization
problems we are able to use scalarization methods (compare Chap. 5) in order to
derive solution procedures.
Lemma 15.4.7 is used in Ehrgott, Ide, Schöbel [159] for deriving solution
procedures for uncertain vector optimization problems in the special case X D Rn ,
q
Y D Rq , C D RC in (VOP). Using Lemma 15.4.7, we apply the knowledge of
deterministic vector optimization (compare Sect. 2.4) and single objective robust
optimization for deriving solution procedures for generating uC nf0g -robust solutions
of an uncertain vector optimization problem P .U /. The most common approach
for computing properly minimal solutions for a deterministic vector optimization
problem in finite dimensional image spaces is the weighted sum scalarization (see
Definition 2.4.4, (a)). The general idea is to form a scalar optimization problem
by multiplying each objective function with a positive weight and summing up
the weighted objectives. The scalarized problem .Py / for a given (deterministic)
vector optimization problem Min.f .S /; C / (see (VOP)) with y 2 C C n f0g (see
Sect. 2.4) is
min y .f .x//: (Py )

x2S
We now reduce the robust counterpart of an uncertain vector-valued optimization

problem to a single objective uncertain optimization problem in order to be able
to generate uC nf0g -robust solutions by computing robust optimal solutions of the
uncertain single objective problem applying Theorem 2.4.11, (b). To this end,
we introduce the robust version of the weighted sum scalarization problem of an
uncertain vector-valued optimization problem P .U / as
min sup y .f .x; //: (P Uy )

x2S 2U
The following result is shown by Ehrgott, Ide, Schöbel [159, Theorem 4.3] for
q
the case X D Rn , Y D Rq , C D RC . We present the result in a more general
setting where X is a linear space, Y is a linear topological space and C Y is a
proper closed convex and pointed cone.
Theorem 15.4.9. Consider an uncertain vector optimization problem P .U /. The
following statements hold:
(a) If x 0 2 S is a unique optimal solution of .P Uy / for some y 2 C C n f0g, then

x 0 is uC -robust for P .U /.
(b) If x 0 2 S is an optimal solution of .P Uy / for some y 2 C # and
max2U y .f .x; // exists for all x 2 S , then x 0 is uC nf0g -robust for P .U /.
(c) If x 0 2 S is an optimal solution of .P Uy / for some y 2 C C n f0g and
max2U y .f .x; // exists for all x 2 S , then x 0 is uint C -robust for P .U /.
Proof. Suppose that x 0 is not uC -robust (uC nf0g -robust , uint C -robust , respec-
tively) for P .U /. Then there exists an element x 2 S n fx 0 g such that
F .x/ F .x 0 / Q; (15.40)
for Q D C , (Q D .C n f0g/, Q D int C , respectively). This implies
8 2 U 9 2 U W f .x; / 2 f .x 0 ; / Q:
taking into account Lemma 15.4.7, (15.37).

Choose now y 2 C C nf0g for Q D C (y 2 C # for Q D C nf0g, y 2 C C nf0g
for Q D int C , respectively) arbitrary but fixed.
H) 8 2 U 9 2 U W y .f .x; // .<; <; respectively/ y .f .x 0 ; //

” 8 2 U W y .f .x; // .<; <; respectively/ sup y .f .x 0 ; 0 //
0 2U
” sup y .f .x; 0 // .<; <; respectively/ sup y .f .x 0 ; 0 //:

0 2U 0 2U
The last inequalities hold because for (b) and (c) max 0 2U y .f .x; // exists. But
this means that x 0 is not the unique optimal (an optimal, an optimal, respectively)
solution of .P Uy / for y 2 C C n f0g (y 2 C # , y 2 C C n f0g, respectively). ut
Remark 15.4.10. In Theorem 15.4.9 (b) we consider y 2 C # . Under our assump-
tions concerning the cone C and if we assume additionally Y D Rq we have C # ¤ ;
(compare [214, Theorem 2.2.12], [214, Example 2.2.16] and Sect. 2.2). Moreover,
if Y is a Hausdorff locally convex space, C Y is a proper convex cone and C has
a base B with 0 … cl B, then C # ¤ ; (compare [214, Theorem 2.2.12]).
Based on Theorem 15.4.9, Ehrgott, Ide and Schöbel [159] derived the following
algorithm for computing uC -robust (uC nf0g -robust, uint C -robust, respectively)
solutions of the uncertain vector-valued optimization problem P .U / for the special
q
case X D Rn , Y D Rq and C D RC . We present the algorithm for the more
general case where X is a linear space, Y is a linear topological space and C Y
is a proper closed convex and pointed cone.
716 15 Applications
Algorithm 1 for calculating (uC j uC nf0g j uint C )-robust solutions of P .U /:

Input: Uncertain vector-valued problem P .U /, solution sets Opt uC D
Opt u D Opt uint C D ;.
C nf0g
Step 1: Choose a set C C C n f0g.
Step 2: If C D ;: STOP. Output: Set of uC -robust solutions Opt uC , set of
uC nf0g -robust solutions Opt u , set of uint C -robust solutions Opt uint C for
C nf0g
P .U /.
Step 3: Choose y 2 C . Set C WD C n fy g.
Step 4: Find an optimal solution x 0 of .P Uy /.
a) If x 0 is a unique optimal solution of .P Uy /, then x 0 is uC -robust for P .U /,
thus
Opt uC WD Opt uC [ fx 0 g:
b) If max2U y .f .x; // exists for all x 2 S and y 2 C # , then x 0 is uC nf0g -
robust for P .U /, thus
Opt u WD Opt u [ fx 0 g:
C nf0g C nf0g
c) If max2U y .f .x; // exists for all x 2 S , then x 0 is uint C -robust for P .U /,
thus
Opt uint C WD Opt uint C [ fx 0 g:
Step 5: Go to Step 2.
The following algorithm is an interactive procedure for solving the uncertain
vector-valued optimization problem P .U /. In this approach we reduce the scalar-
ized problem to a 1-parametric optimization problem. The algorithm works in the
case that X is a linear space, Y is a linear topological space and C Y is a proper
closed convex and pointed cone.
Algorithm 2 for calculating a single desired (uC j uC nf0g j uint C )-robust
solution of P .U /:
Input: Uncertain vector-valued problem P .U /.
Step 1: Choose a nonempty set C C C n f0g.
Step 2: Choose yN 2 C .
Step 3: Find an optimal solution x 0 of .P UyN /.
a) If x 0 is a unique optimal solution of .P UyN /, then x 0 is uC -robust for P .U /.
b) If max2U yN .f .x; // exists for all x 2 S and yN 2 C # , then x 0 is uC nf0g -
robust for P .U /.
c) If max2U yN .f .x; // exists for all x 2 S , then x 0 is uint C -robust for P .U /.
If x 0 is accepted by the decision-maker, then Stop. Output: x 0 . Otherwise, go

to Step 4.
Step 4: Put k D 0, t0 D 0. Choose yO 2 C , yO ¤ yN . Go to Step 5.
Step 5: Choose tkC1 with tk < tkC1 1 and compute an optimal solution x kC1
of
min sup .yN C tkC1 .yO yN //.f .x; // (P UyN CtkC1 .yO yN / )
x2S 2U
and use x k as starting point. If an optimal solution of .P UyN CtkC1 .yO yN / / cannot
be found for t > tk , then go to Step 2. Otherwise, go to Step 6.
Step 6: The point x kC1 is to be evaluated by the decision-maker. If it is accepted
by the decision-maker, then Stop. Output: x kC1 . Otherwise, go to Step 7.
Step 7: If tkC1 D 1, then go to Step 2. Otherwise, set k D k C 1 and go to Step 5.
Furthermore, another approach based on scalarization by means of nonlinear
functionals of type (5.5) is given in Ehrgott, Ide, Schöbel [159]. In [159] the "-
constraint method is used. In Chap. 5 we discussed the functional
'A;k 0 W Y ! R; 'A;k 0 .y/ WD infft 2 R j y 2 tk0 C Ag;
where A is a nonempty subset of a real linear topological space Y and k 0 ¤ 0 an

element of Y (compare (5.5)).
q
The "-constraint method for the case X D Rn , Y D Rq and C D RC was
introduced by Haimes [230] (compare Eichfelder [161] and Ehrgott [158]). The
idea of the "-constraint method is to choose one of the functions fi (i 2 f1; qg)
as objective function and to take for the other functions fj (j ¤ i ) upper bounds
given by a vector " 2 Rq . This "-constraint method is a special case of this general
nonlinear scalarization method by means of the functional (5.5) if we put

1 for l D i
k 0 D .k10 ; : : : ; kq0 /T ; where kl0 WD (15.41)
0 for l ¤ i;

q N N N N N 0 for l D i;
A WD RC C b; with b D .b ; : : : ; b / ; b WD
1 q T l
(15.42)
"l for l ¤ i:
For a deterministic vector-valued optimization problem (VOP), i 2 f1; qg and

" 2 Rq the "-constraint optimization problem is defined as
min fi .x/ s.t. fj .x/ "j 8j ¤ i; x 2 S: (P";i )
We now use the "-constraint method to reduce a vector-valued uncertain opti-

mization problem P .U / to a single objective uncertain optimization problem. For
doing this we define the robust "-constraint version .P U";i / of P .U /, thus an "-
constraint problem for an uncertain vector-valued optimization problem:
718 15 Applications
min sup fi .x; / s.t. fj .x; / "j 8j ¤ i; 8 2 U ; x 2 S: (P U";i )

2U
In the following theorem the relationships between solutions of .P U";i / and

robust strict minimality (robust weak minimality) are presented (see [159]).
Theorem 15.4.11. Consider an uncertain vector-valued optimization problem
q
P .U / with X D Rn , Y D Rq and C D RC , then the following statements hold.
(a) If x 0 is the unique optimal solution of .P U";i / for some " 2 Rq and some
i 2 f1; ; qg, then x 0 is uC -robust for P .U /.
(b) If x 0 2 S is an optimal solution of .P U";i / for some " 2 Rq and
max2U fi .x; / exists for all x 2 S , then x 0 is uint C -robust for P .U /.
Proof. (a): Suppose that x 0 is not uC -robust for P .U /. Then there exists an
element x 0 2 S with
F .x 0 / uQ F .x 0 /;
q
where uQ is given with respect to Q D C D RC .
q
With Lemma 15.4.7, (15.36) and (15.37), we get for Q D C D RC
8 2 U 9 2 U W f .x 0 ; / 2 f .x 0 ; / RC :
q
This implies
sup f .x 0 ; / sup f .x 0 ; /
2U 2U
and
8 2 U 9 2 U W fj .x 0 ; / fj .x 0 ; / "j .j ¤ i /:
This means that x 0 is feasible for .P U";i / and has an equal or better objective
function value than x 0 , in contradiction to the assumption that x 0 is the unique
optimal solution of .P U";i /.
(b): Suppose that x 0 is not uint C -robust for P .U /. Then there exists an element
x 0 2 S with
F .x 0 / uQ F .x 0 /;
q
where uQ is given with respect to Q D C D int RC .
Again, with Lemma 15.4.7, (15.36) and (15.37), we can conclude
8 2 U 9 2 U W f .x 0 ; / 2 f .x 0 ; / int RC :
q
This yields
max f .x 0 ; / < max f .x 0 ; /

2U 2U
and
8 2 U 9 2 U W fj .x 0 ; / < fj .x 0 ; / "j .j ¤ i /:
This means that x 0 is feasible for .P U";i / and has a better objective function
value that x 0 , in contradiction to the assumption that x 0 is the unique optimal
solution of .P U";i /. t
u
This leads to the following algorithm for computing uC -robust (uint C -robust )
solutions of the uncertain vector-valued optimization problem P .U / with X D Rn ,
q
Y D Rq and C D RC .
Algorithm 3 for calculating (uC j uint C )-robust solutions of P .U / based on

"-constraint method:
Input: Uncertain vector-valued problem P .U /, solution sets Opt uC D

Opt uint C D ;.
Step 1: Choose a set E Rq .
Step 2: If E D ;: STOP. Output: Set of uC -robust solutions Opt uC and set of
uint C -robust solutions Opt uint C for P .U /.
Step 3: Choose " 2 E . Set E WD E n f"g.
Step 4: For every i 2 f1; qg find an optimal solution x 0 .i / of .P U";i /.
a) If x 0 .i / is a unique optimal solution of .P U";i /, then x 0 .i / is uC -robust for
P .U /, thus
Opt uC WD Opt uC [ fx 0 .i /g:
b) If max2U fi .x; / exists for all x 2 S , then x 0 .i / is uint C -robust for P .U /,

thus
Opt uint C WD Opt uint C [ fx 0 .i /g:
Remark 15.4.12. Further concepts for robustness and corresponding algorithms for
solving uncertain vector-valued optimization problems P .U / are developed in Ide,
Köbis, Kuroiwa, Schöbel and Tammer [273].
Remark 15.4.13. The algorithms presented in this section can be considered as
algorithms for solving special classes of set-valued optimization problems where
the solution concept is given by the upper set less order relation (Definition 2.6.9).
720 15 Applications
15.4.2 lC -Robustness
In this section we use the l-type set-relation lQ like in Definition 2.6.9 with Q D
C , Q D C n f0g and Q D int C , respectively, instead of lC :
A lQ B W” A C Q B;
where A; B Y are arbitrarily chosen sets. If we are dealing with Q D int C we

suppose int C ¤ ;. Consider again the set-valued map F given by (15.32). Using
this notation we derive the new concept of lQ -robustness, defined analogously to
uQ -robustness (Definitions 15.4.1 and 15.4.2):
Definition 15.4.14. Given an uncertain multi-objective optimization problem
P .U /, a solution x 0 2 S is called lQ -robust (with Q D C , Q D C n f0g
and Q D int C ) for P .U /, if there is no x 2 S n fx 0 g such that
F .x/ lQ F .x 0 /:
Remark 15.4.15. The lQ -robustness (with Q D C , Q D C n f0g and Q D int C ,

respectively) can be interpreted as an optimistic approach. The lQ -robustness is
an alternative tool for the decision maker for obtaining solutions of another type to
an uncertain multi-objective optimization problem. This rather optimistic approach
focuses on the lower bound of a set F .x/ N for the comparison with another set F .x 0 /.
In particular, in the case Q D C , a point x 0 2 S is called a lC -solution if there
is no other point xN 2 S such that F .x 0 / is a subset of F .x/ N C C . Contrary to
the uQ -robustness approach, the lQ -robustness (with Q D C , Q D C n f0g and
Q D int C , respectively) is hence not a worst-case concept, thus the decision maker
is not considered to be risk averse but risk affine. This optimistic concept thus hedges
against perturbations in the best-case scenarios.
For calculating lQ -robust solutions again scalarization method is helpful, but in
order to later on compute lQ -robust solutions to P .U /, we define a new weighted
sum problem in a general setting:
Let y 2 C C n f0g (y 2 C # , respectively). Consider the weighted sum
scalarization problem
min inf y .f .x; //: (P Uyl )

x2S 2U
Theorem 15.4.16. Consider an uncertain vector optimization problem P .U /. The

following statements hold:
(a) If x 0 is a unique optimal solution of (P Uyl ) for some y 2 C C n f0g, then x 0
is a lC -robust solution for P .U /.
(b) If x 0 is an optimal solution of (P Uyl ) for some y 2 C # and

min2U y .f .x; // exists for all x 2 S , then x 0 is a lC nf0g -robust solution
for P .U /.
(c) If x 0 is an optimal solution of (P Uyl ) for some y 2 C C n f0g and
min2U y .f .x; // exists for all x 2 S , then x 0 is a lint C -robust solution
for P .U /.
Proof. Suppose x 0 is not lQ -robust for Q D C (Q D C n f0g, Q D int C ,
respectively). Consequently, there exists an xN 2 S n fx 0 g such that F .x/
N CQ
F .x 0 / for Q D C (Q D C n f0g, Q D int C , respectively). That is equivalent to
8 2 U 9 2 U W f .x;
N / C Q 3 f .x 0 ; /
” 8 2 U 9 2 U W f .x;
N / 2 f .x 0 ; / Q (15.43)
Now choose y 2 C C n f0g for Q D C (y 2 C # for Q D C n f0g, y 2 C C n f0g

for Q D int C , respectively) arbitrary, but fixed. Hence, we obtain from (15.43)
H) 8 2 U 9 2 U W y .f .x; // .<; <; respectively/ y .f .x 0 ; //

H) 8 2 U W inf y .f .x; // .<; <; respectively/ y .f .x 0 ; //
2U
H) inf y .f .x; // .<; <; respectively/ inf y .f .x 0 ; //;

2U 2U
in contradiction to the assumptions. t

u
Based on these results, we are able to present the following algorithm that
computes (lC j lC nf0g j lint C )-robust solutions to P .U /:
Algorithm 4 for deriving (lC j lC nf0g j lint C )-robust solutions of P .U /:

Input & Step 1-5: Analogous to Algorithm 1, only replacing .P Uy / by
(P Uyl ) and replacing
max2U y .f .x 0 ; // by min2U y .f .x 0 ; //.
The next algorithm computes (lC j lC nf0g j lint C )-robust solutions via scalar-
ization by means of linear continuous functionals:
Algorithm 5 for calculating a single desired (lC j lC nf0g j lint C )-robust
solution of P .U /:
Input & Step 1-7: Analogous to Algorithm 2, only replacing .P UyN / by
.P UyNl /,
max2U y .f .x 0 ; // by min2U y .f .x 0 ; // and .P UyN CtkC1 .yO yN / / by
.P UyNl Ct .yO yN / /.
kC1
722 15 Applications
15.4.3 sC -Robustness
Now, we use the set less order relation sQ with Q D C , Q D C n f0g and Q D
int C , respectively (compare Definition 2.6.7) for A; B Y arbitrarily chosen sets:
A sQ B W” A B Q and A C Q B:
If we are dealing with Q D int C we suppose int C ¤ ;.

We can now introduce the concept of sQ -robustness (with Q D C , Q D C n f0g
and Q D int C , respectively):
Definition 15.4.17. A solution x 0 of P .U / is called (sC j sC nf0g j sint C )-robust
if there is no xN 2 S n fx 0 g such that
N sQ F .x 0 /
F .x/
for Q D C (Q D C n f0g, Q D int C , respectively).

Remark 15.4.18. Note that a lC -robust solution is as well sC -robust by definition.
The same assertion holds for a uC -robust solution.
The concept of sC -robustness can be interpreted in the following way: In a
situation where it is not clear if one should follow a risk affine or risk averse strategy
(e.g., the decision maker is not at hand or wants to get a feeling for the variety of
the solutions) this concept might be helpful as it calculates solutions which reflect
these different strategies. Therefore, this concept can serve as a pre-selection before
deciding a definite strategy.
Computing sC -robust solutions is possible with the help of the following
optimization problem:
!
inf2U y .f .x; // biobj
h.x/ WD ! v min (P Uy )
sup2U y .f .x; // x2S
with y 2 C C n f0g .y 2 C # , respectively). For this bi-objective optimization

problem we use the solution concept introduced in Definition 2.4.2. Especially, we
biobj
say that x 0 2 S is a weakly efficient element of the problem (P Uy ) if h.x 0 / 2
WMin.h.S /; R2C /. Moreover, x 0 2 S is called a strictly efficient element of the
biobj
problem (P Uy ), if there is no x 2 S n fx 0 g such that h.x/ 2 h.x 0 / R2C . We
show the following theorem:
Theorem 15.4.19. Consider an uncertain multi-objective optimization problem
P .U /. The following statements hold:
(a) If x 0 is strictly efficient for problem (P Uy ) for some y 2 C C n f0g, then x 0
biobj
is sC -robust.
(b) If x 0 is weakly efficient for problem (P Uy ) for some y 2 C C n f0g and

biobj
min2U y .f .x; // and max2U y .f .x; // exist for all x 2 S and the
chosen weight y 2 C C n f0g, then x 0 is sint C -robust.
(c) If x 0 is weakly efficient for problem (P Uy ) for some y 2 C # and
biobj

min2U y .f .x; // and max2U y .f .x; // exist for all x 2 S and the
chosen weight y 2 C # , then x 0 is sC nf0g -robust.
Proof. Let x 0 be strictly efficient (weakly efficient, weakly efficient) for problem
(P Uy ) with some y 2 C C n f0g (y 2 C C n f0g, y 2 C # , respectively), i.e.,
biobj
there is no xN 2 S n fx 0 g such that
inf y .f .x; // .<; <; respectively/ inf y .f .x 0 ; //

2U 2U
and sup y .f .x; // .<; <; respectively/ sup y .f .x 0 ; //:

2U 2U
Now suppose x 0 is not (sC j sint C j sC nf0g )-robust. Then there exists an xN 2
S n fx 0 g such that
N C Q F .x 0 / and F .x/
F .x/ N F .x 0 / Q
for Q D C (Q D int C , Q D C n f0g). That implies
9xN 2 S nfx 0 g W 81 ; 2 2 U 91 ; 2 2 U W f .x;

N 1 / C Q 3 f .x 0 ; 1 /
N 2 / 2 f .x 0 ; 2 / Q
and f .x; (15.44)
for Q D C (Q D int C , Q D C nf0g). Choose now y 2 C C nf0g (y 2 C C nf0g,

y 2 C # ) as in problem (P Uy ). We obtain from (15.44)
biobj
9xN 2 S nfx 0 gW81 ; 2 2 U 91 ; 2 2 U W y .f .x; 1 // . < ; < ; respectively/

y .f .x 0 ; 1 // and y .f .x; 2 // .<; <; respectively/y .f .x 0 ; 2 //
) inf y .f .x; // .<; <; respectively/ inf y .f .x 0 ; //
2U 2U
and sup y .f .x; // .<; <; respectively/ sup y .f .x 0 ; //:

2U 2U
The last two strict inequalities hold because the minimum and maximum exist. But
this is a contradiction to the assumption. t
u
Based on these observations, we can formulate the following algorithm for
computing sC -robust solutions to P .U /.
724 15 Applications
Algorithm 6 for computing (sC j sC nf0g j sint C )-robust solutions using a
biobj
family of problems (P Uy ):
Input & Step 1-3: Analogous to Algorithm 1.

