Text
                    Jean-Pierre Aubin
Helene Frankowska
Set-Valued Analysis
With 11 Illustrations
1990
Birkhauser
Boston - Basel · Berlin


Jean-Pierre Aubin and Helene Frankowska CEREMADE (Centre de Recherche de Mathematiques de la Decision) Universite de Paris-Dauphine and CNRS F-75775 Paris Cedex 16 France AMS Subject Classification (1985): 26A27, 26A51, 26E25, 28B20, 28D05, 34A60, 49A52, 54C60, 54C65, 58C06, 58C07, 58C30, 93C15, 93C30. Library of Congress Cataloging-in-Publication Data Aubin, Jean-Pierre. Set-valued analysis / Jean-Pierre Aubin, Helene Frankowska. p. cm.—(Systems & control ; v. 2) Includes bibliographical references. ISBN 0-8176-3478-9 (alk. paper) 1. Set-valued maps. I. Frankowska, Helene. II. Title. ΙΠ. Series. QA611.3.A93 1990 514'.32—dc20 89-48854 Printed on acid-free paper. © Birkhauser Boston, 1990 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior permission of the copyright owner. ISBN 0-8176-3478-9 ISBN 3-7643-3478-9 Photocomposed text prepared using the authors' TeX files. Printed and bound by R.R. Donnelley & Sons, Harrisonburg, Virginia. Printed in the U.S.A. 987654321
This book is dedicated to C. Olech and Polish Mathematicians1 who contributed so much to set-valued analysis. 1 "It was a common belief that cultivation of science and the growth of its potential would somehow guarantee the maintenance of the nation" wrote Kura- towski about the'situation of Poland before 1918.
I will share all [my results] with you whenever you wish and do so without any ambition, from which I am more exempt and more distant than any man in the world. Pierre de Fermata Fermat was one of the most important "innovators in the history of mathematics. Newton himself recognized explicitly that he got the hint of the differential calculus from Fermat's method of building tangents devised half a century earlier. This same method is used in our book to build a differential calculus for set-valued maps. Fermat was also the one who discovered that the derivative of a (polynomial) function vanishes when it reaches an extremum. (This is Fermat's Rule, which remains the main strategy for obtaining necessary conditions of op- timality, from mathematical programming to calculus of variations to optimal control). Fermat also was the first to discover the "principle of least time" in optics, the prototype of the variational principles governing so many physical and mechanical laws. He shared independently with Descartes the invention of analytic geometry and with Pascal the creation of the mathematical theory of probability. His achievements in number theory overshadowed his other contributions, and the Last Fermat Theorem remains a challenge. He was on top of that a poet, a linguist, ... and made his living as a lawyer! ain his answer to a first letter from Mersenne inviting him to share his findings with the Parisian mathematicians, which put an end to Fermat's isolation in Toulouse, in 1636.
- "^ЗНХЖ Li ЬЛЬиШ P. DE FERMAT
Epigraph Who needs set-valued analysis? Everyone, we are tempted to say, and we shall state our case. This strong conviction — born out of accumulated experience in using it in control theory and differential games, mathematical economics and game theory, biomathematics, qualitative physics and viability theory — led us to devote time and effort to share some basic material which is used over and over. One can no longer afford the luxury of studying only well posed problems in Hadamard's sense: ΠΙ posed problems, inverse problems and many other unorthodox problems under other names are popping up in every domain of activity, whenever the existence of a solution may fail for some data, whenever uniqueness of the solution is at stake. Requiring that maps should be always single-valued, and even bijective, is too costly an attitude, above all in many applied fields, where we are not free to make such assumptions. This was indeed recognized during the three first decades of this century by the founders of "Functional Calculus": Painleve, Hausdorff, Bouligand, Kuratowski to quote only a few. In his important book Topologie, Kuratowski gave set-valued maps their proper status. Set-valued maps were abandoned by the authors of Bourbaki's volume Topologie Generale, who chose to restrict their study to single-valued maps, regarding set-valued maps as single-valued maps from a set to the power set of another set, or factorizing single-valued maps to make them bijective. This is not always the solution, for, by so doing, many important structural properties may be unfortunately lost; others are useless artifacts, making life more difficult rather than more simple. These points of view, which were widely disseminated all over the world
χ after World War II, misled many of us into unnecessary detours (often towards culs de sac), encouraging the perception that direct routes were too arduous, or worse, that they did not exist. Hence, set-valued analysis inherited the undeserved image of being something difficult and mysterious and, consequently, was regarded as a mathematical curiosity, to be left in the hands of mathematicians who like to generalize for the sake of generalizing, without proper motivations. In contrast, as it turned, out, the need for set-valued analysis in solving problems arising in other fields of knowledge — control theory, economics and management, biology and systems sciences, artificial intelligence, etc. — was pressing enough to help mathematicians overcome the kind of recalcitrance felt towards set-valued analysis. In view of such a wide variety of motivating applications, it is fortunate that most of the basic results of the chapters of "single-valued" analysis can be adapted to the set-valued realm. These include: β Limits and Continuity β Linear Functional Analysis β Nonlinear Functional Analysis (existence and approximation of solutions to equations and inclusions) β Tangents and Normals β Differentiation of Maps β Gradients of Functions and the Fermat Rule β Convergence of Maps β Measures and Integration β Differential Equations The set-valued version of this list is nothing other than the outline of this book. Our account of set-valued analysis is by no means exhaustive: it is just an introduction. The choice of the material has been dictated
XI by our experience in applying these results in control and viability theory. However, we tried very hard to make this presentation as clear as possible, to help the reader become familiar with the main tools. We did not restrict our exposition to the framework of finite dimensional vector-spaces, since set-valued analysis is also useful for solving problems involving partial differential equations or inclusions. But whenever the proofs of the finite-dimensional and infinite- dimensional statements are quite different, two proofs are provided. This book presents only the tools, without mentioning applications. Some applications (to control theory and viability theory in particular) are presented in companion texts2. Jean-Pierre Aubin Helene Frankbwska Paris September 21, 1989 2Viability Theory [50] by Aubin, and the forthcoming books Set-Valued Analysis and Control Theory [195] by Frankowska and Set-Valued Analysis and Subdifferential Calculus [355] by Rockafellar and Wets.
Acknowledgments We are indebted to. CEREMADE (Centre de Recherches de Mathematiques de la Decision), Universite de Paris-Dauphin.e, the Scuola Normale Dl Pisa3 and the Systems and Decision Sciences4 of IIASA (International Institute for Applied Systems Analysis) and their members5, who provided us with the intellectual environment which made possible the writing of this book6. We are also grateful to C. Hess, M. Valadier, 0. Dordan, R. Du- luc, M. Quincampoix and C. Vigneron for their numerous suggestions for improving the manuscript7. We are happy to publish this monograph in the new series Systems and Control: Foundations and Applications of Birkhauser. 3and in particular, G. da Prato. especially, its head, A. Kurzhanski, as well as the dedicated and competent librarians of IIASA. 5C. Byrnes, R. Rockafellar and R. Wets among many other colleagues 6 "...[a mathematician] does need a proper atmosphere; this proper atmosphere can be created only by the cultivation of common topics." Janiszewski wrote in 1918 in On the Needs of Mathematics in Poland, a program which in our mind can still be used all over the world. 7Fmally, as it is customary, each author is grateful to her/his coauthor for his/her wonderful typing of the manuscript.
Contents Introduction 1 1 Continuity of Set-Valued Maps 15 1.1 Limits of Sets 16 1.1.1 Definitions 16 1.1.2 The Compactness Theorem 23 1.1.3 The Duality Theorem 24 1.1.4 Convex Hull of Limits 26 1.2 Calculus of Limits 27 1.2.1 Direct Images 28 1.2.2 Inverse Images 30 1.3 Set-Valued Maps 33 Ί.4 Continuity of Set-Valued maps 38 1.4.1 Definitions 38 1.4.2 Generic Continuity 44 1.4.3 Example: Parametrized Set-Valued Maps ... 46 1.4.4 Marginal Maps 48 1.5 Lower Semi-Continuity Criteria 49 2 Closed Convex Processes 55 2.1 Definitions 56 2.2 Open Mapping and Closed Graph Theorems 57 2.3 Uniform Boundedness Theorem 61 2.4 The Bipolar Theorem 62 2.5 Transposition of Closed Convex Process 67 2.6 Upper Hemicontinuous Maps 74
xiv 3 Existence and Stability of an Equilibrium 77 3.1 Ky Fan's Inequality 80 3.2 Equilibrium and Fixed Point Theorems 83 3.2.1 The Equilibrium Theorem 83 3.2.2 Fixed Point Theorems 86 3.2.3 The Leray-Schauder Theorem 89 3.3 Ekeland's Variational Principle 91 3.4 Constrained Inverse Function Theorem 93 3.4.1 Derivatives of Single-Valued Maps 93 3.4.2 Constrained Inverse Function Theorems .... 94 3.4.3 Pointwise Stability Conditions 101 3.4.4 Local Uniqueness 103 3.5 Monotone and Maximal Monotone Maps 104 3.5.1 Monotone Maps 104 3.5.2 Maximal Monotone Maps 106 3.5.3 Yosida Approximations Ill 3.6 Eigenvectors of Closed Convex Processes 114 4 Tangent Cones 117 4.1 Tangent Cones to a Subset 121 4.1.1 Contingent Cones 121 4.1.2 Elementary Properties of Contingent Cones . . 125 4.1.3 Adjacent and Clarke Tangent Cones 126 4.1.4 Sleek Subsets 130 4.1.5 Limits of Contingent Cones; Finite Dimensional Case 130 4.1.6 Limits of Contingent Cones; Infinite Dimensional Case 132 4.2 Tangent Cones to Convex Sets 138 4.3 Calculus of Tangent Cones 146 4.3.1 Intersection and Inverse Image 146 4.3.2 Example: Tangent cones to subsets defined by equality and inequality constraints 150 4.3.3 Direct Image 153 4.4 Normal Cones 156 4.5 Other Tangent Cones 159 4.5.1 Convex Kernel of a Cone 159 4.5.2 Paratingent Cones 160
XV 4.5.3 Hypertangent Cones 164 4.5.4 A Menagerie of Tangent Cones 165 4.6 Tangent Cones to Sequences of Sets 166 4.7 Higher Order Tangent Sets 171 5 Derivatives of Set-Valued Maps 179 5.1 Contingent Derivatives 181 5.2 Adjacent and Circatangent Derivatives 189 5.2.1 Definitions and Elementary Properties 189 5.2.2 Limits of Differential Quotients 191 5.2.3 Derivatives of monotone operators 194 5.3 Chain Rules 196 5.4 Inverse Set-Valued Map Theorem 203 5.4.1 Stability and Approximation of Inclusions . . . 203 5.4.2 Localization of Inverse Images 206 5.4.3 The Equilibrium Map 207 5.4.4 Local Injectivity 209 5.5 Qualitative Solutions 210 5.6 Higher Order Derivatives 215 6 Epiderivatives of Extended Functions 219 6.1 Contingent Epiderivatives 222 6.1.1 Extended Functions and their Epigraphs .... 222 6.1.2 Contingent Epiderivatives 224 6.1.3 Fermat and Ekeland Rules 232 6.1.4 Elementary Properties 234 6.2 Other Epiderivatives 236 6.2.1 Adjacent and Circatangent Epiderivatives . . . 236 6.2.2 Other Convex Epiderivatives 240 6.3 Epidifferential Calculus 242 6.4 Generalized Gradient 248 6.4.1 Subdifferentials and Generalized Gradients . . 248 6.4.2 Limits of Subdifferentials and Gradients . . . .251 6.4.3 Local Subdifferentials and Superdifferentials . . 252 6.4.4 Remarks 254 6.5 Convex Functions 255 6.6 Higher Order Epiderivatives 259
6.6.1 Second Order Epiderivatives of Moreau-Yosida Approximations 263 Graphical & Epigraphical Convergence 265 7.1 Graphical Limits 267 7.1.1 Definitions 267 7.1.2 Graphical Convergence of Closed Convex Processes 269 7.1.3 Monotone and Maximal Monotone Maps .... 270 7.2 Convergence Theorems 270 7.3 EpiUmits 274 7.3.1 Definitions and Elementary Properties 274 7.3.2 Convergence of Infima and Minimizers 281 7.3.3 Variational Systems 284 7.4 EpiUmits of Sums and Composition Products 286 7.5 Conjugate Functions of Epilimits 289 7.6 Graphical Convergence of Gradients 294 7.6.1 Convergence of Gradients of Smooth Functions 295 7.6.2 Convergence of Subdifferentials of Convex Functions 297 7.7 Asymptotic Epiderivatives 300 Measurability and Integration of Set-Valued Maps 303 8.1 Measurable Set-Valued Maps 306 8.2 Calculus of Measurable Maps 310 8.3 Proof of the Characterization Theorem 319 8.4 Limits of Measurable Maps and Selections 322 8.5 Tangent Cones in Lebesgue Spaces 324 8.6 Integral of Set-Valued Maps 326 8.7 Proofs of the Convexity of the Integral 333 8.7.1 Finite dimensional case 333 8.7.2 Infinite Dimensional Case 340 8.8 The Bang-Bang Principle 343 8.9 Invariant Measures & Poincare's Recurrence Theorem 346 8.9.1 Linear Extension of Set-Valued Maps 346 8.9.2 Invariant Measures 350
xvii 9 Selections and Parametrization 353 9.1 Case of lower semicontinuous maps 355 9.2 Case of upper semicontinuous maps 358 9.3 Minimal Selection 360 9.4 The Steiner Selection 364 9.4.1 Steiner Points of Convex Compact Sets .... 365 9.4.2 The Intersection Lemma 369 9.4.3 Lipschitz Selections of Lipschitz Maps 372 9.5 Selections of Caratheodory maps - 373 9.6 Caratheodory Parametrization 376 9.7 Measurable/Lipschitz Parametrization 379 10 Differential Inclusions 383 10.1 The Viability Theorem 387 10.1.1 Solutions to Differential Inclusions 388 10.1.2 Statements of the Viability Theorems 389 10.1.3 Viability Kernels 392 10.1.4 Viability and Equilibria 393 10.2 Applications of the Viability Theorem 394 10.2.1 Linear Differential Inclusions 395 10.2.2 Lyapunov Functions 395 10.2.3 Tracking a Differential Inclusion 398 10.3 Nonlinear Semi-Groups 399 10.4 Filippov's Theorem 400 10.5 Derivatives of the Solution Map 403 Bibliographical Comments 411 Bibliography 421 Index 457
XIX List of Figures 1.1 Example of Upper and Lower Limits of Sets 18 1.2 Graph of a Set-Valued Map and of its Inverse 35 1.3 Semicontinuous and Noncontinuous Maps 40 3.1 Example of Monotone and Maximal Monotone Maps . 104 4.1 Contingent Cone at a Boundary Point may be the Whole Space 122 4.2 The Graph of T[afi](·) 123 4.3 Counterexample: Tangent Cone to the Intersection . . 142 4.4 The Menagerie of Tangent Cones: they may be all different 161 5.1 Graph of the Contingent Derivative 183 6.1 Epigraph of the Contingent Derivative 227 9.1 Approximate Selection of an Upper Semicontinuous Map359 9.2 Illustration of the proof 370 List of Tables 2.1 Properties of Support Functions 66 4.1 Properties of Contingent Cones 124 4.2 Properties of Adjacent Tangent Cones 129 4.3 Properties of Tangent Cones to Convex Sets 141 4.4 % Properties of Contingent Cones to Derivable Sets in Finite Dimensional Spaces 152 4.5 Properties of Normal Cones In Finite Dimensional Vector Spaces 158 4.6 Properties of mi/l-order Contingent Sets 172 4.7 Properties of mi/l-order Adjacent Sets 174
It is a fact that in mathematical sciences there has been a reluctance to deal with sequences of sets and set-valued maps. Despite the emergence of exciting new vistas for the applications of mathematics, our long familiarity with sequences (of elements) and with (single-valued) maps has perhaps been so deeply rooted in traditional mathematical conceptualizations that it has appeared easier to sacrifice the breadth of some problems or some simple underlying structure in order to avoid set-valued maps. For this reason, we begin by providing examples of natural and/or general problems involving set-valued maps before giving a rough description of the results presented in the pages that follow. Examples of Set-Valued Maps 1. First, we encounter set-valued maps each time we face ill-posed problems or inverse problems, i.e., problems for which either the existence of a solution or its uniqueness is not guaranteed for some data: Set-valued maps allow us to get away from the restriction that a map is bijective when we want to solve an equation. Indeed, the first natural instance when set-valued maps occur is the inverse /-1 of a single-valued map / from X to Y. We always can define /-1 as a set-valued map which associates with any data у the (possibly empty) set of solutions Г Чу) ■= {x€X\f(x)=y} to the equation f(x) = y. Of the three commandments of Hadamard's tablets, existence, uniqueness and stability, we shall only retain the stability requirements, which can be encapsulated in adequate definitions of continuity of jf"1: This is one of the topics of the first chapter.
2 Introduction 2. Taking into account uncertainties, disturbances, modeling errors, etc., leads naturally to set-valued maps and inclusions. They also arise when we wish to treat a problem qualitatively, by looking for solutions common to a set of data, sharing the same (qualitative) properties. Set-valued analysis should play an important role in the new field of qualitative physics, a rapidly growing branch of Artificial Intelligence. ■ 3. Problems with constraints also yield specific set-valued maps: Solving the equation f(x) = y, where the solution χ is required to belong to a subset K, amounts to solving the inclusion /|к(%) = У where f\x, the restriction of / to K, is regarded as the set-valued map associating with χ the point f(x) when χ £ К and the empty set when χ is not in K. 4. Unilateral problems in mechanics were formulated in the framework of variational inequalities (also called "generalized equations" by some authors), which are again inclusions in disguise. Their solution by Stampacchia and J.-L. Lions in the sixties gave a new impetus to set-valued maps, with a different vocabulary. 5. Set-valued maps provide a useful framework for control theory, since the early contributions of Wazewski and Filippov in the beginning of the sixties. Such set-valued maps, called parametrized maps, are associated with a family of maps χ ι-> f(x, и) from X to Υ when и ranges over a set U(x) of parameters. The (single-valued) map / describes the dynamics of the system: It associates with the state χ of the system and the control и the velocity f(x, u) of the system. The set-valued map U describes a feedback map assigning to the state χ the subset U(x) of admissible controls. Hence the map F which associates with each state χ the subset F(x) of feasible velocities is defined by: F(x) := f(x,U(x)) = {f(x,u)}ueU{x)
Examples of Set-Valued Maps 3 So, the control system governed by the family of parametrized differential equations x'(t) = f(x(t),u(t)) where u(t)eU(x(t)) is actually governed by the differential inclusion x'(t) e F(x(t)) 6. Optimization provides examples of problems where uniqueness of the solution is naturally lacking: Let ЖЬеа function from Χ χ Υ to R. We consider the family of minimization problems VyGY; V(y) := ЫМх,у) parametrized by parameters y. The function V is called the marginal (or performance or value) function. For every у G Y, let G(y) := {x€X\W(x,y) = V(y)} be the subset of solutions to our minimization problem. One of the main issues of optimization theory is to study the set-valued map G (nonvacuity, continuity and differentiability in a suitable sense, and so on.) We shall call G the marginal map. It is no wonder that game theory and mathematical economics use set-valued maps in a natural way. 7. Another source of strong motivations came from optimization and mathematical programming, when necessary conditions (the Fermat rule1, stating that the derivative of a function vanishes at points where it achieves an extremum) were needed to replace optimization problems by the resolution of equations. 1 "Je desire seulement qu'il [Descartes] sache que nos questions de Maximis et Minimis et de Tangentibus linearum curvarum sont parfaites depuis huit ou dix ans et que plusieurs personnes qui les ont vues depuis cinq ou six ans le peuvent temoigner", Fermat wrote in 1638 when Descartes accused that the Methodus de Maxima et Minima was just due to luck and trial and error! ("a tatons et par rencontre".)
4 Introduction The Fermat rule is indeed one idea which was revisited and enhanced again and again under different names: — The Euler-Lagrange equations, when dealing with problems of the calculus of variations; — Lagrange and Kuhn-Tucker multipliers, when state constraints were added to optimization problems; — The Pontriagin principle when dealing with optimal control- problems. After the advances of Functional Analysis, it was time to uncover the common fact behind all these results. It is still and always the Fermat rule, provided we are able to "differentiate" larger and larger classes of functions beyond differentiable functions. The crucial revolution in the history of the concept of gradients happened in the sixties when J.- J. Moreau and R. T. Rock- afellar proposed in the framework of convex analysis the notion of subdifferential of a convex function, which is no longer an element, but a set of "subgradients". 8. The use of set-valued maps in mathematical economics and game theory started when von Neumann asked for an extension of the Brouwer Fixed Point Theorem to set-valued maps, which was needed for finding noncooperative equilibria for n- person games, for instance. This was achieved with the famous Kakutani Fixed-Point Theorem, in the forties. It has been used by Arrow and Debreu in the early fifties to provide the long-expected proof of the existence of a Walrasian equilibrium price. While this achievement made set-valued maps popular among mathematical economists, it was not until the challenges raised by optimization, control theory and unilateral problems in mechanics at the beginning of the sixties that renewed motivations arose to study set-valued maps, as an important subject in its own rite. This was the time when Zarantonello introduced monotone maps, which cover many important nonlinear single-valued or
Properties of Set-Valued Maps 5 set-valued maps of the Calculus of Variations. Properties of Set-Valued Maps Having briefly indicated the importance of set-valued maps in a wide spectrum of applications and of fundamental mathematics, we paint a broad picture of their properties. We follow, in doing so, the outline of the book. о Limits and Continuity Limits of sets were introduced by Painleve in the first years of this century, just after Prechet axiomatized in 1906 the concept of ^-spaces (on which a notion of limit is defined2.) Studying limits of sets together with limits of elements may have been very natural in this context. The topological ideas are, indeed, quite simple and straightforward. In the same way that topological concepts are based on the notions of limits and cluster points of sequences of elements, their set-valued analogues are rooted in the concepts of lower and upper limits of sequences of sets, which are, so to speak, "thick" limits and cluster points respectively: The lower limit of a sequence of subsets Kn is the set of limits of sequences of elements xn G Kn, and the upper limit is the set of cluster points of such sequences. We mentioned already that stability is the only requirement that we retain to study ill posed or inverse problems. Stability is a catch word which means that the set of solutions depends continuously upon the data. How can we proceed to define continuity of set-valued maps? If we try to adapt to the set-valued case the two equivalent definitions of continuity of single-valued maps, we obtain two notions which are no longer equivalent! 2In his famous thesis, judged at the time far too much abstract; it was published in the Rendiconti Cir. Mat. di Palermo! Lebesgue wrote: "....set theory was placed outside the pale of mathematics by the high priests of analytic functions... Set theory, which developed from the theory of analytical functions, could prove useful to its elder sister and could show people of good will its qualities and richness".
Introduction This unfortunate situation led to two concepts of semiconti- nuity of set-valued maps, introduced at the beginning of the thirties by Bouligand and Kuratowski: Lower and upper semi- continuity. These issues were developed in the monographs of Hausdorff and Kuratowski and thoroughly investigated at this time by many (but not exclusively) Polish and French mathematicians. For some reason, just a while later, set-valued maps yielded the way to single-valued maps: A set-valued map was viewed at the time as a single-valued map from a set to the power set of another set. However, as it turned out, the structures exported to power sets were too poor, and specific information was indeed wasted by doing so. For instance, when we regard a set-valued map as a single- valued map from one set to the power set of the other (supplied with any one of the topologies we can think of), we arrive at continuity concepts which are stronger than both lower and upper semicontinuity, introducing parasitic artifacts. For example, using such topologies to differentiate set-valued maps, leads to such strong requirements, that most set-valued maps would become nondifferentiable. This is the reason why we shall start this book with the study of limits and leave aside the examination of topologies on power sets. Furthermore, we shall renew history, by regarding a map not ... as a map, but as a graph (a subset of the product of the departure and the arrival sets), reestablishing some symmetry by putting these two sets on the same footing. This brings us back to the source of analytical geometry, at the time of F. Viete, P. de Fermat and R. Descartes, before the concept of function and map evolved from the one of curves and graphs. To regard a map as a graph is our constant and basic point of view throughout this book (which has been called the graphical approach.) For instance, closed maps, that is maps with closed graph, shall play a starring role in this book. It is a weaker property than
Properties of Set-Valued Maps 7 continuity or even, upper semicontinuity, very familiar and thus easy to check, common to both set-valued maps and their inverses. © Linear Functional Analysis What is the set-valued version of a continuous linear operator? Remembering that the graph of a continuous linear operator is a closed vector subspace, we are tempted to single out the maps whose graphs are closed linear subspaces (called linear processes.) This generalization is not bold enough, since dealing exclusively with closed vector subspaces is still too restrictive: We need to use the notion of closed convex cone, which is a kind of vector subspace in which it is forbidden to use subtraction. These cones enjoy many properties of the vector subspaces. For this reason, we select the closed convex processes, i.e., the maps whose graphs are closed convex cones, as the candidates to play the part of set-valued linear maps. We shall see later that derivatives of some set-valued maps are closed convex processes, which is a desirable property for a derivative3. Indeed, the two basic theorems on continuous linear operators due to Banach, the Closed Graph Theorem (equivalent to the Open Mapping Principle) and the Banach- Steinhaus Theorem, can be adapted to closed convex processes. The first one states that a closed convex process defined on the whole space is continuous, and the second states that pointwise bounded families of closed convex processes are bounded — a prerequisite for studying the convergence of closed convex processes. But most important of all, one can transpose closed convex processes and use the benefits of duality theory, based on the Bipolar Theorem. Closed convex processes also possess eigenvectors and invariant spaces and cones. They truly deserve the status of linear set- valued maps. 3keeping us in line with Hadamard's linearity bill enforced by Frechet that a derivative should be linear with respect to the increment.
8 Introduction © Nonlinear Functional Analysis We are convinced that many problems can be regarded as inclusions given F : Χ -^ Υ and у €Y, find χ € X such that F(x) Э у Most theorems on existence of solutions to nonlinear equations can be extended to the case of inclusions. For example, this is the case for the Brouwer Fixed Point Theorem, whose generalization to set-valued maps is the famous Kakutani Fixed-Point Theorem. We shall prove an equivalent statement, called the Equilibrium Theorem, which provides the existence of an equilibrium of a set-valued map, a solution to the inclusion F(x) Э 0. Of course, for applications, we need not only to solve such a problem, but also to approximate its solutions by solutions to approximate problems: given Fn : Xn -^ Yn, yn € Yn, find xn € Xn with Fn(xn) Э yn where Xn and Yn are subspaces of X and Y. The famous Lax's principle of numerical analysis states that convergence of the data yn to y, consistency of Fn to F and stability of the Fn's imply the convergence of approximate solutions. As an important special case, when the spaces Xn = X, Yn = Y and the maps Fn = F are constant, this principle boils down to the statement of the Inverse-Function Theorem. Convergence, consistency and stability, which were originally defined in the framework of linear equations, can be extended to this general case by introducing adequate notions of convergence and derivatives of set-valued maps. For instance, the concept of consistency is nothing other than the fact that the graph of F is the lower limit of the graphs of the approximate maps Fn, while stability is the boundedness of the inverses of the derivatives of the maps Fn. This provides a first motivation for devising a set-valued differential calculus. For that purpose, we need to begin with the study of tangents.
Properties of Set-Valued Maps 9 β Tangents and Normals The concept of tangency has been overshadowed in some sense by the requirement that the space of tangent vectors must be a vector space, so that the original idea became concealed after its formal implementation in differential geometry. If we come back to the idea underlying the notion of tangency to a subset К at some point χ G K, we are tempted to form "thick" differential quotients K-x h and to take (in various ways) their limits when h > 0 goes to 0. We obtain in this way a variety of closed cones made of what we call tangent vectors. The most popular of these tangent cones is for the time the contingent cone introduced in the thirties by Bouligand, (which is the upper limit of these differential quotients.) Some of these tangent cones are closed convex cones, and they enjoy a property which is the natural extension of linearity (without subtraction.) These tangent cones possess a rich calculus which justifies their use in many questions, mainly in problems with state con- ' straints: They are involved in the sufficient conditions for the existence of an equilibrium and for the stability of solutions to equations with constraints. They also appear in the formulation of necessary conditions in optimization problems with constraints and play a key role in viability theory. In order to define space of normals, which in differential geometry consists of vectors othogonal to the tangent vector space, we are led to introduce the dual concept of normal cones to any subset. © Differentiation of Maps We already mentioned that the concept of stability in the Inverse-Function Theorem requires the notion of derivative of a set-valued map, leading to the question: How do we formulate this concept?
10 Introduction The idea is very simple and goes back to the prehistory of the differential calculus, when Pierre de Fermat introduced in the first half of the seventeenth century the concept of tangent to the graph of a function: The tangent space to the graph of a function fata point (x, y) of its graph is the line of slope f'(x), i.e., the graph of the linear function и ι-> f'(x)u. It is possible to implement this idea for any set-valued map F, since we have introduced a way to implement the tangency for any subset of a normed space. Therefore, in the framework of a given problem, we can regard a tangent cone to the graph of the set-valued map F at some point (x, y) of its graph as the graph of the associated "derivative" of F at this point (x,y). Derivatives built in this way from the various choices of tangent cones are called graphical derivatives and the calculus of tangent cones can be transferred to a set-valued differential calculus, including chain rules. With such derivatives of set-valued maps in our hands, we can linearize set-valued problems for approximating them by linearized ones. The latter involving closed convex processes, this strategy provides ways for transferring some properties of linear set-valued maps to nonlinear maps. β Gradients of Functions and the Fermat Rule The particular case of real-valued functions deserves a study by itself for taking into account the order relation of real numbers. We are led to do so whenever we look for a minimizer of a function or when we study the monotone behavior of a function along a solution to a differential equation or inclusion (Lyapunov property.) The set-valued approach indicates the route: We associate with a function V the set-valued map Vj defined by VT(s) := [V(x),+oo[ whose graph is the epigraph of V. The graphs of the derivatives of such set-valued maps Vj are the epigraphs of functions which are called epiderivatives. We
Properties of Set- Valued Maps 11 discover that they are close relatives of the directional derivatives introduced by Dini, who was among the first to revolt against the rigidity imposed by the heirs of Cauchy4. As far as optimization is concerned, the Fermat Rule can be extended to any function by using these epiderivatives. Since they enjoy a rich calculus, we obtain in this way many necessary conditions for a minimum. This can be done by transferring the set-valued differential calculus to what can be called an epidifferential calculus. By duality, we associate with each of the epiderivatives a concept of generalized gradient: It is in general a subset of elements, reduced to the usual gradient whenever the function is differentiable in the usual way. In this framework, the Fermat Rule becomes: // a point achieves the minimum of a function, then it is an equilibrium of the generalized gradient, i.e., the generalized gradient at an optimal point contains 0. © Convergence of Maps What about the convergence of a sequence of set-valued maps The first idea which comes to mind is to extend the various notions of uniform convergence of single-valued maps, regarded as a map from one space to another. Since we know how to deal with limits of sets, it is again natural to use the graphical approach and to study the upper and lower limits of graphs. We follow this hint and study graphical upper and lower limits of a sequence of set-valued maps as the maps whose graphs are the upper and lower limits of the graphs. 4Cauchy, however, had the merit to formalize the concept of limits, continuity and differentiability. His definitions have been canonized ever since: A function was allowed to be differentiated only if the differential quotients were converging to the derivative for the pointwise convergence topology. The need to use nondif- ferentiable functions has been felt several times. By Bouligand, with the notions of contingent and paratingent, by Dini who also broke Hadamard's linearity law, by L. Schwartz and S. Sobolev, with the discovery of weak derivatives of functions and distributions. But each of these extensions was devised for specific purposes (solving partial differential equations, for instance.)
12 Introduction When we deal with real-valued functions, we are led to use the set-valued maps νπ| associated with functions Vn, whose graphs are epigraphs of the functions Vn. The graphical limits of the set-valued maps νπ| induce what we call epigmphical limits of the functions Vn. This concept is closely related to G- convergence introduced by de Giorgi and has been extensively used in the study of stability of optimization problems5. An interesting question arises: What are the connections between the (epigraphical) convergence of a sequence of functions and the (graphical) convergence of their gradients? We shall answer such questions. ® Measures and Integration We encounter measurable maps whenever we deal with models of systems having measurable data, and in particular when we deal with random set-valued variables (an issue we shall not address in this book.) Another important instance where measurable set-valued maps do arise is in the linearization of a control system (or a differential inclusion) along a solution. Hence, we cannot escape the burden of studying measurable maps, which are the maps whose graphs are measurable, and checking in particular that all the standard operations preserve measurability. We also need measurability for defining integrals of set-valued maps. Integrals of set-valued maps are involved in many con- vexification (also called relaxation) problems, since roughly speaking the integral of a measurable set-valued map is always convex. This property was in fact the original motivation to introduce the integral of set-valued maps in mathematical economics and game theory (with a continuum of players.) We shall also address the basic questions of ergodic theory, extending to set-valued maps the Poincare Recurrence Theorem 5See the forthcoming book Set-Valued Analysis and Subdifferential Calculus [355] by Rockafellar and Wets.
Properties of Set-Valued Maps 13 and the existence of invariant measures. β Differential Inclusions Control Theory on one hand, and the evolution of macro- systems under uncertainty on the other hand, constitute very strong motivations for extending differential and partial differential equations to differential and partial differential inclusions. We shall provide only an introductory survey of some basic results, since covering this material requires books by themselves6. We shall state some existence theorems, show that the set of solutions depends continuously upon the initial data, describe some properties of Lyapunov functions (which, thanks to the concept of epiderivatives, can even be taken lower semicontinu- ous), relate the graphical derivatives of the solution map to the solution map of variational inclusions (which are linearizations of the differential inclusion along a solution) and state some applications of the Viability Theorem. β Selections and Parametrization We cannot escape in fine answering two natural questions: Can we find selections of set-valued maps inheriting their regularity properties? Are set-valued maps parumetrizablel We shall be able to show that measurable set-valued maps do have measurable selections and that continuous (Caratheodory, Lipschitz) maps do have continuous (Caratheodory, Lipschitz) selections under severe restrictions: The images of the set- valued map must be convex. Actually, a Lipschitz set-valued map F with closed convex images is parametrizable in the sense that there exists a "control space" U and a Lipschitz map / : X x U ι-> Χ such that Vx, F(x) = {f(x,u)}ueu 6See for instances the books Differential Inclusions [33] by Aubin & Cel- lina, Viability Theory [50] by Aubin and Set-Valued Analysis and Control Theory [195] by Frankowska.
14 Introduction Therefore, this limited class of set-valued maps is in essence made of families of single-valued maps.
Chapter 1 Continuity of Set-Valued Maps Introduction This chapter is devoted to the elementary topological properties of sequences of sets and set-valued maps. We begin the first section by extending concepts of limits and cluster points of sequences of elements to sequences of sets. These set-valued analogues have been introduced by Painleve in 1902 under the names of upper and lower limits of sets. They were then popularized by Kuratowski in his famous and influential book Topologie and thus, often Christianized Kuratowski lower and upper limits of sequences of sets1. We shall call them simply lower and upper limits: The lower limit of a sequence of subsets Kn is the set of limits of sequences of elements xn G Kn and the upper limit is the set of cluster points of such sequences. Some elementary properties of lower and upper limits are investigated in Section 1, whereas their calculus is presented in Section 2. Set-valued maps can be regarded as "continuous sequences." But before adapting the above concepts to this case in Section 4, we gather in Section 3 the main definitions concerning set-valued maps. lHad they been called Painlevd instead of Kuratowski, a easier name to pronounce indeed, these concepts could have been more popular.... This at least is the opinion of one of the authors. 15
16 1- Continuity of Set-Valued Maps Upper and lower limits of set-valued maps are naturally closely related to continuity concepts. Recall that for single-valued maps, continuity amounts to mapping converging sequences io converging sequences or, equivalently, to the celebrated incantation: Ve>0, 3£>0,.... Unfortunately, this characterization is no longer true for set- valued maps: There are two distinct ways to extend the concept of continuity: The first one, encompassing the above "convergence" idea, is called lower semicontinuity. The second, extending the "V e, 3 δ — definition," leads to the so-called upper semicontinuity of set-valued maps2. Kuratowski showed however that semicontinuous maps are generically continuous. Lower semicontinuous maps are closely related to lower limits, but the situation is further complicated by the fact that upper limits characterize the maps with closed graph and are not always upper semicontinuous maps. These subtleties are explained in Section 4. It happens that set-valued maps with closed graph taking their values in a compact set are upper semicontinuous. Hence we have an easy way to verify upper semicontinuity of a map. Lower semicontinuity is not as simple to check: the last section is devoted to several useful lower semicontinuity criteria. The definitions introduced in this chapter are basic and will be used throughout the book. 1.1 Limits of Sets 1.1.1 Definitions Let X be a metric space supplied with a distance d. When К is a subset of X, we denote by άχ(χ) := d(x,K) := inf d(x,y) 2These concepts should not be confused with the upper and lower semicontinuity of real-valued functions.
1.1. Limits of Sets 17 the distance from χ to K, where we set d(x,ty) := +00. The ball of radius r > 0 around К in X is denoted by Bx(K,r) := {x€X \ d(x,K)<r} When there is no ambiguity, we set B(K,r) := Bx(K,r) When X is a Banach space whose unit ball is denoted by В (or Β χ if the space must be mentioned), we observe that Bx(K,r) = K + rBx The balls B(K, r) are neighborhoods of K. When К is compact, each neighborhood of К contains such a ball around K. Limits of sets have been introduced by Painleve3 in 1902, as it is reported by his student Zoretti4. They have been popularized by Kuratowski in his famous book Topologie and thus, often called Kuratowski lower and upper limits of sequences of sets. ' Definition 1.1.1 Let (-Kn)neN be a sequence of subsets of a metric space X. We say that the subset LimsuPj^ooATn := ix&X | liminf d(x,Kn) = 0 \ is the upper limit of the sequence Kn and that the subset LiminfTl_too.Kn := {x € X \ limn^00d(x, Kn) = 0} is its lower limit. A subset К is said to be the limit or the set limit of the sequence Kn if К = Limmfn-^ooUTn = LimsuPn^^Kn =: Ь1тп_юоЛГп 3"Je dirai qu'un point a appartient a I'ensemble limite de Ea si, quelque petits que soient les deux nombres r et e, sous la condition \a — Qo| < e, le cercle de centre α et de rayon r renferme des points appartenant a certains des Ea." who wrote: "Je ne veux pas terminer ce travail sans adresser tous mes remer- ciements a tous mes maitres de 1'Ecole Normale dont la sollicitude a mon egard ne s'est jamais ralentie, et dont les encouragements m'ont ete si precieux. D'ailleurs, dans ce milieu normalien, tout ce qui vous entoure est un enseignement, ...."
18 1- Continuity of Set-Valued Maps Figure 1.1: Example of Upper and Lower Limits of Sets . ff»:={£}x[0,l] if η is even υ Я»:=фх[-1,0] if η is odd и LimimV+oo-Κπ = {0} & Limswpn^^Kn = [-1 .+1] Lower and upper limits are obviously closed. We also see at once that Liminfn^ooUTn с Limsup^oo-Kn and that the upper limits and lower limits of the subsets Kn and of their closures Kn do coincide, since d(x,Kn) = d(x,Kn). Any decreasing sequence of subsets Kn has a limit, which is the intersection of their closures: if Kn с Km when η >m, then Limn-юоКп = Q Kn 7l>0 An upper limit may be empty (no subsequence of elements xn G Kn has a cluster point.) Concerning sequences of singleta {xn}, the set limit, when it exists, is either empty (the sequence of elements xn is not converging), or is a singleton made of the limit of the sequence. It is easy to check that: Proposition 1.1.2 If (Kn)ne^ is α sequence of subsets of α metric space, then Linimfn->oo-Kn is the set of limits of sequences xn G Kn and LimsuPj^oQ-ftTn is the set of cluster points of sequences xn G Kn,
1.1. Limits of Sets 19 i.e., of limits of subsequences xn/ G Kn/. The upper limit is also equal to the subset of cluster points of "approximate" sequences satisfying: V e > О, Э Ν(ε) such that V η > Ν(ε), χη € B(Kn, e) Remark — Replacing the balls of a metric space by neighborhoods and the sequences of a metric space by generalized sequences, we can extend the concepts of upper and lower limits to generalized sequences of subsets of a topological space X. We recall that a set Μ supplied with a preorder У is directed if every finite subset has an upper bound. A generalized sequence is a map μ G Μ t-> χμ G X An element χ G X is the limit of (χμ)μ£Μ if; f°r every neighborhood V of x: there exists μο G Μ such that χμ belongs to V for all μ У μ0. An element χ is said to be a cluster point of this generalized sequence if, for every neighborhood V of a; and every μ G M, there exists ν У_ μ such that xv G V. We recall that a single-valued map / from a topological space X to another Υ is continuous at χ if it maps any generalized sequence converging to a; to a generalized sequence converging to fix) and that a subset К С X is compact if and only if every generalized sequence of elements of К has a cluster point. (See for instance [148, Section 1.7].) The main difference with sequences of a metric space is that a converging generalized sequence is not necessarily bounded. (We can also replace generalized sequences by "filters.") - The upper limit of a generalized sequence of subsets Κμ where μ ranges over a directed subset M. is the set of cluster points of generalized sequences χμ of elements of Κμ and the lower limit is the set of limits of such generalized sequences. We may need the above extension when dealing with weak topologies of a Banach space X and of its dual denoted by X*. We say that the bilinear map < ·, · > (i>,x) G X* x X ь-+< p,x > := p(x) is the duality pairing. We recall that the weakened topology σ(Χ, Χ*) of X is defined by the semi-norms Pm [x) ■= sup | < q, χ > | q&M when Μ := {qi,... ,qi} ranges over the finite subsets of X*. The weak —*· topology σ{Χ*ι Χ) of the dual X* is defined by the semi-norms pN{q) := sup | <q,y>\ yeN
20 1 - Continuity of Set-Valued Maps when N := {x\,..., xn} ranges over the finite subsets of X. In other words, a generalized sequence of elements χμ e X converges weakly_to χ e X if and only if for any q e X*,'< ς,χμ > converges to < q, χ > and a sequence ρμ e X* converges weakly to ρ if and only if for any у e -X", < ρμ, у > converges to < p, у >. The key facts that can be recalled now are that X* is still the dual of X supplied with the weakened topology, that X is the dual of X* supplied with the weak-* topology, and that the bounded subsets of the dual X* are weakly relatively compact. However, in general X is a closed subspace of the bidual X** of X, which is the dual of the Banach space X*, endowed with the norm ||p||* := sup | <p,x> | IMI<i The space X is called reflexive if X = X**. In this case, it enjoys both the properties of a Banach space and of the dual of a Banach space, including the weak compactness of the unit ball. Π We could for instance use the fact that the weakly compact subsets of the dual X* of a separable5 Banach space are metrizable (for the weak-* topology) (see [I48, Theorem 5.6.3]) to avoid generalized sequences. Instead, for the sake of simplicity, we shall use the concept of sequentially weak upper limit: Definition 1.1.3 Let us consider a sequence of subsets Kn of the dual of a Banach space. We shall say that the subset σ - Limsup^eoATn of weak-* limits of subsequences of elements xn G Kn is the sequentially weak upper limit of the subsets Kn. In this way, we can present lower and upper limits in the framework of metric spaces or of (countable) sequences, leaving to the interested reader the task of checking some extensions to the case of non metrizable topological spaces. 5This means that there is countable basis of open subsets, or, equivalently, that there exists a countable basis spanning a dense vector space. The separability concept goes back to Frechet.
1.1. Limits of Sets 21 We also point out the quite impressive equivalent formulation of upper and lower limits which follows from Proposition 1.1.2: Limsup^oolfn = Π U Kn = Π Π U Β(Κη,ε) Ν>0η>Ν ε>0Ν>0η>Ν and Limmin^Kn = Π U Π Β(Κη,ε) ε>0Ν>0η>Ν and single out the following property of upper limits: Theorem 1.1.4 Let К be a subset of a metric space X satisfying the following property: for any neighborhood U of K, 3 N such that V η > Ν, Kn С U Then Limsupjj tooKn С К. Conversely, if X is compact, then the upper limit lAmsw£>n^OQKn enjoys the above property (and thus, is the smallest closed subset satisfying it J Proof — The first statement is obvious. The second one is a consequence of the following more general result: Proposition 1.1.5 Let us consider sequences of subsets Ln and Mn of a metric space and assume that there exists a compact subset Μ satisfying the following property: for any neighborhood W of M, 3 N such that V η > iV, Mn С W Then, for any neighborhood U of Μ Π (LiiaswPn^^Ln), there exists an integer N such that Ln Π Mn С U whenever n> N. Proof — If the neighborhood U contains M, the result follows from the assumption on M. Otherwise, by taking an open neighborhood U, the subset К := M\U is not empty, disjoint of Limsupj^ooLn and is compact by assumption. Let у belong to K. Since у does not belong to LimsuPj^ooLn, there exist ey > 0 and Ny such that, for all η > Ny,y does not belong to B(Ln,ey). The subset К being compact, it can be covered by ρ
22 1 - Continuity of Set-Valued Maps balls B(yi,eyi). This implies that for all η > No := maxj^i,...^iV^ and. . .-: V := \jB(yi,eyi) the intersections Lni~)V are empty. On the,other hand, W := U U V being a neighborhood of M, we deduce from the assumption that there exists JVi such that Vn>JVi, M„ С WUV Therefore LnnMncU for all η > max(iV0) iVi). □ Remark — If Μ is not compact, but just closed, the conclusion of the proposition remains true for any neighborhood U of Μ Π (LimsupTl_HX1LTl) whose complement in Μ is compact. □ We also provide a useful technical lemma: Lemma 1.1.6 Let us consider α sequence of subsets Ln С Ζ of α metric space Ζ and a sequence of subsets Mn С Υ of a compact metric space Y. Let ψ : Ζ χ Υ t-> R be an upper semicontinuous function. We denote by M$ the upper limit of the Mn and Lb the lower limit of the Ln. Then umsuPn-^oo ( SUP inf <P{z,y)) < SUP ίηίμ <Р(*,У) \уемп zeb" / Уеми zeu· Proof— Let у belong to M". Since φ is upper semicontinuous, we know that for any ε > 0 and any ζ £ Lb, there exist neighborhoods N(z) of ζ and N(y) of у such that Уг'ёВД, Vy'eiV(y), φμ,ν') < <p{z,y) + e/2 In particular, by taking zy £ Lb such that <f{Zy,y) < inf tp{z,y)+e/2 zEL* and approximating zy by elements z™ £ Ln, we infer that there exists Ny > 0 satisfying Vn>iVy, \fy'£N(y), φ{ζ",υ') < <p(zy,y) + e/2 and thus, that for any у £ Μ", V η > Ny, V у' £ N {у), inf φ(ζ, у') < inf p(z, у) + ε zELn zELb
1.1. Limits of Sets 23 On the other hand, the compact set M" can be covered by η neighborhoods N(yi) so that, by Theorem 1.1.4, there exists an integer NQ such that Vn>JV0, Mn С (J N(yi) i=l,...,n Set N :=■ maxi=o,...,n Nyi. Then, for all η > N and у € Mn, у belongs to some N(yi), so that, inf <p{z,y) < inf ip(z,yi)+e < sup inf ¥>(z,2/)+e ^ebn zgbb Уем> zeib Hence we have proved that for any ε > 0, there exists Ν > Ό such that \/n>N, sup inf ip(z,y) < sup inf </?(z, ί/) + ε □ уемп zeL" уеми геьь 1.1.2 The Compactness Theorem The Bolzano-Weierstrass Compactness Theorem was adapted in 1927 to the set-valued framework: Theorem 1.1.7 (Zarankiewicz) Every sequence of subsets Kn of a separable metric space X contains a subsequence which has a (possibly empty) limit. Proof — Since X is separable, there exists a countable family of open subsets Um satisfying the following property: V open subset U, V χ G U, 3 Um such that χ e Um С U Let us consider a sequence of subsets Kn. We shall construct a sequence of subsequences (K„ )n>o by induction. For τη = 0, we set Kn := Kn. Assume that the m — 1 first subsequences , 0 < ρ < m — 1 have been constructed. Consider the mth open subset Um. Then either for every subsequence Tlj, Um П (Limsup^i^-1)) φ 0 in which case we set Kj := Kj , or there exists a subsequence rij such that Um П (Limsup^i^-1)) = 0
24 1 - Continuity of Set-Valued Maps in which case we set Kj := Кщ '. (The choice of such a subsequence does not matter.) These sequences [Kn ) being constructed, we extract the \ / n>0 diagonal subsequence Dn := Kn . We claim that it has a set limit. If not, there would exist x0 € Limsup^oo-Dn h x0 φ LiminfTl_too.DTl The latter condition means that there exists an open neighborhood U of xq and a subsequence Dn. such that U Π Dnj = 0 for any j. Let us fix an open subset Um such that xq € Um С U. We thus deduce that Um Π (himswpj^^Dn.) = 0 Since for rij > m, Dnj := Кщ' = Щт~ ' for some pj, we observe that Dn, is a subsequence of the sequence [Knm~~ ') , the upper 3 \ In>0 limit of which is disjoint from Um. By the very construction of [Kn ) , we infer that \ / n>0 к^ = лг^-1) and consequently, that Um П (Limsup^i^H) = Um П (Limsup^i^-1)) = 0 Since Dn := Knn' = Kp^' for some pn, we deduce that the sequence (Dn)n>m is a subsequence of the sequence (Ют ). . Thus xq € Limsup^oo-Dn С Limsupj^^Kj С X \ Um which contradicts the fact that xq belongs to Um- n 1.1.3 The Duality Theorem For closed convex cones Kn, upper limits and lower limits can be exchanged by duality.
1.1. Limits of Sets 25 We introduce the (negative) polar cones to subsets К с X and L С X* defined by K~ := {pGX* \ V χ G Κ, < ρ, χ > < 0} ала L~ := {ж € A" | Vp€L, <p,x>< 0} Let σ — LimsuPj^ooAT" denote the sequentially weak upper limit of the polar cones K~. Theorem 1.1.8 Let (Kn)ne^ be a sequence of closed convex cones of a Banach space X. Then Limmfn^ooATn = (σ - Limsupn_+00iir~)~ Proof — Inclusion Liminfn-юоЛГп С (σ -Limsup^^Ar")"" is obvious: If χ G Liminin^coKn is the limit of a sequence of elements xn G Kn and ρ G σ - Limsupn^oolf" is the weak-* limit of a subsequence pni G K~,, inequalities < ρηι, xn> > < 0 imply that < p,x >< 0. Conversely, assume that some χ G (a-Limsup^^UT-)- does not belong to the lower limit Limmin^ooKn. Then there exists e > 0 and a subsequence (again denoted (ΑΓπ)πεΝ) such that V π > 0, (χ + ε Β) Π Κη = 0 The Separation Theorem implies the existence of elements pn G X* of norm equal to one such that σκη(ρη) <<Pn,x> -ε||ρη|| = < Ρη,χ > -ε The subsets Kn being cones, we deduce that pn G K~ and that σκη(Ρη) = 0. Then a subsequence (again denoted) pn converges weakly—* to some p, which thus belongs to σ — lAmsn^^^K^, so that < pyx >< 0. Inequalities 0 < (ρη,Χη) -ε imply that e < 0, a contradiction. □
26 1 - Continuity of Set-Valued Maps _ 1.1.4 Convex Hull of Limits Since the distance function to a subset of a normed space is convex if and only if this subset is convex, we infer that the lower limit of а sequence of convex subsets is closed and convex. It is useful to have a characterization of the closed convex hull of upper limit. We denote by co(K) the convex hull oiK and by co(K) its closed convex hull. Lemma 1.1.9 Let us consider a sequence of subsets Kn contained in a bounded subset of a finite dimensional vector space X. Then со (Limsup^^-Kn) = f] со MJ Kn N>0 \n>N Proof — The closed convex hull of the upper limit is obviously contained in the closed convex subset A := We have to prove that it is equal to it when the dimension of X is finite and the subsets Kn are contained in a bounded set. Since an element ж of Л is the limit of a subsequence of convex combinations ν ν of elements of (Jn>N Kn and since the dimension of X is an integer p, Caratheodory's Theorem allows us to write that ρ ΣΝ 3=0 where ρ Nj > N, of > 0, Σα? = 1 j=o and a;^ belong to K^y The vector aN of p + 1 components a? contains a subsequence (again denoted by) aN. which converges to some nonnegative vector a oi p + 1 components aj such that J2^=o аз = 1- The subsets Kn being contained in a given compact subset, we can extract successively subsequences (again denoted by) х$[. converging to elements Xj, which belong to the upper limit of the subsets Kn. Hence χ is equal to the convex combination YTj=q ajXj, and the lemma is proved. □
1.2. Calculus of Limits 27 1.2 Calculus of Limits We begin by pointing out the following obvious properties: Proposition 1.2.1 Let Kn, Ln, Кгп, (i = Ι,.,.,ρ) be sequences of subsets of α metric space. Then i) Limsupn^oo(iin П Ln) С Limsupn_>00iin Π Limsupn_>00Ln ii) Liminfn-yooC-Kn HLn) С Liminfn^oouTn П Liminin^ooLn Hi) Limsup^^(Kn U Ln) - Limswpn^coKn U Limsupn_tooLn гг;) Liminfn_>oo (-^n U L„) э Liminfn_>00l£n U Limmfn^ooLr, υ) Limsup^oo EELi ^n С Π?=ι Limsup^^ii^ i) Liminfn-чоо Ш=1 •K'n = Π?=ι Liminfn-»oo^ We need also to relate direct and inverse images of upper and lower limits of a sequence of subsets to the upper and lower limits of their direct and inverse images. We mention now the obvious relations and postpone the proofs of criteria which transform the following inclusions to equalities. Proposition 1.2.2 Let Kn be a sequence of subsets of a metric space X, Mn be a sequence of subsets of a metric space Υ and f : X ι—> У be a (single-valued) continuous map. Then ' i) fCLimsuPn^Kn) с UxDawprtr^OQf(Kn) ii) /(Limmfn^ooATn) с Limmin^oo f(Kn) Hi) Limsupjj^QQ iv) Liminfn-^o/^Mn) С jf"1 (Liminf^ooMn) The question arises of providing converse results under adequate assumptions.
28 1 - Continuity of Set-Valued Maps 1.2.1 Direct Images We begin with the direct images of upper limits: We easily obtain equalities when / is proper. We recall that a continuous single-valued map from a metric space X to a metric space Υ is proper if and only if one of the following equivalent statements If f(xn) converges in Y, then xn has a cluster point or i) f maps closed subsets to closed subsets ii) V compact Μ С Υ, f~1(M) is compact holds true. Proposition 1.2.3 We posit the assumptions of Proposition 1.2.2. Let us assume that f is proper, then /(Limsupn_tooi;!rn) = Ьш1зирп_>00/(ЛГП) Furthermore, if f is surjective, we obtain f1 (LimsuPn^ooMn) = Limsupn_>0o/~1(M„) The proof is a simple consequence of definitions and is omitted. In the case of continuous linear operators, we obtain a more specific criterion using the polar set K°:={p€X* | \/x € К, <р,х><1} of a subset К of X. When X and Υ are Banach spaces and / is a continuous linear operator A G C(X, Y), we obtain the following result: Theorem 1.2.4 LetX and Υ be Banach spaces, (-Kn)neN be a sequence of subsets of X and A G £(X, Y) be a continuous linear operator satisfying 0 € Int (lm(A*) + (J Π κ) (1.1) V Ν>0η>Ν )
1.2. Calculus of Limits 29 — If X is α finite dimensional space, then Limsupn_>0oA(.K'n) = A(UmS}xpn_,00Kn) — If X is reflexive, then σ - UmsnVn^^AiKn) = Α(σ - Limsupn^00Kn) Proof — Observe that the first statement follows from the second one and Proposition 1.2.2 i). Hence we have to prove only the second claim. Let us consider a sequence xn G Kn such that a subsequence of elements A(xn) (again denoted A(xn)) converges weakly to some у in Y. We shall check that (xn)neN has a weak cluster point, by showing that it is weakly bounded, and thus, weakly relatively compact. Assumption (l.l) implies that the set ЩА*) + (J Π ΚΙ Ν>0η>Ν contains a ball around zero of radius 7 > 0. Then any ρ G X*, ||p||* < 7 can be written, ρ := A*q + r, where q G Y*, r € \J f] K°n N>0n>N Therefore, there exists N > 0 such that r € Ππ>;ν·^π ап^ consequently, suVn>N <p,xn> = supn>^(< q, Axn > + < r,xn >) < < 8vpn>N(\\q\\\\Axn\\+a\xpxeKn <r,x>) < swpn>N \\q\\\\Axn\\ + 1 < +00 since the weakly converging sequence Axn is bounded. Then xn remains in a bounded subset, which is weakly relatively compact. Hence it has a cluster point χ which belongs to σ — lAmsxvon^00Kn. Since A is continuous from X to Y, when they are supplied with their weak topologies, Ax is a weak cluster point of the sequence of elements Axn, which converges weakly to у = Ax. Hence σ - lAmsnVn^^AiKn) С Α(σ - Limsupn_tooi;!rn) The opposite inclusion being obvious, the proof ensues. □
30 1- Continuity of Set-Valued Maps 1.2.2 Inverse Images We study next the case of inverse images of limits. We begin first with a simple criterion which holds true when the subsets Mn are convex and the map / is a continuous linear operator. Proposition 1.2.5 Let us consider two Banach spaces X and Υ, a continuous linear operator A G jC(X,Y) and a sequence of convex subsets Mn С Υ. We assume that the "constraint qualification assumption" 3 7 > 0, с > 0 such that yB С сА(Вх) - Mn holds true for large enough η 's. Then Liminfn_tooJ4"1(Mn) = ^"^Liminfn-^oo-Wn) < w Limsup^ooA-^Mn) = ^(Limsup^^Mn) This statement is a consequence of a more precise result: Proposition 1.2.6 Let us consider two Banach spaces X and Υ, a continuous linear operator A G C(X,Y) and two sequences of convex subsets Ln С X and Mn С Υ. Assume that there exist 7 > 0, с > 0, N > 0 such that V η > Ν, jB С A(Ln Π cBx) - Mn Then Liminfn_too(LnnJ4"1(Mn)) = (Liminfn_t0obn)nJ4"1(Liminfn_tooMn) Limsupn_too(LnnA-1(Mn)) D (Limsupn_tooLn)nJ4-1(Liminfn_tooMn) and Limswpn^00(LnnA~1(Mn)) D (Liminfn_tooLn)nA"1(Limsupn_tooMn) Proof — The inclusion Liminfn_too(LnnA-1(Mn)) с (Liminfn_tooLn)nA-1(Liminfn_tooMn)
1.2. Calculus of Limits 31 being obvious, let us prove the other one, by checking that any χ in the intersection Liminfn—y oo Ln is the limit of a sequence of elements xn belonging to Ln such that A(xn) belongs to Mn. We know that χ can be approximated by elements un e Ln and that A(x) can be approximated by elements vn € Mn. Then en := ||A(?xn) -vn\\ converges to 0 and θη := 7/(7 +en) €]0,1[ converges to 1 and satisfies θηεη = (l — #n)7- Therefore, Θη(νη-Α(ηη)) e θηεηΒ = {ΐ-θη)ΊΒ С (l-9n)(A(LnncBx)-Mn) Consequently, there exist elements u'n £ LnD cBx and v'n G Mn such that A {9nun + (1 - 9n)u'n) = θηυη + (1 - θη)ν'η If we set xn := 9nun + (1 - вп)и'п we observe that xn belongs to Ln and that A(xn) belongs to Mn, for these subsets are convex. Furthermore, ||яп-ип|| = (1 - 9n)\\un - u'n\\ converges to 0 since the sequences un and u'n are bounded, the first one because it is convergent to и and the second because it is bounded by с by assumption. Hence xn converges to x. The proof of the second and third statements is analogous and left as an exercise» □ Remark — Actually, Proposition 1.2.5 is equivalent to Proposition 1.2.6: The former used with 1 x A and Ln x Mn implies obviously the latter. □
32 1 - Continuity of Set-Valued Maps We can prove a similar result for non convex subsets and nonlinear maps, which is based on extensions of Graves' Inverse Function Theorem 3.4.2 proved in Chapter 3. It states that if f is α continuous map from an open subset of a Banach space X to a Banach space Y, continuously differentiable at xq, the surjectivity of f'(xo) implies the existence of a constant I > 0 such that, whenever у lies in a neighborhood of f(xo), the equation f(x) = у has a solution satisfying ||s-soil < l\\f(xo)-y\\ .(1.2)" The simplest extension reads: Proposition 1.2.7 Let us assume thatX andY are Banach spaces, that the map f : X ι—> У is continuously differentiable at some point χ of f~l (Liminfn^ooiWn), and that f'(x) is surjective. Then χ belongs to the lower limit LiminfTl_too/_1(MTl), and a similar statement holds true for upper limits. Proof — Set у = f(x) and consider a sequence of elements yn G Mn converging to y. Then (1.2) implies the existence of a solution xn to the equation f(xn) = Уп satisfying inequality \\x — xn\\ < l\\y — yn\\ for η large enough. Hence χ being the limit of xn € /_1(Μπ) belongs to the lower limit of the inverse images of the subsets Mn. The proof of the second statement is analogous. □ We shall generalize the above inverse function theorem in Chapter 3 and thus be able to weaken the surjectivity assumption of f'(x) by assuming a transversality condition bearing on / and the subsets Mn. These conditions involve the concept of contingent cone Тмп(Уп) to a subset Mn at some point yn € Mn, which will be studied in Chapter 4. The contingent cone Тм{у) is defined by ν €= TM(y) <=> limine+ M = 0 /i->o+ h However, this is a natural place to state this important extension despite the technical complexity of this transversality assumption: Theorem 1.2.8 Let X and Υ be two Banach spaces. We consider a sequence of closed subsets Mn С Υ and a map f : Χ ι—► У continuously differentiable at some point χ € jT1 (Liminfn-юоМп)
1.3. Set-Valued Maps 33 Let us assume that there exist constants с > 0 and η > 0 such that V yn € B(f(x),V) П МП1 Бу с с/'(а:)(ВД - ZWn(yn) JTien χ belongs to Limmin->oo f~1(Mn) and a similar statement holds true for upper limits. This result is contained in the following, more general, Theorem 1.2.9 Let X and Υ be two Banach spaces. We consider a sequence of closed subsets Mn С Υ and a map f : Χ η-» Υ dif- ferentiable on a neighborhood U of f~1(Limmin->coMn) (respectively /_1(LimsupTl_HX1MTl)) such that the derivatives f'(x) are uniformly bounded on Ы. Let us assume that there exist constants с > 0 and η > 0 such that for every χ G U, V yn € B(f(x),V) П Mn, By С cf(x)(Bx) - TMn{yn) Then Liminfn-t0o/"1(Mn) = /^(LimhuV^Mn) respectively Limsupn^oo/'H^fn) = /-1(LimsuPn->oo-Mn) We postpone the proof of the last theorem to Chapter 3. 1.3 Set-Valued Maps Sequences of subsets can be regarded as set-valued maps defined on the set N of integers. Naturally, we can replace N by a metric (or even, topological) space X, and sequences of subsets n ~> Kn by set-valued maps χ ~~* F(x) (which associates with a point χ a subset F(x)) and adapt the definition of upper and lower limits to this case, called the continuous case. Before proceeding further, we recall in this section the basic definitions dealing with set-valued maps, also called multifunctions, multivalued functions, point to set maps or correspondences.