Step 4: Find a set of weakly efficient solutions SOLwe .y / of (P Uy ).
biobj

Step 5: If SOLwe .y / D ;, then go to Step 2.
Step 6: Choose xN 2 SOLwe .y /. Set SOLwe .y / WD SOLwe .y / n fxg.
N
biobj
a) If x is a strictly efficient solution of (P Uy ), then x is sC -robust for P .U /,
thus
OptC WD OptC [ fxg:
b) If x is weakly efficient for problem (P Uy ) and y 2 C # and

biobj
min2U y .f .x; // and max2U y .f .x; // exist for all x 2 S and
the chosen weight y 2 C # , then x is sC nf0g -robust for P .U /, thus
OptC nf0g WD OptC nf0g [ fxg:
c) If x is a weakly efficient solution of (P Uy ) and max2U y .f .x; // and

biobj
min2U y .f .x; // exist for all x 2 S , then x is sint C -robust for P .U /,
thus
Optint C WD Optint C [ fxg:
In the following we present an algorithm that computes sC -robust solutions

biobj
while varying the weights in the vector of objectives of problem (P Uy ).
Algorithm 7 for computing (sC j sC nf0g j sint C )-robust solutions using a
biobj
family of problems (P Uy ):
Input & Step 1-3 & Step 5-7: Analogous to Algorithm 2, only replacing
biobj biobj
.P UyN / by .P UyN / and .P UyN CtkC1 .yO yN / / by .P UyN Ct .yO yN / /.
kC1
Step 4: Analogous to Step 4 of Algorithm 5.
15.4.4 Algorithms for Solving Special Classes of Set-Valued

In the following we will explain that the algorithms presented in Sect. 15.4 can be
used for solving special classes of set-valued optimization problems.
Having a close look at all the concepts of robustness, we can see that in fact all
of these are set-valued optimization problems.
Consider a set-valued optimization problem of the form
minimize F .x/; subject to x 2 S; (SP )
with some given preorder and a set-valued objective map F W S Y we can see
the following:
If the preorder is given by lC , uC , or sC with some proper closed convex
pointed cone C Y and F .x/ can be parametrized by parameters 2 U with
some set U in the way that
F .x/ WD ff .x; / j 2 U g;
where f W X U 7! Y , then the set-valued optimization problem .SP )

is equivalent to finding -robust solutions to the uncertain vector optimization
problem P .U / and can therefore be solved by using one of the respective
algorithms presented in Sect. 15.4.
We revealed strong connections between set-valued optimization and uncertain
vector optimization. Furthermore, we derived the results in a more general setting
as in [159] and [274]. In particular, we provided solution algorithms for a certain
class of set-valued optimization problems. It seems possible to extend this class of
problems to a more general one, but this is future work and of interest for the next
steps in this area of research.
Moreover, Sect. 15.4 makes clear that finding robust solutions to uncertain
vector optimization problems can be interpreted as an application of set-valued
optimization. Thus, robust solutions to uncertain vector optimization problems
can be obtained by using the solution techniques from set-valued optimization.
Formulating concrete algorithms of this kind is another topic for future research.
Appendix
Very often, order structures are defined on subsets of linear spaces. We will consider
only real linear spaces, so later on we shall omit the word real.
Let X be a linear space and ; ¤ X0 X ; recall that X0 is a linear subspace of
X if
x C y 2 X0 for all x; y 2 X0 and
; 2 R. This shows that X0 is stable
w.r.t. addition and multiplication by scalars, and X0 endowed with these operations
is itself a linear space. Also recall that the nonempty set M X is affine (or a
linear manifold) if
x C .1
/y 2 M for all x; y 2 M and
2 R; it is known
that M is affine iff there exist x0 2 X and X0 a linear subspace of X such that
M D x0 C X0 (X0 being called the parallel linear subspace to M ).
Definition A.1. Let X be a nonempty set. We say that X is a (real) linear space
or a set equipped with linear structure if an addition (that is, a mapping C W
X X ! X ) and a multiplication by scalars (that is, a mapping W R X ! X )
are defined satisfying the following conditions:
(a) 8 x; y; z 2 X W .x C y/ C z D x C .y C z/ (associativity),
(b) 8 x; y 2 X W x C y D y C x (commutativity),
(c) 9 0 2 X; 8 x 2 X W x C 0 D x (null element),
(d) 8 x 2 X; 9 x 0 2 X W x C x 0 D 0; we write x 0 D x,
(e) 8 x; y 2 X; 8
2 R W
.x C y/ D
x C
y,
(f) 8 x 2 X; 8
; 2 R W .
C /x D
x C x,
(g) 8 x 2 X; 8
; 2 R W
.x/ D .
/x,
(h) 8 x 2 X W 1x D x (unity element).
A linear space is often called a vector space. Defining x C y WD .x1 C
y1 ; : : : ; xn C yn / and
x WD .
x1 ; : : : ;
xn / for x; y 2 Rn and
2 R, we have
that Rn is a linear space.
In the sequel we shall use the following notation: Let A; B be nonempty subsets
of a linear space X , ; ¤ R; x 2 X , and
2 R; then
A C B WD fx C y j x 2 A; y 2 Bg; A C ; WD ; C A WD ;; x C A WD fxg C A;

DOI 10.1007/978-3-642-54265-7
728 Appendix
and
A WD f x j 2 ; x 2 Ag;
A D f
g A; ; D ;:
In particular, A B D A C .1/B D fx y j x 2 A; y 2 Bg. If, moreover, I is a

nonempty set and ; ¤ Ai X for every i 2 I , we have
! !
[ [ \ \

.A C B/ D
A C
B;
Ai D
Ai ;
Ai D
Ai :
i 2I i 2I i 2I i 2I
Often it is useful to have also a topological structure on the set under considera-
tion.
Definition A.2. Let X be a nonempty set. A topology on X is a family of subsets
of X satisfying the following conditions:
(a) every union of sets of belongs to ,
(b) every finite intersection of sets of belongs to ,
(c) the empty set ; and the whole set X belong to .
The elements of are called open sets, and X equipped with is called a
topological space and is denoted by .X; /.
Important examples of topological spaces are furnished by metric spaces. Recall
that a metric on a nonempty set X is a mapping W X X ! RC having the
properties
8 x; y 2 X W .x; y/ D 0 , x D y;
8 x; y 2 X W .x; y/ D .y; x/;
8 x; y; z 2 X W .x; z/ .x; y/ C .y; z/:
The set X equipped with the metric is called a metric space and is denoted by
.X; /.
Having the metric space .X; /, x 2 X and r > 0, consider the (closed) ball of
center x and radius r W
B.x; r/ WD B .x; r/ WD fy 2 X j .x; y/ rg:
Define
WD f;g [ fD X j 8 x 2 D; 9 r > 0 W B.x; r/ Dg:
It is easy to verify that is a topology on X , and so X becomes a topological

space; a metric space .X; / will always be endowed with the topology if not
Appendix 729
stated explicitly otherwise. Note that B.x; r/ is a closed set, while the open ball
fy 2 X j .x; y/ < rg is an open set.
Taking X D Rn and .x; y/ the usual Euclidean distance between x 2 Rn and
y 2 Rn , is a metric on Rn , and the open sets w.r.t. are the usual open sets of
Rn . As in Rn , in an arbitrary topological space .X; / we define other topological
notions, that of neighborhood being very useful.
Definition A.3. Let .X; / be a topological space and x 2 X . The subset U of X
is a neighborhood of x (relative to ) if there exists an open set O 2 such that
x 2 O U . The class of all neighborhoods of x will be denoted by N .x/, or
simply N .x/. A subset B.x/ of N .x/ is called a base of neighborhoods (filter
base) of x relative to if for every U 2 N .x/ there exists V 2 B.x/ such that
V U . One says that .X; / is first-countable if every element x 2 X has an (at
most) countable neighborhood base.
For the topological space .X; / and a neighborhood base B.x/ of x for every
x 2 X , the family fB.x/ j x 2 X g has the following properties:
(NB1) 8 x 2 X; 8 U 2 B.x/ W x 2 U ,
(NB2) 8 x 2 X , 8 U1 ; U2 2 B.x/; 9 U3 2 B.x/ W U3 U1 \ U2 ,
(NB3) 8 x 2 X; 8 U 2 B.x/; 9 V 2 B.x/; 8 y 2 V; 9 W 2 B.y/ W W U .
Conversely, if for every x 2 X one has a nonempty family B.x/ P.X /
such that fB.x/ j x 2 X g satisfies the conditions (NB1)–(NB3), then there exists
a unique topology on X such that B.x/ is a neighborhood base for x, for every
x 2 X ; moreover,
8 x 2 X W N .x/ D fU X j 9 V 2 B.x/ W V U g:
In fact, the topology .X; / of a metric space is constructed in this manner: The
family fB0 .x/ j x 2 X g with B0 .x/ WD fB .x; 1=n/ j n 2 N n f0gg satisfies
the conditions (NB1)–(NB3) above; the topology defined by fB0 .x/ j x 2 X g is
nothing else but the topology defined above. Since B0 .x/ is countable for every
x 2 X , we obtain that every metric space is first-countable.
As seen above (property (NB2)), taking B.x/ a neighborhood base of x 2 .X; /
and considering the relation on B.x/ defined by U1 U2 iff U1 U2 , then for
all U1 ; U2 2 B.x/ there exists U3 2 B.x/ such that U3 U1 and U3 U2 . This is
the prototype of a directed set. So, we say that the nonempty set I is directed by
if is a partial order on I and
8 i1 ; i2 2 I; 9 i3 2 I W i3 i1 and i3 i2 :
Note that if .I 0 ; 0 / and .I 00 ; 00 / are directed sets, then I 0 I 00 is directed by

the relation defined by .i 0 ; i 00 / .j 0 ; j 00 / iff i 0 0 j 0 and i 00 00 j 00 . Another
important example of directed set is N endowed with its usual order
. A subset J
of a directed set .I; / is called cofinal if for every i 2 I there exists j 2 J such
that j i ; it follows immediately that .J; / is directed if J is a cofinal subset of I .
730 Appendix
The above discussion shows that .B.x/; / is directed if B.x/ is a neighborhood

base for x 2 .X; /. This consideration is the base of the following definition.
Definition A.4. A mapping ' W I ! X , where .I; / is a directed set, is called
a net or a generalized sequence of X ; generally, the net ' is denoted by .xi /i 2I ,
where xi WD '.i /. When X is equipped with a topology we say that the net
.xi /i 2I X converges to x 2 X if
8 V 2 N .x/; 9 iV 2 I; 8 i iV W xi 2 V:

We denote this fact by .xi / ! x or simply .xi / ! x, and x is called a limit of .xi /.
Lemma A.5. Let .X; / be a topological space. Then every convergent net of X
has a unique limit if and only if
8 x; y 2 X; x ¤ y; 9 U 2 N .x/; V 2 N .y/ W U \ V D ;; (A.1)
i.e., .X; / is Hausdorff.

Proof. Suppose that (A.1) does not hold. Then there exist x; y 2 X , x ¤ y, such
that
8 U 2 N .x/; 8 V 2 N .y/; 9 xU;V 2 U \ V:
As seen above, I WD N .x/ N .y/ is directed, and so .xU;V /.U;V /2I is a net in X .

It is obvious that .xU;V / ! x and .xU;V / ! y, which shows that the limit of nets is
not unique in this case.
Suppose now that (A.1) holds and assume that the net .xi /i 2I X converges to
x and y with x ¤ y. Take U 2 N .x/ and V 2 N .y/ such that U \ V D ;. Then
there exist iU ; iV 2 I such that xi 2 U for i iU and xi 2 V for i iV . Because
I is directed, there exists i0 2 I such that i0 iU and i0 iV . We thus obtain the
contradiction that xi0 2 U \ V . The proof is complete. t
u
A useful notion is that of subnet; so, we say that the net .yk /k2K is a subnet of
the net .xi /i 2I if there exists a mapping W .K; / ! .I; / such that yk D x .k/
for every k 2 K and for every i 2 I there exists ki 2 K such that .k/ i for
k ki . If J is a cofinal subset of the directed set .I; /, then .xi /i 2J is a subnet of
the net .xi /i 2I . With the help of nets and subnets it is possible to characterize several
topological notions as the closure of a set, the cluster points, the compact subsets,
and the continuity of mappings between topological spaces in the same manner
as these notions are characterized in Rk by using sequences and subsequences.
For example, a subset A of the topological space .X; / is closed if and only if
x 2 A whenever the net .xi /i 2I A converges to x 2 X , but generally, taking
sequences is not sufficient; we say that A is sequentially closed if x 2 A whenever
the sequence .xn /n2N A converges to x 2 X .
Appendix 731
Recall that the function f W .X; / ! .Y; / is continuous at x 2 X if
8 V 2 N .f .x//; 9 U 2 N .x/; 8 x 0 2 U W f .x 0 / 2 V I
f is continuous if f is continuous at every x 2 X . When X is a linear space and

is a topology on X , it is important to have compatibility between the linear and
topological structures of X .
Definition A.6. Let X be a linear space endowed with a topology . We say that
.X; / is a topological linear space or topological vector space (t.l.s. or t.v.s. for
short) if both operations on X (the addition and the multiplication by scalars) are
continuous; in this case is called a linear topology on X .
Since these operations are defined on product spaces, we recall that for two
topological spaces .X1 ; 1 / and .X2 ; 2 /, there exists a unique topology on X1 X2 ,
denoted by 1 2 , with the property that
B.x1 ; x2 / WD fU1 U2 j U1 2 N1 .x1 /; U2 2 N2 .x2 /g
is a neighborhood base of .x1 ; x2 / w.r.t. 1 2 for every .x1 ; x2 / 2 X1 X2 ;

1 2 is called the product topology on X1 X2 . Of course, in Definition A.6
the topology on X X is , and the topology on R X is 0 , where
0 is the usual topology of R. It is easy to see that when .X; / is a topological
linear space, a 2 X and
2 R n f0g, the mappings Ta , H
W X ! X defined by
Ta .x/ D aCx, H
.x/ WD
x, are bijective and continuous with continuous inverses
i.e., they are homeomorphisms). It follows, that V 2 N .0/ iff a C V 2 N .a/
for every a 2 X and every neighborhood of 0 is absorbing (A X is absorbing
if for every x 2 X there exists ı > 0 such that Œı; ı x A, or equivalently, for
every x 2 X there exists ı > 0 such that Œ0; ı x A). It is well known that the
origin 0 of a t.v.s. .X; / has a neighborhood base formed by balanced absorbing
sets (; ¤ A X is balanced if Œ1; 1 A D A). When X is a t.v.s., the class of
all balanced neighborhoods of 0 2 X will be denoted by NX . Taking into account
the formula
\
cl A D fA C V j V 2 Bg (A.2)
for the closure of a subset A of the t.l.s. .X; /, where B is a neighborhood base for
0, one obtains that the origin of a topological linear space has a neighborhood base
formed by closed, balanced, and absorbing sets.
Note also that a topological linear space .X; / is Hausdorff iff for every x 2
X n f0g there exists V 2 NX such that x … V , or equivalently (see (A.2)), the set
f0g is closed.
Further examples of topological linear spaces are given in the next section.
References
1. Abadie, J.: On the Kuhn-Tucker theorem. In: Nonlinear Programming (NATO Summer
School, Menton, 1964), pp. 19–36. North-Holland, Amsterdam (1967)
2. Abdouni, B.E., Thibault, L.: Optimality conditions for problems with set-valued objectives.
J. Appl. Anal. 2, 183–201 (1996)
3. Alonso, M., Rodríguez-Marín, L.: Set-relations and optimality conditions in set-valued maps.
Nonlinear Anal. 63(8), 1167–1179 (2005)
4. Altangerel, L., Boţ, R.I., Wanka, G.: Conjugate duality in vector optimization and some
applications to the vector variational inequality. J. Math. Anal. Appl. 329(2), 1010–1035
(2007)
5. Amahroq, T., Taa, A.: Sufficient conditions of optimality for multiobjective optimization
problems with -paraconvex data. Studia Math. 124(3), 239–247 (1997)
6. Amahroq, T., Thibault, L.: On proto-differentiability and strict proto-differentiability of
multifunctions of feasible points in perturbed optimization problems. Numer. Funct. Anal.
Optim. 16, 1293–1307 (1995)
7. Anger, B., Lembcke, J.: Hahn-Banach type theorems for hypolinear functionals. Math. Ann.
209, 127–151 (1974)
8. Anh, N.L., Khanh, P.Q., Tung, L.T.: Variational sets: calculus and applications to nonsmooth
vector optimization. Nonlinear Anal. 74(6), 2358–2379 (2011)
9. Apetrii, M., Durea, M., Strugariu, R.: On subregularity properties of set-valued mappings. Set
Valued Var. Anal. 21(1), 93–126 (2013)
10. Aragón Artacho, F.J., Mordukhovich, B.S.: Enhanced metric regularity and Lipschitzian
properties of variational systems. J. Global Optim. 50(1), 145–167 (2011)
11. Arrow, K.J., Hurwicz, L., Uzawa, H.: Studies in linear and non-linear programming. With
contributions by H. B. Chenery, S. M. Johnson, S. Karlin, T. Marschak, R. M. Solow. Stanford
Mathematical Studies in the Social Sciences, vol. II. Stanford University Press, Stanford,
California (1958)
12. Artzner, P., Delbaen, F., Eber, J.M., Heath, D.: Coherent measures of risk. Math. Finance 9,
203–228 (1999)
13. Attouch, H., Riahi, H.: Stability results for Ekeland’s -variational principle and cone
extremal solutions. Math. Oper. Res. 18(1), 173–201 (1993)
14. Aubin, J.P.: Contingent derivatives of set-valued maps and existence of solutions to nonlinear
inclusions and differential inclusions. In: Mathematical analysis and applications, Part A,
Adv. in Math. Suppl. Stud., vol. 7, pp. 159–229. Academic Press, New York (1981)
15. Aubin, J.P., Cellina, A.: Differential inclusions. Set-Valued Maps and Viability Theory
[Grundlehren der Mathematischen Wissenschaften (Fundamental Principles of Mathematical
Sciences)], vol. 264. Springer, Berlin (1984)

DOI 10.1007/978-3-642-54265-7
734 References
16. Aubin, J.P., Ekeland, I.: Applied nonlinear analysis. Pure and Applied Mathematics
(New York). Wiley, New York (1984)
17. Aubin, J.P., Frankowska, H.: On inverse function theorems for set-valued maps. J. Math.
Pures Appl. (9) 66(1), 71–89 (1987)
18. Aubin, J.P., Frankowska, H.: Set-valued analysis. Systems & Control: Foundations &
Applications, vol. 2. Birkhäuser Boston, Boston (1990)
19. Auslender, A., Cominetti, R.: A comparative study of multifunction differentiability with
applications in mathematical programming. Math. Oper. Res. 16(2), 240–258 (1991)
20. Avakov, E.R.: Necessary conditions for an extremum for smooth abnormal problems with
constraints of equality and inequality type. Mat. Zametki 45(6), 3–11, 110 (1989)
21. Azé, D.: A unified theory for metric regularity of multifunctions. J. Convex Anal. 13(2),
225–252 (2006)
22. Baier, J., Jahn, J.: On subdifferentials of set-valued maps. J. Optim. Theory Appl. 100, 233–
240 (1999)
23. Banks, H.T., Jacobs, M.Q.: A differential calculus for multifunctions. J. Math. Anal. Appl.
29, 246–272 (1970)
24. Bao, T.Q., Mordukhovich, B.S.: Existence of minimizers and necessary conditions in set-
valued optimization with equilibrium constraints. Appl. Math. 52, 452–562 (2007)
25. Bao, T.Q., Mordukhovich, B.S.: Variational principles for set-valued mappings with applica-
tions to multiobjective optimization. Control Cybern. 36(3), 531–562 (2007)
26. Bao, T.Q., Mordukhovich, B.S.: Necessary conditions for super minimizers in constrained
multiobjective optimization. J. Global Optim. 43(4), 533–552 (2009)
27. Bao, T.Q., Mordukhovich, B.S.: Relative Pareto minimizers for multiobjective problems:
existence and optimality conditions. Math. Program. 122(2, Ser. A), 301–347 (2010)
28. Bao, T.Q., Mordukhovich, B.S.: Set-valued optimization in welfare economics. Adv. Math.
Econ. 13, 113–153 (2010)
29. Bao, T.Q., Tammer, C.: Lagrange necessary conditions for Pareto minimizers in Asplund
spaces and applications. Nonlinear Anal. 75(3), 1089–1103 (2012)
30. Barbagallo, A., Cojocaru, M.G.: Continuity of solutions for parametric variational inequalities
in Banach space. J. Math. Anal. Appl. 351(2), 707–720 (2009)
31. Barber, C.B., Dobkin, D.P., Huhdanpaa, H.: The quickhull algorithm for convex hulls. ACM
Trans. Math. Software 22(4), 469–483 (1996)
32. Barbu, V., Precupanu, T.: Convexity and optimization in Banach spaces, Mathematics and its
Applications (East European Series), vol. 10, romanian edn. D. Reidel Publishing, Dordrecht
(1986)
33. Bauschke, H.H., Combettes, P.L.: Convex analysis and monotone operator theory in Hilbert
spaces, with a foreword by Hédy Attouch. CMS Books in Mathematics/Ouvrages de
Mathématiques de la SMC. Springer, New York (2011)
34. Bazaraa, M., Goode, J., Shetty, C.: Constraint qualifications revisited. Manag. Sci. 18, 567–
573 (1972)
35. Bazaraa, M.S., Goode, J.J.: Necessary optimality criteria in mathematical programming in
normed linear spaces. J. Optim. Theory Appl. 11, 235–244 (1973)
36. Bazaraa, M.S., Goode, J.J., Nashed, M.Z.: On the cones of tangents with applications to
mathematical programming. J. Optim. Theory Appl. 13, 389–426 (1974)
37. Bazaraa, M.S., Shetty, C.M., Goode, J.J., Nashed, M.Z.: Nonlinear programming without dif-
ferentiability in Banach spaces: necessary and sufficient constraint qualification. Applicable
Anal. 5(3), 165–173 (1975/76)
38. Bazarra, M., Shetty, C.: Foundations of Optimization. Springer, New York (1976)
39. Bednarczuk, E.M., Song, W.: Contingent epiderivative and its applications to set-valued
optimization. Control Cybern. 27, 375–386 (1998)
40. Bednarczuk, E.M., Zagrodny, D.: Vector variational principle. Arch. Math. (Basel) 93(6),
577–586 (2009)
41. Bellaassali, S., Jourani, A.: Lagrange multipliers for multiobjective programs with a general
preference. Set Valued Anal. 16(2-3), 229–243 (2008)
References 735
42. Ben-Israel, A., Ben-Tal, A., Zlobec, S.: Optimality in Nonlinear Programming: A Feasible
Directions Approach. Wiley, New York (1981)
43. Ben-Tal, A., Ghaoui, L.E., Nemirovski, A.: Robust Optimization. Princeton University Press,
Princeton and Oxford (2009)
44. Ben-Tal, A., Nemirovski, A.: Robust convex optimization. Math. Oper. Res. 23(4), 769–805
(1998)
45. Ben-Tal, A., Zowe, J.: A unified theory of first and second order conditions for extremum
problems in topological vector spaces. Math. Program. Stud. (19), 39–76 (1982)
46. Benson, H.: An outer approximation algorithm for generating all efficient extreme points in
the outcome set of a multiple objective linear programming problem. J. Global Optim. 13,
1–24 (1998)
47. Benson, H.P.: On a domination property for vector maximization with respect to cones. J.
Optim. Theory Appl. 39(1), 125–132 (1983)
48. Bianchi, M., Kassay, G., Pini, R.: Ekeland’s principle for vector equilibrium problems.
Nonlinear Anal. 66(7), 1454–1464 (2007)
49. Bigolin, F., Greco, G.H.: Geometric characterizations of C 1 manifolds in Euclidean spaces
by tangent cones. J. Math. Anal. Appl. 396(1), 145–163 (2012)
50. Birge, J.R., Louveaux, F.: Introduction to Stochastic Programming, 2nd edn. Springer Series
in Operations Research and Financial Engineering. Springer, New York (2011)
51. Bitran, G.R., Magnanti, T.L.: The structure of admissible points with respect to cone
dominance. J. Optim. Theory Appl. 29(4), 573–614 (1979)
52. Boţ, R., Wanka, G.: An analysis of some dual problems in multiobjective optimization. I, II.
Optimization 53(3), 281–324 (2004)
53. Bonnans, J.F., Shapiro, A.: Perturbation analysis of optimization problems. Springer Series
in Operations Research. Springer, New York (2000)
54. Bonnisseau, J.M., Cornet, B.: Existence of equilibria when firms follow bounded losses
pricing rules. J. Math. Econ. 17, 119–147 (1988)
55. Bonnisseau, J.M., Crettez, B.: On the characterization of efficient production vectors. Econ.
Theory 31, 213–223 (2007)
56. Borwein, J.M.: Multivalued convexity and optimization: A unified approach to inequality and
equality constraints. Math. Program. 13, 183–199 (1977)
57. Borwein, J.M.: Proper efficient points for maximizations with respect to cones. SIAM J.
Control Optim. 15(1), 57–63 (1977)
58. Borwein, J.M.: Tangent cones, starshape and convexity. Internat. J. Math. Math. Sci. 1(4),
459–477 (1978)
59. Borwein, J.M.: Weak tangent cones and optimization in a Banach space. SIAM J. Control
Optim. 16(3), 512–522 (1978)
60. Borwein, J.M.: Stability and regular points of inequality systems. J. Optim. Theory Appl.
48(1), 9–52 (1986)
61. Borwein, J.M.: Convex cones, minimality notions, and consequences. In: Recent Advances
and Historical Development of Vector Optimization (Darmstadt, 1986), Lecture Notes in
Economics and Mathematical Systems, vol. 294, pp. 62–85. Springer, Berlin (1987)
62. Borwein, J.M.: Epi-Lipschitz-like sets in Banach space: theorems and examples. Nonlinear
Anal. 11(10), 1207–1217 (1987)
63. Borwein, J.M., Lucet, Y., Mordukhovich, B.: Compactly epi-Lipschitzian convex sets and
functions in normed spaces. J. Convex Anal. 7(2), 375–393 (2000)
64. Borwein, J.M., O’Brien, R.: Tangent cones and convexity. Can. Math. Bull. 19(3), 257–261
(1976)
65. Borwein, J.M., Strójwas, H.M.: The hypertangent cone. Nonlinear Anal. 13(2), 125–144
(1989)
66. Borwein, J.M., Zhu, Q.J.: Techniques of variational analysis. CMS Books in Mathemat-
ics/Ouvrages de Mathématiques de la SMC, vol. 20. Springer, New York (2005)
67. Borwein, J.M.: Continuity and differentiability properties of convex operators. Proc. Lond.
Math. Soc. (3) 44(3), 420–444 (1982)
736 References
68. Borwein, J.M.: On the existence of Pareto efficient points. Math. Oper. Res. 8(1), 64–73
(1983)
69. Borwein, J.M., Lewis, A.S.: Convex analysis and nonlinear optimization. Theory and
examples. CMS Books in Mathematics/Ouvrages de Mathématiques de la SMC. 3. Springer,
New York (2000)
70. Borwein, J.M., Strójwas, H.M.: Tangential approximations. Nonlinear Anal. 9(12), 1347–
1366 (1985)
71. Borwein, J.M., Zhuang, D.: Super efficiency in vector optimization. Trans. Am. Math. Soc.
338(1), 105–122 (1993)
72. Boţ, R.I., Csetnek, E., Wanka, G.: Regularity conditions via quasi-relative interior in convex
programming. SIAM J. Optim. 19, 217–233 (2008)
73. Boţ, R.I., Grad, S.M., Wanka, G.: Duality in vector optimization. Vector Optimization.
Springer, Berlin (2009)
74. Bouligand, G.: Sur l’existence des demi-tangentes à une courbe de Jordan. Fundamenta Math.
15, 215–218 (1030)
75. Bouligand, G.: Introduction á la géométrie Infinitseminale Directe. Gauthier-Villars, Paris
(1932)
76. Bourbaki, N.: Eléments de mathématique. XVIII. Première partie: Les structures fonda-
mentales de l’analyse. Livre V: Espaces vectoriels topologiques. Chapitre III: Espaces
d’applications linéaires continues. Chapitre IV: La dualité dans les espaces vectoriels
topologiques. Chapitre V: Espaces hilbertiens. Actualités Sci. Ind., no. 1229. Hermann &
Cie, Paris (1955)
77. Branke, J.: Creating robust solutions by means of evolutionary algorithms. In: Eiben, E.,
Bäck, T., Schenauer, M., Schwefel, H.P. (eds.) Parallel Problem Solving from Nature, Lecture
Notes in Computer Science, pp. 119–128. Springer, Berlin, Heidelberg (1998)
78. Breckner, W.W.: Dualität bei Optimierungsaufgaben in halbgeordneten topologischen Vek-
torräumen. i. Math. Rev. Anal. Numér. Théor. Approx. 1, 5–35 (1972)
79. Breckner, W.W., Kassay, G.: A systematization of convexity concepts for sets and functions.
J. Convex Anal. 4(1), 109–127 (1997)
80. Bremner, D., Fukuda, K., Marzetta, A.: Primal-dual methods for vertex and facet enumeration.
Discrete Comput. Geom. 20(3), 333–357 (1998)
81. Brumelle, S.L.: Convex operators and supports. Math. Oper. Res. 3(2), 171–175 (1978)
82. Cambini, A., Martein, L., Vlach, M.: Second order tangent sets and optimality conditions.
Math. Jpn. 49(3), 451–461 (1999)
83. Canon, M.D., Cullum Jr., C.D., Polak, E.: Theory of Optimal Control and Mathematical
Programming. McGraw-Hill, New York (1970)
84. Castellani, M.: A necessary second-order optimality condition in nonsmooth mathematical
programming. Oper. Res. Lett. 19(2), 79–86 (1996)
85. Castellani, M., Pappalardo, M.: First-order cone approximations and necessary optimality
conditions. Optimization 35(2), 113–126 (1995)
86. Castellani, M., Pappalardo, M.: Local second-order approximations and applications in
optimization. Optimization 37(4), 305–321 (1996)
87. Cesari, L., Suryanarayana, M.B.: Existence theorems for Pareto optimization; multivalued
and Banach space valued functionals. Trans. Am. Math. Soc. 244, 37–65 (1978)
88. Chai, Y., Liu, S., Li, Y.: The existence of contingent epiderivative for a set-valued mapping
and vector variational-like inequalities. J. Inequal. Appl. 2013(352), 13 (2013)
89. Chen, C.R., Li, S.J.: Different conjugate dual problems in vector optimization and their
relations. J. Optim. Theory Appl. 140(3), 443–461 (2009)
90. Chen, G.Y., Goh, C.J., Yang, X.Q.: On gap functions for vector variational inequalities. In:
Vector variational inequalities and vector equilibria, Nonconvex Optim. Appl., vol. 38, pp.
55–72. Kluwer Academic, Dordrecht (2000)
91. Chen, G.Y., Huang, X., Yang, X.: Vector optimization. Lecture Notes in Economics and
Mathematical Systems, vol. 541. Springer, Berlin (2005)
References 737
92. Chen, G.Y., Jahn, J.: Optimality conditions for set-valued optimization problems. Math.
Methods Oper. Res. 48(2), 187–200 (1998)
93. Chen, G.Y., Wang, Y.Y.: Generalized Hahn-Banach theorems and subdifferential of set-valued
mapping. J. Syst. Sci. Math. Sci. 5(3), 223–230 (1985)
94. Chen, X., Nashed, Z., Qi, L.: Convergence of Newton’s method for singular smooth and
nonsmooth equations using adaptive outer inverses. SIAM J. Optim. 7(2), 445–462 (1997)
95. Chew, K.L.: Maximal points with respect to cone dominance in Banach spaces and their
existence. J. Optim. Theory Appl. 44(1), 1–53 (1984)
96. Chiang, Y., Yao, J.C.: Vector variational inequalities and the .S/C condition. J. Optim.
Theory Appl. 123(2), 271–290 (2004)
97. Chuong, T.D.: Clarke coderivatives of efficient point multifunctions in parametric vector
optimization. Nonlinear Anal. 74(1), 273–285 (2011)
98. Chuong, T.D., Yao, J.C.: Coderivatives of efficient point multifunctions in parameteric vector
optimization. Taiwanese J. Math. 13(6A), 1671–1693 (2009)
99. Chuong, T.D., Yao, J.C.: Generalized Clarke epiderivatives of parametric vector optimization
problems. J. Optim. Theory Appl. 146(1), 77–94 (2010)
100. Clarke, F.H.: Optimization and nonsmooth analysis. Canadian Mathematical Society Series
of Monographs and Advanced Texts. Wiley, New York (1983)
101. Clarke, F.H., Stern, R.J., Wolenski, P.R.: Subgradient criteria for monotonicity, the Lipschitz
condition, and convexity. Can. J. Math. 45(6), 1167–1183 (1993)
102. Combari, C., Laghdir, M., Thibault, L.: Sous-différentiels de fonctions convexes composées.
Ann. Sci. Math. Québec 18(2), 119–148 (1994)
103. Cominetti, R.: Metric regularity, tangent sets, and second-order optimality conditions. Appl.
Math. Optim. 21(3), 265–287 (1990)
104. Constantin, E.: Higher order necessary and sufficient conditions for optimality. Panamer.
Math. J. 14(3), 1–25 (2004)
105. Constantin, E.: An application of higher order tangent cones to flow-invariance. Int. J. Appl.
Math. Sci. 2(1), 121–129 (2005)
106. Constantin, E.: Higher order necessary conditions in smooth constrained optimization. In:
Communicating Mathematics, Contemp. Math., vol. 479, pp. 41–49. Amer. Math. Soc.,
Providence, RI (2009)
107. Constantin, E.: Second-order optimality conditions for problems with locally Lipschitz data
via tangential directions. Comm. Appl. Nonlinear Anal. 18(2), 75–84 (2011)
108. Conway, J.B.: A course in functional analysis, Graduate Texts in Mathematics, vol. 96, 2nd
edn. Springer, New York (1990)
109. Corley, H.: Optimality conditions for maximization of set-valued functions. J. Optim. Theory
Appl. 58, 1–10 (1988)
110. Corley, H.W.: An existence result for maximizations with respect to cones. J. Optim. Theory
Appl. 31(2), 277–281 (1980)
111. Corley, H.W.: Duality theory for maximizations with respect to cones. J. Math. Anal. Appl.
84, 560–568 (1981)
112. Corley, H.W.: Existence and Lagrangian duality for maximizations of set-valued functions. J.
113. Cornet, B.: Regularity properties of open tangent cones. Math. Program. Stud. (30), 17–33
(1987)
114. Correa, R., Jofré, A., Thibault, L.: Characterization of lower semicontinuous convex func-
tions. Proc. Am. Math. Soc. 116(1), 67–72 (1992)
115. Correa, R., Jofré, A., Thibault, L.: Subdifferential monotonicity as characterization of convex
functions. Numer. Funct. Anal. Optim. 15(5-6), 531–535 (1994)
116. Crespi, G.P., Ginchev, I., Rocca, M.: First-order optimality conditions in constrained set-
valued optimization. Pac. J. Optim. 2(2), 225–239 (2006)
117. Crespi, G.P., Ginchev, I., Rocca, M.: First-order optimality conditions in set-valued optimiza-
tion. Math. Methods Oper. Res. 63(1), 87–106 (2006)
738 References
118. Crespi, G.P., Ginchev, I., Rocca, M.: Existence of solutions of Minty type scalar and vector
variational inequalities. Optimization 58(7), 791–808 (2009)
119. Dancs, S.: Generalized tangent cone and an optimization problem in a normed space. J.
120. Dancs, S., Hegedüs, M., Medvegyev, P.: A general ordering and fixed-point principle in
complete metric space. Acta Sci. Math. (Szeged) 46(1-4), 381–388 (1983)
121. Daniele, P.: Dynamic networks and evolutionary variational inequalities. New Dimensions in
Networks. Edward Elgar Publishing, Cheltenham (2006)
122. Daniele, P., Maugeri, A.: Vector variational inequalities and modelling of a continuum traffic
equilibrium problem. In: Vector variational inequalities and vector equilibria, Nonconvex
Optim. Appl., vol. 38, pp. 97–111. Kluwer Academic, Dordrecht (2000)
123. Daniilidis, A., Hadjisavvas, N.: Characterization of nonsmooth semistrictly quasiconvex and
strictly quasiconvex functions. J. Optim. Theory Appl. 102(3), 525–536 (1999)
124. Daniilidis, A., Pang, J.C.H.: Continuity and differentiability of set-valued maps revisited in
the light of tame geometry. J. Lond. Math. Soc. (2) 83(3), 637–658 (2011)
125. Das, P.C.: Application of Dubovitskii-Milyutin formalism to optimal settling problem with
constraints. In: Optimization and optimal control (Proc. Conf., Oberwolfach, 1974), pp. 44–
58. Lecture Notes in Math., Vol. 477. Springer, Berlin (1975)
126. Dauer, J., Saleh, O.: A characterization of proper minimal points as solutions of sublinear
optimization problems. J. Math. Anal. Appl. 178, 227–246 (1993)
127. Day, M.M.: Normed Linear Spaces, 3rd edn. Springer, New York (1973). Ergebnisse der
Mathematik und ihrer Grenzgebiete, Band 21
128. De Blasi, F.S.: On the differentiability of multifunctions. Pac. J. Math. 66(1), 67–81 (1976)
129. Dedieu, J.P.: Critères de fermeture pour l’image d’un fermé non convexe par une multiappli-
cation. C. R. Acad. Sci. Paris Sér. A-B 287(14), A941–A943 (1978)
130. Dellnitz, M., Witting, K.: Computation of robust Pareto points. Int. J. Comput. Sci. Math.
2(3), 243–266 (2009)
131. Demyanov, V.F., Rubinov, A.M.: Constructive nonsmooth analysis. Approximation & Opti-
mization, vol. 7. Peter Lang, Frankfurt am Main (1995)
132. Dentcheva, D.: Approximations, expansions and univalued representations of multifunctions.
Nonlinear Anal. 45(1, Ser. A: Theory Methods), 85–108 (2001)
133. Dep, K., Gupta, H.: Introducing robustness in multi-objective optimization. Evol. Comput.
14(4), 463–494 (2006)
134. Di, S.: Classical optimality conditions under weaker assumptions. SIAM J. Optim. 6(1),
178–197 (1996)
135. Di, S., Poliquin, R.: Contingent cone to a set defined by equality and inequality constraints at
a Fréchet differentiable point. J. Optim. Theory Appl. 81(3), 469–478 (1994)
136. Di, S., Poliquin, R.A.: Optimality conditions for max-type nonsmooth minimization prob-
lems. Optimization 41(3), 219–236 (1997)
137. Dias, S., Smirnov, G.: On the Newton method for set-valued maps. Nonlinear Anal. 75(3),
1219–1230 (2012)
138. Dien, P.H.: Locally Lipschitzian set-valued maps and generalized extremal problems with
inclusion constraints. Acta Math. Vietnam. 8(1), 109–122 (1984) (1983)
139. Djafari Rouhani, B., Khan, A.A.: On the embedding of variational inequalities. Proc. Am.
Math. Soc. 131(12), 3861–3871 (2003)
140. Dolecki, S.: Tangency and differentiation: some applications of convergence theory. Ann.
Mat. Pura Appl. (4) 130, 223–255 (1982)
141. Dolecki, S.: Tangency and differentiation: marginal functions. Adv. Appl. Math. 11(4), 389–
411 (1990)
142. Dolecki, S., Greco, G.H.: Towards historical roots of necessary conditions of optimality:
regula of Peano. Control Cybern. 36(3), 491–518 (2007)
143. Dolecki, S., Greco, G.H.: Tangency vis-à-vis differentiability by Peano, Severi and Guareschi.
J. Convex Anal. 18(2), 301–339 (2011)
References 739
144. Dolecki, S., Malivert, C.: Polarities and stability in vector optimization. In: Recent Advances
and Historical Development of Vector Optimization (Darmstadt, 1986), Lecture Notes in
Economics and Mathematical Systems, vol. 294, pp. 96–113. Springer, Berlin (1987)
145. Dolecki, S., Malivert, C.: General duality in vector optimization. Optimization 27(1-2), 97–
119 (1993)
146. Dontchev, A.L., Rockafellar, R.T.: Implicit functions and solution mappings. Springer
Monographs in Mathematics. Springer, Dordrecht (2009)
147. Du, W.S.: The extension of continuity and generalized results for set-valued maps in convex
analysis. Nonlinear Anal. 71(7-8), 3176–3184 (2009)
148. Dubovickiı̆, A.J., Miljutin, A.A.: Extremal problems with constraints. Ž. Vyčisl. Mat. i Mat.
Fiz. 5, 395–453 (1965)
149. Durea, M., Dutta, J.: Lagrange multipliers for Pareto minima in general Banach spaces. Pac.
J. Optim. 4, 447–463 (2008)
150. Durea, M., Dutta, J., Tammer, C.: Bounded sets of Lagrange multipliers for vector optimiza-
tion problems in infinite dimension. J. Math. Anal. Appl. 348, 589–606 (2008)
151. Durea, M., Dutta, J., Tammer, C.: Lagrange multipliers for -pareto solutions in vector
optimization with nonsolid cones in banach spaces. J. Optim. Theory Appl. 145, 196–211
(2010)
152. Durea, M., Dutta, J., Tammer, C.: Stability properties of KK T points in vector optimization.
Optimization 60(7), 823–838 (2011)
153. Durea, M., Strugariu, R.: On parametric vector optimization via metric regularity of constraint
systems. Math. Methods Oper. Res. 74(3), 409–425 (2011)
154. Durea, M., Strugariu, R.: Openness stability and implicit multifunction theorems: applications
to variational systems. Nonlinear Anal. 75(3), 1246–1259 (2012)
155. Durea, M., Strugariu, R.: Calculus of tangent sets and derivatives of set-valued maps under
metric subregularity conditions. J. Global Optim. 56(2), 587–603 (2013)
156. Durea, M., Tammer, C.: Fuzzy necessary optimality conditions for vector optimization
problems. Optimization 58, 449–467 (2009)
157. Dutta, J., Tammer, C.: Lagrangian conditions for vector optimization in Banach spaces. Math.
Methods Oper. Res. 64, 521–541 (2006)
158. Ehrgott, M.: Multicriteria Optimization, 2nd edn. Springer, Berlin (2005)
159. Ehrgott, M., Ide, J., Schöbel, A.: Minmax robustness for multi-objective optimization
problems. Eur. J. Oper. Res. 239, 17–31 (2014)
160. Ehrgott, M., Löhne, A., Shao, L.: A dual variant of Benson’s outer approximation algorithm.
J. Glob. Optim. 52(4), 757–778 (2012)
161. Eichfelder, G.: Adaptive scalarization methods in multiobjective optimization. Vector
Optimization. Springer, Berlin (2008)
162. Eichfelder, G.: Optimal elements in vector optimization with a variable ordering structure. J.
163. Eichfelder, G.: Variable ordering structures in vector optimization. Habilitation Thesis,
University Erlangen-Nürnberg (2011)
164. Eichfelder, G.: Variable ordering structures in vector optimization. In: Recent Developments
in Vector Optimization, pp. 95–126. Springer, Berlin (2012)
165. Eichfelder, G.: Numerical procedures in multiobjective optimization with variable ordering
structures. J. Optim. Theory Appl. (2013). DOI:10.1007/s10957-013-0267-y
166. Eichfelder, G.: Variable Ordering Structures in Vector Optimization. Springer, Berlin (2014)
167. Eichfelder, G., Gerlach, T.: Characterization of proper optimal elements with variable
ordering structures. Preprint-Series of the Institute of Mathematics, Ilmenau University of
Technology, Germany No. M 14/01 (2014)
168. Eichfelder, G., Ha, T.X.D.: Optimality conditions for vector optimization problems with
variable ordering structures. Optimization 62(5), 597–627 (2013)
169. Eichfelder, G., Jahn, J.: Vector optimization problems and their solution concepts. In: Recent
Developments in Vector Optimization, pp. 1–27. Springer, Berlin (2012)
740 References
170. Eichfelder, G., Kasimbeyli, R.: Properly optimal elements in vector optimization with variable
ordering structures. J. Global Optim. (2013). DOI:10.1007/s10898-013-0132-4
171. Ekeland, I.: On the variational principle. J. Math. Anal. Appl. 47, 324–353 (1974)
172. Ekeland, I., Temam, R.: Convex Analysis and Variational Problems. North-Holland,
Amsterdam (1976)
173. El Ghaoui, L., Lebret, H.: Robust solutions to least-squares problems with uncertain data.
SIAM J. Matrix Anal. Appl. 18(4), 1035–1064 (1997)
174. El Maghri, M.: Pareto-Fenchel -subdifferential sum rule and -efficiency. Optim. Lett. 6(4),
763–781 (2012)
175. El Maghri, M., Laghdir, M.: Pareto subdifferential calculus for convex vector mappings and
applications to vector optimization. SIAM J. Optim. 19(4), 1970–1994 (2008)
176. Ellaia, R., Hassouni, A.: Characterization of nonsmooth functions through their generalized
gradients. Optimization 22(3), 401–416 (1991)
177. Elster, K.H., Elster, R.: Vector variational inequality and geometric vector optimization. In:
Variational Inequalities and Network Equilibrium Problems (Erice, 1994), pp. 55–67. Plenum,
New York (1995)
178. Elster, K.H., Thierfelder, J.: On cone approximation of sets. In: Mathematical Methods in
Operations Research, Lectures Presented at the Summer School of Opearetions Research,
Primorsko, pp. 33–59 (1984)
179. Elster, K.H., Thierfelder, J.: The general concept of cone approximations in nondifferentiable
optimization. In: Nondifferentiable Optimization: Motivations and Applications (Sopron,
1984), Lecture Notes in Econom. and Math. Systems, vol. 255, pp. 170–189. Springer, Berlin
(1985)
180. Elster, K.H., Thierfelder, J.: Abstract cone approximations and generalized differentiability
in nonsmooth optimization. Optimization 19(3), 315–341 (1988)
181. Elster, K.H., Thierfelder, J.: On cone approximations and generalized directional derivatives.
In: Nonsmooth Optimization and Related Topics (Erice, 1988), Ettore Majorana Internat. Sci.
Ser. Phys. Sci., vol. 43, pp. 133–154. Plenum, New York (1989)
182. Federer, H.: Curvature measures. Trans. Am. Math. Soc. 93, 418–491 (1959)
183. Fernández, F.R., Hinojosa, M.A., Puerto, J.: Set-valued TU-games. Eur. J. Oper. Res. 159(1),
181–195 (2004)
184. Ferro, F.: An optimization result for set-valued mappings and a stability property in vector
problems with constraints. J. Optim. Theory Appl. 90(1), 63–77 (1996)
185. Ferro, F.: Optimization and stability results through cone lower semicontinuity. Set Valued
Anal. 5(4), 365–375 (1997)
186. Fischetti, M., Monaci, M.: Light robustness. In: Ahuja, R.K., Moehring, R., Zaroliagis, C.
(eds.) Robust and Online Large-Scale Optimization, no. 5868 in Lecture Notes in Computer
Science, pp. 61–84. Springer, Berlin/Heidelberg (2009)
187. Flett, T.M.: Differential Analysis. Cambridge University Press, Cambridge (1980)
188. Flores-Bazan, F.: Optimality conditions in non-convex set-valued optimization. Math.
Methods Oper. Res. 53, 403–417 (2001)
189. Flores-Bazán, F., Hernández, E., Novo, V.: Characterizing efficiency without linear structure:
a unified approach. J. Global Optim. 41(1), 43–60 (2008)
190. Föllmer, H., Schied, A.: Convex measures of risk and trading constraints. Finance Stoch. 4,
429–447 (2002)
191. Föllmer, H., Schied, A.: Stochastic Finance. Walter de Gruyter, Berlin (2004)
192. Furukawa, N., Yoshinaga, Y.: Higher-order variational sets, variational derivatives and higher-
order necessary conditions in abstract mathematical programming. Bull. Inform. Cybern.
23(1-2), 9–40 (1988)
193. Gajek, L., Zagrodny, D.: Countably orderable sets and their applications in optimization.
Optimization 26(3-4), 287–301 (1992)
194. Gajek, L., Zagrodny, D.: Weierstrass theorem for monotonically semicontinuous functions.
Optimization 29(3), 199–203 (1994)
References 741
195. Gast, S.: Network flow problems with uncertain input data in the context of supply chain
management applications. Ph.D. thesis, University of Wuppertal (2010)
196. Gerstewitz(Tammer), C.: Nichtkonvexe Dualität in der Vektoroptimierung. Wissenschaftliche
Zeitschrift der TH Leuna-Merseburg 25(3), 357–364 (1983)
197. Gerstewitz(Tammer), C., Iwanow, E.: Dualität für nichtkonvexe Vektoroptimierungsprob-
leme. Wiss. Z. Tech. Hochsch. Ilmenau 2, 61–81 (1985)
198. Gerth, C., Weidner, P.: Nonconvex separation theorems and some applications in vector
optimization. J. Optim. Theory Appl. 67(2), 297–320 (1990)
199. Giannessi, F.: Theorems of alternative, quadratic programs and complementarity problems.
In: Variational Inequalities and Complementarity Problems (Proc. Internat. School, Erice,
1978), pp. 151–186. Wiley, Chichester (1980)
200. Giannessi, F. (ed.): Vector variational inequalities and vector equilibria, mathematical theo-
ries. In: Nonconvex Optimization and Its Applications, vol. 38. Kluwer Academic, Dordrecht
(2000)
201. Giannessi, F., Khan, A.A.: Regularization of non-coercive quasi variational inequalities.
Control Cybern. 29(1), 91–110 (2000)
202. Giannessi, F., Khan, A.A.: On the envelope of a variational inequality. In: Nonlinear Analysis
and Variational Problems, Springer Optim. Appl., vol. 35, pp. 285–293. Springer, New York
(2010)
203. Giannessi, F., Mastroeni, G., Yang, X.Q.: A survey on vector variational inequalities. Boll.
Unione Mat. Ital. (9) 2(1), 225–237 (2009)
204. Giannessi, F., Mastroeni, G., Yang, X.Q.: Survey on vector complementarity problems. J.
Global Optim. 53(1), 53–67 (2012)
205. Ginchev, I., Guerraggio, A., Rocca, M.: Geoffrion type characterization of higher-order
properly efficient points in vector optimization. J. Math. Anal. Appl. 328(2), 780–788 (2007)
206. Ginchev, I., Guerraggio, A., Rocca, M.: Higher order properly efficient points in vector
optimization. In: Generalized Convexity and Related Topics, Lecture Notes in Econom. and
Math. Systems, vol. 583, pp. 227–245. Springer, Berlin (2007)
207. Giorgi, G., Guerraggio, A.: On a characterization of Clarke’s tangent cone. J. Optim. Theory
Appl. 74(2), 369–372 (1992)
208. Giorgi, G., Guerraggio, A.: On the notion of tangent cone in mathematical programming.
Optimization 25(1), 11–23 (1992)
209. Giorgi, G., Jiménez, B., Novo, V.: An overview of second order tangent sets and their
application to vector optimization. Bol. Soc. Esp. Mat. Apl. SeMA (52), 73–96 (2010)
210. Girejko, E., Malinowska, A.B., Torres, D.F.M.: The contingent epiderivative and the calculus
of variations on time scales. Optimization 61(3), 251–264 (2012)
211. Girsanov, I.V.: Lectures on mathematical theory of extremum problems. Springer, Berlin
(1972), edited by B. T. Poljak, Translated from the Russian by D. Louvish, Lecture Notes in
Economics and Mathematical Systems, Vol. 67
212. Goerigk, M., Schöbel, A.: A scenario-based approach for robust linear optimization. In:
Proceedings of the 1st International ICST Conference on Practice and Theory of Algorithms
in (Computer) Systems (TAPAS), Lecture Notes in Computer Science, pp. 139–150. Springer,
New York (2011)
213. Gong, X.H.: Optimality conditions for Henig and globally proper efficient solutions with
ordering cone has empty interior. J. Math. Anal. 307, 12–37 (2005)
214. Göpfert, A., Riahi, H., Tammer, C., Zălinescu, C.: Variational methods in partially ordered
spaces. CMS Books in Mathematics/Ouvrages de Mathématiques de la SMC, vol. 17.
Springer, New York (2003)
215. Göpfert, A., Tammer, C., Zălinescu, C.: On the vectorial Ekeland’s variational principle and
minimal points in product spaces. Nonlinear Anal. 39(7, Ser. A: Theory Methods), 909–922
(2000)
216. Gorokhovik, V.V., Zabreiko, P.P.: On Fréchet differentiability of multifunctions. Optimization
54(4-5), 391–409 (2005)
742 References
217. Gorokhovik, V.V, Gorokhovik, S.Ya.: A criterion of the global epi-Lipschitz property of sets
(Russian). Izvestiya Natsional’no Akademii Nauk Belarusi. Seriya Fiziko-Matematicheskikh
Nauk, pp. 118–120 (1995)
218. Götz, A., Jahn, J.: The Lagrange multiplier rule in set-valued optimization. SIAM J. Optim.
10(2), 331–344 (1999)
219. Guerraggio, A., Molno, E., Zaffaroni, A.: On the notion of proper efficiency in vector
optimization. J. Optim. Theory Appl. 82, 1–21 (1994)
220. Gutiérrez, C., Jiménez, B., Novo, V.: A set-valued Ekeland’s variational principle in vector
optimization. SIAM J. Control Optim. 47(2), 883–903 (2008)
221. Gwinner, J.: On a new class of differential variational inequalities and a stability result. Math.
Program. 139(1-2, Ser. B), 205–221 (2013)
222. Gwinner, J., Oettli, W.: Theorems of the alternative and duality for inf-sup problems. Math.
Oper. Res. 19(1), 238–256 (1994)
223. Ha, T.: Optimality conditions for several types of efficient solutions of set-valued optimization
problems. In: Nonlinear Analysis and variational Problems, Springer Optim. Appl., vol. 35,
pp. 305–324. Springer, New York (2010)
224. Ha, T.X.D.: A note on a class of cones ensuring the existence of efficient points in bounded
complete sets. Optimization 31(2), 141–152 (1994)
225. Ha, T.X.D.: On the existence of efficient points in locally convex spaces. J. Global Optim.
4(3), 265–278 (1994)
226. Ha, T.X.D.: Lagrange multipliers for set-valued optimization problems associated with
coderivatives. J. Math. Anal. Appl. 311(2), 647–663 (2005)
227. Ha, T.X.D.: Some variants of the Ekeland variational principle for a set-valued map. J. Optim.
Theory Appl. 124(1), 187–206 (2005)
228. Ha, T.X.D.: Optimality conditions for several types of efficient solutions of set-valued
optimization problems. In: Nonlinear Analysis and Variational Problems, Springer Optim.
Appl., vol. 35, pp. 305–324. Springer, New York (2010)
229. Ha, T.X.D.: The Fermat rule and Lagrange multiplier rule for various efficient solutions of set-
valued optimization problems expressed in terms of coderivatives. In: Recent Developments
in Vector Optimization, pp. 417–466. Springer, Berlin (2012)
230. Haimes, Y., Lasdon, L.S., Wismer, D.A.: On a bicriterion formulation of the problems of
integrated system identification and system optimization. IEEE Trans. Syst. Man Cybern. 1,
296–297 (1971)
231. Halkin, H.: A satisfactory treatment of equality and operator constraints in the Dubovitskii-
Milyutin optimization formalism. J. Optim. Theory Appl. 6, 138–149 (1970)
232. Halkin, H.: Necessary conditions for optimal control problems with differentiable or non-
differentiable data. In: Mathematical Control Theory (Proc. Conf., Australian Nat. Univ.,
Canberra, 1977), Lecture Notes in Math., vol. 680, pp. 77–118. Springer, Berlin (1978)
233. Hamel, A.: Remarks to an equivalent formulation of Ekeland’s variational principle. Opti-
mization 31(3), 233–238 (1994)
234. Hamel, A., Heyde, F., Löhne, A., Tammer, C., Winkler, K.: Closing the duality gap in linear
vector optimization. J. Convex Anal. 11(1), 163–178 (2004)
235. Hamel, A.H.: Equivalents to Ekeland’s variational principle in uniform spaces. Nonlinear
Anal. 62(5), 913–924 (2005)
236. Hamel, A.H., Löhne, A.: Lagrange duality in set optimization. Submitted for publication
(2012), http://arxiv.org/abs/1207.4433v1
237. Hamel, A.H., Tammer, C.: Minimal elements for product orders. Optimization 57(2), 263–
275 (2008)
238. Han, W.: A posteriori error analysis via duality theory, with applications in modeling and
numerical approximations. In: Advances in Mechanics and Mathematics, vol. 8. Springer,
New York (2005)
239. Hartley, R.: On cone-efficiency, cone-convexity and cone-compactness. SIAM J. Appl. Math.
34(2), 211–222 (1978)
References 743
240. Hazen, G.B., Morin, T.L.: Optimality conditions in nonconical multiple-objective program-
ming. J. Optim. Theory Appl. 40(1), 25–60 (1983)
241. Henig, M.I.: Existence and characterization of efficient decisions with respect to cones. Math.
Program. 23(1), 111–116 (1982)
242. Henig, M.I.: Proper efficiency with respect to cones. J. Optim. Theory Appl. 36(3), 387–407
(1982)
243. Henrion, R., Jourani, A., Outrata, J.: On the calmness of a class of multifunctions. SIAM J.
Optim. 13(2), 603–618 (electronic) (2002)
244. Hernández, E., Jiménez, B., Novo, V.: Benson proper efficiency in set-valued optimization
on real linear spaces. In: Recent Advances in Optimization, Lecture Notes in Econom. and
245. Hernández, E., Jiménez, B., Novo, V.: Weak and proper efficiency in set-valued optimization
on real linear spaces. J. Convex Anal. 14(2), 275–296 (2007)
246. Hernández, E., Khan, A.A., Rodríguez-Marín, L., Sama, M.: Computation formulas and
multiplier rules for graphical derivatives in separable banach spaces. Nonlinear Anal. 71,
4241–4250 (2009)
247. Hernández, E., Löhne, A., Rodríguez-Marín, L., Tammer, C.: Lagrange duality in vector
optimization - a simplified approach based on complete lattices. Preprint (2009)
248. Hernández, E., Löhne, A., Rodríguez-Marín, L., Tammer, C.: Lagrange duality, stability and
subdifferentials in vector optimization. Optimization 62(3), 415–428 (2013)
249. Hernández, E., Rodríguez-Marín, L.: Duality in set optimization with set-valued maps. Pac.
J. Optim. 3(2), 245–255 (2007)
250. Hernández, E., Rodríguez-Marín, L.: Existence theorems for set optimization problems.
Nonlinear Anal. 67(6), 1726–1736 (2007)
251. Hernández, E., Rodríguez-Marín, L.: Lagrangian duality in set-valued optimization. J. Optim.
Theory Appl. 134(1), 119–134 (2007)
252. Hernández, E., Rodríguez-Marín, L.: Nonconvex scalarization in set optimization with set-
valued maps. J. Math. Anal. Appl. 325(1), 1–18 (2007)
253. Hernández, E., Rodríguez-Marín, L., Sama, M.: On solutions of set-valued optimization
problems. Comput. Math. Appl. 60(5), 1401–1408 (2010)
254. Hestenes, M.R.: Calculus of Variations and Optimal Control Theory. Wiley, New York (1966)
255. Hestenes, M.R.: Optimization Theory. Wiley-Interscience, New York (1975)
256. Heyde, F.: Coherent risk measures and vector optimization. In: Küfer, K.H., Rommelfanger,
H., Tammer, C., Winkler, K. (eds.) Multicriteria Decision Making and Fuzzy Systems, pp. 3–
12. SHAKER, Aachen (2006)
257. Heyde, F., Löhne, A.: Solution concepts in vector optimization: a fresh look at an old story.
Optimization 60(12), 1421–1440 (2011)
258. Heyde, F., Löhne, A., Tammer, C.: Set-valued duality theory for multiple objective linear
programs and application to mathematical finance. Math. Methods Oper. Res. 69(1), 159–
179 (2009)
259. Heyde, F., Löhne, A., Tammer, C.: The attainment of the solution of the dual program in
vertices for vectorial linear programs. In: Vincent, B. et al. (eds.) Multiobjective Programming
and Goal Programming. Theoretical Results and Practical Applications. Lecture Notes in
Economics and Mathematical Systems, vol. 618, pp. 13–24. Springer, New York (2009)
260. Heyde, F., Schrage, C.: Continuity concepts for set-valued functions and a fundamental
duality formula for set-valued optimization. J. Math. Anal. Appl. 397(2), 772–784 (2013)
261. Hiriart-Urruty, J.B.: New concepts in nondifferentiable programming. Bull. Soc. Math.
France Mém 60, 57–85 (1979)
262. Hiriart-Urruty, J.B.: Tangent cones, generalized gradients and mathematical programming in
Banach spaces. Math. Oper. Res. 4(1), 79–97 (1979)
263. Hiriart-Urruty, J.B.: A note on the mean value theorem for convex functions. Boll. Un. Mat.
Ital. B (5) 17(2), 765–775 (1980)
264. Hites, R., De Smet, Y., Risse, N., Salazar-Neumann, M., Vincke, P.: About the applicability
of MCDA to some robustness problems. Eur. J. Oper. Res. 174, 322–332 (2006)
744 References
265. Hoffmann, K.H., Kornstaedt, H.J.: Higher-order necessary conditions in abstract mathemati-
cal programming. J. Optim. Theory Appl. 26(4), 533–568 (1978)
266. Hoheisel, T., Kanzow, C., Mordukhovich, B.S., Phan, H.: Generalized Newton’s method
based on graphical derivatives. Nonlinear Anal. 75(3), 1324–1340 (2012)
267. Hu, S., Papageorgiou, N.S.: Handbook of multivalued analysis. Vol. I, Mathematics and Its
Applications, vol. 419. Kluwer Academic, Dordrecht (1997)
268. Hu, S., Papageorgiou, N.S.: Handbook of multivalued analysis. Vol. II, Mathematics and Its
269. Huang, H.: The Lagrange multiplier rule for super efficiency in vector optimization. J. Math.
Anal. Appl. 342, 503–513 (2008)
270. Huang, H., Li, R.: Global error bounds for -paraconvex multifunctions. Set Valued Var.
Anal. 19(3), 487–504 (2011)
271. Hurwicz, L.: Programming in linear space. In: Studies in Linear and Non-linear Program-
ming. Stanford University Press, Stanford, California (1958)
272. Huy, N.Q., Mordukhovich, B., Yao, J.C.: Coderivatives of frontier and solution maps in
parameteric multiobjective optimization. Taiwanese J. Math. 13(6A), 1671–1693 (2009)
273. Ide, J., Köbis, E. Kuroiwa, D., Schöbel, A., Tammer, C.: The relationship between multicrite-
ria robustness concepts and set-valued optimization. Fixed Point Theory Appl. 83 (2014)
274. Ide, J., Köbis, E.: Concepts of robustness for multi-objective optimization problems based on
set order relations (2013)
275. Ioffe, A.D.: Calculus of Dini subdifferentials of functions and contingent coderivatives of
set-valued maps. Nonlinear Anal. 8(5), 517–539 (1984)
276. Ioffe, A.D.: Approximate subdifferentials and applications. II. Mathematika 33(1), 111–128
(1986)
277. Ioffe, A.D.: Approximate subdifferentials and applications. III. The metric theory. Mathe-
matika 36(1), 1–38 (1989)
278. Ioffe, A.D.: Variational analysis and mathematical economics 1: Subdifferential calculus and
the second theorem of welfare economics. In: Advances in Mathematical Economics, Vol. 12.
Adv. Math. Econ., vol. 12, pp. 71–95. Springer, Tokyo (2009)
279. Ioffe, A.D., Tihomirov, V.M.: Theory of extremal problems. Studies in Mathematics and Its
Applications, vol. 6. North-Holland, Amsterdam (1979)
280. Isac, G.: Sur l’existence de l’optimum de Pareto. Riv. Mat. Univ. Parma (4) 9, 303–325 (1984)
(1983)
281. Isac, G.: Pareto optimization in infinite-dimensional spaces: the importance of nuclear cones.
J. Math. Anal. Appl. 182(2), 393–404 (1994)
282. Isac, G.: The Ekeland’s principle and the Pareto "-efficiency. In: Multi-objective Pro-
gramming and Goal Programming: Theories and Applications. Proceedings of the First
International Conference, MOPGP (Portsmouth, UK, 1994), Lecture Notes in Econom. and
283. Isac, G., Khan, A.: Second-order optimality conditions in set-valued optimization by a new
tangential derivative. Acta Math. Vietnam. 34, 81–90 (2009)
284. Isac, G., Khan, A.A.: Dubovitskii-Milyutin approach in set-valued optimization. SIAM J.
Control Optim. 47(1), 144–162 (2008)
285. Isac, G., Tammer, C.: Nuclear and full nuclear cones in product spaces: Pareto efficiency and
an Ekeland type variational principle. Positivity 9(3), 511–539 (2005)
286. Isermann, H.: Proper efficiency and the linear vector maximization problem. Oper. Res. 22,
189–191 (1974)
287. Isermann, H.: On some relations between a dual pair of multiple objective linear programs.
Z. Oper. Res., Ser. A 22, 33–41 (1978)
288. Jachymski, J.R.: Caristi’s fixed point theorem and selections of set-valued contractions. J.
Math. Anal. Appl. 227(1), 55–67 (1998)
289. Jadamba, B., Khan, A.A., Raciti, F.: Regularization of stochastic variational inequalities and
a comparison of an Lp and a sample-path approach. Nonlinear Anal. 94, 65–83 (2014)
290. Jahn, J.: Duality in vector optimization. Math. Program. 25, 343–353 (1983)
References 745
291. Jahn, J.: Existence theorems in vector optimization. J. Optim. Theory Appl. 50(3), 397–406
(1986)
292. Jahn, J.: Mathematical Vector Optimization in Partially Ordered Linear Spaces. Verlag Peter
Lang, Frankfurt am Main (1986)
293. Jahn, J.: Vector Optimization. Theory, Applications and Extensions. Springer, Berlin (2004)
294. Jahn, J.: A derivative-free descent method in set optimization. Preprint-Reihe Angewandte
Mathematik, Friedrich-Alexander-Universität Erlangen-Nürnberg 377 (2014)
295. Jahn, J., Ha, T.X.D.: New order relations in set optimization. J. Optim. Theory Appl. 148(2),
209–236 (2011)
296. Jahn, J., Khan, A.A.: Existence theorems and characterizations of generalized contingent
epiderivatives. J. Nonlinear Convex Anal. 3(3), 315–330 (2002)
297. Jahn, J., Khan, A.A.: Generalized contingent epiderivatives in set-valued optimization:
optimality conditions. Numer. Funct. Anal. Optim. 23(7-8), 807–831 (2002)
298. Jahn, J., Khan, A.A.: The existence of contingent epiderivatives for set-valued maps. Appl.
Math. Lett. 16(8), 1179–1185 (2003)
299. Jahn, J., Khan, A.A.: Some calculus rules for contingent epiderivatives. Optimization 52(2),
113–125 (2003)
300. Jahn, J., Khan, A.A., Zeilinger, P.: Second-order optimality conditions in set optimization. J.
301. Jahn, J., Rauh, R.: Contingent epiderivatives and set-valued optimization. Math. Methods
Oper. Res. 46(2), 193–211 (1997)
302. Jahn, J., Sachs, E.: Generalized quasiconvex mappings and vector optimization. SIAM J.
Control Optim. 24(2), 306–322 (1986)
303. Jameson, G.: Ordered linear spaces. Lecture Notes in Mathematics, Vol. 141. Springer, Berlin
(1970)
304. Jiménez, B., Novo, V.: Second order necessary conditions in set constrained differentiable
vector optimization. Math. Methods Oper. Res. 58(2), 299–317 (2003)
305. Jiménez, B., Novo, V.: Optimality conditions in differentiable vector optimization via second-
order tangent sets. Appl. Math. Optim. 49(2), 123–144 (2004)
306. Jiménez, B., Novo, V.: Characterization of the cone of attainable directions. J. Optim. Theory
Appl. 131(3), 493–499 (2006)
307. Jofré, A., Penot, J.P.: Comparing new notions of tangent cones. J. Lond. Math. Soc. (2) 40(2),
280–290 (1989)
308. Jourani, A.: Regularity and strong sufficient optimality conditions in differentiable optimiza-
tion problems. Numer. Funct. Anal. Optim. 14(1-2), 69–87 (1993)
309. Jourani, A.: Metric regularity and second-order necessary optimality conditions for minimiza-
tion problems under inclusion constraints. J. Optim. Theory Appl. 81(1), 97–120 (1994)
310. Jourani, A.: Open mapping theorem and inversion theorem for -paraconvex multivalued
mappings and applications. Studia Math. 117(2), 123–136 (1996)
311. Jourani, A.: Necessary conditions for extremality and separation theorems with applications
to multiobjective optimization. Optimization 44(4), 327–350 (1998)
312. Jourani, A., Thibault, L.: Chain rules for coderivatives of multivalued mappings in Banach
spaces. Proc. Am. Math. Soc. 126, 1479–1485 (1998)
313. Kalashnikov, V.V., Jadamba, B., Khan, A.A.: First and second order optimality condition in
set-optimization. In: Dempe, S., et al. (eds.), Optimization with Multivalued Mappings, pp.
265–276. Springer, New York (2006)
314. Kasimbeyli, R.: Radial epiderivatives and set-valued optimization. Optimization 58, 521–534
(2009)
315. Kawasaki, H.: Conjugate relations and weak subdifferentials of relations. Math. Oper. Res. 6,
593–607 (1981)
316. Kawasaki, H.: A duality theorem in multiobjective nonlinear programming. Math. Oper. Res.
7, 95–110 (1982)
317. Kawasaki, H.: An envelope-like effect of infinitely many inequality constraints on second-
order necessary conditions for minimization problems. Math. Program. 41(1, (Ser. A)), 73–96
(1988)
746 References
318. Khan, A.A., Raciti, F.: A multiplier rule in set-valued optimization. Bull. Aust. Math. Soc.
68, 93–100 (2003)
319. Khan, A.A., Sama, M.: A multiplier rule for stable problems in vector optimization. J. Convex
Anal. 19(2), 525–539 (2012)
320. Khan, A.A., Sama, M.: Optimal control of multivalued quasi variational inequalities. Nonlin-
ear Anal. 75(3), 1419–1428 (2012)