34 1- Continuity of Set-Valued Maps Definition 1.3.1 Let X andY be metric spaces. A set-valued map F from X to Υ is characterized by its graph Graph(F), the subset of the product space Χ Χ Υ defined by Graph(-F) := {(x,y) £ X xY \ y£ F(x)} We shall say that F(x) is the image or the value of F at x. A set-valued map is said to be nontrivial if its graph is not empty, i.e., if there exists at least an element χ £ X such that F(x) is not empty. We say that F is strict if all images F(x) are hot empty. The domain of F is the subset of elements χ £ X such that F(x) is not empty: Dom(F) := {x£X \ F(x) φ 0} The image of F is the union of the images (or values) F(x), when χ ranges over X: lm(F) := (J F(x) xeX The inverse F~1 of F is the set-valued map from Υ to X defined by xeF'^iy) <=> y€F(x) <=> (x, y) € Graph(F) The domain of F is thus the image of -F-1 and coincides with the projection of the graph onto the space X and, in a symmetric way, the image of F is equal to the domain of -F"1 and to the projection of the graph of F onto the space Υ. If К is a subset of X, we denote by F\k the restriction of F to K, defined by F\K(x) ■= I F(X) 'lixGK *lK{X) ' \ 0 if x i К Let V be a property of a subset (for instance, closed, convex, etc..) Since we shall emphasize the symmetric interpretation of a set-valued map as a graph (instead of a map from a set to another one), we shall say as a general rule that a set-valued map satisfies property V if and only if its graph satisfies it. For instance, a set-valued map is said to be closed (respectively convex, closed convex, measurable, monotone, maximal monotone)
1.3. Set-Valued Maps 35 Figure 1.2: Graph of a Set-Valued Map and of its Inverse Dom(F) 0 Graph and Domain of a Set-Valued Map Graph and Image of its Inverse
36 1 — Continuity of Set-Valued Maps if and only if its graph is closed (respectively convex, closed convex, measurable, monotone, maximal monotone.) If the images of a set-valued map F are closed, convex, bounded, compact, and so on, we say that F is closed-valued, convex-valued, bounded-valued, compact-valued, and so on. When * denotes an operation on subsets, we use the same notation for the operation on set-valued maps, which is defined by Fi*F2: x ~> Fi(x)*F2(.x) We define in that way Fi Π F2, F\ U F2, F\ + F2 (in vector spaces), etc. Similarly, if α is a map from the subsets of Υ to the subsets of Y, we define a(F) ·. χ -^ a(F(x)) For instance, we shall use F, co(F), etc., to denote the set-valued maps χ ~> F(x), χ ~> co(F(x)), etc. We shall write F С G <i=^ Graph(F) С Graph(G) and say that G is an extension of F. Let us mention the following elementary properties: ' г) Р(КгиК2) = FiKJUFfa) ii) F(KinK2) С ^(ΛΓι)η^(ΛΓ2) Hi) F(X\K) D lm(F)\F(K) iv) КгСК2 =» F{K{) С F(K2) There are two manners to define the inverse image by a set-valued map F of a subset M: ' i) F-\M) := {x | F(x)DM фЩ < ^ ii) F+1(M) := {x \ F(x) С М} The subset F_1(M) is called the inverse image of Μ by F and F+1(M) is called the core of Μ by F.
1.3. Set-Valued Maps 37 They naturally coincide when F is single-valued. We observe at once that F+1(V\M) = XXF-^M) L· F-^YXM) = X\F+1(M) One can conceive as well two dual ways for defining composition products of set-valued maps (which coincide when the maps are single-valued): Definition 1.3.2 Let Χ, Υ, Ζ be metric spaces and G:X ~> Υ L· H:Y ~> Ζ be set-valued maps. 1 — The usual composition product (called simply the product,) Η о G : X ~> Ζ of Η and G at χ is defined by (HoG)(x) := (J Η (у) yeG{x) 2 — The square product HUG : X ~> Ζ of Η and G at χ is defined by {HUG){x) := Π Η (у) yeG{x) Let 1 denote the identity map from one set to itself. We deduce the following formulas ' Graph(tfoG) = (G x l)71(Graph(F)) = (lxtf)(Graph(G)) ( Graph(tfDG) = (G x l)+1(Graph(tf)) as well as formulas which state that the inverse of a product is the product of the inverses (in reverse Order): ' i) {HoG)-\z) = G-\H~\z)) = (G-1 о Я"1) (ζ) ii) (tfDG)"1^) = G+Htf-1^))
38 1- Continuity of Set-Valued Maps We also observe that ' i) χ € (tfDG)"1^) <=*> G(s) С Я"1^) < и) х е (С-^пЯ"1)^) <=► Я"1^) с G(s) and thus, that G{x) = H'\z) «=► χ e (G^Dff-^WniffDG)-1^) Let us also point out the following relations: When Μ is a subset of Z, then ' i) {HoG)~\M) = G-^H-^M)) < ^ и) (ЯоС)+1(М) = С+1(Я+1(М)) 1.4 Continuity of Set-Valued maps The concepts of semi-continuous maps have been introduced in 1932 by G. Bouligand6 and K. Kuratowski7. We begin with the upper semicontinuous set-valued maps: 1.4.1 Definitions In this section, Χ, Υ, Ζ denote metric spaces. We describe the concepts of semicontinuous set-valued maps introduced by Bouligand, Kuratowski and Wilson in the early thirties. Definition 1.4.1 A set-valued map F : Χ ~> Υ is called upper semicontinuous at χ £ Dom(F) if and only if for any neighborhood U of F{x), 3 η > 0 such that V x' € Βχ(χ,η), F(x') С U. 6who wrote: "Peut-on rendre plus profond hommage a la memoire de Rene Baire qu'en poursuivant les consequences d'une idee degagee par lui et dont l'importance se revele chaque jour accrue, la semi-continuite? Elle echappa tout le XIXе siecle aux adeptes de la theorie des fonctions, et a plus forte raison, aux purs geometres, qui s'adonnaient a des occupations moins subtiles." 7 who also wrote: "D'apres Monsieur Baire, une fontion est dite semi-continue superieurement La notion de semi-continuite dont nous allons nous servir ici est tout a fait analogue a celle-ci, mais concerne le cas ou la fonction F(x) admet comme valeurs des sous-ensembles fermes...."
1.4. Continuity of Set-Valued maps 39 It is said to be upper semicontinuous if and only if it is upper semi- continuous at any point of Dom(F). When F(x) is compact, F is upper semicontinuous at χ if and only if Ve > 0, 3 η > 0 such that Vs' G Βχ(χ,η), F(x') с BY(F(x),e) We observe that this definition is a natural adaptation of the definition of a continuous single-valued map. Why then do we use the adjective upper semicontinuous instead of continuous? One of the reasons is that the celebrated characterization of continuous maps — a map f is continuous at χ if and only if it maps sequences converging to χ to sequences converging to f(x) — does not hold true any longer in the set-valued case. Indeed, the set-valued version of this characterization leads to the following definition. Definition 1.4.2 A set-valued map F : X ~> У is called lower semi- continuous at χ £ Dom(.F) if and only if for any у G F(x) and for any sequence of elements xn G Dom(F) converging to x, there exists a sequence of elements yn G F(xn) converging to y. It is said to be lower semicontinuous if it is lower semicontinuous at every point χ G Dom(F). Actually, as in the single-valued case, this definition is equivalent to: For any open subset U <ZY such that U П F(x) φ 0, 3 η > 0 such that V χ' G Βχ(χ, η), F(x') Π U φ 0 Unfortunately, there exist set-valued maps which enjoy one property without satisfying the other. Examples — The set-valued map i<\ : R -^> R defined by *Л*) - \ {0} if χ = 0 · is lower semicontinuous at zero but not upper semicontinuous at zero. The set-valued map F2 : R ~> R defined by Fo(x) - I {0} if X ф ° Ы*)- | μ1)+1] if x = 0
40 1- Continuity of Set-Valued Maps Figure 1.3: Semicontinuous and Noncontinuous Maps . 0 , lower s.c. h not upper s.c. . 0 upper s.c. & not lower s.c. is upper semicontinuous at zero but not lower semicontinuous at zero.' □ We are therefore led to introduce still another Definition 1.4.3 We shall say that set-valued map F is continuous at χ if it is both upper semicontinuous and lower semicontinuous at x, and that it is continuous if and only if it is continuous at every point o/Dom(-F). Remark — We can use the concepts of inverse images ала cores to characterize upper and lower semicontinuous maps: Proposition 1.4.4 A set-valued map F :X ^Y is upper semicontinuous at χ if the core of any neighborhood ofF(x) is a neighborhood of χ and a set-valued map is lower semicontinuous at χ if the inverse image of any open subset intersecting F(x) is a neighborhood of x. Hence, F is upper semicontinuous if and only if the core of any open subset is open and it is lower semicontinuous if and only if the inverse image of any open subset is open. IfOom(F) is closed, then F is lower semicontinuous if and only if the core of any closed subset is closed and F is upper semicontinuous if and only if the inverse image of any closed subset is closed. We shall also need to adapt to the set-valued case the concept of Lipschitz applications.
1.4. Continuity of Set-Valued maps 41 Definition 1.4.5 When X and Υ are normed spaces, we shall say that F : Χ ~> Υ is Lipschitz around χ £ X if there exist a positive constant I and a neighborhood U С Dom(F) of χ such that \/X!,X2&U, F(xi) С F(x2) + l\\xi - Χ2\\Βγ In this case F is also called Lipschitz (or I—Lipschitz) on U. If у is given in F(x), F is said pseudo-Lipschitz around (x,y) € Graph(F) if there exist a positive constant I and neighborhoods U С Dom(-F) of χ andV of у such that Vii,i2eW, F(xi)nV С F{x2) + l\\xi - x2\\BY We can also adapt the definitions of upper and lower limits to the case of a set-valued map F : X ~»· Y. For that purpose, the notation x1 -+p χ means that x' remains in Dom(.F) and converges to x. Definition 1.4.6 When F : Χ ~> Υ is a set-valued map, we say that IAmsu-ox,^xF{x') := \y £ Υ \ liminf d(y,F(x')) = θ) is the upper limit of F(x') when x' —> χ and that Um.mix,^xF{x') := \y <Ξ Υ | lim d(y,F(x')) = θ] is the lower limit of F(x') when x' —> x. They are obviously closed. We also see at once that Liminix/^xF(x') С F(x) С Limsnpx,^xF(x') Remark — The use of the concept of filter would avoid duplicating these definitions in the discrete and continuous cases. We have preferred this longer, hopefully more pedagogical, presentation. D Remark — As in the case of sequences of sets, we can easily obtain the following characterization of upper and lower limits. Set К := Dom(F). Then Limsupa^I^a/) = Γ\η>0{]χ>ζΒκ(χ,η)Ρ(χ') = Πε>ο ί\>ο ϋχ>ζΒκ(χ,η) B(F(x'), ε)
42 1 - Continuity of Set-Valued Maps and Lmnnix^xF(x') = Π U Π B{F{x'),e) □ ε>0 7?>0χ'ςβκ(χ,7?) The connections between semi-continuity of set-valued maps and semi limits are given by Proposition 1.4.7 A point (x, y) belongs to the closure of the graph of α set-valued map F : Χ ~> Υ if and only if у € Limsup3.,_t3.F(a;') and F is lower semicontinuous at χ £ Dom(-F) if and only if F(x) С Liminf-c'-^-F^') Then we can measure the lack of closedness (of the graph) or the lack of lower semicontinuity by the discrepancy between the sets F(x), L]mmix>^>xF(x') and Limsup3./_t3.ir(a;/) Remark — This proposition led several authors to call upper semi- continuous maps the ones which are closed in our terminology. Naturally, these two concepts coincide for compact-valued maps, since Theorem 1.1.4 can be easily adapted to the case of set-valued maps: Proposition 1.4.8 The graph of an upper semicontinuous set-valued map F : X ~> У with closed domain and closed values is closed. The converse is true if we assume that Υ is compact. More generally, the "continuous version" of Proposition 1.1.5 reads: Proposition 1.4.9 Let F and G be two set-valued maps from X to Y. Assume that F is closed, that G{x) is compact and that G is upper semicontinuous at χ G Dom(F Π G). Then F Π G is upper semicontinuous at χ. The above proposition provides an easy way to construct upper semi- continuous set-valued maps, by "cutting" set-valued maps with closed graph by set-valued maps with ball values, the radius of which are upper semi- continuous (real-valued) functions:
1.4. Continuity of Set-Valued maps 43 Corollary 1.4.10 Let F : X ^У be α closed set-valued map and r:X U R+ he an upper semicontinuous function. If the dimension of Υ is finite , then the cut set-valued map Fr : Χ ~> Υ defined by Fr(x):=F(x)Dr(x)B is upper semicontinuous. This is due to the fact that χ ~> r{x)B is upper semicontinuous (respectively lower semicontinuous, Lipschitz) whenever r is upper semicontinuous (respectively lower semicontinuous, Lipschitz.) Remark — We shall prove later (Theorem 2.2.5) that the set-valued analogues of continuous linear operators, called closed convex processes, from a Banach space to another are also upper semicontinuous (and even, Lipschitz): this is an extension of Banach's Closed Graph Theorem. □ We can extend the concept of proper maps to set-valued maps in the following way: Definition 1.4.11 (Proper Set-Valued Map) Let us consider a closed set-valued map F : X ~> Y. The two following properties are obviously equivalent: 1. — The projection πγ : Graph(F) иУ is proper 2. — If a sequence yn £ F{xn) converges in Υ, then the sequence (xn)neN has a cluster point. If they are satisfied, we shall say that the set-valued map F : X -^У is proper. When F is proper, the images F{K) of closed subsets К С X are closed and the inverse images F~l(M) of compact subsets Μ С Υ are compact. Indeed, we can write i) F{K) = πγ (Graph(F) n(Kx Y)) ii) F-X{M) = έχ (Graph^nTiy^M)) In particular, the image of a proper set-valued map is closed.
44 1 - Continuity of Set-Valued Maps Proposition 1.4.12 Let us assume that X is locally compact and that for every compact subset К С X, the graph of the restriction F\k ofF : X ^>Y to К is compact. Then 1. — F is upper semicontinuous 2. — F~l is proper (and thus, its domain is closed) Proof— Since for all a; ε Dom(F), there exists a compact neighbor- -hood К of x, the restriction F\k is upper semicontinuous on K, having a compact graph. Then F is upper semicontinuous at x. Let a sequence xn € F~l{yn) с Dom(F) converge to some x. We have to check that the sequence yn has a cluster point. Since the convergent sequence (£n)neN remains in a compact subset K, and since the pair (хп,Уп) belongs to the graph of F\k, which is compact, the sequence (xn, yn) has a cluster point of the form [x, y), which belongs to the graph of F. Hence у £ F~1(x) is a cluster point of (j/n)^eN- D 1.4.2 Generic Continuity Although the notions of upper semicontinuity and lower semicontinuity are distinct, they generically coincide, i.e., a semicontinuous map is continuous on a residual. Recall that a residual of a metric space X is a countable intersection of dense open subsets An с X. Countable intersections of residuals are residuals. Baire's Theorem9, states that a residual of a complete metric space is dense. A property which is true for every element of a residual is called generic. Theorem 1.4.13 (Generic Continuity) Let F be a set-valued map from a complete metric space X to a complete separable metric space Υ. 1. — If F is upper semicontinuous, it is continuous on a residual ofX. 2. — If F is lower semicontinuous with compact values, it is continuous on a residual of X. 3. — If F is lower semicontinuous with closed values, then there exists a residual R of X such that Viefi, Limsupa.^iV) = F(x) proved in 1899...
1.4. Continuity of Set-Valued maps 45 Proof— Since У is a separable metric space, there exists a countable family of open subsets Vn С Υ, stable by finite union satisfying the following property: V open subset V С Υ, V у ε V, 3 V„ such that у ε V£ с V 1. — Assume first that F is upper semicontinuous. We associate with any open subset Vn the subset Ln := F-X(K) = {x£X \ F(x)n%^4>} which is closed by Proposition 1.4.4 since the Vn are closed. We then note that if χ ε X is such that V ra e N, x&Ln => xe Int(Ln) then F is lower semicontinuous at x. Indeed, let V С Υ be any open subset such that F{x) Π V φ 0. Then there exists η ε Ν and ν ε f(i)ni; с F(a;)nv Thus χ € Ln and therefore it belongs to Int(jLn.) Since Ln = F-ЦЩ С F-lQ>) we deduce that F is lower semicontinuous at x. Therefore the subset Db of points where F is not lower semicontinuous is contained in the countable union of the subsets dLn, that are closed subsets with empty interior. By taking the complements, we infer that the complement of Db, which is the set of elements at which F is continuous, contains a residual. 2. — If F is lower semicontinuous, then the subsets Kn := F+1(K) = {xeX \ F{x) cK} are closed by Proposition 1.4.4. If χ ε Χ is such that VneN, x<=Kn => χ ε Ы(Кп) F is upper semicontinuous at x. Indeed, let an open set V С У be such that F(x) с V. For every у ε F(x), let V„u be such that у ε Vny С V. Since F has compact images, F{x) can be covered by a finite number of sets Vnj ··= Vny., j = 1, ···, m. Thus τη - - τη F{x) С (J Щ := (J Vnj С V
46 1 - Continuity of Set-Valued Maps The family of Vn being stable by finite union, we deduce that for somen, χ € F+l (pVnjj = Kn С F+X(V) Thus, a; £ Int(F+1(V) and we have proved that F is upper semicontinuous at x. Therefore the subset D$ of points where F is not upper semicontinuous is contained in the countable union of the subsets dKn := Kn\ Int(.fiTn). The interiors of these subsets are empty because they are closed. By taking the complements, we deduce that the complement of D", which is the set of elements at which F is continuous, contains a residual. 3. — We follow Choquet's proof to deduce the third statement. It is based on the fact that any complete separable metric space Υ is homeo- morphic to a subset Zq of a compact metric space Z. Denote by φ the homeomorphism from Υ onto Zq, and set G0 := φ ο F : X ~» Z0 k G(x) := G0(x) the closure of Gq{x) in Z. Since F is lower semicontinuous, so is Go and G. By the preceding statement, there exists a residual R of X such that G is upper semicontinuous from R to Z. Therefore, for any ε > 0, there exists η such that, for any χ' ε Β(χ,η), G0(x') С G(x') С B(G(x),e) = B(G0(x),e) On the other hand, Go(a;) = G(x) Π Zq since the images Go(a;) are closed in Zq for the induced topology. Therefore MxGR, Limsupx,_xGo(a/) = Go(a;) Since ψ is an homeomorphism, F enjoys the same property. □ Remark — It is easy to construct a lower semicontinuous set-valued map F (with closed but noncompact values) which is nowhere upper semi- continuous: Take X = R, Υ = R2 and F(t) := {{x, y)\y = tx}. Π 1.4.3 Example: Parametrized Set-Valued Maps Let us consider three metric spaces Χ, Υ and Z, a set-valued map U:X ~> Ζ
1.4. Continuity of Set-Valued maps 47 and a single-valued map / : Graph (17) ■-» Υ We associate with these data the set-valued map F : Χ ~> Υ defined by Vx€X, F{x) := (f(x, u))ueu{x) Proposition 1.4.14 Assume that f is continuous from Graph(!7) toY. — If U is lower semicontinuous, so is F. — If U is upper semicontinuous with compact values, so is F. Proof — Let us consider a sequence xn £ Oom(F) converging to χ € Dom(-F) and take у := f(x,u) belonging to F(x), where и €U(x). Since U is lower semicontinuous, there exists a sequence un G U(xn) converging to u. Then the sequence yn = f(xn,un), where yn belongs to F(xn), converges to у because / is continuous. Hence F is lower semicontinuous. Let us fix χ € Dom(l7), e > 0 and consider the neighborhood B(F(x), e) of F(x). Since it is a neighborhood of each f(x, u) when и € U(χ), the continuity of / implies the existence of щ > 0 and 6U > 0 such that V (x1, v) £ Graph([/) П (B(x, η») χ B{u, 6U)), f{x', v) € B{F(x), e) The subset U(x) being compact, it can be covered by ρ balls В(щ, 6щ). Since U is upper semicontinuous, there exists 770 > 0 such that ρ Vx' € Β(χ,η0), U(x') С \jB(Ui,6Ui) i=l We take η := min(77o,mini=i)...)P щ{) > 0. Then we infer that Vs'__e Β(χ,ή), F{x') с B(F(x),e) i.e., that F is upper semicontinuous at x. □
48 1- Continuity of Set-Valued Maps 1.4.4 Marginal Maps Definition 1.4.15 (Marginal Functions) Let us consider α set- valued map F : Χ ~> Υ and a function f : Graph(F) i-> R. We associate with them the marginal function jiIhRu {+°o} defined by g{x) := sup f(x,y) yeF(x) Theorem 1.4.16 (Maximum Theorem) Let metric spaces X, Y, a set-valued map F : Χ ~> Υ and a function f : Graph(F) i-> R be given. — If f and F are lower semicontinuous, so is the marginal function. — // / and F are upper semicontinuous and if the values of F are compact, so is the marginal function. Proof — Let us" consider a sequence xn converging to x, fix λ < g(x) and choose у € F(x) such that λ < f(x,y). Then there exist elements yn G F(xn) converging to у (because F is lower semi- continuous) and we know that f{xn,yn) < g{%n)· Since / is lower semicontinuous, we infer that λ < f(x,y) < liminf/(a;n,yn) < liminf^(a;n) η—юо τι—юо By letting λ converge to g(x), the claim ensues. For proving the second statement, pick χ € X and fix e > 0. Since / is upper semicontinuous, we can associate with any у G F(x) open neighborhoods V(y) of у and Uy{x) of χ such that Vue Uy{x) and ν £V(y), f{u,v) < f{x,y) + e (1.4) Since F(x) is compact, it can be covered by η neighborhoods V(yi), i = 1, · · · ,p, the union of which makes up a neighborhood of F(x). Then there exists a neighborhood Uq{x) such that Vx'€ U0(x), F(x') С \JV(yi) i=l
1.5. Lower Semi-Continuity Criteria 49 because F is upper semicontinuous. By taking и in the neighborhood ρ U{x) := UQ{x) П f]UVi{x) we observe that \fueU(x), Vv£F(u), f(u,v)< sup f{x,yi)+e < g{x)+e i=l,---,p (thanks to (1.4)) and we deduce that \fu£U(x), g{u) < g{x) + e □ We will use quite often the following corollary·. Corollary 1.4.17 // α set-valued map F is lower semicontinuous (resp. upper semicontinuous with compact values), then the function (x,y) ι—» d(y,F(x)) is upper semicontinuous (resp. lower semicontinuous) on Dom(-F). 1.5 Lower Semi-Continuity Criteria We provide in this section lower semicontinuity criteria, some of them being adaptations of results on lower limits of sequences to the continuous case. Proposition 1.5.1 Consider a metric space X, two normed spaces Υ and Z, two set-valued maps G and F from X to Υ and Ζ respectively, and a (single-valued) map f from Χ χ Ζ to Υ satisfying the following assumptions: . i) G and F are lower semicontinuous with convex values < ii) f is continuous Hi) V χ £ X, и η-► f(x, и) is affine We posit the following condition: Vx £X, 37>0,£>0, c>0, r>0 such that V x' £ B(x, δ) we have ΊΒΥ С f{x',F{x')nrBz)-G{x')
50 1 — Continuity of Set-Valued Maps Then the set-valued map R:X~^Z defined by R(x) :=-{u€F(x) | f(x,u)eG(x)} (1.5) is lower semicontinuous with nonempty convex values. The proof is a trivial adaptation of the one of Proposition 1.2.6. We state now another condition which is less symmetric. Proposition 1.5.2 Consider a metric space X, two normed spaces Υ and Z, two set-valued maps G and F from X to Υ and Ζ respectively and a (single-valued) map f from X x Ζ to Υ such that ' i) F is lower semicontmuous with convex values ii) f is continuous < Hi) V χ G X, u i-> f(x, u) is affine iv) V χ G X, G{x) is convex and its interior is nonempty' v) the graph of the map X э χ ~» Int(G(a;)) is open We posit the following condition: Vx eX, 3u€ F(x) such that f(x, и) е Iht(G(s)) (1.6) Then the set-valued map R defined by (1.5) is lower semicontinuous with convex values. Proof 1. — We introduce the set-valued map S : Χ ~> Ζ defined by S(x) ={u£ F(x) | f(x,u) e Int(G(s))} С R{x) Assumption (1.6) implies that S(x) is not empty. We claim that S is lower semicontinuous. Indeed, if xn —> χ and if и belongs to S(x) С F(x), there exists a sequence of elements un G F(xn) which converges to и because F is lower semicontinuous. Since {xn-,f{xn-,un)) converges to (x, f(x, u)) € Graph(Int(G(·))) by continuity of / and since the graph of Int(G(·)) is open, the elements f(xn-,un) belong to Int(G(xn)) for η large enough and thus, the elements un belong to S(xn) and converge to u.
1.5. Lower Semi-Continuity Criteria 51 2. — Convexity of F(x) and G(x) implies that S(x) = R(x). Indeed, let us fix и € R{x) and ад € 5(ж). Then г/g := Quo + (l — 6)u belongs to S(x) when θ e]0,1[, because C?(:e) is convex and f(x,uo) belongs to the interior of G(x), so that for every θ e]0, l[, f{x, u) + 0(/(s, u0) - f{x, u)) = f(x, у + вуо- ву) = f(x, υθ) belongs to the interior of G(x). Then и is the limit of vq when θ > 0 converges to 0. 3. — The theorem follows because the closure of any lower semicontinuous set-valued map is still lower semicontinuous. Π We extend now the above lower semicontinuity criterion to infinite intersection of set-valued maps. Theorem 1.5.3 Let us consider α metric space X, normed vector- spaces Υ and Ζ and set-valued maps F : XxY ~> Ζ and Η : X ~> Y. We assume that i) F is lower semicontinuous with convex values ii) Η is upper semicontinuous with compact values and that there exist positive constants 7, δ, с such that for every single- valued map е:Ун jB we have Vx'e Β(χ,δ), сВП П (F(x',y)-e(y)) φ 0 (1.7) вея(х') Then the set-valued map G : Χ ~> Ζ defined by Vx€l, G{x) := f) F(x,y) y£H(x) is lower semicontinuous (with nonempty convex images.) Remark — When the set-valued map F is locally bounded (in the sense that it maps some neighborhood of each point to a bounded subset), we do not need the constant с and we can replace (1.7) by \/χ'€Β(χ,δ), Π (F(x',y)-e(y)) φ 0 D yetf(x')
52 1 - Continuity of Set-Valued Maps Proof— Let us choose any sequence of elements xn € Dom(-F) converging to χ and ζ € G(x). We have to approximate ζ by elements zn eG(xn). We introduce the following numbers: en := sup d(z,F(xn,y))/2 (1.8) yeH(xn) Now, let us choose for each у £ H(xn) an element un(y) € F(xn,y) satisfying ll«-«n(y)ll < Щг,Р(рп,у)) < en and set θη := 7/(7 + en). Consequently, 6n(«-«n(y)) € 6„enS = {1-θη)ΊΒ So that there exists ап(у) € 7В such that θη(ζ - un(y)) = (1 - вп)оп(у) Therefore, assumption (1.7) implies the existence for all η large enough of elements wn G cB and elements vn(y) € -Ρ(2;π, у) such that an{y) = vn{y) ~ Wn for all у € Я(яп). Hence we can write 0n(z - Un(y)) = (1 - 6n)(vn(y) - wn) So that the common value: zn := θηζ + (1 - ^η)ωπ = 9nun(y) + (1 - 6n)vn{y) does not depend on y, belongs to all F(xn,y) (by convexity), converges to ζ because \\z-zn\\ = (l-en)\\z-wn\\ < (1-θη)(\\ζ\\+α) and, because 1 — θη = en/(-y + en) converges to 0 for en, converges to 0 thanks to the following lemma. □ Lemma 1.5.4 Let us assume that F is lower semicontinuous and that Η is upper semicontinuous with compact images. Then the numbers en defined by (1.8) converge to 0.
1.5. Lower Semi-Continuity Criteria 53 Proof— Since F is lower semicontinuous, Corollary 1.4.17 to the Maximum Theorem implies that the function (x,y,z) н-> d(z,F{x,y)) is upper semicontinuous. Therefore, for any e > 0 and any у £ Η {χ), there exist an integer Ny and a neighborhood Vy of у such that Vy'eVj,, Vn>Ny, d{z,F(xn,y')) < ε (1.9) because d(z,F(x,y)) = 0. Hence the compact set H{x) can be covered by ρ neighborhoods Vyi. Furthermore, Η being upper semicontinuous, there exists an integer iVo such that, Vn>iV0, H(xn) с (J Vy. i=l, ...,p Set N := тахг-=о,...,р NVi. Then, for all η > N and у G Я(жТ1), у belongs to some Vyi, so that, by (1.9), d(z,F(xn,y)) < ε. Thus, Vn>iV, en := sup d(z,F(xn,y))/2 < e/2 y£H{xn) i.e., our lemma is proved. □ For set-valued maps with non convex images, we deduce from Theorem 1.2.9 its continuous version: Theorem 1.5.5 Let G : Χ ~> Ζ be α closed lower semicontinuous set-valued map from a metric space X to a Banach space Ζ and f : X xY н-> Ζ a continuous (single-valued) map, where Υ is another Banach space. Let us assume that f is differentiable with respect to у and that there exist constants с > 0 and η > 0 such that ' Vi£B(i0,i)), je%,i?), 2£B(/(a:o,5o),i))nG(i) < Bz С cfy(x,y)(BY)-TG[x)(z) (1.10) Then the set-valued map R defined by R(x) := {y€ Υ | f(x,y)€G(x)} is lower semicontinuous at xq-
Chapter 2 Closed Convex Processes Introduction Naturally, the first question which arises is: What are the set- valued analogues of continuous linear operators! Since the graph of a continuous linear operator A £ £(X, Y) is a (closed) vector subspace of Χ χ Y, it is quite natural to regard set-valued maps, with closed convex cones as their graphs, as these set-valued analogues. Such set-valued maps are called closed convex processes1 and the maps the graph of which are vector subspaces are called linear processes. The main class of examples of closed convex processes is provided by derivatives of set-valued maps which are introduced and studied exhaustively in Chapter 5. We shall prove that closed convex processes enjoy (almost) all properties of continuous linear operators, including Banach's Open Mapping and Closed Graph Theorems (Section 2) and the Uniform Boundedness Theorem (Section 3.) As continuous linear operators, closed convex processes can be transposed and the Bipolar Theorem can be adapted to closed convex processes. They thus enjoy the benefits of a duality theory exposed in Section 5. For instance, a duality criterion of invariance by a 'The term "process" has been coined by R.T. Rockafellar for denoting maps the graph of which are cones in a study of economic "processes" (with constant return to scale.) 55
56 2- Closed Convex Processes closed convex process is given in this section for linear processes and in Chapter 4 (Section 2) in the general case. We shall also provide in Chapter 3 a theorem on existence of eigenvectors of closed convex processes (Theorem 3.6.2.) For proving these results, we recall in Section 4 the Bipolar Theorem for continuous linear operators, the Closed Range Theorem and the properties of support functions. The latter allow characterization of closed convex subsets by an (infinite) family of linear inequalities thanks to a version of the Hahn-Banach Separation Theorem. It also enables one to deal with the class of upper semicontinuous convex positively homogeneous functions instead of handling closed convex subsets. We conclude the chapter with a short section on upper hemicon- tinuous set-valued maps, characterized by the upper semicontinuity of the support functions of their values, a more familiar property. This characterization is mainly useful for set-valued maps with closed convex images. 2.1 Definitions Let us introduce the set-valued analogues of continuous linear operators, which are the closed convex processes. Definition 2.1.1 (Closed Convex Process) Let F : Χ ~> Υ be a set-valued map from a normed space X to a normed space Y. We shall say that F is — convex if its graph is convex — closed if its graph is closed — a process (or positively homogeneous,) if its graph is a cone — a linear process if its graph is a vector subspace. Hence a closed convex process is a set-valued map whose graph is a closed convex cone. We shall see that most of the properties of continuous linear operators are enjoyed by closed convex processes. Let us begin with the following obvious statements.
2.2. Open Mapping and Closed Graph 57 Lemma 2.1.2 A set-valued map F is convex if and only if ' Vsi, X2 € Dom(F), VA e [0,1], < AF(xi) + (1 - AjFCsa) С F(Asi + (1 - A)s2) It is a process if and only if \/x e X, V λ > 0, AF(s) = F(As) and 0 e -F(O) and a convex process if and only if it is a process satisfying Vsi, ж2 € X, FCsiJ + F^) С F(x1+x2) We observe that the domain and the image of a closed convex process are convex cones (not necessarily closed.) The main examples of closed processes are provided by contingent derivatives of set-valued maps that we shall introduce in Chapter 5. We associate with a closed convex process its norm defined in the following way. Definition 2.1.3 (Norm of a Closed Convex Process) Let F : X~> Υ be a closed convex process. Its norm ||.F|| is equal to ' \\F\\ ··= supieDom(F)d(0,F(a;))/||a;|| = suPxeDom(F)mW(x)IM|/IM| (2.1) = suPxeDom(F)nB 'miv eF{x) 11ν11 2.2 Open Mapping and Closed Graph Theorems The Banach Open Mapping Theorem can be extended to closed convex processes: Theorem 2.2.1 (Open Mapping) Let Χ, Υ be Banach spaces. Assume that a closed convex process F : Χ ~> Υ is surjective (in the sense that Jm(F) =Y). Then -F-1 is Lipschitz:
58 2- Closed Convex Processes There exists a constant I > 0 such that, for all x\ G F 1(yi) and for any ?/2 € У, we can find a solution x^ G F~1(y2J. satisfying: \\xi -ж2|| < l\\yi — 2/21| Actually, this theorem holds true for closed convex maps: Theorem 2.2.2 (Robinson-Ursescu) LetX, Υ be Banach spaces, F : Χ ~> Υ be a closed convex set-valued map. Suppose that yo belongs to the interior of the image of F and let xq G F~1(yo). Then there exist positive constants I and 7 such that for any у G г/о + 7-В, there exists a solution χ to the inclusion F(x) Э у satisfying \\x-x0\\ < l\\y-yo\\ Proof — For simplicity, we prove this theorem only when the Banach space X is reflexive2. Let us introduce the function ρ defined by p(y) := inf ||s-soil = dix^F-^y)) (2.2) xeF-Цу) (It takes the value +00 when у does not belong to Im(F).) Since the set-valued map F is convex, this function is obviously convex. Assume for a while it is lower semicontinuous. Then, Baire's Theorem implies that ρ is continuous3 on the interior of Im(F), which is not empty by assumption. Since continuous convex functions are locally Lipschitz, there exists a ball of radius 7 > 0 centered at yo and a constant I' > 0 such that for all у in this ball, llp(y)ll = llp(y)-p(yo)|| < %-уо|| 2See for instance [35, Theorem 3.3.1] for the nonreflexive case and the original papers of Robinson and Ursescu. 3Let us recall this result: The domain of ρ is the union of the sections Sn := {y I p(y) < n) which are closed because ρ is lower semicontinuous. Since the interior of the domain of ρ is not empty, the interior of one of these sections is not empty. So, the convex function ρ being bounded on an open subset Int(5n) for some n, it is continuous (and even, locally Lipschitz) on the interior of its domain.
2.2. Open Mapping and Closed Graph 59 because p(yo) = 0. Therefore d^F-'iy)) < i'||y-yo|| By setting I = 21' we end the proof. It remains to check that the function ρ is lower semicontinuous. Π Lemma 2.2.3 Let us consider α reflexive Banach space X, a Ba- nach space Y, a closed convex process F : X ~> У'. Then the function ρ defined by (2.2) is lower semicontinuous. Proof — It is enough to prove that nonempty sections {y I p(y) < λ} are closed. Let us consider a sequence of elements yn of such a section converging to some y. Then, by reflexivity of X, there exists Xn € F 1(yn) satisfying || Xn -CO II = р(Уп) < λ. Hence the elements xn remain in the ball xq + \B. Since X is reflexive, there exists a (weak) cluster point χ of the sequence xn. Then (x, y) is a weak cluster point of the sequence (хп,Уп), which thus belongs to the graph of F because, being closed and convex, it is closed in Χ χ Υ when X is supplied with the weak topology and Υ with the norm topology. Since the elements xn belong to the ball xq + \B, which is weakly closed, the cluster point χ belongs also to this ball, so that p{y) < \\x — xo\\ < λ Π Proof of Theorem 2.2.1 — We take xq = 0, yo = 0 in Theorem 2.2.2. To say that 0 belongs to the interior of the cone Im(F) amounts to saying that Im(F) is equal to Y, i.e., that F is surjective. By Theorem 2.2.2 for a constant I > 0 we have: Vy € У, 3x € F~1(y) such that ||x|| < 2||y|| Fix ?/i, г/2 € Υ and let us choose any solution x\ € F~1{y{) and e e F~1(y2 - 2/1) satisfying ||e|| < l\\yi— 2/2II· Then X2 := xi + e belongs to i?_1(2/2), since F is a convex process and satisfies the estimate ||zi-S2|| = ||e|| < i||yi-y2|| °
60 2- Closed Convex Processes Corollary 2.2.4 Let us consider Banach spaces X, Y, a continuous linear operator A G C(X, Y) and a_ closed convex subset К С X. Let us assume that there exists xq G К such that Axq G Int(A(K)). Then there exist positive constants I and j such that for any у G A(xq) + 7-B, there exists a- solution χ £ К to the equation Ax = у satisfying \\x — xq\\ < l\\y — A(xq)\\. - —*-- Remark — Actually, Corollary 2.2.4 is equivalent to Theorem 2.2.2; we apply it with К := Graph(F) and A := πγ. □ Since the restriction F := Α\κ of a continuous linear operator A G C(X, Y) to a closed convex cone is a closed convex process, we obtain the following consequence: Corollary 2.2.5 Let us consider Banach spaces X, Y, a continuous linear operator A G C(X, Y) and a closed convex cone К С X such that A(K) = Y. Then the set-valued map у ~> A~1{y) Π Κ is Lipschitz: there exists a positive constant I such that, Vyi, 2/2 € Y, А-1{У1)ПК С А-1{у2)ПК + i||yi-y2||£ As in the case of continuous linear operators, the Open Mapping Theorem is equivalent to the Closed Graph Theorem, which can be stated as follows. Theorem 2.2.6 (Closed Graph Theorem) A closed convex process F from a Banach space X to another Υ whose domain is the whole space is Lipschitz: there exists a (Lipschitz) constant I > 0 such that Vxu x2 GX, F(xi) С F(x2) + i||si-s2||-B (2.3) Thus, the norm of F is finite whenever Dom(F) = X. Proof — It is sufficient to apply the Open Mapping Theorem 2.2.1 to the closed convex process -F-1. □
2.3. Uniform Boundedness 61 2.3 Uniform Boundedness Theorem We can now adapt the Uniform Boundedness Theorem to the case of closed convex processes. Theorem 2.3.1 (Uniform Boundedness) Let X and Υ be Ba- nach spaces and Fh be a family of closed convex processes from X to Y, "pointwise bounded" in the sense that V χ € X, 3yh € Fh(x) such that вирЦу^Ц < +oo (2.4) h Then this family is "uniformly bounded" in the sense that sup 11iX11 < +oo h Hence we can speak of bounded families of closed convex processes, without specifying whether it is pointwise or uniform. Proof — Let us consider the positively homogeneous convex lower semicontinuous functions ph defined by ph{x) := inf \\y\\ = d(0,Fh(x)) j/eFh(x) (which are lower semicontinuous because the closed convex processes Fh are Lipschitz) and the function ρ defined by У χ £ Χ, ρ(χ) := supphfa) h Assumption (2.4) implies that this function ρ is finite. Since it is also positively homogeneous, convex and lower semicontinuous (being the supremum of such functions), it is continuous at 0. Hence there exists a constant I such that saphd(0,Fh(x)) = p(x) < l\\x\\, i.e., \\Fh\\ < l< oo. □ The following consequence of Theorem 2.3.1 extends to closed convex processes the following useful convergence result. Theorem 2.3.2 (Crossed Convergence) Consider a metric space U, Banach spaces Χ, Υ and a set-valued map associating to each
62 2- Closed Convex Processes и £ U a closed convex process F(u) : X ~> Y. Let us assume that the family of closed convex processes F(u) is pointwise bounded. Then the following conditions are equivalent:' i) the map и ~> Graph(.F(ii)) is lower semicontinuous < ii) the map (u,x) ~> F{u)(x) is lower semicontinuous Proof — For proving that г) implies ii), let us consider a sequence of elements (un, xn) converging to (u, x) and an element у € F(u)(x). We have to approximate it by elements yn £F(un)(xn). Since и ~> Gr&ph(F(u)) is lower semicontinuous, we can approximate (x, y) by elements (хп,Уп) £ Graph(-F(un)). By the pointwise boundedness assumption and Theorem 2.3.1, there exist I > 0 and solutions fn € F{un){xn — xn) satisfying ||/π|| Ъ. ΙΊ\Χη 2·π|| The right hand side of the above inequality converges to zero when η goes to infinity. Because F(un) is a convex process, the element Уп '■= Уп + fn does belong to F(un)(xn). Consequently, yn converging to y, we deduce that the set-valued map (u, x) ~> F(u)(x) is lower semicontinuous at (u, x). The converse is obviously true even when the family (F(u))ueu is unbounded. Π 2.4 The Bipolar Theorem Definition 2.4.1 Let К be a nonempty subset of a Banach space X. We associate with any continuous linear form ρ £ X* σκ{ρ) '■= σ(Κ,ρ) := sup < ρ, χ> € RU{+oo} хек The function σκ ■ X* nRU {+oo} is called the support function of K. Its domain is a convex cone called the barrier cone denoted by b{K) := Όοτη(σκ) := {реГ| σκ{ρ) < oo} (2.5)
2.4. The Bipolar Theorem 63 We say that the subsets of X* defined by ' i) K° ii) Κ- ίη) K+ iv) K1 = {p & X*\ σκ(ρ) < 1} = {ρ & X*\ σκ{ρ) < 0} = -κ- = {ρ € Χ* I V χ € К, < ρ, χ > = 0} are the polar set, (negative) polar cone, positive polar cone and orthogonal of К respectively. When L С I*, we define the polar set L° С X as the subset of elements χ £ X (and not X**) satisfying < ρ, χ >< 1 for all ρ G L. The polar cone L~ С X and the orthogonal L1 С X of L are defined in the same way. The subsets K°° := {K°)° С X к К— := {К')' С X are called respectively the bipolar set and bipolar cone of a subset К С X and the subspace K±J- := {К1)1- С X the biorthogonal of jr. It is clear that K° is a closed convex subset containing 0, that K~ is a closed convex cone, that AT is a closed subspace of X* and that K1 = К~ПК+ С K~ С K° С b(K) Examples о When К = {χ}, then σκ(ρ) — <p,x> © When К = Дх, then σΒχ (ρ) = ||ρ||* ® If AT is a cone, then , . / 0 if ?€Г σ*ω = ( +oo if pi К' D When AT = 0, we set σ%(ρ) = — oo for every ρ G X*. The Separation Theorem can be stated in the following way:
64 2- Closed Convex-Processes Theorem 2.4.2 (Separation theorem) Let К be a nonempty subset of a Banach space X. Its closed convex hull is characterized by linear constraint inequalities in the following way: co(K) ={ieI|Vj)6l*, <ji,i>< σκ{ρ)} Furthermore, there is a bijective correspondance between nonempty closed convex subsets of X and nontrivial lower semicontinuous positively homogeneous convex functions on X*. Remark — The Separation Theorem holds true not only in Banach spaces, but in any Hausdorff locally convex topological vector-space. In particular, we can use it when X is supplied with the weakened topology. The geometrical interpretation can be stated as follows: the closed convex hull of a nonempty subset is the intersection of all closed half-spaces containing it. □ We observe that a subset К is bounded if and only if its support function is finite. We mention the following consequence, known as the Bipolar theorem. Theorem 2.4.3 (Bipolar Theorem) Let X,Y be Banach spaces and К С X. The bipolar cone К is the closed convex cone spanned by K. If A G C(X, Y) is a continuous linear operator from X to Υ and К is a subset of X, then (A(k))- = a*-\k-) where A* denotes the transpose of A. Thus the closed cone spanned by A(K) is equal to (A* {K~~)j . We provide now a simple criterion which implies that the image of a closed subset is closed. Theorem 2.4.4 (Closed Range Theorem) Let X be a Banach space, Υ be a reflexive space, К С X be a weakly closed subset and A G C(X, Y) a continuous linear operator satisfying lm(A*) + b{K) = X* (2.6)
2.4. The Bipolar Theorem 65 Then the image A(K) is closed. In particular, if К is a closed convex cone and if bn(A*) + K- = X* then A(K) = (a*-\k-))~ Proof — Let us consider a sequence of elements xn G К such that A(xn) converges to some у in Y. We shall check that this sequence is weakly bounded, and thus, weakly relatively compact. Let us take for that purpose any ρ € X*, which can, by assumption (2.6), be written ρ := A*q + r, where q<=Y*, r G b(K). Therefore, supn < p,xn > = supn(< q, Axn > + < r,xn >) < < supn(||g||||.Aa:n|| + aK{r)) < +oo since the converging sequence (Axn)nej^ is bounded. Therefore the sequence (£n)neN has a weak cluster point χ which belongs to К since it is weakly closed. D Remark — Actually, arguments of the above proof imply that, under assumptions of the Closed Range Theorem, every sequence xn € К such that A(xn) is weakly converging, has a weak cluster point □ For the convenience of the reader, we list below some useful calculus of support functions and barrier cones4. 4 See [35, Chapter 3] for instance.
66 2- Closed Convex Processes Table 2.1: Properties of Support Functions. > If К С L С X, then 6(L) С b(K) and σ^ < aL > If AT; С X, i £ I, then №^))сае/Р«) o"(co((Jie/ Ki)iP) = suPie/ °X-Cp) > Ιΐ KiCXi, (i = 1,..., n), then ΚΠ?=ι^) = Π?=ιΚ^) *(Π?=ι *i> (Pi. ■ ■ ■. Pn)) = Σ?=ι ^Х-Ы > If A e ДА", У), then 6(Α(ϋΤ)) = Α*_16(ϋΤ) °аЩ)^ = σκ{Α*Ρ) > If Κι and Й2 are contained in X, then 6(ΛΓι + K2) = b(Ki) П 6(ϋΤ2) σκ1+κ2 (ρ) = σΚι (ρ) + σκ2 {ρ) In particular, if К С X and P is a cone, then b(K + P) = b(K) П Ρ- and σκ+ρ(ρ) = σχ(ρ) if ρ £ P~ and +oo if not (5) > If L С X and Μ С У are closed convex subsets and A G C{X, Y) is a continuous linear operator such that the following constraint qualification condition 0 € Int(M - A(L)) holds true, then 6(L П A-1 (M)) = b(L) + A*b(M) and VpeifLnA-^M)). 3 5 e У* such that οί,οα-ΜλοΟρ) =<tl(p- A*q) + σΜ{Ά) = miqeY*(aL(p - A*q) + σΜ{θ)) (5)a) > If Μ С У is a closed convex subset and if A € C{X, У) is a continuous linear operator such that 0 € Int(Im(A) - M), then ^^(M)) = A*b{M) and, for every ρ £ 6(Л_1(М)), there exists g € 6(M) such that σΛ-ι(Μ)(ρ) = σΜ(?) = ΐηίΑ*9=ρ(σΜ(?)) (5)b) > If K\ and ΛΓ2 are closed convex subsets of X such that 0 € Int(UTi - ЯГ2), then 6(JTi η K2) = b{K{) + b{K2) and V ρ € b(Ki Π ΑΓ2), 3pP,(i= 1,2) such that с^пАГгСр) = σΚι{θΊ) + VK2{qi) = να.ΐρ=ρι+ρ2{σΚλ{Ρι) + σκ2{Ρ2)) (1) (2) (3) (4)a) (4)b)
2.5. Transposition 67 2.5 Transposition of Closed Convex Process Definition 2.5.1 (Transpose of a Process) Let X,Y be Banach spaces, F : Χ ~> Υ be a process. Its left-transpose (in short, its transpose,) F* is the closed convex process from Y* to X* defined by ρ € F*(q) <==> У χ £ X, У у £ F(x), <p,x>~< < q,y > (2.7) In particular, the transpose F* of a linear process F is defined by p £ F*(q) <{=> Ух £ X, У у £ F(x), <p,x> = < q,y > The graph of the transpose F* of F is related to the polar cone of the graph of F in the following way: Lemma 2.5.2 (Graph of the Transpose) Let us consider Banach spaces Χ, Υ and let F : X ~> У be a process. Then (q,p) € Graph(F*) «=!►(?,-?) € (Graph(F))" In the case of linear processes, we observe that ρ G F*(q) if and only if (p,—q) belongs to Graph(F)-1- and we see at once that the bitranspose of a closed linear process F coincides with F. The definition of a bitranspose of a convex process is not symmetric: If G : Y* ~> X* is a convex process, we define its transpose G* : Χ ~> Υ by the formula (-y,x) € (Graph(G))- (instead of the formula (y, —x) € (Graph(G))" obtained by exchanging the roles of X and Υ*, Υ and X* respectively.) With this definition, the bitranspose of a closed convex process F coincides with F. We provide now a formula for transposing the product of closed convex processes.
68 2- Closed Convex Processes Theorem 2.5.3 (Transpose of a Product) Let W, Χ, Υ, Ζ be Ba- nach spaces, F be a closed convex process from X to Υ', A G C{W, X) and В G C(Y, Z) be continuous linear operators. Assume that lm(A)-Oom(F) = X (2.8) Then the transpose of BFA is equal to: (BFA)* = A*F*B* Proof — First, we prove that the formula (BF)* = F*B* always holds true, since the graph oiBF is equal to (1 xi?)Graph(.F). Consequently, thanks to the Bipolar Theorem 2.4.3 ((1 χ B)Graph(F))~ = (1 x B)*'1 (Graph(F)-)~ so that (p, —q) belongs to Graph(i?.F)"~ if and only if (p, —B*q) belongs to Graph(.F)"~, i.e., if and only if ρ belongs to F*{B*q). The graph of FA being equal to (A x l)_1Graph(.F), we need to assume the constraint qualification property Im(A x 1) - Graph(F) = Χ χ Υ (2.9) for deducing from the properties of polar cones that (Graph(FA))- = (A* x 1) (Graph(F)-) If this is the case, we infer that r G (FA)*(q) if and only if there exists ρ € F*(q) such that r = A*p. It remains now to check that assumption (2.8) implies the constraint qualification property (2.9.) Indeed, let (x, y) belong to Χ χ Υ. Since χ can be expressed as χ = Axq — χι where xq belongs to W and x\ to the domain of F, we can write (x,y) = (AxQ,yQ)- (xi,yi) where y\ G F(x\) and yo = у + y\. Hence (χ,у) belongs to Im(A χ l)-Graph(F). We then deduce from the above proof the conclusion of the theorem. □
2.5. Transposition 69 Corollary 2.5.4 Let X,Y and Ζ be Banach spaces, F a closed convex process from X to Z, G a closed convex process from Υ to Ζ and A£C(X,Y). If A(Oom{F))-Oom(G) = Υ (2.10) then {F + GA)* = F* + A*G*. Proof— We set H(x) := F(x) + G{Ax), B(y,z) := у + ζ so that the set-valued map Η can be written Η = B(F χ G)(l χ A). Since Dom(.F χ G) is equal to Dom(.F) χ Dom(G), assumption (2.10) implies that Im(l χ A) - Oom(F xG) = XxY Therefore from Theorem 2.5.3 follows that H* = (1 χ A)*(F χ G)*B*. Since {F χ G)* = F* xG* к (1 χ A)* = 1 + A* we infer that H* = F* + A*G*. Π Corollary 2.5.5 (Transpose of the Restriction) LetX,Y be Banach spaces, F : Χ ~» Υ be a closed convex process and К С X be a closed convex cone. Assume that K-Oom{F) = X Then the transpose of the restriction F\k of F to К is given by (F\k) (q) - | 0 otherwise Proof— We apply Corollary 2.5.4 with A = 1 and G defined by G(x) = [ {0} if X € K 10 otherwise whose domain is К and whose transpose is the constant set-valued map denned by G*{q) =K~. Π We shall adapt to the case of closed convex processes the Bipolar Theorem 2.4.3. To this end, we begin by stating the following simple result.
70 2- Closed Convex Processes Proposition 2.5.6 Let X,Y be Banach spaces and F : X ~> У be a process. Then (Im(F))- = -F*~\0) к F(0) = (Dom(F*))+ Therefore, if F is a convex process, the image of F is dense if and only if the kernel F* (0) of its transpose is equal to 0. Furthermore a convex process F is surjective if F* (0) = {0} and either the dimension of Υ is finite or the image of F is closed. Proof — To say that q belongs to the polar cone of the image of F amounts to saying that the pair (0, q) belongs to the polar cone of the graph of F, i.e., that (—q, 0) belongs to the graph of the transpose F*, in other words, that 0 belongs to F*(—q). The proof of the second statement is naturally analogous. □ The extension of the Bipolar Theorem 2.4.3 to closed convex processes is then a consequence of Corollary 2.5.5 and Proposition 2.5.6. Theorem 2.5.7 (Bipolar Theorem) Consider Banach spaces Χ, Υ and let F : Χ ~> Υ be a closed convex process, and К С X be a cone satisfying Dom(.F) — К = X. Then (F{K)y = -F*'1 (K+) Proof — We apply Proposition 2.5.6 to the restriction F\k whose image is F(K) and whose transpose is F*(·) + K~ thanks to Corollary 2.5.5. D The above condition is obviously satisfied when the domain of F is the whole space. In this case, we obtain Corollary 2.5.8 Let F : Χ ~> Υ be a strict closed convex process. Then Dom(F*) = F(0)+ and F* is upper hemicontinuous (see Definition 2.6.2 below) with bounded closed convex images, mapping the unit ball to the ball of radius \\F\\. In particular, F*(0) = {0}. The restriction of F* to the vector space Oom(F*) П (Dom(F*)) is single-valued and linear.
2.5. Transposition 71 Proof — We observe that V? € Dom(F*), sup ||p|| < ||F||||g|| p£F*(q) because for all χ € Dom(.F) = X and for all ρ G F*(q), we have, by- definitions of the transpose and the norm of a closed convex process \\p\\ := suPxeB <P,X> < supieB miyeF{x) <q,y> < < supxejB infj,ejF(x) ||e||||y|| = ||ЛНЫ1 Then F* maps bounded sets to bounded sets. Therefore, the cone F*(0) being bounded, is equal to {0}. Since the domain of F is the whole space, the assumptions of Proposition 2.6.4 below are met, so that for all χ £ X, the function q i-> a(F*(q),x) is upper semicon- tinuous. The domain of F* is closed, thanks to the Closed Range Theorem 2.4.4 applied to the projection πγ* from Χ* χ Υ* and the cone Graph(F*). Then Dom(F*) = πγ* (Graph(F*)) is closed because, the domain of F being equal to X, 1т((тгк*)*) - (Graph(F*))~ = X xY This and Proposition 2.5.6 imply that Dom(F*) = F(0)+. If q belongs to both Dom(.F*) and -Dom(.F*), then F*(q) + F*{-q) С F*(0) = {0} Hence F*(q) contains only one element and F*(q) = —F*(—q). D Example: Case of linear processes When F : Χ ~> Υ is a linear process from a Banach space X to another Y, then, from the very definition of the adjoint, it follows that F* is also a linear process from Y* into X*. Hence its domain is a subspace of Y* and, by Proposition 2.5.6, Dom(.F*) = F^)1. In general the space Dom(.F*) is not closed. However this is always the case when Υ has a finite dimension.
72 2- Closed Convex Processes Furthermore Corollary 2.5.8 implies that when F is strict, then Dom(P*) = F(0)± and F* is a linear operator from the subspace Ρ(Ο)1· into X*. When X = Y, a closed subspace Ρ С Dom(P) is called invariant under F if F{P) С Р. We have the following Proposition 2.5.9 Assume that Dom(P*) is closed. If a closed sub- space Ρ is invariant under F, then its orthogonal space P± is invariant under F*. Consequently, if both Dom(P) and Dom(P*) are closed, then a closed subspace Ρ is invariant under F, if and only if its orthogonal space P^ is invariant under F*. Proof— Let Ρ С Dom(P) be an invariant closed subspace. Then P(0) С Р and therefore P± С P(0)± = Dom(P*)±± = Dom(P*) Fixing q € P1 and ρ € F*(q), for every χ € Ρ, у & F(x) С Р we have <p,x> = < q,y > = 0 Thus ρ € Ρ1- and we have proved that if a subspace Ρ is invariant under F, then its orthogonal P-1 is invariant under the adjoint process F*. Since F = F** the last statement follows. Π For every χ £ X define recursively F1{x) = F{x) and for every integer η > 1, Fn+1{x) = F(Fn(x)) Then for every n, Fn(0) is a subspace of X and Pn(0) С Pn+1(0) Consider the subspace Q = U ^п(°) π>1
2.5. Transposition 73 Clearly F(Q) С Q. If F is strict, then it is Lipschitz on X. Thus F(Q) С Q and therefore in this case Q is the smallest closed subspace of X invariant under F. We also have the following implication Fk(0) = Fk+1(0) ==» VreeN, Fk(0) = Fk+n(0) (2.11) In particular this yields that when X = Rn, then Q = Fn(0). Indeed in this case {Fk(0)}k>i is a nondecreasing sequence of subspaces of Rn satisfying (2.11.) The dimension of Rn being equal to re, the last claim follows. Thus .Fn(0) is the smallest subspace of Rn invariant under F. О The sum of two closed convex processes or the product BF is not necessarily closed. We have to provide sufficient conditions implying that they are still closed. Proposition 2.5.10 (Closed Graph Criterion) Let Χ, Υ and Ζ be reflexive Banach spaces, F : X ~> Υ and G : Χ ~> Ζ be closed convex processes and В G C(Y, Z) be a continuous linear operator. If B*(Dom(G*)) - Oom(F*) = Y* then the convex process BF + G is closed. Proof — This is a consequence of the Closed Range Theorem 2.4.4 with К := Graph(.F χ G) and the continuous linear operator A defined by A(x, y, z) := (x, By + ζ). Π Proposition 2.5.11 Let X,Y be Banach spaces, G : Χ ~> Υ be a closed convex process and Ρ С X and Q С Υ be closed convex cones. Let us consider the convex process F defined by - .pM ._ ί G{x) + Q ifxeP [ > - \ 0 if χ i Ρ It is closed when we suppose that Dom(G*) + Q- = У* If we assume that Dom(G) — Ρ = X, then its transpose is given by F*(n) ._ ί G*(q)+P- ifq € Q+ {q) " 10 ifqi Q+
74 2- Closed Convex Processes Proof — We observe that F is the sum of the closed convex process G and the closed convex process Η defined by H{x) := Q if χ € Ρ and 0 if χ £ Ρ, whose domain is Ρ and whose transpose is defined by H*(q) := P~ if q € Q+ and 0 if not. Corollary 2.5.4 ends the proof. □ 2.6 Upper Hemicontimious Maps We associate with a set-valued map F from a metric space X to a normed space Υ the family of functions χ >-^ a(F(x),p) := sup <p,y> yeF{x) indexed by the continuous linear functionals ρ £Y*. We observe that Corollary 2.6.1 // a set-valued map F from a metric space X to a normed space Υ is weakly upper semicontinuous and has compact values (resp. lower semicontinuous), then the function (x,q) € XxY* ^ a{F{x),q) is upper semicontinuous (resp. lower semicontinuous). Therefore, it is quite convenient to introduce the following definition. Definition 2.6.2 (Upper Hemicontinuous Map) We shall say that a set-valued map F : Χ ~> Υ is upper hemicontinuous at xq G Dom(.F) if and only if for any ρ G Y*, the function χ н-> a(F(x),p) is upper semicontinuous at xq. It is said to be upper hemicontinuous if and only if it is upper hemicontinuous at every point ofOom(F). Proposition 2.6.3 The graph of an upper hemicontinuous set-valued map with closed convex values is closed. Proof— Let us consider elements (хп,Уп) € Graph(-F) converging to a pair (x,y). Then, for every ρ €Υ*, <p,y>= lim <p,yn>< limsupa(F(xn),p) < a(F(x),p) η~>0° 71-ЮО
2.6. Upper Hemicontinuous Maps 75 by the upper semicontinuity ofiM a(F(x),p). This inequality implies that у € F(x) since these subsets are closed and convex, thanks to the Separation Theorem 2.4.2. We have shown that (x, y) belongs to Graph(.F), which ends the proof. D It is useful to compare the support functions of F(x) and F*(q). Proposition 2.6.4 (Support Function of the Transpose) Let Χ, Υ be Banach spaces and F : Χ ~> Υ be a closed convex process. Then for every χ G X and q £ Y* a(F*(q),x) + a(F(x),-q) < 0 Furthermore for every xQ in the interior of the domain of F there exists po € F*(q0) such that < pQ, x0 > is equal to the common value a{F*(q0),x0) = -а{Р{хо),-до) (2.12) In the same way, if qQ belongs to the interior of the domain of F*, there exists y0 € F(x0) such that < q0,y0 > is equal to the common value (2.12.) Therefore for every χ € Dom{F), the function q н-> a(F*(q),x) is upper semicontinuous on the interior of the domain of F* and for every q £ Dom{F*) the function χ н-> a(F(x), -q) is upper semicontinuous on the interior of the domain of F. Proof— The first claim follows from the very definition of the support function. Denoting by π χ the projector from Χ χ Υ to X, we can write a(F(x0),qo) = a(GT&ph(F)mrx1(x0),(0>qo)) We apply the formula (5) of Table 2.1 on the support functions of an intersection and inverse image with L := Graph(.F), Μ := {xq} and A := πχ. The constraint qualification assumption is satisfied because 0 belongs to the interior of KX(GT&ph(F)) - x0 = Dom(F) - x0 by assumption. Then a(F(x0),qo) = M{<p,xQ> +a(Graph(F),(0,g0)-7Γ^(ρ)))
76 2- Closed Convex Processes We observe that πχ(ρ) = (ρ,Ο) and that a(Graph(F),(-Pi?o)) = 0_ if and only if ρ € i?*(^o)) so that (2.12) ensues. Since the function —a(F(xo), ■) is upper semicontinuous, the proof of the last claim ensues. □
Chapter 3 Existence and Stability of quilihrium Introduction This chapter is devoted to the major questions of nonlinear functional analysis: provide criteria allowing to solve1 a (nonlinear) equation or an inclusion (with or without constraints) and, also, approximate their solutions and study their stability. We tried to present the shortest introduction to this so important field by choosing only the theorems we judge the most powerful, i.e., the theorems from which we obtain the most results with the least efforts. Historically, nonlinear analysis is based on two fixed point theorems: » Banach-Picard Successive Approximation Theorem for problems set on complete metric spaces β Brouwer Fixed Point Theorem in a convex compact environment Using these tools amounts to transforming the problem under investigation to a fixed point problem, for which one can apply one of 1 Apparently Cauchy was the first to prove an existence theorem, instead of assuming a priori that the solution exists and compute it, risking in this way to obtain paradoxes. 77
78 3- Existence ала Stability of ал Equilibrium these fixed point theorems or their numerous variations or extensions. The "fixed point" format being quite rigid, one encounters the risk of more or less considerable loss of information by doing so. It also often happens that this transformation may require additional assumptions and useless technical difficulties. This is the reason why so many statements logically equivalent to the Brouwer Fixed Point Theorem — constituting the corpus of nonlinear analysis — have been designed to be readily adapted to classes of specific problems. Still, among these equivalent results, one can look for those statements which, in some sense, incorporate more labor-value2, and therefore, might be more useful. This is the case of the Ky Fan and Ekeland Theorems, which provide less elegant but more efficient statements than the corresponding Fixed Point Theorems. These two more flexible and powerful theorems, equivalent to each of the above Fixed Point Theorems, appeared at the beginning of the seventies: β The Ky Fan Inequality, in 1972 β Ekeland's Variational Principle, in 1974. We thus begin by deducing Ky Fan's Inequality from the Brouwer Fixed Point Theorem in the first section, and use it to derive in Section 2 not only the traditional fixed point theorems (including the Kakutani-Fan Fixed Point Theorem for set-valued maps), but also a Constrained Equilibrium Theorem, providing an equilibrium 2Although two statements V and Q may well be equivalent, it is common experience that the proof of one of the implications, say V => Q, is more difficult or involves deeper results than the proof of the converse. In this case, one can say that Q incorporates more labor value than V and thus, expect as a general rule that the statement Q may be more useful than V. One could still follow Aristotle's observation that most theorems of geometry have valid converses, so that one can go leisurely from V => Q (called analysis) to Q =>■ V. However, more serious greek mathematicians, conscious of the pitfalls, used to add supplementary conditions (called diorismoi.) The reverse inference was called synthesis. By the way, the most pleasant or intuitive statements are quite often the ones with the least labor value.
3.0. Introduction 79 ж of a set-valued map F (i.e., a solution to the inclusion 0 € F(x) satisfying the constraints described by χ € К.) Besides standard mild conditions on F, we shall assume that К is a convex compact viability domain of F; this means that we can find at every point χ of the compact convex subset К an element ν € F (x) tangent to К at x. We then prove Ekeland's Variational Principle in Section 3, a result we shall use extensively in this book, above all for establishing the calculus of tangent cones, of derivatives of set-valued maps and epiderivatives of extended functions in the forthcoming chapters. But we start to apply it in Section 4 to extend Graves' Inverse Function Theorem to constrained problems of the form: К being a closed subset describing the constraints, look for a solution xq G К to the equation f(xo) = yo- In the unconstrained case, Graves' Theorem states that if / is C1 at xq and if f'(xo) is surjective, then the equation f(x) = у still has a solution for any right-hand side у close enough to f(xo), and the set-valued map /-1 behaves in a Lipschitz way. When constraints requiring that χ belongs to К are present, we "differentiate" them by using the contingent cone Тк{х) to К at x, which is a most reasonable way to translate the concept of tangency for an arbitrary subset. Roughly speaking, if /'(·) is continuous at xq, if Τκ{·) is lower semicontinuous at xq and if /'{хо)(Тк(хо)) is the whole space, then the equation f(x) = у has a solution χ € К for any у in a neighborhood of f(xo). This stability property, often regarded as an existence theorem of a solution to nonlinear problems, is actually a particular case of an approximation procedure, and conceals the famous Lax Principle: {Consistency and Stability imply Convergence for approximating solutions to given problems, which underlies numerical functional analysis. The constrained problem is approximated by the following constrained problems: find xn € Kn satisfying fn{xn) = Уп
80 3- Existence and Stability of an Equilibrium Consistency means that in some sense, fn converges to / and Kn to K. Stability means that linear equations find un € TKn{xn) satisfying f^(xn)un = vn have solutions satisfying the stability condition ||ιιπ|| < c||wn|| and Convergence means that the convergence of the right-hand sides yn to ?/o implies the convergence of approximate solutions xn to xq. These theorems will be extended to inclusions as soon as we have defined the derivatives of set-valued maps in Chapter 5. We describe in Section 5 the basic results on monotone and maximal monotone maps, which share many properties with continuous linear operators. Maximal monotone maps can be approximated by single-valued Lipschitz maps (Yosida approximations), a feature that is very useful3. We provide in Section 6 a result on existence of eigenvalues and eigenvectors of a closed convex process. 3.1 Ky Fan's Inequality After what we said in the introduction, we state and prove the Ky Fan4 inequality. Theorem 3.1.1 (Ky Ean Inequality) Let К be a compact convex subset of a Banach space5 X and φ : Χ χ X \—t R ie α function satisfying i) У у € К, х 1-^><р(х,у) is lower semicontinuous < ii) Ух £ К, г/(-► (^(ж,?/) is concave (3.1) ^ in) У у € К, ip{y,y) < 0 Then, there exists χ G K, a solution to У у € Κ, φ{χ,ν) < 0 (3.2) 3We shall use it in Chapter 9 to construct regular selections of set-valued maps. 4A student of Prechet. He visited Paris for one year in 1939, with a small grant and the map of Paris' metro. He then had to survive there during the whole war. 5Actually, the proof we give shows that this theorem remains true for any Hausdorff locally convex topological vector space and in particular, when the Banach space X is endowed with the weak topology.
3.1. Ky Fan's Inequality- Si Proof — We shall prove it first in the finite-dimensional case, from which we shall derive the proof in Banach spaces. — The finite-dimensional case. We shall derive a contradiction from the negation of the conclusion: V χ £ K, 3y € К such that φ(χ, у) > О so that К can be covered by the subsets Vy := {x € K\ φ(χ, у) > 0} which are open by assumption (3.1) г). Since К is compact, it can be covered by η such open subsets Vyi. Let us consider a continuous partition of unity6 (скг)г=1,...,п associated with this open covering of К and define the map / : К н-► X by η Ух £ К, f(x) := Y^ai(x)yi г=1 It maps К to itself because К is convex and the elements yi belong to K. It is also continuous, so that Brouwer's Fixed Point Theorem implies the existence of a fixed point у = f(y) G К of /. Assumption (3.1) ii) imply that π π г=1 г=1 Let us introduce I(V) ■= {» = l,...,n| a{(y) > 0} It is not empty because Σά-ι щ{у) = 1. Furthermore η Y^ai(y)<p(y,yi) = Σ ац(у)<р(у,у{) > 0 г=1 iel(y) 6 A continuous partition of unity associated with a covering of К by η open Subsets Vi is a sequence of η continuous maps са : К н-> R such that, η Vie Jf, ^аг(ж) = 1, Vi = l,...,n, m(x) > 0 & support(ai) С Vi Such continuous partitions of unity do exist when К is a compact metric space.
82 3- Existence and Stability of an Equilibrium because, whenever г belongs to I(y), щ{у) > 0, so that у belongs to Vyv and thus, by the very definition of this subset, if{y,yi) > 0. Hence, we have proved that ip(y, y) is strictly positive, a contradiction of assumption (3.1) Hi). — The infinite-dimensional case. Let us introduce the family S of finite subsets Μ := {у ι,... ,ym} С К of К and the number υ := sup inf max (p{x,yi) This number is nonpositive. Indeed, the function φΜ (λ, μ) defined on Sm χ Sm (where the simplex Sm is a subset of elements (Ab ..., Xm) ε Rm such that Xi > 0 and Υ^χ λί = 1) by m η ψΜ{μ,Χ) ■■= Σ)μί¥Έλί2/ί>2/.ί) ί=1 ί=1 satisfies the assumption of Ky Fan's Theorem in finite dimension (it is obviously lower semicontinuous with respect to λ and linear with respect to μ, and ψΜ(μ,μ) < 0 because φ is concave with respect to у and by assumption (3.1) iii). Hence there exists λ £ Sm such that m У μ £ 5m, J2^M^Vj) = ΨΜβ,μ) < 0 where we set χ := ]Γ^ι λΐί/ΐ e co(M) С К. We then deduce that τη vM ■■= inf max <p(x,yj) < max ^(ϊ,^·) < sup У^МФ&Уз) ^ ° xeKyjeM yjEM ues™rr J = l and thus, that ν = supMeS vm is nonpositive. It remains now to prove the existence of ж £ if such that sup φ(χ, у) < ν уек This results from the following Lemma 3.1.2 Let К С X be α compact subset of a topological space X, L be any subset and φ : Κ χ L н-> R a function such that χ н-> φ(χ, у) is lower semicontinuous for every у £ L. Let S be the family of finite subsets of L. Then there exists χ £ К satisfying sup</j(5, y) < υ := sup inf max φ(χ, yj) yeL MesxeKvoEM
3.2. Equilibrium and Fixed Point Theorems 83 Proof — We introduce the subsets Sy := {χ ε Κ Ι ψ{χ,ν) < ν) They are nonempty closed subsets of the compact К because </?(·, у) is lower semicontinuous. The very definition of υ implies that these subsets enjoy the finite intersection property: indeed, consider η such subsets Sy. and set Μ := {г/ι,. · · ,yn}· The compactness of К and the lower semicontinuity of </?(·, y) implies the existence of an element xM ε Κ minimizing the function тах^емУ^Уг)) which then satisfies max φίχΜ,νί) = inf max φ(χ,уЛ < ν ViEM xeKyiEM so that it belongs to Γ)Γ=ι ■%/; ■ Hence there exists some χ in their nonempty intersection f]yeK Sy, which is a solution to the desired inequality. D 3.2 Equilibrium and Fixed Point Theorems 3.2.1 The Equilibrium Theorem We are ready to derive from the Ky Fan inequality the Constrained Equilibrium Theorem. It provides sufficient conditions for an upper semicontinuous set-valued map F from a Banach space X to closed convex subsets of X to have an equilibrium satisfying constraints, i.e., a solution Τ to the inclusion 0 € F(x) required to belong to a given closed subset K. Such a set К defining the constraints will be assumed to be at least convex and compact, properties which are usually demanded for any existence theorem derived from the Brouwer Fixed Point Theorem7. But we have to provide a further assumption which links the set К with the set-valued map F: we shall require that К be a viability domain of the set-valued map F. We recall that the tangent cone Tk(x) to a convex subset К at χ € К is the closed cone spanned by Κ — χ, which is convex: im*)= UK~X h>o h 7up to the reduction to this case, using coerciveness to trade-off with compactness, as we shall see at the end of this section, for instance.
84 3- Existence and Stability of art Equilibrium (see Section 2 of Chapter 4 for a detailed exposition of tangent cones to convex subsets.) A subset К С Dom(F) is said to be a viability domain of F if and only if Vx€K, F{x)nTK(x) φ 0 This means that for any point χ € AT, there exists at least a direction υ € F(x) which is tangent to К at x. We shall see in Chapter 10 that this means also that from any initial state xq € K, there exists a solution x(-) to the differential inclusion x' € F(x), x(0) = x0 which is viable in К in the sense that x(t) € К for any t > 0. Theorem 3.2.1 (Equilibrium Theorem) Assume thatX is a'Ba- nach space8 and that F : X ~> X is an upper hemicontinuous set- valued map with closed convex images. If К С X is a convex compact viability domain of F, then it contains an equilibrium of F. Proof — We proceed by contradiction, assuming that the conclusion is false. Hence, for any χ € К, 0 does not belong to F(x). Since the images of F are closed and convex, the Hahn-Banach Separation Theorem implies that there exists px € X* such that a{F{x),px)<0. By setting Vp := {x €K\ a{F{x),p) < 0} the negation of the existence of an equilibrium of F in К implies that К can be covered by the subsets Vp when ρ ranges over the dual of X. These subsets are open by the very definition of upper hemicontinuity of F. So К can be covered by η such open subsets VPi. Let us consider a continuous partition of unity (αί)ί=ι,...,π associated with 8 Actually, the proof we give shows that this fundamental theorem remains true for any HausdorfF locally convex topological vector space and in particular for a Banach space endowed with the weak topology.
3.2. Equilibrium and Fixed Point Theorems 85 this finite open covering and introduce the function φ : Κ χ Κ —* R defined by η Ч>{х,У) ■= Y^Oii{x)<Pi,x-y> i=l Being continuous with respect to χ and affine with respect to y, the assumptions of Ky Fan's Inequality (Theorem 3.1.1) are satisfied. Hence there exists χ € К such that for ρ ·.= Σ?=1 oii(x)pi we have Vy€K, φ{χ,ν) = <p,x-y > < 0 The above inequality means that —p belongs to the polar cone Τκ{χ)~ of the convex subset К at x. Since К is a viability domain of F, there exists υ € F(~x) Γ\Τκ(χ), and thus a(F{x),p) > <p,v > > 0 We set I(x) := {i = l,...,n\ oti(x) > 0} which is not empty. Hence a(F(x),p)< Σ <*i@MF@),Pi) < 0 because, for any i € I(x), щ{х) > 0, and thus, χ belongs to the subset VPi, which means precisely that a(F(x),pi) < 0. The latter inequality is then a contradiction of the previous one. □ By modifying slightly the proof of the Equilibrium Theorem, we can prove the existence of zeros of a set-valued map from a Banach space X to another Banach space Y. Theorem 3.2.2 (Solvability Theorem) Let К be α convex compact subset of a Banach space X and F be an upper hemicontinuous set-valued map with closed convex values from К to another Banach space Y. Let us consider also a continuous map В : К н-> £(X, Υ). If К, F and В are related by the condition Vx(=K, F(x) П B(x)TK(x) Φ 0 then (i) 3 χ ε Κ such that 0 ε F{x) ii) My ε Κ, 3χ ε Κ such that B(x)y ε B(x)x + F(x)
86 3- Existence ала Stability of ал Equilibrium Proof — The proof of the existence of ал equilibrium χ £ К of F is the same as the one of the Equilibrium Theorem, where we define the function φ by η Ч>(х,У) ·"= Σ<*ί{χ) <Pi,B(x)(x-y) > Ky Fan's Inequality thus implies the existence of χ € К such that -B{x)*p € TK(x)~ so that, taking a sequence un £ Τχ{χ) such that B(x)un converges to some ν £ F(x)nB(x)TK{x) (which exists by the tangential condition), we infer that f a(F(x),p) \ > <p,v> = limn-^oo < P, B{x)un > = limn-^oo < B(x)*p,un > > 0 This inequality is contradicted as in the proof of Theorem 3.2.1. Take now у € if and introduce the set-valued map G : X-^Y defined by G{x) := F(x) + B{x)(x - y) which also satisfies the assumptions of our theorem. Then there exists a zero χ £ К of G, which is a solution to the inclusion B(x)y £ B(x)x+F(x). Π 3.2.2 Fixed Point Theorems These theorems are equivalent to the Brouwer Fixed Point Theorem; we shall prove here only this equivalence9. We begin by showing that Theorem 3.2.1 implies the Kakutani Fixed Point Theorem10, which is the set-valued version of the Brouwer 9See [45, Appendix B] for a proof of the Brouwer Fixed Point Theorem based on Sperner's Lemma and [35, Chapter II] for a proof based on differential geometry. 10called Ky Fan's Fixed Point Theorem in infinite dimensional spaces. The story began in 1910 with the Brouwer Fixed Point Theorem, which was proved later in 1926 via the Three Polish Lemma, the three Poles being Knaster, Kuratowski and Mazurkiewicz. Knaster saw the connection between Sperner's Lemma and the fixed point theorem, Mazurkiewicz provided a proof corrected by Kuratowski. The extension to Banach spaces was proved in 1930 by their colleague Schauder. Von Neumann did need the set-valued version of this Fixed Point Theorem in game theory, which was proved by Kakutani in 1941.
3.2. Equilibrium and Fixed Point Theorems 87 Fixed Point Theorem. Theorem 3.2.3 (Kakutani Fixed Point Theorem) Let К be а convex compact subset of a Banach space X and G : Χ ~> Κ be an upper hemicontinuous set-valued map with nonempty closed convex values. Then G has a fixed point11 χ € if. Π G(x). Proof— We set F(x) := G(x)—x, which is also upper hemicontinuous with convex values. Since К is convex, then К — χ С Тк{х), and since G(K) С К, we deduce that К is a viability domain of F because F{x) С K-x С Тк{х) Hence there exists a viable equilibrium χ € К of F, which is a fixed point of G. □ Actually, we do not need to assume that G maps К to itself. It is enough to assume that if is a viability domain of F := G — 1, which can be written in the following form \/x<=K, G{x)n(x + TK{x)) φ 0 (3.3) This leads to the following Definition 3.2.4 (Inward L· Outward Maps) A map G : Κ'~> Χ satisfying property (3.3) is said to be inward. It is called outward if Vx&K, G(x)n(x-TK{x)) φ 0 Since if is a viability domain of F := G — 1 when G is inward and of Л := 1 — G when G is outward, and since the equilibria of F and F_ are fixed points of G, we obtain the useful о Theorem 3.2.5 Let К be a convex compact subset of a Banach space X and G : Κ ~> Χ be an upper hemicontinuous map with nonempty closed convex values. If G is either inward or outward, it has a fixed point ~x £ Kr\G{x) nwhich can be regarded as an equilibrium for the discrete set-valued dynamical system xn+i e G{xn).
88 3- Existence and Stability of an Equilibrium Since the Kakutani Fixed Point Theorem implies obviously the Вгошгег Fixed Point Theorem, we have proved that all these statements are equivalent. □ Remark — We can deduce from the Viability Theorem and other properties of differential inclusions two other results on the existence of an equilibrium of a set-valued map F. We shall mention them in Chapter 10. Π We now provide a convenient sufficient condition implying that a set- valued map G is outward. We say that δσκ(ρ) := {x e К | <p,x>= σκ{ρ)} is the extremal face of К at p. Observe that χ ε δσκ(ρ) if and only if peTK(x)~. Proposition 3.2.6 Let К be a nonempty closed convex subset of X and let G be a set-valued map from К to К satisfying \/p£X*,Vx£daK(p), G(x)ndaK(p) φ 0 Then G is outward. In particular, this is the case of a map G satisfying Vpel*, G(daK(p)) с δσκ(ρ) Proof — Let χ ε Θσκ(ρ) and Уо е G{x)f\daK{p) be fixed. Hence, ν = ya — χ belongs to G(x) — χ and — Тк(х) because, Vpe TK(x)~, <p,v> = <p,y0 > - <p,x> = σκ{ρ)-σκ{ρ) =0 D We can relax the assumption that К is compact, replacing it by a coerciveness assumption. Theorem 3.2.7 Let К be a closed convex subset of a finite dimensional space and let F be an upper hemicontinuous map with nonempty compact convex images satisfying the coerciveness assumption limsupCT(F(a;),a;) < 0 (3.4) ll«ll-.oo хек We posit the tangential condition \fx £ К, F{x)r\TK{x) φ 0
3.2. Equilibrium and Fixed Point Theorems 89 Then there exists an equilibrium χ £ К of F. If we posit the stronger coerciveness assumption, a(F(x),x) hmsup v I (' ; = -oo (3,5) IMI—oo \\X\\ хек Then for all у £ К, there exists a solution χ ς. Κ to the inclusion у £ x-F{x). Proof— Coerciveness assumption (3.4) implies that there exist e > 0 and a > 0 such that sup a(F(x),x) < -e < 0 INI>« хек This implies that for all a; £ if with ||ж|| < a, F(x) с Тав{х). By taking о the number a large enough so that Κ Π α Β^ 0, we know that Ткпав{х) = TK(x)nTaB{x) (see Chapter 4, Section 2.) So the tangential condition implies that УхеКПаВ, F{x)nTKnaB{x) φ 0 It is sufficient to apply Theorem 3.2.1. To prove the second part of the theorem, we fix у £ if and replace F by the map G defined by G{x) = F{x) + y-x which satisfies the coerciveness assumption (3.4) whenever F satisfies the stronger coerciveness assumption (3.5.) Π 3.2.3 The Leray-Schauder Theorem We can deduce the set-valued-version of the Leray-Schauder Theorem12 on the existence of stationary points from Theorem 3.2.1 by Poincare's continuation method. We take X = Rn and К to be a compact convex subset with a nonempty interior, so that the boundary 8K = K\ Int(UT) of К is distinct from K. proved in 1934,
90 3- Existence and Stability of an Equilibrium Theorem 3.2.8 (Leray-Schauder) Let us consider α compact convex subset К С Rn with nonempty interior and an upper hemicontin- uous set-valued map F from [0,1] x К to Rn, with nonempty closed convex values. Suppose the set-valued map χ —» F(0, x) satisfies the tangential condition \/x € dK, F(0,x)nTK(x) φ 0 (3.6) and Vx € dK, VA € [0,1[, 0 i F(X,x) (3.7) Then there exists χ € К such that 0 €'F(l,x). Proof — We suppose that the conclusion is false and derive a contradiction. We set N := dK, a closed subset of К and introduce the subset Μ := {χ € K\ 3 λ € [0,1] satisfying 0 € F(\,x)} The subset Μ is nonempty because it contains a solution χ € К to the inclusion 0 € F(0,x), which exists by Theorem 3.2.1 (thanks to assumption (3.6).) It is closed, since the graph of F is closed. The intersection Μ Π Ν is empty : If χ € N and if λ € [0,1[, assumption (3.7) implies that χ φ Μ. Consider a continuous function ψ from К to [0,1] ( л = ^>AQ Ψ{Χ) ' d{x,M) + d{x,N) which is equal to zero on N and to one on Μ and define the set-valued map G on К by G{x) := F(<p(x),x) Then G is clearly upper hemicontinuous with nonempty closed convex values. It coincides with F(0, ■) on N = dK and, consequently, satisfies the assumptions of Theorem 3.2.1. Hence, there exists a solution χ € К to the inclusion 0 € G{x) = Ρ{ψ{χ),χ) But this implies that χ € Μ and, therefore, ψ(χ) = 1, so that 0 € F(l, x) which is a contradiction. □
3.3. Ekeland's Variational Principle 91 3.3 Ekeland's Variational Principle Ekeland's Theorem, we are about to prove, provides an approximate minimizer of a bounded from below lower semicontinuous function in a given neighborhood of a point. This localization property is very- useful and explains the importance of this result which has been used extensively since its discovery in 1974. Theorem 3.3.1 (Ekeland's Variational Principle) Let y:lHRu{+oo} be α lower semicontinuous nontrivial extended bounded from below function defined on a complete metric space X. Let xq € Dom(V) and ε > 0 be fixed. Then there exists χ € X, a solution to (i) V{x) + ed{x0,x) < V{x0) (3.8) ii) Vi^i, V{x) < V(x) + ed(x,x) Proof— We can take e := 1 and assume that V is non negative by replacing it by V(-) — inf V if needed. We associate with V the set-valued map F : X ~> X defined by F(x) := {y € X | V(y) + d(x,y) < V(x)} The subsets F(x) are closed and F is a preorder (given by у У- х Ч=> у € F{x)) in the sense that г) Vi e Dom(V), χ € F(x) (reflexivity) ii) у € F{x) =^ F{y) С F(x) (transitivity) Indeed, the second condition is obvious when χ fi Dom(V), since in this case F(x) = X. Assume now that V{x) is finite. By taking у € F(x), ζ € F(y), adding inequalities V(z) + d(y,z) < V{y) L· V(y) + d(x,y) < V(x) and using the triangular inequality, we obtain V(z) + d(z, x) < V(x), i.e., ζ € F(x).
92 3- Existence and Stability of an-Equilibrium Let us define the function g on Dom(V) by -~Vi/"€ Dom(V), 3(y) :=" inf V(z) It is clear that VyeF(x), d(x',y) < V{x)-g{x) so that the diameter supy ζς.ρ{χ) Ну — z\\ °f F(x) is no^ greater than 2(V(x)-g(x)). Define now the following sequence starting at xq: we take xn+i € F(xn) such that V(xn+1) < g(xn) + 2"n Since F(xn+i) С F{xn) by transitivity, we deduce that g(xn) < gfan+i) On the other hand, inequality g{y) < V(y) implies that g{xn+i) < V{xn+i) < g{xn) + 2~n < g{xn+i) + 2~~n and thus, that 0 < V(xn+1) - g(xn+1) < 2-n Therefore, the diameters of the closed subsets F(xn) converge to 0. Since the sequence of these closed subsets is nonincreasing and since the metric space is complete, we infer that Π F(xn) = {x} 7l>0 This implies that χ € F(xo), i.e., inequality (3.8) г). We also observe that F(x) = {x}. Indeed, χ belonging to all the subsets F(xn), then, by transitivity, F(x) is contained in all the ir(xTl)'s, and thus, in {x}. Consequently, when χ φ χ, then χ does not belong to F(x), i.e., V{x) + d{x,x) > V(x) This is inequality (3.8) и). ■ □
3.4. Constrained Inverse Function Theorem 93 3.4 Constrained Inverse Function Theorem 3.4.1 Derivatives of Single-Valued Maps We recall first some classical definitions and notations. Definition 3.4.1 Let Ω be an open subset of a normed space X and f-.П ι-+ Υ be a single-valued map from Ω to a normed space13 Y. We denote by Vhf(x):v - {f{x + hv)-f{x))/h 13The concepts of functional and of derivatives of functionals are due to Volterra in 1887. Then Gateaux, in a note written in 1913 and published in 1919 after his death during the First World War, introduced the concept of first variation Sf{x):=\-^-f(x + X6x) L d\ Ι λ=0 As we can see, Gateaux's definition did not involve linearity with respect to the increment — long before our time when we are at last ready not to require linearity anymore in order to go forward — and for that, was criticized by Hadamard and his former student, Frechet. Mathematicians of this period still asked many properties for the derivatives of functionals and were not ready to give them away. But Gateaux was forgiven, and even rewarded posthumously since linearity was introduced in the definition of the Gateaux derivative! Frechet proposed his concept of derivative (with the mandatory linearity) for a function as early as in 1912 in the case of functions and in 1925 for maps from a normed space to another one. However, he did not choose the stronger concept (called strong or strict differentiability) derived from a definition of the derivative of a function that Peano suggested in 1892 f{x) := Hm /fri)-/fo) on the ground of convincing physical motivations. This track was not used until 1961 (by S. Leach to give an improved statement of the Inverse Function Theorem.) However, Peano's point of view on how to take limits is very much the one which underlies the concepts of Clarke tangent cones and paratingent cones which will be studied in Chapter 4. Since then, the history of derivatives in Banach spaces is a kind of mathematical striptease, where one by one, and very shyly, the required properties of the derivative of a functional are taken away. We shall go quite far in Chapter 6 to leave the derivatives with the bare minimum. This led to a menagerie of concepts: strong or weak Frechet and Gateaux derivatives, Hadamard, bounded (Suchomlinov), locally uniform (Vainberg) derivatives, directional semiderivatives or derivatives from the right, Sindalowsky
94 3-—Existence and Stability of an Equilibrium the differential quotient of / at χ in the direction v. If the differential quotients V/i,/(:e)(i/) converge to a limit denoted by df(x)(v)~, then df(x)(v) is called the Dini directional derivative of f at χ in the direction v. If the map υ ι—» df(x)(v) is linear and continuous, we say that f is Gateaux differentiable at χ and the continuous linear operator f'{x) : υ н-» df{x){v) =: f'{x)v is called the derivative of f at x. When f is real-valued, we say that f'(x) € X* is the gradient of f at x. If furthermore Шп /Ы-/(я)-/'0с')О/-и) = 0 y->x \\y — x\\ then f is said to be Prechet differentiable at x. When the derivative у ι—» f'(y) is continuous at x, f is called continuously differentiable (or of class C1) at x. Recall that if / is continuously differentiable, it is Prechet differentiable and when / is Prechet differentiable, it is continuous. Naturally, there are many other ways to take limits14 of the differential quotients, using the weak topologies of either X or Υ or both, taking limsup's or liminf's, etc. 3.4.2 Constrained Inverse Function Theorems Let us consider now a Banach space X, a normed space Y, a closed subset К С X and a continuous (single-valued) map f : К i-^Y. or Muravev variations, and the ones we shall study in Chapter 6 and quote in the bibliographical comments of this chapter. We shall face a similar situation in the set-valued case. This is quite natural, though, because each problem demands its own amount of properties that the derivative should enjoy, i.e., its own degree of regularity. Without going too far by always requiring minimal assumptions, some problems could not be solved by sticking to the richest structure. The right balance between generality and readability is naturally a subjective choice. 14too many of them having names of French mathematicians..., according at least to some American ones.
3.4. Constrained Inverse Function Theorem 95 We shall deduce from Ekeland's Theorem a "nonlinear extension" of Graves' Theorem 3.4.2, which allows one to solve (locally) a "constrained problem" of the form find χ € К such that f(x) = у for any right-hand side у in a neighborhood of some У0 := f(xo) € /(IT) When К is equal to the whole space X, a solution to this unconstrained problem is given by Graves' version of the Inverse-Function Theorem15: Theorem 3.4.2 (Graves' Theorem) Let Χ, Υ be Banach spaces and f : Χ ι—> Υ a continuous (single-valued) map. We assume that f is continuously differentiable on a neighborhood of xq and that f'{xo) is surjective Then the set-valued map у ~»· /-1(y) is pseudo-Lipschitz around {f{xo),xo). Proof — It is an obvious consequence of Theorem 3.4.10 below with К := X. □ For constrained problems, i.e., when К is no longer the whole space, we need to introduce right now the concept of contingent cone, the exhaustive study of which is the purpose of Chapter 4. The contingent cone Тк{х) to К С X at χ is the upper limit of the subsets {K — x)/h when h —> 0+: ν € Tk(x) if and only if liminf — , V) = 0 /i-»o+ h Graves' Inverse-Function Theorem can be extended to our constrained problem provided that an adequate transversality condition holds true: 15This result is sometimes called in the literature Ljusternik's theorem. However Graves was the first to state it in 1950,, while Ljusternik investigated in 1934 tangent spaces to level sets of a function. Both Graves and Ljusternik used in their proofs an approximation procedure which is a standard technique of functional analysis.
96 3- Existence and Stability of an Equilibrium Theorem 3.4.3 (Constrained Inverse Function Theorem) Lei X be α Banach space and Υ be a normed space. We consider a (single-valued) continuous map f : X *-*Y, a closed subset К С X and an element xq of K. We assume that f is Frechet differentiable on a neighborhood of xo and we posit the following transversality assumption: there exist constants с > 0, a € [0, l[ and η > 0 such that Vx € ΚΠΒ(χ0,η), Βγ С f'(x)(TK(x)ncBx) + aBY Then f(xo) belongs to the interior of f(K) and the set-valued map у ~» /~1(y) Π Κ is pseudo-Lipschitz around (f(xo),xo). This theorem is a consequence of the still more general Theorem 3.4.5 below: Indeed, not only are we interested in knowing whether a solution to the constrained problem does exist, but we wish to approximate it by solutions to the approximate problems find xn € Kn such that fn{xn) = Уп where yn converges to yo, fn converges to / and Kn to К in the following sense: Definition 3.4.4 We say that the maps fn are consistent with f on Kn at xq Ε Κ if there exist elements XQn € Kn converging to xq such that fn{%0n) converges to f(xo). We have to add to this consistency assumption a stability assumption to be able to approximate the solution of the constrained problem: Theorem 3.4.5 Let X be a Banach space and Υ a normed space. Consider a sequence of continuous single-valued maps fn from X to Y, a sequence of closed subsets Kn С X and an element xq in the lower limit of the subsets Kn such that the maps fn are consistent with f on Kn at xq. We assume that fn are Frechet differentiable on a neighborhood of xq and verify the following stability assumption:
3.4. Constrained Inverse Function Theorem 97 there exist constants с > Ο, a E [0, l[ and η > 0 suc/i that \/x € ЛГпПБ(хо,?7), -Ву С /;(х)(Г/сп(х)ПсВл:) + аВу (3.9) JTien ί/iere exist I > 0 and 7 > 0 such that for any xqu € Вкп(хо, η) satisfying fn{xQn) € B(f(xQ),j), we have Vyn € J3(/(so),7)i ^(^./„"'(ynjnlfn) < i||l/n-/n(a;on)|| When the functions fn = f and the subsets ΛΓπ ξ Κ are const ant, the conclusion of the theorem means simply that /-1 is pseudo- Lipschitz at (f(xo), xo), so that Theorem 3.4.3 follows obviously from it. Before proving the above theorem we show that it implies Theorem 1.2.9 from Chapter 1. Proof of Theorem 1.2.9 — Let uq € /-1 ( Liminf^ooMn): There exists a sequence of elements vqu € Mn converging to f(uo). We shall construct a sequence of elements un € f~1(Mn) converging to ?xo thanks to Theorem 3.4.5. Indeed, we apply it with Χ χ Υ playing the role of X, the map fQl:{u,v) i-> f(u) - υ playing the role of / and fn, the sets Kn := Χ χ Mn, XQn = (uq, von) and yn = 0. The transversality assumption of Theorem 1.2.9 implies that for every χ € U Βγ С f(x)(dBx)-TM71(y)ndBY (with d = csupieW \\f'(x)\\ + 1) which can be written By С (fei)'(x)(TKn(x)n2dBx) Hence the transversality condition is satisfied with a = 0. Therefore, there exist a constant I > 0 and solutions Xn = (un,vn) € X x"Mn to%(fei){xn) = f{un)-vn = 0 satisfying {IIEn - xon\\ = ||"o —ип|| + ||υοη — vn\\ < J||l/n-(/ei)(son)|| = l\\f(uo)-von\\
98 3- Existence and Stability of ал Equilibrium Hence un Ε f г(Мп) and converges to щ, so that uq belongs to the lower limit of the inverse images /~1(Μπ). □ Proof of Theorem 3.4.5 — We choose ρ > 0, ε > 0 such that So 1-a — < ε < η с and consider elements xon satisfying ||EOn-Eo|| < η/3 & ||/πΟθπ) - 2/0|| < 7?/3 and elements yn G B(fn(x0n),p). By Ekeland's Variational Principle 3.3.1 applied to the function V{x) := \\Vn-Ux)\\ on the complete metric space Kn, we know that there exists a solution xn £ Kn to 0 hn - fn(£n)\\ + e\\xn - xon\\ < \\Уп - fn(xon)\\ < ii) Vi„ e /Q, ||yn -/n(a?n)|| < ||yn-/n(a:n)|| + e||a;n-in|| (3.10) We deduce from inequality (3.10) г) that \\2n-xon\\ < -Ьп - fn(xon)\\ < - < ? ε ε ο so that \\xn - xo\\ < η/З + \\xon - жо|| < 2η/3 Stability assumption (3.9) implies that there exist un Ε Ткп(хп) and wn Ε Υ satisfying г) Уп - fn{xn) = fn{xn)un + wn H) \\un\\ < c\\yn- fn(xn)\\ & ||wn|| < a\\yn - fn(xn)\\ By definition of the contingent cone, there exist elements hp > 0 and ep Ε Χ converging to 0+ and 0 respectively such that, fn being Prechet differentiable, Xn "= Xn τ fipUn -f- π,ρβρ Ε li-n fn(xn) = fn{xn) + hvUn(xn){un + ep) + e(hp))
3.4. Constrained Inverse Function Theorem 99 where e{hv) converges to 0 with hp. By taking in inequality (3.10) it) such an xn, by observing that yn-fn(xn) = (1 ~ hp)(yn - fn(xn)) + hpwn - hp{f'Jxn)ev + e{hv)) we deduce that hp\\yn -/n(2n)|| < ^pll^nll +hp\\fL(xn)ep + e{hp)\\ + ehp\\un + ep\\ Dividing by hv > 0 and letting ρ —* +oo, we get: \\yn - fn{xn)\\ < ||юп|| + е||ип|| < (" + ec)||y„ -/n(in)|| Since we have chosen e such that а + ее < 1, we infer that xn is a solution to xn e Kn & fn{xn) = Уп satisfying -X0n\\ < ~\\Уп— fn(x0n)\\ from which the error estimate follows. Observe also that by letting e converge to c/(l — a), we obtain the estimate dixonJn^y^nKn) < c||y„-/n(a:on)||/(l-aO Π Remark — This theorem extends to nonlinear constrained problems Lax's principle that consistency and stability imply convergence. Due to the importance of this principle in numerical analysis, we reformulate it in this framework: Theorem 3.4.6 Let us consider a sequence of continuous single- valued maps fn from a Banach space X to a normed space Y, a sequence of closed subsets Kn С X and an element x0 £ Liminfn-H.00 ifn such that the maps fn are consistent with f on Kn at xq. Assume that fn are Frechet differentiable on a neighborhood of xo and that stability assumption (3.9) holds true.
100 3- Existence and Stability of an Equilibriwx Then there exists α constant I > 0 such that: if yn converges to f(xo), then xq can be approximated by solutions Xn € Kn to the equation fn{xn) — Уп satisfying the following erroi estimate: Given any "reference " sequence of elements xon € Kn converging to xq such that fn{xon) converges to f(xo)i for all η large enough \\xn - xq\\ < K\\xq - xon\\ + ||/(zo) - /η(^οη|| + \\yn - yo\\) Remark — Case of Linear Operators When we consider the case of continuous linear operators /n := An Ε £{X, Y), we can restrict yn to belong to the subset An(Kn) in the proof of Theorem 3.4.5 and thus, to relax somewhat the stability assumption. Indeed, in this case, 2/n — Anxn £ An[Kn — xn) When xn £ Kn, we denote by SKn{xn) ■ the cone spanned by Kn — xn: Theorem 3.4.7 (Inverse Stability Theorem) LetX be a Banach space and Υ a normed space. Consider a sequence of continuous linear operators An ε £(X, Y) and a sequence of closed subsets Kn С X such that An are consistent with Α ε £{X, Y) on Kn at xq ε Liminfn_00i;sTn. We posit the following stability assumption: there exist constants с > 0, α ζ. [0,1[ and η > 0 such that I ^ Xn ε ΚηΠΒ{χ0,η), { (з.п) ( AnSKn(xn)r\BY С Ап(ТкЛхп)ПсВх) + аВу Then there exist I > 0 and 7 > 0 such that for any x0n £ Вкп(ха,гу) satisfying Anxon ε Β(Αχο,ί) we have ν Уп 6 An {Κη)ΐΛΒ{Αχ0,Ί), d(xQri,A-1(yn)r\Kn) < l\\ In particular, for fixed Kn and An, we obtain: = U Λη Χγ, h>0 h
3.4. Constrained Inverse Function Theorem 101 Corollary 3.4.8 (Criterion of Pseudo-Lipschitzeanity) Consider any closed subset К of a Banach space X and let A £ C{X,Y) be a continuous linear operator from X to a normed space Y. Assume that for some x0 £ К, there exist constants с > 0, a £ [0,1[ and η > 0 such that \fx £ ΚηΒ(χ0,η), ASK{x)C\BY с А(Тк(х)ПсВх) + аВу (3.12) Then the set-valued map A{K) Э у ~> A-1(y)nK is pSeudo-Lipschitz around (Axq,xq). 3.4.3 Pointwise Stability Conditions The presence of the constant a G [0, l[ in the transversality conditions of the Inverse Function Theorem 3.4.3 is useful because it enables its deduction from a weaker pointwise condition and a kind of "continuity." Namely, we introduce the following definition: Definition 3.4.9 (Sleek and Uniformly Sleek Sets) A closed subset К is called sleek at xq G К if the cone-valued map К Э χ -^ TK{x) is lower semicontinuous-at xq. It is called uniformly sleek at xq € К lim f sup d(u,TK{x))\ = 0 ϊ->ϊ° \ιι6Τ*(ζ0)ηΒ* / We shall prove in Theorem 4.1.8 that the contingent cone Τκ{χο) ίο a subset К sleek at xq is convex. Theorem 3.4.10 (Pointwise Inverse Function Theorem) Let X and Υ be Banach spaces, К be a closed subset of X and f : X i-+Y be a continuously differentiable at xq G- К map. Assume that f'(xo)TK(xo) = Υ 1. — If К is uniformly sleek atxQ, then f(xo) belongs to the interior of f(K) and the set-valued map у ~> Г1{у)пк
102 3- Existence and Stability of an Equilibrium is pseudo- Lipschitz around (f(xo),xo)· 2. — If the dimension of Υ is finite, it is sufficient to assume" that К is sleek at xq to obtain the same conclusion. Proof — Replacing К by a closed neigEborhood of xq in K, we may assume that / is continuous on K. We have to prove that in both cases, the stability assumption is satisfied. 1. — The proof of the first case is easy. There exists a constant с > 0 such that, for all υ in the unit sphere Sy-, there exists a solution uq to the equation f'(xo)u = υ such that ||uo|| < c\\v\\, thanks to Robinson-Ursescu's Theorem (because Tk(xq) is a closed convex cone.) Since К is uniformly sleek at xq and /'(·) is continuous at xq, we can associate with any e > 0 an η > 0 such that, for all uq G X and all χ G Bk(xo,v)> there exists и G Тк{х) satisfying II"-«oil < e||u0|| & H/'(aO-/'(so)|| ^ ε Hence any ν G Sy can be written ν = f'(x)u + w where и G Tk{x), \\u\\ < (1 + e)c||u|| and w = f'(xo)uo — f'(x)u. Thus IN < e||u0||(||/'(so)||+l + e) < εφ||(||/'(χο)||+ l + e) Taking e > 0 small enough and using Theorem 3.4.3 we complete the proof of the first statement. 2. — When the dimension of Υ is finite, the unit sphere Sy is compact. We know from the Robinson-Ursescu Theorem that for any Vi € Sy, there exits a solution uoi € Tk{xq) to the equation f'(xo)uoi = iH such that ||здг|| < CIK|| = c. Fix e > 0. Since К is sleek at xq, for any Vi, there exists щ > 0 such that, for all χ € Βκ{χο,η), we can find щ G Тк(х) satisfying ||uj - uoj|| < е||здг|| The map /'(·) being continuous at xq, there exists щ > 0 such that ||/'(х)-/'Ы11<е. We can cover Sy by ρ balls В(щ,е) so that, by taking m щ °,<p
3.4. Constrained Inverse Function Theorem 103 we obtain that for any ν € Sy and any χ € Βκ(χο,η), there exist щ £Тк{х) and Wi eY satisfying ν = f'(x)ui+Wi, \\щ\\ < c(l+e), Wi = v-Vi+f'(xo)u^-f'(x)ui As above, we can show that Nil < e(l + c||/'(so)|| + (l+e)c) and use Theorem 3.4.3 with e small enough. □ 3.4.4 Local Uniqueness We conclude this section with a local uniqueness criterion of a solution to a constrained nonlinear problem: Proposition 3.4.11 Let К be α subset of α finite dimensional vector- space X and f : Χ ι—» Υ a single-valued map to a normed space Y. Let xq G К and assume that f is Frechet differentiable at xq . If Тк(х0)П ber(/'(s0)) = {0} then xq is the only local solution χ € К to the equation f(x) = f(xo), in the sense that there exists a neighborhood Μ{xq) ofxQ such that \/x € λί{χ0)ηΚ, χ^χο, f{x) φ /(so) Proof — Assume the contrary: for all η > 0, there exists xn € К П B(xo,l/n), xn φ xo such that J(in) = /(^o)· Let us set hn '■= \\χη — xo\\ > 0 which converges to 0 and Xn xq Vn ■■= 7 " hn Since vn belongs to the unit sphere, which is compact, a subsequence (again denoted) vn converges to some element ν of the unit sphere.
104 3- Existence and Stability of an Equilibrium Figure 3.1: Example of Monotone ала Maximal Monotone Maps , 0 / ) Monotone Map , 0 / I Maximal Monotone Map This limit ν belongs also to the contingent cone Tk{xq) because, for all η > 0, xo + hnvn = xn belongs to K. Furthermore, / being Frechet differentiable at xo, we obtain the equation f'{xo)vn = e{hn) where e(hn) converges to 0 with hn. We thus infer that ν belongs also to the kernel of f'(xo)· Since υ φ 0, we derive a contradiction of the assumption. □ 3.5 Monotone and Maximal Monotone Maps In this section, we assume once and for all that X is a Eilbert space identified with its dual. 3.5.1 Monotone Maps Definition 3.5.1 We shall say that a set-valued map A from X to X is monotone if its graph is monotone in the sense that V (χ,ρ) Ε Graph(A), V (y, q) G Graph(A), <p-q,x-y>> 0 Example The map {{x - 1} if χ < 0 {-1,+1} if χ = 0 {x + 1} if χ > 0
3.5. Monotone and Maximal Monotone Maps 105 is a monotone map as well as the map A(x) = < {x - 1} if χ < 0 [-1,1] if χ = 0 {x + 1} if χ > 0 The terminology comes from the monotone maps from R to R. For instance, a nondecreasing map F from R to R is monotone. More generally, if / is a nondecreasing map from R to R, the map χ н-> A(x) := \f(x—), f(x+)] Π R is monotone. Let us give another example of monotone map. Let / be a non- decreasing function from an interval Dom(/) С R to R. Let Ω С Rn be open and X := L2(fi). Define the set-valued map A from X to *by for almost allu€fi, Α(χ(·))(ω) : = /(χ(ω)) It is clear that A is monotone. We shall see in Chapter 6 that the subdifferential map χ ι—> dV(x) of a nontrivial convex function V : X —> RU {+00} is monotone. □ Since monotonicity is a property bearing on the graph of A, A is monotone if and only if A~l is monotone If A and В are monotone and λ > 0, μ > 0, then \A + μΒ is monotone. If we supply Χ χ Χ with the product of the strong topology and the weak topology, we can see that the closure of a monotone graph is still the graph of a monotone map. -" We begin by mentioning the following characterization: Proposition-.3.5.2 A set-valued map A from X to X is monotone if and only if fbr every λ > 0 V(z,p), (2/, ?) € Graph(A), ||s-y|| < ||s -y + X(p - q)\\ (3.13) Proof — We compute \\x - У + \(p - q)\\2 = \\x-y\\2 + \2\\p-q\\2 + 2\<p-q,x-y>
106 3- Existence ала Stability of an Equilibrium Hence, if A is monotone, inequality (3.13) ensues; conversely, if (3.13) holds true, we deduce that λ2||ρ - ill2 + 2X<p-q,x-y>> 0 and obtain monotonicity by dividing by λ > 0 and letting λ converge to zero. □ Remark — We observe that condition (3.13) does not involve the scalar product; therefore, it can be-used on Banach spaces. In this case, maps A satisfying (3.13) are called accretive maps. □ We associate with the map A its resolvents VAX), Jx := (1 + XA)-1 An important consequence of property (3.13) is given in Proposition 3.5.3 The resolvent J\ of a monotone map A is a single-valued nonexpansive map from Im(l + XA) to X. Proof— Let χ G J\{u) and у € J\{v). We can write и € x + XA(x), ν € y + \A(y) Property (3.13) implies that (x-y)+X^—x —j к - 2/|| < = \\u — v\\ This shows that J\ is nonexpansive and, by taking и = v, that J\ contains a unique point. □ 3.5.2 Maximal Monotone Maps We shall characterize monotone maps A such that VA > 0, lm(I + XA) = X They are the maximal monotone maps. Definition 3.5.4 A monotone set-valued map A is maximal if there is no other monotone set-valued map whose graph strictly contains the graph of A.
3.5. Monotone and Maximal Monotone Maps 107 We begin by pointing out the following: A set-valued map is maximal monotone if and only if its inverse A"1 is maximal monotone. Also, the graph of any monotone set-valued map is contained in the graph of a maximal monotone set-valued map by Zorn's lemma, because the union of an increasing family of graphs of monotone set-valued maps is the graph of a set-valued monotone map. The following characterization of maximal monotone maps provides a useful and manageable way for recognizing that an element и belongs to A{x). Proposition 3.5.5 A necessary and sufficient condition for a set- valued map A to be maximal monotone is that the property V (y, v) € Graph(A), <u-v,x-y > > 0 is equivalent to и £ A(x) Proposition 3.5.6 Let A be maximal monotone. Then 1. — Its images A(x) are closed and convex 2. — Its graph is strongly-weakly closed in the sense that if xn converges to χ and if un G A(xn) converges weakly to u, then и € A{x). Proof — 1. — By the preceding proposition, A(x) is the intersection of the closed half-spaces {u£ X \ <u-v,x-y>>0} when (y, v) ranges over the graph of A. Hence, A(x) is closed and convex. 2. — Let xn converge to χ and let un G A(xn) converge weakly to u. Let us choose (y, v) in the graph of A. Inequalities < un - v, xn - у > > 0
108 3- Existence and Stability of an Equilibrium imply, by going to the limit, inequalities < u — ν, χ — у > > 0 Hence, и £ A(x) by Proposition 3.5.5. □ Proposition 3.5.7 // the restriction of α monotone single-valued map A : X ι—> X to finite dimensional vector subspaces is continuous, then A is maximal monotone. Proof — Let χ £ X and и £ X such that (u - A(y), x-y) > 0 for all у £ X To show that A is maximal monotone, we have to check that и = A(x) (Proposition 3.5.5.) For that purpose, we take16 у = x — \(z—x) where λ G]0, l[ and ζ £ X. The above inequality becomes (u - A(x - \(z - x)), z-x) > 0 for all ζ € X By letting λ converge to zero and by the continuity of A on finite dimensional vector subspaces, we deduce that \f ζ £ X, {u-A(x),z-x) > 0 that is и = A(x). D The Minty Theorem we are about to prove provides a very important characterization of maximal monotone maps. Theorem 3.5.8 (Minty) A monotone map A is maximal if and only if the map 1 + A is surjective. Proof— 1. — Assume that 1 + A is surjective. Let χ £ X and и £ X satisfy V (y,v) £ Graph(A), <u-v,x-y>>0 (3.14) 16This method is known under the name of Minty's trick.
3.5. Monotone and Maximal Monotone Maps 109 By Proposition 3.5.5, we have to prove that и G A(x). Since 1 + A is surjective, we can choose у in (3.14) to be a solution г/о of the inclusion u + x € yo +A(yo) Let vq € A(yo) such that гх + χ = г/о + зд. Then, by (3.14), ||я-2/о||2 = <х-уо,х-уо> = - <u-v0,x -yo> < 0 Hence, x = yo, and thus и = v0 € Л(г/0) = ^(ж) 2. — Assume now that A is maximal monotone. Let у € X. We have to prove that there exists χ such that у ζ. χ + A(x). It is sufficient to choose у = 0, since this amounts to replacing A by x —> — г/ + Л (ж), which is also maximal monotone. So we have to show that there exists χ such that —x G A(x). By Proposition 3.5.5, we must prove that there exists χ satisfying V (г/, ν) € Graph(A), < -χ - ν, χ - у > > 0 For convenience, set φ(χ; (г/, υ)) :=< χ + υ, χ — у >, or, equivalently, ip(x;(y,v)) := \\x\\2+ < χ, ν - у > - < ν, у > We have to check that there exists χ such that V(y,u) € Graph(A), φ(χ;(ν,υ)) < 0 (3.15) Fix (г/о, г;о) to be any point in the graph of A. If a solution to (3.15) exists, it certainly belongs to L := '{x £ Dom(A) | φ(χ· (г/0, υ0)) < 0} The map χ —> (^(ж; (г/о, vq)) is quadratic. Consequently, the set {x | <^(ж;(г/0,г;о)) < 0} is convex, closed, and bounded, hence, weakly compact. Let L denote its intersection with Dom(A). Then co(L) is weakly compact.
110 3- Existence and Stability of an Equilibrium Also, the maps χ ι-> φ(χ; (у, υ)) are convex and (strongly) continuous. _ Therefore, their lower sections are convex and closed, hence weakly closed. It follows that the above maps are weakly lower semi- continuous. We use Lemma 3.1.2: Set <S to be the family of all finite subsets Μ := {{yi,vi),...,(yn,vn)} of Graph(A). Then there exists χ £ W{L) such that suP(»,«)6Graph(A)V(^ (»'")) < supMe5 infz6gB(x,) тах(=1,...)П φ{χ; (у», «()) < supMe5 Мхесо[уи^Уп) maxi=if...,n ψ{χ; (у», ъ)) < SUpMe5 infz6co(»i,...,»n) SUPm£5" Σ"=1 /ijVfa (Vj,Vj)) < supMe5 infAe5n sup^n ψμ{\ μ) where for all λ, μ in the simplex Sn η η ψΜ{\μ) ■= Σμιφ{β{\); (yj,Vj)), and β(\) := J^XiVi 3=1 г=1 Every function ψΜ is continuous with respect to λ and ψΜ{μ, μ) = E£fc=i A*Vfc < β(μ) + vh Ук-Уг> = Σα=ι ^Vfc < β(μ), у к -уг>+ Σ£*=ι μιμΗ <ъ,ук-уг> The first term is zero for reasons of symmetry, while the second can be written E%k=iμίμ'ΐ < miУк-Уг> + Σ£ί=ι/*V < -щ,Уз -yi> = E%k=iμίμ]ζ <Уг-ук,Ук-Уг> < о since A is monotone. Hence, assumptions of the Ky Fan inequality (Theorem 3.1.1) are satisfied. Therefore, for all M, inf sup ψΜ{Κμ) < 0 so that inequality (3.15) holds true. We have shown that ж is a solution to 0 G χ + A(x). □
3.5. Monotone and Maximal Monotone Maps 111 3.5.3 Yosida Approximations Suppose that A is maximal monotone. We show that A can be approximated in some sense by single-valued maps Αχ that are also maximal monotone. These maps, called Yosida approximations, play an important role, thanks to: Theorem 3.5.9 Let A be maximal monotone. Then for every λ > 0, the resolvent J\ = (l + \A)~~l is a nonexpansive single-valued map from X to X and the map A\ := (1 — J\)/\ satisfies Г i) Ух € X, Ax(x) e A{Jx{x)) I ii) Αχ is Lipschitz with constant l/λ and maximal monotone Let m(A(x)) denote the element of A{x) with the smallest norm. We also have VxGDom(A), \\Ax{x)-m{A{x))\\2 < \\m(A(x))f - \\Ax(x)\\2 and for all χ G Dom(A), (i) J\{x) converges to χ when λ —> 0+ ii) A\(x) converges to m{A{x)) when λ —> 0+ Definition 3.5.10 The maps A\ are called the Yosida Approximations of A. Proof 1. — Let Xi (i = 1,2) be solutions to the inclusions Уг&Хг + Щхг) (« = 1, 2) So Уг = χι + Xvi where Vi G A(xi). We obtain II2/I-2/2II2 = \\xi ~ X2 + λ(νι - V2)\\2 < = \\xi-x2\\2 + \2\\т-щ\\2+ 2X <υι-υ2,χι-Χ2> . > \Ы - X2W2 + >?\\vi - v2\\2
112 3- Existence and Stability of an Equilibrium Hence, Hzi-ajall < II2/1-2/2II & ||«1-«з|| < llyi-yall/λ (3.16) By taking ?/i = ?/2) (3.16) proves the uniqueness of the solution. We note that Xi = J\Vi and Vi = A\{yi) Hence, inequalities (3.16) prove that J\ and Ax are Lipschitz with constants 1 and l/λ, respectively. 2. — By the very definitions of J\ and Αχ, we have Му) = \(У ~ Ыу)) € A(J\(V)) for all у е X Therefore, since yi = J\{yi) + XA\{y{), we obtain < Ax(yi) - А\{У2),У1 ~ 2/2 > - = < Ax{y1)-Ax{y2),Jx{y1) - Jx{y2) > +X\\Ax(yi) - Ax(y2)f k > Х\МУ1)-МУ2)\\2 > 0 Hence, Αχ is monotone (and by Proposition 3.5.7, maximal monotone.) 3. — Let χ € Dom (A.) Then \\Αχ(χ)-τη(Α(χ)ψ = \\Αχ(χψ + \\m(A(x))\\2 - 2 < Αχ(χ),τη(Α(χ)) > = ||m(A(s))||2 - \\Ax(x)\\2 - 2 < AA(s),m(A(s)) - Ax(x) > Using that A is monotone, that т(Л(ж)) € Л(ж) and A\(:c) £ Л(7л(ж)), we obtain < AA(s),m(A(s))-AA(s) >= γ < m(A(x))-Ax(x),x-Jx(x) > > 0
3.5. Monotone and Maxima,! Monotone Maps 113 Therefore, we have proved inequality \\Ax(x) - m(A(x))\\2 < \\m(A(x))f - \\Ax(x)\\2 (3.17) 4. — Then when χ G Dom(A), we have \\х-Ых)\\ = λ||Αλ(χ)|| < A||m(A(x))|| therefore, J\(x) converges to χ when λ converges to zero. 5. — We deduce that у = A\(x) is a solution to the equation у € A(x — \y). Indeed, setting ζ = χ — Xy, this equation becomes χ £ ζ + \A(z). Hence, ζ = Jx{x) к у = {x-Jx{x))/\ = Ax{x) This remark implies that Αμ+χ{χ) = (Αμ)χ{χ) Indeed, у = Αμ+χ{χ) is a solution to the equation у G A(x — \y — μy)■, then у € Αμ(χ — Ay). Applying again the preceding remark to the Yosida approximation Αμ, which is maximal monotone, we deduce that у = (Αμ)χ(χ). 6. — Now we use inequality (3.17), replacing A by Αμ. Since πι(Αμ(χ)) = Αμ(χ), we obtain ||Αμ+λ(χ)-Αμ(χ)||2 < ||Αμ(χ)||2-||Αλ+μ(χ)||2 Then the sequence ||Αμ(α;)||2 is monotone and bounded from above by ||т(Л(ж))||2, so that it converges to some real number α when A —> 0+. This implies that lim ||Αμ+λ(χ)-Αμ(χ)||2 < а-а = О Hence, Αχ(χ) satisfies the Cauchy criterion and converges to some element ν in X. Since A\(x) G A(Jx(x)) and the graph of Л is closed, we deduce that ν £ A(x). Also IMI = lim ||Αλ(χ)|| < ||m(A(x))|| A—*U4~
114 3- Existence and Stability of an Equilibrium Since A(x) is closed and convex, the projection of zero onto A(x) is unique and consequently, ν = m{A{x)). Therefore, A\(x) converges to m(A{x)) for all χ € Dom(A). D We shall provide an existence and uniqueness theorem for differential inclusions the right-hand side of which is minus a maximal monotone in Chapter 10. Remark — We shall study in Chapter 5, Section 2 the derivatives of monotone maps and in Chapter 7, Section 1, the graphical convergence of maximal monotone maps. □ 3.6 Eigenvectors of Closed Convex Processes We shall prove now that a closed convex process F does have an eigenvector in cones with compact soles. Let К be a closed convex cone in Rn. We recall that the following conditions are equivalent: ' г) К is spanned by a convex compact set disjoint from 0 < ii) the interior of the polar cone K+ is not empty Hi) S := {x € K\ < po,x > = 1} where po € Int(UT+) spans К The compact convex subset S is called the sole, and such closed convex cone are called cones with compact sole. Definition 3.6.1 We say that a solution χ to the problem χ £ Κ, χ ^ 0, AGR & Xx £ F(x) (3.18) is an eigenvector χ G К of the closed convex process F associated with an eigenvalue λ. Theorem 3.6.2 Let X be a finite dimensional vector-space and F : X ~»· X be a closed convex process. Assume that a closed convex cone К С X enjoys the following properties: г) К has a compact sole ii) К is a viability domain of F Hi) the norm \\R\\ of the map R defined by : \/x<EK, R(x) := F(x)nTK(x) φ 0 is finite
3.6- Eigenvectors 115 Then there exists a nonzero eigenvector χ G К of the closed convex process F associated with an eigenvalue λ. Proof — We associate with the closed convex cone К and an element po € Int(.ftT+) the "compact sole" S := {x € K\ <po,x > = 1} By Lemma 4.2.5 below, its tangent cone is equal to Ts(x) = {v€TK(x)\ <po,v > = 0} For an element у G S the projection w(y) onto the orthogonal hy- perplane to po is defined by VzG-X", w{y)z := z- <p0,z> у We then remark that: V2/G5, ш(у)Тк(у) С Ts{y) (3.19) Indeed, by Lemma 4.2.5 below, if и € Тк(у), then w(y)u :=u-<p0,u>y € {K + Ry) + Ry С К+ Ry = TK(y) because К is a closed convex cone. Furthermore < po,и > = < po, и > - < po, и >< po,y>= 0 because < po,y >= 1. We deduce that w{y)u belongs to Ts{y). Let us associate with the closed convex process -F the set-valued map G defined on the compact sole S by G{y) := w{y){F{y) П ||Д||В) It is obviously a set-valued map with closed convex images contained in the ball ||-R||-B, which is compact. Since its graph is closed, we deduce that G is upper semicontin- jaous from S to X. By (3.19), S is a viability domain of the set-valued map G since the cone К is a viability domain of the closed convex process F.
116 3- Existence and Stability of an Equilibrium Since the sole S is a convex compact subset, we can apply Equilibrium Theorem 3.2.1 to the set-valued map G. There exists an equilibrium χ € S of G, i.e., a solution to 0 € G(x), in other words, a solution to χ € S, 0 = y— < po, у > χ where у € F(x) By setting λ :=< po) У >i we see that the pair (λ,χ) is a solution to inclusion (3.18.) □ Remark — Other theorems dealing with positive eigenvalues and an extension .of Perron-Probenius's Theorem to closed convex processes can be found in [35, Chapter 3]. □
Chapter 4 Tangent Cones Introduction We already had to use tangent cones in the preceding chapter, both for solving inclusions under constraints and studying stability and convergence of approximations of their solutions. They also play an important role in viability theory and in optimization. It is time now to conduct a thorough investigation of their properties in order to apply theorems involving tangent cones. There are many ways to describe technically the concept of tan- gency to a given subset К at a point χ € К. The idea can well be expressed in terms of constraints: We regard a (closed) subset К as the subset of elements of the whole space X obeying given constraints. Pick any direction ν G X and start from χ in the direction v, ranging over the line χ + hv when h > 0. We would like to distinguish those directions which, for small h, do not lead us too far away from K. Such directions encapsulate the idea of tangency. But how can we translate rigorously this idea in mathematical terms? It certainly involves the distance άκ(χ + hv) from the points x + hv to K, which should tend to 0 faster than h, in the sense that άκ(χ + hv)/h should converge to 0 with h. Naturally, but unfortunately, there are many ways of taking such limits; we can take either the liminf or the limsup, we can require 117
118 4- Tangent Cones a little more uniformity in both cases. We feel like opening the door of a menagerie of tangents, and facing the choice of a favorite pet! We are in the situation of the particle physicists when so many unsuspected particles popped up, destroying the harmonious pre- particle description of matter. We also face a whole collection of more or less colored "quark cones", each one having a charm of its own. Even though the subjective choice of a given category of tangents can be justified by accumulated experience, the reasons to emphasize one concept at a given time can very well disappear later on. This is why we shall use the concept of "tangent cone" as a generic term to designate any of the particular implementations inherited from history (as the contingent and paratingent cones since Bouli- gand.) Fortunately, most of these different implementations coincide for familiar classes of subsets (differential manifolds, convex subsets.) An analogous situation happened with the concept of derivative of a function, rigorously defined after Cauchy and Weierstrass, and severely taught to generations of young, properly brainwashed mathematicians. The acquired dogma were difficult to overrun when Sobolev (very timidly) and then L. Schwartz (much more boldly) introduced the weak derivatives leading on to derivatives in the sense of distributions, when Dini involved liminf and limsup to let differential quotients converge (more on this in next chapters) and when Bouli- gand introduced the concepts of contingent and paratingent sets. Furthermore, even for the usual derivatives, different ways of taking the limits and various topologies were used, giving rise to a manifold of concepts (Gateaux, Hadamard, Frechet, C1, etc., derivatives) which were adapted more or less reluctantly, each problem requiring its particular brand of derivative. Geometers became eagerly "differential geometers" one century ago, and would hardly accept leaving the luxurious environment where the sets of tangents are comfortable vector spaces, with smoothness all around them. But non smooth sets do exist, unfortunately, as soon as objects of investigation are no longer freely given; they are encountered in numerous problems as the result of some operations. This may destroy smoothness, so that we lose the possibility of imposing a priori
4.0. Introduction 119 regularity conditions . Many of us did actually encounter them. The pressure coming from many problems2 in control theory and differential games, in economy theory, in biological regulation, in systems theory, etc., led mathematicians from many countries and schools, motivated by different problems,.to investigate various ways of overcoming regularity requirements. The first step away from differential geometry was to adapt many of the results to the case of convex subsets, after Fenchel, J.-J. Moreau and Rockafellar. It was a surprise to see that tangent cones and normal cones to convex subsets shared many properties of tangent and normal spaces to differentiable manifolds. After being born with Bouligand in the thirties, and then revived during the early days of control theory, the problem of defining and using tangent cones to arbitrary subsets got a new start in 1975, when Clarke introduced a tangent cone Ск{х) to a set К at χ G К which he proved to be always a closed convex cone. The price of this nice property, however, was quite high since this tangent cone may often be too small or even reduced to the singleton 0. At that time, notwithstanding this cost, it was a good candidate to be "the" tangent cone that was waited for as a messiah! Further investigations both in viability theory and control theory suggested however that, for the time being at least, the tangent cone which was the most often adequate was the contingent cone Тк{%) introduced by Bouligand: it is the upper limit of the differential quotients3 {K — x)/h when h —> 0+. It happens however that if the set-valued map у ~> Тк{у) is lower semicontinuous at x, then the contingent cone Тк{х) coincides 1However Differential Geometers can find solace in the 2,600 year old Xeno- phanes of Colophon's statement that God was spherical, unique, motionless and "shaking" all things by the power of thought. He made fun of the anthropomorphic character of the gods of the Olympian theology, arguing that cows would make their gods in their own image. Of greater philosophic interest, he stated that men can have no certain knowledge, only opinions. 2 Just remember to begin with that the intersection of two smooth sets may no longer be smooth. 3whereas the tangent cone С к (х) is the lower limit of the differential quotients {K — x')/h, when h —> 0+ and χ' —>κ χ.