321. Khan, A.A., Tammer, C.: Generalized Dubovitskii-Milyutin approach in set-valued optimiza-
tion. Vietnam J. Math. 40(2-3), 285–304 (2012)
322. Khan, A.A., Tammer, C.: Second-order optimality conditions in set-valued optimization via
asymptotic derivatives. Optimization 62(6), 743–758 (2013)
323. Khan, A.A., Ward, D.E.: Toward second-order sensitivity analysis in set-valued optimization.
J. Nonlinear Convex Anal. 13(1), 65–83 (2012)
324. Khan, M.A.: Ioffe’s normal cone and the foundations of welfare economics: the infinite-
dimensional theory. J. Math. Anal. Appl. 161(1), 284–298 (1991)
325. Khanh, P., Tuan, N.: Higher-order variational sets and higher-order optimality conditions for
proper efficiency in set-valued nonsmooth vector optimization. J. Optim. Theory Appl. 139,
243–261 (2008)
326. Khanh, P., Tuan, N.: Variational sets of multi-valued mappings and a unified study of
optimality condition. J. Optim. Theory Appl. 139, 45–67 (2008)
327. Kinderlehrer, D., Stampacchia, G.: An introduction to variational inequalities and their
applications. In: Classics in Applied Mathematics, vol. 31. Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, PA (2000). Reprint of the 1980 original
328. Klamroth, K., Köbis, E., Schöbel, A., Tammer, C.: A unified approach for different concepts
of robustness and stochastic programming via non-linear scalarizing functionals. Optimiza-
tion 62(5), 649–671 (2013)
329. Klatte, D., Kummer, B.: Nonsmooth equations in optimization. Nonconvex Optimization and
Its Applications, vol. 60. Kluwer Academic, Dordrecht (2002)
330. Klose, J.: Sensitivity analysis using the tangent derivative. Numer. Funct. Anal. Optim. 13(1-
2), 143–153 (1992)
331. Konnov, I.: Combined relaxation methods for variational inequalities. Lecture Notes in
Economics and Mathematical Systems, vol. 495. Springer, Berlin (2001)
332. Kotarski, W.: Further generalization of the Dubovicki-Milutin theorem. J. Optim. Theory
Appl. 54(3), 565–573 (1987)
333. Kotarski, W.: Optimal control of a system governed by a parabolic equation with an infinite
number of variables. J. Optim. Theory Appl. 60(1), 33–41 (1989)
334. Kotarski, W.: On some specification of the Dubovicki-Milutin theorem for Pareto optimal
problems. Nonlinear Anal. 14(3), 287–291 (1990)
335. Kouvelis, P., Sayin, S.: Algorithm robust for the bicriteria discrete optimization problem. Ann.
Oper. Res. 147, 71–85 (2006)
336. Kouvelis, P., Yu, G.: Robust Discrete Optimization and Its Applications. Kluwer Academic,
Dordrecht (1997)
337. Kowalewski, A., Kotarski, W.: Application of Milutin-Dubovicki’s method to solving an
optimal control problem for hyperbolic systems. Probl. Control Inform. Theory/Problemy
Upravlen. Teor. Inform. 9(3), 183–193 (1980)
338. Krabs, W.: Optimization and Approximation. Wiley, Chichester (1979)
339. Krasnosel0 skiı̆, M.A.: Polozhitelnye resheniya operatornykh uravnenii. Gosudarstv. Izdat.
Fiz.-Mat. Lit., Moscow (1962)
340. Krasnosel0 skiı̆, M.A.: Positive solutions of operator equations. Translated from the Russian
by Richard E. Flaherty; edited by Leo F. Boron. P. Noordhoff Ltd. Groningen (1964)
341. Křivan, V.: On the intersection of contingent cones. J. Optim. Theory Appl. 70(2), 397–404
(1991)
342. Kruger, A.J., Morduhovič, B.Š.: Extremal points and the Euler equation in nonsmooth
optimization problems. Dokl. Akad. Nauk BSSR 24(8), 684–687, 763 (1980)
References 747
343. Kruger, A.Y.: On the extremality of set systems. Dokl. Nats. Akad. Nauk Belarusi 42(1),
24–28, 123 (1998)
344. Kuk, H., Tanino, T., Tanaka, M.: Sensitivity analysis in parametrized convex vector optimiza-
tion. J. Math. Anal. Appl. 202(2), 511–522 (1996)
345. Kuk, H., Tanino, T., Tanaka, M.: Sensitivity analysis in vector optimization. J. Optim. Theory
Appl. 89(3), 713–730 (1996)
346. Kurcyusz, S.: On the existence and non-existence Lagrange multipliers in Banach spaces. J.
347. Kuroiwa, D.: Existence Theorems of Set Optimization with Set-Valued Maps. Manuscript
Shimane University, Japan (1997)
348. Kuroiwa, D.: Lagrange duality of set-valued optimization with natural criteria.
Sūrikaisekikenkyūsho Kōkyūroku (1068), 164–170 (1998)
349. Kuroiwa, D.: Natural Criteria of Set-Valued Optimization. Manuscript Shimane University,
Japan (1068), 164–170 (1998)
350. Kuroiwa, D.: On natural criteria in set-valued optimization. RIMS Kokyuroku 1048, 86–92
(1998)
351. Kuroiwa, D.: Some duality theorems of set-valued optimization with natural criteria. In:
Tanaka, T. (ed.) Proceedings of the International Conference on Nonlinear Analysis and
Convex Analysis, pp. 221–228. World Scientific, Singapore (1999)
352. Kuroiwa, D.: Existence theorems of set optimization with set-valued maps. J. Inf. Optim. Sci.
24(1), 73–84 (2003)
353. Kuroiwa, D.: On derivatives of Set-Valued Maps in Set Optimization. Manuscript Shimane
University, Japan (2008)
354. Kuroiwa, D.: On derivatives of set-valued maps and optimality conditions for set optimization.
J. Nonlinear Convex Anal. 10(1), 41–50 (2009)
355. Kuroiwa, D., Lee, G.M.: On robust multiobjective optimization. Vietnam J. Math. 40(2-3),
305–317 (2012)
356. Kuroiwa, D., Lee, G.M.: On robust convex multiobjective optimization. Accepted to J.
Nonlinear Convex Anal. (2013)
357. Kuroiwa, D., Nuriya, T.: A generalized embedding vector space in set optimization. In:
Proceedings of the Fourth Intemational Conference on Nonlinear Analysis and Convex
Analysis, pp. 297–304 (2006)
358. Kusraev, A.G., Kutateladze, S.S.: Subdifferentials: theory and applications. In: Mathematics
and Its Applications, vol. 323. Kluwer Academic, Dordrecht (1995). Translated from the
Russian
359. Laghdir, M.: Some remarks on subdifferentiability of convex functions. Appl. Math. E-Notes
5, 150–156 (electronic) (2005)
360. Lalitha, C.S., Arora, R.: Conjugate maps, subgradients and conjugate duality in set-valued
optimization. Numer. Funct. Anal. Optim. 28(7-8), 897–909 (2007)
361. Lalitha, C.S., Arora, R.: Weak Clarke epiderivative in set-valued optimization. J. Math. Anal.
Appl. 342(1), 704–714 (2008)
362. Lasiecka, I.: Conical approximations of sets in optimization theory. Control Cybern. 4(3-4),
39–58 (1975)
363. Laurent, P.J.: Approximation et Optimisation, Collection Enseignement des Sciences, No. 13.
Hermann, Paris (1972)
364. Ledzewicz, U., Schaettler, H.: Second-order conditions for extremum problems with nonreg-
ular equality constraints. J. Optim. Theory Appl. 86(1), 113–144 (1995)
365. Ledzewicz, U., Schättler, H.: A second-order Dubovitskii-Milyutin theory and applications
to control. In: Advances in Nonlinear Dynamics. Stability Control Theory Methods Appl.,
vol. 5, pp. 179–192. Gordon and Breach, Amsterdam (1997)
366. Ledzewicz, U., Schättler, H.: High-order extended maximum principles for optimal control
problems with non-regular constraints. In: Optimal control (Gainesville, FL, 1997), Appl.
Optim., vol. 15, pp. 298–325. Kluwer Academic, Dordrecht (1998)
748 References
367. Ledzewicz, U., Schättler, H.: A high-order generalization of the Lyusternik theorem. Nonlin-
ear Anal. 34(6), 793–815 (1998)
368. Ledzewicz, U., Schättler, H.: High-order approximations and generalized necessary condi-
tions for optimality. SIAM J. Control Optim. 37(1), 33–53 (1999)
369. Ledzewicz, U., Schättler, H.: A high-order generalized local maximum principle. SIAM J.
Control Optim. 38(3), 823–854 (2000)
370. Ledzewicz, U., Walczak, S.: On the Lusternik theorem for nonsmooth operators. Nonlinear
Anal. 22(2), 121–128 (1994)
371. Ledzewicz, U., Walczak, S.: Generalizations of the Lusternik theorem and applications to
nonsmooth and abnormal problems in optimization and optimal control. In: World Congress
of Nonlinear Analysts ’92, Vol. I–IV (Tampa, FL, 1992), pp. 2583–2594. de Gruyter, Berlin
(1996)
372. Ledzewicz-Kowalewska, U.: On some specification of the Dubovitskiı̆-Milyutin method.
Nonlinear Anal. 10(12), 1367–1371 (1986)
373. Ledzewicz-Kowalewska, U.: Application of the method of contractor directions to the
Dubovitskiı̆-Milyutin formalism. J. Math. Anal. Appl. 125(1), 174–184 (1987)
374. Ledzewicz-Kowalewska, U.: Application of some specification of the Dubovitskiı̆-Milyutin
method to problems of optimal control. Nonlinear Anal. 12(2), 101–108 (1988)
375. Leung, S.F.: An economic application of the Dubovitskii-Milyutin version of the maximum
principle. Optim. Control Appl. Methods 28(6), 435–449 (2007)
376. Levy, A.B., Mordukhovich, B.S.: Coderivatives in parametric optimization. Math. Program.
99(2, Ser. A), 311–327 (2004)
377. Li, J., Mastroeni, G.: Vector variational inequalities involving set-valued mappings via
scalarization with applications to error bounds for gap functions. J. Optim. Theory Appl.
145(2), 355–372 (2010)
378. Li, M.H., Li, S.J.: Second-order differential and sensitivity properties of weak vector
variational inequalities. J. Optim. Theory Appl. 144(1), 76–87 (2010)
379. Li, S.J., Chen, C.R., Wu, S.Y.: Conjugate dual problems in constrained set-valued optimiza-
tion and applications. Eur. J. Oper. Res. 196(1), 21–32 (2009)
380. Li, S.J., Meng, K.W., Penot, J.P.: Calculus rules for derivatives of multimaps. Set Valued Var.
Anal. 17(1), 21–39 (2009)
381. Li, S.J., Sun, X.K., Zhai, J.: Second-order contingent derivatives of set-valued mappings with
application to set-valued optimization. Appl. Math. Comput. 218(12), 6874–6886 (2012)
382. Li, S.J., Sun, X.K., Zhu, S.K.: Higher-order optimality conditions for strict minimality in
set-valued optimization. J. Nonlinear Convex Anal. 13(2), 281–291 (2012)
383. Li, S.J., Yan, H., Chen, G.Y.: Differential and sensitivity properties of gap functions for vector
variational inequalities. Math. Methods Oper. Res. 57(3), 377–391 (2003)
384. Li, S.J., Yang, X.Q., Chen, G.y.: Vector Ekeland variational principle. In: Vector Variational
Inequalities and Vector Equilibria. Nonconvex Optim. Appl., vol. 38, pp. 321–333. Kluwer
Academic, Dordrecht (2000)
385. Li, S.J., Zhu, S.K., Li, X.B.: Second-order optimality conditions for strict efficiency of
constrained set-valued optimization. J. Optim. Theory Appl. 155(2), 534–557 (2012)
386. Li, T., Xu, Y.: The strictly efficient subgradient of set-valued optimisation. Bull. Aust. Math.
Soc. 75(3), 361–371 (2007)
387. Li, Z.F., Chen, G.Y.: Lagrangian multipliers, saddle points, and duality in vector optimization
of set-valued maps. J. Math. Anal. Appl. 215(2), 297–316 (1997)
388. Liebchen, C., Lübbecke, M., Möhring, R.H., Stiller, S.: The concept of recoverable robust-
ness, linear programming recovery, and railway applications. In: Ahuja, R.K.m Möhring, R.,
Zaroliagis, C. (eds.) Robust and online large-scale optimization. Lecture Note on Computer
Science, vol. 5868. Springer, New York (2009)
389. Lin, L.J.: Optimization of set-valued functions. J. Math. Anal. Appl. 186(1), 30–51 (1994)
390. Lin, L.J., Chuang, C.S.: Existence theorems for variational inclusion problems and the set-
valued vector Ekeland variational principle in a complete metric space. Nonlinear Anal. 70(7),
2665–2672 (2009)
References 749
391. Liu, C.G., Ng, K.F.: Ekeland’s variational principle for set-valued functions. SIAM J. Optim.
21(1), 41–56 (2011)
392. Liu, J., Song, W.: On proper efficiencies in locally convex spaces—a survey. Acta Math.
Vietnam. 26(3), 301–312 (2001)
393. Löhne, A.: Optimization with set relations. Dissertation, Martin-Luther-University Halle-
Wittenberg (2005)
394. Löhne, A.: Optimization with set relations: Conjugate duality. Optimization 54(3), 265–282
(2005)
395. Löhne, A.: Vector Optimization with Infimum and Supremum. Springer, Berlin (2011)
396. Löhne, A., Schrage, C.: An algorithm to solve polyhedral convex set-valued optimization
problems. Optimization 62(1), 131–141 (2013)
397. Löhne, A., Tammer, C.: A new approach to duality in vector optimization. Optimization
56(1-2), 221–239 (2007)
398. Löhne, A., Tammer, C.: Lagrange duality in vector optimization. In: Löhne, A., Riedrich, T.,
Tammer, C. (eds.) Festschrift zum 75. Geburtstag von Prof. Dr. Alfred Göpfert, pp. 17–33.
Preprint Series Institut of Mathematics, Martin-Luther-University Halle-Wittenberg No. 07
(2009)
399. Long, X.J., Peng, J.W., Wong, M.M.: Generalized radial epiderivatives and nonconvex set-
valued optimization problems. Appl. Anal. 91(10), 1891–1900 (2012)
400. Lu.c, D.T.: Contingent derivatives of set-valued maps and applications to vector optimization.
Math. Program. 50(1, (Ser. A)), 99–111 (1991)
401. Luc, D.T.: An existence theorem in vector optimization. Math. Oper. Res. 14(4), 693–699
(1989)
402. Luc, D.T.: Theory of vector optimization. Lecture Notes in Economics and Mathematical
Systems, vol. 319. Springer, Berlin (1989)
403. Luc, D.T.: Theory of vector optimization. Lecture Notes in Economics and Mathematical
Systems, vol. 319. Springer, Berlin (1989)
404. Lu.c, D.T.: Recession cones and the domination property in vector optimization. Math.
Program. 49(1, (Ser. A)), 113–122 (1990/91)
405. Luc, D.T., Malivert, C.: Invex optimization problems. Bull. Aust. Math. Soc. 46, 47–66 (1992)
406. Luc, D.T., Swaminathan, S.: A characterization of convex functions. Nonlinear Anal. 20(6),
697–701 (1993)
407. Luc, D.T., Tan, N.X., Tinh, P.N.: Convex vector functions and their subdifferential. Acta
Math. Vietnam. 23(1), 107–127 (1998)
408. Luenberger, D.G.: New optimality principles for economic efficiency and equilibrium. J.
409. Makarov, E.K., Rachkovski, N.N.: Unified representation of proper efficiency by means of
dilating cones. J. Optim. Theory Appl. 101(1), 141–165 (1999)
410. Malcolm, G.G., Mordukhovich, B.S.: Pareto optimality in nonconvex economies with infinite-
dimensional commodity spaces. J. Global Optim. 20(3-4), 323–346 (2001)
411. Malivert, C.: Fenchel duality in vector optimization. In: Advances in Optimization (Lam-
brecht, 1991). Lecture Notes in Econom. and Math. Systems, vol. 382, pp. 420–438. Springer,
Berlin (1992)
412. Malivert, C., Penot, J.P., Théra, M.: Un prolongement du théorème de Hahn-Banach. C. R.
Acad. Sci. Paris Sér. A-B 286(3), A165–A168 (1978)
413. Martin, D.H., Gardner, R.J., Watkins, G.G.: Indicating cones and the intersection principle for
tangential approximants in abstract multiplier rules. J. Optim. Theory Appl. 33(4), 515–537
(1981)
414. Martin, D.H., Watkins, G.G.: Cores of tangent cones and Clarke’s tangent cone. Math. Oper.
Res. 10(4), 565–575 (1985)
415. Mas-Colell, A.: Pareto optima and equilibria: the finite-dimensional case. In: Advances in
Equilibrium Theory (Indianapolis, Ind., 1984). Lecture Notes in Econom. and Math. Systems,
vol. 244, pp. 25–42. Springer, Berlin (1985)
750 References
416. Mastroeni, G.: On Minty vector variational inequality. In: Vector Variational Inequalities
and Vector Equilibria. Nonconvex Optim. Appl., vol. 38, pp. 351–361. Kluwer Academic,
Dordrecht (2000)
417. Megginson, R.E.: An introduction to Banach space theory. Graduate Texts in Mathematics,
vol. 183. Springer, New York (1998)
418. Meng, K.W., Li, S.J.: Differential and sensitivity properties of gap functions for Minty vector
variational inequalities. J. Math. Anal. Appl. 337(1), 386–398 (2008)
419. Meng, Z.Q.: Hahn-Banach theorem of set-valued map. Appl. Math. Mech. 19(1), 55–61
(1998)
420. Michel, P., Penot, J.P.: Calcul sous-différentiel pour des fonctions lipschitziennes et non
lipschitziennes. C. R. Acad. Sci. Paris Sér. I Math. 298(12), 269–272 (1984)
421. Minchenko, L., Volosevich, A.: Strongly differentiable multifunctions and directional differ-
entiability of marginal functions. In: Quasidifferentiability and Related Topics. Nonconvex
Optim. Appl., vol. 43, pp. 163–171. Kluwer Academic, Dordrecht (2000)
422. Minh, N.B., Tan, N.X.: On the C -Lipschitz continuities and C -approximations of multivalued
mappings. Vietnam J. Math. 30(4), 343–363 (2002)
423. Minh, N.B., Tan, N.X.: On the continuity of vector convex multivalued functions. Acta Math.
Vietnam. 27(1), 13–25 (2002)
424. Mishra, S.K., Noor, M.A.: On vector variational-like inequality problems. J. Math. Anal.
Appl. 311(1), 69–75 (2005)
425. Mordukhovich, B.S.: Maximum principle in the problem of time optimal response with
nonsmooth constraints. Prikl. Mat. Meh. 40(6), 1014–1023 (1976)
426. Mordukhovich, B.S.: Metric approximations and necessary conditions for optimality for
general classes of nonsmooth extremal problems, (russian). Dokl. Akad. Nauk SSSR 254,
1072–1076 (1980)
427. Mordukhovich, B.S.: Generalized differential calculus for nonsmooth and set-valued map-
pings. J. Math. Anal. Appl. 183(1), 250–288 (1994)
428. Mordukhovich, B.S.: An abstract extremal principle with applications to welfare economics.
J. Math. Anal. Appl. 251(1), 187–216 (2000)
429. Mordukhovich, B.S.: Nonlinear prices in nonconvex economies with classical Pareto and
strong Pareto optimal allocations. Positivity 9(3), 541–568 (2005)
430. Mordukhovich, B.S.: Variational Analysis and Generalized Differentiation, Vol. I: Basic
Theory, Vol. II: Applications. Springer, Berlin (2006)
431. Mordukhovich, B.S.: Variational analysis and generalized differentiation, Vol. II: Applica-
tions, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathe-
matical Sciences], vol. 331. Springer, Berlin (2006)
432. Mordukhovich, B.S., Shao, Y.H.: Nonsmooth sequential analysis in Asplund spaces. Trans.
Am. Math. Soc. 348(4), 1235–1280 (1996)
433. Mordukhovich, B.S., Treiman, J.S., Zhu, Q.J.: An extended extremal principle with applica-
tions to multiobjective optimization. SIAM J. Optim. 14(2), 359–379 (electronic) (2003)
434. Mordukhovich, B.S., Wang, B.: Necessary optimality and suboptimality conditions in
nondifferentiable programming via variational principles. SIAM J. Control Optim. 41, 623–
640 (2002)
435. Motreanu, D., Pavel, N.: Tangency, flow invariance for differential equations, and optimiza-
tion problems. Monographs and Textbooks in Pure and Applied Mathematics, vol. 219.
Marcel Dekker, New York (1999)
436. Motreanu, D., Pavel, N.H.: Quasitangent vectors in flow-invariance and optimization prob-
lems on Banach manifolds. J. Math. Anal. Appl. 88(1), 116–132 (1982)
437. Nakayama, H.: Duality in multi-objective optimization. In: Gal, T., Stewart, T., Hanne, T.
(eds.) Multicriteria Decision Making, pp. 3–29. Kluwer Academic, Boston/Dordrecht/London
(1999)
438. Nam, N.M.: Coderivatives of normal cone mappings and Lipschitzian stability of parametric
variational inequalities. Nonlinear Anal. 73(7), 2271–2282 (2010)
References 751
439. Nehse, R.: The Hahn-Banach property and equivalent conditions. Comment. Math. Univ.
Carolin. 19(1), 165–177 (1978)
440. Németh, A.B.: A nonconvex vector minimization problem. Nonlinear Anal. 10(7), 669–678
(1986)
441. Ng, K.F., Zheng, X.Y.: Existence of efficient points in vector optimization and generalized
Bishop-Phelps theorem. J. Optim. Theory Appl. 115(1), 29–47 (2002). Dedicated to Professor
Wolfram Stadler
442. Nieuwenhuis, J.W.: Supremal points and generalized duality. Math. Operationsforsch. Statist.
Ser. Optim. 11(1), 41–59 (1980)
443. Nishnianidze, Z.G.: Fixed points of monotone multivalued operators. Soobshch. Akad. Nauk
Gruzin. SSR 114(3), 489–491 (1984)
444. Nurminskiı̆, A.E.: On the differentiability of set-valued mappings. Kibernetika (Kiev) (5),
46–48 (1978)
445. Oettli, W.: Optimality conditions for programming problems involving multivalued mappings.
In: Modern Applied Mathematics (Bonn, 1979), pp. 195–226. North-Holland, Amsterdam
(1982)
446. Ortega, J.M., Rheinboldt, W.C.: Iterative Solution of Nonlinear Equations in Several Vari-
ables. Academic Press, New York (1970)
447. Ovcharova, N., Gwinner, J.: On Moreau-Yosida approximation and on stability of second-
order subdifferentials of Clarke’s type. J. Convex Anal. 14(2), 413–432 (2007)
448. Ovcharova, N., Gwinner, J.: Stability in generalized differentiability based on a set conver-
gence principle. Set Valued Anal. 15(4), 317–329 (2007)
449. Palata, J.: A survey of conical approximations used in the optimization. Optimization 20,
147–161 (1989)
450. Páles, Z., Zeidan, V.: Nonsmooth optimum problems with constraints. SIAM J. Control
Optim. 32(5), 1476–1502 (1994)
451. Páles, Z., Zeidan, V.: Critical and critical tangent cones in optimization problems. Set Valued
Anal. 12(1-2), 241–258 (2004)
452. Páles, Z., Zeidan, V.: First- and second-order optimality conditions for a strong local
minimum in control problems with pure state constraints. Nonlinear Anal. 67(8), 2506–2526
(2007)
453. Pallaschke, D., Rolewicz, S.: Foundations of mathematical optimization. Mathematics and Its
454. Pascoletti, A., Serafini, P.: Scalarizing vector optimization problems. J. Opt. Theory Appl.
42, 499–524 (1984)
455. Pavel, N.: Second order differential equations on closed subsets of a Banach space. Boll. Un.
Mat. Ital. (4) 12(3), 348–353 (1975)
456. Pavel, N.H., Huang, Y.K., Kim, J.K.: Higher order necessary conditions for optimization.
Libertas Math. 14, 41–50 (1994)
457. Pavel, N.H., Ursescu, C.: Flow-invariant sets for autonomous second order differential
equations and applications in mechanics. Nonlinear Anal. 6(1), 35–74 (1982)
458. Peng, J., Lee, H.W.J., Rong, W., Yang, X.: A generalization of Hahn-Banach extension
theorem. J. Math. Anal. Appl. 302(2), 441–449 (2005)
459. Pennanen, T.: Graph-convex mappings and K-convex functions. J. Convex Anal. 6(2), 235–
266 (1999)
460. Penot, J.P.: L’optimisation à la Pareto: deux ou trois choses que je sais d’elle. In: Structures
Economiques et Econométrie (Lyon, 1978), pp. 4.14–4.33. Lyon (1978)
461. Penot, J.P.: A characterization of tangential regularity. Nonlinear Anal. 5(6), 625–643 (1981)
462. Penot, J.P.: Differentiability of relations and differential stability of perturbed optimization
problems. SIAM J. Control Optim. 22(4), 529–551 (1984)
752 References
463. Penot, J.P.: Variations on the theme of nonsmooth analysis: another subdifferential. In:
Nondifferentiable Optimization: Motivations and Applications (Sopron, 1984). Lecture Notes
in Econom. and Math. Systems, vol. 255, pp. 41–54. Springer, Berlin (1985)
464. Penot, J.P.: Erratum: “Differentiability of relations and differential stability of perturbed
optimization problems” [SIAM J. Control Optim. 22(4), 529–551 (1984); MR0747968
(85i:49041)]. SIAM J. Control Optim. 26(4), 997–998 (1988)
465. Penot, J.P.: Second-order generalized derivatives: comparisons of two types of epi-derivatives.
In: Advances in Optimization (Lambrecht, 1991), Lecture Notes in Econom. and Math.
Systems, vol. 382, pp. 52–76. Springer, Berlin (1992)
466. Penot, J.P.: Second-order generalized derivatives: comparisons of two types of epi-derivatives.
In: Advances in Optimization (Lambrecht, 1991), Lecture Notes in Econom. and Math.
Systems, vol. 382, pp. 52–76. Springer, Berlin (1992)
467. Penot, J.P.: Optimality conditions in mathematical programming and composite optimization.
Math. Program. 67(2, Ser. A), 225–245 (1994)
468. Penot, J.P.: Second-order conditions for optimization problems with constraints. SIAM
J. Control Optim. 37(1), 303–318 (electronic) (1999)
469. Penot, J.P.: Recent advances on second-order optimality conditions. In: Optimization (Namur,
1998), Lecture Notes in Econom. and Math. Systems, vol. 481, pp. 357–380. Springer, Berlin
(2000)
470. Penot, J.P.: Calculus without derivatives. Graduate Texts in Mathematics, vol. 266. Springer,
New York (2013)
471. Penot, J.P., Sach, P.H.: Generalized monotonicity of subdifferentials and generalized convex-
ity. J. Optim. Theory Appl. 94(1), 251–262 (1997)
472. Penot, J.P., Sterna-Karwat, A.: Parametrized multicriteria optimization: continuity and closed-
ness of optimal multifunctions. J. Math. Anal. Appl. 120(1), 150–168 (1986)
473. Penot, J.P., Théra, M.: Semicontinuous mappings in general topology. Arch. Math. (Basel)
38(2), 158–166 (1982)
474. Penot, J.P., Zălinescu, C.: Harmonic sum and duality. J. Convex Anal. 7(1), 95–113 (2000)
475. Peressini, A.L.: Ordered Topological Vector Spaces. Harper & Row Publishers, New York
(1967)
476. Perny, P., Spanjaard, O., Storme, L.X.: A decision-theoretic approach to robust optimization.
Ann. Oper. Res. 147, 317–341 (2006)
477. Plotnikov, A.V.: Differentiation of multivalued mappings. The T -derivative. Ukraïn. Mat. Zh.
52(8), 1119–1126 (2000)
478. Popovici, N., Rocca, M.: Scalarization and decomposition of vector variational inequalities
governed by bifunctions. Optimization 62(6), 735–742 (2013)
479. Postolică, V.: Existence conditions of efficient points for multifunctions with values in locally
convex spaces. Stud. Cerc. Mat. 41(4), 325–331 (1989)
480. Postolică, V.: New existence results for efficient points in locally convex spaces ordered by
supernormal cones. J. Global Optim. 3(2), 233–242 (1993)
481. Pühl, H., Schirotzek, W.: Linear semi-openness and the Lyusternik theorem. Eur. J. Oper.
Res. 157(1), 16–27 (2004)
482. Qiu, J.H.: Dual characterization and scalarization for Benson proper efficiency. SIAM J.
Optim. 19, 144–162 (2008)
483. Qiu, J.H.: A generalized Ekeland vector variational principle and its applications in optimiza-
tion. Nonlinear Anal. 71(10), 4705–4717 (2009)
484. Qiu, J.H.: Set-valued quasi-metrics and a general Ekeland’s variational principle in vector
optimization. SIAM J. Control Optim. 51(2), 1350–1371 (2013)
485. Raffin, C.: Sur les programmes convexes définis dans des espaces vectoriels topologiques.
Ann. Inst. Fourier (Grenoble) 20(fasc. 1), 457–491 (1970)
486. Ricceri, B.: Remarks: “On multifunctions with convex graph” [Atti Accad. Naz. Lincei Rend.
Cl. Sci. Fis. Mat. Natur. (8) 77(3-4), 64–70 (1984, 1985); MR0884937 (88d:54021)]. Arch.
Math. (Basel) 52(5), 519–520 (1989)
References 753
487. Rigby, L.: Contribution to Dubovitskiy and Milyutin’s optimization formalism. In: Optimiza-
tion Techniques Modeling and Optimization in the Service of Man. Part 2. Lecture Notes in
Computer Science, vol. 41. Proceedings of the 7th IFIP Conference Nice, 8–12 September
1975, pp. 438–453. Springer, Heidelberg (1976)
488. Rocca, M.: Well-posed vector optimization problems and vector variational inequalities. J.
Inf. Optim. Sci. 27(2), 259–270 (2006)
489. Rockafellar, R.T.: Convex Analysis. Princeton Mathematical Series, No. 28. Princeton
University Press, Princeton (1970)
490. Rockafellar, R.T.: Conjugate duality and optimization. Society for Industrial and Applied
Mathematics. VI (1974), cBMS-NSF Regional Conference Series in Applied Mathematics,
vol. 16. SIAM, Philadelphia
491. Rockafellar, R.T.: Clarke’s tangent cones and the boundaries of closed sets in Rn . Nonlinear
Anal. Theory Methods Appl. 3, 145–154 (1979)
492. Rockafellar, R.T.: Generalized directional derivatives and subgradients of nonconvex func-
tions. Can. J. Math. 32(2), 257–280 (1980)
493. Rockafellar, R.T.: The Theory of Subgradients and Its Applications to Problems of Optimiza-
tion, R & E, vol. 1. Heldermann Verlag, Berlin (1981)
494. Rockafellar, R.T.: First- and second-order epi-differentiability in nonlinear programming.
Trans. Am. Math. Soc. 307(1), 75–108 (1988)
495. Rockafellar, R.T.: Proto-differentiability of set-valued mappings and its applications in
optimization, analyse non linéaire (Perpignan, 1987). Ann. Inst. H. Poincaré Anal. Non
Linéaire 6(suppl.), 449–482 (1989)
496. Rockafellar, R.T., Uryasev, S.: Optimization of conditional value-at-risk. J. Risk 2, 21–41
(2000)
497. Rockafellar, R.T., Uryasev, S.: Conditional value-at-risk for general loss distributions. J.
Bank. Finance 26, 1443–1471 (2002)
498. Rockafellar, R.T., Uryasev, S., Zabarankin, M.: Deviation measures in risk analysis and
optimization. Preprint 7 (2002)
499. Rockafellar, R.T., Wets, R.J.B.: Variational Analysis. Grundlehren der Mathematischen
Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 317. Springer,
Berlin (1998)
500. Rodríguez-Marín, L., Sama, M.: About contingent epiderivatives. J. Math. Anal. Appl.
327(2), 745–762 (2007)
501. Rodríguez-Marín, L., Sama, M.: .; C /-contingent derivatives of set-valued maps. J. Math.
Anal. Appl. 335(2), 974–989 (2007)
502. Rodríguez-Marín, L., Sama, M.: Variational characterization of the contingent epiderivative.
J. Math. Anal. Appl. 335(2), 1374–1382 (2007)
503. Rodríguez-Marín, L., Sama, M.: Epidifferentiability of the map of infima in Hilbert spaces.
J. Math. Anal. Appl. 342(1), 371–385 (2008)
504. Rodríguez-Marín, L., Sama, M.: w -contingent epiderivatives in reflexive spaces. Nonlinear
Anal. 68(12), 3780–3788 (2008)
505. Rolewicz, S.: On paraconvex multifunctions. In: Third Symposium on Operations Research
(Univ. Mannheim, Mannheim, 1978), Section I, Operations Res. Verfahren, vol. 31, pp. 539–
546. Hain, Königstein/Ts (1979)
506. Rolewicz, S.: On -paraconvex multifunctions. Math. Jpn. 24(3), 293–300 (1979/80)
507. Rolewicz, S.: On ˛./-monotone multifunctions and differentiability of strongly ˛./-
paraconvex functions, well-posedness in optimization and related topics (Warsaw, 2001).
Control Cybern. 31(3), 601–619 (2002)
508. Rouhani, B.D., Khan, A.A., Raciti, F.: Penalization and regularization for multivalued pseudo-
monotone variational inequalities with Mosco approximation on constraint sets. J. Global
Optim. 40(1-3), 147–153 (2008)
509. Rubinov, A., Gasimov, R.: Scalarization and nonlinear scalar duality for vector optimization
with preferences that are not necessarily a pre-order relation. J. Global Optim. 29(4), 455–477
(2004)
754 References
510. Rubinov, A., Singer, I.: Topical and sub-topical functions. downward sets and abstract
convexity. Optimization 50, 307–351 (2001)
511. Rubinov, A.M.: Sublinear operators and their applications. Uspehi Mat. Nauk 32(4(196)),
113–174, 287 (1977)
512. Ruszczyński, A., Shapiro, A.: Optimization of convex risk functions. Math. Oper. Res. 31(3),
433–452 (2006)
513. Sach, P.: Nearly subconvexlike set-valued maps and vector optimization problems. J. Optim.
Theory Appl. 119, 335–356 (2003)
514. Sách, P.H.: Differentiability of set-valued maps in Banach spaces. Math. Nachr. 139, 215–235
(1988)
515. Sach, P.H.: Sufficient conditions for generalized convex set-valued maps. Optimization 37(4),
293–304 (1996)
516. Sach, P.H.: Another characterization of convexity for set-valued maps. Numer. Funct. Anal.
Optim. 20(3-4), 341–351 (1999)
517. Sach, P.H.: New generalized convexity notion for set-valued maps and application to vector
optimization. J. Optim. Theory Appl. 125(1), 157–179 (2005)
518. Sach, P.H., Yen, N.D.: Convexity criteria for set-valued maps. Set Valued Anal. 5(1), 37–45
(1997)
519. Sama, M.: Some remarks on the existence and computation of contingent epiderivatives.
Nonlinear Anal. 71(7-8), 2997–3007 (2009)
520. Sama, M.: The role of directional compactness in the existence and computation of contingent
epiderivatives. J. Math. Anal. Appl. 372(1), 262–272 (2010)
521. Sawaragi, Y., Nakayama, H., Tanino, T.: Theory of multiobjective optimization. Mathematics
in Science and Engineering, Vol. 176. Academic Press, Orlando (Harcourt Brace Jovanovich,
Publishers). (1985)
522. Sayin, S., Kouvelis, P.: The multiobjective discrete optimization problem: A weighted min-
max two-stage optimization approach and a bicriteria algorithm. Manag. Sci. 51, 1572–1581
(2005)
523. Schinas, J., Boudourides, M.: Higher order differentiability of multifunctions. Nonlinear
Anal. 5(5), 509–516 (1981)
524. Schirotzek, W.: Nonsmooth Analysis. Universitext. Springer, Berlin (2007)
525. Schönfeld, P.: Some duality theorems for the non-linear vector maximum problem.
Unternehmensforschung 14, 51–63 (1970)
526. Scott, E.M., Saltelli, A., Sörensen, T.: Practical experience in applying sensitivity and
uncertainty analysis. In: Sensitivity Analysis, Wiley Ser. Probab. Stat., pp. 267–274. Wiley,
Chichester (2000)
527. Severi, F.: Sulla differenziabilità totale delle funzioni di più variabili reali. Ann. Mat. Pura
Appl. 13(1), 1–35 (1934)
528. Shao, L., Ehrgott, M.: Approximately solving multiobjective linear programmes in objective
space and an application in radiotherapy treatment planning. Math. Methods Oper. Res. 68(2),
257–276 (2008)
529. Shao, L., Ehrgott, M.: Approximating the nondominated set of an MOLP by approximately
solving its dual problem. Math. Methods Oper. Res. 68(3), 469–492 (2008)
530. Shi, D.S.: Contingent derivative of the perturbation map in multiobjective optimization.
J. Optim. Theory Appl. 70,no.(2), 385–396 (1991)
531. Shi, D.S.: Sensitivity analysis in convex vector optimization. J. Optim. Theory Appl. 77,
145–159 (1993)
532. da Silva, M.R.: Nonnegative order iterates of Bernstein polynomials and their limiting
semigroup. Portugal. Math. 42(3), 225–248 (1983/84)
533. Silverman, R.J., Yen, T.: The Hahn-Banach theorem and the least upper bound property.
Trans. Am. Math. Soc. 90, 523–526 (1959)
534. Simons, S.: Extended and sandwich versions of the Hahn-Banach theorem. J. Math. Anal.
Appl. 21, 112–122 (1968)
References 755
535. Song, W.: Conjugate duality in set-valued vector optimization. J. Math. Anal. Appl. 216(1),
265–283 (1997)
536. Song, W.: A generalization of Fenchel duality in set-valued vector optimization. Math.
Methods Oper. Res. 48(2), 259–272 (1998)
537. Song, W.: Characterizations of some remarkable classes of cones. J. Math. Anal. Appl.
279(1), 308–316 (2003)
538. Song, W.: Weak subdifferential of set-valued mappings. Optimization 52(3), 263–276 (2003)
539. Sonntag, Y., Zălinescu, C.: Comparison of existence results for efficient points. J. Optim.
Theory Appl. 105, 161–188 (2000)
540. Soyster, A.: Convex programming with set-inclusive constraints and applications to inexact
linear programming. Oper. Res. 21, 1154–1157 (1973)
541. Steponavičė, I., Miettinen, K.: Survey on multiobjective robustness for simulation-based
optimization. In: Talk at the 21st International Symposium on Mathematical Programming,
August 19–24, 2012. Berlin, Germany (2012)
542. Sterna-Karwat, A.: Remarks on convex cones. J. Optim. Theory Appl. 59(2), 335–340 (1988)
543. Stewart, T., Bandte, O., Braun, H., Chakraborti, N., Ehrgott, M., Göbelt, M., Jin, Y.,
Nakayama, H., Poles, S., Stefano, D.D.: Real-world applications of multiobjective optimiza-
tion. In: Branke, J., Deb, K., Miettinen, K., Slowinski, R. (eds.) Multiobjective Optimization:
Interactive and Evolutionary Approaches. Lecture Notes in Computer Science, vol. 5252,
pp. 285–327. Springer, Berlin (2008)
544. Stiller, S.: Extending concepts of reliability: Network creation games, real-time scheduling,
and robust optimization. Ph.D. thesis, TU Berlin (2009)
545. Treiman, J.S.: Shrinking generalized gradients. Nonlinear Anal. 12,no.(12), 1429–1450
(1988)
546. Studniarski, M.: Application of the Dubovitskii-Milyutin method to some locally convex
extremal problems. Bull. Soc. Sci. Lett. Łódź 29(6), 8 (1979)
547. Taa, A.: Necessary and sufficient conditions for multiobjective optimization problems.
Optimization 36, 97–104 (1996)
548. Taa, A.: Set-valued derivatives of multifunctions and optimality conditions. Numer. Funct.
Anal. Optim. 19(1-2), 121–140 (1998)
549. Taa, A.: Subdifferentials of multifunctions and Lagrange multipliers for multiobjective
optimization. J. Math. Anal. Appl. 283(2), 398–415 (2003)
550. Taa, A.: -subdifferentials of set-valued maps and -weak Pareto optimality for multiobjective
optimization. Math. Methods Oper. Res. 62(2), 187–209 (2005)
551. Taa, A.: On subdifferential calculus for set-valued mappings and optimality conditions.
Nonlinear Anal. 74(18), 7312–7324 (2011)
552. Taa, A.: Second-order conditions for nonsmooth multiobjective optimization problems with
inclusion constraints. J. Global Optim. 50(2), 271–291 (2011)
553. Tagawa, S.: Optimierung mit mengenwertigen Abbildungen. Dissertation, Univ. Mannheim,
Mannheim (1978)
554. Tagawa, S.: Optimierung mit mengenwertigen Abbildungen. In: Third Symposium on Oper-
ations Research (Univ. Mannheim, Mannheim, 1978), Section I, Operations Res. Verfahren,
vol. 31, pp. 619–629. Hain, Königstein/Ts (1979)
555. Takahashi, W.: Existence theorems generalizing fixed point theorems for multivalued map-
pings. In: Fixed point theory and applications (Marseille, 1989), Pitman Res. Notes Math.
Ser., vol. 252, pp. 397–406. Longman Sci. Tech., Harlow (1991)
556. Tam, B., Ehrgott, M., Ryan, D., Zakeri, G.: A comparison of stochastic programming and
bi-objective optimisation approaches to robust airline crew scheduling. OR Spectrum 33(1),
49–75 (2011)
557. Tammer, C.: Lagrange duality in vector optimization. (Lagrange-Dualität in der Vektoropti-
mierung.). Wiss. Z. Tech. Hochsch. Ilmenau 37(3), 71–88 (1991)
558. Tammer, C., Zălinescu, C.: Vector variational principles for set-valued functions. Optimiza-
tion 60(7), 839–857 (2011)
756 References
559. Tammer, C., Zălinescu, C.: Vector variational principles for set-valued functions. In: Recent
Developments in Vector Optimization, pp. 367–415. Springer, Berlin (2012)
560. Tammer, C., Zălinescu, C.: Lipschitz properties of the scalarization function and applications.
Optimization 59, 305–319 (2010)
561. Tanaka, T.: Some minimax problems of vector-valued functions. J. Optim. Theory Appl.
59(3), 505–524 (1988)
562. Tanaka, T., Kuroiwa, D.: The convexity of A and B assures int A C B D int.A C B/. Appl.
Math. Lett. 6(1), 83–86 (1993)
563. Tanino, T.: On supremum of a set in a multidimensional space. J. Math. Anal. Appl. 130(2),
386–397 (1988)
564. Tanino, T.: Sensitivity analysis in multiobjective optimization. J. Optim. Theory Appl. 56(3),
479–499 (1988)
565. Tanino, T.: Stability and sensitivity analysis in convex vector optimization. SIAM J. Control
Optim. 26(3), 521–536 (1988)
566. Tanino, T.: Conjugate duality in vector optimization. J. Math. Anal. Appl. 167(1), 84–97
(1992)
567. Tanino, T.: Conjugate duality in vector optimization. J. Math. Anal. Appl. 167(1), 84–97
(1992)
568. Tanino, T., Sawaragi, Y.: Duality theory in multiobjective programming. J. Optim. Theory
Appl. 27(4), 509–529 (1979)
569. Tanino, T., Sawaragi, Y.: Conjugate maps and duality in multiobjective optimization.
J. Optim. Theory Appl. 31(4), 473–499 (1980)
570. Tasset, T.N.: Lagrange multipliers for set-valued functions when ordering cones have empty
interior. ProQuest LLC, Ann Arbor, MI (2010), thesis (Ph.D.)–University of Colorado at
Boulder
571. Thibault, L.: Tangent cones and quasi-interiorly tangent cones to multifunctions. Trans. Am.
Math. Soc. 277(2), 601–621 (1983)
572. Thierfelder, J.: Nonvertical affine manifolds and separation theorems in product spaces. Math.
Nachr. 151, 329–344 (1991)
573. Thierfelder, J.: Separation theorems for sets in product spaces and equivalent assertions.
Kybernetika (Prague) 27(6), 522–534 (1991)
574. Treiman, J.S.: Characterization of Clarke’s tangent and normal cones in finite and infinite
dimensions. Nonlinear Anal. 7(7), 771–783 (1983)
575. Turinici, M.: Maximal elements in a class of order complete metric spaces. Math. Jpn. 25(5),
511–517 (1980)
576. Ursescu, C.: Multifunctions with convex closed graph. Czechoslovak Math. J. 25(100)(3),
438–441 (1975)
577. Ursescu, C.: Tangent sets’ calculus and necessary conditions for extremality. SIAM J. Control
Optim. 20(4), 563–574 (1982)
578. Ursescu, C.: On the calculus of some topologically open tangent sets. Boll. Un. Mat. Ital.
A (7) 1(2), 145–151 (1987)
579. Ursescu, C.: Adjacent and tangent regularity. Optimization 30(2), 105–118 (1994)
580. Valadier, M.: Sous-différentiabilité de fonctions convexes à valeurs dans un espace vectoriel
ordonné. Math. Scand. 30, 65–74 (1972)
581. Vangeldère, J.: Sur une famille d’ensembles particuliers d’un espace vectoriel. Bull. Soc.
Roy. Sci. Liège 38, 158–170 (1969)
582. Varaiya, P.: Nonlinear programming in Banach spaces. SIAM J. Appl. Math. 15, 284–293
(1967)
583. Vîrsan, C.: Necessary conditions of optimality of second order for the optimal control
problem with functional constraints. Rev. Roumaine Math. Pures Appl. 18, 767–791 (1973)
584. Vlach, M.: Approximation operators in optimization theory. Zeitschrift Oper. Res. 25, 15–23
(1981)
585. Walczak, S.: Some properties of cones in normed spaces and their application to investigating
extremal problems. J. Optim. Theory Appl. 42, 561–582 (1984)
References 757
586. Wang, Q., Li, S., Teo, K.: Higher-order optimality conditions for weakly efficient solutions in
nonconvex set-valued optimization. Optim. Lett. 4, 425–437 (2010)
587. Wang, Q., Yu, G.: Second-order optimality conditions for set-valued optimization problems
under Benson proper efficiency. Abstr. Appl. Anal. Art. ID 432,963, 16 (2011)
588. Wang, S.S.: A separation theorem for a convex cone on an ordered vector space and its
applications. Acta Math. Appl. Sinica 9(3), 309–318 (1986)
589. Wang, Y.T.: A characterization of paratingent cone and P -subderivative with applications in
nonsmooth analysis. Acta Math. Sinica (N.S.) 7(2), 181–192 (1991)
590. Ward, D.: Convex subcones of the contingent cone in nonsmooth calculus and optimization.
Trans. Am. Math. Soc. 302(2), 661–682 (1987)
591. Ward, D.: Isotone tangent cones and nonsmooth optimization. Optimization 18(6), 769–783
(1987)
592. Ward, D.: The quantificational tangent cones. Can. J. Math. 40(3), 666–694 (1988)
593. Ward, D.: Which subgradients have sum formulas? Nonlinear Anal. 12(11), 1231–1243
(1988)
594. Ward, D.: Calculus for parabolic second-order derivatives. Set Valued Anal. 1(3), 213–246
(1993)
595. Ward, D.: A chain rule for parabolic second-order epiderivatives. Optimization 28(3-4), 223–
236 (1994)
596. Ward, D.E.: A chain rule for first- and second-order epiderivatives and hypoderivatives.
J. Math. Anal. Appl. 348(1), 324–336 (2008)
597. Ward, D.E., Borwein, J.M.: Nonsmooth calculus in finite dimensions. SIAM J. Control
Optim. 25, 1312–1340 (1987)
598. Watkins, G.G.: Nonsmooth Milyutin-Dubovitskiı̆ theory and Clarke’s tangent cone. Math.
Oper. Res. 11(1), 70–80 (1986)
599. Webster, R.: Convexity. Oxford Science Publications. The Clarendon Press Oxford University
Press, New York (1994)
600. Xia, Z.Q., Wang, M.Z., Zhang, L.W.: Directional derivative of a class of set-valued mappings
and its application. J. Convex Anal. 10(1), 211–227 (2003)
601. Xu, Y., Song, X.: The relationship between ic-cone-convexness and nearly cone-
subconvexlikeness. Appl. Math. Lett. 24(9), 1622–1624 (2011)
602. Yang, X.M., Li, D., Wang, S.Y.: Near-subconvexlikeness in vector optimization with set-
valued functions. J. Optim. Theory Appl. 110(2), 413–427 (2001)
603. Yang, X.Q.: A Hahn-Banach theorem in ordered linear spaces and its applications. Optimiza-
tion 25(1), 1–9 (1992)
604. Yao, J.C., Yen, N.D.: Coderivative calculation related to a parametric affine variational
inequality. I. Basic calculations. Acta Math. Vietnam. 34(1), 157–172 (2009)
605. Young, R.C.: The algebra of many-valued quantities. Math. Ann. 104(1), 260–290 (1931)
606. Yu, P.: Cone convexity, cone extreme points, and nondominated solutions in decision
problems with multiobjectives. J. Optim. Theory Appl. 14, 319–377 (1974)
607. Zaffaroni, A.: Degrees of efficiency and degrees of minimality. SIAM J. Control Optim. 42,
1071–1086 (2003)
608. Zagrodny, D.: General sufficient conditions for the convexity of a function. Z. Anal.
Anwendungen 11(2), 277–283 (1992)
609. Zălinescu, C.: The Fenchel–Rockafellar duality theory for mathematical programming in
oreder-complete vector lattices and applications. Tech. Rep. 45, INCREST, Bucharest (1980)
610. Zălinescu, C.: Duality for vectorial nonconvex optimization by convexification and applica-
tions. An. Ştiinţ. Univ. Al. I. Cuza Iaşi Secţ. I a Mat. 29(3, suppl.), 15–34 (1983)
611. Zălinescu, C.: Optimality conditions in infinite dimensional convex programming
(Romanian). Ph.D. thesis, University Alexandru Ioan Cuza Iasi, Iasi, Romania (1983)
612. Zălinescu, C.: Stability for a class of nonlinear optimization problems and applications. In:
Nonsmooth Optimization and Related Topics (Erice, 1988), Ettore Majorana Internat. Sci.
Ser. Phys. Sci., vol. 43, pp. 437–458. Plenum, New York (1989)
758 References
613. Zălinescu, C.: A result on sets, with applications to vector optimization. Z. Oper. Res. 35(4),
291–298 (1991)
614. Zălinescu, C.: Convex Analysis in General Vector Spaces. World Scientific, River Edge
(2002)
615. Zălinescu, C.: Hahn-Banach extension theorems for multifunctions revisited. Math. Methods
Oper. Res. 68(3), 493–508 (2008)
616. Duality for vectorial convex optimization, conjugate operators and subdifferentials. The
continuous case. Illmenau University, Eisenach (1984), Internationale Tagung Mathematische
Optimierung – Theorie und Anwendungen
617. Zeidler, E.: Nonlinear Functional Analysis and its Applications. Springer, New York (1990)
618. Zheng, X.Y.: Proper efficiency in locally convex topological vector spaces. J. Optim. Theory
Appl. 94(2), 469–486 (1997)
619. Zheng, X.Y.: Scalarization of Henig proper efficient points in a normed space. J. Optim.
Theory Appl. 105(1), 233–247 (2000)
620. Zheng, X.Y., Ng, K.F.: The Fermat rule for multifunctions on Banach spaces. Math. Program.
104(1, Ser. A), 69–90 (2005)
621. Zheng, X.Y., Ng, K.F.: The Lagrange multiplier rule for multifunctions in Banach spaces.
SIAM J. Optim. 17(4), 1154–1175 (electronic) (2006)
622. Zheng, X.Y., Ng, K.F.: Metric subregularity and calmness for nonconvex generalized
equations in Banach spaces. SIAM J. Optim. 20(5), 2119–2136 (2010)
623. Zhou, Z.A., Yang, X.M., Peng, J.W.: -strict subdifferentials of set-valued maps and
optimality conditions. Nonlinear Anal. 75(9), 3761–3775 (2012)
624. Zhu, Q.J.: Hamiltonian necessary conditions for a multiobjective optimal control problem
with endpoint constraints. SIAM J. Control Optim. 39(1), 97–112 (2000)
625. Zhu, Q.J.: Nonconvex separation theorem for multifunctions, subdifferential calculus and
applications. Set Valued Anal. 12(1-2), 275–290 (2004)
626. Zhu, S.K., Li, S.J., Teo, K.L.: Second-order Karush–Kuhn–Tucker optimality conditions for
set-valued optimization. J. Global Optim. 58(4), 673–692 (2014)
627. Zhuang, D.: Density results for proper efficiencies. SIAM J. Control Optim. 32(1), 51–58
(1994)
628. Zowe, J.: A remark on a regularity assumption for the mathematical programming problem in
Banach spaces. J. Optim. Theory Appl. 25(3), 375–414 (1978)
629. Zowe, J.: Sandwich theorems for convex operators with values in an ordered vector space.
J. Math. Anal. Appl. 66(2), 282–296 (1978)
Index
Additively separate, 431 basic, normal, Mordukhovich, 105