120 4- Tangent Cones with Ck(x) and is thus convex. This fortunate situation deserves a name: we say that K_ is sleek at _x. Both smooth manifolds and closed convex subsets are sleek: This, and elementary properties of the contingent cones, are presented in Section 1. Section 2 is devoted to the tangent cones to convex subsets, where we recall (most often, without proof) the properties of the tangent cones from convex analysis. The main motivation for introducing sleek subsets (and/or Clarke tangent cones) is that they play a crucial role in the transversality assumptions that we need for stating that a tangent cone to an inverse image (respectively an intersection) is equal to the inverse image (respectively the intersection) of the tangent cones. These crucial formulas are indeed a consequence of the Constrained Inverse Function Theorem from Chapter 3, and are proved in Section 3. We then devote Section 4 to dual concepts: the normal space to a differentiable manifold being defined as the orthogonal space to the tangent space, the normal cone to a convex subset as the polar cone to the tangent cone, we consider the polar cones to the tangent cones and translate by duality the properties of the corresponding tangent cones to normal cones. We shall not dwell on this, since many other monographs emphasize this view point, couched in the traditional Lagrange formulation of necessary conditions for optimality. In this sense, one can say that this book is "orthogonal" to most of the texts devoted to optimization and nonsmooth analysis! Section 5, which can be omitted in a first reading, deals with other tangent cones. When convexity is required, we can use what we call the convex kernel of the contingent cone, which is a convex cone lying between the Clarke tangent cone and the contingent cone, and which is large enough: it is the intersection of the maximal closed convex cones contained in the contingent cone. We then examine the paratingent cones introduced also by Bouli- gand, which contain the contingent cones. The graph of the paratingent map is closed (an important property indeed.) We prove Cho- quet's Theorem which states that the contingent and paratingent cones are generically equal.
4.1. Tangent Cones to a Subset 121 We study in Section 6 the limits of tangent cones Ткп{хп) to a sequence of subsets Kn and relate them to the upper and lower limits of differential quotients (Kn — xn)/hn. Since we can regard tangent cones to a subset К as first-order approximations of K, we conclude this chapter with a short introduction to higher order approximations, which are various limits when h i-> 0+ of "differential quotients" К — χ — hu\ — ■■ ■ — hrn~1vm-i _ They enjoy many of the properties of first-order approximations. 4.1 Tangent Cones to a Subset 4.1.1 Contingent Cones We begin with a presentation of the contingent cones: Definition 4.1.1 (Contingent Cones) Let К с X be α subset of α normed vector space X and χ € К belong to the closure of K. The contingent4 cone Тк{х) is defined by TK{x) := {v\limw.fdK(x + hv)/h = 0} We see at once that the contingent cone Тк{х) is the upper limit of the subsets (K — x)/h, so that Τχ(χ) is a closed cone. 4from the Latin contingere, to touch on all sides, introduced by G. Bouligand in the 30's. This term was already used by R. Descartes, in a 1638 letter to Mersenne criticizing P. de Fermat's method on tangents: "Puis, outre cela, sa regie pretendue n'est pas universelle comme il lui semble, et elle ne peut s'etendre a aucune des questions qui sont un peu difficiles, mais seulement aux plus aisees, ainsi qu'il pourra eprouver si, apres 1'avoir mieux digeree, il tache de s'en servir pour trouver les contingentes, par exemple, de la ligne courbe BDN, que je suppose etre telle qu'en quelque lieu de sa circonference qu'on prenne le point B, ayant tire la perpendiculaire ВС, les deux cubes des deux lignes ВС et CD soient ensemble egaux au parallelepipede des deux memes lignes ВС, CD et de la ligne donnee P." This is quite a long and nasty sentence indeed.
122 4 - Tangent Cones Figure 4.1: Contingent Cone at a Boundary Point may be the Whole Space _ /^ \ ° . ~~2^) ~^) Subset К such that TK{0) = X Denote by SK(x) := U K-x h>0 h the cone spanned by К — χ. Hence the contingent cone Тк(х) is contained in Sk(x)- It is very convenient to have the following characterization of this cone in terms of sequences: {υ € Тк{х) if and only if 3 hn —»· 0 + and 3 υη —»· υ such that V n, ж + hnvn € AT We also observe that if χ € Int(ur), then Гя-(х) = X This situation may also happen when χ does not belong to the interior of К (see Figure 4.1.) Unfortunately, the graph of the set-valued map К Э χ ~> Тк{х) С X is not necessarily closed. This important property is lacking whenever inequality constraints are involved in the definition of K, as the
4.1. Tangent Cones to a Subset 123 Figure 4.2: The Graph of T[a>b](·) ΤΓ following simple example, where К := [а,Ь], shows (see Figure 4.2.) We shall see in a while that this set-valued map is closed when К is described by equality constraints and that it is lower semicontinuous when К is convex. Example Let us consider a normed space X, a continuous map д = (gi,---,9P) -X >-> R-p and the subset К of X denned by inequality constraints К := {χ € X I gi(x) > 0, i= l,...,p} Fix χ in K. We denote by I(x) ■- {i=i,...,p| 9i(x) = 0} the subset of active constraints. We observe that Тк{х) = X whenever I(x) = 0 and that, otherwise, inclusion TK(x) С {и £ X I Vie I(x), < д-(х), и > > 0} holds true when g is Frechet differentiable at ж. If we posit moreover the constraint qualification assumption: 3 v0 £ X such that V г € I(x), < g'i(x), v0 > > 0
124 ■ —4 - Tangent Cones Table 4.1: Properties of Contingent Cones. > If К С L and χ € K, then Тк(х) С TL(x) > If Kid X, (i = 1,. .., n) and ж € Ц ^i> then where /(ж) := {г | χ € ^} > If ΛΓί С Xj, (г = 1, ... j n) and χι € if,;, then 2nr=ilf«^1,"-,a:n^ C ПГ=1^(Жг) > IfgeVpf.y), ififCl,xeIaBdMc7, then д'{х){Тк{х)) С Гв(к?(5(х)) Tff-i(i0(s) С ^(хНад^а:)) о If Κι С X, (г = 1,..., η) and x € fli ^'> tnen then the contingent cone Гк:(а;) = {u ε Χ | Vie I(x), <д'{(х),и>> 0} Indeed, let и satisfy < д[{х), и >> 0 for any г ε I(x). For г φ, I(x), strict inequalities 5i(a;) > 0 imply that for some а > 0, we have Vfte [Ο,α], Vi ^ /(ж), gi(x + hu) > 0 Consider first the case when (^(a;),u) > 0 for any г ε I{x)- Then V г ε Ι(χ), gi{x + hu) = gi(x + hu) - gi{x) = < д{(х), u> + hei(h) where Si(h) converges to 0 with h. This implies that gi{x + hu) > 0 for h small enough and all г ε I(x), and thus, for all г = 1,... ,p. Then such an element и belongs to the contingent cone Тк(х)· Consider now the general case. By assumption, we deduce that for any β ε]0,1[, υ,β := (1 — P)u + βυο satisfies strict inequalities (д[(х), up) > 0 for any г ε 1{χ) and, by what precedes, belongs also to the contingent cone Tk{x)· Letting β converge to 0, we infer that the limit и of the up's belongs also to the contingent cone Τχ(χ). Π (1) (2) (3) (4) (5)
4.1. Tangent Cones to a Subset 125 4.1.2 Elementary Properties of Contingent Cones Proposition 4.1.2 Assume that the Banach space X is smooth, i.e., that the norm of X is Gateaux differentiable off the origin, and denote by J(x) its gradient at x. Let К be a closed subset of X. We set Пк-(у) := {ζ € K\ \\z - ?/|| = dK(y)} Let у £ К be such that Лк(у) is n°t empty. Then Vz€UK(y), \/v£cd(TK(z)), (J(y-z),v) < 0 Proof — Fix υ e Tk{z): There exist hn > 0 and vn e X converging to 0 and ν respectively such that ζ + hnVn belongs to К for all η > 0. Since the function || · || is convex and Gateaux differentiable off the origin, the following inequality holds true: for every ?/i ^ 0 and 2/2, < J(yi),V2 - 2/1 > < II2/2II - H2/1II We thus deduce that < J(y - ζ - hnvn),vn > < = < J(y - ζ - hnvn), y-z-(y-z- hnVn) > /hn < (\\y-z\\-\\y-'z-hnvn\\)/hn < 0 Since J is continuous from X to X* supplied with the weak-* topol- " ogy, we infer that J{y — ζ — hnvn) converges weakly to J{y — z). Hence < J{y — ζ — hnvn),vn > converges to < J(y — ζ),υ > because vn converges strongly to v, so that \/v€TK(z), <J{y-z),v><0 and thus the same inequality holds true for any ν in the closed convex hull of TK(z). □ We introduce the Dubovitskij-Miljutin cone defined by 5which is a map from X to the unit sphere of X*, called the duality map, satisfying in particular < J(x),x >= \\x\\. The duality map is continuous from X to X* supplied with the weak-* topology.
126 4 .-- Tangent Cones Definition 4.1.3 The "Dubovitskij-Miljutin tangent cone" Dk(x) to К at χ € К is defined by: υ € Dk{x) if and only if 3 e > 0 such that χ + ]0,ε](υ+εΒ) С К This cone is the complement to the contingent cone to the complement: Lemma 4.1.4 Let χ belong to the boundary of К and К denote the complement of К'. Then the complement of the contingent cone Tg(x) to К at χ is the "Dubovitskij-Miljutin cone" Ώχ(χ) to К at x. Remark — The last example shows that when К := {χ I gj(x) > 0, j = 1,... ,p} then the open subset {v\ <sj(i),0 > 0, j £l(x)} is contained in Dk{x)· α 4.1.3 Adjacent and Clarke Tangent Cones To go further by having equalities in some of the formulas of Table 4.1, we need more assumptions in order to use the Constrained Inverse Function Theorems of the preceding chapter. These further assumptions are regularity assumptions, which require that the contingent cone Тк(х), which is the upper limit of the subsets (K — x)/h when h —>■ 0+, coincides with the lower limit of these subsets when /i-+0+ and even, of the subsets (K — y)/h when both h —> 0+ and у £ К converge to x. This is the reason why we have to introduce two more tangent cones: Definition 4.1.5 Let К С X be a subset of a normed vector space X and χ £ К belong to the closure of K.
4.1. Tangent Cones to a Subset 127 1. — the intermediate or adjacent6 cone Т\{х) is defined by Ί*κ(χ) := {v\ lim dK{x + hv)/h = 0} 2. — the Clarke7 tangent cone or circatangent cone С к {я) is defined by CK{x) ■■= {v | lim dK(x' + hv)/h = 0} h—>0+,i'—+Д-1 where —+K denotes the convergence in K. We shall say that a subset К С X is derivable at χ ς. Κ if and only ifl\(x) = Тк(х) and tangentially regular at χ г/Тк(х) = Ск(х)· We see at once that these tangent cones are lower limits Ί%(χ) = Ummih^Q+—τ— h and К — χ' Ск(х) = Liminf/l_t0+, кэх'^х—τ— so that they are closed cones, that Ск(х) С Ί*κ(χ) С Тк(х) с 5^М that these tangent cones to К and the closure К of К do coincide and that if χ e Int(UT), then CK{x) = X We shall see later that the converse is true when the dimension of X is finite. These tangent cones may all be different: see Figure 4.5.2 of Section 4.5.2 for an example when these tangent cones as well as other ones are distinct. In the example of the set of Figure 4.1, the 6from the Latin adjacere, to lie near, also used under the name of derivable cone by R. T. Rockafellar. 7from the Canadian Frank H. Clarke; we shall use the adjective circatangent to mention properties derived from this tangent cone, for instance, circatangent derivatives and epiderivatives.
128 4 - Tangent Cones adjacent and the contingent cones at zero coincide (with the whole space) and the cone Ск{0) is equal to the horizontal axis R χ {0}. It is very convenient to use the following characterization of these cones in terms of sequences. f ν € Тьк{х) if and only if V hn -* 0+, 1 3vn^*v such that V η, χ + hnvn € К and ν € Ck{x) if and only if V hn —»· 0+, V xn —*к x, 3vn —* ν such that V η, χη + /ιπυπ € AT They are not necessarily equal (see Figure 4.1) for instance. Let us single out an astonishing fact: the tangent cone Ск{х) is always a closed convex cone. Proposition 4.1.6 The tangent cone Ск{х) is a closed convex cone satisfying the following properties CK(x) + TK(x) С Tk(x) L· Ск{х)+Тьк{х) С Тьк(х) Proof— Let vi and V2 belong to Ск(х)· To prove that vi +щ belongs to this cone, let us choose any sequence hn > 0 converging to 0 and any sequence of elements xn € К converging to x. There exists a sequence of elements v\n converging to щ such that the elements xin := xn + hnvin do belong to К for all n. But since the sequence x\n does also converge to χ in K, there exists a sequence of elements V2n converging to V2 such that V П, Xin + hnV2n = Xn + Κ(υίη + V2n) & К This implies that v\ +V2 belongs to Ск(х) because the sequence of elements v\n + V2n converges to υχ + V2. The proof of the two other inclusions is analogous and left as an exercise. □ Unfortunately, the price to pay for enjoying this convexity property of the Clarke tangent cones is that they may often be reduced to the trivial cone {0}.
4.1. Tangent Cones to a Subset 129 Table 4.2: Properties of Adjacent Tangent Cones. > If К С L and χ £ К, then Ί+Κ(χ) С lj{x) > If K{ С X, (г = 1,..., η) and χ G (Ji ^i, then where I(x) := {i|a; G iQ} > li Ki С Xi, (i = 1,. .., n) and Xi G ifj, then > IfggC^y), if К С Χ, χ € F and M С У, then tf'COCUOO) с 2j(i0GKaO) > If Ki С X and lefli^i (i = l,...,n), then As an example, consider the subset ЛГ := {(x,y)eR2||x| = |y|} We see that Tk(0) = 2^(0) = ЯГ and C^(0) = {0} But we shall show in just a moment that the Clarke tangent cone and the contingent cone do coincide at those points χ where К is sleek, i.e., where the set-valued map χ ~»· Тк{х) is lower semicon- tinuous. Hence the cone Ск{х) can be seen as a "regularization" of the contingent cone Тк{х). The elementary properties of the adjacent cones are the same than the ones of contingent cones, except for the formula of the intermediate tangent cone to the product, which is the product of the intermediate tangent cones, and the formula for the union. They are summarized in Table 4.2. (1) (2) (3) (4) (5)
130 4 - Tangent Cones 4.1.4 Sleek Subsets We recall the definition of sleek subsets: Definition 4.1.7 We shall say that a closed subset К is sleek at xq € К if the cone-valued map К Э χ ~> TK{x) is lower semicontinuous at xq and that it is sleek if it is sleek at every point of K. Theorem 4.1.8 (Tangent Cones to Sleek Subsets) Let K~be a closed subset of a Banach space. If К is sleek at χ G K, then the contingent and Clarke tangent cones to К at χ do coincide, and consequently, are convex. It follows from the "lower semicontinuous version" of the next result: Theorem 4.1.9 Let К be a nonempty closed subset of a Banach space X and xq belong to К. Then LiminfI_tic3;or^(a:) С Cr{xo) because if К is sleek at xq, then Tk{xq) С Liming кХ0Тк {х) С Ck{xq) С Тк(х0) Before proving Theorem 4.1.9 in full generality, we shall prove stronger versions first in the case of finite dimensional vector-spaces, and then, in the case of Hilbert spaces and more generally, in the case of uniformly smooth Banach spaces X such that the norm of its dual X* is Prechet differentiable away from zero. 4.1.5 Limits of Contingent Cones; Finite Dimensional Case Theorem 4.1.10 Let X be a finite dimensional vector-space and К be a closed subset of X. Then for every χ £ К lAnnxiiy^KXTK{y) = Limmiy^KXcd{TK{y)) = CK{x) Consequently, К is sleek at χ if and only if it is tangentially regular at x.
4.1. Tangent Cones to a Subset 131 Proof 1. — Let us take ν φ 0 in the lower limit of the closed convex hull of the contingent cones Τκ{ζ) when ζ converges to χ G K: This means that for any e > 0, there exists η > 0 such that for V ζ£Β{χ,η)Γ]Κ, Β(ν,ε)ΠΈοΤκ{ζ) φ 0 For proving that ν G Ск(х), we take any y£B{x,V/4)nK, t < η/4\\υ\\, zEUK{y + tv) and we introduce the function д defined by g(t) := ак{у + tv), which, being Lipschitz, is almost everywhere differentiable. Let ί€[0,τ,/4|Η] be any point where g is differentiable. We claim that g'(t) < ε whenever g(t) > 0, i.e., whenever у + tv does not belong to K. Let h > 0 be small enough. Since dK(y + (t + h)v)-dK(y + tv) < ||y-z + nt,||-||y-z|| and since the euclidian norm is convex and differentiable off 0, we infer that ' 9(t + h)-g(t) < \\y-z + hv\\-\\y-z\\ < h(^j±j^,v) < { Ь fi \\\y-z+hv\\ \\v-z\\>v/ + n \W^>O/ We check next that ζ G B(x, η) Π Κ because \\z — x\\ < \\z — у — tv\\ + ||?/ — χ +iw|| < 2||y — a;+ίϋ|| < η whenever у G Β(χ,η/4) and t < 77/4||w||. Therefore, there exists w G co(Tk{z)) such that \\v — w\\ < e, so that, by Proposition 4.1.2, ' (9(t + h)-g(t))/h У - e + \\\y-z+hv\\ "Й'7
132 4 - Tangent Cones Taking the limit when h —»· 0+ we deduce that g'{t) < ε whenever g{t) > 0. Hence the claim is proved. We prove next that g(t) < et. This being obviously true when g{t) = 0, assume that g(t) > 0. Let ίο := sup{T€[O,i]|0(T) = 0} Since g is continuous, g(to) is equal to 0, so that we can write g(t) = f g'{r)dr < e(t - i0) < et because g'{r) < ε for any r €]ίο>ί]· This shows that υ G Ск{х)· 2. — Let us prove now that Ск{%) is contained in the lower limit of the contingent cones Тк(у) when у converges to χ in K. ' Indeed, let υ belong to Ск(х) and be a sequence of elements xn € К converging to x. Then, for all e > 0, there exist N and β > 0 such that, for all 0 < h < β, η > Ν ак(хп + hv) < Ηε Let us associate with such xn elements y% € К satisfying \\Уп~хп -Ml ^ ^ε We set υ^:= (y^ — xn)/h. Since ||ϋ£ —ϋ|| < ε and since the dimension of the space X is finite, there exists a cluster point vn G υ+εΒ of the sequence (υ^)/ι>ο· Such a vn belongs to the contingent cone Τκ{%η)· Hence υ is the limit of the elements vn. □ 4.1.6 Limits of Contingent Cones; Infinite Dimensional Case The proof of the first part of the above theorem can readily be extended to the case of Hilbert spaces when К is weakly closed. This is too strong an assumption, since we need to use such a result for closed subsets. We can overcome this difficulty and extend Theorem 4.1.Ю
4.1. Tangent Cones to a Subset 133 to uniformly smooth8 Banach spaces by using the following Edel- stein Theorem9 (see [35, Theorem 5.7.13. p.294] for instance), which states that in Hilbert spaces and some Banach spaces, we can approximate any point by another which has a unique projection of best approximation on a closed subset K. Theorem 4.1.11 Assume that X is reflexive and that the norms of X and X* are Frechet differentiable off the origin. Let К be a closed subset of X. Then there exists a dense subset D of X of points у with a unique projection on K. In infinite dimensional spaces, we have to replace the contingent cones Tk{x) by the weak contingent cones Τχ(χ) defined by Vh — X υ € Tpc{x) <£=Φ- υ is a weak cluster point of —-— when h —> 0+ h where ун belongs to K. We observe that Proposition 4.1.2 can be extended to the case when the contingent cones Τκ {ζ) are replaced by the weak contingent cones T%(z). Proposition 4.1.12 Assume that the norm of the Banach space X is Frechet differentiable off the origin and denote by J(x) its gradient at x. Let К be a closed subset of X and у ф К be such that Цк-(у) is not empty. Then Vz€nK(y), \/v€co(T£(z)), <J(y-z),v><0 Proof — Since the norm is Frechet differentiable at у — ζ φ 0, we know that for any e > 0, there exists η > 0 such that whenever IIHI ^ Vi \\\y — ζ + w\\ — \\y — z\\— < J(y — z),w > I < e||w|| 8A Banach space is uniformly smooth if and only if its norm is uniformly Frechet differentiable in the sense that there exists e(h) converging to 0 with h > 0 such that V χ φ 0, ж 6 Б, υ SB, \\\χ + HI " N1 -h< J(x),v >| < he(h) In this case, J is uniformly continuous from X to X* supplied with the norm topology. Uniformly smooth Banach spaces are reflexive. Lebesgue spaces Lp(0) and Sobolev spaces Wm,p(n) are uniformly smooth for 1 < ρ < +oo. 9which can be derived from Ekeland's Variational Principle.
134 4 - Tangent Cones Let ν belong to the contingent cone T£(z). Then there exist sequences hn > 0 converging to 0 and vn g X converging weakly to ν such that ζ + hnvn belongs to К for any η > 0, so that \\y ~z|| < \\y-z- hnvn\\ For η large enough, the norm of w := hnvn is smaller than or equal to r\ since vn is bounded by a constant с and hn converges to 0. The above inequalities imply < J(y - z),vn > --— , . < (||y -z -hnvn\\ - \\y -z\\ + hn < J(y - z),vn >)/hn < e\\vn\\ < ec and thus, by letting vn converge weakly to v, that < J(y — z),v >< ec for any e > 0. Hence < J(y — z),v >< 0 for every υ € Τχ(ζ), and thus, for any ν in its closed convex hull. □ Theorem 4:1.13 Assume that X is uniformly smooth10 and that the norm of X* is Frechet differentiable off the origin. Let К be a closed subset of X. Then Liminfy-^ Тк{у) = 1лшш1у^кхсо{Т^{у)) = Ск(х) Proof — The proof is similar to the one of Theorem 4.1.10 1. — To prove the first inclusion, let us take ν φ 0 in the lower limit of the closed convex hull of the weak contingent cones Τχ(ζ) when ζ € К converges to χ € К: This means that for any e > 0, there exists η > 0 such that V ζ€Β(χ,η)ηΚ, Β{υ,ε)Γ)ΈδΤκ{ζ) φ 0 We have to show that it belongs to Ск(х)· For that purpose, we take у € B(x, 77/6) П K, t < 77/6||t^|| and we introduce the function g(t) := dK(y + tv) 10Actually, we only need that the norm of X is Frechet differentiable off the origin and that J is uniformly continuous from the unit sphere of X to X* supplied with the weak-* topology.
4.1. Tangent Cones to a Subset 135 which, being Lipschitz, is almost everywhere differentiable. Let t € [0,η/6\\υ\\] be any point where д is differentiable. We claim that g'{t) < ε whenever g(t) > 0. Let ft > 0 be small enough. By Edelstein's Theorem 4.1.11, we can associate with any ft > 0 an element ун € В (у + tv, ft2) which has a unique projection Zh G K. Therefore, f dK(y + {t + h)v)-dK{y + tv) < dK(yh + hv)-dK(yh) + 2h2 \ < \\Vh - Zh + hv\\ - \\yh - zh\\ + 2h2 Let J : Χ η-> J5* be the gradient of the norm of X. Since X is uniformly smooth, we infer that there exists e{h) converging to 0 with h such that \\yh -zh + hv\\ - \\yh - zh\\ < h< J(yh - zh),v > +he{h) Finally, we check that Zh € Β(χ,η) because \W -z|| < \\zh -Vh\\ + \\yh -y -tv\\ + ||y-a;+iu|| < dK{y+tv) + 2h2 + ||y-a;+iu|| < 2(\\y - χ + tv\\ + h?) < η whenever h < -у/77/6, у € Β(χ,η/6) and t < rj/6||u||. Therefore, there exists Vh € W(T^(zh)) such that \\v — υ^\\ < ε, so that, by Proposition 4.1.12, ' g(t + ft) - g(t) = dK(y+tv + hv) -dK(y + tv) < ft < J(yh-zh),vh > +Λ(||ϋ-ϋ/»||+ε(Λ)+2Λ) < ft(e + e(ft) + 2ft) so that g'(t) < ε whenever g(t) > 0. This yields that g(t) < et in the same way as in the proof of Theorem 4.1.10 and shows that ν £ Ск{х)·
136 4 - Tangent Cones 2. — The second inclusion is a consequence of Lemma 4.1.14 Let X be α reflexive Banach space, К С X a closed subset and xq belong to K. Then Ck{xq) С итт£х->кХ0Тк(х) Proof — Let υ belong to Ck(xq) and be a sequence of elements xn € К converging to xq. Fix e > 0 and let N and β > 0 be such that, for all 0 < h < β, η > Ν ак(хп + hv) < he Let us associate with such xn elements y% £ К satisfying \\уЬ-хп-кь\\ < 2Ле We set v% ■= {Уп — xn)/h. Since \\v^ - v\\ < 2e and since the space is reflexive, there exists a weak cluster point vn € ν + 2εΒ of the sequence (w^)/l>0. Such a vn belongs to the contingent cone Τχ(χη). Hence ν is the (strong) limit of the elements vn. □ Proof of Theorem 4.1.9 — Since the lower limit of the contingent cones Tk(x) when χ £ К converges to xq in К is a cone, it is enough to show that any υ G Limmix-yкХ0Тк(х) with norm equal to 1 belongs to Ck{xq)· Fix such a v. Then for any e > 0, there exists iV > 1 such that Vn>iV, VzeBjf(i0ll/n), Β{υ,ε/2)ΓλΤκ{ζ)φ% (4.1) Assume for a while that υ does not belong to Ск(%о)· This means that there exist e G]0, l/4[ and sequences xn € Βχ(χο, 1/2η), К е]0, l/4n[ with {xn + hnB{v,e))DK = 0 (4.2) Fix iV such that (4.1) holds true, n> N and set δ := e/(l + ε). We use Ekeland's Variational Principle (Theorem 3.3.1) with the closed subset Q := Kn{xn + \0,hn]B{v,e))3xn
4.1. Tangent Cones to a Subset 137 and the continuous function V defined on Q by V(x) := — \\x — xn\\, which is bounded from below on Q: There exists then a solution ζ G Q to the «^-minimization problem Vi£Q, Цж-ЖпЦ < \\z —xn\\ +δ\\ζ —x\\ (4.3) We shall exhibit an element χ € Q which, plugged in the above inequality, yields a contradiction. For that purpose, we observe that ζ can be written ζ = xn + (1 — a)hnwn with a € [0, l] and ||wn —ϋ|| < ε. We first infer from (4.2) that α φ 0. We check next that z + [0,ahn]B(v,e) С xn + [0,hn]B(v,e) (4.4) Indeed, the subset xn + [0, hn]B(v, e) being convex, ζ + [0, ahn]B(v, e) = (1 - a)(xn + hnwn) + a(xn + [0, hn]B(v, e)) Cxn+ [0,hn]B(v,e) Finally, ζ belongs to the ball Вк(хо, Vn)- Indeed, ll^-^nll < (1-α)/ιη(||ϋ||+ε) an,d thus, the norm of ν being equal to one, II2 — xq\\ < \\z — xn\\ + \\xn — xq\\ < hn(l + e)+ l/2n < 1/n Property (4.1) implies the existence of и € Τκ(ζ) Π Β(υ,ε/2). Therefore there exist sequences uv converging to и and kp > 0 converging to 0 such that ζ + kpup belongs to K. By (4.4), it also belongs to xn + [0, hn]B(v, ε) for ρ large enough. Hence there exists ρ such that ζ + kpup € Q and up € Β{υ, ε). Taking χ := ζ + kpup in inequality (4.3), we get ((1 - a)hn + kp)\\wn\\ - kp\\up-wn\\ < < \\((1 - a)hn + kp)wn + kp(up - wn)\\ = ||(1 - a)hnwn + kpup)\\ = \\x - xn\\ < (1 - α)/ιπ||^π|| +<5&ρ||ϊχρ|
138 4 - Tangent Cones and thus, using the fact that \\up — wn\\ < 2e, Αφ||^π|| < 5Α;ρ||ϊχρ|| + kp\\up — wn\\ < kp(6(l + e) + 2e) Dividing by kp > 0, we obtain 1-е < \\wn\\ < <5(l + e) + 2e < 3e Hence e > 1/4, the looked for contradiction of the choice of е. П 4.2 Tangent Cones to Convex Sets For convex subsets K, the situation is dramatically simplified by the fact that the Clarke tangent cones and the contingent cones coincide with the closed cone spanned by К — χ: Proposition 4.2.1 (Tangent Cones to Convex Sets) Let us assume that К is convex. Then the contingent cone Тк(х) to К at χ is convex and CK(x) = TbK(x) = TK(x) = SK(x) We shall denote by Тк(х) the common value of these cones, and call it the tangent cone to the convex subset К at x. Proof — We begin by stating the following consequence of convexity: V ν € SK(x), 3 h > 0, such that V t € [0, h], x + tv € К since we can write that for any t € [0, h] x + tv = ί 1 - — ] χ + —(χ + hv) is a convex combination of elements of K. It is enough to prove that Sk(x) is contained in Ск(х)- Let v '■= (y — x)/h belong to Sk{x) (where у € К and h > 0) and let us consider sequences of elements hn > 0 and xn € К converging to 0 and χ respectively. We see that vn ·.= (y — xn)/h converges to υ and that τ. ί hn\ hn xn + hnvn = [1--г]хп + -гУ € К
4.2. Tangent Cones to Convex Sets 139 since it is a convex combination of elements of K. D The negative polar cone of the tangent cone Тк(х) to a convex subset, is called the normal cone to К at χ and is denoted by NK(x) := TK(x)- = SK(x)- It can be easily characterized by: Nr(x) = {p € X* | max < p, у > = <p, χ >} We observe that ί г) NK(x) С b(K) ) ii) the "asymptotic cone" Ъ{К)~~ С Тк{х) Theorem 4.2.2 Any closed convex subset of a Banach space is sleek. Proof— Let К be a closed subset of a Banach space X. We begin by proving that the graph of the set-valued map К э χ --> NK(x) is closed in Χ χ Χ*, when X is supplied with the norm topology and X*' with the weak-* topology: Let us consider sequences of elements xn € К converging to χ and pn € Νκ(χη) converging weakly to p. Then inequalities У у € К, <рп,У><<Рп,Хп> imply by passing to the limit inequalities V у € К, <p,y><<p,x> which state that ρ belongs to Νκ(χ). Hence the graph is closed, so that the set-valued map Τκ{·) is lower semicontinuous thanks to Proposition 1.1.8. □ It may be useful to characterize the interior of the tangent cone to a convex subset.
140 4 - Tangent Cones Proposition 4.2.3 (Interior of a Tangent Cone) Assume that the interior of a convex subset К С X is not empty. Then л><Л n J Furthermore, the graph of the set-valued map К Э χ ~> Ы(Тк(х)) is open. For the convenience of the reader, we list in Table 4.3 some useful formulas of the calculus of tangent cones to convex subsets derived from convex analysis (see [35, Section 4.1.].) The subsets K, Ki, L, M, ... are assumed to be convex, the spaces Χ, Υ are Banach spaces and £(X, Y) denotes the space of continuous linear operators. Remark — Property TKlnK2(x) = ТК1(х)ПТк2{х) is false when assumption 0 € Int(UTi — K-i) is not satisfied. Take for instance two balls K\ and Кч tangent at a point x. The tangent cone to the intersection {x} is reduced to {0}, whereas the intersection of the tangent cones is a hyperplane. This shows that we cannot dispense with the constraint qualification assumptions in the calculus of tangent cones to inverse images and intersections. □ We provide examples of tangent cones to some specific convex subsets: Examples — 1. We observe first that an element ν belongs to Tr^ (x) if and only if V{ > 0 whenever χι = 0. — 2. We denote by Sn := 1хеЩ. | f^Xi = ll the probability simplex.
4.2. Tangent Cones to Convex Sets 141 Table 4.3: Properties of Tangent Cones to Convex Sets. (1) > If χ € К С L С X, then Тк(х) С TL(x) & NL(x) c Νκ(χ) (3) > If Xi € Ki С Xi, (i = 1, ...,n), then ΓΠ»=ι^(ζι,..·,ζη) = Π?=ι?№;) (4)a) > If Л € C(X, Y) and ж € if С X, then TA{K)(Ax) = A{T_K{x)) NA(K)(Ax) = A*-*NK(x) (4)b) > If Ku K2 С X, χ» € gj, г = 1,2, then Τκ,+κ^Χ! + X2) = ΤΚ}(χΐ)+Τκ2(Χ2) Νκλ+κ2{χι +χϊ) = ^κ1(χι)ηΝΚ2(χ2) In particular, if x\ € К and жг belongs to a subspace Ρ of X, then 2k+p(xi + S2) = TKl(Xl) + P Nk+p(xi + X2) = ΝκΙχ^ΠΡ1 (5) > If L С -X" and Μ С Υ are closed convex subsets and A eC(X,Y) satisfies the constraint qualification assumption 0 € Int(M - A(L)), then, for every χ € L Π A-1 (M), Гьпл->(М) = TL{x)nA-lTM{Ax) NLnA->(M) = NL(x)+A*NM(Ax) (5)a) > If Μ С У is cZoseci convex and if Л € £(X, У) satisfies 0 € Int(Im(A) - M), then for χ € Л_1(М) ΤΑ-4Μ)(χ) = A^TiaiAx) Na->(M)(x) = A*NM(Ax) (5)b) > If Κι, K2 С X are closed convex and satisfy 0 € Int(UTi - K2), then for any χ € Κι Π Κ2 TKinK2(x) = ТК1(х)ПТк2(х) NKlnK2(x) = ΝΚι(χ) + Νκ2(χ) (5)c) > If Κ{ С X, (г = 1,...,η), are closed and convex, χ € f)i=i К and if there exists 7 > 0 satisfying \/xi such that \\xi\\ < 7, f|?-i(-fQ - ^i) ^ 0, then Τ^ιΚί(χ) = (T=i2kf(x)
142 4 - Tangent Cones Figure 4.3: Counterexample: Tangent Cone to the Intersection LnM = {x} ™ $ II W II / TM{x) TLnM(x) = {o} Lemma 4.2.4 The contingent cone Ts"(x) to Sn at χ Ε Sn is the cone of elements ν G Rn satisfying ^2,Vi = 0 L· V{ > 0 whenever ж,- = О (4.5) ΐ=ι Proof— Let us take υ G Tsn(x). There exist sequences hp > 0 converging to 0 and vv converging to υ such that yp ;= χ + /ιρυρ belongs to Sn for any ρ > 0. Then η - / η η > Σϋ* = f Σ 2/ρ.· - Σ жр.· ο 1 ,υΡ \i = l i = l / so that J2?=i vi = 0- On *^е other hand, if Xi = 0, then "Pi 2/Pi.//ip > 0 so that V{ > 0. Conversely, let us take w satisfying (4.5) and deduce that у := χ -\- hv belongs to the simplex for h small enough. First, the sum of the y,· is obviously equal to 1. Second, y,- > 0 when X{ = 0 because in
4.2. Tangent Cones to Convex Sets 143 this case V{ is nonnegative, and, when χι > 0, it is sufficient to take h < Xi/\vi\ for having yi > 0. Hence у does belong to the simplex. □ In the case of general convex cones, we have: Lemma 4.2.5 Let К С X be α convex cone of α normed space X and χ € K. Then TK(x) = K + Kx. Furthermore, ρ € Nk{x) <=> x € Κ, ρ € K~ к <p,x> = 0 <=ϊ χ € NK- (p) where NK-(p) := {x € K\ V q € K~, < q - p,x > < 0}. Assume that po € Int(UT+) and set S := {x € K\ <p0,x>= 1} Then Ts(x) = {veTK(x)\ <Po,v>= 0} (4.6) Proof — Indeed, K + Kx = K-R+x с ТК[х) because, for any ζ € Κ, λ > 0, V/i<l/A, x + h(z-\x) = (l-h\)x + hz € К Conversely, ν €. Τκ(χ), being the limit of vn such that x + hnvn € K, belongs to К + Ra: for all n, because Vn € K — x/hn с K-R+x To say that ρ € Nk(x) means that < ρ, χ >= σκ(ρ), which is equal to 0 if and only if ρ € К", so that the second statement of the lemma ensues. Observe that S can be written in the form К Пр^1(1) and the constraint qualification assumption 0 € Int(po(K) — 1) is satisfied because po(K) is a cone of R containing 1. We then deduce that Ts(x) = ТК(х)Про1Т{1}(1)
144 4 - Tangent Cones i.e., formula (4.6.) D Remark — If we assume that K~ + {x}~ = X* and that X is reflexive, then Тк(х) = К + ~Rx thanks to the Closed Range Theorem 2.4.4. D The next result plays an important role in control theory for proving that controllability and observability for closed convex process are dual concepts: Theorem 4.2.6 (Polar of a Viability Domain) Let X be α reflexive Banach space, F : X ~> X be a closed convex process whose domain is the whole space and К be a closed convex cone. The following statements are equivalent: ' i) \/x&K, F(x) С TK(x) < ^ it) VqeK+, F*(q) П TK+(q) φ 0 where K+ := -K~ = {p € X* \ Vχ € К, <р,х>> 0}. Proof — The first condition is equivalent to УхЕК, Vq€-(TK(x))-, sup <-g,y><0 (4.7) yeF{x) But we know from the latter lemma that q € — (Τκ(χ))~~ = —Νκ(%) if and only if x € NK-(-q) = (К-+Щ)- = (К++Щ)+ = (TK+(q))+ On the other hand, since the domain of F is the whole space X, Proposition 2.6.4 implies that sup < — q, у > + sup < ρ, χ > = 0 y£F(x) peF*{q) Therefore, condition (4.7) is equivalent to Vq€K+, Ух€ (TK+(q))+, sup <p,x>>0 (4.8) p£F*{q)
4.2. T&ngent Cones to Convex Sets 145 Now, since the images of F*(q) are weakly compact, the Separation Theorem implies that the intersection of F*(q) and TK+(q) is not empty, (i.e., that 0 belongs to the closed convex subset F*(q) - Τκ+(ΐ)), if and only if Ух EX, 0 < a{F*{q)-TK+{q\x) It is enough to observe that this statement is equivalent to (4.8), because K M' K W } \ +oo if x i (TK+(q))+n Definition 4.2.7 (Pseudo-Convex & Star-Shaped Sets) A set К of a normed space X is said to be star-shaped around χ ζ Κ if ЧуеК, VAe[0,l], x + X{y-x) e К and pseudo-convex at χ € К if and only if either one of the equivalent properties i) TK(x) = S^(x) =: ΰ^ψ ii) К С χ + Τκ(χ) holds true. We observe the following Lemma 4.2.8 If К С X is star-shaped around χ e K, then it is pseudo- convex at this point. Convex subsets are star-shaped around each of their elements and thus, share with them some properties. For instance: Proposition 4.2.9 If Χ, Υ are normed spaces, A € £(X,Y) is a continuous linear operator and if К С X is pseudo-convex at some point χ of K, then ATK(x) = TA{K)(Ax) In particular, when К is convex, we have for every у € A(K) Π АЩф)) = ТА{к){у) хекг\А~1(у)
146 4 - Tangent Cones Proof — Take ν € TA(K) (Ax) and let hn > 0 and i)„ 6 У be sequences converging to 0 and ν respectively such that Ax + hnvn € A(K) С Ах + А(Тк(х)) Therefore the vn's belong to А(Тк(х)) and so their limit belongs to the closure of this cone. □ 4.3 Calculus of Tangent Cones By using the tangent cone Ск(х) instead of sleekness, we can state the following version of the "pointwise inverse function theorem", which follows from Theorem 4.1.10 and the proof of Theorem 3.4.10. Theorem 4.3.1 (Pointwise Inverse Mapping Theorem) LetX be α Banach space, К be a closed subset of Χ, Υ be a finite dimensional vector-space f : X ι—> У be continuously differentiable around an element xq € K. If f'(xo)(CK(xQ)) = Υ then the set-valued map у ~> f"1(y) Π Κ is pseudo-Lipschitz around (f(x0),xo)- In particular, we single out the following consequence: Corollary 4.3.2 If К is a closed subset of a finite dimensional vector-space X, then xq € Int(UT) if and only if Ck(xq) = X. If X is a Banach space, xq € Int(UT) if and only if there exist constants a € [0, l[, η > 0 such that Vx€ BK(x0,η), Βχ С Тк(х) + αΒχ D 4.3.1 Intersection and Inverse Image We are ready to derive the basic results of the calculus of tangent cones.
4.3. Calculus of Tangent Cones 147 Theorem 4.3.3 Let X andY be Banach spaces, L с X and Μ С Υ be closed subsets, f : X ι-> Υ be continuously differentiable around an element xq € L Π f"1(M). If Υ is finite dimensional, we posit the pointwise transversality condition f(xo)(CL(xo)) - CM(f(xo)) = Υ Then Ti(xo)nf'(x0r1(TiI(f(xo))) = Tinrl{M)(x0) and CL(xo)nf'(xo)"1(CM(f(xo))) С CLnf-i{M)(x0) Otherwise, to obtain the same results in the case of any Banach space, we have to replace the pointwise transversality condition by the local transversality condition.· there exist constants с > 0, a € [0, l[ and η > 0 such that \/χ€ΐΠΒ(χ0,η), \/y€MnB(f(x0),v) BY с f(x) (TbL(x) П cBx) - TM(y) + aBY . Furthermore the above assumption yields that the set Lflf~1(M) is sleek (respectively derivable) whenever L and Μ are sleek (respectively derivable.) Proof — Let us prove for instance the inclusion for the Clarke tangent cones. Consider the closed subset К := Lf\f-l{M) and take any sequence of elements xn € К which converges to x. Let us pick any и € Cl(xq) such that f'(xo)u € Cm(/(^o))· Hence for any sequences hn > 0 and xn € К converging to 0 and xq respectively, there exist sequences un and vn converging to и and f'(xo)u respectively such that, for all n > 0, xn + Кип € L к f{xn) + hnvn € Μ
148 4 - Tangent Cones We now apply Theorem 4.3.1 to the subset L χ Μ of Χ χ Υ and the continuous map / θ 1 associating to any (x, y) the element f(x) — y.. It is obvious that the transversality condition f'(xo)(CL(xo))-CM(f(x0)) = Υ implies the surjectivity assumption of Theorem 4.3.1. The pair (xn + hnun, f(xn) + hnvn) belongs to L χ Μ and (f Bl)(xn + hnun,f(xn) + hnvn) converges to 0 because / is continuous at x. Therefore, by Theorem 4.3.1, there exist I > 0 and a solution (xn, i/n) £ I x Μ to the equation Cfei)(i„,y„) = о (i.e., yn = f(xn)) such that ||a;n + hnun - xn\\ + \\f(xn) + hnvn - yn\\ < lhn\\{f(xn + Kun) - f(xn))/h η Vn\\ Hence un := (xn — xn)/hn converges to u, and for all η > 0, we know that xn + hnun belongs to L Π f"1 (M) because xn + hnun = xn and f(xn + hnun) = yn. When У is an arbitrary Banach space we apply Theorem 3.4.3 in exactly the same way. Π We list now three useful corollaries of this theorem: Corollary 4.3.4 Assume that Χ, Υ are Banach spaces, Μ с Υ is a closed subset and that f : Χ η-» Υ is a continuously differentiable map around an element xq € f"1(M). When the dimension of Υ is finite, we suppose that bn(/'(x0)) + CM(f(x0)) = Υ Then Tf-4M)(xo) = f'(xo)-1 (TM(f(xo))) Furthermore 7/-1(m)(*o) = f'(xo)-1 (TbM(f(x0)))
4.3, Calculus of Tangent Cones 149 and Cf-4M)(xo) D f'ixo)-1 (CM(f(xo))) Otherwise, to obtain the same result in the case of a Banach space Y, we have to assume that there exist constants с > 0, a € [0, l[ and η > 0 such that ( \/χ£Β(χ0,η), VyeB(f(xQ),rinM \ By С cf(x)(Bx)+TM(y)+aBY Corollary 4.3.5 Let K\ and K2 be closed subsets of a Banach space X andx € K\ П Яг- If the dimension of X is finite, we assume that CKl(x) ~ CK2(x) = X Then Ί*Κι{χ)ηΤΚ2{χ) С ТК1пк2(х) & Τ*Κι(χ)ηΊ\2(χ) = TbKinK2(x) and СкЛ00) п Ск2(х) С CKlnK2(x) Otherwise, the same result holds true when X is a Banach space if we suppose that there exist constants с > 0, a € [0,1[ and η > 0 such that Vi eKi Γ)Β(χ0,η), Vy £Κ2Γ)Β(χ0,η) By С (тьК1(х)ПсВх)-ТК2(у) + аВу Finally, for a finite intersection, we can state: Corollary 4.3.6 Let us consider n closed subsets Щ of a Banach space X and xq € Π£=ι Κί· When the dimension of X is finite, we assume that n \/Vl,...,vneX, f](CKi(xo)-Vi) Φ 0 i=l Then pifeCzo) = T^^xo)
150 4 - Tangent Cones and η If the dimension of X is infinite, we assume that there exist constants с > 0, a € [0, l[ and η > 0 such that ' VxitEKiPl B(xq, η), V V{ <E X, < 3 wi € X, 3 и € Π?=ι (Гл:{ (zj) - «i - ϊ«ΐ) such that ||u|| < cmaxi=i,...,n ||ϋ;|| & ||ϊ«ϊ|| ^ qmaxi=i1...1n |K|| to get the same conclusions. Proof — It is enough to apply Theorem 4.3.3 to Μ = {0}, the subset Hf=1Ki x D, where D :== {(x,.. .,x)}xex, and the map / defined on this set by f(x, y) = χ — у. □ 4.3.2 Example: Tangent cones to subsets defined by equality and inequality constraints Consider a closed subset L of a Banach space X and two continuously different iable maps g := (gi,...,gp):X ^ W L· h := (hlt..., hq) : X н-> № defined on an open neighborhood of L. Let К be the subset of L defined by the constraints Κ := {χ € L | gi(x) > 0, i = l,...,p k hj{x) = 0, j = 1,...,?} We denote by I(x) := {i = 1,..., ρ \ gi{x) = 0} the subset of active constraints. Proposition 4.3.7 Let us posit the following transversality condition at a given χ € К: ' 0 ti(x)CL(x) = Hq < ii) 3 wo € Cl(x) such that h'(x)vo = 0 and Vi€/(s), <gi(x),v0> > 0
4.3. Calculus of Tangent Cones 151 Then и belongs to the contingent cone to К at χ if and only if и belongs to the contingent cone to L at χ and satisfies the constraints Vi€/(s), <gi(x),u>> 0 & Vj = l,...,g, h'a{x)u = 0 Proof— It is clear that for every υ € Тк(х) С Tl(x), equations h'Ax)v = 0, j = 1,..., q are satisfied and inequalities < д[{х), ν >> 0 are satisfied for all active constraints i € I(x)· The converse inclusion is a direct consequence of Theorem 4.3.3 with Υ := Rp χ Rq, f = gxh, Μ := Rp+x {0} It is sufficient to check that the assumption implies the transversality condition (g'(x) χ ti(x)) (CL(x))-TRP+x{Q}(g(x),0) =R"xR« Indeed, take (y, z) € Rp x R9. There exists a solution υ € Cl(x) to the equation h'(x)v = ζ by the first assumption. Let a := mm <д[{х),у0> and λ := max(0,yi - < g[(x),v >,...,yp- < g'p(x),v >)/a We set ai '■= < 9i(x), V + \V0> -yi By construction, ai > 0 for any i € /(ж), so that a := (αϊ,... ,αρ) belongs to the contingent cone to R+ to g{x). Hence w := Xvq + υ belongs to the convex cone Cl(x) and is a solution to the equations g'(x)w — a = у and h'{x)w = &, so that (У, z) € (g'{x) x ti[x))u - TRP+x{0}(g(x), h(x)^j Π Remark — Observe that the above proof implies that if L = X in Proposition 4.3.7, then an element и € Ск(х) if and only if Vi€/(s), <g[{x),u>> 0 & ν;' = 1,...,σ, ^-(s)u = 0 □
152 4 - Tangent Cones Table 4.4: Properties of Contingent Cones to Derivable Sets in Finite Dimensional Spaces. (5) > If L С X and Μ С Υ are closed derivable subsets, / € ^(Χ,Υ) is a continuously differentiable map and χ € L П f~1(M) satisfies the transversality condition f(x)CL(x)-CM(f(x))=Y,then Tinf-4M)(x) = П(х) П/'(х^ТмМх)) (5)a) > If Μ с У is a closed derivable subset, / € Сг(Х,У) is a continuously differentiable map and χ € f-^M) satisfies bn(/'(x)) - CW№)) = Y, then Tf-4M)(x) = fix^TMifix)) (5)b) > If K\ and K2 are closed derivable subsets contained in Χ, χ € К1ПК2 satisfies Скг(х) — Ск2(х) = X-> then ТкгпкЛх) = TKl(x)nTK2(x) (5)c) > If Ki С X, (i = 1,... , n) are closed derivable subsets and χ € Hi Ki satisfies Vt/i = 1,..., η, ΠΓ=ι (C*, (s) - «i) ^ 0, then
4.3. Calculus of Tangent Cones 153 4.3.3 Direct Image Let us consider now two normed vector spaces X and Y, a subset К С X and a differentiable single-valued map / from X to Y. We saw that for any у € f(K), we have D Г(*У?к(х)) С Tf{K)(y) and U f'(x)№c№) С T){K){y) xeKnf-Цу) It is not that easy to find elegant sufficient conditions implying the equality U f'(x)(TK(x)) = Tf{K){y) xeKnf-^y) Theorem 4.3.8 Let X be α finite dimensional vector-space, Υ a normed space, Ω С X be an open subset containing К and f : Ω ι-> Υ be a single-valued map satisfying f(K) Э у ~> Г\у)ПК is pseudo-Lipschitz at {f{xo),xo): there exists a constant I > 0 such that, for any у € f{K) close to f(xo), а(х0,ГЧу)ПК) < Z||y-/(x0)|| Then, if f is Frechet differentiable at xq, we obtain the equality f'(xo)TK(xo) = Tf{K)(f(x0)) Proof— Let υ belong to Т^(^(/(жо))· Then there exist sequences of elements hn > 0 and vn converging to 0 and ν respectively such that /(so) + hnvn = f(xn) € f(K) (4.9) The point is to choose solutions xn € К to the above equation (4.9) such that a subsequence of un := (xn — xo)/hn converges to some u. Such an element и belongs to the contingent cone Tk{xq) and is a solution to the equation f'{xo)u = v.
154 4 - Tangent Cones Since the set-valued map f(K) Э у — К П Г1 (У) is pseudo-Lipschitz at (f(xo),xo) by assumption, there exist a constant I and solutions xn € К to the equation (4.9) such that llzo-Znll < l\\f{xo) - f(xo) - hnvn\\ = ^π||υπ|| Therefore, the sequence of elements un is bounded, so that a subsequence (again denoted) un converges to some u. □ Remark — Therefore, any sufficient condition implying that for some xq € K, the set-valued map 1(к)эу ^ кпг\у) is pseudo-Lipschitz at (f(xo)> ^o) will automatically imply the above equality between the contingent cone to the image and the closure of the image of the contingent cone. One of them is given by the Criterion of Pseudo-Lipschitzeanity (See Corollary 3.4.8.) □ Actually, when the dimension of X is finite, we have another criterion: Proposition 4.3.9 Let Χ, Υ be normed spaces, К С Χ, Α € C(X, Υ) be a continuous linear operator and yo € Y. If the dimension of X is finite and if for some xq € К П А~1(уо) кет(А)ПТк(х0) - {0} then A(TK(x0)) = ТА{К)(А(х0)). Proof— Let υ belong to TA(K)(A(xq))■ There exist hn > 0 converging to 0, vn € Υ converging to υ and xn € К such that A(xQ) + hnvn = A{xn) = A(xQ + hnun) where we set un := (xn — xo)/hn. If the sequence un is bounded, being in a finite dimensional vector-space, a subsequence converges to some u, which belongs to
4.3. Calculus of Tangent Cones 155 Τκ(χο)ι ancl satisfies Au = v. If not, a subsequence (again denoted) \\un\\ would go to oo. Then a subsequence of elements un = un/\\un\\, satisfying equations Aun = υπ/||?χπ||, converges to some и in the unit sphere, which is a solution to the equation Au = 0 and which belongs to the contingent cone Tk(xq)· This is impossible by assumption. □ When the map is no longer linear and the dimension of X still finite, we can prove: Proposition 4.3.10 Let К be α subset of α finite dimensional vector- space Χ, Υ be a normed space, f : X —> Υ a Frechet differentiable function and у о € f{K), xq € Κ Π f~1(yo). Assume that for some ε > 0, the subset f~1(B(yo,e) Π Κ) is bounded. Then condition Ух€Г1(Уо)ПК, kei(f'(x))nTK(x) = {0} implies __U /ΌΌ(Γ*0Ό) = TfiK)(yo) xeKnf-i{yo) Proof— Let ν belong to Tf^(yo) be different from 0. There exist hn > 0 converging to 0, vn € Υ converging to υ and xn € К such that yo+hnvn = f(xn), which are then bounded by assumption. Then a subsequence (again denoted) xn converges to some χ € К satisfying f(x) = y0. 'We set an := \\xn — x\\ and un := (xn — χ)/απ. Then an converges to 0 and the sequence un being bounded in a finite dimensional vector-space, has a subsequence converging to some u, which belongs to TK(x). On the other hand, f(x + OnUn) = f(x) + anf'(x)u + αηε{αη) = yo + hnvn Hence f'{x)u + e(an) = hnvn/an "The left hand side being bounded and vn converging to ν φ 0, we deduce that the sequence hn/an is bounded, so that a subsequence converges to some λ. Therefore f'(x)u = Xv. By assumption, λ must be different from 0. Consequently, υ = f'{x)u/\ € f'(x)(TK(x)) D
156 4 - Tangent Cones 4.4 Normal Cones We introduce in this section the dual concept of tangent cones: the normal cones. We begin with the following characterization of the polar of the contingent cone: Proposition 4.4.1 Let К be a subset of a finite dimensional vector- space X. Then ρ € (Тк(х))~ if and only ' Ve>0, 3 77 > 0 such that My € Κ Π Β(χ,η), (4.10) <p,y-x > < e||y-a;|| Proof — Let ρ satisfy above property (4.10) and υ € Тк(х). Then there exist hn converging to 0 and vn converging to υ such that у := χ + hnvn € Κ Π Β(χ, η) for η large enough. Consequently, inequalities < p, vn >< ε imply by taking the limit that < ρ, υ >< e for all e > 0. Hence < ρ, υ >< 0, so that any element ρ satisfying the above property belongs to the polar cone of Тк(х), even in the case of normed spaces. Conversely, assume that ρ violates property (4.10): There exist e > 0 and a sequence of elements xn € К converging to χ such that (p,xn-x) > er|| Xji X\\ We set hn := \\xn — x\\, which converges to 0, and vn := (xn — x)/hn. These elements belonging to the unit sphere, a subsequence (again denoted) vn converges to some v. By definition, this limit belongs to Tk(x), so that < ρ, υ >< 0. But our choice implies that < p, vn » e, so that < p,v >> e, a contradiction. Π Notice that Proposition 4.1.2 implies that when the Banach space X is smooth, then for any у ф К such that Ик(у) is not empty and for any ζ € Лк(у), J(y — ζ) belongs to the Τκ{ζ)~. The most important justification for dealing with dual concepts, is the availability of the one to one correspondences between closed convex cones and their polar cones, continuous linear operators and
4.4. Normal Cones 157 their transposes, lower semicontinuous convex functions and their conjugates. This requires one to deal with convex notions. Since the Clarke tangent cone is convex, it can then be characterized by its polar cone, which, by analogy with the case of smooth manifolds, can be regarded as the normal cone. Definition 4.4.2 Let χ belong to К С X. We shall say that the (negative) pofar cone NK(x) ■= CK(x)~ = {p€X* | Vv€CK(x), <P,v><0} is the (Clarke) normal cone to К at x. The normal cone Νχ(χ) is equal to the whole space whenever Ck(x) is reduced to zero. The normal cone is related to the closed convex hull of the upper limit of the polar cones to the contingent cones: Theorem 4.4.3 Let X be a finite dimensional vector-space and К be a closed subset of X. Then NK(x) = со (Limsupj,.^,. (TK(y))~) This a consequence of Theorem 4.1.10 and the Duality Theorem 1.1.8. In infinite dimensional spaces, we deduce from Theorem 4.1.13 its dual version Theorem 4.4.4 Assume that X is uniformly smooth and that the norm of X* is Frechet differentiable off the origin. Let К be a closed subset of X. Then NK(x) = со (σ - Limsup^^ {T^{y))~) Unfortunately, since the contingent and intermediate cones are" not convex, duality may just be a blissful wish, since many problems, which are not smooth by nature, force us. to*use naturally these larger tangent cones. The price to pay in terms of loss of information for playing with duality, just to be able to conserve some familiar classical formulation, is indeed too high in many situations. Therefore, these dual concepts we have presented are recommended for convex, or more generally, sleek subsets.
158 4 - Tangent Cones Table 4.5: Properties of Normal Cones In Finite Dimensional Vector Spaces. (3) > If Xi € Ki С Χ{, (i = l,...,ri), then (5) > If L с -X" and Μ С У are cZosed , s € L П /-1(М), / € C1^, У) is continuously differentiable and the transversality condition }'{x)CL{x) - CM(f(x)) = Υ holds true, then NLnf-4M)(x) С NL(x) + f'(x)*NM(f(x)) (5)a) > If Μ С У is a closed subset and if / € C1(X, У) is a continuously differentiable map such that Im(f'(x)) - CM(f(x)) = Y, then Nf-4M)(x) С f(x)*NM(f(x)) (5)b) > If K\ and K2 are closed subsets contained in X and ж € K\ П Яг satisfies (7^ (ж) — Сд:2(:е) = X, then NKlnib(x) С iV^(a;)+iV^2(a;) (5)с) > Ιϊ Щ С X, (г = 1,..., η) are closed, χ €.f]{Ki and V t* = 1,..., n, |T=i (С*< (s) - t/i) φ 0, then
4.5. Other Tangent Cones 159 4.5 Other Tangent Cones 4.5.1 Convex Kernel of a Cone Let us recall that the Minkowski difference К Q L of two subsets К and L of a vector space is the subset f\xeL{K — x) of elements ζ such that L + zCK. Let Γ be a closed cone. We observe that the Minkowski difference ΤθΤ := {y<=X\ y + TcT} is α closed convex cone contained in T. Indeed, it is obviously a cone, contained in Τ (because 0 £ T) and is convex: If у and ζ belong to Τ θ Τ, then y+{z + T) С у + Т с Т so that у + ζ belongs also to Τ θ Τ. Definition 4.5.1 When Τ is a closed cone, the Minkowski difference ΤθΤ is called the convex kernel (or the asymptotic cone) of Τ and is denoted ЪуТ°°. The convex kernel provides an explicit convex subcone of T, which may not be the largest one, but is the intersection of the maximal11 closed convex subcones of T: Proposition 4.5.2 The convex kernel of a closed cone Τ is the intersection of the maximal closed convex cones contained in T. Proof— Let V denote the family of the maximal closed convex cones contained in T. Then T°° С f] Ρ P£T> Indeed, if χ belongs to T°° and Ρ ε V, then the closed convex cone Ρ + Л+х is a closed convex cone containing Ρ and contained in Τ because χ is in the Minkowski difference. The maximal character of Ρ implies that it is equal to Ρ + R+x, and thus, that ж £ P. On the other hand, \Jp = т pev 11 We say that a closed convex cone К С Г is maximal if for any other closed convex cone Ρ satisfying К С Ρ С Г, we have К = P.
160 4 - Tangent Cones because for any χ ε Τ, the closed convex cone R+x is contained in a cone Ρ £ V-, thanks to Zorn's Lemma: Indeed, for any family of closed convex cones Ki С Τ totally ordered for inclusion, (J{ Ki '1Έ> a closed convex subcone of Γ which is an upper bound of the K{'s. To prove that f]PevP С T°°, pick any χ ε C\pevp anc^ апУ У G ^· Since у belongs to some PeP, then x + y ε P + P С Ρ С Γ π' " Proposition 4.5.3 -Lei К be α closed cone in α normed space and K°° its convex kernel. Then CK(Q) = K°° к 4χβΚ, СК(х) D K°° + Rx Proof— Fix χ ε Κ. Take any element λχ+y ε Έίχ+Κ°° where у ε Κ°° and consider any sequence of elements hn > 0 converging to 0 and xn ε Κ converging to x. Since xn + hn(Xxn + y) = (1 + Xhn)xn + hny ε Κ we infer that Xx + y belongs to Ск(х). Consider now ν ε C^(0) and assume for a moment that it does not belong to K°°. Then there exists χ ζ. Κ such that χ + ν £ К, and thus, ε > 0 with (χ + ν + εΒ) П К = 0 But for h > 0 small enough, we know that there exists e ε εΒ with Лж + h(v + e) = /1(2; + щ + e) ε if Since if is a cone, x + v + e belongs to K, a contradiction. □ We can interpret the second statement of Proposition 4.1.6 by saying that the Clarke tangent cone is contained in the convex kernels of the contingent cone and adjacent cone. If one is interested in dealing with convex subcones of the contingent cone without requiring regularity, it may be advantageous to use the convex kernel of the contingent cone. 4.5.2 Paratingent Cones Bouligand had also introduced in the thirties the concept of paratingent cone Ρχ {x) to К at χ ε Κ, which can be regarded as a symmetric concept to the one of Clarke tangent cone, in the sense that it is defined as Κ -χ' Ρκ(χ) := Limsuph_f0+x,_fifX—-—
4.5. Other Tangent Cones 161 Figure 4.4: The Menagerie of Тал gent Cones: they may be all different K- := N~ K+ := {{-X,X)}X>Q {(~,~)}n>0 Subset К := K~ U K+ Clarke Tangent Cone СКЩ Adjacent Cone 7^(0) Contingent Cone Τκ{ϋ) Limsupa._t0rj!C(a:) Paiatingent Cone Pftr(O) Paiatingent Cone Pg (0) Paratingent Cone P£~ (0)
162 4 - Tangent Cones Actually, this cone may be too large to be useful: For instance, take X := R2 and if:=Rx R+. Then TK(x,0) = CK{x,Q) = Κ and PK(x,0) = X Therefore, we are led to introduce a more precise concept. Definition 4.5.4 (Paratingent Cone) Let К С X be α subset of α normed space X, L a subset of X and χ £ Lf\K. The Bouligand paratingent cone Ρχ(χ) to К relative to L at χ is defined by pk(x) ■= lv\ liminf da{x' + hv)/h=ti\ У /ι->0+, x'-*lx ) When L = K, we set PK(x) := Pg(x). In other words, _ К — ν VxGLnK, P%(x) = Limsuph_>0+iy_>z<x—— One of the current choices of L is the boundary dK of К or the complement KolK. Lemma 4.5.5 If К is open, then ЧхедК, PfcK{x) = Pg{x) Proof — Since ΡχΚ(χ) С P£(a;), we have to show that any ν ε P# (ж), i.e., satisfying 3 xn —»·~ x, 3 hn —> 0+, 3 vn —► ν such that xn + hnvn ε Κ belongs also to ΡχΚ(χ). Since К is open and χ £ К, there exists kn e]0, hn[ such that yn := xn + knVn belongs to dK and converges to χ in dK. Take In '■= hn — kn > 0, which converges to 0 and note that Уп ~i~ bnVn = Xn ~f- flnVn belongs to К for any n. Hence υ does belongs to P^K(x). □ We observe at once that VselnL, p£(x) = -p£{x) and in particular, that PK(x) = —Рк(х) is symmetric.
4.5. Other Tangent Cones 163 Proposition 4.5.6 Let L and К be closed subsets of a normed space X. The graph of the set-valued map P% : К П L ~» X is closed: VxeKHL, Limsupy^LXP%(y) с Р£(х) Proof— Consider a sequence of elements (xn,vn) of the graph of the restriction of P£(·) to Κ Π L converging to (x,v). There are sequences xmn £ L converging to xn in L, hmn converging to 0+ and vmn converging to vn such that xmn + hmnVmn belongs to К for every n. We then can associate with any к > 0 integers m,k and nk such that max(||a;„fc -x)),))xmknk -s„J|, hmkTlk, \\vnk -v\\,\\vmk7lk -υηκ\\) < 1/k By taking Ук := ХткПк-1 "к := hmk7lk, Vk := Vmk7lk we observe that y^ £ L converges to x, hk > 0 converges to 0, υ к converges to υ and that ук + hkVk belongs to К for every к > 0. Hence υ belongs to Pk{x). □ A celebrated theorem of G. Choquet states that the contingent cone TK(x) and the paratingent cone ΡχΚ(χ) coincide on a residual12 of the boundary dK. Theorem 4.5.7 (Choquet's Theorem) Let us consider a separable Ba- nach space X, К С X and let L be a closed subset of X. Then there exists a residual R С L of L such that Vieinfl, TK(x) = p£(x) Consequently, the graph of the restriction ofTx(·) to Kf\R is closed. Proof — By definition of the upper limit, we have: ' ρκ(χ) = (\,δ,η>ο UyeBl(,,,) Ue]o,i] ((K ~ y)/h + eB) = a>0n ^кф^К-у)1ь\+еВ) = Пп>о Limsupy_^x [Ufce]0,i/n] (K ~ v)/h 12Recall that a residual is a countable intersection of dense open subsets An. Countable intersections of residuals are residuals. Baire's Theorem states that α residual of a complete metric space is dense. A property which is true for every element of a residual is called generic.
164 4 - Tangent Cones Set - Sn(y) := - (J (K-y)/h he]o,i/n] The set-valued map Sn : L Э у ~> 5n(y) is obviously lower semicontinuous with closed values. By Theorem 1.4.13, there exists a residual An с L such that \/x<=An, Limswpy^^Sniy) = Sn(x) Take R := f]n>0 An, which is still a residual. Hence VzEjR, Limsupy^ LXSn(y) = Sn(x) so that Vx<=R, P%{x) = P| Sn{x) = LimBupfc_o+(-K'-a:)/'i = Τκ(χ) ρ n>0 The paratingent cones Ρχ(χ) are complementary convex, i.e., their complements are convex, and the contingent cones are generically complementary convex. This follows from the next subsection. 4.5.3 Hypertangent Cones We introduce the hypertangent cone to if at a; defined as follows: Definition 4.5.8 (Hypertangent Cone) Let χ belong to a subset К of a normed space X. The hypertangent cone Нк{х) to К at χ ζ. Κ is the set of elements υ £ X such that there exist ε > 0, δ > 0 and η > 0 for which Βκ{χ,η) + [Ό,δ]{υ + εΒ) С К In the same way that the Dubovitskij-Miljutin cone (see Definition 4.1.3) to the complement of К is the complement of the contingent cone to K, it is easy to check the following Proposition 4.5.9 The hypertangent cone is an open convex cone contained in the Dubovitskij-Miljutin cone and satisfying HK{x)+TvK{x) с DK(x) Furthermore, V a; £ дК, the complement of Ρχ(χ) is equal to Hg(x) = —Ηκ(χ)
4.5. Other Tangent Cones 165 This ала Choquet's Theorem 4.5.7 imply the following generic regularity result: Theorem 4.5.10 (Shi Shuzhong) Let X be α separable Banach space, К С X be an open subset. Then there exists a residual R с дК of the boundary of К such that VxedK, TK(x) = P%K(x) = -(X\HK(x)) In particular, Τχ(·) is generically complementary convex on dK. Proof— Indeed, Choquet's Theorem 4.5.7 implies that there exists a residual of dK on which Τκ(·) and ΡχΚ(·) coincide. Furthermore, К being open by the last lemma, ΡχΚ(χ) is equal to P§(x) = —P%(x). We it then use the fact that Ρϊί(χ) is the complement of Нк{х)· D it 4.5.4 A Menagerie of Tangent Cones A. D. Ioffe has introduced a systematic way to classify the four main tangent cones which we have introduced, i.e., the paratingent, contingent, adjacent and circatangent cones, as well as the Dubovitskij-Miljutin and hypertan- gent cones. Let К and L be subsets of a normed space X and χ £ Κ Л L. We denote by the notation ]{an —> a) the fact that an = a for all n. The subset TkRS(K,x) denotes the cone of elements υ £ Χ satisfying Q(xn -+l x), R(hn -> 0+), S(vn —> v), xn + hnvneK where Q(-),S(·) £ {V, 3, f} and R{-) £ {V, 3}. Hence the cones we have introduced can be written ' TK(x) τνκ{χ) CK{x) PkW DK(x) . HK(x) = %^X) = 4^{K'X) = Tfa{K,x) = Tii3^K,x) = TU{K>X) = ^ν,ν,νί-^!37)
166 4 - Tangent Cones The largest of these cones (when L = K) is the paratingent cone Tif-g 3(if, x) and the smallest is the hypertangent cone T^VV(K, x). The proofof Proposition 4.1.6 implies that the cones T^ s(K,x) are convex and are contained in the convex kernel of TqRS(K:x): T^s(K,x) + T§RS(K,x) с T«RS{K,x) We also observe the following complementary property: recall that К denotes the complement of К and set V := 3, 3 = V and f = t· Then the complement of TqRS(K, x) is equal to Т~~~-(К,х) In infinite dimensional space, the complexity is compounded by the fact that (at least for the properties £(·)) we can use either the strong or weak convergence.... It is clear that we can meet in the proofs of theorems various instances of these tangent cones, but few of them enjoy useful properties. As far as the experience of the authors and some of their colleagues is concerned, the contingent cone plays a major role both in optimization theory and viability theory, the use of the Clarke tangent cone is necessary when lower semicontinuity is required or convexity is needed (in this case, it may be still better to use convex kernels whenever the Clarke tangent cone is degenerate.) The paratingent cones are the only ones whose graphs are closed, a property which is quite important indeed, and which can be transmitted generically thanks to Choquet's Theorem. The introduction of the cones T^R s(K,x) can also be justified as particular cases of asymptotic tangent cones to (constant) sequences of subsets. This is the topic of the next section. 4.6 Tangent Cones to Sequences of Sets Let X be a normed vector space. We consider a sequence of subsets Kn с X of X and their upper and lower limits denoted by if" and КУ respectively. We shall adapt the concepts of contingent cones and intermediate cones to sequences of subsets Kn in the following way: Definition 4.6.1 We shall say that the circatangent cone to the sequence (Kn)n at χ £ КУ is the set С\к Лх) defined by ^(Kn)(x) := ■L,iminfn_).00i/fn3Xn_).x^n_>o+—r
4.6. Tangent Cones to Sequences of Sets 167 IfxGK^, we shall say that is the paratingent cone to the sequence (Kn)n at x. Remark — When we consider a constant sequence Kn := K, we see that circatangent cone C^KnJx) coincides with the Clarke tangent cone CK{x). In the case of the constant sequence Kn := K, the paratingent cone PL· Jx) coincides with the paratingent cone Рк(х)- □ The upper limit of the paratingent cones to a sequence of Kn is contained in the paratingent cone to the sequence of the Kn's: Proposition 4.6.2 Let us consider a sequence of subsets Kn С X of X and their upper limit KK Then LimsupKnBXn^xTKn{xn) С UmsupKn3Xn^xPKJxn) с PlKn){x) Proof — It is the same as the one of Proposition 4.5.6. □ We provide a characterization of the polar cone to the paratingent cone to a sequence of subsets: Proposition 4.6.3 Let X be a finite dimensional vector-space, Kn be a sequence of closed subsets of X. Then ρ ε ί PL· Лх)) if and only if there exists a sequence xn £ Kn converging to χ such that V e > 0, 3N: η>ϋ such that Vn>N, V xn £ Kn ΠΒ(χη,η), < ρ, χη — xn > < e\\x η Xn || Proof— Let ρ satisfy above property (4.11) and ν £ PL· Jx). Then there exist hn converging to 0 and vn converging to ν such that for η large enough xn ■= Xn + hnVn £ Kn ηΒ(χη:η) Prom inequalities < p: vn >< ε we deduce that < ρ, ν >< ε for all ε > 0. Hence < ρ, ν >< 0, so that pe (P(W*))~ even in the case of normed spaces.
168 4 - Tangent Cones Conversely, assume that ρ S PL· Ax) violates property (4.11): There exist ε > 0 and a sequence xn £ Kn converging to χ such that VP) 3?n ^n) •*> £ 11^n Xn\\ Set hn := \\xn— Xn\\ (which converges to 0) and vn := (xn — Xn)/hn. These elements belonging to the unit sphere, a subsequence (again denoted) vn converges to some υ £ PL· Ax), so that < ρ, υ >< 0. But our choice implies that < ρ,υη » ε, so that <ρ:υ >> ε, a contradiction. Π The circatangent cone to a sequence of subsets enjoys the same properties as the Clarke tangent cones: it is always a closed convex cone and is actually equal to the lower limit of the contingent cones in finite dimensional spaces: Theorem 4.6.4 Let us consider α sequence of closed subsets Kn С X of α Banach space X and an element xq = linin_юо хп (where xn £ Kn) of their lower limit K^. Then 1. — The circatangent cone С\к Лхо) is always a closed convex cone 2. — The lower limit of the contingent cones Тлп(хп) is contained in the circatangent cone: Liminfn_>0o,i<:„3a:„-»a:T,Kn(a;Tl) С C^K^{xQ) and equality holds when the dimension of X is finite. Proof of the first statement — It is similar to the one of Proposition 4.1.6. Proof of the second statement — Pick an element ν £ LiminfA-n3a!n_a:orK-T1(a;n) with norm equal to one. This means that for any ε > 0, there exists N > 1 such that Vn > N: Vz £ BKJx0:l/n): Β(υ,ε/2)Γ\ΤΚη(ζ) φ 0 (4.12) We shall derive a contradiction by assuming that υ does not belong to C$KJxo). This means that there exist ε £J0, l/4[ and sequences Xn £ B/cfro.l/Zn), /ine]0,l/4n[
4.6. Tangent Cones to Sequences of Sets 169 with {xn + hnB(v,e))nKn = 0 (4.13) Fix N such that (4.12) holds true, η > N and set δ := ε/(1 + ε). We use Ekeland's Variational Principle (Theorem 3.3.1) in exactly the same way as in the proof of Theorem 4.1.9 to verify that (4.13) cannot hold. Proof of the third statement — As for Lemma 4.1.14, the third statement follows from the second one and: Lemma 4.6.5 Let Kn be nonempty closed subsets of α reflexive Banach space and xq = ϋπ^^^ xn (where xn ε Κη). Then C\Kn){x0) С Liminf^g^ Proof— The proof is analogous to the one of Lemma 4.1.14. Π When the subsets Kn are convex, we obtain the following relations: Proposition 4.6.6 Let us consider a sequence of closed convex subsets Kn of a reflexive Banach space X and its upper and lower limits K^ and K^. Then for every χ ε Кь: Ткъ{х) С Ъ]хаЫКпЭХп^хТКп(хп) and Τκί{χ) С LimsupKn3Xn^xTKn(xn) Assume now that X is a Hilbert space and that the lower limit is equal to the sequentially weak upper limit σ — Limsupn_>00i;srTl and denote by К the common value. Then Liminfn^oo^sz^xTVJzn) = TK(x) Proof — 1. Let us take χ ε Kb and ν £ SKb(x), the cone spanned by КУ —x. Hence there exists h > 0 such that χ + hv ε Кь. Consider any sequence of elements xn belonging to Kn and converging to x: and let yn denote the projection of x+hv onto Kn. Since yn converges „to χ + hv, the sequence of elements vn := (yn — xn)/h converges to v. Furthermore, for hn ε]0, h[ xn + hnvn = ί 1 - -γ J xn + -~yn ε Kn and we infer that vn belongs to the contingent cone TKn(xn).
170 4 - Tangent Cones Hence г/, the limit of the vn's, belongs to the lower limit of the tangent cones TKn(xn)· — 2. Let χ belong to K$ and υ := (у — x)/h be chosen arbitrarily in the cone spanned by K^—x. Denoting by xn and yn the projections onto Kn of χ and χ + hv respectively, we infer that vn := (yn — xn)/h belongs to the tangent cone Ткп (xn) and converges to v. — 3. We deduce from Corollary 7.6.5 below that NK(x) = a-Limsupn_>00ii<Ti3XTi_>xiVA:„(a;n) By transposition, Theorem 1.1.8 implies that L\mioiTl^00iKn3Xn^xTKn(xn) = TK{x) □ We shall derive from the Stability Theorem a formula for the circatan- gent cone of an inverse image. Theorem 4.6.7 Let X and Υ be two Banach spaces and A £ C(X,Y) be a continuous linear operator. Consider sequences of closed subsets Ln с X and Mn С Υ. Let us assume that there exist constants с > 0, a £ [0,1[ and η > 0 such that \fxn £ Kn ПΒ(χ0,η), yn ε ΜηΠΒ(Αχ0,η) BY С A (T*n(хп) ПсВх)- Гм„(у„) + aBY Let x belong to the lower limit of the sequence Ln Π A~1(Mn). Then C\Ln){x)nA-'C\Mn){Ax) с СЬпА-чт(х) Proof — Take any sequence of elements xn ε Ln Г\А-1{Мп) which converges to χ and any и ε C$L Jx) such that Au ε CiM ΛΑχ). Hence for any sequence hn > 0, there exist sequences un and vn converging to и and Au respectively such that, for all n > 0, xn + h к Axn + hnvn ε Mn We now apply Theorem 3.4.5 to the subsets Ln χ Mn oiXxY and the continuous linear operators Α θ 1 associating to any (x: y) the element Ax — y. The stability assumptions of this Theorem are obviously satisfied. The pair (xn + hnun, Axn + hnvn) belongs to Ln χ Mn and (Α θ l)(xn + hnun, Axn + hnvn) converges to 0
4.7. Higher Order Tangent Sets 171 Therefore, by Theorem 3.4.5, there exist I > 0 ала (xn, yn) £ Ln χ Μη such that, for η large enough, (Α θ l)(xn, yn) = 0 and \\xn + hnun — xn\\ + \\Axn + hnvn -yn\\ < lhn\\Aun — vn — 0|| Hence un := (xn — xn)/hn converges to u: and we observe that for all ,n > 0, xn + hnun belongs to Ln П A~1(MTl). □ 4.7 Higher Order Tangent Sets The contingent cones are upper limits of the sets (K — x)/h when h > 0 converges to 0. We may need more accurate approximations, and consider the sets of cluster points of mi/l-order differential quotients К — x — hv\ — ··· — hm~1vm^-i when h > 0 converges to 0. Definition 4.7.1 Let χ belong to α subset К of α normed space X and Vi,... ,vm-i be elements of X. We say that the subset m(m), ч τ- K-X-hVx hm~1Vm-l T^ '{x,v1,...vm-.1) := Limsuph_+0+ — is the mi/l-order contingent subset of К at (χ,υι,... ,vm_i). We see at once that Т^-'(х) = Тк(х) is the contingent cone to if at a; and that if T^(x,Vi,...,vm-i) is not empty, then necessarily, υ1£Τ£){χ),υ2€Τ£)(χ,υ1), ... .VieT^'ti,!)!,...,»^) Naturally, we may need more regularity and ask for stronger limits of the subsets (K—x — hv\ hm~1vm-i)/hm, as in the case of first order approximations. Definition 4.7.2 Let χ belong to a subset К of a normed space X and Vi,.. .,vm-i be elements of X. We say that the subset mbfmb \ τ- ■ с K — X—hVi km~1Vm-i T%m)(x,v1,...vm-1) := Limmf^o+ r^ —
172 4- Tangent Cones Table 4.6: Properties of m -order Contingent Sets (i) (2) (3) (5) > > > > If К С L and χ ζ. Κ, then If Ki С X, (г = 1,..., те) ала а; £ (J4 ifi, then Γυ?)1ΛΓ.·(:Ε'υι'···'υ'"-1) = ^еЦх)Тк!](х^1,-- where I(x) := {г \ χ £ ifj} If Ki С -Xi, (г = 1,..., η) and э^ £ ifi, then m(77l) / ГТ" ЛГ ^'Εΐ' * ' ' ' 'Ет1' ^1,1) ■ ■ ■ ι ^Ί,η» ■ ■ " ) ^m—1,1) · ■ ■ С ГК=1Г*7 (xi>vl,ii--->Vm-l,i) _ If Κι С -X", (г = 1,..., η) and χ £ |~|{ i^i, then Τ^^Ζ,Τ*,...,«„_!) С [^l^Wl,- ■,^m-l), ι Vm—Ι,η) ■,Vm-l) is ihe mi/l-order adjacent subset o/ Uf at (χ, νχ,... , um._i) and that r,(m), ν τ -c K-y-hvx hm~1vm-i Lfc [x,v1}. ..vm-\) := Limint/l_)o+,s-lKi r^ is the mth -order circatangent subset of К at (x,V\,... ,vm_i). We shall say that К is mi/l-derivable at x, Vi,..., um_i if the mth-order contingent and adjacent sets coincide. They are closed subsets satisfying, for any A > 0, T^\x, At/X> Α2υ2,..., Ат-Чт_!) = XmT^\x,v1,v2,... ,t/m_i) Τ^ζ,λ^,λΧ... , Α™"1^-!) = A-^Wi,^... ,t/m_i) (4.14) As for the mi/l-order contingent sets,- the adjacent set T^(m)(a;,vi,...,vm_i) (respectively the circatangent set C™ (x, Vi,..., um_i)) is empty if one of the г/j's does not belong to T^°' (x, vi,... ,Vj-\) (resp. C^'(x,Vi,... ,Vj-i)) for j = 1,... , m — 1.)
4.7. Higher Order Tangent Sets 173 We observe that m—l) С T^m)(a;,iii + i;i,...,iim_i + i;m_i) and , Hi,. .. , _-m_l ,Vl,...,Vm_i) С C*gl){x,ul + vi,...,um--i+vm-i) So that, taking into account (4.14), we obtain the following property of m"1-order circatangent sets: v/л η ХтС(£1)(х,и1,...,ит„1) + цтС(£1)(х,у1,...,ут^1) VA,M>0, _____ r(m)( Aui + μν-ί Am~1um_i +μ7η~1υ7η_ι C °* {X' λ + μ '···' (Α + μ)--1 j We also observe that f Cj^l)(a;,i;i,...'ym_i) [ := Liminf^-*o+,у-*кх,^-*!>!,. .,f^_1-*t>m_1 so that if (^"^(ζ,υι,... ,Vj-i) ^ 0, then 0 ε (^(ζ,υι,... ,vm_i). Example — Consider for instance the mi/l-order contingent set Тду (ж). We obtain that for every χ ε Int R™ and all Vi ε Rn r^i-r.vi,...,^.!) = Rn For every χ from the boundary of R™ we have «m ε Tg? (_,«-.■■ ■.«·»--) = T$g\x,v1,...,vm-1) if and only if vmi > 0 whenever a^ = vit = · ■ ■ = v^m_1)i = 0 and vmie'Ri£xi = vli = --- = V(k-!){ =0, vki > 0 for 0 < fc < m - 1 Π K-y-hv[ hm-1-' 'm-1
174 4- Tangent Cones Table 4.7: Properties of mth-order Adjacent Sets. (i) (2) (3) (5) > > > > If К С L and χ ζ. Κ, then If i^ С X,~ (i = 1,..., n) and a; e Ц -Ki. then where /(ж) := {г | a; £ Ki} If -FQ С Хг-, (г = 1,..., η) and э^ £ ifj, then ГТ" AT ''El' ' ' ' '•Еп'г'111' · ' · ι^Ι,η' ■ · · ι^πι—Ι,Ι) · · · >Vm—l,n) If Ki С X, (г = 1,..., η) and χ £ f|i -K"ii then The formulas enjoyed by the contingent and adjacent cones can be easily extended to the mi/l-order contingent and adjacent sets. They are presented in Table 4.6 and Table 4.7. However, the formulas for the direct and inverse images involve higher ordeT derivatives and are presented in the case of second order sets for simplicity. Proposition 4.7.3 Let Χ, Υ be normed spaces and let д £ C2(X, Y), К С Χ, χ £ ~Κ and Μ с Υ be given. Then, for any υλ £ TK{x) \/v2€T^)(x,v1), g'^T^ix^) С rs((2i)(5(a;)l5'(a;)t/i)-^"(a;)(«il«i) and thus ί-1(Μ)(^ι) С д'(хГ1(т^(д(х),д'(х)у1)-У"(х)(у1,у1) lrl(M)1 ' 1J yy ' \ M w^i'V \·υ)"ί) 2 The same results hold true for the second order adjacent sets. Remark — When g = A £ £(X, Y) is a continuous linear operator, the higher order derivatives vanish and we obtain the formulas ' А (г£т) (a;, «χ,... ,«„_!)) С 1^jc){Ax,Av1,...,Avn-i) < ^ A (r£m)(a;,«!,...,«„,_!)) С 3*((™))(Aa;>A«i>...>At/m_i) □
4.7. Higher Order Tangent Sets 175 In order to obtain the converse inclusions, we need to assume a transver- sality assumption. Theorem 4.7.4 Let X and Υ be Banach spaces, L с X and Μ С Υ be closed subsets, f : X >—> Υ be twice continuously differentiable around an element x0 £ Ln f~l{M). If Υ is a finite dimensional vector-space, we posit the pointwise transver- sality condition ?Ы(СьЫ) ~ CM(f(xQ)) = Υ Then ^'^■«On/'W"1 (^(/(so),/'(:co)ui)- У"(х0)(щ,щ)) С ■42) nnz-W^i) and, when Μ is twice derivable, we have the equality instead of the inclusion. Furthermore т¥2)(х0,и1)пГ(х0)-1 (T^2)(f(x0)J'(x0)u,) - y"(x0)(uuUl)) = 3ffi}-W*o,«i) Otherwise, to obtain the same results in the case of any Banach space, we have to replace the pointwise transversality condition by the local transver- sality condition: there exist constants с > 0, a € [0,1[ and η > 0 such that νχ&ΣΓ\Β{χ0:η), УуеМпВ(/(а;о),ч) BY С f'(x) (ТЦх) П cBx) - TM(y) + αΒγ Proof — Let us prove for instance the equality for the second order adjacent sets. Consider the closed subset К := Ln f~1(M). We already know from Theorem 4.3.3 that every щ ε TbL{x0) П f {хоУ1 {TbM{f{x0))) belongs to T^nf_i,Ms(xo). Let us choose such an element u\ and pick any Щ S t£(2)(:eo,?Xi) such that f'(x0)u2 £ Ttf (/(zo),/'Mui) - |/"(a;o)(ui,Ui) Hence for any sequence hn > 0 converging to 0, there exist sequences u2n and v2n converging to u2 and f'(xo)u2 + \f"{xo){u\,U\) respectively such that, for all π > 0, x0 + hnii-L + h\u2n e L к f(x0) + /in/'(a;o)"i + h\v2n ε Μ
176 4- Tangent Cones We apply now-Theorem 4.3.1 to the subset L χ Μ of Χ χ Υ and the continuous map / θ 1 associating to any (ж, у) the element f(x) — y. It is obvious that the transversality condition f'(xo)(CL(xo»-CM(f(x0)) = Υ implies the assumption of Theorem 4.3.1. On the other hand vj(/ θ l)(x0 + hnm + h2nu2n, f(x0) + hnf'{x0)ui. + h2nv2n) converges to 0 because / is continuously differentiable at xq. Therefore, by Theorem 4.3.1, there exist I > 0 and a solution (xn, yn) ε L χ Μ to the equation (/θ 1) (£„,£„) = 0 (i.e., yn = f(xn)) such that \\x0 + hnU! + h2nu2n - xn\\ + \\f(xo) + hnf'ixo)^ + h2nv2n - yn\\ < lh2n ||(/(a;o + Нпщ + h2nu2n) - f(x0) - /i„/'(a;o)"i)/^ - ^2n|| Hence u2n := (xn — xq — /ι„ϊΧι)//ι^ converges to u2, and for all π > 0, we know that xq + hnu-i + ^u2n belongs to L Π f~1(M) because Xq + hnU! + h2nUn = Xn and f(x0 + НпЩ + h2nun) = yn. To prove the converse inclusion, fix u2 £ Т^.^^Лх^.щ). Then for every sequence hn —> 0+, there exists a sequence u2n —> u2 such that x0 + hnui + h2nu2n £ L7 f(x0 + hnU! + h2nu2n) ε Μ Thus u2 ε Τ^'^ο,ϊΧι) and since f(xQ + Ки-^ + h2nu2n) f{xo) + hnf'(x0)ui + ^hlf"{x0)(uuui) + Η2ηΓ(χ0)η2η + ε{Ηη)Ηΐ where e(hn) converges to zero with hn, we deduce that /'Wh2 + J/"W(4i,iii) e Т^2)(/(ж0),/'(а;о)щ) П
4.7- Higher Order Tangent Sets 177 Example: Second order contingent sets to subsets defined by equality and inequality constraints Consider a closed subset L of a Banach space X and two twice continuously differentiable maps 9 ··= (51, · · · ,9p) ■· X ^ RP & h := (hi, ...,hq):X~B? defined on an open neighborhood of L. Let К be the subset of L defined by the constraints Κ := {χ £ L I д{(х) > 0, ί = 1,... ,p &/y(a;) = 0, J = l,...,g} and let /(a;) := {»=ll...,p|5i(a;) = 0} denote the subset of active constraints. We posit the following transversality condition at a given χ ε Κ: ' i) ti(x)CL(x) = R9 ii) 3 vq £ Cl(x) such that h'(x)vo = 0 & Vie I(x), < g[(x), vq > > 0 By Proposition 4.3.7, υ,χ belongs to the contingent cone to if at a; if and only if u\ belongs to the contingent cone to L at χ and satisfies the constraints VieJ(s), <д[{х),щ >> 0 & V j = 1,...,ςτ, /i^(a;)u1=0 Proposition 4.7.5 -Lei us sei 7<2>(a;, ux) ;= {i £ /(a;) | < s<(z), ux > = 0} Then U2 belongs to the second order contingent set to К at (x, ui) if and only if v,2 belongs to the second order contingent set to L at (χ, ΐίχ) and satisfies the constraints V i<=№(x,ui), <9'г{х),и2>+\д'!(х)(иищ) > О and Vj = l,...,5, ti-(x)u2 + \ti-{x){ui,ui) = $ Proof — Define Υ = Β? χ R», / = (5, /ι), M = R^xR« The proof follows by exactly the same arguments as the ones of Proposition 4.3.7, after observing that the previous example yields that Μ is twice derivable and using Theorem 4.7.4. D
Chapter 5 Derivatives of Set-Valued Maps Introduction We need to differentiate set-valued maps as much as we need to differentiate single-valued maps, for extending the Inverse Function Theorem to set-valued maps for instance, and for many other reasons. How should we go about it? The idea is very simple, and goes back to the prehistory of the differential calculus, when Pierre de Fermat introduced in the first half of the seventeenth century the concept of a tangent to the graph of a function: the tangent space to the graph of a function f at a point (x, y) of its graph is the line of slope f'(x), i.e., the graph of the linear function и н-> f'(x)u. It is possible to implement this idea for any set-valued map F since we have introduced concepts of tangency for any subset of a normed space. For instance, we can regard the contingent cone to the graph of the set-valued map F at some point (x, y) of its graph as the graph of the associated "contingent derivative" of F at this point (z,y). Since the contingent cone is at least... a cone, the contingent derivative is at least positively homogeneous (or a process). This is what remains of the familiar, but luxurious, requirement of linearity. However, when the contingent cone happens to be convex (this 179
180 5- Derivatives of Set-Valued Maps is the case when the graph is sleek), the contingent derivative is a closed convex process, i.e., a set-valued analogue of a continuous linear operator.- The graphical derivatives associated to the Clarke tangent cone, which we call here the circatangent derivatives, are also closed convex processes. They are.presented in the second section together with the adjacent derivatives, the graphs of which are the adjacent (or intermediate) tangent cones. Hence the graphical derivatives which are attached to each such tangent cone correspond to different regularity- requirements. It is also possible to define them as adequate limits of differential quotients. The limits, involving "limsupinf's", that would not have been proposed by any sensible or even sober person in a right frame of mind had they not been derived from the "graphical approach." They are still pointwise limits, which puts them in a different class than the distributional derivatives1 introduced by Laurent Schwartz in the early fifties. We shall see in Chapter 7 that they are also adequate "graphical limits" of differential quotients. Chain rules are provided in the third section. With this differential calculus at hand, we are able to approximate solutions to an inclusion F(x) э у by solutions to approximate inclusions Fn(xn) Э у η by adapting Lax's principle: Convergence of the data yn to y, consistency of Fn to F and stability of the Fn's imply the convergence of some solutions xn's to the solution x. Consistency means that (x, y) belongs to the lower limit of the graphs of the maps Fn. Stability means here that the norms of the 'which were designed to make sense out of Dirac's operational calculus and solve partial differential equations. In order to keep the linearity of the differential operators, the limit of differential quotients is taken in spaces of distributions endowed with topologies much weaker than the pointwise convergence topology, used ever since Cauchy's rigorous definition.
5.1. Contingent Derivatives 181 inverses DFn(xn,yn)~l are bounded when the pairs (хп,Уп) range over a neighborhood of (x,y). We shall deduce that, if the right- hand side уn converges to y, there exist approximate solutions xn € -P1-1 (Уп) which converge to x, and we shall provide an error estimate. When the maps Fn = F are constant, this principle boils down to the statement of the Inverse-Function Theorem. We consider also an application to Qualitative Analysis, a field started in mathematical economics under the name of "comparative statics" and taken over by computer scientists as a domain of Artificial Intelligence. For instance, if the vector-spaces X and Υ are finite dimensional, we are just interested in the possibility of solving an equation or an inclusion F(x) Э у for prescribed sign vectors. We provide a dual criterion allowing us to answer this question. Finally, by using high-order contingent sets to the graph of a set-valued map F, we define in Section 6 high-order derivatives of a set-valued map. We derive in Chapter 10 variational inclusions by linearizing a differential inclusion, the solution map of which contains the contingent derivative of the solution map of the original differential inclusion. 5.1 Contingent Derivatives By coming back to the original point of view proposed by Fermat, we are able to define geometrically derivatives of set-valued maps from the choice of tangent cones to the graphs, even though they yield very strange limits of differential quotients. Definition 5.1.1 Let Χ, Υ be normed spaces and F : Χ ~> Υ be a set-valued map. The contingent derivative DF(x,y) of F at (x,y) £Graph(F) is the set-valued map from XtoY defined by Giaph(DF(x,y)) := TGTaph{F)(x,y) When F := f is single-valued, we set Df(x) := Df(x,f(x)).
182 5 - Derivatives of Set-Valued Maps We shall say that F is sleek at (x, y) G Graph(-F) if the map Graph(F) Э (x',y') ~~> Giwh(DF(x',y')) is lower semicontinuous at (x,y) (i.e., if the graph of F is sleek at (z,y))· It is said to be derivable at (x, y) if Graph(-F) is derivable at that point. The set-valued map F is sleek (respectively derivable) if it is sleek (respectively derivable) at every point of its graph. Finally, we shall say that F is pseudo-convex at (x, y) G Graph(F) if and only if Ух' £ Dom(-F), F(x') С DF(x,y)(x' - x) +y (i.e., if its graph is pseudo-convex at (x,y)). Naturally, the contingent derivative is a closed convex process whenever F is sleek at (x,y). We can easily compute the derivative of the inverse of a set-valued map F (or even of a non injective single-valued map): The contingent derivative of the inverse of a set-valued map F is the inverse of the contingent derivative: D(F~l)(y,x) = DF(x,y)-1 The restriction F := f\x of a single-valued map / to a subset К С X provides an example of a set-valued map defined by ί\κ(χ) \ f(x) ifz e К :~ 10 ifs 0 К for which we obtain the following formula: If / is Prechet differentiable around a point χ G K, then the contingent derivative of the restriction of f to К is the restriction of the derivative to the contingent cone: D(f\K)(x) = D(f\K)(x,f(x)) = f(x)\TK{x) Actually, this follows from the useful
5.1. Contingent Derivatives Figure 5.1: Graph of the Contingent Derivative
184 5- Derivatives of Set-Valued Maps Proposition 5.1.2 Let X and Υ be normed spaces, f be a single- valued map from an open subset Ω С X to Υ, Μ : Χ ~> Υ be a set-valued map. Define the set-valued map F : X ~> У by: V χ € X, F(x) := f(x) - M(x) If f is Frechet differentiable at χ € Ω Π Dom(M), then for every У € F(x) DF(x,y)(u) = f'(x)u-DM(x,f(x)-y)(u) Proof— Let ν belong to DF(x,y)(u). Then there exist hn > 0 converging to 0 and sequences un and vn converging to и and ν respectively such that for every η У + hnvn € f(x + hnun) - M(x + hnun) Since f(x + hnun) = f(x) + hn(f'(x)u + e(hn)) where e(hn) converges to 0 with hn we get f(x) - у + hn(f (x)u - vn + e(hn)) € M(x + hnun) and thus f'{x)u - ν € DM(x, f(x) - y){u). Conversely, assume that f'(x)u—v belongs to DM(x, f(x)—y)(u). Hence, there exist a sequence hn > 0 converging to 0 and sequences un and wn converging to и and f'(x)u — ν such that f(x) - у + hnwn € M(x + hnun) Then for some e(hn) -* 0, the sequence vn := f'(x)u + e(hn) — wn converges to ν and satisfies y + hnvn € f(x + hnun) - M(x + hnun), □ When we add "state constraints" L, this result becomes: Proposition 5.1.3 Let Χ, Υ be normed spaces, f : Ω (-► Υ be a single-valued map, where Ω С X is open, Μ : Χ ~> Υ be a set-valued map and L С X. Let F : X ~> У be the set-valued map defined by: f(x) — M(x) when χ € L 0 when χ £ L F(x) :=
5.1. Contingent Derivatives 185 If f is Freehet differentiable at χ € Ω Π Oom(F), then for every У € F(x), ПР(т „ν,Λ г [ fl{-x)u ~ DM(X> f№ ~ M") when u € Tl^ V*ix>yKU)C\ 0 when u£TL(x) Equality holds true when we assume that either L or Μ is derivable at χ and Μ is Lipschitz at x. In particular if Μ is constant, then \fu € TL(x), DF(x,y)(u) = f'(x)u-TM(f(x)-y) Remark — The above two Propositions remain true if / is locally Lipschitz and Gateaux differentiable. Π Proof— Let ν belong to DF(x, y)(u). Then there exist hn > 0 converging to 0 and sequences un and vn converging to и and ν respectively such that for all η χ + hnun € L & у + hnvn € f(x + hnun) - M(x + hnun) This implies that и belongs to Tl(x)- Since we can write f(x + hnun) = f(x)+hn(f'(x)u + e(hn)) where e(hn) converges to 0 with hn, we get f(x) - у + hn(f'(x)u -vn+ e(hn)) € M(x + hnun) and thus f'(x)u — ν does belong to DM(x, f(x) — y)(u). Conversely, assume for instance that Μ is derivable, that и belongs to Tl(x) and that f'(x)u — ν belongs to DM(x, f(x) — y)(u). Hence, there exist a sequence hn > 0 converging to 0 and sequences un, й~п and wn converging to и, и and f'{x)u — ν respectively, such that χ + hnun belongs to L and f(x) — у + hnwn to M(x + hnu~n). Since Μ is Lipschitz at χ there exists a sequence wn converging to f(x)u — ν satisfying -V η > 1, f(x) -y + hnwn € M(x + hnun)
186 5- Denvatives of Set-Valued Maps Then for some e(hn) —► 0, the sequence vn := f'{x)u + e(hn) — wn converges to ν and satisfies У + hnvn € f(x + hnun) - M(x + hnun) □ Our first task is to characterize these contingent derivatives by- adequate limits of differential quotients. We derive that the contingent derivative in the direction и is the upper limit of the differential quotients Fix + hu1) — у DF(x,y)(<u) = Limsuph_>0+iJi/_>Ji — from Proposition 5.1.4 Let Χ, Υ be normed spaces, F : Χ ~> Υ be a set-valued map and let (x, y) € Graph(-F). Then / pfx -\-hu') — у \ ν € DF(x,y)(u) <=> liminffc^o+.u'-fu^ ( «, ^ —J = 0 If χ € Int(Dom(F)) and F is Lipschitz around x, then υ € DF(x,y)(u) <^> limiafrffv, (x+ u'y) = Q /1-Ю+ \ h j If moreover the dimension of Υ is finite and I denotes the Lipschitz constant of F at x, then Oom(DF(x,y)) = X and DF(x,y) is I - Lipschitz Proof — The first two statements being obvious, let us check the last one. Let и belong to X and I denote the Lipschitz constant of F on a neighborhood of x. Then, for all h > 0 small enough and У € F(x), y>€ F(x) С F(x + hu) + lh\\u\\B Hence there exists yh € F(x + hu) such that vn := (уь, — y)/h belongs to Z||u||J5, which is compact. Therefore the sequence Vh has a cluster point v, which belongs to DF(x,y)(u). Fix any Щ, u2 € X and v\ € DF(x,y)(ui). To end the proof we have to show that there exists v2 € DF(x,y)(u2) such that
5.2. Contingent Derivatives 187 \\vi — щ\\ < ^ ||iii — иг||. Consider hn —> 0+ and v\n —> v\ satisfying у + hnvin € F(x + hnui) and let zn € -F(x + /1^2) be such that for all large n, II^n - 2/ - hnvin\\ < lhn \\m - U21| (they exist by the Lipschitz continuity of F at x.) Taking a subsequence and keeping the same notations, we may assume that the bounded sequence (zn — y)/hn converge to some г/2 € DF(x,y)(u2). Clearly г/2 satisfies the required estimates. Π Remark — Domain of the Derivative It is quite useful to relate the tangent cones to the domain of a set-valued map to the domain of its derivative. We always have Oom(DF(x,y)) с TOom{F)(x) If F is pseudo-convex at (x,y) £ Graph(F), (and in particular, if F is convex), then equality Oom(DF(x,y)) = TOom{F)(x) holds true, because Dom(F) = Tr^Graph(F) where ττ χ is the projection from Χ χ Υ onto X. We thus deduce from Proposition 4.2.9 that ' Oom(DF(x,y)) = irxGTaph(DF(x,y)) < = nxTGTa:pbiF)(x,y) = ?;x(Graph(ir))(a;) = TOom{F)(x) Π Remark — Kernel of the Contingent Derivative The kernel of the contingent derivative is related to the contingent cone to the inverse image: Тр-Цу)(х) С kerDF(x,y) := DF(x,y)-\0) Equality holds true in this formula whenever F~l is pseudo-Lipschitz around (2/, ζ).
188 5- Derivatives of Set-Valued Maps To prove the converse inclusion, let и belong to the kernel of DF(x, y). Then there exist sequences hn > 0 converging to 0, un and vn converging respectively to и and 0 satisfying Vn, y + hnVn e F(x + hnun) Since F~l is pseudo-Lipschitz around (y,x), there exist Ζ > 0 and an element x\ 6 F~l(y) such that \\Xn~ (x + hnun)\\ < l\\y - (y + hnvn)\\ = lhn\\vn\\ Hence, by setting u\ := (x\ — x)/hnt we see that x + hnuln = x\ £ F~\y) and that u\ converges to и because ||u^ - un\\ < l\\vn\\ and because vn converges to 0. Therefore we have proved that и belongs to the contingent cone to F~1(y) at x. □ Remark — Lower Semicontinuously Differentiable Maps Definition 5.1.5 (Lower Semicontinuous Differentiability) LetX, Υ be normed spaces and F : Χ ~» Υ be a set-valued map. We say that F is lower semicontinuously differentiable at (x,y) &Graph(F) if the set-valued map (x,y,u) ε Graph(F) xX-, DF(x,y)(u) is lower semicontinuous. Observe that it implies that F is sleek at (xo,yo)· The converse needs further assumptions. We derive for instance from the "Cross-Convergence" Theorem 2.3.2 the following criterion: Proposition 5.1.6 Assume that X and Υ are Banach spaces and that F is sleek on some neighborhood U of(xo,yo) & Graph(F). If the boundedness property Vn el, sup inf \\v\\ < -t-oo (*,») e wnGraph(F) »е^(*.»)(«) holds true, then the set-valued map (x,y,u) ε Graph(F) χ Χ ^ DF(x,y)(u) is lower semicontinuous on (Uf) Graph(F)) χ Χ.
5.2. Adjacent and Circatangent Derivatives 189 5.2 Adjacent and Circatangent Derivatives 5.2.1 Definitions and Elementary Properties Naturally, we can also associate with any other concept of tangent cone a concept of derivative. Since we were led to introduce Clarke and adjacent (or intermediate) tangent cones, we can introduce two more graphical derivatives: Definition 5.2.1 Let Χ, Υ be normed spaces, F : Χ ~> Υ be a set-valued map and у € F(x). 1. — the adjacent derivative D^F(x, y) is the set-valued map from XtoY defined by Graph (D»F(x,y)) := lfcraph(F)(*,y) 2. — the circatangent derivative CF(x, y) is the set-valued map from XtoY defined by Graph(CF(x,y)) := CGTaph{F)(x,y) When F := f is single-valued, we set Dbf(x) := Dbf(xJ(x)), Cf(x) := Cf(x,f(x)) We see at once that these graphical derivatives are closed processes, that Vu, CF(x,y)(u) С DbF(x,y)(u) С DF(x,y)(u) When / is single-valued, we observe that the circatangent and adjacent derivatives Cf(x)(u) and D* f(x)(u) are either empty or contain only one element. In particular, i) Dbf{x){u) = f'(x)u if / is Frechet differentiable at χ ii) Cf(x)(u) = f{x)u if / is continuously differentiable at χ Naturally, the circatangent derivative is always a closed convex process.
190 5 - Derivatives of Set-Valued Maps Remark — Since CF(x, y) is a closed convex process, it is equal to its bitranspose. The transpose CF(x,y)*:Y*"~> X*' has been called the codifferential of F at (x, y). It is involved whenever dual conditions are used. Instead of founding a "direct" differential calculus on tangent cones, as we did, it is possible to start with any concept of normal cone ^Qraph(F) (x> У) *° ^е S^Ph of F at a point (x,y). We can associate with it the corresponding "codifferential" dF(x, y)* defined by ρ € dF(x,y)*(q) if and only if (p, -q) € NGraph(F) (x, y) When such a normal cone is not convex, it can no longer be defined as polar of a tangent cone, so that such codifferential,, in general, is not the transpose of a graphical derivative. □ As for the contingent derivative, a graphical derivative of the inverse of a set-valued map F is the inverse of the derivative: ' i) D\F~l){y,x) = D^F^y)-1 < ii) C{F~l){y,x) = CF{x,y)-1 If К is a subset of X and / : X н-> Υ is a single-valued continuously differentiable around a point χ € К map, then the derivative of the restriction is the restriction of the derivative to the corresponding tangent cone: ■ i) D\f\K)(x) := Db(f\K)(x,f(x)) = Г(х)\тьк{х) < . H) C(f\K)(x) := C(f\K)(x,f(x)) = f(x)\cK(x) Actually, these formulas follow from: Proposition 5.2.2 Consider normed spaces Χ, Υ, a single-valued map f from an open subset Ω С X to Υ and a set-valued map Μ : X ~> Υ. Define the set-valued map F : Χ ~> Υ by: V χ € X, F(x) := f(x) - M(x)
5.2. Adjacent and Circatangent Derivatives 191 If f is Frechet differentiable at χ € Ω Π Dom(M); then for every У € F(x) DbF(x,y)(u) = f'(x)u-DbM(x,f(x)-y)(u) and if f is continuously differentiable at x, then CF(x,y)(u) = f'(x)u-CM(xJ(x)-y)(u) The proof is analogous to the one of Proposition 5.1.2 and is left as an exercise. Proposition 5.2.3 Let X andY be normed spaces, fa single-valued map from an open subset Ω С X to Υ, Μ : Χ ~> Υ be a set-valued map and L С X. Consider the set-valued map F : Χ ~> Υ defined by: , χ _ J f(x) — M(x) when χ € L [X) :~ { 0 when x<£L If f is Frechet differentiable at χ e Ω Π Dom(F) and Μ is Lipschitz at x, then for every у € F(x) the adjacent derivative of F at (x, y) is equal to n*F(r ,Λί,Λ-ί Г(ф-0ЬМ(хЛх)-у)(и) when и€ТЦх) DF{x,y){u)-^ 0 ^η uiTbdx) The same formula holds true for a map f continuously differentible at x, the circatangent derivatives of F and Μ and the Clarke tangent cone to L. The proof is the same as the one of Proposition 5.1.3 and these statements remain true when / is Lipschitz at χ and Gateaux differentiable. □ 5.2.2 Limits of Differential Quotients In order to characterize adjacent and circatangent derivatives in terms of limits of differential quotients, we need the concept of "lim sup inf" (or Γ-convergence) of functions of two variables introduced independently by several authors (de Giorgi, Rockafellar, ...):
192 5- Denvatives of Set- Valued Maps- Definition 5.2.4 (Lim sup inf) LetL and Μ be two metric spaces and φ : L χ Μ l·-* Л be a function. We set lim sup^/^a; ϊτιίυΐ->υφ(χ', у') := sup inf sup inf ф(х', у') e>0 V>0 χ'βΒ{χ,η) y'eB(y,e) We write (x', y') -±p (x, y) to denote that (x1, y') converges to (x, y) _ while remaining in the graph of F. Hence, by translating the definition of the intermediate and the Clarke tangent cones, we obtain the following characterizations: Proposition 5.2.5 Let (x, y) belong to the graph of a set-balued map F : X ~> У from a normed space X to another Y. Then ν belongs to V*F(x,y)(u) if and only if < lim sup^0+ Mu^ud (v, Р^')~У) = 0 and ν belongs to CF(x,y)(u) if and only if < k lim sup/l_>0+i vtf)^F{Xiy) miu>->ud (υ, Εί^ψ}^ = 0 If F is Lipschitz around χ € Int(Dom(-F)); then the formulas become much simpler: υ belongs to D^F(x,y)(u) if and only if and ν belongs to CF(x,y)(u) if and only if Um/l_>0+[ (x>,y>)->F{x,y) d (v, n*'+hv)-y'^ = 0 Remark: Kernel of the Derivative As for the contingent derivative, the kernels of other derivatives are linked to the associated tangent cones to the inverse image: Тр-г(у)(х) С kerDbF(a;,y) := D*F(x,yrl(0)
5.2. Adjacent and Circatangent Derivatives 193 ала equality holds true in this formula whenever F""1 is pseudo-Lipschitz around (y,x). In this case, we also have: kerCF(x,y) с Ср-цу)(х) □ The proofs are left as an exercise. Proposition 5.2.6 Let us assume that the images of F are convex and that F is Lipschitz around x. Then for any (x, y) € Graph(-F) the images of the adjacent derivative D F(x, y) are convex and DbF(x,y)(0) = TF{x)(y) DbF{x,y){u) + D^F{x,ym = D*F(x,y)(u) Proof— Let vi and V2 belong to DF(x,y)(u). Then, for any sequence hn > 0 converging to 0, there exist sequences u\n and U2n converging to и and sequences v\n and V2n converging to v\ and V2 respectively such that V n, у + hnVin € F(x + hnUin) (i = 1,2) Since F is Lipschitz around x, there exists I > 0 such that for all η large enough, y + hnV2n € F(x + hnUin) + lhn\\u2n — Uin\\ so that we can find another sequence V3n converging to V2 such that у + hnV3n С F(x + hnu\n) Now, F(x + hnu\n) being convex, we deduce that for all λ € [0,1], у + hn{\vin + (1 - λ)υ3π) € F(x + hnum) Since Xvin + (1 — \)vsn converges to Xvi + (1 — λ)г/2, this element belongs to D]'F(x,y)(u). By Proposition 5.2.5, we see that ν € DbF(x,y)(0) if and only if d(v, (F(x) — y)/h) converges to 0, i.e., ν belongs to the adjacent tangent cone to F(x) at y. Since F(x) is convex, it coincides with the tangent cone.
194 5- Derivatives of Set-Valued Maps Since 0€DbF(x,y)(0) we obtain that Vuj DbF(x,y)(u) С DbF(x,y)(u) + DbF(x,y)(0) To prove the opposite inclusion fix ν € DbF(x,y)(u) & w € DbF(x,y)(0) Let hn —► 0+, vn —► υ be such that V n, у + /ιη^π € -F(:c + /inu) By convexity of -F(:e), there exist wn -* w such that for η large enough, у +y/h^wn € -F(a;). Then, by the Lipschitz continuity of F, for all large η and for some w'n, we have у + VK.w'n £ -Р(ж +/ιπω); ||и4-гоп|| < Z\AnlMI Thus (1 -y/K)(y + hnvn) + y/K(y +y/h^w'n) ' = у + hn(vn + w'n) -y/Khnvn = у + hn(v + w) + hne(hn) € F(x + hnu) where e{hn) converges to 0. Hence r a- л. ( , F(x + hnu)-y\ lim dist [v + w, — - = 0 n->DO V К J This ends the proof. D 5.2.3 Derivatives of monotone operators Let X be a Hubert space (identified with its dual.) We recall that a set-valued map F : X ~> X is monotone if and only if V(z,p), (y,q) € Graph(-F), <p-q, x-y>> 0 (see Chapter 3 for an introduction to monotone and maximal monotone maps.) We also recall that when F is monotone, its resolvent J := (1 + -F)-1 is single-valued and Lipschitz (with constant equal to 1) on its domain. Therefore, we can easily compute derivatives of F in terms of the derivatives of its resolvent.
5.2. Adjacent and Circatangent Derivatives 195 Proposition 5.2.7 Let X be α Hilbert space identified with its dual X*. We supply X* with the weak-* topology. If F : X ~> X* is monotone, then the contingent derivative DF(x,p) at every pair (x, p) € Graph(F) is semi positive-definite in the sense that V(u,r) € Gr&ph(DF(x,p)), <r,u>>0 Furthermore, the following statements are equivalent: ( a) r € DF(x,p)(u) [ b) и € DJ(x+p)(r + u) All the above statements remain true for the adjacent and circatangent derivatives. Proof — The first claim is obvious, since < г, и > is the limit of a sequence < rn, un >, where ρ + hnrn € F(x + hnun), (because un converges to и strongly and rn converges to r weakly) and since h\ < rn, un > = < χ + hnun - x,p + hnrn - ρ > > 0 For proving the second statement, we observe that ρ belongs to F(x) if and only if χ = J (χ + ρ), because χ + ρ € (1 + F)(x) and because J is the inverse of (1 + F). We also deduce from the latter that the two following statements are equivalent: ii) r + u € D(l + F)(x,x+p)(u) ii) и € DJ{x + p,x)(r + u) = DJ{x+p){r + u) On the other hand, Proposition 5.1.3 implies that D(l + F)(x,x+p)(u) = u + DF(x,p)(u) We have obtained the formula we were looking for. Π Since the cone-valued map Νκ associating with any χ € К the normal cone Νκ(χ)~^Ό a closed convex subset is maximal monotone (because the normal cone is the subdifferential of the indicator of K), and since its resolvent is the projector of best approximation onto K, we deduce the following corollary:
196 5 - Derivatives of Set-Valued Maps Corollary 5.2.8 Let К be α closed convex subset of α Hilbert space and ρ belong to the normal cone Nk(x) to К at some χ € К. Denote by Пк the best approximation projector onto K. Then, -q € DNK(x,p)(u) <=> и € DTlK(x+p)(u + q) 5.3 Chain Rules We derive from the calculus of tangent cones the associated calculus of derivatives of set-valued maps. We begin naturally by the chain rule for computing the composition product of a set-valued map G : Χ ~> Υ and a set-valued map Η : Υ ~> Ζ. We shall need the following result: Proposition 5.3.1 Let Χ, Υ be normed spaces, F : Χ ~> Υ be a set-valued map and К be a subset of X. Assume that F is Lipschitz around some χ € К. Then, for any у € F(x), we have DbF(x,y)(TK(x)) С TF[K){y) As a consequence, we deduce that when Μ is a subset of Υ and у € Μ, then ТРщм){х) С DbF(x,y)+1(TM(y)) (5.1) Proof — Take и in TK(x) and ν € DbF(x,y)(u). Then there exist sequences hn > 0 converging to 0, u\n and U2n converging to и and vn converging to ν such that χ + hnuin € К & y + hnvn € F(x + hnU2n) Since F is Lipschitz around χ with a Lipschitz constant I, we deduce that y + hnvn € F(x + hnuin) + lhn\\uin - U2n\\ so that there exists another sequence υ* converging to ν such that y + hnv*n € F(x + hnuin) С F{K) This implies that ν belongs to the contingent cone to F{K) at y.
5.3. Chain Rules 197 Consider now К := F+1(M). Since F(F+1(M)) is contained in M, we deduce that ДьЯя>2/)(?>+1(м)(:е)) С TF(F+4M))(y) С Гм(2/) from which formula (5.1) ensues. Π Remark — Naturally, we can show in the same way that for a Lipschitz map F the formula DF(x,y)TbK(x) С TF{K){y) is also true whenever у € F(x). Π We begin by the following simple result: Theorem 5.3.2 Let us consider normed spaces Χ, Υ, Ζ, a set- valued map G : X ~> У and a set-valued map Ή. :Υ ^* Ζ. 1. — Let us assume that Η is Lipschitz around y, where у € G(x). Then, for any ζ € H{y), we have ( DbH(y,z)oDG(x,y) С D(HoG)(x,z) \ DH(y,z)oD]>G(x,y) С D(HoG)(x,z) 2. — If G := g is single-valued and Frechet differentiable at x, we obtain Vz€H(g(x)), D(Hg)(x,z)(u) С DH(g(x),z)(g\x)u) and the equality holds true when Η is Lipschitz around g(x). Proof — We apply Proposition 5.3.1 to equality Graph(tf о G) = (1 χ Я) (Graph(G)) for proving the first statement. The second one follows from Graph(tfoff) = (g χ l)-1 (Graph(tf)) ТгЧк)(х) С /Ή"1 СВД(х))) □
198" 5- Derivatives of Set-Valued Maps Remark — Let us assume that G is Lipschitz around x. Then, for all у € G(x) and ζ € (HU\G)(x), we have D(HnG)(x,z) С DH(y,z)nDbG(x,y) because we can apply Proposition 5.3.1 to Μ := Graph(fT) and because Graph(tfDG) = (G χ 1)+1 (Graph(tf)) We state now a more powerful result. Theorem 5.3.3 Let Χ, Ζ be Banach spaces and Υ be a finite dimensional vector-space. Consider set-valued maps G : Χ ~> Υ and Η : Υ ~> Ζ. Fix x0 € Dom(G), y0 € G(x0) П Oom(H) and z0eH(y0). If G and Η are closed and satisfy the transversality condition lm(CG(x0,y0))-Oom(CH(y0,z0)) = Υ then ' i) DbH(y0,z0)oDG(x0,y0) С D(HoG)(x0,z0) < ii) DbH(yo,z0)oD]'G(xo,yo) С Db(HoG)(xQ,zQ) ^ Hi) CH(y0,z0)oCG(x0)y0) С C(HoG)(x0,z0) Proof — If we denote by ω the continuous linear operator from XxYxYxZ to Υ associating to (x, yi, y2, ζ) the element y\ — уг and by πχχζ the canonical projection from Χ χ Υ χ Υ χ Ζ onto Χ χ Ζ. Observe that Graph(tf ο G) = πχχΖ ((Graph(G) χ Graph(tf)) η иГ^О)) We apply Theorem 4.3.3 with L = Graph(G) x Graph(#), / = ω and Μ = {0}.
5.3. Chain Rules 199 Consequently, we deduce that, for instance, (Graph(L>bG(z0,yo)) x Graph(L>btf (yo,zQ)j) П ω~ι(0) = feraphtGJ^'IO) X ГСгарЬ(я)(2/o^o)) П^(0) (5.2) = ^raphiGjxGraphiff)^»'У°'У°'z°) nω_1(°) = Г(СгарЬ(С)хСгарЬ(я))пш-1(о)^0'Уо'Уо'z^ By applying the projection ж χ χΖ to both sides of these equalities, we prove that ' Graph (ubtf (yo, z0) о D*G(x0, yQ)) = πχχζ ((СгарЬ(£>ьС(г0, Ы) * Graph(L>btf (j/q, z0))) Π uT^O)) = πχχζ (^r(Graph(G)xGraph(ff))no;-i(o)^0'2/0'2/0'2;o)y C r7rxx2r((Graph(G)xGraph(H))no;-i(o))^0' z^> ..= rGraph(ffoG)(S0'z°) = Graph(ub(FoG)(o;o^o)) The proof of the other statements is similar and left as an exercise. D Remark — We recall that the norm of a process (positively homogeneous set-valued map) A:Y^>X is defined by ||A|| := sup inf \\x\\ IMI<i хЕА(у) If Υ is any Banach space, we have to assume that there exist constants
200 5 — Derivatives of Set-Valued Maps с > 0, а £ [0,1[ ала η > О such that ' V(s,yi) £ Graph(G)nB((:co,2/o),77), V(ifc,z) £ Graph(H)nB((y0,*)),q), < By с Jm(DbG(x,yi))ncBY- Oom(DH(y2,z)) + αΒγ and . \\DbG(x,yi)\\<ck \\D{H^){z,y2)\\<c to prove the above formulas. Indeed, the above assumption implies the second transversality condition of Theorem 4.3.3. For any υ £ Υ, we can find v\ £ lm(DbG(x, yi)) and v2 £ Dom(DH(y2,z)) such that ν = vi — v2 + e, \\v\\\ < с||г?||, ||e|| < α\\υ\\ Hence ||г/2|| < (с + 1 + а)||г/|| and there exist u £ i^Gfa yO-^vO & ω £ DH(y2,z)(v2) such that ||?x|| < c||t/i|| and ||ω|| < ο||υ2||. Therefore, ν = uj(u,vi,v2,w) + e where (u,vi,v2,w) belongs to the contingent cone to the product of the graphs of G and Η and e £ «By. Consequently, we can derive properties (5.2) and conclude the proof. □ We obtain converse inclusions under more severe assumptions: Proposition 5.3.4 Let Χ, Υ, Ζ be normed spaces. Consider set- valued maps G : X ~> У', Η :Υ ~» Ζ and assume that G is pseudo- convex at (жо)Уо) € Graph(G) and Η is pseudo-convex at (yo,zo) 6 Graph(#). Then D(HoG)(x0,z0) с DH(yQ,zQ)oDG(x0,y0) Proof— Let ω and πχχζ be the operators denned in the proof of Theorem 5.3.3. When the graph of G is pseudo-convex at (x, y) and the graph of Η is pseudo-convex at (y, z), we derive the above property from Proposition 4.2.9 applied in the following situation:
5.3. Chain Rules 201 ' Graph(.D(tf о G)(x0, zq)) = rGraph(HoG)(s0>2b) = cl [πχχΖ (^r(Graph(G)xGraph(H))no;-i(o)(:cO'2/o>2/o>2;o)JJ С cl {πΧχΖ ((Graph(L>G(z0)2/o)) x Graph(UF(y0,^))) Πω"1^))) = d(Graph(^(yo,*o)°£G(xo,yo))) Remark — We can deduce the conclusions of this proposition from any other criterion implying that the contingent cones to the images by πχχζ axe the closures of the images of the contingent cones. Π Remark — Let X be a normed space and K, L be subsets of X. Recall that the Bouligand paratingent cone Ρχ(χ) to К relative to L at χ £ L Π Κ is defined by P%(x) := {v\ liminf dK(x' + hv)/h = 0} and that we set Рк(х) := Рк(.х)· •We can obtain upper estimates of the composition product by using the pamtingent derivatives defined by saying that the graph of the paratingent derivative PF(x, y) of F at (x, y) is the paratingent cone to the graph of F at (x,y): Graph(PF(a;)2/)) := PGraph(F) (x> v) Hence, for any и £ X, the paratingent derivative in the direction и is equal to Fix' + hu') — y' PF{x,y){u) = Limsuph_>0+iU,_>Ui(x,i2/,)_>j;,(Xi2/) We can also define the lop-sided paratingent derivatives PxF(x,y) and PyF(x, y) in the following way: ' i) P*F(x,y)(u) = LimSuph_t0+|M,_tMiy,_tF(x)/^yb^ < ^ ii) pyF(x,y)(u) = Umsuph^Q+iU^UiX^F_i(v)XF^'+hh^--y
202 5 — Derivatives of Set-Valued Maps Theorem 5.3.5 Let X,Y be normed spaces. Assume that G : Χ -^ Υ is Lipschitz around x. If Υ is a finite dimensional vector-space and G(x) is bounded, then D{HoG){x,z) С (J P*H(y,z)oP*G(x,y) yeG(a;) Proof— Let w belong to D(H о G)(x,z)(u): there exist sequences hn > 0, un and wn converging to 0, и and w respectively such that V n, z + hnwn £ H(yn) where yn ε G(x + hnun) Since G is Lipschitz around x, there exist I > 0 ала elements y° ε G(x) such that, for η laxge enough, У П У ΤΙ .· η II II Til II νη := —;: satisfies \\υη\\ < l\\un\\ hn Furthermore, G{x) being relatively compact, a subsequence (again denoted) y° converges to some y. We can also extract a subsequence (again denoted) vn which converges to some v, since this sequence is bounded and the dimension of Υ is finite. Prom the relations yn + hnvn ε G(x + hnun) к z + hnWn ε #(y° + hnvn) we infer that w belongs to PzH(y, z)(v) and υ to PxG(x,y)(u). Π It is not difficult to check that for all у ε G(x), и ε Dom(CG(a;,y)), P(HDG)(x,z)(u) С (PH(y,z)nCG(x,y))(u) Indeed, let и ε Oom(CG(x, у)) and w belong to P(HoG)(x,z)(u). Hence there exist a sequence hn > 0 converging to 0 and sequences of elements (.Χη,Ζη) ε Graph(#DG), un and wn converging to (x, z), и and w respectively such that Vn>0, zn + hnwn ε f] H(y) у £ G(xn+hnun) The set-valued map G being Lipschitz, there exists a sequence of elements yn ε G{xn) converging to y. By definition of the square product, we know that zn ε H{yn) (because zn ε (HoG)(xn).) Take now any υ in CG(x,y)(u). Since G is Lipschitz around x, there exists a sequence of elements vn converging to ν such that V η > 0, уn + hnvn ε G(xn + hnun)
5.4. Inverse Set-Valued Map Theorem 203 Therefore, V η > 0, zn+ hnwn £ H{yn + hnvn) so that we infer w ε PH(y,z)(v) Since this is true for every element υ of CG(x, y)(u), we deduce that w ε Π PH(y,z){v) = (PH(2/^)nCG(^l2/))(u) Π veCG(x,y)(u) 5.4 Inverse Set-Valued Map Theorem 5.4.1 Stability and Approximation of Inclusions Let Χ, Υ be Banach spaces and F : X ~> Υ and Fn : Χ ~> Υ be set-valued maps with closed graphs. We are ready to extend Lax's Principle to the approximation of a solution to inclusion find xq € X such that F(xq) Э yo by solutions to the approximate inclusions find xn € X such that Fn(xn) Э yn Definition 5.4.1 We shall say that the set-valued maps Fn are consistent at (xo,yo) € Graph(F) if there exist sequences of (xon,yon) € Graph(Fn) such that XQn converges to xq and y$n converges to y$. We shall say that a family of set-valued maps Fn is stable around (x, y) if there exist constants с > 0, α£ [0,1[ and η > 0 such that * {xn> Уп ) e Gr&ph(Fn)nB((x0,y0),v), Mv e Y, < 3un € X, 3wn <E Υ with ν £ DFn(xn,yn)(un)+ wn and \\un\\ < c\\v\\ & ||гоп|| < a\\v\\
204 5 - Derivatives of Set-Valued Maps Theorem 5.4.2 Let us consider Banach spaces X and Υ, a sequence of closed set-valued maps Fn : X ~» У, yo &Y and a solution xq to the inclusion F(xq) Э vq. Assume that the set-valued maps Fn are consistent at (xq, vq) and that the family of Fn 's is stable. Then there exists a constant I such that, for any sequence of elements (xonfyon) € Graph(i1Tl) converging to (xo,yo) and any sequence yn converging to yo, we have for η large enough d\xon,F~1(yn)j < 1\\уоп-Уп\\ As a consequence, we obtain the following error estimate: for any sequence yn converging to yo, ^(^Oi-FirHyn)) < I (d((xo,yo), Graph(.Fn)) + ||yo-yn||) . so that xq can be approximated by solutions xn G F~l(yn). Proof — We apply Theorem 3.4.5 with X replaced by X x Y, Kn by СгарЬ(^п), fn by the projection ΐΐγ from Χ χ Υ onto Υ. We have to prove that the stability assumption implies transversality assumption (3.9) of Theorem 3.4.5, i.e., that for all ν € Υ, there exist (un, vn) in the contingent cone Tqj.^^p ,(ιπ, yn) and wn G Υ satisfying ν = vn + wn, max(\\un\\,\\vn\\) < c\\v\\, \\wn\\ < α\\υ\\ This information is provided by our stability assumption since the contingent cone to the graph is the graph of the contingent derivative and the norm of vn = ν — wn is smaller than or equal to (1 + α)||υ||· D When the set-valued maps Fn are all equal to F, we obtain: Theorem 5.4.3 (Inverse Set-Valued Map Theorem) Consider α closed set-valued map F : X ~> У', an element (xq, yo) of its graph and let us assume that there exist constants с > 0, a G [0,1[ and
5.4. Inverse Set-Valued Map Theorem 205 η > 0 such that ' V(s,y) € Graph(F) П ^((zo.yo), ту), V υ e У, 13u € -X", Зги € ysuch that υ € DF(x,y)(u) + w and ||u|| < с||г>|| & HI < a||v|| Then г/о belongs to the interior of the image of F and F'"1 is pseudo- Lipschitz around (yo,xo). We can extend Theorem 3.4.10 to the case of set-valued maps by introducing an adequate definition of uniformly sleek maps, which are the set-valued analogues of the continuously differentiable maps. Definition 5.4.4 (Strongly Sleek Maps) We shall say that F is uniformly sleek at a point (xq, 2/0) G Graph(F) if its graph is uniformly sleek at this point, i.e., if sup d((u,v),Grapb(DF(x,y))) (u,v)eGTeLph(DF(xQ,yQ))r)(BxB) converges to 0 when (x,y) converges to (xo,yo) in Graph(F). With this definition, we can state a natural set-valued version of Graves' Theorem Theorem 5.4.5 Let us consider Banach spaces X and Υ, a closed set-valued map F : X -^ У and an element (xo,yo) of its graph. Let us assume that DF(xq, г/о) is surjective 1. — If F is uniformly sleek at (xo,yo), then 2/0 belongs to the interior of the image of F and F'"1 is pseudo-Lipschitz around (Уо,хо)- 2. — // the dimension of Υ is finite, it is sufficient to assume that F is sleek at (xo,yo) to get the same conclusions. The proof follows from Theorem 3.4.10. The very same proof yields
206 5 — Derivatives of Set-Valued Maps Theorem. 5.4.6 Consider Banach spaces X and Y, a closed set- valued map F : Χ ~> Υ and an element (xo,yo) of its graph. Let us assume that the dimension of Υ is finite and Cir(xo)2/o) is surjective Then ?/o belongs to the interior of the image of F and FT1 is pseudo- Lipschitz around (yo,xo). 5.4.2 Localization of Inverse Images In this section we provide some estimates of the inverse image F'"1 (г/*). Theorem 5.4.7 Let F be a set-valued map from a normed space X to a normed space Υ and (x*,y*) belong to its graph. 1. — Assume that the dimension of X is finite. Then, for any closed cone Ρ satisfying kei(DF(x*,y*))nP = {0} there exists ε > 0 such that? F-1(y*)n(x* + e(PnS)) = {x*} In particular ifker(DF(x*,y*)) — {0}, then F-\y*)t\B{x\e) = {x*} 2. — Assume that F~1(y*) is convex^. Then F'^y*) С x* + kei(DF(x*,y*)) In particular ifkei(DF(x*,y*)) = {0}, then F~1(y*) = {x*}- 2If i1-1 is pseudo-Lipchitz at (j/, ж), it is possible to show that conversely, the property P"V)n(i'+£(Pnfl)) = {χ*} implies that кег(Л^(а;*, у*)) П Int(P) = 0. 3or even, pseudo-convex at x*.