Adjacent cone, 120 Fréchet, 105
Adjacent derivative, 407 Commutativity, 727
Adjacent epiderivative, 446 Compactly approximable, 419
Adjacently epidifferentiable, 446 Compactly Gamma-paraconvex, 545
Allocation Comparable elements, 12
feasible, 701 Cone, 15
Pareto-type optimal, 702 asymptotic, 361
A-scalarized subgradient, 528 barrier, 231
Asplund space, 104 based, 19
Associativity, 727 contingent, 647
Asymptotical compactness condition, continuous positive dual, 16
579 convex, 15
Asymptotically closed, 598 correct, 360
Daniell, 18, 19, 357
sequentially, 357
Base (for a cone), 19 Fréchet normal, 103
Base of neighborhoods, 729 generalized normal, 104
Bouligand tangent cone, 113 nontrivial, 15
Bound normal, 17
greatest lower, 55 ./-, 360
least upper, 55 pointed, 15
lower, 14 proper, 15
upper, 14 quasi-interior of C C , 16
B-subgradient, 528 recession, 219
regular, 19
reproducing, 15
Calm set-valued map, 405
supernormal, 19
Chain, 351
well-based, 19
Clarke derivative, 407
Cone of adherent dispalcements, 113
Clarke epiderivative, 446
Cone of admissible directions, 111
Clarke epidifferentiable, 446
Cone of attainable directions, 120
Clarke tangent cone, 126
Cone of epi Lipschitzian directions, 125
Closed radial cone, 135
Cone of feasible directions, 111
Closure (of a set), 17
Cone of interior displacements, 113
Coderivative