5.4. Inverse Set-Valued Map Theorem 207 Proof — Assume that the first statement is false. Then there exists a sequence of elements xn G F_1(y*) converging to x* such that xn — χ* Φ 0 belong to P. Setting hn := \\xn — x*\\ which converges to 0+ and un := (xn — x*)/hn belonging to the unit sphere and to the cone P, we infer that a subsequence (again denoted) un converges to some element и £ Ρ οι norm 1 because the dimension of X is finite. On the other hand, inclusions y* + hn0 € F(xn) = F(x* + hnun) imply that 0 € DF(x*,y*)(u), i.e., that и belongs'to the kernel of DF(x*,y*). This is impossible. By taking Ρ = X, we obtain the local uniqueness of a solution x* to y* G F(x). The second statement follows readily from the definition, since we observe that the convexity of the inverse image implies that F'^y*) С χ* + 2>-1(у.)(а:*) С х*+kei(DF(x*,y*)) D 5.4.3 The Equilibrium Map Let us consider two finite dimensional spaces X and Λ and a set- valued map F:Ax X ^ X We denote by Ε : Λ ~»· X the equilibrium map defined by νλ€Λ, E(\) := {xeX\0€F(\,x)} which associates with any parameter λ e Λ the set of equilibria of F(\, ·). We can derive some information on this equilibrium map from the knowledge of its contingent derivative. Proposition 5.4.8 Let Λ and X be two finite dimensional vector- spaces and F : Α χ Χ ^ Χ be a closed set-valued map. Assume that the circatangent derivatives CF(\, x, 0) are surjective for every (X,x) belonging to the graph of the equilibrium map E. Then 1. — The contingent derivative of the equilibrium map is the equilibrium map of the contingent derivative: 0 € Ι?ί,(λΙχΙ0)(μ,υ) <^> ν e DE{\,x){u)
208 5 — Derivatives of Set-Valued Maps and 0 € CF(X,x,0)fa,v) =D- υ € CE(X,x)(p) 2. — // for any μ £ A, there exists an equilibrium υ of CF(X, χ, 0)(μ, ·), then the equilibrium map is pseudo-Lipschitz around (X,x). 3. — Set Q{X,x) := {v€X\0€DF(\,x,0)(0,v)} Then, for any equilibrium χ £ E(\) and any closed cone Ρ satisfying Ρ Π Q(X,x) = {0}, there exists ε > 0 such that E(\) П (x + ε(Ρ П B)) = {χ} and in particular, an equilibrium χ £ E(\) is locally unique whenever 0 is the only equilibrium of DF(\,x, 0)(0, ·). 4- If E(\) is convex, then E(X) С χ + Q(\,x). Proof — We observe at once that the graph of the equilibrium map Ε is the inverse image of the graph of F by the continuous linear operator π € £(Λ χ Χ, Λ χ Χ χ Χ) defined by V(A,i)eAxI, π(Χ,χ) := (λ,χ,Ο) Because the circatangent derivative CF(\,x,0) is surjective, the transversality condition holds true: 1т(тг)-ССт&рЪ{р)(тг(\,х)) = AxX-Graph(CF(\,x,0)) = AxXxX Therefore Corollary 4.3.4 yields Graph(DE(X,x)) = π"1 (GTaph(DF(\,x,0))) and Giaph(CE(\,x)) D π"1 (Giaph(CF(\,x,0))) which are the formulas we were looking for. The second statement follows from Theorem 5.4.6 applied to E"1. We then apply Theorem 5.4.7 to the inverse of Ε to derive the localization properties of E(\). Π
5.4. Inverse Set-Valued Map Theorem 209 5.4.4 Local Injectivity Definition 5.4.9 Let Χ, Υ be norrned spaces and F : Χ ~»· Υ be a set-valued map. We shall say that it is locally injective around x* if and only if there exists a neighborhood N(x*) such that Vn ^i2e N(x*), F(xi)nF(x2) = 0 It is said to be (globally) injective if we can take for neighborhood N(x*) the whole domain of F. Theorem 5.4.10 Let F be a set-valued map from a finite dimensional vector-space X to a norrned space Υ and (x*,y*) belong to its graph. Assume that there exists 7 > 0 such that F(x* + ηΒ) is relatively compact and that F has a closed graph. If for all у G F(x*) the kernels of the paratingent derivatives PF(x*,y) of F at (x*,y) are equal to {0}, then F is locally injective around x*. Proof — Assume that F is not locally injective. Then there exist sequences of elements х\пф 3>2n, converging to x* and yn satisfying Vn>0, yn € F(xln) П F(x2n) Let us set hn := \\х\п — X2n\, which converges to 0, and Un ■= (Xln ~ X2n)/hn The elements un do belong to the unit sphere, which is compact. Hence a subsequence (again denoted) un does converge to some и different from 0. Then for all large η yn € F(xln) Π F(x2n) = F(x2n + hnun) Π F(x2n) С F(x*+-yB) we deduce that a subsequence (again denoted) yn converges to some у € F(x*) (because Graph(F) is closed.) Since the above relation implies V η > 0, yn + hn0 £ F(x2n + Кип)
210 5 — Derivatives of Set-Valued Maps we deduce that . 0 € PF(x*,y)(u) _ ...... Hence we have proved the existence of a non zero element of the kernel of PF(x*, y), which is a contradiction. □ 5.5 Qualitative Solutions This application deals with problems motivated by economics in the sixties (comparative statics) and, recently, by a domain of Artificial Intelligence known under the name of "qualitative simulation" or "qualitative physics." It concerns "sign solutions" (or confluence solutions, to adopt the terminology used in Artificial Intelligence) of nonlinear equations and inclusions. We shall provide a criterion involving adequate linearizations of the sign problem, which allows to obtain sign solutions for both the original problem and its linearization. Let us consider a closed set-valued map F : Rn ~> Rm and the problem find χ £ Rn such that F(x) Э у We associate with Rn the η-dimensional confluence space 1Zn defined by Kn := { -, 0, + }n whose elements are denoted by a := (αϊ,..., an)· Given a multi-sign b ε TZm, does there exist a "qualitative solution" a £ 7гп, defined by: 3 χ &y € F{x) such that sign (ж) = a h sign(y) = b (5.3) We provide a criterion4 for answering this question. Before stating our first theorem, it is convenient to introduce the notations Qn(a) := R£ := {ν ε Kn \ sign of (Vi) = a;} Qn(a) := аЩ := {ν <= Kn \ sign of fa) = a; or 0} 4As in numerical analysis, which deals both with approximating infinite- dimensional problems by finite-dimensional ones and with solving them, problems of qualitative analysis arise at two levels: the passage from "quantitative to qualitative" and algorithms for solving the latter. Only the first aspect is investigated here in the framework of solving this inclusion.
5.5. Qualitative Solutions 211 The cone <2*(a) is defined by: (Vi > 0 if a; = + Vi < 0 if di = - Vi £ R if aj = 0 Observe that <2n (a) is the closure of Qn (a) and that the positive polar cone of Qn{a) is the cone Q*n{a). Theorem 5.5.1 Let α ε Пп, bell™, F : Rn ~> Rm be α closed set-valued map. Consider xQ &Oom(F) П Qn(a) satisfying F(xQ)r\Qm(b) φ%. If the criterion {(0,0) is the only solution (p, g) to ρ ε CF(xQ,yQ)*(q) (5.4) satisfying ρ ε Q*(a) & g ε -Q^(b) holds true, then a solves the qualitative inclusion 3x ε Qn{a) such that F(x) nQm{b) φφ (5.5) Proof— Pick τ/ο ε i^o) П Qm(b), then (a;o,yo)_belongs to the intersection of the graph of F and the closed convex cone Qn(a) χ Qm(b), so that {{xo,ya), (ха,Уа)) is a solution in Graph(F) χ (Qn(a) χ Qm{b)) to the equation (zi,2/i)-(z2,2/2) = ° Let leR" denote the unit vector 1 := (1,..., 1) and set αϊ := (аД)^!,...,! We shall prove that for some ε > 0 there exists a solution (zi.yi) ε Graph(-F), (za.ito) 6 Qn(a) xQm{b) to the equation (»i,J/i)-(a;2,2/2) = е(а1,Ы) So that Zi = 2Γ2+εα1 belongs to Qn(a) and j/i = y2+ebl belongs to Qm(b) and to F(x1). For that purpose, we apply Theorem 5.4.6, which states that a solution tothe above equation does exist provided the assumption CGraph(F)fro, i/b) - CqM Ы x C^m(b)(yQ) = R" χ Rm is satisfied.
212 5- Derivatives of Set-Valued Maps Since xq belongs to Qn(a) := a72™, the cone C-q гаЛхо) coincides with the tangent cone, which contains Qn (a) (recall that when_ Q is a_ convex cone, TQ(x) = cl(<3 + a;R) D Q.) In the same way, С^(Ь)(у0) D Qm(b). The above assumption would follow from СгарЦСЩздо)) - (QM x Qm(b)) = R" x Rm By polarity, this is equivalent to the condition (G™ph(CF(xQ,yQ))rn(Qn(a)xQm(b))+ = {0} which is nothing other than condition (5.4.) Π In the single-valued case, we obtain: Corollary 5.5.2 Lei_a ε ΊΙη, b ε 1lm, F : Rnj-> Rm be a closed set- valued map. Let xq ε Qn(a) be such that f(xo) ε Qn{b) and assume that f is continuously differentiable at xq. Then the criterion 0 is the only solution q to q ε —Qm{b) & f'(xa)*Q 6 Qn(a) ; implies that a solves the qualitative equation 3 χ such that sign(x) = a & sign(/(a;)) = b Remark — When A is a closed convex process, we can apply the above theorem with (xQ,yQ) = (0,0) ε Graph(A)n(Qn(a)xgm(b)) Since CA(Q, 0) = A (because the tangent cone to the closed convex cone Graph(A) at the origin is equal to this closed convex cone), we deduce from the above theorem the following consequence: Corollary 5.5.3 Let α ε 72.™, b ε 1Zm and A be a closed convex process5 fromBJ1 toRm. If {(0,0) is the only solution (p, q) to ρ ε A*(q) such that ρ ε Q*n{a) h q ε -Q*m{b) then a solves the qualitative inclusion 3i£ Qn(a) such that A(x)nQm(b) φ 0 5Closed convex processes provide a way to represent uncertainty for linear operators analogous to "sign matrices." They possess better mathematical properties than sign matrices. In particular, since we cannot extend "sign matrix" to nonlinear maps, we can not say that a sign matrix is the derivative of a sign operator. Observe that this corollary provides the same criterion for both the initial inclusion and its set-valued linearization.
5.5. Qualitative Solutions 213 Therefore, criterion (5.4) implies the existence of a qualitative solution to both problem (5.5) ала the lineaxized problem 3u ε Qn(a) such that CF(xa,ya)(u) Π Qm(b) /|D Let us consider the case when F : Rn ~> Rm is the set-valued map defined by: T?( \ = I f(x) ~ M№ when x e L [X) : \ 0 when χ i L where / : Rn i-+ Rm is continuous, L is a closed subset of Rn and Μ : Rn ~> Rm is a closed set-valued map. Corollary 5.5.4 Let a <=Tln, b <= Um, Μ : Kn ^ Rm be a closed set- valued map and L с Rn be a closed set. Assume that xq and j/q satisfy xQ £ Qn(a)nL & yo ε {f{xQ)-M{xQ))nQm{b) that f is continuously differentiable at xq, Μ is Lipschitz at xq and that CL{xQ) -Dom{CM{xQ,yQ - f{xQ))) = Rn The criterion (0,0) is the only solution (p, q) to ■ Ρ e f'{xo)*{q) + CM(xa,ya - /{χ0))*(-9) + Νφο) (5-6) _ such that ρ ε Ql(a) & q ε -Q*m{b) implies that a solves the qualitative equation 3i£ Qn(a)ilL & у ε Qm{b) such that у ε f(x) - M{x) It follows from Theorem 5.5.1 and Proposition 5.2.3. The particular case when Μ is constant yields Corollary 5.5.5 Let α ε Пп, b ε 1lm and L с Rn, Μ С Rm be closed sets. Assume that xq and j/q satisfy xQ ε α»η£ & уо ε (/Ы-м)пдт(Ь) Assume that f is continuously differentiable at xq. Then the criterion (0 is the only solution q to q £ -Q*m(b)nNM(f(xa)-yQ) & /'Ы*«ед;(а)-%Ы implies that a solves the qualitative equation 3x ε (3„(a)ni & у ε <Эт(Ь) such that /(ж) ε y + М
214 5 — _ Derivatives of Set-Valued Maps Another example is provided by qualitative solutions to variational inequalities. Example Variational Inequalities Let us consider a closed convex subset К С Rn and a C1-map / from a neighborhood of К to Rn. An element yQ ε Rn being given, we recall that an element xQ ε Κ is a solution to the variational inequalities V χ ε -ЙГ, < f(xQ) - 2/0, xa - χ > < 0 if and only if xq is a solution to the inclusion 2/o € f{xa) + NK{xQ) where Nk(xq) is the normal cone to К at xq. Theorem 5.5.6 Let α ε Пп, bell™ and let xQ ε Qn(a) ПК, yQ ε Q^b) satisfy Vχ ε Κ, < f(xQ) -ya,xo -χ > < 0 The criterion (0,0) is the only solution (p, q) to ■ Ρ e f'(xQ)*(q) + CNK(xQ,yQ-f(x0))*(-q) (5.7) _ satisfying ρ ε Q*n(a) & q £ -Q*m{b) implies that a solves the qualitative variational inequalities (3x ε Qn{a)nK & у ε Qm{b) such that \fx ε Κ, <f{x)-y,x-x>< 0 Proof — This a consequence of Proposition 5.1.2 with M(x) := -NK{x) and Theorem 5.5.1. Π Remark — Recall that Proposition 5.2.7 provides a characterization of the circatangent derivative of the normal cone map К Э χ ~> Νκ (x) through the circatangent derivative of the best approximation projection πχ to the closed convex subset K: q ε CNK{x,p){u) <£=^ и ε CirK(x + p)(u + q) which can be used to check criterion (5.7.)
5.6. Higher Order Derivatives 215 5.6 Higher Order Derivatives Naturally, by taking mi/l-order contingent sets to a graph of a set-valued map F : X ~> Y, we obtain concepts of high-order derivatives. Definition 5.6.1 Let Χ, Υ be normed spaces and F : Χ ~> Υ be a set- valued map. The mi/l-order contingent derivative DF(x, y) ofF at (x, y) £ Graph(.F) is the set-valued map from X to Υ defined by Graph(D(m)-F(a;,2/,Ui, i>i,..., Jim-i,Vi)) ··= T^)a:ph{F)(x,y,u1,v1,...,um-1,vm-1) We say that it is m—th derivable if its graph is m—th derivable. This definition implies right away that D(m)(F-l){y,x,vi,Ui,...,vm-i,um-i) = (D(m)F(x,y,Ui,vl,...,um-i,vm-i)yi For simplicity, we shall restrict ourselves to second order contingent derivatives of set-valued maps. If if is a subset of X and / is a single-valued map which is twice continuously differentiable around a point χ ζ. Κ, then the second order contingent derivative of the restriction of f to К is given by the formula: D<V{f\K){x,Ul){u2) = fl(24/l*)fo/(a0,ui,/'(a0ui)(u2) = f'{x)u2 + \f"{x){ui,ui) whenever u2 eT^'(x,Ui) It is empty when u2 ^ T^\x,Ui). The proof of this fact is similar to the proof of the following proposition. Proposition 5.6.2 Let X and Υ be normed spaces, f be a twice continuously differentiable single-valued map from an open subset ilcX to Y, Μ : X ~~y γ be a set-valued map and L с X. Let F : Χ ~> Υ be the set-valued map defined by: ■pi \ — j f(x) ~ M(x) when χ ε L [x) '— 1 0 when x<£L
216 5- Derivatives of Set-Valued Maps Then, for each χ £ nnDom(.F), у £ F(x) and V\ £ DF(x,y)(ui), the second order derivative D^ F{x,y,u\,Vi)(u2) in the direction^ £ T^ '(x,Ui) is contained in /'(aO«2 + |/'Ή(«ι,«ι) - D^M(x,f(x)-y,u1,f'(x)u1-v1)(u2) Furthermore D^F(x,y,Ui,Vi)(u2) is empty whenever u2 £ Т^'(х,щ). Equality holds true when we assume that either L or Μ is twice derivable and Μ is Lipschitz. Proof— Let v2 belong to D^F(x,y,ui,vi)(u2). Then there exist hn > 0 converging to 0 ала sequences U2n and V2n converging to u2 and v2 respectively such that χ + hnU\ + h^u2n belongs to L and Vn>0, y + hnvi+hlv2n £ f(x + hnui + h2lu2n)-M(x + hnui + h2lu2Tl) This implies that u2 belongs to Tl(x,ui). Since / is twice continuously differentiable (f(x + hnui + hlu2n) = f(x) + ft™/'(:c)ui + h2n(f'(x)u2 + £/"(aO(ui,Ui)) + e(hn)) where e(hn) converges to 0 with hn. Then f{x) -y + M/'(aOui - "ι) + h2n{f'(x)u2 + \f"{x){uuщ) -v2 + e{hn)) £ M(x + hnui + hlu2n) and thus f'(x)u2 + \f"{x)(u\,u-\) — v2 does belong to D^M(x,f(x)-y,ul,f'(x)ul-vl)(u2) Conversely, assume for instance that Μ is twice derivable, that u2 belongs to T"l{x,U\) and that f{x)u2 + ^f"{x){uuul)-v2 £ D^M{x,f{x)-y,ul,f'{x)u1-vl)(u2) Hence, there exist a sequence hn > 0 converging to 0 and sequences u2n and w2n converging to u2 and f'{x)u2 + ^f"(x)(ui,Ui) - v2 respectively such that χ + hnui + h\u2n £ L and f{x)-y + hn(f'{x)ui-v1) + hlw2n £ M(x + hnui + h2nu2n)
5.6. Higher Order Derivatives 217 (we have used the Lipschitz continuity of M). Then for some e(hn) —> 0, the sequence V2n ■= f'{x)u2 + ^f"{x)(ui,u1)+£{hn)-w2n converges to v2 and satisfies у + hnvi + hlv2n e f(x + hnui + h2nu2n) - M(x + hnui + hlu2n) □
Chapter 6 Epiderivatives of Extended Functions Introduction When we consider real-valued functions, the order relation on R is involved in optimization (for rinding a minimum) or in Lyapunov functions (which decrease along solutions to differential equations or inclusions.) These are reasons why we have to associate with a real-valued function V the set-valued map Vf defined by VT(s) := F(a;)+R+ whose graph is the epigraph of V and consider not only the contingent derivative of V, but also the contingent derivative of Vj. We shall check in Section 1 that its graph is the epigraph of a lower semicontinuous function, which is called the contingent epiderivative D^V(x) of V at x. For instance, the Fermat rule1 states: if χ minimizes V, then DiV(x)(v) > 0 for all ν £ X, 'which should be called Oresme-Fermat Rule, Nicholas Oresme (1323-1382), Bishop of Lisieux, 300 years ahead of his time, anticipated the great Fermat himself by observing in his treatises De Proportionibus Proporiionum and Algo- rismus Proporiionum that near a maximum, the increment of a variable quantity becomes 0. (He also anticipated the coordinate system in another book Tractatus de Latitudine Formarum, Copernicus' views and was only one step to the logarithms.) He did not hesitate to attack repeatedly the extravagant presumptions of astrologers (called at the time mathematicians!..., the mathematicians were then 219
220 6- Epideriv&tives of Extended Functions so that necessary conditions in optimization and optimal control can be derived from this simple rule, provided that a rich and exhaustive calculus of contingent epiderivative is available. Section 3 is thus devoted to such an epidifferential calculus. As for derivatives of set-valued maps, an epidifferential calculus involves not only the use of contingent epiderivatives, but of other epiderivatives2 corresponding to various tangent cones. So, adjacent D\V{x) and circatangent С|У(ж) epiderivatives are presented in the second section, as well as convex epiderivatives. We study in Section 4 dual concepts: the subdifferential doV(x) and the generalized gradient dV(x) of an extended function defined respectively by d0V(x) := {p € X* | Vν € Χ, <ρ,ν>< ΌΊν(χ)(ν)} dV(x) := {peX*\Vv€X, <p,v>< C}V{x){v)} They naturally coincide when V is convex or when it is continuously differentiable (in this case, they are made up of the gradient V'(x).) The subdifferential is always contained in the generalized gradient, and even more, in the case of finite dimensional vector-spaces, the upper limit of the subdifferentials doV(xn) when xn converges to χ is contained in the generalized gradient dV(x). Again in finite dimensional vector-spaces, the subdifferential coincides with the (local) subdifferential d-V(x) (used in the concept of viscosity solutions of Hamilton-Jacobi equations), which is the subset of ρ G X* such that liminf VW-VM-<P,*-E2L > о X '•E 11 uif JLi 11 This still increases the population of our menagerie with species doomed to disappear through Darwinian evolution for lack of duality, since their use does not allow recovery of the original information ου epiderivatives. called geometers), whose persuasive arguments influenced affairs of the state, and he even criticized the Pope. A clear thinker, he translated Aristotle's books into French and commented on them with extremely original views. 2This procedure is also known under the name of the epigraphical approach.
6.0. Introduction 221 The use of generalized gradients for functions is then recommended in a "convex environment, i.e., for convex, and more generally, episleek functions (with convex epigraphs.) However, there are other reasons than the exploitation of duality for using these dual concepts. One is their familiarity with more classical concepts. For usual functions on Hilbert spaces, there is a canonical identification between, say, a derivative of a differentiable function and its gradient, and it became traditional to formulate many results in terms of gradients and normal cones. Think of the traditional role of the Lagrange multipliers in constrained optimization, for instance, which appeals to generalizations stated in a dual form (Kuhn-Tucker rule, maximum principles in optimal control.) Another reason lies in the fact that the first attempts to generalize the concept of gradients used limiting procedures. Since it seems easier to take limits of elements (the gradients, for instance) than limits of functionals (the associated directional derivatives, for example), many generalizations of the concept of gradient dealt with set of limits or cluster points of gradients taken in various ways. These approaches are still pursued (in the framework of "proximal analysis", for instance), but we do not dwell on them in this book. When we deal with convex functions, which are the extended functions with convex epigraphs, all these epiderivatives coincide. Furthermore, lower semicontinuous convex functions enjoy duality properties: We recall in Section 5 what the Fenchel conjugate of a real-valued function is and that this Fenchel conjugacy is a bijective correspondence between lower semicontinuous convex functions on a Banach space and its dual. It plays a role similar to the one played by the Legendre transform, since the inverse of the subdifferential map of a convex function is the subdifferential map of the conjugate function. These facts are at the origin of duality theory in optimization. We also recall here the Fenchel Theorem which provides a criterion for the existence of solutions to convex minimization problems, from which one can derive a formula for computing the conjugate of the composition of a function with a continuous linear operator. We finally point out that the subdifferential map of a convex
222 6 - Epideriv&tives of Extended Functions function is a maximal monotone map. Its Yosida approximation is the gradient of a continuously differentiable convex function, called the Moreau-Yosida approximation. We conclude this chapter with a presentation of high order epi- derivatives, defined from the high-order derivatives of the set-valued map Vf associated with an extended function. We shall also estimate the second-order contingent epiderivative of the Moreau-Yosida approximation of a convex function. 6.1 Contingent Epiderivatives 6.1.1 Extended Functions and their Epigraphs Let us consider an extended function V : X i-> R U {±00} whose domain Dom(F) := {x € X \ V(x) φ ±οο} is not empty. (Such a function is said to be proper in convex and non smooth analysis. We shall rather say that it is nontrivial in order to avoid confusion with proper maps.) Any function V defined on a subset К С X can be regarded as the extended function Vk equal to V on К and to +00 outside of K, whose domain is K. An extended function is characterized by its epigraph £p(V) := {(χ,λ) € Χ χ R I V(x) < X} An extended function V is convex (resp. positively homogeneous) if and only if its epigraph is convex (resp. a cone.) We also observe that any positively homogeneous extended function is non trivial whenever V(0) φ —oo. In this case, V(0) = 0. The hypograph of a function W : X —> R U {±00} is defined in a symmetric way: Hp(W) := {(χ,λ) ε Xxn\W(x) > A} = -£p(-W) Such a function is concave if and only if — W is convex, i.e., if and only if its hypograph is convex.
6.1. Contingent Epiderivatives 223 The main examples of extended functions are the indicators φκ of subsets К defined by . , , /θ if χ € К ФкМ ■■= { +oo if not The indicator φκ is lower semicontinuous if and only if К is closed and ψ χ is convex if and only if К is convex. One can regard the sum V + φκ as the restriction of V to K. We recall the convention inf(0) := +00. Lemma 6.1.1 Consider a function {:Ih RU{±oo}. Its epigraph is closed if and only if \/v € Χ, δ(ν) = liminf<5(?/) V1—*V Assume that the epigraph of δ is a closed cone. Then the following conditions are equivalent: i) \fv € Χ, δ(ν) > -oo ii) <5(0) = 0 Hi) (0,-1) $ Ερ{δ) Proof — Assume that the epigraph of δ is closed and pick ν € X. There exists a sequence of elements vn converging to ν such that lim δ(νη) — liminf <5(г/) η—юо υ'—*ν Hence, for any λ > lim inf„/_>„ δ(υ'), there exist iV such that, for all η > Ν, δ(νη) < λ, i.e., such that (г;п, λ) € ερ(δ). By taking the limit, we infer that δ(ν) < λ, and thus, that δ(ν) < liminfi;'_+i;5(?;')· The converse statement is obvious. Suppose next that the epigraph of δ is a cone. Then it contains (0, 0) and 5(0) < 0. The statements ii) and Hi) are clearly equivalent. If i) holds true and <5(0) < 0, then belongs to the epigraph of δ, as well as all (0, —λ), and (by letting λ —> +oo) we deduce that <5(0) = —00, so that г) implies ii).
224 6 - Epidenvatives of Extended Functions To end the proof, assume that <5(0) = 0 and that for some v, δ(ν) = -oo. Then, for any e > 0, the pair (v, —l/ε) belongs to the epigraph of δ, as well as the pairs (εν, —1): By letting e converge to 0, we infer that (0, —1) belongs also to the epigraph, since it is closed. Hence <5(0) < 0, a contradiction. □ 6.1.2 Contingent Epiderivatives We can also regard an extended function V as the set-valued map V : X ~> R defined by V(x) : j V(x) if ж € Dom(V) 0 if χ i Dom(V) so that we can define in the usual way the contingent derivatives of V at ж € Dom(V): BV(i)(ti) := DV(x,V(x))(u) = {v | liminf/^o+y-ni \V(x + hu') - V(x) - hv\/h = 0} However, minimization problems and Lyapunov functions involve obviously the order relation on R. Hence, when dealing with such problems, it is convenient to consider two new set-valued maps V| and Vj defined in the following way: i) VT(x) := { ii) Vi(x) : V(x) + ~R+ if ж € Dom(V) 0 if V(x) = +oo R if V(x) = -oo V(x) - R+ if ж € Dom(V) 0 if V(x) = -oo R if V(x) = +00 We see at once that Graph(VT) = Sp{V) L· Graph(V|) = Hp(V) Therefore, we are led naturally to associate with these two set-valued maps Vf and Vj their contingent derivatives. We observe that the
6.1. Contingent Epiderivatives 225 values of the contingent derivative of Vj are obviously half lines in the sense that VA > V(x), Vii € Dom(DVT(a;,A)), £>VT(:c,A)(?x) = DVT(:c,A)(?x)+R+ Definition 6.1.2 Let V:Ih>RU {i°°} be α nontrivial extended function and χ belong to its domain. We shall say that the function D-\V(x) from X to RU {±00} defined by Vu €l, D}V(x)(u) := inf{i;|7; € DV^x, V(x))(u)} is the contingent epiderivative ofV at χ in the direction u. The function V is said to be contingently epidifferentiable at χ if its contingent epiderivative never takes the value —00. It is said to be episleek (at x) if its epigraph is sleek (at (x, V(x))). Actually, the contingent epiderivative can be characterized as a limit of differential quotients: Proposition 6.1.3 Let V : Χ i-> RU{±oo} be a nontrivial extended function and χ belong to its domain. Then η лп и ^ ν ■ f V(x + hu')-V(x) D-tVixjiu) = hmmf — -1 ^-J- 1 ' /ι->0+,ιι'->ιι h The function V is contingently epidifferentiable at χ if and only if ΌΊν(χ)(0) =0. Proof — Assume first that D^V(x)(u) < +00. By the very definition of the contingent epiderivative, for any λ > D^V(x)(u), there exists some ν € DV^(x, V(x))(u) smaller than λ. Therefore, we can associate with any e > 0 elements h €]0, e[ and (u1, v') € B((u, v), e) satisfying V(x) + hv' > V(x + hu'). So that V(x + hu') - V(x) η Hence the liminf of the differential quotients (V(x + hu') — V(x))/h is smaller or equal to D^V(x)(u).
226 6 - Epiderivatives of Extended Functions Conversely, let hn > 0 converging to 0, un converging to и and λη > V(x + hnun) be such that the differential quotients vn '·= (K-V(x))/hn converge to .. . c V(x + hu')-V(x) ν := hmmf — -1 ^-L h—>0+, и'—щ h Then the pairs (x + hnun, V(x) + hnvn) belong to the graph of Vj, so that ν € DV^(x)(u). Hence .. . , V{x + hu')-V{x) ,, hmmf -f ^ > D}V{x)(u) Consider now the case when D^V(x)(u) = +oo, i.e., the case when the subset DVT(i, V(x))(u) = 0. Then for any ν € R, for any hn > 0 converging to 0 and any sequence un converging to u, we have V(x + hnun) > V(x) + hnv for η large enough, so that the liminf of the differential quotients is equal to +oo. □ We define in a symmetric way the contingent hypoderivative DiV(x) from X to R U {±oo} of V : X н-> R U {±00} at a point χ of its domain by D{V{x){u) := -DT(-V)(s)(u) = limsup У(ж + hu') ~ V(x) h-*a+,u'—>u "■ We could also have defined the contingent epiderivative of a function by taking the contingent cone to its epigraph: Proposition 6.1/4 Let V : X (-► RU{±oo} be a nontrivial extended function and χ belong to its domain. Then the contingent cone to the epigraph of У at (ж, V(x)) is the epigraph of the contingent epiderivative of V at x: εΡ(ΌΊν(χ)) = T£p{v)(x,V(x))
6.1. Contingent Epideiivatives 227 Figure 6.1: Epigraph of the Contingent Derivative
228 6 - Epiderivatives of Extended Functions Consequently, the epigraph of the contingent epiderivative at χ is a closed cone. The contingent epiderivative D^V{x) is then lower semicontinuous and positively homogeneous whenever V is contingently epidifferentiable at x. Furthermore, Vw > V[x), T£p{v)(x,w) С TOom{v)(x) χ R and Vw > V(x), Oom(DxV(x)) χ R С T£p{v)(x,w) If the restriction of V to its domain is upper semicontinuous, then for all w > V(x), T£p{v)(x,w) = TOom{v)(x) χ R Proof 1. — The first claim follows from the very definition of the contingent epiderivative. Fix w > V(x). Let us assume that (u, v) belongs to Тщу){х, w). We infer that there exist sequences un, vn and hn > 0 converging to u, ν and 0 such that w + hnvn > V(x + hnun) We thus deduce that и belongs to the contingent cone to the domain of V at x, and thus, that T£p(y){x, w) С Tj)om,v,(x) χ R. 2. — If и belongs to the domain of the contingent epiderivative of V at x, if w > V(x) and if ν is any real number, we check that (u,v) belongs to T£p(yj(x,w). Indeed, there exist sequences of elements hn > 0, un and vn converging to 0, и and D-\V(x){u) respectively such that (x + hnun, V{x) + hnvn) € Sp(V) But we can write (x+hnun,w+hnv) = (x+hnun,V(x)+hnvn)+(0,w-V(x)+hn(v-vn)) Since w — V(x) + hn(y — vn) is strictly positive when /j,T1 is small enough, we infer that (x + hnun, w + hnv) belongs to the epigraph of V, i.e., that (u,v) belongs to the cone T£p(y)(x,w).
6.1. Contingent Epiderivatives 229 3. — Let w be strictly larger than V{x) and и belong to T-£)0m/V\(x)■ Then there exist sequences un and hn > 0 converging to и and 0 such that V(x + hnun) < +oo for all n. When V is upper semicontinuous on its domain, for all 0 < e < (w-V(x))/2 there exists η > 0 such that, for all /ιπ||ϊχη|| < η, we obtain V{x + hnun) < V{x) + ε < w - ε Let ν be given arbitrarily in R. Then, for any hn > 0 when ν > 0 or for any hn €]0, 73jj[ when ν < 0, inequality ги — e < ги + hnv implies that V(x + hnun) < w + hnv, i.e., that the pair (u, v) belongs to TSp(y)(x,w). Π Proposition 6.1.5 LetV : Χ ι—» RU{±oo} be an extended function and χ belong to its domain. Take any и € Oom(D^V(x))nOom(DiV(x)) Then {ВтПх)(и), !>№)(«)} С DV(x)(u) С [ВтК(х)(и),и^(х)(и)] Proof— Since Graph(V) = £p(V) П Wp(V), we deduce that the inclusions TGraphooO^H) С V)^^^^^^^1)) can be translated into Graph(DV(a;)) С £p(Z>TF(s)) П Ир(В^(х)) from which the inclusion DV(x)(u) С [Βτν(α:)(«),.Ο^(α0(«)] follows. Since the contingent epiderivative of У at ж in the direction и is equal to ,ч/ч ,. η VYx + ftu') - V(x) DrV(x)(u) := liminf — ^ ^ 1 Ч /i-+0+,u'-ni /l
230 6 - Epiderivatives of Extended Functions we see that D^V(x)(u) is the limit of a subsequence of differential quotients (y(*+fcO-y(*?j wnen fc_ 0+ and thus, that ΌΊν(χ)(η) belongs to DV(x)(u). The same is true for the contingent hypo derivative. □ Theorem 6.1.6 Let V : X i-> R U {±00} be an extended function which is continuous on a neighborhood of χ € Dom(V). Then for every и €' DomCUTV(s)) aBom{DiV{x)) we have DV(x)(u) = [D1V(x)(u),DlV(x)(u)} Proof— By Proposition 6.1.6, it remains to show that when V is continuous on a neighborhood of x, the inclusion ΊΈχν)(χ>ν(χ)) ПГНр(у)(х,К(х)) С TGTaph(y)(x,V(x)) holds true. Pick any (u, v) in the intersection: There exist sequences h^ > 0 and h!j^> 0 converging to 0+, and sequences of pairs (u£, ?4) and (uj^, v^) converging to (u, v) such that V{x + hjui)-V{x) T V{x + hjui)-V{x) , hi ~ Vn & Л* We claim that it is enough to show that for every η large enough, there exists λη € [0,1] such that, setting К ··= (l-\n)hl + \nhi and (l-Xn)hlul + \nh^n (l-An)ftTt4 + Awft*4 Un ;= & ^ ;= we have the following equality V(x + frn"n) - V(x) hT = Vn Indeed hn > 0 converges to 0+ and (un, vn) converges to (u, v). So, (u, v) belongs to the contingent cone to the graph of V.
6.1. Contingent Epiderivatives 231 To find such λπ, consider for all large η the continuous functions ψη : [0,1] —> R defined by [ν(χ + (ΐ- λ)4«τ + λ/44) - (v(x) + (i - \)hlvl + xh&i) and observe that ψη(0) < 0 & ψη(1) > 0 Since ψη is continuous, there exists λπ € [0,1] such that φη{λη) = 0. Then λπ satisfies.:the required property. □ The contingent epiderivative coincides with the directional derivative < V'(x),u > when V is Frechet differentiable. If V is Frechet differentiable at a point χ € AT, then the contingent epiderivative of the restriction is the restriction of the derivative to the contingent cone: The formulas become much more simple when V is Lipschitz: the contingent epiderivative coincides with the lower Dini derivative : Proposition 6.1.7 Let us assume that V:IhRU {±oo} is Lipschitz at a point χ of its domain. Then V(x + hu) — V(x) D-[V(x)(u) = liminf — (the lower Dini derivative) h—*0+ fl and satisfies for some i>0: Vu e X, \D<tV(x)(u)\ < l\\u\\ Remark — There are other intimate connections between contingent cones and contingent epiderivatives. Let фк be the indicator of a subset K. Then it is easy to check that Di(1>k)(x) = VTk(x) Therefore we can either derive properties of the epiderivatives from properties of the tangent cones to epigraphs, as we did, or take the opposite approach by using the above formula. D
232 6 - Epiderivatives of Extended Functions There is also ал obvious link between the contingent cone and the contingent epiderivative of the distance function to К since we can write for every χ 6 K: — TK(x) = {υ £ X | D}dK{x){v) = 0} We also mention the following estimate of the contingent epiderivative of the distance function to K: Proposition 6.1.8 Let К be α closed subset of α normed vector-space and Пйг(у) be the set of projections of у onto K, i.e., the subset of ζ £ Д" such that \\y — z|| = dfc(y)- Then when Т1к(у) Φ 0> we have the following inequalities: D^dK{y){v) < d(v,TK(UK(y))) Proof — We begin by proving this inequality when у belongs to K. Indeed, for all ω £ X, inequality di((.y + hv) < ак(у + hw) + h\\v — w\\ implies that D^dfc(y)(v) = liminf djc{y + hv)/h < \\v — w\\ when w £ Тк{у) h—*Q+ Assume next that у £ K. Choose ζ £ TiK(y). Then dK(y + hv)-dK(y) < \\y-z\\ +dK(z + hv)-dK(.y) = dK(z + hv) Prom the first part of the proof, we deduce that Dj άκ(ζ)(ν) < d(v, Τκ(ζ)), and consequently, that D-\dK(y)(v) < d{v,TK{z).) Π 6.1.3 Fermat and Ekeland Rules Since we can define the contingent epiderivative of any extended function V:XhRu {+oo}, we can extend the "Fermat rule" to any minimization problem. Theorem 6.1.9 (Fermat Rule) Let V : X н-> R U {+00} be a nontrivial extended function on a normed space X and χ € Dom(V) a local minimizer ofV on X. Then χ is ά solution to the variational inequalities: Vu € X, 0 < D}V{x){u)
6.1. Contingent Epiderivatives 233 Proof — The proof is naturally obvious: We write that for all v! € X and all h > 0, 0 < (V(x + hu')-V(x))/h and we take the lim inf when h converges to 0 and v! to и. П The converse statement is not true without further assumptions such as convexity or, more generally, pseudo-convexity, or second order conditions: Definition 6.1.10 An extended function V : X i-> R U {±00} is called pseudo-convex atx€. Dom(V) if its epigraph is pseudo-convex at (x, V(x)), i.e., if Vy € Χ, ΌΊν(χ)(ν-χ) < V(y)-V(x) We infer easily from this definition the converse of the Fermat Rule: Proposition 6.1.11 Let V : X н-> R U {±00} and χ € Dom(V) satisfy 0 < D-[V{x){u) for all и € X. If V is pseudo-convex at x, then χ achieves the minimum ofV. What is not obvious is the use of this Fermat rule for more and more general problems, when the function V is built from other simpler functions and involves constraints. The search for necessary conditions for a minimum requires quite a rich calculus of contingent epiderivatives which provides estimates of D^V(x)(u). In particular, when constraints (of the type χ € К) are involved, the fact that the epiderivative of the restriction to К is the restriction of the epiderivative to the contingent cone TV (ж), allows one to write necessary conditions using also contingent cones to constraint sets (or in the dual form, using gradients and polars of the contingent cones.) In the same way, it is easy to derive an epidifferential version of Ekeland's Variational Principle: Theorem 6.1.12 Let X be a Banach space, V : X (-► R+ U {+00} be a nontrivial lower semicontinuous bounded from below function
234 6 - Epiderivatives of Extended Functions and xq € Dom(V) be a given point of its domain. Then, for any ε > 0, there exists a solution χε € Dom(V) to: (i) V{xe) +e\\xe-xo\\ < V{xq) (6-1) it) Vu € Χ, Ο < ΌΊν(χε)(η) + ε\\η\\ 6.1.4 Elementary Properties We present below some formulas concerning epiderivatives which allow exploitation of the Fermat and Ekeland rules. 1. — Sum and composition Proposition 6.1.13 Let us consider two Banach spaces X, Y, a single-valued map f : X н-» У and two extended functions V and W from X and Υ to RU {+00} respectively. Let xq belong to the domain of the function V+Wof and assume that f is Frechet differentidble at xq. Then D}V{x0){u)+D}W{f{x0)){f'{x0)u) < D}{V + W о f){x0){u) In particular, if К is a subset of X and if xq € К nDom(V), then Vu € TK(xo), D^V(x0){u) < ΌΊ(ν\κ)(χο)(η) 2. — Supremum of Functions Let us consider now a family of functions Vi-.X ^ RU{±oo}, (»€/) and let us associate with it the function U defined by U(x) := supVi(x) We set I(x) := {i € / | Ц,(х) = U(x)}. The following estimate is obvious whenever I(x) is not empty and χ € Dom(i7)·. У и € X, sup D}Vi(x){u) < D^U{x){u) (6.2) i£l{x)
6.1. Contingent Epiderivatives 235 because, for all г € I(x), Vi(x + hu) - Vj{x) U(x + hu) - U(x) h ~ h 3. — Marginal Function Let us consider normed spaces X and Υ and an extended function U:X хУн->Ни{±оо}. We associate with it the marginal function I^iIhRu {±°o} defined by V(x) := mfU(x,y) Let π denote the projection from XxYxHtoXxH. We observe that: kEP(U) С Ep(V) С irEp(U) The first inclusion is obvious. The very definition of the infimum implies that for every e > 0 and every (χ, λ) € Ep(V), there exists ye € Υ such that (x, ye, λ + e) belongs to Ep(U). Taking projections and passing to the limit, the second inclusion ensues. Let xq € X be given. Suppose that there exists yo €. Υ which achieves the minimum of U(xq, ·) on Y: V(xq) = U(x0,yo) φ ±οο By taking contingent cones to the above inclusions, we obtain the inequality Vu € X, D-tV(xo)(u) < liminf finf D-tU(xo,yo){u',v)] u1—>u \v& J because -KSp^D^Uixo.yo)) = тгТ£р^(х0,у0,и(хо,уо)) С Τπ{ερ{υ))(χ0,ν(χο)) = T£p{v)(xq,V(xq)) = Sp(D}V{xQ)) can be easily translated into the desired inequality. Π In order to obtain the equality sign in some of the above formulas, we need further assumptions, such as the functions are episleek or pseudo-convex, which actually require the introduction of adjacent and circatangent epiderivatives. These are the topics of the next section.
236 6 - Epiderivatives of Extended Functions 6.2 Other Epiderivatives We recall the definition of "lim sup inf" (or Γ-convergence) of functions of two variables introduced in Definition 5.2.4: ]imsupa!,_KBiiifB#_n,0(a:/Jy/) := sup inf sup inf Mx\y') e>0 J7>° i'eB(i,ij) v'€B(y,e) In the same way, we set lim in£xi_>x sup4i_>4<j>(x', y') := inf sup inf sup ф(х',у') ε>0 η>0 χ'£Β{χ,η) у'еВ{у,е) 6.2.1 Adjacent and Circatangent Epiderivatives We can associate with the adjacent and Clarke tangent cones the adjacent and circatangent derivatives of V at χ G Dom(V) of an extended function V : Χ i-> R U {±00} regarded as the set-valued map V : X ~> R taking empty values outside of the domain of V: ' DW(x)(u) := DW(x,V(x))(u) = < {v I lim sup/l_>0+ 'miui->u \V(x + hu') — V(x) — hv\/h = 0} and ' CV(i)(u) := CV(x,V{x)){u) = < {v I lim sup^o+,χ'-χ nuv^ \V{x' + hu') - V{x) - hv\/h = 0} We are also naturally led to associate with the two set-valued maps Vf and Vj, defined in the preceding section, their adjacent and circatangent derivatives at points (x,V(x)). Definition 6.2.1 (Epiderivatives) Let V : X 1-» R U {±00} be а nontrividl extended function and χ belong to its domain. We shall say that the functions D^V(x) and C^V(x) from X to R U {±00} defined respectively by ' i) Vuel, D\V{x){u) := mi\v\v£D]'V}{x,V{x)){u)} < w ii) Vuel, C}V{x){u) := inf {υ \υ€ CVt(i,V(i))(u)}
6.2. Other Epiderivatives 237 are the adjacent and circatangent epiderivatives of V at χ in the direction u. These epiderivatives can also be characterized as limits of differential quotients. Proposition 6.2.2 Let V : X i-> RU{±oo} be a nontrivial extended function and χ belong to its domain. Then ь , w ч ι- · r V(x + hu')-V(x) D]V{x)[u) = limsup^_t0+ and V(x' + hu') — λ' CiV(x)(u) = limsuph^0+iXl^xy(xl)<x^V(x)mfu>^u We define in a symmetric way the adjacent and circatangent hypoderiva- tives D\V(x) and CiV{x) from X to RU {±00} D\V(x)(u) = -D\{-V){x){u) := lim infh_0+ supu,^uV{x+^ ~V{x) and CiV(x)(u) = -C^-V){x){u) := lim rciih-+Q+,x>->x,v{x')>\'-*v(x) sup^^V^a;' + hu') - X')/h Naturally, when V is contingently epidifferentiable, the adjacent and circatangent epiderivatives are lower semicontinuous and positively homogeneous, and the circatangent epiderivative is moreover convex. They coincide with the directional derivatives < V'{x), и > when V is respectively Frechet and continuously differentiable. Remark — It is also possible to derive another interpretation of the epiderivatives in terms of epilimits of the following differential quotients r, τ,/ w ч V(x + hu)-V(x) VhV(x) : и н-> VhV(x)(u) := \ __ _ ; K-L _;_ as we shall see in Chapter 7. □
238 6 - Epiderivatives of Extended Functions Let К С Χ, χ € Κ and V be Frechet differentiable at x. Then the adjacent epiderivative of the restriction V\k is the restriction of the derivative to the adjacent cone. When V is continuously differentiable at a point χ G K, the similar statement holds true for the circatangent derivative: л nt/,π \r \r \ f <V'(x),u> tiu€lt(x) ,) Щтк)(*)(и) := |+0oH' ifnot «0 Cf №)(«)(.) := {^'W'°> !ίη„ΓΚΜ The formulas become much more simple when V is Lipschitz. Proposition 6.2.3 Let us assume that V:1hRU {i°°} is Lipschitz around a point χ of its domain. Then ρτ/γ„.ν„Λ — 4m sup - — - /ι->0+ UtV(a;)(u) = limsup 1 ■ " ■ h is the upper Dini derivative andr плп \< \ ν V{x' + hu)-V{x') C^V(x)(u) = limsup — j- —L /i->0+,a:'->3: "· Furthermore, for some neighborhood U of x, i) the map (y,u) gM χ Χ ι-> C^V(y)(u) is upper semicontinuous ii) the map и ι—» C^V(x)(u) is Lipschitz on X { Hi) Vu€X, C<t(-V)(x)(u) = CfV{x)(-u) We also observe the following relations between tangent cones to epigraphs and epigraphs of epiderivatives.· Proposition 6.2.4 Let V : X i-> RU{±oo} be a nontrivial extended function and χ belong to its domain. 'Used also with the notation V°(x, u) by Clarke and VT (x, u) by Rockafellar.
6.2. Other Epiderivatives 239 Then the tangent cones to the epigraph ofV at (x, V(x)) are the epigraphs of the corresponding epiderivatives ofV at x: Sp{D\v{x)) =Ί*εήν){χ,ν{χ)) к Sp{C}V{x)) = CSp{v){x,V{x)) and, when χ G К С Χ, D\0>k)(x) = Ψτ*κ(χ) & βΐ(Ψκ)(χ) = ФсК{х) The relations between tangent cones and epiderivatives of the distance functions are given in: Proposition 6.2.5 Let К be a subset of a normed space X, у G X апаИк(у) be the set of projections of у onto K. Then ifUfciy) φ 0, i) D\dK{y){v) < d(v,T\{IiK{y))) (6.3) ii) C}dK{y){v) < d(v,CK(nK(y))) Proof — The proof of the first statement is analogous to the proof of Proposition 6.1.8. For proving the second inequality, we take ζ G Ик{у) and w G Ck{z)· We observe that when у <£ K, VzeUK(y), У x€K, \\z-x\\ < 2\\y-x\\ Hence s^Ph<a, \\ν-χ\\<β(άκ(ν + hv) - dK{y))/h < sup^<Qi \\ζ-χ\\<2β dK{z + hw)/h + ||v - w\\ so that for every w G Ck{z), С-\ак{у){у) < ||υ—го||. Since ζ G Πκ(?/) and w G С к (ζ) were chosen arbitrarily, our proof ensues. Π The adjacent derivatives enjoy Proposition 6.2.6 Let V : X ^> Ru{±oo} be an extended function continuous on a neighborhood of χ G Int(Dom(V)). The values of the adjacent derivative at χ are convex. Furthermore, for any и G T>om{D\V{x))f\-Dom{D\v{x)) we have DbV(x)(u) = [D\V{x){u),D\V{x){u)\
240 6 - Epiderivatives of Extended Functions Proof — As for the contingent case, we can check that for any и e Oom(D\V{x)) Π Oom(D\V{x)), we have {D\v(x)(u),D\v{x){u)} С DbV{x){u) С [D\V{x)(u),D\v{x){u)} Let us take now two elements vi and V2 in D^V(x)(u). Pick any λ € [0,1]. We shall show that (1 - λ)υι + λυ2 belongs to DbV{x){u) and thus, that D^V(x)(u) is an interval. By definition of the adjacent derivative, we know that there exist sequences u\h and и^н converging to и when h —> 0+ such that У(х + hulh) - V(x) V(x + hu2h) - V{x) /ι h Since У is continuous on a neighborhood of x, for h small enough, it maps the interval χ + h[uih,U2h] to a connected subset, i.e., an interval which contains V(x + huih) and V(x + hu^h)· Then there exists ги^ € [ϊχι^, U2h\ such that (l-A)V(s + ftui/l) + AF(s + u2/l) = V(x + hwh) Hence Wh converges to и and hm —^ ^-^ = (1 — λ)υι + Лг/г This ends the proof. D 6.2.2 Other Convex Epiderivatives Convexity of the circatangent epiderivative counts among its attractive features. But, when the function is not episleek, there are other positively homogeneous convex lower semicontinuous functions lying between the contingent epiderivative and the circatangent epiderivative, among which one can choose whenever convexity is a prerequisite to apply duality arguments. Definition 6.2.7 Let 1A;IhRU {^°°} be α nontrivial extended function and χ belong to its domain. We shall say that a lower semicontinuous convex positively homogeneous function 6V{x) :X ^ RU{±oo}
g.2. Other Epiderivatives 241 satisfying D^V{x){u) < 6V{x){u) < C^V{x){u) (6.4) is a convex epiderivative. It is said to be minimal if there is no other strictly smaller convex epiderivative. All these convex epiderivatives coincide when V is episleek. Convex epiderivatives exist whenever V is contingently epidiffer- entiable at χ and do not exist if the circatangent epiderivative at χ takes the value —oo (or, equivalently, CfV(x)(0) < 0.) The epigraphs of the convex epiderivatives are the closed convex cones lying between the Clarke tangent cone and the contingent cone to the epigraph of У at ж. Among the possible choices of such convex epiderivatives, we can single out the convex epiderivative Όψν{χ) whose epigraph is the convex kernel of the epigraph of the contingent epiderivative. It satisfies Vuel, D^V{x){u) < D^°V{x){u) = suPveOom{DiV{a:)№V(xKu + v)-DlV(x)(v)) ^ fyV(x)(u) Denote by Vy(x) the family of minimal convex epiderivatives 6V(x) of V at x. Since the convex kernel of the contingent cone is the intersection of its maximal closed convex subcones (by Proposition 4.5.2), we deduce that D^V{x) is the supremum of the minimal convex epiderivatives: VueX, DfV{x)(u) = sup oY(x)(u) 8V(x)eVv(x) Also, the convex kernel of a cone Ρ being equal to the Clarke tangent cone to Ρ at the origin (Proposition 4.5.3), we obtain the formula VueX, D^V(x)(u) = CT(DTy(^))(0)(u) Remark — Paratingent epiderivatives We can also define the paratingent epiderivative P^V(x) of an extended function У:1н RU {±oo} at χ € Dom(y) by saying that the epigraph of the paratingent epiderivative is the paratingent cone -P-pD,7?.. {x>V{x)).
242 6 - Epiderivatives of Extended Functions One can then check that P^V(x)(v) = hminf —i r1 — h-*0+,y-*x,V(y)-+V(x),u'-*u h We deduce from Choquet's Theorem 4.5.7 the following result due to Shi Shuzhong: Theorem 6.2.8 Assume that V : X н-> R is continuous. Then there exists a residual R on which the contingent and paratingent epiderivatives do coincide. Proof — Since Graph(V) is closed, Choquet's Theorem implies that there exists a residual 1Z of the graph of V on which the contingent and paratingent cones coincide. Let R denote the projection Ώ.1Ζ of the residual 11 С Graph(y) С Χ χ R to X. The continuity of V implies that the restriction of the projector Π to the graph of V is bicontinuous, so that R is a residual of the domain of V. D 6.3 Epidifferential Calculus Is the epiderivative of the sum of two functions equal to the epi- derivative of the sum? We answer here this question and study as well the epiderivative of the composition product. Theorem 6.3.1 Let us consider two finite dimensional vector-spaces X and Υ, a continuous single-valued map f : X >—» Υ and two extended lower semicontinuous functions V and W from X and Υ to R U {+oo} respectively. Let xq belong to the domain of the function U := V + Wof We assume that f is continuously differentiable around xq, that V andW are contingently epidifferentiable at xq and f(xo) respectively and that the following transversality condition: Dom(CTW(/(s0))) - /'Ы (Dom(CTF(s0))) = Υ holds true. Then the epiderivatives ofU satisfy the estimates: ( Ϊ) D}U{xQ){u) < D\V(x0)(u) + DiW(f(xoW(x0)u) ii) D\U(x0)(u) = D\V(xQ)(u) + D\W(f(x0))(f'(xo)u) { Hi) CTl7(so)(«) < CiV(x0)(u) + C7TW(/(so))(/'(so)«)
q.3. Epidifferential Calculus 243 In particular, if К is a closed subset of X and if xq G К П Dom(V) satisfies Dom(CTT/(z0)) - CK(xQ) = X then, if V is contingently epidifferentiable at xq, ■ i) Vu G TbK(xQ), Di(V\K)(xo)(u) < DiV(x0)(u) < ii) Vug 2k(so), D\{V\k){xq){u) = D\V{x0){u) . Hi) Vug Ck{xo), Ci(V\k)(xo)(u) < CiV{x0){u) Proof — We shall prove the formula only in the case of cir- catangent epiderivatives. If we set ' К := Sp(V) x Sp{W) xR С IxRxFxRxR < G(x,a,y,b,c) := (f(x)-y,a + b-c) H(x,a,y,b,c) := (x,c) we can write Sp{U) = H{K П (Г1 (0,0)) We shall use Theorem 4.3.3 to estimate tangent cones to КпС_1(0,0). We first observe that the assumptions of our theorem imply the corresponding transversality conditions of Theorem 4.3.3. Let us set z0 = (xo,V(xo)J(xo),W(f(x0)),U(x0)) Then we deduce that CK(zo)nG>(z0)-1(0,0) с CKnG- Чо^о) It remains to show that this inclusion implies the desired inequality. Fix и G Όοτα(ΟΊν{χ0)) such that f'{x0)u G Dom(CTW(/(z0))) and define A = CiV(x0)(u), μ = CiW(f(xQ))(f'(xo)u)
244 6 - Epiderivatives of Extended Functions Hence (η,\,ί'(χο)η,μ,λ + μ) € СкЫ ^G'(zq)-1 (0,0) and thus, it belongs to the Clarke tangent cone to К П G_1(0,0) at z0. This means that for any sequence hn > 0 converging to 0 and any sequence zn ■= {хп,ап,Уп,Ьп,Сп) € АГП(?-1(0,0) converging to zq, there exist elements wn := (un, \n, υη, μη, un) converging to (u, λ, f'(xo)u, μ, λ + μ) satisfying Vn>0, zn + hnwn e finG_1(0,0) Since ζπ and ζπ + /ιπ^π belong to G~1(0,0), we infer that j[xn) = Уп> J(xn "i" h>nun) = Уп "Ь 11"пУп1 Cn = ап "Ь 0π, ί'π = λπ +'μη and using that ζπ + /ιηιοη belongs to if, we deduce that У(жп + hnun) -an W(yn + /ιπυπ) - 6n 7 s An к s μ-η. tin hn Consequently, U(xn + hnun) - Cn 7 S ^Π /in Since (un, vn) converges to (u, λ + μ), we finally obtain CiU(x0)(u) < \ + μ = CiV(x0)(u) + CiW(f(xoW'(xo)u) □ Remark — If Χ, Υ are Banach spaces, the conclusions remain true when we replace the transversality assumption by the following stability assumption: there exist constants с > 0, a £ [0,1[ and η > 0 such that, for all n, г) \fx G Dom(y)nB(a;o,,7), Vt/ ё Dom(W) Π Β{/{χ0),η) Βγ С /'(го) (Dom(2^V(aO) П сВх\ - Dom(DTW(y)) + aBy й) raP1IeDom(B}v(*))li?TV(aO(u)l/llull ^ с "Ό suPueDomc/wO/)) Ι^Τ^ΜΜΙ/ΝΙ ^ с (6.5)
6.3. Epidifferential Calculus 245 For that purpose, we have to check that the second transversality assumption of Theorem 4.3.3 is satisfied, i.e., that there exists a constant d > 0 such that, for all n, for all (x,a,y,b,c) € К close to zq = {xa,V(xQ),f(x0),W{f{x0)),U{xQ)) for all (z, A) G Υ x R, there exist (ω, μ, υ, ι/, δ) € Τκ(χ, α, y, b, c) and e such that !i) ζ = f'(x)u-v + e & A = μ + ν - δ и) ||е|| < а(И + |λ|) & И + ||V|| + |μ| + \V\ + \δ\ < c'())z\) + |λ|) Assumptions (6.5) imply right away that there exist и € Oom(D^V(x)), ν £ Oom(D^W(y)) and e satisfying (г) ζ = f'{x)u - ν + e, ii) ||e|| < (φ|| & ||U|| < φ|| & N < {l + a + c\\f'(x)\\)\\z\\ Let us now take μ := с||и||, г/ := с)Н|, 5 := μ + ι/-λ We deduce from (6.5)m) that (u, μ) belongs to £p(Db~V(x)), that (v, u) belongs to Ep(D^W(y)) and that D\V{x){u) + D^W{y){v) < с(И + И|) = μ + ν = λ + δ Consequently 1*1 < (|A| + c(|N + ||«||)) < c'(N| + |A|) D We now provide formulas for computing the epiderivative of the supremum of a finite number of extended functions. Let us consider a finite family of functions V- :X ι-» RU{±oo}, (i G/) with which we associate the function U defined by U(x) := max.Vi{x) We set I{x) :={»€/ | Ц(х) = U{x)}
246 6 - Epiderivatives of Extended Functions Proposition 6.3.2 Consider α finite dimensional vector-space X_-, and η extended functions VI : X >—» R U {+00}. If the following^ transversality assumption at xq € Dom(i7) holds true η V щ e Χ, Π (Вот(СтИ(:со)) - щ) φ 0 i=l then η У и e Π DomCDlfViCxo)), £>bl7(so)(«) = max 1>Ы(а:о)(и) ί=ι ' ie/(x0) and η У и € Π Dom(CTK-(^o)), CU{xq){u) < max CTVi(so)(«) i=i ie/(xo) Proof— Since the dimension of X is finite and since the epigraph of U is the intersection of the epigraphs of the η functions Vi, we can use Corollary 4.3.6, stating that if for all pairs (щ, \i), η f](c£p{Vi)(x0,U(xo))-(ui,\i)) φ 0 г'=1 then Т*£р{и)(хо,иЫ) = (\Ttp{Vi)(xo,U(xo)) The former property follows immediately from the assumption η f|(Dom(CTVi(so))-«i) φ 0 i=l The left-hand side of the latter formula is the epigraph of the adjacent epiderivative of U at xq. Since Tb£p{V{)(xoMxo)) D T^om{Vi)(x0)x-R whenever Vi(xo) < U(xq), i.e., whenever i £ I{xo), we deduce that η Vu€ i]Oom(D\Vi(x0)), DbU(x0)(u) < max D\Vi(x0)(u)
6.3. Epidifferential Calculus 247 The proof of the second statement is analogous. D We now complete formulas for the epiderivatives of marginal functions: Consider two normed vector spaces X and Υ and an extended function U :X xY ь-* Ru{±oo} with which we associate the marginal function V : X t-^ILU {+00} defined by V(x) := inf U(x,y) Proposition 6.3.3 Let us consider two normed vector spaces X and Υ, an extended function [/;Xx7hRU {i°°} a^d its marginal function V. Let xq G Dom(V) and suppose that there exists yo € Υ which achieves the minimum of U(xq, ·) onY: V{x0) = U(x0,yo) If U is pseudo-convex at (xo,yo), then У и £ X, О^У(хо){и) = liminf ( inf D-[U{xQ,yQ){ul,ν) u'—*u \v£Y ' Proof — Let π denote the projection from XxYxRto.X'xR. We recall that: kEP{U) С Sp(V) С ttSp{U) We deduce these statements from any criterion implying that the contingent cones to the images are the closures of the images of the contingent cones. This is the case of the epigraph of a pseudo-convex function, which is pseudo-convex. Then, equalities {тгЕр(0^и(х0,Уо)) = πΤερψ)(χα, yo,U{x0,y0)) = Τπ{ερ{υ))(χ0,ν(χ0)) = TSp{v)(x0,V{xo)) = ερ(ΌΊν(χ0)) can be easily translated into the desired result. □
248 6- Epiderivatives of Extended Functions- 6.4 Generalized Gradient 6.4.1 Subdifferentials arid Generalized Gradients We devote this section to dual concepts of epiderivatives of extended functions. When a function V is differentiable at x, its gradient V'(x), being a continuous linear functional, is therefore an element V'(x) € X* of the dual of X: Vv€X, <V'(x),v>= DfV(x)(v) When V is no longer different iable, but contingently epidifferen- tiable, we can still introduce subgradients of У at ж, which are those continuous linear functionals ρ € -X"* satisfying V«el, <p,v > < DfV(x){v) which constitute the (possibly empty) closed convex subset dQV(x) := {p£X*\Vv£X, <ρ,υ> < ΌΊν(χ)(υ)} Definition 6.4.1 (Subdifferential) Let V : X ■-» R U {±00} be a nontrivial extended function and χ belong to its domain. Assume that V is contingently epidifferentiable at x. The subset d0V{x) С X* is called the subdifferential ofV at χ (or contingent generalized gradient) and its elements are called the subgradients. We observe that when the contingent epiderivative D^V(x)(-) is convex, it is the support function of the subdifferential doV(x). When it is not convex, it is convenient to regard any convex epiderivative 6V(x)(·) (see Definition 6.2.7) as the support function Vv€X, 6V{x){v) = a(dsV(x),v) of the closed convex subset dsV(x) defined by d6V{x) := {p£X*\ Vv€X, <p,v>< 6V{x)(v)} (6.6)
6.4. Generalized Gradient 249 Definition 6.4.2 Let ViIhRu {±°o} be α nontrivial extended function and χ belong to its domain. Assume that V is contingently epidifferentiable at x. The (Clarke) generalized gradient of V at χ is defined by dV{x) := {p£X*\ V«el, <P,u>< СтУ(ж)(и)} More generally, we say that the closed convex subset dsV(x) associated to a convex epiderivative 6V(x) through formula (6.6) is the δ-generalized gradient of У at x. Naturally, we observe that d0V(x) С d6V{x) с dV{x) (6.7) Another way to state that ρ G doV(x) is a subgradient is to say that the pair (p, — 1) belongs to the polar cone to the contingent cone to the epigraph of V at (x, V(x)): (p,-l) € (T£p{v)(x,V(x))Y In the same way, to state that ρ € dV{x) amounts to saying that the pair (p, — 1) belongs to the (Clarke) normal cone N£p(V)(x, V(x)) to the epigraph of V at (x, V(x)). •Remark — Recall that Oy{x) denotes the family of minimal convex epiderivatives 6V(x) of У at i. The 5-generalized gradients dsV(x) associated with minimal convex epiderivatives 6V(x) axe the minimal δ- generalized gradients containing the subdifferential daV(x). If we denote by 9оо^(а;) the generalized gradient associated with the convex kernel of the contingent epiderivative, which is the supremum of the minimal convex epiderivatives, we obtain the formula dQV(x) с dooVix) = со (J dsV(x) ) с dV(x) Naturally, when V is Prechet differentiable at x, then ΌΊν{χ){ν) =<V'{x),v> < 6V(x)(v) so that the subdifferential 8qV{x) is reduced to the gradient V'{x).
250 6 - Epiderivatives of Extended Functions When V is continuously differentiable at x, the circatangent epi- derivative coincides with the directional derivative < V'(x),· >, so that the «^-generalized gradients are reduced to the only gradient: dsV(x) = {V'(x)} when V is continuously differentiable at χ If У is continuously differentiable around a point χ G K, then the generalized gradient of the restriction is the sum of the gradient and the normal cone: d{V\K){x) = V'(x) + NK(x) We also note that the generalized gradient of the indicator of a subset is the normal cone: θφκ(χ) = NK(x) The Fermat and Ekeland rules can be formulated in a ^-generalized gradient way: Theorem 6.4.3 (Fermat Rule) Let V : X i-> RU{+oo} be a non- trivial extended function defined on a normed space and χ £ Dom(V) be a local minimizer of V on X. Then it is a solution to the inclusions: 0 e d0V{x) С d5V{x) С dV(x) for any convex epiderivative 6V(x) ofV at x. Theorem 6.4.4 (Ekeland Rule) Consider a Banach space X, a nontrivial lower semicontinuous bounded from below function 1^:IhR+u {+oo} and let xq G Dom(V) be a given point of its domain. Assume that V is contingently epidifferentiable and let 6V(x) denote any convex epiderivative of V at χ. Then, for any ε > 0, there exists a solution χε £ Dom(V) to: i) V(xe) +ε\\χε - x0\\ < V(xq) (6.8) ii) 0 € θδν{χε)+εΒ* С dV{xe)+£B*
6,4. Generalized Gradient 251 Proof— Indeed, Theorem 6.1.12 implies the existence of χε € X satisfying (6.8) г) ала Vv€X, 0 < DTV(se)(t/)+e||t/|| Therefore, for any convex epiderivative 6V(xe), we deduce that 0 < 6V(xe)(v)+e\\v\\ = σ{δδν{χε),ν) + σ{εΒ*,ν) = σ(θδν(χε) + eS*,υ) D 6.4.2 Limits of Subdifferentials and Gradients Not only the subdifferential doV(x) is contained in the Clarke generalized gradient dV{x), but, in a finite dimensional vector-space, the upper limit of subdifferentials doV(xn) when also a subset of this generalized gradient: Theorem 6.4.5 Let X be α finite dimensional vector-space and ^:IhRu {+oo} an extended lower semicontinuous function contingently epidifferen- tiable in a neighborhood of χ £ Dom(V). Then Limsup^,^,))^^)) d0V{x') с dV{x) Proof — Indeed, let xn be any subsequence converging to χ such that V(xn) converges to V(x) and ρ be the limit of a subsequence of subgradients pn G doV(xn). Since (Pn,-1) € {Tsp(v)(xn,V{xn))) we deduce from Theorem 4.4.3 that [p, —1) belongs to the (Clarke) normal cone to the epigraph of V at (x, V(x)). This means that ρ belongs to the generalized gradient dV{x). □ In Hubert spaces and some Banach spaces, one can prove that the limits of the gradients at points where V is differentiable belong to the generalized gradient: . _
252 6 - Epiderivatives of Extended Functions Theorem 6.4.6 Consider α uniformly smooth Banach space X such that the norm of X* is Frechet differentiable off the origin and an extended lower semicontinuous function V : Χ \—► RU {+00} contingently epidifferentiable on. a neighborhood of χ £ Dom(V). Let a sequence of elements {xn)n>i converging to χ be such that V is Frechet differentiable at xn and V(xn) converges to V{x). Then σ - Iimsupn_too{V'/(sn)} С dV{x) Proof— We claim that the pair (V'(xn), —1) belongs to the polar cone to the weak contingent cone Τξ ,vJxn, V(xn)), defined in Section 4.1.5. by (u, v) € T£ ,Vs{xn, V(xn)) <=$■ (u, v) is a weak cluster point of < (uh,Vh) € (Sp(V) - (xn, V(xn))) /h when ft-»0+ Indeed, fix (u,v) € T^p,vJxn,V{xn)). Since V is Frechet differentiable at xn, V{xn + huh) = V(xn) +h< V\xn),uh > +he(h) < V(xn) + hvh where e(h) converges to zero with h. Since continuous linear func- tionals are weakly continuous, we deduce that <V'(xn),u>-v = ({V'{xn),-l),(u,v)) < 0 Now, assume that ρ is the weak-* limit of a subsequence of the gradients V'(xn). Then the pair (p, — 1) belongs to the sequentially weak upper limit of the polar cones to the weak contingent cones: (p,-l) € ^-Limsupn_too(2fp(vj(snjF(sn))) By Theorem 4.4.4, it is contained in the normal cone N£p^(x, V(x)). This means that ρ belongs to the generalized gradient dV(x). □ 6.4.3 Local Subdifferentials and Super differentials In finite dimensional vector-spaces, the subdifferential coincides with the local subdifferential d-V{x) defined in the following way:
6.4. Generalized Gradient 253 Definition 6.4.7 Let X be α normed space, V : X i-> R U {±00} and xq € Dom(V). We shall say that the subset d-V(xa) of elements ρ G X* satisfying ^^ΥΜζΙψζ^φ^ι^^ > о (6,9) z-zo If-zoII is the local subdifferential4 of V at жо· In a symmetric way, the local superdifferential d+V(xo) ofV at xq is defined by d+V(x0) := -d-(-V)(x0) It is the subset of elements ρ € X* satisfying V{x)- V{xq)~ <p,x-x0> limsup—- r. -. < 0 1—no IF -EoH Clearly super- and subdifferential are closed (possibly empty) convex sets. When V is Frechet differentiable, they coincide with the gradient of V. Proposition 6.4.8 Let X be a normed space, V:IhRU {±00} be a nontrivial extended function and xq G Dom(V). Then the local subdifferential d-V(x) is contained in the subdifferential doV(x). These two subsets coincide when the dimension of X is finite. Proof— Let us take ρ € d~V(xo) and υ € X. There exist sequences of elements hn > 0 and vn £ X converging to 0 and υ respectively such that (V(xo + hnvn) — V(xo))/hn converges to D^V(xq)(v). Therefore ' D^V(x0)(v)- <p,v> < = Нтп_>00(У(жо + hnvn) ~ V{xq) -hn<p, vn >)/hn k > \\v\\]iminix->X0(V(x)-V(xQ)-<p,x-x0>)/\\x-x0\\ > 0 4This concept, used by Crandall & P.-L. Lions for defining viscosity solutions to Hamilton-Jacobi equations, has been called "subdifferential." We add here the adjective "local" to avoid confusion with the concept of subdifferential (or contingent generalized gradient), although these two concepts are equivalent in finite dimensional vector-spaces.
254 6 - Epidenvatives of Extended Functions so that < ρ,υ >< DjV(xo)(v) for every υ € X. . Conversely, let ρ satisfy < p,v >< D^V(xq)(v) for every υ G X and assume that X is finite dimensional. There exists a sequence of elements xn converging to xq such that {]iminfx->XQ(V(x) - V(xq)- <p,x- xq >)/\\x - xq\\ = linin-foo(V(a;n) - V(xq)- <p,xn- xo >)/\\xn - xo\\ Taking a subsequence if needed, we may assume that the sequence of elements vn :— (xn — xo)/\xn — xq\\ of the unit sphere (which is compact) converges to an element v. Then for hn := ||a;n — xq\\, we have Нтп_>00(У(а;п) - V(x0)- <p,xn- x0 >)/||a:n - x0\\ ' — Итп_>оо(У(ж0 + hnvn) - V(xq) -hn< p,vn >)/hn > D}V{xQ){v) - <ρ,υ>> Ο so that ρ satisfies property (6.9.) □ 6.4.4 Remarks The table of formulas on support functions allows one to translate the properties of the circatangent epiderivatives into corresponding properties of the generalized gradient and vice-versa. For instance, Proposition 6.2.3 can be restated in the following form. Proposition 6.4.9 When V: Jf hRU {^°°} zs locally Lipschitz on the interior of its domain, then the generalized gradient satisfies: • the function (x,u) € Χ χ Int(Dom(V)) (-► a(dV(x),u) is upper semicontinuous, so that dV(-) is upper hemicontinuous J «Vie Int(Dom(V)), dV(x) is nonempty bounded closed convex ^.VxS Int(Dom(l/)), dV{x) = -d{-V{x))
6.5. Convex functions 255 Remark — When a function V is continuously differentiable at x, its gradient V'(x) being a continuous linear functional, it is both an element V'(x) e X* of the dual and the image V'(a;)*(+1) of +1 by the transpose V'(x)* of V'(x), a linear operator from R to X*. When V is no longer continuously differentiable, the generalized gradient remains intimately connected with the transpose of the circatan- gent derivative CV^(x,V(x)), (which, shall we recall, is the codifferential CVf(x, V{x)Y) of the set-valued map VT at (x, V{x)). Namely: Proposition 6.4.10 Let V : X i-> R и {±00} and χ € Dom(V). The generalized gradient is the value at 1 of the codifferential o/Vj at (x, V{x)). Proof — Indeed, the codifferential is a closed convex process from R to X* which, being positively homogeneous, needs to be defined only at the points —1,0 and +1. We obtain from the definitions i) CVi(x,V(x))*(-l) = 0 ii) CVT(a;,V(aO)*(0) = (Dom(CTV)(a;))~ =: d°°V(x) (6.10) Hi) CVT(a;,V(aO)*(+l) = dV{x) □ 6.5 Convex Functions Convex functions enjoy further properties. We already mentioned that an extended function is convex (respectively lower semicontin- uous) if and only if its epigraph is convex (respectively closed.) A convex function defined on a finite dimensional vector-space is locally Lipschitz on the interior of its domain and a convex lower semicon- tinuous function on a Banach space is also locally Lipschitz on the interior of its domain. Proposition 6.5.1 When the function V:IhRU {+00} is convex, the contingent, adjacent and circatangent epiderivatives coincide and are equal to D}V{x){u) = ШЫ(Ы^^±^Щ 1 Ч М ' u'-m \h>0 h J Furthermore, when χ belongs to the interior of the domain of V, Vii £ I, DiV(x)(u) = mf — j- K-± is finite
256 6 - Epiderivatives of Extended Functions When V is convex, the generalized gradient coincides with the subdif- ferential introduced by Moreau and Rockafellar for convex functions: dV{x) = {p € X*\ V у £ X, <p,y-x> < V(y) - V(x)} There is more to that: lower semicontinuous convex functions enjoy duality properties. In the same way that we associated with cones their polar cones, with closed convex processes their transposes, we can, following Fenchel, associate with lower semicontinuous convex functions conjugate functions for the same reasons, and with the same success. Definition 6.5.2 Let V : X —> RU{+oo} be any nontrivial extended function defined on a Banach space X. We associate with it its conjugate function V* : X* -»RU {+°°} defined on the dual of X by Vp € X*, V*{p) := sup{<p,x>-V{x)) xeX Its biconjugate V** :IhRU {ioo} is defined by V**(x) := sup {<p,x> ~V*(p)) pe x* For instance, the conjugate function of the indicator ψ χ of a subset К is the support function αχ. We deduce from the definition the following convenient inequality Vxel, ρ € X\ <p,x>< V(x) + V*(p) known as Fenchel's Inequality. The epigraphs of the conjugate and biconjugate functions being closed convex subsets, the conjugate function is lower semicontinuous and convex and so is its biconjugate when it never takes the value -oo. We observe that Vx £ X, V**(x) < V{x) If equality holds, then V is convex and lower semicontinuous. The converse statement, a consequence of the Hahn-Banach Separation Theorem, is the first basic theorem of convex analysis:
6.5. Convex functions 257 Theorem 6.5.3 A nontrivial extended function V : X —> RU{+oo} is convex and lower semicontinuous if and only if it coincides with its biconjugate. In this case, the conjugate function V* is поп trivial. So, the Fenchel correspondence associating with any function V its conjugate V* is a one to one correspondence between the sets of nontrivial lower semicontinuous convex functions defined on X and its dual X*. This fact is at the root of duality theory in convex optimization. We deduce at once the following characterization of the subdif- ferential: Proposition 6.5.4 Let V : X —> RU{+oo} be a nontrivial extended convex function defined on a Banach space X. Then ρ € dV{x) <=> <p,x>= V{x) + V*(p) If moreover the function V is lower semicontinuous, then the inverse of the subdifferential dV(-) is the subdifferential dV*(·) of the conjugate function: ρ € dV{x) Φ=» χ € dV*(p) Since - V*(0) = inf-cgx V(x), the Fermat rule becomes: Theorem 6.5.5 Let K:1h»RU {+°°} be a nontrivial convex extended function defined on a normed space X. Then χ £ X achieves the minimum ofV on X if and only if 0 € dV(x) or, equivalents, V и € X, 0 < D}V^c){u) If moreover the function V is lower semicontinuous, then dV*(0) is the set of minimizers of V. The second fundamental theorem of convex analysis, also a consequence of the Separation Theorem, provides a criterion (weak constraint qualification) for computing the conjugate function of the sum and the composition product of convex functions: Theorem 6.5.6 (Fenchel) Let X and Υ be reflexive Banach spaces, A € C(X,Y) a continuous linear operator and V:I-»RU {+00}
258 6 - Epiderivatives of Extended Functions and W : Υ —> R U {+00} be convex lower semicontinuous functions. Let U :=V + WoA. We posit the weak constraint qualification assumption5: - - 0 € lnt(A(Dom(V)) - Oom(W)) Then Dom([T) = Dom(l/*) + A*Oom(W*) and, for every ρ G Dom(i7*)J there exists a solution q G Y* to the minimization problem U*{p) = {V*(p - A*q) + W*{q) = inf {V*{p - A*q) + W*(q)) When ρ = 0, we deduce from the fact that t/*(0) = -inf I7(x) and from the above theorem the duality relation: for any χ € X, q € inf (F(s) + W(Ax)) + inf (V^f-A*?) + W*{q)) = 0 We observe that the subdifferential map dV(·) of a nontrivial lower semicontinuous convex function is obviously monotone. When X is a Hubert space, we can prove that dV(-) is maximal monotone. This is a consequence of Minty's Theorem 3.5.8 and the following Moreau Theorem: Theorem 6.5.7 (Moreau) Let X be a Hilbert space identified with its dual and V :X ^ RU{+oo} be a nontrivial lower semicontinuous convex function. For every λ > 0, define the Moreau-Yosida approximation V\ by Vx € X, Vx{x) := mf (v(y) + ^||s - y||2) < V(x) 5We observe that U is non trivial if and only if 0 e A{Oom{V)) - Dom(W)· In this case, inequality U*[p) < inf,ey* (V*(p — A*q) + W*(q)) is always true.
6.6. Higher Order Epiderivatives 259 This minimization problem has a unique solution J\x: Vx(x) = V(Jxx) + ±\\x - Jxxf which is the resolvent of the maximal monotone map dV(-). Furthermore, V\ is convex, continuously differentiable, and converges pointwise to V on Dom(V) when λ —> 0+. The Yosida approximation defined by A\x := (x- J\x)/\, is both equal to the gradient ofV\ at χ and to a subgradient of V at J\x: Axx = V{{x) к Axx € dV(Jxx) Finally, this implies that the domain of dV(-) is dense in Dom(V). 6.6 Higher Order Epiderivatives Let us consider a normed space X and a nontrivial extended function V ;X h+ Ru{+oo} We can define m^-order epiderivatives by taking mt/l-order derivatives of the set-valued map Vj. We observe that vm belongs to jD(m)VT(a;,2/,Ui,wi,... ,um-i,vm-\){um) if and only if vm > liminf^_>o+)li/ji_>lim -^ (V{x + hUi+... + hT^Um-i + hmu'm) -y-hVi-...- hr^Vm-l) Denote by D^'V(x,ui,... ,Uj) the infimum of the subset i?u)VT (a;, V{x), UuD^Vix)^),..., D^~l)V{x, uu ..., ty-Ofa)) Ы which is equal to D\jS>V(x, Ui,...,uj) := lim inffc_0+i u/_Uj. £ < (V{x + Atii + ... + ft^S-i + A^) - ΣίΖα tiD\l)V{x,Ul,...,u,))
260 - 6- Epidenvatives of Extended Functions By taking successively у = V{x);-V\ := D^V(x,Ui), etc., we can check recursively that for any j, ' 2?ϋ)ντ (x,V(x),u1,D]V(x){u1),...,D<f~1)V{x,ul,...,Uj-1)) (Uj) D\3)V{x, uu..., uj), +001 Definition 6.6.1 Let us consider α normed space X and a nontrivial extended function F:Xb»RU {+00}. Let χ € Dom(y) and U\ € Dom(£)jV(x)),..., Urn—ι e Dom(l)jm 1V(a;,wi,...,wm_2,·)) be given. Then D^V(x,ui,...,um) ■= liminf^_>o+,u'm-^Um t^ < _ (y{x + hu1 + ...hm-1um-1+hmu^)-YZ^hlD\l)V{x,u1,...,ui)) is called the mt,l-order contingent epiderivative of V at χ in the directions Ml, U2, ..., Um. We define in the same way т*к-отает adjacent and circatangent epi- derivatives D\{m)V{x,Ui,...,um) h C^m)V(x,uu...,um) using the corresponding m—th order derivatives of set-valued maps. For m = 2, we obtain D^V(x,uuu2) := limЫк-*о+!и'2->и2 {V(x + hui + h2u'2) - V{x) - hD^V{x){ul)) /h? We observe right away that £>}т)^(а;,«1,...,«т-1,·) = Ψτ™[Χί^ „_,)(·) and that, for every χ G К ) '^=> D(fn'dK(x,Ui,...,um-i:um) = 0 The mt/l-order contingent epiderivative enjoys the same kind of calculus as those of first-order.
6.6. Higher Order Epiderivatives 261 Theorem 6.6.2 Let us consider two finite dimensional vector-spaces X and Y, a continuous single-valued map f : X \-^ Y, and two nontrivial extended functions V and W from X and Υ to R U {+00} respectively. Let хй belong to the domain of the function U := V + W о f. We assume that f is twice continuously differentiable around xq, that V and W are contingently epidifferentiable at xQ and f(xQ) respectively and that the following transversality condition: Oom(CiW(f(x0))) - f'(x0) (Oom(CiV(xQ))) = Υ holds true. Then the second-order epiderivatives of U satisfy the estimates: Df)U{xu,Ui,u2) < Db}{2)V{xu,ui,u2) + D\2)W(f{xQ)J,{xQ)u1,f,(xQ)u2 + £/"(:βο)(«ι,«ι)) and Dbi{2)U(xa,u1,u2) = D^2)V{xu,uuu2) + Ό){2)]νϋ(χ0),Γ(χ0)^,Πχ0)^ + |/"(a;o)(«i,«i)) In particular, if К is a closed subset of X and ifxQ € K(lOom(V) satisfies Оот(СтУЫ) - CK{xQ) = X then, ifV is contingently epidifferentiable at xq, i) VueTf^ctti), D{2)(V\K)(x0,Ul,u2) = D{2)V(xQ,Ul,u2) [ ii) УиеТ%2)(х0,щ), d\{2){V\k){xu,uuu2) = D\(2)V{xQ,uuu2) Proof— The proof is exactly the same as the first-order Theorem 6.3.1, but we use Theorem 4.7.4 instead of Theorem 4.3.3. □ We can make more precise the Fermat rule by using mt/l-order contingent epiderivatives. Indeed, we saw that if V : X —> R U {+00} achieves its minimum at some xq G Dom(^), then for every u\ e X, we have DfV(xQ)(ui) > 0. We can now distinguish the directions u\ at which D^V{xq){u\) = 0. Then we see that for such directions, Di 'V{xq){u\,u2) > 0, and so on.
262 6- Epiderivatives of Extended Functions Proposition 6.6.3 Let us consider a~ normed space X and a nontrivial extended function V : X j—> R U {+00}. Assume that xq achieves the minimum of V. Then Vtti,...,Um-i such that V j — 1,.. . ,m — 1, D^ V(xq,u\, ... ,Uj) = 0, < we have VtimeX, 0 < D^m'V(xa,ui:... ,um-i,um) We can prove that second order conditions are sufficient in finite dimensional vector-spaces: Theorem 6.6.4 Consider a normed space X and a nontrivial extended function 1/:IhRU {+00}. Assume that xq e Dom(y) satisfies ( i) \/u € X, 0 < D^V{xa)(u) { (6.11) { ii) \/u^Q such that D^V{xu){u) = 0, D^]V{xu, щи) > О Then xq is a weak local minimizer of V in the sense that for every и in the unit sphere S, there exists eu > 0 such that, for any h £ [0, eu] and any e € euB, V{xQ) < V{xQ + hu + h2u + h2e) Proof — Assume the contrary: there exist и € S and sequences of hn > 0 and en ζ Χ converging to 0 such that V(xQ + hnu + h2nu + hlen) - V(xQ) < 0 By letting η —► oo, we deduce that D^V(xQ)(u) < 0. Together with assumption (6.11) г), this implies that D^V(xQ)(u) = 0. We then infer from the definition of the second-order contingent epiderivative that Df]V{xQ,u,u) < 0 which is a contradiction of assumption (6.11) ii). D Remark — A local minimizer is always a weak local one. Stronger conditions on the second order epiderivative imply stronger conditions. □
6.6. Higher Order Epiderivatives 263 6.6.1 Second Order Epiderivatives of Moreau-Yosida Approximations Let X be a Hilbert space identified with its dual and V :X h+ Ru{+oo} be a nontrivial lower semicontmuous convex function. We recall that it is the pointwise limit when A —> 0+ of C1 convex functions V\, the Moreau- Yosida approximations, defined by Ух eX, Vx(x) := mf (v(y) + ±\)x - yf^j This minimization problem has a unique solution 3\x: Vx(x) = V(Jxx)+±-\\x-Jxxf Furthermore, A\x := (x — J\x)/X is both equal to the gradient of V\ at χ and belongs to a subgradient of V at J\x: Αχχ = V{(x) & Αχχ e dV{Jxx) Finally, we associate with any ρ £ dV(x) the normal cone NgV^{j>) to the subdifferential of V at χ at some ρ € dV(x), equal to {v e X\ <p,v>= D-[V{x){v)} For instance, when V = ψκ is the indicator of a closed convex subset K, then Vx{x) := d2K{x)/2\ & Jxx = -κκ{χ) is the projection of best approximation of a; to Χ and ΝΝκ(πκχ)(χ-πκχ) ■= {ν € Τκ(πκχ)\ < χ - πκ(χ),υ > = 0} Theorem 6.6.5 Let X be a Hilbert space identified with its dual and V : Χ η-► R U {+oo} be a nontrivial lower semicontinuous convex function. Then, for any χ £ X, A > 0, for every щ, ui, V2 & X and Ы e NdV{Jx(x))(Ax{x)) we have О^Ух{х,иии2) < D^V{Jxx,v1,v2)+<Axx,U2-V2>+ — \\ui-vl\\2
264 6- Epiderivatives of Extended Functions In particular, by taking V = ψκ we obtain Corollary 6.6.6 Let К be α closed convex subset of α Hilbert space X. Then, for any U\, u2, v\ £ Тк{кк{х)) satisfying < χ — πκ{χ), ν\ > = 0 and V2 £Tj( (x,vi), inequality D2d2K(x,Ui,u2) < 2<x-nK(x),u2-v2> + )\щ - vi\\2 holds true. Proof — We observe that V\{x + hui + h2u2) < V{J\x + hvi + h2v2) + \\x - J\x + /i(wi - vi) + h2{u2 - υ2)\\2/2λ < = V{J\x + hvi + h2v2) + \\x - Jxx\\2/2X + h < {x - J\x)/X, щ - vi > +h2 (< (x - Jxx), u2 - v2 > /A + ))щ - Vl ||2/2A + e(h)) where e(h) converges to 0 with h. We now take into account that Αχ = (1-Λ)/λ & D^Vx(x)(Ul) =<Axx)Ul> and that D\V{J\x){vi) =< Αχχ,νι > whenever νλ € NdV(Jx{x))(Ax(x)) Therefore, we obtain ' (Vx(x + hu, + h2u2) - Vx(x) - hDiVx(x)(Ul)) /h2 - < (V{Jxx + hvi+h2v2)-V{Jxx)-hD]V{Jxx){vl))/h2 + <Axx,u2-v2>+\\u1-v1))2/2X + e(h) We then take the liminf when h —> 0+ in the above inequalities to obtain the formula. □
Chapter 7 Graphical & Epigraphical Convergence Introduction Here are some basic notions on graphical limits of single-valued and/or set-valued maps and epigraphical limits (or epilimits) of extended functions. It is one of the basic themes of this book to regard maps, and more generally, set-valued maps, not only as maps from one space to another, but as graphs, in order to characterize them in an intrinsic and symmetric way. This point of view, which goes back to the prehistory of analysis with Fermat and Descartes dealing with curves rather than functions, has been let aside for a long time. In particular, as far as limits of functions and maps are concerned, generations of mathematicians have been accustomed to deal with many concepts of convergence of functions, from pointwise to uniform, but all based on the fact that a map is a map, and not a graph. When dealing with limits of maps, either single-valued or set- valued, it is quite advantageous to replace pointwise convergence by graphical convergence: Instead of studying (more or less uniform) limits of images of maps, we consider the limits of their graphs1. Then the graphs of 1This point of view, regarding maps as graphs, has already been used in 265
266 7- Graphical L· Epigraphical Convergence the upper and lower graphical limits of a sequence of set-valued maps Fn are the upper and lower limits of the graphs of the FnS. One of the main reasons for doing so is to treat on the same footing a map and its inverse. This is quite important in approximation theory, where the problem is to derive pointwise convergence of the inverses from the pointwise convergence of the maps. Indeed, we already proved an extension of Lax's principle to the general case of inclusions (Theorem 5.4.2) stating that stability, convergence of the data and consistency imply the convergence of the solutions: Consistency of set-valued maps Fn at (хо,Уо) means simply that {x0> Уо) belongs to their graphical lower limit. Hence, a first problem is to investigate the connections between graphical and pointwise convergence. We then relate in Section 3 the concepts of graphical convergence of set-valued maps to the concepts of epigraphical limits of functions. These concepts have recently been successful in overcoming the failure of pointwise convergence in many problems of calculus of variations, optimization, stochastic programming2, etc. The use of this concept is mandatory whenever the order relation of the real line comes into play, as in optimization or Lyapunov stability for instance. In such cases, a function is replaced by the set-valued map obtained by adding to it the positive cone R+ (for minimization), whose graph is thus the epigraph of the function. Therefore, the convergence of the graphs of such set-valued maps is the convergence of the epigraphs of the associated functions. We present just a selection of issues on epiconvergence, among which are some formulas dealing with epilimits of sums and composition products of functions (Section 4) and conjugates of epilimits (Section 5.) We answer in Section 6 the problem of the graphical convergence of gradients of differentiable functions and generalized gradients of lower semicontinuous functions. In the last section, we study the "asymptotic derivatives" of a sequence of functions. Chapter 5 to build a differential calculus of set-valued maps based on "graphical derivatives." 2see for instance the book [30] and its bibliography.
7.L Graphical Limits 267 7.1 Graphical Limits 7.1.1 Definitions We shall use the concepts of upper and lower limits of sets to define graphical convergence of set-valued maps. Definition 7.1.1 (Graphical Convergence) Let us consider metric spaces Χ, Υ and a sequence of set-valued maps Fn : X ~> У. The set-valued maps ΙΛπτπ_>0οί1π and Lim n-»oo-Pn from X to Υ defined by (i) Graph (Lim8 п-юо-Рп) := LimsupTl_>00Graph(i;,Tl) ii) Graph(LimbTl_>oo.FTl) := LiminfTl_>ooGraph(FTl) are called the (graphical) upper and lower limits of the set-valued maps Fn respectively. If the graphical lower and upper limits coincide, it is called the graphical limit. To say that (x,y) belongs to the graphical lower limit of set-valued maps Fn amounts to saying that they are consistent at (x,y) (see Definition 5.4.1.) We provide a more explicit characterization of these graphical upper and lower limits, which follows immediately from the definitions: Proposition 7.1.2 Consider metric spaces Χ, Υ and a sequence of set-valued maps Fn : X ~> У. Then у € (UiJn^oaFn^ (i) if and only if у is the limit of a subsequence of elements yni G F{xni) where xn/ converges to x. Furthermore у € (Limbn_>00i1n)(a;) if and only if у is the limit of a sequence of elements yn € F(xn), where xn converges to x.
268 7 Graphical <fe Epigraphical Convergence The pointwise limit / of single-valued maps fn (when it exists) is a selection of the graphical upper limit of the fn. The latter is equal to f when fn remains in an equicontinuous subset: Indeed, in this case, any limit of elements (xn, fn(xn)')--bemg of the form (z,/(z)) belongs to the graph of /. We point out two useful formulas: Proposition 7.1.3 Let us consider a sequence of set-valued maps Fn:X^Y. Then (Lim^eJ^Xz) D n£>oLimsuPn->oo-FM£(z,e)) (Limbn_>00i1n)(z) С D£>0Liminfn->ooFn(B(x,e)) These formulas can be regarded as relating graphical convergence with some kind of "almost pointwise convergence." But can we compare the graphical convergence of Fn and the "pointwise convergence" of Fn, i.e., the upper and lower limits of the subsets Fn{x)l The following statement provides the easy answers. Proposition 7.1.4 Let us consider metric spaces Χ, Υ and a sequence of set-valued maps Fn '. Χ ~> Υ■ Then the relations {LuJn^ooFnj (z) = Limsupn_>00)In_>I Fn(xn) and, setting Kn '■= Oom(Fn), (Limbn_00.FnJ (z) D Limmfn-.oo,!,,-»^! Fn(xn) B(Fn(xn),e)) hold true. The missing equality holds true under more assumptions: Theorem 7.1.5 Let X and Υ be Banach spaces. Consider a sequence of closed set-valued maps Fn '■ X ~~>Y and its lower graphical limit Lim n->oo-Fn· Let ?/o € (Limbn_>ooFn)(z0)
7.1. Graphical Limits 269 and 'assume that there exist constants с > 0 and η > 0 such that f V(znjy„) € Graph(-Fn)nJ3((zo,2/o),?7), [ ||£>-Fn(zn,2/n)ll := snpueXMveDFn{Xniy7i){u)\\v\\/\\u\\ < с Then, for any sequence xn converging to xq, we have ?/o € Liminfn_>00Fn(a;n) Consequently, whenever this assumption is satisfied for every г/о € (Limbn_>00i;in)(:Eo)) then (Limbn_>00i;in)(2;o) = Liminfn_»Oo,xn_»l0i<1n(3:n) Proof — This is just Theorem 5.4.2 of Chapter 5 applied to the set-valued maps F~l with a = 0. □ 7.1.2 Graphical Convergence of Closed Convex Processes Let us consider a sequence of closed convex processes Fn from a Banach space X to a Banach space Υ and their transposes F£ : Y* ^ X*. Graphical'lower and upper convergences are exchanged by transposition: Proposition 7.1.6 Let Χ, Υ be Banach spaces and Fn : Χ ~> Υ be closed convex processes. We supply X* and Y* with the weak-* topologies. Then the graphical lower limit Lim n-*ooFn of the sequence (Fn)n is the transpose of the sequentially weak graphical upper limit of the transposes F£: Limb n^ooFn = [cr-Lia^n^ooF^j Proof— It follows from Proposition 1.1.8 applied to the closed convex cones Kn := Graph(Fn): The lower limit of the jFfn's is equal to the negative polar cone of the sequentially weak upper limit of the cones K~. This can be translated by saying that the graph of the graphical lower Umit is the negative polar of the graph of the sequentially weak graphical upper Umit of the transposes F*. Π
270 7 Graphical <fe Epigraphical Convergence 7.1.3 Monotone and Maximal Monotone Maps Let X be a Hubert space, the dual of which is identified with X and F : X ^л X be a monotone map. If the graphical lower limit Lim n-*oaFn of a sequence of monotone maps Fn is maximal monotone, then it is actually the graphical limit of this sequence: Proposition 7.1.7 (Convergence of Monotone Maps) Let X be a Hubert space and let its dual X* be supplied with the weak-* topology. Consider monotone set-valued maps Fn : X ~> X* and a maximal monotone map F:X^X*. If F is contained in the graphical lower limit Lim n-*ooFn °f the Fn 's, then F is actually the graphical limit of the Fn 's. Proof— We have to prove that the graphical upper limit Lim"n_>00.FTl of the set-valued maps Fn is contained in F. Let ρ belong to (Lim"n_>00FTl)(a;). Hence the pair (x, p) is a cluster point of a sequence of elements (xn,Pn) of the graph of Fn. Take now any pair (y, q) in the graph of F. Since Graph(F) С Graph (Liml'n_00.Fn) we know that there exists a sequence of elements (г/„, qn) of the graph of Fn converging to (г/, q). The monotonicity of the set-valued maps Fn implies the inequalities <Рп-Цп , xn-yn>> ϋ Going to the limit, we deduce that V (г/, q) G Graph(F), <p-q, x-y>> 0 Therefore ρ belongs to F(x) because of the maximality of the graph of F among monotone graphs. D 7.2 Convergence Theorems We present in this section two versions of a Convergence Theorem which plays a key role in the study of differential inclusions (see Chapter 10 below.) Let α(·) be a measurable strictly positive function from an open subset Ω С Rn to R+. We denote by L1(fi; У, a) the space of classes of measurable functions integrable for the measure a(u>)du>.
7.2. The Convergence Theorem 271 Theorem 7.2.1 (Convergence Theorem I) Let X be a topological vector space, Υ be a finite dimensional vector space and Fn be a sequence of nontrivial set-valued maps from X to Υ. We assume that for any χ € Oom(F), there exists a neighborhood V of χ such that \Jn>i Fn(V) is bounded. Let us consider measurable Junctions xm and ym from Ω to X and Υ respectively, satisfying: for almost all ω € Ω and for all neighborhoodU ofO in the product space X x Y, there exists Μ := Μ(ω,Κ) such that Vm > M, {хт{ш),ут{ш)) € Graph(.Fm)+W (7.1) If we assume that i) xm{·) converges almost everywhere to a function x(-) ' H) Ут(·) € Ll(£l;Y, a) and converges weakly in L1(Q;Y, a) to a function ?/(·) 6 L1(D,;Y,a) then for almost all ω € Ω, у (ω) € coF^(x(uj)). Proof — Let us recall that in a Banach space ^(Ω,',Υ,α) in our case), the closure of a convex subset (for the normed topology) coincides with its weak closure for the weakened topology3 σ (l\Q; Y, a), L°°(fi; Y\ a"1)) We apply this result: for every m, the function y(-) belongs to the weak closure of the convex hull co({yp(-)}p>m.) It coincides with the 3By definition of the weakened topology σ(Χ, Χ*), the continuous linear functional and the weakly continuous linear functional coincide. Therefore, the closed half-spaces and weakly closed half-spaces are the same. The Hahn-Banach Separation Theorem, which holds true in Hausdorff locally convex topological vector spaces, states that closed convex subsets are the intersections of the closed half-spaces containing them. Since the weakened topology is locally convex, we then deduce that closed convex subsets and weakly closed convex subsets do coincide. This result is known as Mazur's Theorem. When X is not reflexive, closed convex subsets of the dual supplied with the weak topology σ(Χ*, X) may be different from their closure for the weak topology σ(Χ\Χ**).
272 7 Graphical <fe Epigraphical Convergence (strong) closure of co({?/p(-)}p>m). Hence we can choose functions CO %(') := J2 атУр(·) € С0({?/р(-)}р>ш) p=m (where the coefficients a^ are positive or equal to 0 but for a finite number of them, and where ]C^Un am = 1) which converge strongly to ?/(·) in ^(Ω-,Υ,α). This implies that the sequence a(-)vm(·) con- verges strongly to the function a(-)y(·) in L1 (Ω; Υ"), since the operator of multiplication by α(·) is continuous from ^(Ω,-,Υ,α) to L1(fi;Y). Thus, there exists another subsequence (again denoted by) wm(·) such that4 for almost all ω € Ω, а(ш)ут(ш) converges to a(u))y(ω) Since the function α(·) is strictly positive, we deduce that for almost all ω € Ω, vm(u>) converges to у (ω) Let ω £ Ω, such that а;т(ш) converges to χ(ω) in X and υτη(ω) converges to у (ω) in Υ. By assumption (7.1), we know that for any neighborhood U := VxW 4Strong convergence of a sequence in Lebesgue spaces Lp (1 < ρ < +oo) implies that some subsequence converges almost everywhere. Let us consider indeed a sequence of functions /„ converging strongly to a function / in Lp. We can associate with it a subsequence fnK satisfying ||/»fc -/||lp < Tk\ ■■■ < nk < nk+1 < ... Therefore, the series of integrals °° Γ Σ НЛчИ-/И||5-с^ < +~ fc=i ^ is convergent. The Monotone Convergence Theorem implies that the series oo ΣΐΙ/»*(ω)-/(ω)||^ fc=i converges almost everywhere. For every ω where this series converges, we infer that the general term converges to 0.
7.2. The Convergence Theorem 273 of (0,0) eX xY, there exists N such that for all q> N, хд{ш) € χ(ω)+ν/2 к {хд{ш),уд{ш)) е Graph(Fg) +U/2 Hence Уд(ш) € Fq(x{u>) + V)+W Consequently, Vm > N, ут{ш) e col \J{Fq{x{a>) + V) + W)\ \q>n ) Then its limit у (ω) belongs to these subsets for all N and all neighborhoods of the origin V, W in X and Υ respectively, so that yMe Π co[{j{Fq{x{u) + V) + W)\ N>0, V, W \q>N ) By assumption, there exist neighborhoods of the origin V and W such that the subsets Fq(x(u) + V) + W are contained in a bounded subset. Hence 2/H€ Π cd[\J(Fq(x^) + V/N)+W/N)\ N>0 \q>N / Since the dimension of Υ is finite, Lemma 1.1.9 implies that у (ω) belongs to the closed convex hull of the upper limit of the subsets Fq(x(u) + V/N)+W/N i.e., the closed convex hull of F^(x(uj)). □ The "continuous version" of the convergence theorem can be stated for any Banach space in the following way: Theorem 7.2.2 (Convergence Theorem II) Let X be a topological vector space, Υ a Banach space and F be a nontrivial set-valued map from X to Υ. We assume that F is upper semicontinuous on its domain. - -
274 7 Graphical <fe Epigraphical Convergence Let us consider measurable functions xm(·) and ym{·) from Ω to X and Υ respectively, satisfying: for almost all ω € Ω and for all neighborhoods Ы of 0 in the product space X x Y, there exists Μ := Μ{ω,14) such that Vm > M, (im(w),ym(w)) € Graph(_F)+W If we assume that i) xm{·) converges almost everywhere to a function x(-) < и) Ут{·) € L1 (Ω; У, α) and converges weakly in L1 (Ω; У, a) to a function у € ^(Ω;!'', α) for almost all u> € Ω such that χ[ω) 6 Вот(^), ?/(α;) € ~cb~F(x(uj).) Proof — The proof is analogous to that of the previous theorem with Fn := F. With the same arguments, we arrive at the conclusion that for almost all ω Ε Ω such that χ (ω) € Dom(F), all neighborhoods V and W of the origin in X and Υ respectively, у (ω) € co(F(x(u) + V) + Щ Since F is upper semicontinuous at χ(ω), for every e > 0, there exists a neighborhood V such that F(x(w) + V) С F(x(u>)) + eJ5. Taking W = eB, we obtain у(ш) € co{F(x(u;)) + 2eB Hence, e > 0 being arbitrary, у(ш) belongs to coF(x(uj)). D 7.3 Epilimits 7.3.1 Definitions and Elementary Properties Let us consider a metric space X and a sequence of extended functions Vn: X н-> R U {±oo}
7.3. Epilimits 275 whose domains Dom(Fn) := {χ € X | Vn(x) φ ±οο} are not empty. Taking into account the order relation of R, as we did in Chapter 6, when we introduced epiderivatives, we associate with each extended function Vn two new set-valued maps Vnj and Vnj. defined in the following way: 0 VnT : < it) VnJ. : We see at once that Graph(VnT) = Ep{Vn) L· Graph(Vnx) = Hp{Vn) Using the concept of graphical upper and lower limits with these associated set-valued maps, we come up with the concepts of epi and hypo convergence, which are thus obtained by taking upper and lower limits of their epigraphs and hypographs. Definition 7.3.1 (Epilimits) Let us consider a metric space X and a sequence of extended functions Vn : X >—> R U {±00} whose domains are not empty. We shall say that 1. — the function lim^n^coVn whose epigraph is the lower limit of the epigraphs of the functions Vn SpQimfn-^ooVn) := Ыт.т.{п->со£р(Уп) is the upper epilimit of the functions Vn 2. — the function Ivcain^ooV-n, whose epigraph is the upper limit of the epigraphs of the functions Vn Vn(x) + R+ if χ € Dom(14) 0 if Vn(x) = +00 R if Vn{x) = —00 Vn{x) - R+ if χ € Dom(14) < 0 if Vn(x) = -00 R if Vn(x) = +00 £y(lim"p_ool4) := Limsupn_>co£p(Kl)
276 7 Graphical 8ε Epigraphical Convergence - is the lower epilimit of the functions Vn 3. — the function lim|n_»ooV^ whose hypograph is the lower limit of the hypographs of the functions Vn Vn) := Liminfn-.ooWp(Vn) is the lower hypo-limit of the functions Vn 4- — the function limi n->coVn whose hypograph is the upper limit of the hypographs of the functions Vn Hpilvailn-^oVn) := Limsupnl>coWp(14) is the upper hypo-limit of the functions Vn- If the upper and lower epilimits coincide at some point xq, then we shall say that the common value (limTn_0oKl)(a;o) := (Ит^г-юоКг) (ю) = \Jun\n->aoVn) (z0) is the epilimit of the sequence of functions Vn at xq, and we define the hypolimit lim|n_»ooVn at xq in the same way. The terminology concerning the epilimits seems at odds with the . choice of the limits; the upper epilimit is associated with the lower limit5. However, they are consistent in the case of hypo-limits. This is due to the analytical definitions of these epilimits, involving the concepts of Γ-convergence and Urn sup inf introduced in Definition 5.2.4: Ит sup^j. из£у>^уф{х', у') := sup inf sup inf ф{х',у') ■ ε>0 V>0 χ'£Β(χ,η) У £В{у,е) The concept of Urn inf sup is defined in a symmetric way, arid the adaptation to sequences (or filters) is straightforward: For instance lim sup^^ Ыу>^уфп{у') := sup inf sup inf фп(у') ε>0 W>° n>N J/'eB(i/,e) 5This would never have happened had "Convex Analysis"; been "Concave Analysis", and then, had maximization problems taken precedence over minimization ones. Which is not the case, despite economists who prefer to maximize profits, utilities and all these nice things.
7.3. EpUimits 277 Proposition 7.3.2 Let us consider a metric space X and a sequence of extended functions Vn : X i-> R U {±00} whose domains are not empty. We obtain the following formulas: i) (lim^n^ooVn) (z0) = limsup^^ infz_I0 Vn{x) ii) (lim^ool^J (x0) = liminfn-oo.i-ioKiiz) Hi) (iwaln^ooVnj {xq) = liminfn-00 snpx_X0Vn(x) ^ iv) (limJn-ooVk) (zo) = limsupn_>OOi3._>3.0Fn(z) Proof — We shall check these formulas for epigraphical convergence only. We set V^ := l\m^n-^c>oVn and show first that 4' limsup inf Vn(x) < Vt(zo) It is enough to consider the case when V^{xq) < +00. To compute the value of V» at xq, we use the fact that for every λ > V}(xq), there exist sequences of elements xn converging to xq and λπ to λ such that λη > Yn{xn)· Therefore, for all e > 0 and 77 > 0, there exists N such that, for all η > iV, we have and thus inf Vn{x) < Vn{xn) < λη < λ + ε \\Χ—Χ0\\<η sup inf Vn(x) < A + e n>N \\x-xo\\<v from which we deduce that I I I VA > W(zd), limsup inf Vn{x) < A Ί η—>oo x *XQ and thus, .that 1: :_£ τ/ с„\ ^ τ, limsup inf Vn{x) < VUxq) n—юо x—+x0
278 7 Graphical <fe Epigraphical Convergence 2. — Conversely, to prove the other inequality, we have to show that for any λο > limsup inf Vn(x) φ +οο η—>oo x ^o the pair (xq, λο) belongs to the epigraph of VJ., i.e., to the lower limits of the epigraphs of the functions Vn. But, by the very definition of the infimum, we deduce that for all e > 0 and η > 0, there exist N such that, for all η > Ν, we can find an element xn € Β(χο,η) satisfying Vn{xn) < inf Vn(x)+e < sup inf Vn(x)+e < λ0 + ε ||i-io||<7? n>N ||a:—aioll < V By taking ε = η = 1/n and setting λπ := Xq + 1/п, we have proved that xn converges to xq, λπ to λο and that Vn{xn) < \n for all n. 3. — We set vl· := limL^ool^ and prove next that liminf VJx) < Vf(xQ) n—>oo,i—>io ' It is obvious when vHxq) = +oo. Otherwise, let us estimate any λ > Vf(x0). We know that for any ε > 0, η > 0 and N > 0, there exist n' > N, (xn/, \n/) in the epigraph of Vn' satisfying ll^n'-^oll < V & Vn'{xn') ^ λη/ < λ + ε We deduce that inf Vn{x) < λ + ε п>АГ,||х-хо||<»? Hence liminf VJx) < λ π—>οο,ι->xq By letting λ —> νί(ΐο) we obtain liminf VJx) < v!(xq)
7.3. EpUimits 279 4. — To show the opposite inequality Vf(xo) < liminf Vn(x) we begin by noting that it is always true when liminf Vn(x) = +oo n—>oo,x—>xo We conclude by observing that the pair (xq, X\) where λι > liminf Vn(x) n—+ca,x^xo belongs to the epigraph of Vj, because, by the very definition of the Km inf, we can construct a subsequence of elements (again denoted) (xn, \n) of the epigraphs of Vn converging to (xq, λι). □ It may be useful to store the following inequalities i\n->ooVn < lm^n^oaVn < limj i) limL_ooKi < limL^ool^ < limL^ooV^ (7.2) ii) lim^n_ooKi < linixn-.ooVn < ^a\n-^ooVn Remark — We have defined the concepts of epilimits from the limits of the epigraphs. Conversely, we can recover the limits of subsets from the epilimits of their indicators: Proposition 7.3.3 Let us consider a metric space X and a sequence of subsets Kn С X and their indicators ψκη· Let K^ and K^ denote the upper and lower limits of the Kn's. Then О Фк> = ^т^п-*ооФкп is the upper epilimit ofij)Kn ii) φΚί = Уш\п-*°оФк„ is the lower epilimit of-ψκη The proof is left as ал exercise. □ Prom the Inverse Stability Theorem 7.1.5, we can derive criteria which imply some equalities in formulas (7.2.)
280 7 Graphical 8ε Epigrapbical Convergence Proposition 7.3.4 Let us consider α metric space X and a sequence of extended lower semicontinuous functions Vn: X *-^> RU{+oo} whose domains are not empty. - - -' Let xq € X and assume that there exist constants с > 0, N > 0 and η > 0 such that, for alln> Ν, χ £ Β(χο,η), the domains of the contingent epiderivatives D^Vn(x) are equal to the whole space and sup |2?TV„(:c)(u)| < с (7.3) Hl=i Then - - . n—ί-ехэ or, equivalently, limsup Vn(x) = limsup inf Vn(x) n—+00,χ—*xq n—+oo x~*xo Proof — Set V{ := lim^-ooVn. We apply Theorem 7.1.5 to the set-valued maps Fn defined by Fn{x) := VnT(a;) It is easy to check that assumption (7.3) implies the stability assumption of Theorem 7.1.5. This is straightforward when A„ = V(xn), since in this case ))DFn(x,Xn))) := sup inf |μ| = sup |£>TV„(a;)(u)| < с ||u||=l μΕθΡη(χ,Χη)(η) N1=1 When A„ > V(xn), Proposition 6.1.4 implies that Dom(DTVn(a;)) χ R С Graph(DFn(a;, λη)) so that \\DFn(x,Xn))):= sup inf )μ\ = 0 ||u||=l ^еВ^(1,Ап)(и) Hence, there exists a constant с such that, for all n > Ν, χ £ Dom(Vn) close to xq and A„ > Vn(x) close to V^(xq), we have ||№(a;,An)|| < с Now, Theorem 7.1.5 implies that V^(xo), which by definition belongs to the graphical lower limit of the Fn's, also belongs to the lower limit of the V„T(a;) when п-юо and χ —* xq- Therefore: limsup Vn(x) < V^(xQ) n—юо,а:—Ho The left hand side of this inequality is equal to V^(xo)· This, together with inequalities (7.2), implies that V^(xq) = V^(xq). □
7.3. Epilimits 281 7.3.2 Convergence of Infima and Minimizers We expect that the infimum of an epilimit is closely related to the limit of the infima. To prove this fact, we recall that an extended function V is said to be inf-compact or lower semicompact if for any λ € R, the level set {x € X | V(x) < X} ' is relatively compact. A sequence of functions Vn is called sequentially inf-compact if for any λ € R, every subsequence of elements {xn> € X | Vn,{xn>) < X} remains in a relatively compact set. Proposition 7.3.5 (Limits of Infima) Let us consider a metric space X and a sequence of extended functions Vn : Χ ι—» R U {+00} whose domains are not empty. Then limsup(inf Vn) < inf (Нт^-юсТ^. If we assume that the functions Vn are sequentially inf-compact, then inf (limL-^ooFn) < lim inf (inf Vn) Consequently, if V is the epilimit of a sequence of sequentially inf- compact functions Vn, then i) inf У = lim^ootinf Vn) ii) Limsup^oo-f?/ | Vn{y) = inf 14} С {z0 | V{x0) = inf V} Proof 1. — Set УТЬ := lun^n^ooVn The first inequality holds true whenever V* = +00. Otherwise, let λ > inf V}_h§ fixed and xq be chosen such that ^(zo) < λ. We know that for all η > 0, there exists N > 0 such that sup inf Vn(x) < X n>NxeB{x0,v)
282 7 Graphical 8ε Epigraphical Convergence Therefore, sup (infl^i) < sup inf Vn(x) < λ n>N n>N χ£Β{χ0,η) Hence it is enough to let N go to oo and λ to V^(xq). 2. — Set Taking a subsequence and keeping the same notations, we may assume that liminfiinf Vn) = lim (inf Vn) n—>oo n—>oo Consider first the case when liminfiinfVn) > —oo 71—>CO Let us choose for any η an element xn such that Vn(xn) < inf!4 + l/n They remain in a relatively compact subset since the functions Vn are sequentially inf-compact. So a subsequence of xn does converge to some xq. Therefore, infV^ < vf(xo) = liminfn_»oo,i-.ioKi(a;) < < liminfTl_>ooHl(2;Tl) = liminfn^ooCmf Vn) When liminfiinf Vn) = —oo n—>oo we can choose xn € X in such a way that Vn(xn) —> —oo. Sequential inf-compactness thus implies that a subsequence of xn converges to some xq. Then, for all μ > 0, the pair (ζθ) —μ) belongs to the upper limit of the epigraphs of the Ws, i.e., ( lim^n_>00l/n)(a;o) = -oo 3. — Equality г) is a consequence of the first two inequalities. Observe that the last statement holds true if the left-hand side of inclusion ii) is empty. Otherwise, consider an element xq € Limsupn_too{y | Vn{y) = infVn}
7.3. Epilimits 283 which is the limit of a subsequence of elements xnt such that Vni {xn') = inf Vni. Then i) implies that У\-(xo) < liminf Vni{xni) = inf У so that νγ(ΐο) < inf У. Since V(xq) = Vf(xo) by assumption, we have proved that xq minimizes V. Π · Unfortunately, there are counter-examples for the property that the set of minimizers of the upper epilimit is the lower limit of the sets of minimizers of the functions Vn. However, Theorem 7.1.5 provides some results about the lower limits of level sets: Proposition 7.3.6 Let us consider a metric space X and a sequence of extended lower semicontinuous functions Vn : X н-> R U {±00} whose domains are not empty. Let xq belong to the domain of their upper epilimit and assume that there exist N > 0 and constants с > 0, η > 0 such that, for all n> Ν, χ € B(xq, 77) Π Dom(Fn), (Ϊ) 3 u~ € cBx such that D^Vn{x){u^) = -1 (7.4) ii) 3 и+ € cBx such that DfVn(x)(u+) = +1 Then there exists a constant I such that, for any sequence of elements XQn converging to xq such thatVn(xQn) converges to (limi-n-^ooV^J {xq), and for any \n converging to V^(xq), there exist xn € X satisfying Vn(xn) < \n, Ha-n-zonll < l\\n-Vn{xQn)\ Proof — Set Vj := Ип^п-юоКг We apply Theorem 5.4.2 to the set-valued maps Fn(x) := Vn-\{x), which are obviously consistent at (xq,V^(xq)) by definition of*Vp SetA0 = FTb(z0). Assumption (7.4) implies that the set-valued maps Fn are stable: There exist constants η > 0 and с > 0 such that, for any (ιη,μη) € GTaph(Fn)nB((x0,\o),v)
284 7 Graphical L· Epigrapbical Convergence for any υ € R, there exists un € X such that ϋ 6 DFn(xn, μη)(ηη) L· \\un\\ < c\v\ Indeed, if μη > Vn(xn), we can take un = 0. If μη = Vn(xn), we take ν = vu% if ν > 0 and υ = —vu~ if ν < 0. Therefore, Theorem 5.4.2 implies that for any λπ approximating λο, we can find xn € F^r{\n) satisfying the required error estimates. D Remark — Observe that stability assumptions (7.4) imply that for all η > N and all xn € Β(χο,η), Xn) ι ел since the Fermat rule is violated. This excludes the case when the level sets are set of minimizers. □ 7.3.3 Variational Systems Naturally, we can introduce the same definitions for "continuous" parameters и GU, where U is another metric space. In such a framework, we consider a family of extended functions Vii £ U, V(u,-):X ^ Ru{±oo} depending upon the parameter u, with which we associate the set-valued maps V(u,x)+n+ if (u,x) £ Dom(y) i) V(it)T(a;) := { 0 if V{u,x) = +oo R if V(u,x) = —oo V(ii,a;)-R+ if (u,x) £ Dom(y) гг) V(u)j.(a;) := { 0 if V(u,x) = -oo R if V(u,x) = +oo There is a subtle, but important, difference between the extended function V:(u,x) £ U χ X н-> RU{±oo} for which the variables и and χ are on the same footing, and the set-valued maps V(-)T ·· и £ U — V(u)T
7.3. EpUimits 285 for which the variables play a different role: u, the role of a parameter and x the role of the variable. This type of situation occurs for instance in optimization where the order relation on R involves the variable x, when we minimize the function depending on the parameter и with respect to x. To emphasize this difference, we borrow from Rockafellar and Wets the following terminology: a set-valued map V(-)1- is called a variational system. Definition 7.3.7 (Variational system) Let us consider metric spaces X, U and a variational systemV(-)^ defined on UxX. We shall say that V(·, ·) (or V(-)f, to be precise), is 1. — upper epicontinuous at и if the set-valued map u' ~> V(u')j is lower semicontinuous at и 2. — lower epicontinuous at и i/V(u)1- = limsup „/^„V^ii') τ 3. — epicontinuous at и if it is both lower and upper epicontinuous at u. The definition of upper and lower hypocontinuity are naturally symmetric. As in the discrete case, these concepts can be characterized in an analytic way: Proposition 7.3.8 Let us consider metric spaces X and U and a variational system V(-)T defined on U x X. Then V is • 1. — upper epicontinuous at и if and only if for any χ £ Oom.(V (и, ·)) V(и,χ) := limsup inf V(u',x') 2. — lower epicontinuous at и if and only if for any χ £ Ooxn{V(u, ·)) V(u,x) = lira inf V(u',x') uf —*u,xf —*x 3. — epicontinuous at и if and only if for any χ £ Dom(y(u, ·)) V(u,x) = limsup inf V(u',x') = liminf V(u,x) I .TV b.T ΊΙ· 17) T·' IT 11' —HL.X' —*X Example — Definition 6;1.2-©f epiderivatives provides another interpretation of the epiderivatives in terms of epilimits of the difference quotients V(x + hu)-V(x) VhV(x) := и ^ h
286 7 Graphical <fe Epigraphies! Convergence 1. — the contingent epiderivative is the lower epilimit of the difference quotients VhV(x) when h —> 0+ 2. — the adjacent epiderivative is the upper epilimit of the" difference quotients VhV(x) when h —> 0+ 3. — the circatangent epiderivative is the upper epilimit of the difference quotients и ^ (V(x' + hu) - \')/h when h —> 0+ and (x\ λ') G £p(V) converges to (x, V(x)). □ 7.4 Epilimits of Sums and Composition Products This section is devoted to the upper and lower epilimits of the functions Un := Vn + Wn о А in terms of the upper and lower epilimits of the functions Vn and Wn. Theorem 7.4.1 Let us consider Banach spaces X, Y, a continuous linear operator A £ £(X, Y), two sequences of extended functions Vn and Wn from X and Υ to R U {+00} respectively and define Un := Vn + WnoA 1. — Then for every Xq £ X, the inequality (limeTn_>00Vn)(a;o) + (limeTn_i00Wrn)(Aa;o) < (limeTn_>00[/n)(a;o) (7.5) holds true whenever the left hand side is well defined. 2. — If in addition Vn and Wn are lower semicontinuous and the following stability assumption holds true: there exist constants с > 0, a £ [0,1[ and η > 0 such that, for all n, i) \/x £ Όοτη(νη)Γ\Β{χ0,η), V у £ Dom(Wn)n Β(Αχ0,η) BY с A(Oom(D\Vn{x))ncBx\ -Oom{D^Wn{y)) + aBY < ") snPueOom(D)vn(x))\Di\Vn(x)(u)\/M ^ c Hi) snpveOom{DiWn{y)) \DiWn(y){v)\/\\v\\ < с (7.6)
7.4. Epilimits of Sums and Composition Product 287 then the upper epHim.it \т\^п^,хип satisfies: n—+00 ип)Ы < (1Ц τι—+oo whenever the right hand side is well defined. Consequently, if the sequences of functions Vn and Wn have epilimits VT and WT at xQ and AxQ respectively and if V^(xQ) + W^Axq) is well defined, then the functions Un have an epilimit U-\ at Xq and Щ{х0) = V1(x0) + W1(Ax0) Proof— For simplicity, we adopt the notations Vf := lim^^Vn & V? := lim^^^Vn for the functions Vn as well as for the functions Un and Wn- Since inequality (7.5) holds true when one of the values Vf(x0) and W^(Ax0) is equal to -00, or Ι/Ηχο) is equal to +00, we have to check the formula when the two first values are larger than —00 and the last one is smaller than +00. Then, by definition of the lim inf, there is a finite number p, an integer N and a positive number η such that Vn > N, \/x £ Β(χ0,η), Vn{x) > p, Wn(Ax) > ρ By definition of the lower epilimit, the pair (хо,Щ(хо)) is the limit of a subsequence (again denoted by) (arn,cn) satisfying Vn{xn) + Wn{Axn) < cn The two above inequalities imply that the sequences of real numbers an := Vn(xn) and bn := cn — an are bounded. Hence, subsequences (again denoted) an and bn do converge to a and b satisfying Uf(x0) = a + b, a > Vf(x0), b > W\{AxQ) To prove the second statement, we begin by observing that if we set ' i) Kn := Sp(Vn) χ Sp{Wn) xR с XxRxFxRxR < ii) G(x,a,y,b,c) := (Ax — y, a + b — c) Hi) H(x,a,y,b,c) := (x,c) then we can write Sp(Un) = ff^nG-^O.O))
288 7 Graphical L·Epigrapb'ical Convergence Therefore,„it is-sufficient to show that the lower limit of the subsets Kn Π G-1(0,0)_ contains the intersection of the lower limit of the subsets Kn with G~i(0,0). For that purpose, we shall use Theorem 3.4.5. Let us consider any sequence of elements (χοη,αοη) e £p(Vn), (y0n,bQn) £ Sp(Wn) converging respectively to (xq, VHxq)) and (Axq, W*(Axq)) and let us set con := aon + bon. We observe that the elements (xon,aon)yon,bon)con) belong to Kn and that С(жо„,аоп,2/Оп)Ьоп,СОп) = {Αχαη — 2/Οτι,Ο) " converges to (0,0). We begin by checking that the assumptions (7.6) of our Theorem imply the stability assumption (3.9) of Theorem 3.4.5, i.e., that there exists a constant с > 0 such that, for all n, V (x,a,y,b,c) ε Kn close to (xQ, V^(x0),yo,W\(y0),v) for all (ζ, Α) ε X x R, there exist (u, μ, ν, у, δ) ε Τκ„ {х, а, У, Ь, с) and e ε Χ satisfying г) ζ = Аи — ν + е & Χ = μ + u — δ , ii) ||e|| < α(|Μ| + |λ|) ^ ш) |Н| + |И| + и + и + |5|<с(|И| + |А|) Assumptions (7.6) г) and гг) imply right away that there exist u ε Oora{D\Vn{x)), ν ε Dom(DTWn(y)) & e ε Χ satisfying ζ = Au-v + e, ||e|| < a(\\z\\ + |λ|), ||U|| < φ|| Then Ml < 1И1 + 1|е|| + ИИН| < 7(ΙΝΙ + λ) for some positive 7 independent of (z, A). Let us take now μ := c||u||, и := φ\\, δ := c(||u|| + ||v||) - A We deduce from (7.6) иг) that (η,μ) ε Ep{DiVn){x), {υ,ν) ε Sp{D}Wn){y)
7.5. Conjugate Functions of Epilimits 289 and that DiVn{x){u) + DiWn{y){v) < c(|H| + ||w||) = λ + δ Consequently \δ) < |A| + C(|H| + ||«||) < c'(\\z)\ + \X\) The conclusion of Theorem 3.4.5 being available, we then know that there exist elements (£П)аП)2/П)ЬП)Сп) 6 Кn satisfying ' * G(xn,an,yn,bn,cn) = 0 and ||in-a;on|| + ||yn-2/On|| + |an-aon| + |bn-bon| + |c'n-con| <Ί))Λψθη-νοη)\ Hence lirrin-*^ \an — aon\ = lim,^,» \bn — bon\ = 0, so that wn and bn converge to, Xf(xa) and ty*(xo) respectively. Furthermore, inequalities Ul{xn) < Si +U imply tllat Ы(хо)'< Vhx'a) + W\Aa). Q • 1 \i '· ■ ■ \ ι ' ' « ι : , 7.5 Conjugate Functions of Epilimits • ϊ i ι \ 1 i Ί Let us consider a Banach space X, a sequence of nontrivial convex I lower semicontinuous functions Vn:X i-> R U {+oo} and their conjugate functions vf.X* i-> Ru{+oo} III * ' \ • '. ' ί . : The purpose of this section is to characterize the upperj epilimit of the! functions Vn. ! We can still define conjugate of an extended function U:X *-+ RU{±oo} by the formula U*(p) = s\ip(<p,x> -U(x)) x'ex
290 7 Graphical <fe Epigrapbical Convergence We observe that U* = +00 when U(x) = —00 for some χ G X and that if U = +oo, its conjugate U* = —00. When X* is supplied with the weak-* topology, we naturally adapt the definition of lower epilimit of a sequence of extended functions Wn : X* i-> R U {±oo} to the concept of sequentially weak lower epilimit: The epigraph of σ — limS.n_,ooWn is .by definition the sequentially weak upper limit of the epigraphs of the functions Wn. Theorem 7.5.1 (Mosco) Consider a Banach space X and a sequence of nontrivial convex lower semicontinuous functions Vn: X н-> R U {+oo} Assume that the upper epilimit of the functions Vn is not trivial. Then it is equal to the conjugate of the sequentially weak lower epilimit of the sequence of conjugate functions V£ : lim^n_ooVn = (σ - ΙκΏ^η-χχ,ν*) Proof — We set Vf := lim'fn-.oo^n, W := a-lmi^^V* Then for any ρ G X*, there exists a subsequence of elements (ρηι, ληι) € Sp{y*i) converging weakly to (p, W(p)). We observe first that inequality W*[x) < Vf(x) always holds true: Indeed, it holds true when V^(x) = +00 or W = +00, because, in this case, W* = —00. Assume now that Vt (ж) < +oo and pick ρ such that W(p) < +00. Then there exist sequences of elements xn and μη > Vn(xn) converging (strongly) to χ and V}(x) respectively and subsequences pn> and λπ/ > V*i(pn') converging weakly to ρ and W(p) respectively Therefore, by taking the limit when n' —> 00, Fenchel inequalities < Ρη',Χη' > < Vni{xni) + V*,(pnt) < μη' + Xn' imply <p,x> -W(p) < Vf(x)
7-5. Conjugate Functions of Epilimits 291 from which our claim follows. If for some q € X*, W{q) = -oo, then for all χ € X, W*(x) = +00. Consequently, in this case, inequality Vf(x) < W*(x) (7.7) is satisfied. It still holds true when V*(x) = —00. It remains to prove that it is verified when W never takes the value —00 and V^(x) > —00. To this end, we shall check that λ < W*(x) for all λ < V^(x). Fix such (ж,λ). The latter statement means that the pair (x,\) does not belong to the epigraph of V?, which is, by definition, the lower limit of the epigraphs of functions Vn. Therefore, there exist e > 0 and a subsequence of functions (again denoted by) Vn such that ((x + eB)x(X + e[-ll+l]))neP(Vn) = 0 We infer from the Hahn-Banach Separation Theorem that there exist elements (pn, —an) GPxR satisfying ||pn|| + \an\ = 1 and V (χη,μη) € £p(Vn), (ρη,Χη) - θίημη < (pn,x) - 0ίη\ - ε (7.8) This inequality implies that an > 0 because, if an < 0, it is enough to take xn G Dom(14) and let μ\ := Vn(xn) + к —> +oo with к to derive the contradiction 00 < est. We proceed now in three steps. 1. — Assume that lim infn-»00 απ is strictly positive. Then there exist δ > 0 and a subsequence (again denoted by) an satisfying an > δ. Therefore, we can divide inequality (7.8) by an > 0. Setting qn ;= pn/an and taking the supremum over the xn G Oom.(Vn), we deduce that, V*(qn) < < <ln,x > -λ - ε/αη < < qn,x > -λ On the other hand, ||^π|| < 1/5, so that there exists a subsequence (again denoted) qn converging weakly to some cluster point q. Since < qn, χ > -X converges to < q, χ > -λ, we infer that W(q) < (q,x)-X
292 7 Graphical L· Epigraphical Convergence and thus, that λ < W*(x). 2. — This happens at least when χ belongs to the domain of Vp because, in this case, we claim that αη > ε/4ρ for η large enough (where ρ := V^(x) — λ > 0.) Indeed, there exist sequences of elements x$n and μη > V(xon) converging strongly to χ and V^(x) respectively. Therefore, for η large enough, \\x — --xbnW < ε/2 and μη < V^(x) + p. By taking such x0n and μη in inequality (7.8), we obtain 0 < (pn, x ~ xon) + θίη(μη - λ) - ε < ||ж - жоп|| + 071(^(0;) - λ + ρ) -ε < 2αηρ-ε/2 Therefore, inequality (7.7) holds true whenever V+(x) < +00. Furthermore, we deduce from the first step that the domain of W is not empty. 3. — Consider now the case when V+(x) = +00. Either liminfπ_4θο απ > 0 and, by the first step, λ < W*(x), or liminfn-xxian = 0, and a subsequence (again denoted by) an converges to 0. In this case, we take an element q in the domain of W, which is not empty by the second step. Then there exist subsequences (again denoted) qn and pn > V*(qn) which converge weakly to some q and W(q) respectively. Next, we add FenchePs inequality (ίη,Χη) < Vn(xn) + V*(qn) to inequalities (7.8) multiplied by μ > 0 and we obtain Vin€ Dom(14), <qn + μρ < (1 + μαη)νη(χη) + Pn+ < μΡη, x > -μαπλ - με Dividing by 1 + μαη and taking the supremum over the domain of Vn, yields y* f<ln + μΡη\ < pn+ < μρπ, x > -μαηλ - με η\1 + μαη ) ~ 1 + μαη
7.5. Conjugate Functions of EpUimits 293 On the other hand, knowing that ||pn|| < 1, there exists still a subsequence (again denoted by) pn converging weakly to some cluster point p. The right hand side of the above inequality converges to W(q)+ < μρ,χ > —με. Hence, by going to the limit, we infer that ΪΥ(ς + μρ) < W(q) + < μρ,χ > -με Therefore, taking into account that <ς + μρ,χ> -Wiq + μρ) < W*(x) we obtain με - W(q) + < q,x > < W*(x) Since e > 0, we deduce that W*(x) > λ by taking μ large enough. Therefore, in both cases, λ < W*(x), so that, by letting λ —> +oo, we infer that W*(x) = +oo. □ By taking the conjugates, we obtain the formula (lim^ooVy* = (σ - lim^-^ooV^) Applying it to the indicators Vn := φκη of closed convex subsets Kn, the conjugate of which are the support functions of these subsets, we obtain a formula for the support function of the lower limit of a sequence of subsets Kn: 9 Corollary 7.5.2..Let X be α Banach space and Kn С X be a sequence of nonempty closed convex subsets. Then the support function of their lower limit К is the biconjugate of the sequentially weak lower limit of the support functions σκη of the subsets Kn: σ{Κ\·) = (a-lim»„.w7tfn (·))** When the space X is reflexive, we can apply Theorem 7.5.1 to the conjugate functions V£ to obtain the following consequence: Theorem 7.5.3 Consider a reflexive Banach space X and a sequence of nontrivial convex lower semicontinuous functions Vn '■ X >-^ Ru{+oo}.