DOI 10.1007/978-3-642-54265-7
760 Index
Cone of interior tangent vectors, 111 Face, 687

Cone of linear tangent vectors, 111 exposed, 687
Cone of quasi-interior directions, 120 proper, 687
Cone of radial interior displacements, 111 Facet, 687
Cones of the opposite sense, 559 Family of C -proper sets, 236
Cones of the same sense, 559 Family of equivalence classes, 236
Cone-valued map Feasible tangent cone, 111
pointed, 70 Feasible tangent set of order m, 211
Contingent cone, 113 Fréchet normally regular, 625
Contingent derivative, 400 Free disposal, 219
Contingent epiderivative, 445 strong, 219
Contingently C-absorbing, 437 Function
Contingently epidifferentiable, 445 benefit, 218
Contingently-quasi-convex, 543 B-monotone, 220
Contingent set of order m, 211 strictly, 222
Convergence C -˛-convex, 27
Kuratowski–Painlevé, 83 C -convex, 27
net, 730 C -lower continuous (C -l.c.), 86
C-subdifferentiable, 310 C -lower continuous (C -l.c.) at, 86
C-subdifferential, 310 C -lower semicontinuous, 86
C-subgradient, 310 C -mid-convex, 27
C(x)-pseudo-convex, 570 C -mid-quasiconvex, 27
C -monotone, 213
C -nearly convex, 27
Derivable, 130, 416 C -nearly quasiconvex, 27
Derivable cone, 120 continuous, 731
Derivative convex, 214
contingent, 647 C -˛-quasiconvex, 27
Dilation, 41 C -quasiconvex, 27
Dini locally convex, 564 C -upper continuous (C -u.c.), 86
Dini locally regular, 564 C -upper semicontinuous, 86
Disjunction preserving, 147 domain of, 10
Distance, 99 epigraph of, 10
Hausdorff, 100 Hiriart-Urruty’s, 232
Domination property, 312 marginal, 281
monotonically lower semicontinuous,
362
regular, 278
Efficient point separating, 213
super, 30 shortage, 218
Element strictly D-monotone, 213
efficient, 356 strong conjugate, 280
maximal (minimal), 12 sublinear, 214
null, 727 upper continuous (u.c.)
unity, 727 C - (C -u.c.), 86
Embedding approach, 65 value, 281
Epi-Lipschitz-like, 156 Functional
Excess (of A over B), 78 increasing, 241
Existence of minimal solutions, 242 strictly increasing, 241
Extremal principle, 245
approximate extremal principle, 245
exact extremal principle, 245 Generalized contingent epiderivative, 470
"-extremal principle, 245 Generalized proper contingent epiderivative,
Extremal system, 245 470
Index 761
Generalized proper subdifferential, 528 Mangasarian-Fromovitz constraint

Generalized second-order adjacent set, 176 qualification, 164, 200
Generalized second-order contingent set, 176 Manifold
Generalized subdifferential, 528 linear, 14, 727
Generalized subgradient, 528 Map
Generalized weak contingent epiderivative, biconjugate, 309
470 c-biconjugate, 310
Generalized weak subdifferential, 528 c-conjugate, 310
conjugate, 309
Hadamard differentiable, 157 minimal value, 312
Homeomorphism, 731 perturbation, 342, 673
Hull set-valued perturbation, 311
affine, 14 ˚k0 -stable, 242
convex, 14 Mathematical finance, 215
linear, 14 Maximal point
Hypertangent cone, 125 Pareto, 28
weakly, 28
Maximizer, 46
Indicator function, 102, 231 weak, 46
Infimal set, 57 Metric, 728
Infimum of a set, 14, 55 Metrically regular, 426, 646, 647
Interior, 17 Metrically semi-regular, 427
algebraic Metric regularity, 590
relative, 421 Minimal point
Interiorly adjacent cone, 120 Pareto, 28
Interiorly adjacent set of order m, 211 properly
Interiorly contingent cone, 113 Benson, 29
Interiorly contingent set of order m, 211 Borwein, 29
Interiorly prototangent cone, 149 Hartley, 29
Interiorly quasi-strict tangent cone, 156 Henig, 30
Interiorly Treiman tangent cone, 151 Henig global, 30
Intermediate cone, 120 Hurwicz, 29
Isotone, 147 S-properly, 29
Q-minimal, 37
strong, 29
Lattice v-minimal, 44
complete, 55 weakly, 28
Level-set mapping, 67 weakly v-minimal, 44
Limit, 730 Minimizer, 46
inferior, 78 approximate, 73
superior, 78 fully localized, 68
Lineality, 40 Q-minimizer, 46
Linearly semi-open, 426 quasi-weak, 48
Lipschitz v-minimizer, 70
epi-Lipschitz, 228 weak, 46
Lipschitz-like Modified adjacent cone , 126
epigraphically, 101 Modified cone of feasible directions, 124
Lipschitz Properties, 96 Modified cone of interior displacements, 125
Local approximation of second-order, 207 Modified contingent cone, 125
Local cone approximation, 143 Modified feasible tangent cone, 124
Local extremal point, 244 Modified interiorly adjacent cone, 126
Local upper-Lipschitzian selection, 624 Modified radial tangent cone, 124
Lower Dini derivative, 408 Multifunction
Lower semidifferentiable, 430 epigraphical, 101
762 Index
Nagumo cone, 120 primal, 281

Neighborhood, 729 stable, 282
Net, 730 Projection, 9, 647
C -decreasing, 356 Proper subgradient, 528
C -increasing, 356 Property
strictly, 356 domination, 351
convergent, 730 Proto-differentiable, 416
lC -decreasing, 242 Prototangent cone, 149
t -decreasing, 351 Pseudo-convex set-valued map, 403
t -increasing, 351 Pseudo tangent cone, 123
strictly, 351 P-tangent cone, 152
Newton method, 645, 648
damped, 648
Normal cone Quasi-interiorly Clarke derivative, 408
basic, limiting, Mordukhovich, 104 Quasi-strict tangent cone, 156
Clarke, 102
Normally C-minimal point, 616
Radial cone, 111
Radial derivative, 408
Open covering, 426 Radial tangent cone, 110
Optimally positioned sets, 562 Radial tangent set of order m, 210
Order Recession adjacent cone, 148
lexicographic, 16 Recession contingent cone, 148
linear, 12 Relation
partial, 12 antisymmetric, 12
total, 12 binary, 11
Order adjacent set of order m, 211 composition, 12
Order relation inverse, 12
certainly less, 50 order, 11
lower set less, 49 reflexive, 12
minmax certainly less, 52 transitive, 12
minmax certainly nondominated, 53 Risk measure
minmax less, 52 coherent, 216, 217
possible less, 51 value at risk, 216
set less, 48 worst-case, 217
upper set less, 49 Robustness for vector-valued optimization
Oriented distance, 215 problems, 706
Outer tangent cone, 113
Satiation requirement, 704

Pair Scalarizing functional
complementary, 684 conjugate of, 231
Point subdifferential of, 231
efficient, 356 S-derivative, 438
Preference mapping, 701 Second-order adjacent derivative, 488
Preference relation Second-order adjacent set, 173
almost transitive, 68 Second-order asymptotic adjacent cone, 185
closed, 68 Second-order asymptotic adjacent derivative,
locally satiated, 68 489
nonreflexive, 68 Second-order asymptotic contingent cone, 183
Preference set, 701 Second-order asymptotic contingent derivative,
Preorder, 12 489
Problem Second-order asymptotic contingent
dual, 281 epiderivative, 504
Index 763
Second-order asymptotic feasible tangent cone, Separation

182 of nonconvex sets, 224
Second-order asymptotic interiorly adjacent Sequence
cone, 185 generalized, 730
Second-order asymptotic interiorly contingent Set
cone, 183 absorbing, 731
The second-order asymptotic radial tangent affine, 14, 727
cone, 182 asymptotically compact (a.c.), 361
Second-order Clarke’s tangent cone, 187 balanced, 731
Second-order compactly approximable, 497 bounded
Second-order compound adjacent derivative, C -upper (lower), 356
489 lower, 14
Second-order compound Clarke derivative, 489 quasi (C -lower), 91
Second-order compound contingent derivative, upper, 14
488 boundedly order complete, 358
Second-order compound contingent C -bounded, 237
epiderivative, 504 C -closed, 237
Second-order compound generalized C -compact, 237, 360
contingent epiderivative, 505 C -complete, 359
Second-order compound proto-differentiable, sequentially, 359
490 closed
Second-order contingent derivative, 488 sequentially, 730
Second-order contingent epiderivative, 504 closed around x, 104
Second-order contingently C-absorbing, 494 closely convex, 22
Second-order contingent set, 171 cofinal, 729
Second-order derivable, 490 connected, 81
Second-order derivable cone, 206 convex, 14
Second-order directionally compact, 496 ˛-convex, 22
Second-order feasible cone, 202 C -proper, 236
Second-order feasible tangent set, 170 cs-closed, 377
Second-order generalized asymptotic cs-complete, 377
contingent epiderivative, 504 C -semicompact, 360
Second-order generalized contingent directed, 729
epiderivative, 504 epigraph type, 220
Second-order incident cone, 202 full, 17
Second-order indexed adjacent set, 176 K C -bounded, 370
Second-order indexed contingent set, 176 lower bounded, 18
Second-order interiorly adjacent set, 173 mid-convex, 22
Second-order interiorly contingent set, 171 nearly convex, 22
Second-order lower Dini derivative, 489 open, 728
Second-order projective adjacent cone, 204 power, 45
Second-order projective contingent cone, 204 strict sublevel, 10, 26
Second-order proto-differentiable, 490 sublevel, 10, 26
Second-order radial tangent set, 170 well-ordered, 12, 351
Second-order semi-differentiable, 490 Set-valued
Second-order weak projective adjacent cone, upper Lipschitz, 411
204 Set-valued map, 9
Second-order weak projective contingent cone, affinelike, 256
204 C -˛-quasiconvex, C -mid-quasiconvex, C -
Section nearly quasiconvex, C -quasiconvex,
lower, 350 27
upper, 350 C -continuous, 83
Semi-differentiable, 416 Hausdorff, 83
Semi-pseudo-Lipschitz, 427 C -convex, C -˛-convex, 26
764 Index
closed, 80, 100 super, 46

at a point, 80 weak, 46
closed-valued, 80 solution concept, 45
C -lower continuous (C -l.c.), 83 strict approximate "k 0 -minimizer, 74
Hausdorff (H-C -l.c.), 83 Set-valued problem .SPv /
uniformly (at a point on a set), 83 v-minimizer, 70
C -mid-convex, C -nearly convex, weak v-minimizer, 70
26 Slowly convergent sequences, 121
compact, 100 Space
compact (at a point), 80 embedding, 65
composition, 10 extended, 56
connected, 292 least upper bound property, 250
continuous, 78 lineality, 356
Hausdorff, 82 linear, 727
convex, ˛-convex, 25 metric, 728
C -upper bounded weakly, 85 perturbation, 311
C -upper continuous (C -u.c.), 83 quasi-complete, 358
Hausdorff (H-C -u.c.), 83 quotient, 66
domain of, 9 topological, 728
epigraph of, 26 first-countable, 729
graph of, 9 Hausdorff, 730
image of, 9 linear, 731
inner-semicompact, 600 vector, 731
inverse, 9 vector, 727
Lipschitz, 99 Stability of KKT points, 600
Lipschitz-like, 98 Stable, 343
Lipschitz-like around, 98 externally, 312
lower continuous (l.c.), 77 with respect to the perturbation map,
Hausdorff (H-l.c.), 82 312
lower semicontinuous Strictly differentiable, 157
C - (C -l.s.c.), 85 Strictly proto-differentiable, 416
mid-convex, nearly convex, 25 Strictly semi-differentiable, 417
strongly compactly Lipschitzian, 101 Strict normal cone, 162
sublevel, strict sublevel, 27 Structure
upper continuous (u.c.), 77 compatible, 15
Hausdorff (H-u.c.), 81 linear, 727
upper semicontinuous Subdifferentiable, 309
C - (C -u.s.c.), 85 Subdifferential, 309
weakly f-monotone, 71 approximate, 107
Set-valued problem (SP), 45 basic, limiting, Mordukhovich, 105
approximate "k 0 -minimizer, 73 Clarke, 102
fully localized minimizer, 69 Fréchet, 106
fully localized strong minimizer, 69 limiting, 106
fully localized weak minimizer, 68 strong, 276
minimal solution, 53 weak, 288
minimizer, 46 Subgradient, 309
Benson proper, 46 basic, limiting, Mordukhovich, 105
Borwein proper, 46 positive, 342
Hartley proper, 46 strong, 276
Henig global proper, 46 Subnet, 730
Henig proper, 46 Subspace
Hurwicz proper, 46 linear, 727
Q-minimizer, 46 Support functional, 534
S-proper, 46 Supremum of a set, 14, 55
Index 765
Tangent cone, 113 Weak contingent cone, 119, 423

Tangentially regular, 130 Weak contingent derivative, 423
Theorem Weak-domination property, 476
nonconvex separation, 224 Weak feasible tangent cone, 111
Topology, 728 Weak-inverse image, 523
core convex, 261 Weakly efficient solution, 682, 683
linear, 731 Weak subgradient, 528
product, 731 Weak vector variational inequality, 634
Treiman tangent cone, 151 Welfare economics, 701
admissible state, 701
Uniformly positive polar, 624 consumption plans, 701
Upper closure, 57 production strategies, 701
Upper limit
Kuratowski-Painlevé, 103
Upper semidifferentiable, 430 Young-Fenchel inequality, 310
Ursescu cone, 120
Vector variational inequality, 634 Zorn’s lemma, 14

Vertex, 687

2015 Book Set-ValuedOptimization

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2015 Book Set-ValuedOptimization

Uploaded by

Copyright:

Available Formats

Vector Optimization

The series in Vector Optimization contains publications in various ﬁelds of opti-

More information about this series at

ISSN 1867-8971 ISSN 1867-898X (electronic)

© Springer-Verlag Berlin Heidelberg 2015

Printed on acid-free paper

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

(#210443 to Akhtar Khan). The research of C. Zălinescu was supported partially by

Rochester, NY, USA Akhtar A. Khan

R The set of real numbers

Im f D ff .x/ j x 2 X g Image (range) of the function f W

icr A Relative algebraic interior (or

Inf A Inﬁmal set of A  Y with respect

3.5 Limiting Cones and Generalized Differentiability . . . . . . . . . . . . . . . . 103

4.11.3 Second-Order Asymptotic Adjacent Cone

7 Conjugates and Subdifferentials .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 275

11 Derivatives and Epiderivatives of Set-Valued Maps . . . . . . . . . . . . . . . . . . . . 399

12.2.4 Lagrange Multiplier Rules

13.3 Sensitivity Analysis in Set-Valued Optimization

Fig. 2.1 Cone properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 19

Fig. 11.5 (i) A view of epi .F /. (ii) A contradiction

Set-valued optimization is a vibrant and expanding branch of applied mathematics

1.1 Motivating Examples

© Springer-Verlag Berlin Heidelberg 2015 1

Example 1.1.1 (Game Theory).

F .a/ lC F .a/ for some a 2 M H) F .a/ lC F .a/:

lC minimize F .a/; subject to a 2 M: (SP lC )

When Player 2 is non-cooperative (Player 2 wants Player 1 to lose), Player 1

F .a/ uC F .a/ for some a 2 M H) F .a/ uC F .a/:

Then the vector-valued game can again be formulated as a set-valued optimization

uC minimize F .a/; subject to a 2 M: (SP uC )

Further interesting applications of similar set-valued approaches are given in the

minimizeC f .x/ subject to x 2 S D fx 2 M j g.x/ 2 Kg (P)

where M Rn ; C Rp and K Rm are all pointed closed convex cones, and

˚./ D Min.fL.x; /j x 2 M g; C /:

maximizeC ˚./ subject to  2 L : (D)

We derive duality statements for vector optimization problems using a set-valued

We deﬁne a set-valued mapping ˚ W X Y by

˚.x/ D Max.fh .x/;

minimizeQ0 ˚.x/ subject to x 2 Q:

1.2 Book Structure

The organization of this book is as follows:

set-valued optimization. We introduce and discuss the following solution concepts

in Chap. 10. Continuing, we study characterizations of the solutions of set-valued

1.3 Useful Notation

We conclude this chapter by introducing some notation:

dom F WD fx 2 X j F .x/ ¤ ;g;

while its image is

The set-valued map F W X Y is usually identiﬁed with its graph,

graph F WD f.x; y/ 2 X  Y j y 2 F .x/g:

dom F D PrX .graph F /

so, Im F D F .X /. The inverse image by F of the set B F is

F 1 .B/ WD fx 2 X j F .x/ \ B ¤ ;g:

In fact, F 1 .B/ is the image of B given by the inverse set-valued map F 1 W

so F 1 .y/ D F 1 .fyg/. Of course, dom F 1 D Im F and Im F 1 D dom F . We

When  W X W is another set-valued map, the composition of  and F is the

When Y is a linear space and F; F1 ; F2 W X Y , we deﬁne the sum F1 C F2

.F1 C F2 /.x/ WD F1 .x/ C F2 .x/;

with the usual conventions that A C ; WD ; C A WD ;, ; D ; for A Y .

dom f WD fx 2 X j f .x/ 2 Y [ f1gg;

2.1 Order Relations

The following deﬁnition gives the notion of an order relation.

© Springer-Verlag Berlin Heidelberg 2015 11

R 1 WD f.x1 ; x2 / 2 M  M j .x2 ; x1 / 2 Rg;

Inf A Inﬁmal set of A Y with respect

F .a/ lC F .a/ for some a 2 M H) F .a/ lC F .a/:

lC minimize F .a/; subject to a 2 M: (SP lC )

F .a/ uC F .a/ for some a 2 M H) F .a/ uC F .a/:

uC minimize F .a/; subject to a 2 M: (SP uC )

˚./ D Min.fL.x; /j x 2 M g; C /:

maximizeC ˚./ subject to 2 L : (D)

˚.x/ D Max.fh .x/;

graph F WD f.x; y/ 2 X Y j y 2 F .x/g:

When W X W is another set-valued map, the composition of and F is the

R 1 WD f.x1 ; x2 / 2 M M j .x2 ; x1 / 2 Rg;

2. If R is an order relation on M and ; ¤ M0 M , then R0 WD R \ .M0 M0 / is

where x D .x1 ; : : : ; xn /, y D .y1 ; : : : ; yn / and 1; n WD fi 2 N j 1

y1 C y2 if and only if y2 y1 2 C for all y1 ; y2 2 Y: (2.9)

valid for all subsets A; B Y , and the fact that cl C D cl.int C /:

Letting ! y .x/; 0 ! y .x 0 / we get y .˛x C .1 ˛/x 0 /