294 7 Graphical L· Epigraphical Convergence Assume that the upper epilimit of the conjugates V* is not trivial. Then the conjugate of the sequentially weak lower epilimit of the sequence of functions Vn is equal to the upper epilimit of the conjugates y*. (σ-ίπηξη-^ο^η)* = limfn-ooV^ Consequently, we deduce the following: Corollary 7.5.4 Consider a reflexive Banach space X and a sequence of nontrivial lower semicontinuous convex functions Vn:X н-> R U {+oo} Let us assume that its upper epilimit and the sequentially weak lower epilimit do coincide with a nontrivial function V. Then for any χ € Dom(V) and ρ G Dom(V*), there exist sequences xn G X and pn £ X* converging to χ and ρ respectively and satisfying limsupVn(:cn) < V(x), \imsnpV*(pn) < V*(p) n—>oo n—>oo Applying Theorem 7.5.3 to the indicators Vn := φκη of subsets Kn, we obtain a formula for the support function of the sequentially weak upper limit of subsets: Corollary 7.5.5 Let Kn be a sequence of nonempty closed convex subsets of a reflexive Banach space X. Then the support function of the sequentially weak upper limit of the subsets Kn is equal to the upper epilimit of the support functions of the subsets Kn. 7.6 Graphical Convergence of Gradients We study in this section the convergence properties of gradients and generalized gradients of a sequence of functions Vn.
7.6. Graphical Convergence of Gradients 295 7.6.1 Convergence of Gradients of Smooth Functions We begin with Theorem 7.6.1 Let Ω be an open subset of a Banach space X. Consider a sequence of Frechet differentiable functions Vn : Ω ι—» R converging to a function V uniformly on bounded subsets σ/Ω. Assume also that V is bounded from below on bounded subsets ο/Ω. Then the set-valued map д-V : Ω ~> X* is contained in the graphical lower limit of the gradients V^ of the functions Vn: for any χ G Ω and any ρ G d-V(x), there exists a sequence xn converging to χ such that V^(xn) converges to p. If V is bounded from above on bounded subsets of Ω, the set- valued map d+V is also contained in the graphical lower limit of the gradients V^ of the functions Vn. The same holds true if the functions Vn are locally Lipschitz and Gateaux differentiable on Ω. Proof— Let us fix e > 0 and χ G Ω. To say that ρ G d_V(x) amounts to saying that there exists μ G]0, e[ such that Vy € Β(χ,μ), V(x) - V(y)- <p,x-y> < ~\\x - y\\ We may assume that Β(χ,μ) С Ω. Then, the functions Vn converge uniformly to V on the ball Β(χ,μ): For any δ > 0, there exists Ns such that, for all η > N$ and for all у G Β(χ,μ), Vn(y) — V(y) < δ/2. We take δ = με/A and fix η > Νβ. We now apply Ekeland's Theorem to the continuous function У ·-» Vn{y) - <p,y > on the ball Β(χ,μ): There exists xn G Β(χ,μ) satisfying i) Vn(xn)- < P> χη > +ε\\χ -χη\\ < Vn(x)- <p,x> ii) Vy G Β(χ,μ), Vn(xn)- <p,xn> < Vn(y)- <p,y> +e\\xn-y\\ We infer from the first inequality that ε\\χ — xn\\ < δ + V(x) — V(xn)— < ρ,χ — xn > < δ + ~\\x — xn\\
296 7 Graphical L· Epigraphical Convergence ι from which we deduce that ' ; . · i \\x< — Xn\\ < 2δ/ε = μ/2 ) j Sipc.e \cjl belongs to the interior of B(x, μ) and minimizes on this bail the function ; ; ϊ '■ '; ; ί Vn(y)~ < P, У> +ε\\χη - y\\ the Fermat Rule implies that V и € X, 0 < (Vn(xn)-p,u) + e\\u\\ ■ : I \ which means that ||ν^(χπ) —ρ\\ < ε. Hence the pair (x,-p) is the limit of a sequence of elements (xn, V^(xn)) of the graph of V^. Π This statement is actually a consequence of a more precise Theorem 7.6.2 Let us consider α Banach space X and a sequence of extended lower semicontinuous functions Vn : X >—» R U {+00} whose domains contain a ball B(x, a). Assume that the functions Vn converge uniformly to an extended function V on this ball. Assume also that V is bounded from below on B(x,a). Choose any kind of δ-generalized gradients dsVn. Then for any ρ € d-V{x), there exists a^sequence xn converging to χ and a sequence of elements pn € dsVn(xn) converging to p. ; Proof— j We begin the proof as in Theorem 7.6.1, but this time the Fermat Rule implies that i /'·,..' i€ Χ,.Οζ ΡΊνη(χη)(ΐί)- ,< p, и > +e\\u\\ Consequently/, the function] 14' is contingently' epidiffereniiabk a't«:,(. rjLLBtnermbre, w;e deduce that for any convex epidenyatiye OsVh{xA) ■ [and 'for дШ ■ ' , > j , : \ j - . i Vu € X, 0 < δνη(χη)(ιή-<p,u>+e\\u\\ ■it- i i This can be written as i 0 € dsVn(xn) -ρ + εΒ* where. S* denotes the unit ball of -X*. This means that there exists p\i^-BbYii{xn).such, that \\p —. pn\\ < e. Hence the pair (x,p) is the limit of a sequence of elements (xn,Pn) of the graph of dsVn. □
7.6. Graphical Convergence of Gradients 297 7.6.2 Convergence of Subdifferentials of Convex Functions Theorem 7.6.3 Consider α reflexive Banach space X and a sequence of nontrivial lower semicontinuous convex functions Vn: X *-^> RU{+oo}. Let us assume that its upper epilimit and the sequentially weak lower epilimit coincide with a nontrivial function V. Then the subdifferential map dV is contained in the graphical lower limit of the subdifferential maps dVn: 3V С Lim^oo^ More precisely, let ρ G dV(x). Consider sequences XQn and p$n converging to χ and ρ respectively satisfying limsupFn(a;on) < V(x), limsupV^POn) < V*(p) П—+СО 71—>CO (which exist by Corollary 7.5-4-) Then the error estimate d((xon,Pon), Graph(<9Vn)) < yVn(x0n) + V*{p0n)- < p0n, x0n > (7.9) holds true. , Proof — We introduce the "duality lack" δη ■= Vn(x0n) + V*(pQn)- < P0n,XQn > > 0 which converges to 0, since by assumption 0 < limsupn_4005n < < lim sup^oo Vn(xon) + limsup^,*, V*(p0n)- <p,x> < V(x)+V*(p)- <p,x> = 0 If δη = 0, this means that p$n belongs to dVn{xQn). If not, we apply Ekeland's Theorem to the lower semicontinuous function - У ·-> Vn(y)- <p0n,y >
298 7 Graphical &ε Epigraphical Convergence Since limsupj^oo V*(j>on) < +00, there exist constants с and N such that " Vn > TV, VJipon) < с This implies that the functions у ι—> Vn(y) — < pon, у > are bounded from below by — с Hence, we can apply Ekeland's Variational Principle in the generalized gradient form (Theorem 6.4.4): there exists a solution xn satisfying I г) Vn(xn)- < ροη,Χη > +V%n\\ < Vn(xQn)- < pon, x0n > (7.10) ^ it) 0 € dVn{xn) - p0n + л/δ^Β* Inequality (7.10) гг) tells us that there exists pn in dVn(xn) satisfying Upon ~Pn\\ < \β~η On the other hand, inequality (7.10) г) yields ll^On - xn\\ < -7= (Vn(xon) - Vn(xn)+ < pon, Xn> - < POn, XQn >) VOn Taking into account that < Pon, xn> < Vn{xn) + V*(p0n) we obtain \\XQn ~ Xn\\ < -7= (Vn(x0n) + VniPOn)- < ΡΟη,ΧΟη >) = λβ~η VOn Hence inequality (7.9) is proved. Since (ж, ρ) is the limit of (xon, pQn) by definition, we deduce that it is also the limit of the sequence of elements (xn, pn) of the graph of dVn. □ Since dV is a maximal monotone set-valued map when X is a Hubert space, Proposition 7.1.7 implies:
7.6. Graphical Convergence of Gradients 299 Theorem 7.6.4 (Attouch) Let X be α Hilbert space. We supply X with the norm topology and X* with the weak-* topology. Consider a sequence of nontrivial lower semicontinuous convex functions Vn : IhRU{+oo}. Let us assume that its upper epilimit and the sequentially weak lower epilimit do coincide with a nontrivial function denoted by V. Then dv = iW'n-oodVn = iJmin-oodVn In particular, we deduce that under the assumptions of Theorem 7.6.4, the graphical convergence of the subdifferentials dVn to dV implies that for every χ £ X, dV(x) = σ -LimsuO^^^dVnixn) When the functions Vn are the indicators of nonempty closed convex subsets Kn, we obtain the following theorem on the convergence of the sequence of normal cones to these subsets: Corollary 7.6.5 Let X be a Hilbert space and Kn С X be a sequence of nonempty closed convex subsets. We supply X with the norm topology and X* with the weak-* topology. Assume that the lower limit К of the sequence {Kn)n coincides with its sequentially weak upper limit. Then NK(x) = tf-Lnnsupn^0OiXn^a.J\fan(a:n) In order to state that some ρ 6 dV(x) belongs to the lower limit of dVn{xn) for any sequence xn G Oom.(dVn) converging to ж, we need some stability assumptions on the contingent second derivatives of the functions Vn- Naturally, the contingent Hessian d2V(x,p) := D(dV)(x,p) of У at some point (ж, ρ) in the graph of dV is defined as the contingent derivative of the set-valued map dV at {x,p). Proposition 7.6.6 Let X be a Hilbert space. Consider a sequence of nontrivial lower semicontinuous convex functions Vn : X н-» R U {+00}. Let us assume that their upper epilimit and their sequentially weak lower epilimit coincide with a nontrivial function denoted by V.
300 7 Graphical L· Epigraphical Convergence Let ρ belong to dV(x). We posit the following stability assumption: There exist constants с > 0 and η > 0 such that V (xn,Pn) € Graph(<9Vn) Π Β((χ,ρ),η), Oom(d2Vn(xn,Pn)) = X < and w \\d2Vn{Xn,Pn)\\ ■= SUpuejfinfiree2v'„(xBlp„)(«)lkll/ll«ll < C Then ρ € Limb3£n^00iXn-^xdVn{xn). Proof — We apply Theorem 5.4.2 to the set-valued maps Fn := dVn. U 7.7 Asymptotic Epiderivatives We are now able to define asymptotic epiderivatives of a sequence of functions Vn, by taking the asymptotic tangent cones to their epigraphs. "We denote by V\ := Нт^-ооК their upper limit. Definition 7.7.1 (Asymptotic Epiderivatives) Let us consider a sequence of extended functions Vn :IhRU {+oo} whose domains are not empty and an element xq in the domain of the upper epilimit of the functions Vn (i.e., V^{xq) φ ±oo.) We shall say that the function C^{Vn}(xo) defined by C\{Vn}(x0)(u) : = Urn sup (Vn(x+hu') - X)/h n—>oo,x—>xo,Vn(x)<X—*V^(x0),h—*0+ is the asymptotic circatangent epiderivative of the sequence of functions Vn at xq in the direction u. We see at once that the epigraph of C^{Vn}(xo) is the asymptotic circatangent cone to the epigraphs of the functions Vn at (xo,V^(xq)), or, equivalently, that C^{Vn}(xo) is the upper epilimit of the difference quotients u' i-+ (Vn(x+hu')-X)/h when n —> oo & (x,X,h) £ fp(K,) χ R+ converges to (xo,V%(xo),ty
7.7. Asymptotic Epiderivatives 301 We deduce that the asymptotic circatangent epiderivative is a positively homogeneous, lower semicontinuous and convex function from X to R U {±oq}. We shall estimate the asymptotic circatangent epiderivative of a family of functions Un :=Vn + Wn ° A. Theorem 7.7.2 Consider two Banach spaces X and Y, a continuous linear operator A £ £(X, Y) and two sequences of extended lower semicontinuous functions Vn and Wn from X and Υ to R U {+00} respectively. Assume thatx0 belongs to the domain ofV^ and Ax Q to the domain ofW\. We posit the following stability assumption/ there exist constants с > 0, α £ [0,1[ and η > 0 such that, for all n, (7.6) holds true. Then, the asymptotic circatangent epiderivative of the sequence of functions Un -.= Vn + Wn о A satisfies the estimate: C\{Un}(xQ)(u) < &^{Vn}{x0){u) + C\{Wn}{AxQ){Au) (7.11) whenever the right hand side is well defined. Remark — Since the assumptions of this theorem are the same as those of Theorem 7.4.1, we know that Щ(х0) < V^ (x0)+W^ (AxQ) whenever the right hand side is well defined. □ Proof — We apply Theorem 3.4.5, since we have seen in the proof of Theorem 7.4.1 that if we set ' i) Kn := £p{Vn) x Sp(Wn) xR с XxRx^xRxR < ii) G(x, a, y, b, c) := (Ax — y, a + b - c) Hi) H(x,a,y,b,c) := (x,c) we can write Ep(Un) = tf(JfnnG-i(o,0)) The stability assumption being the same as the one of Theorem 7.4.1, we already know that they imply the stability assumptions of Theorem 3.4.5. Hence, we deduce exactly as in the proof of Theorem 7.4.1 that ' C\, (x0, V\{x0), Ax0, W\(AxQ),Ifyxo)) П G~\Q,0) < (so, Vffro), AxQi W\(Ax0), υ\(χ0ή It remains to show that this inclusion implies inequality (7.11.)
302 7 Graphical &ε Epigraphical Convergence Inequality (7.11) being satisfied when one of the members of the right hand side is +oo, we assume that both of them are different from +oo. Let us consider λ > C\{Vn}(x0)(u), μ > C\{Wn}(Ax0)(Au), ν := λ + μ Hence the element (u, A, Au, μ, ν) belongs to СЬкь (χ0,ν$(χο),Αχα,νή(Αχα),ΐή(χο)) nGT^O.O) It then belongs to the asymptotic circatangent cone to Kn Л G~1(0,0). Then, for all sequence hn > 0 converging to zero, there exist elements (un, λ„, μη) converging to (u, λ, μ) such that, for all η > Ν, (xn + hnu + Ηημη, αη + bn + Ηηνη) belongs to Kn П G-\O, 0). Therefore, the pairs (xn + hnun, an + bn + hn(A„ + μη)) belong to the epigraph of Un. Since (iLn,Xn +μη) converges to (u, v), we deduce that C\{Un}(x0)(u) < λ + μ By letting A and μ converge to C^{Vn}(xo)(u) and C^{Wn}(Ax0)(Au) respectively, we infer that C\{Un}(x0)(u) < C\{Vn}(x0)(u) + C\{Wn}(Ax0)(Au) □
Chapter 8 Measurability and Integration of Set»Valued Maps Introduction We meet measurable maps whenever we deal with models of systems having measurable data. Integrals of set-valued maps are involved in many convexification (also called relaxation) problems, since roughly speaking, the integral of a measurable set-valued map is always convex. Another important instance where measurable set-valued maps do arise is related to linearization of differential inclusions along a solution. Consider indeed a Lipschitz set-valued map F : Rn ~> Rn and the differential inclusion χ € Fix) Let χ ·. [0,Г] н-> Rn be a solution to this differential inclusion, i.e., an absolutely continuous map satisfying x'{t) G F(x(t)) almost everywhere in [0, T]. Linearization of this inclusion along x(·) leads to the new differential inclusion: w'(t) € DF(x(t),x'(t))(w(t)) almost everywhere in [0, T) 303
304 8- Measurability and Integration of Set-Valued Maps where DF denotes the contingent derivative of F. Setting G(t,w) := DF{x{t),x'{t)){w) we obtain a set-valued map G merely measurable in time and Lips- chitz in w (see Chapter 10.) We investigate in this chapter some properties of measurable set- valued maps with closed images, which we define in Section 1. Under adequate assumptions, α set-valued map is measurable if and only if its graph is measurable. Measurable maps with nonempty closed images have measurable selections. This result was born as a lemma by von Neumann in 1949 and has found different extensions and applications. It was given in a stronger form by Kuratowski and Ryll-Nardzewski. Castaing has proved that there exists a countable dense family of measurable selections, a useful theorem since countability plays a major role· in many problems of measurability. These two statements constitute the core of what we call the Characterization Theorem for measurable maps, that we prove in Section 3. But it is already used in Section 2 which is devoted to the calculus of measurable maps. It is shown there that the main operations, such as union, intersection, direct and inverse image, lower and upper limits etc., do preserve measurability. The convex hull of a measurable map is measurable, and measurable selections from the convex hull have a Caratheodory representation. This is particularly useful in control theory, when we have to deal with relaxed controls. In Section 4 we investigate selections from limits of measurable maps and in Section 5 tangent cones in Lebesgue spaces. Section 6 is devoted to the concept of integral of set-valued maps introduced by Aumann, who proved that the integral is closed and convex when the set-valued map is measurable, integrably bounded and takes closed values in a finite dimensional space, (He was motivated at the time by problems of cooperative game theory and mathematical economics where convexity properties are missing when the set of players is finite. He then introduced the concept of continuum of players and replaced the sum by the integral.) Convexity of the integral is an immediate consequence of Lya- punov's Convexity Theorem on convexity of the range of a nonatomic
8.0. Introduction 305 vector measure. Similar results concerning convexity of the integral are still valid in the infinite dimensional case (the closure of the integral is convex.) They are proved in Section 7. The integral of a time-dependent set-valued map is also equal to the integral of finite concatenations of extremal selections. This result is sometimes referred as the bang-bang principle. The bang- bang principle we state here is due to Olech. It says in essence, that to integrate a map F : [a, b] ~> Rn, we can restrict the attention to piecewise extremal trajectories with the number of switchings not greater than η +1. This is the topic of Section 8, which does not use the Lyapunov Theorem. The last section deals with the extension to upper semicontinuous maps of the Poincare Recurrence Theorem: Let X be a compact metric space, V(X) denote the set of probability measures on X, F : X ~> X be a closed set-valued map and μ G V(X) an invariant measure of F. For any Borelian subset B, let Boo := Π U F~n{B) N>0 n>N be the subset of points χ such that for all N, there exists η > N such that F~n(x) Π Β φ 0. Then the measure of В Г) В^ is equal to the measure of В. The statement of this theorem is clear as soon as we have defined the notion of an invariant measure of a set-valued map F: Let В denote the σ-algebra of Borel subsets of X. We recall that if F = f is single-valued, an invariant probability measure μ is defined by: VA&B, μ(Α)=μ(Γ1(Α)) When F is set-valued, we cannot extend this definition as it is because А н-> μ(Ρ~ι(Α)) is no longer a measure. However, we shall introduce the following definition: We say that μ is an invariant measure of a closed set-valued map F : X ~> X if and only if VAeB, μ{Α) < μ{Ρ~χ{Α)) (8.1) Indeed we see that, for single-valued maps /, this definition coincides with the classical one when we apply it to both A and its complement. We shall prove the following result:
306 8- Measurability and Integration of Set-Valued Maps Let X be α compact metric space and F : X ~> X be a closed set- valued map with nonempty values. Then there exists an invariant probability measure. 8.1 Measurable Set-Valued Maps We define in this section measurable set-valued maps taking their values in a complete separable metric space, (a Polish space1.) This contains separable Banach spaces and, in particular, Lebesgue spaces LP and Sobolev spaces Wm'p with 1 < ρ < oo. Consider a set Ω and a family A of subsets A of Ω. Recall that A is called a a—algebra if it verifies the following properties: ' i) 0 € A <ii)A€A=^£l\A€A Hi) An € A, n = 1, 2,... ==> Un>i Аг € -4 It is clear that i) and ii) imply that Ω € A and from й), га) follows that the intersection of a countable family of elements of A is still an element of A. We call the pair (Ω, A) a measurable space and the elements of A measurable sets. If Ω is a topological space, then the smallest σ—algebra containing all open sets is called the Borel σ—algebra. We denote it by β(Ω) or simply by β, when it is clear from the context. A map μ : A t—» R U {+00} is called a positive measure if for any sequence of disjoint sets An £ A μ(υη>ιΑη) = Υ^μ{Αη) π>1 A positive measure is σ-finite if Ω is the union of a (countable) sequence of measurable sets of finite measure. The σ—algebra A is 1 Polish space is the name given by Bourbaki to a topological space which is metrizable in such a way that it becomes complete and separable. This space is particularly often discussed in mathematics. Opial wrote: "...this touching tribute to the Polish School of Mathematics testifies most expressively to its merits and to its status in mathematics."
8.1. Measurable Set-Valued Maps 307 complete (or, to be more precise, μ—complete) if for every set A G A satisfying μ{Α) = 0 any subset A\ с A is an element of A. For instance for every open (or closed) subset Ω С Rn, the set of all Lebesgue measurable subsets of Ω is complete (with respect to the Lebesgue measure.) The Lebesgue measure is also σ—finite. In summary, we shall say that the triple (Ω, Α, μ) is a complete σ—finite measure space if μ is a positive σ—finite measure such that A is μ—complete. Consider a complete separable metric space X. A single-valued / : Ω ι—» X is said to be measurable (.Α-measurable, to be more precise) if for every open subset О С X, f~1(0) G A (or equivalently for every closed subset С С X, /-1(C) € A.) We say that a map is simple if it takes only a finite number of values. A simple map is measurable if and only if for every χ £ X, f~1 (x) G A. The following equivalences provide in many instances convenient tools to check measurability of a map: ' i) f is A — measurable ii) f is the pointwise limit of simple measurable maps in) f is the uniform limit of measurable maps assuming a countable number of values One immediately deduces that the pointwise limit of measurable maps is measurable. Measurable set-valued maps with closed images are defined in a similar way·. Definition 8.1.1 Consider a measurable space (Ω, A), a complete separable metric space X and a set-valued map F : Ω ~> Χ with closed images. The map F is called measurable if the inverse image of each open set is a measurable set: for every open subset О С Χ, we have F~l{0) := {ω € tt\F(uj)nO φ 0} € A Remark — The notion of measurability that we use here is sometimes called in the literature weak measurability: a set-valued map is weakly measurable if inverse images of open sets are measurable, in comparison to strong measurability: inverse images of closed sets are measurable. However in the framework we shall deal with (complete σ-finite measure spaces),
308 8- Measurability and Integration of Set-Valued Maps. these two notions do coincide. That is why we simply use the word "measurable." □ The domain of a measurable map is measurable as well as its ■ complement: {ω£Ω| F{uj) = 0}. A measurable set-valued map has a measurable selection: Definition 8.1.2 Let (Ω, Λ) be α measurable space and X a complete separable metric space. Consider a set-valued map F : Ω ~> X. A measurable map /:ΩηΙ satisfying Vu>€ Ω, f{w) € F{lj) is called a measurable selection of F. Theorem 8.1.3 (Measurable Selection) LetX be a complete separable metric space, (Ω, Λ) a measurable space, F a measurable set- valued map from Ω to closed nonempty subsets of X. Then there exists a measurable selection of F. Proof — Fix a countable dense subset {xn}n>i of X. We construct a sequence of measurable maps Д : Ω н-► X, к > 0 taking values in {xn}n>i and converging uniformly to a selection / of F, so that / is measurable. We proceed by induction to construct such a sequence. Denote by d the distance in X. For every ω G Ω, let η > 1 be the о smallest integer such that F(uj)r\ В (xn, 1) Φ 0· We set /ο(ω) = χη· Then /ο(·) is measurable. Furthermore \/ш&П, d{f0{u), F(u>)) < 1 Assume that we already constructed measurable maps fk ■ Ω (->■ {xn}n>i> к = 0,...,m satisfying V0 < fc < m, \/ω € Ω, d(fk(u>),F(u>)) < ~ (8.2) and V0 < к < m-1, d(/feH,/fe+iH) < -lj (8.3)
8.1. Measurable Set-Valued Maps 309 For every n, define Sn = {ω€Ω| fm{u) = xn } The sets Sn are mutually disjoint and Un>i Sn = Ω. Furthermore (8.2) implies that V^ e 5n, F(lu) Π Β (ι»,2~m) ^ 0 Fix a; € Ω and let η be such that ω € 5η. Consider the smallest integer г such that FH П В (xn,2-m) Π 5 (xr> 2-(ш+1)) ^ 0 and set /m+i(o;) = жг. Then d(/mH,/m+iH) < 2~m + 2-(ш+1) < 2~m+1 Moreover d(/m+1 (07)^(0;)) < 2~(™+1) Clearly, this defines a measurable map fm+i ·' Ω ι-* {α;π}π>ι and (8.2), (8.3) hold true with m replaced by m + 1. From inequality (8.3) follows that for all ω G Ω, (fk(u))k>i is a Cauchy sequence in a complete metric space X. Let f{uj) denote the limit of /fe(o;). From (8.3) we deduce that /& converges uniformly to / and from (8.2) that d(/H,F(u;)) = 0 Thus / is measurable and for every ω € Ω, /(ω) € ί^ω). □ The next theorem provides several useful characterizations of measurability. For that purpose, let Λ ® В denote the σ—algebra generated by products Ax B, where A £ Λ and В € В (В is the Borel σ—algebra of a metric space X.)
310 8- Measurability ajad Integration of Set-Valued Maps Theorem 8.1.4 (Characterization Theorem) Let (Ω, Α, μ) be а complete σ—finite measure space, X a complete separable metric space and F : Ω ~> X a set-valued map with поп empty closed images. Then the following properties are equivalent: i) F is measurable ii) The graph of F belongs to A® B Hi) F^iC) € A for every closed set С С I < iv) F~1{B)eA for every Borel set В С X ν) For all x£X the map d(x,F(·)) is measurable vi) There exists a sequence of measurable selections (fn)n>i of F such that V ω £ Ω, F(u) = Un>i /η(ω) A countable family of selections satisfying the latter property is said to be dense. The equivalence of statements i) and ii) is fundamental in the framework of our graphical approach, since it characterizes measurable set-valued maps through the measurability of their graphs. Here too, this happens to be quite important. The equivalence of claims i) and vi), due to Castaing, is also very useful, since countability is basic in many measurability questions. We shall prove this theorem in the third section. But first we use it to provide some properties of measurable set-valued maps. 8.2 Calculus of Measurable Maps We develop in this section the calculus of measurable set-valued maps in complete σ-finite measure spaces. However the reader should be aware that some of the results below can be extended to more general cases. First, we note that semicontinuous maps are measurable:
8.2. Calculus of Measurable Maps 311 Proposition 8.2.1 (Continuity and Measurability) Consider а metric space Ω and a complete σ-finite measure space (Ω,,Α,μ). We assume that A contains all open subsets of Ω. Let X be a complete separable metric space and F : Ω ~> X a set-valued map with closed nonempty images. If F is upper semi-continuous (or lower semi-continuous), then F is measurable. Proof — If F is upper semicontinuous, then for every closed set С С X, F~1{C) is closed. We deduce from Characterization Theorem 8.1.4 that F is measurable. If F is lower semicontinuous, then for every open set О С X, -F_1(C) is open. D Theorem 8.2.2 (Convex Hull) Let(£l,A, μ) be a complete σ-finite measure space, X a separable Banach space and F : Ω ~> Χ a set- valued map with closed images. Then the map Ω Э ω ~> coF(u>) is measurable. To prove this theorem we need the following well known result: We recall that a map φ from Ω χ X to a metric space Υ is called Caratheodory, if for every χ ζ. Χ, φ(·, χ) is measurable and for every ω£ί], ¥>(ω, ■) is continuous. Lemma 8.2.3 Consider two complete separable metric spaces X, Y, a measurable space (Ω, Λ) and a Caratheodory map ψ : Ω χ X н-► Υ. Then for every measurable f : Ω ι—» X, the map ω ι—» φ(ω,ί{ω)) is measurable. We provide its proof for convenience of the reader. 1 Proof — Since /(·) is measurable, there exists a sequence of simple measurable maps fn : Ω (-► X converging pointwise to /. Then ω ι-+ (p(u),fn(u>)) is measurable. Since φ is continuous with respect to the second variable, Va) £ Ω, \υοαοφ(ω,/η(ω)) = φ(ω,/(ω)) Hence φ(ω, /(ω)) is the pointwise limit of measurable maps and thus, is measurable. D
312 8- Measurability and Integration of Set-Valued Maps Proof of Theorem 8.2.2 — It is not restrictive to assume that F.has; nonempty images. Characterization Theorem 8.1.4 implies that there exists a dense sequence of measurable selections (fn)n>i of F. Let Q+ denote all nonnegative rational numbers. Consider the set Λ := |(Ai>...,A„)JA1 € Q+> ]Γλ; = 1, η > ll and the countable family of maps J2f=i Xifi, where (λι,..., λπ) G Λ. By Lemma 8.2.3 these maps are measurable selections of coF. They are also dense in coF. Characterization Theorem 8.1.4 ends the proof. □ Theorem 8.2.4 (Union and Intersection) Let(Q,A, μ) be α complete σ-finite measure space, X a complete separable metric space and Fn : Ω ~> Χ set-valued maps with closed images. Then the maps Ω Э ω ~> ω ~> Η(ω) := Q Fn(uj) π>1 π>1 are measurable. Proof — Fix an open set О С X. Then G~\0) = {^Ω| |J(FnHnO)^0} = \JF-\0) π>1 π>1 To prove the second statement observe that Η has closed images and Graph(tf) = Γ) Graph (Fn) 71>1 By Characterization Theorem 8.1.4, Graph (Fn) belongs to Λ®Β and so does the graph of H, which is measurable, thanks to the same theorem. Π Theorem 8.2.5 (Lower and Upper Limits) Let (Ω,Α,μ) be a complete σ-finite measure space, X a complete separable metric space and Fn : Ω ~> Χ, η > 1 set-valued maps with closed images. Then the maps Ω Э ω ~> LiminfTl_KX1i;iTl(a;) & Ω Э ω ~> Limsupj^oo-Fn^)
8.2. Calculus of Measurable Maps 313 are measurable. Consequently, if for every ω € Ω, the images Fn(u) converge to a subset F(uj), the set-valued map F (called the pointwise limit) is measurable. We need the following Lemma 8.2.6 Consider a measurable space (Ω,Α), complete separable metric spaces Χ, Υ and let g : Ω χ X н-► Υ be a Caratheodory map. Then g is A® B-measurable. Proof — Indeed it is enough to show that there exists a sequence of maps gn : Ω χ Χ ι-> Υ measurable on A ® B, converging pointwise to g. Consider a dense sequence xk € X, к > 1. Fix (ω, χ) € Ω χ Χ. For every η > 1, let k be the smallest integer satisfying о ( 1 X € В \Хк, - \ η and set дп(ш,х) = д(ш,Хк). Then gn converge pointwise to g. Furthermore Vy € Yk := В (zfc'-J \ U B \Xm'n)' 5n^'y) = ^(ω,ζ*;) ^ ' m<k ^ ' But Un>i ^ic = -^ and therefore for each η, ^π is A® B-measurable. D Proof of Theorem 8.2.5 — By Lemma 8.2.6 and Theorem 8.1.4 the map (ω, χ) ι-> d(Fn(u), χ) is A ® 23(X)-measurable. Thus for every ε > 0, Graph(S(Fn(-),£)) := {(ω,χ) € Ω χ X\ d{Fn{u),x) < ε] С Л®В(.Х") This and Characterization Theorem 8.1.4 yield that the map ω ~> B(Fn(uj),e) is measurable Then the result follows from Theorem 8.2.4 and equations i) Liminfn^oo-FnM = Γ\^ιΌν>ιΓ\π>ν Β(Ρη{^)Λ) w ii) Limsupn_too-Fn(^) = f|jv>i Un>jv Ρη{ω) □
314 8- Measurability and Integration of Set-Valued Maps Definition 8.2.7 Consider metric spaces Χ, Υ and a measurable space (Ω,Α). A set-valued map G : Ω χ Χ ~> Υ with closed values is called a Caratheodory map if for every χ £ X the map ω ~> G(u>, x) is measurable and for every ω G Ω the map χ ~> G(u>, x) is continuous. Theorem 8.2.8 (Direct Image) Let (Ω, Α, μ) be a complete σ-finite measure space, X a complete separable metric space and F : Ω ~> X a measurable set-valued map with closed images. Consider a Caratheodory set-valued map G from Ω χ Χ to a complete separable metric space Υ. Then, the map Ω Э ω ~> (?(ω,ί(ω)) is measurable. In particular for every measurable single-valued map ζ : Ω ι—» X, the set-valued map ω ~> G(lj, ζ(ω)) is measurable and for every Caratheodory single-valued map φ from Ω χ X to Υ, the set-valued map Ω Э ω ~> φ(ω, F(u)) is measurable. Proof — It is not restrictive to assume that F has nonempty- images. Prom Characterization Theorem 8.1.4, there exists a dense sequence (fn)n>i of measurable selections of F. We claim that the map ω ~> G{u>, fn{u)) is measurable. Indeed consider a sequence of measurable simple maps fnk from Ω to X, converging pointwise to fn when к —> oo. Then, since fnk are simple, for every к the set-valued map ω ~» ΰ{ω, fnk{u)) is measurable. On the other hand, since G(u>, ·) is continuous, \f ω € Ω, Limfe^ooG^/nic^)) = G(w,/nM) and from Theorem 8.2.5, we deduce that this limit is again a measurable map. On the other hand, since G is continuous with respect to the second variable and (fn)n>i is dense, for every ω£ί] G(u;,F(u;)) = \jG(u,Jn(u,)) 71>1 Theorem 8.2.4 ends the proof. D
8.2. Calculus of Measurable Maps 315 Theorem 8.2.9 (Inverse Image) Consider α complete σ-finite measure space (Ω,Α,μ), complete separable metric spaces X, Y, measurable set-valued maps F : Ω ~> X, G : Ω ~> У with closed images. Let g : Ωχ Χ ι-> У be a Caratheodory map. Then the set-valued map Η defined by Η (ω) := {ieF(w)| д(и,х) € G(w)} is measurable. Consequently, if VW€fi, g(Lj,F(tj)) П G(w) ^ 0 ΐ/ien i/iere eaasis α measurable selection f of F such that for every ωξί], g(uj,f(uj)) belongs to G(uj). Proof — Observe that the map Η has closed images. Define the single-valued map φ: ΩχΧ\-+ΩχΥογ φ(ω, χ) = (ω, д(ш, х)). Then, by Lemma 8.2.6, VS € B(Y), {(ω,χ)\ g(u,x) € B}ei®B(I) Observe that Graph(tf) = Graph(F)n^1(Graph(G)) By Characterization Theorem 8.1.4, Graph(F) e A®B{X) к Graph(G) e A®B{Y) On the other hand for every Ay. В £ A® B(Y), φ-^ΑχΒ) = {(ω,χ)\ д(ш,х) £ В} П{АхХ) £ А®В{Х) Hence for every С С Α®Β{Υ), φ'1 {С) е А®В(Х). Thus Graph(tf) belongs to A ®B(X) and from Characterization Theorem 8.1.4, we deduce that Η is measurable. Finally the Measurable Selection Theorem implies that it has a measurable selection, whenever Η has nonempty images. □ As a consequence, we obtain the very useful
316 8- Measurability and Integration of Set-Valued Maps Theorem 8.2.10 (Pilippov) Consider α complete σ-finite measure space (Ω,Α,μ), complete separable metric spaces Χ, Υ and a measurable set-valued map F : Ω ~> Χ with closed nonempty images. Let g : Ω χ Χ ι—> Υ be a Caratheodory map. Then for every measurable map h : Ω ι—» Υ satisfying h{uj) G g(u>, -F(o;)) for almost all ω G Ω there exists a measurable selection ${ω) G F(u) such that h{uj) = д(ш, f(u>)) for almost all ω G Ω Marginal functions and maps are measurable under the following conditions: Theorem 8.2.11 (Marginal Map) Consider a complete σ-finite measure space (Ω,Α,μ), complete separable metric spaces X, Y, a measurable set-valued map F : Ω ~> Χ with closed nonempty images and a Caratheodory function f : Ω Χ Χ н-► R. Then the marginal function ν : Ω ι—» R U {—oo} defined by Vw£ii, ν(ω) := inf f(uj,x) is measurable. Furthermore, the marginal map R defined by Vwefi, R{u) := {x€F^)\f&,x)= inf }{ш,у)} yeFH is also measurable. We recall the following: Lemma 8.2.12 Consider a measurable space (Ω, Λ). Then for every sequence of measurable real-valued functions fn : Ω >—» R, the extended function ω н-> f{u) := inf/n(u;) G RU{-oo} has a measurable domain of definition and is measurable on it. A similar statement holds true for s\iOn>i fn.
8.2. Calculus of Measurable Maps 317 Proof of Theorem 8.2.11 — Let us consider a dense family of measurable selections (gn)n>i of the map F. Since the maps /(ω, ·) are continuous, we infer that Vw € Ω, ν{ω) = inf f{u>,gn(uj)) n>l Since the maps ω ι-> f(uj,gn(uj)) are measurable by Lemma 8.2.3, we deduce from Lemma 8.2.12 that ν is also measurable. Theorem 8.2.9 on inverses of measurable maps imply that the marginal map R is measurable, since R{u) = {x € F{u;) \ί{ω,χ) = ν{ω)} and because ν is measurable. Π Corollary 8.2.13 Let us consider α complete σ-finite measure space (Ω,Α,μ), a complete separable metric space X, a measurable set- valued map F : Ω ~> Χ with closed images and measurable single- valued maps / : Ω ь> Χ, ρ : Ω н-> R+. Then the following maps are measurable: 1. — the map ω ~> Β(f (ω), ρ(ω)) 2. — the distance function ПЭши <ί(/(α;), -F(o;)) 3. — the projection map Ω Э ω ~> Πρ(ω)(/(α;)) defined by • nFH(/H) := {x € FM| φ;,/Η) = d(f(u),F(u))} Consequently if for every ω£Ω, Πρ(ω)(/(α;)) φ 0, ί/ien ί/iere exists a measurable selection д(ш) G F(u) such that d(f(u>),g(u>)) = d(f(u),F(u>)) Proof— Consider the function д(ш,х) = d(x,f(uj)). By Lemma 8.2.3 it is measurable in ω. Since it is also continuous in χ we may use Theorem 8.2.9 with F = X, G{uS) = [0, ρ(ω)]. To prove the second and third statements, we apply Theorem 8.2.11 to the function (u,x)i-*d(x,f(uj)) □ Theorem 8.2.14 (Support Function) Let us consider a complete σ-finite measure space (Ω, Λ, μ) and let X be a separable Banach
318 8- Measurability and Integration of Set-Valued Maps- space, F : Ω, ~> Χ be a measurable set-valued map with nonempty closed images. Then it has measurable support functions: for every pel*, the function ω ι-> a{F{uS),p) := sup < p, f > is measurable /eF(«) The converse statement holds true if the dual of X is separable and images of F are convex and bounded. Proof — Theorem 8.2.11 implies that the support functions a(F(-),p) are measurable. Assume next that for every ρ € X*, a(p,F(·)) is measurable, that X* is separable and F has closed convex and bounded images. To prove the last statement is enough to verify that for all x, the map ω ι-> d(x, F(uj)) is measurable. We first observe that since F{uj) is bounded, its support function a(F(u>), ·) is continuous. Let pn e Χ*, η > 1 be a dense set of points of the unit sphere of X*. By our assumption, for every η > 1, a(F(-),pn) is measurable. Fix χ £ X. Then, using that F{uj) is closed and convex, we get d(x,F{uj)) = d(0,F{uj)-x) = - inf Μΐ<ι a{F{u) - x,p) = suPHpIL^iCCp-e) -σ(-ΡΉιίΟ) = supn>i((pn,a;) -a{F{u),pn)) Hence d(x,F(·)) is the supremum of measurable functions. Consequently it is measurable. Π The following result is very useful to study relaxation problems in control theory. Theorem 8.2.15 (Caratheodory Representation) Consider a complete σ-finite measure space (Ω,Α,μ), a measurable set-valued map G : Ω ~> Rn with nonempty closed images and a measurable selection /(ω) G co(G(o;)). Then there exist measurable functions \k ; Ω t—» R+ and measurable selections fk{u) € G(u>), к = 0, ...η such that /M = X>fe(u;)/fc(u;) & Σλ*Μ = l fe=0 fe=0
8.3. Proof of the Characterization Theorem 319 Proof — Let Sn+1 denote the set of all points (λ0, ■··, λη) € R++1 such that 2fc=o λ& = 1. Consider the continuous map h : R^+1 x (Rn)n+1 ^ Rn defined by η h(Xo,...,Xn,xo,...,xn) = У^Л^ж^ fe=o and the measurable set-valued map F(u) = Sn+1 χ (G(u;))n+1 By the Caratheodory theorem, /(a;) € /i(-F(a;)). We conclude the proof by applying Theorem 8.2.9 to this set-valued map F with g(u>, x) = h(x) and the map G equal to /. Π 8.3 Proof of the Characterization Theorem Characterization Theorem 8.1.4 is the consequence of both a Theorem due to Castaing and a classical result on the projection of certain classes of Borel subsets. We show first that a map is measurable if and only if it has a countable dense subset of measurable selections. Theorem 8.3.1 (Castaing) Let (Ω,Α) be α measurable space, X a complete separable metric space and F : Ω ~> Χ be a set-valued map with поп empty closed images. Then the statements ί), υ), νϊ) of Characterization Theorem 8.1.4 are equivalent. Proof 1. — We begin by proving that г) =Ф г/г). Consider a dense countable subset {χη}η>ι οι X. For all n, к > 1 define the set-valued map Gnk : Ω ~> Χ by Gn*M = ί Р{Ш)П °B {Xn' ^ if Ρ{ω) П °B {Xn'^ Φ 0 [ F{yS) otherwise
320 8- Measurability and Integration ofSet-Valued Maps and set Fnk(u>) = Gnk{uj). Then Fnk has nonempty closed images and is measurable. This follows from the very definition of Gnk and equality' for every open О С X, F^(0) = G~l(0) Hence, by Theorem 8.1.3, Fnk has a measurable selection'/^. It remains to verify that for every ω £ Ω, the values /nfc(^) are dense in F(u). Fix χ £ F(u>) and ε > 0. Let к > 1 be such that 1/fc < ε/2 and η such that <ί(α;π, ж) < l/к. Hence F(u>)nB(xn,l/k) φ 0 and /nfc(^) € B(xn,l/k). This yields d{fnk(b>),x) < d(fnk{uj),xn) + d(xn,x) < ε Since ε > 0 is arbitrary, the proof of this implication ensues. 2. — We prove now that vi) =*> v). Let /π(·) be as in vi). Fix χ £ X and let d denote the metric of X. Then for every n, the function ω t—» ci(a;, fn{u)) is measurable (thanks to Lemma 8.2.3.) Consequently, the function ω η-► (ίίχ, F(uj)) = inf d(x, fn(uj)) ri>l is also measurable and v) follows. 3. — Finally assume that v) holds true. Then for every χ £ X and r > 0 the set {ω £ Ω\ d(x,F(u)) < r} = F'1 (в (s,r)J £ Λ. Fix an open set О С X. Since X is separable, О is a countable union о of balls Β {χηιΤ·η). Hence F~l{0) = Un^iF-1 (в (xn,rn)\ £ Λ The proof is complete. D To prove the rest of Characterization Theorem 8.1.4. we need a very useful projection property enjoyed by complete measure spaces:
8.3- Proof of the Characterization Theorem 321 Theorem 8.3.2 (Measurable Projection) Let (Ω, Α, μ) be α complete σ-finite measure space, X a complete separable metric space and G € A®B(X). Then its projection is measurable: 7rn(G) := {ω£Ω\3χ£Χ, (ω,χ)€ΰ} € A The proof can be found in [91] ? Proof of Characterization Theorem 8.1.4 — We already know г) Φ=Φ υ) <=$■ г/г). It is clear that iv) =Ф Hi). To show that Hi) =*> г), fix an open subset О С X and define the closed sets Cn = {x € X\d(x,X\0) > l/n} where d states for the distance. Then О = Un>i @η· Consequently F{u>) Π Ο φ 0 if and only if for some η > 1, Ε(ω) Г\Спф%. This yields F-\0) = (J F~\Cn) e A 71>1 and ends the proof of г). We show next that v) implies ii). Observe that Graph (F) = {{ω,χ)€ΩχΧ \ d(x, F(u/)) = 0 } By Lemma 8.2.6 the function (ω, χ) н-> d(x, -F(o;)) is A<g>B measurable and therefore {(ω,χ) £ Ω χ X I d(x, F(u)) = 0} e A® B Thus Graph(F) belongs to A®B. 2It is quite tempting to think that the projection (and therefore any continuous image) of a Borel set should be still a Borel set. This wishful thinking was made once by Lebesgue. Suslin, who had discovered that error, could not believe that the great Lebesgue would commit such a mistake. The discovery of Borel sets whose continuous image is not Borel started the intensive study of analytic sets, i.e., continuous images of Borel sets, which have many interesting properties. As "tartes Tatin", "Rocquefort" and "Sauterne wine", this example shows how frequently mistakes and errors can trigger unexpected advances in both everyday life and sciences.
322 8- Measurability and Integration of Set-Valued Maps We end the proof by assuming that ii) holds true and observing that for every Borel set В £ В we have F~\B) = 7rn(Graph(.F) П (Ω χ В)) But Graph(-F)n(fi χ Β) is an element of Л ® B. Therefore, using Theorem 8.3.2, we deduce iv). 8.4 Limits of Measurable Maps and Selections Let (Ω, S, μ) be a complete σ—finite measure space and X be a separable Banach space. For a measurable map / : Ω ι-> X such that ||/(·)|| belongs to L1(X; R, μ), denote by /Ω /άμ or /Ω /(ω)μ(άω) the integral of /. Recall that two measurable single-valued maps are equal almost everywhere if the set where they are different is of zero measure. We denote by ΙΡ{£ΐ]Χ,μ), 1 < ρ < oo the Banach space of (classes) of measurable maps /:ΩηΙ such that /Ω ||/||ρ <ίμ < oo. Consider a sequence of measurable set-valued maps Κη:Ω Э ω ~» Κη{ω) С X We associate with it the subsets ICn С Ι?(Ω\Χ,μ) (1 < ρ < oo) of selections, defined by /Cn := {x{·) € Lp(£l;X, μ) | for almost all ω € Ω, χ (ω) € Κη(ω)} The purpose of the next theorem is to compare the limits of the sets JCn and the sets of selections x(-) of the limits of the sets Κη{ω). Theorem 8.4.1 Let us assume that the set-valued maps Kn are measurable, have closed images and that ω ι—» supra>1 d(0, Kn{u)) belongs to ^(Ω,-,Χ,μ), where 1 < ρ < oo. Then {x(-) € ΙΡ{Ω\Χ,μ) | for almost alio;, χ(ω) € LiminfTl_tooi;irTl(a;)} - С Liminfn_Kjo/Cn С Limsup^^/Cn С {χ(·) € LP (Ω; Χ, μ) | for almost all ω, χ(ω) € Limsupj^oolfnC't')}
8.4. Limits of Measurable Maps and Selections 323 Consequently, if the subsets Κη{ω) have a limit Κ{ω) for almost all ω € Ω, then the subsets KLn converge to the subset Κ. := {x(·) € £Ρ(Ω;Χ,μ) | for almost alio;, χ(ω) € .Κ" (ω)} Furthermore, if the dimension of X is finite and if LP (Ω; Χ, μ) is supplied with the weak topology, then σ - Limsupn_too/Cn С {χ{·) € 1^(9,;Χ, μ) | for a.a. ω, χ(ω) G co(LimsupTl_tooi;!rTl(a;))} Proof— Let χ(·) belong to the first subset. Then the functions απ(·) defined by αη{ω) := ά(χ(ω),Κη(ω)) are measurable and converge to 0 almost everywhere. Since for almost all ω € Ω, αη(ω) < ||α:(ω)|| + sup<i(0, Κη(ω)) π>1 and since the right-hand side of this inequality belongs to 77(0; R, μ), we deduce from Lebesgue's Theorem that the functions απ(·) do converge to 0 in 77(0; R, μ). Let к G 77(0;R, μ) be a function with strictly positive values. Corollary 8.2.13 allows us to choose a measurable selection zn{·) of Kn(·) such that \\χ(ω) — ζη(ω)\\ < αη(ω) + к(ш)/п It belongs to LP(Ω; R, μ) since for almost all ω € Ω,, \\ζη(ω)\\ < ||α:(ω)|| +αη(ω) + к(ш)/п Therefore ζη(·) belongs to 1Сп and converges to x(-) in 77(Ω;Χ,μ), i.e., x(-) does belong to the lower limit of the subsets K.n. Let us choose some x(-) in the upper limit of the subsets K.n. Then there exists a subsequence of elements ζηι(·) oiiCni converging to x(-) in 77(Ω;Χ,μ). It has a subsequence converging almost everywhere to x(-) and consequently, for almost all ω, χ (ω) belongs to the upper limit of the subsets Κη{ω). Finally, assume that x{·) -belongs to the sequentially weak upper limit of the subsets /Cn. It is a weak limit in ΙΡ(Ω;Χ,μ) of a
324 8- Measurability and Integration of Set-Valued Maps subsequence of functions zn{·) € lCn and also, thanks to Mazur's Theorem, the strong limit of convex combinations υη(·) of elements of the sequence zn{·). Then a subsequence (again denoted by) vn converges almost everywhere to x{·). We conclude as in the proof of Convergence Theorem 7.2.1 that for almost all ω e Ω, χ{ω) G со {L\ms\XOn_^oaKn{u)) □ 8.5 Tangent Cones in Lebesgue Spaces Let (Ω, S, μ) be a complete σ—finite measure space and X be a separable Banach space. Let us consider a measurable set-valued map Κ:Ω Э ω ~> Κ (ω) С Χ We associate with it the subset /С С 1^(0; Χ, μ) of selections defined by /C := { s(-) € Ι^Ω;*, μ) | for almost all ω e Ω, χ(ω) € АГ(а>) } We shall characterize the contingent and adjacent tangent cones to /C in terms of the tangent cones to the subsets Κ (ω). Theorem 8.5.1 Let us assume that the set-valued map К is measurable and has closed images. Then for every χ £ 1С, the set-valued maps: П Э ω ~> Τκ{ω){χ{ω)) к Ω Э ω ~> ТЬК{ш){х{ш)) are measurable. Furthermore |t/(·) € £Ρ(Ω; Χ, μ) | for almost alia;, υ (ω) € Γ^(ω)(χ(ω))} - С Г£(я(·)) С Гк(х(0) С |г/(·) € υ^Ω,-,Χ,μ) | for almost alia;, υ(α>) € Τκ{ω){χ{ω))} This theorem and its numerous applications motivate the introduction of adjacent tangent cones and derivable sets.
8.5. Tangent Cones in Lebesgue Spaces 325 Proof— Fix χ £ 1С and ω ζ. Ω. Let Q+ denote the set of all strictly positive rationale. It is countable and from the very definition of the tangent cones, we can write г*м(.м) =п4 U *м-«сЛ aeQ+ V^e]o,a]nQ+ / and n>0 \<*eQ+ ^e]o,a]nQ+ v " ri / j We deduce the first assertion from Theorem 8.2.4. To prove the second statement, consider υ(·) in the first subset. We have to prove that when h > 0 goes to 0, there exist maps Vh(·) € ΙΡ{Ω\Χ,μ) converging to υ(·) such that for almost all ω € Ω, χ (ω) + hvh{u>) € Κ (ω) Let us set: ак(ш) := άΙυ(ω), —-—-—^j Functions ah are measurable and converge to 0 almost everywhere because for almost all ω G Ω, υ(ω) belongs to the adjacent cone to Κ (ω) at χ (ω). Since for almost all ω, α^(α;) < ||υ(ω)|| and υ € 77(0; R, μ), by the Lebesgue dominated convergence theorem, the functions α/ι(·) do converge to 0 in Ζ^Ω;!?., μ). Fix к G 1^(0; R, μ) with strictly positive values. Corollary 8.2.13 yields that for some Zh € /C Vwsfi, dft/H, ^H-gH^ < αΐι{ω) + ftjfc(a;) We define now the maps υ/ι(·) by Vh{bJ) ■= {zh{u)-x{uj))/h They are measurable, satisfy Vh{w)-v(u)\\ < ah{uj) + hk{uj)
326 8- Measurability and Integration of Set-Valued Maps and thus, converge to v(-) in 1^(0; Χ, μ) since α^(·) converges to 0 in Ζ^(Ω;R, μ). We infer that υ(·) belongs to Г^(ж(·)) because for almost all ω € Ω, ж (ω) + hvh(u) & Κ (ω) Let us choose next some v(-) in the contingent cone to the subset /C. Then there exist sequences hn > 0 and υπ(·) converging respectively to 0 and to x(·) in LFifl; X) and satisfying for almost all ω G Ω, ж(а>) + hnvn{u) & Κ(ω) Then a subsequence (again denoted) υπ(·) converges almost everywhere to υ(·) and consequently, for almost all ω, ν (ω) belongs to the contingent cone to the subset Κ (ω) at χ(ω). Π As an obvious consequence, we obtain Corollary 8.5.2 Let us assume that the set-valued map К is measurable and has closed images. Let x(·) G /C. If for almost every ω£ί], the subsets Κ (ω) are derivable at χ(ω), so is 1С at χ and Tk{x) = {«(■) € ΣΡ(μ·,Χ,μ) | for almost alia;, υ (ω) € Тщы)(х(ш))} 8.6 Integral of Set-Valued Maps Let (Ω, Α, μ) be a complete σ—finite measure space and X a separable Banach space supplied with the norm ||-||. In this section we investigate some properties of the integral of a set-valued map jP from Ω into closed nonempty subsets of X. We denote by Τ the set of all integrable selections of F: Τ = {/ € Ll{Q;X, μ) | /(ω) € F{ui) almost everywhere in Ω} A set-valued map F : Ω ~> Χ is called integrably bounded if there exists a nonnegative function к G £*(Ω; R, μ) such that F{oj) С к(ш)В almost everywhere in Ω In this case every measurable selection of F is an element of Τ thanks to Lebesgue's Theorem. Aumann did suggest definition of the integral of a set-valued map in the following way:
8.6. Integral of Set-Valued Maps 327 Definition 8.6.1 The integral of F on Ω is the set of integrak of integrable selections of F: jf ity := {ja№\ /e^} The integral is convex whenever F has convex images and we shall prove later that its closure is still convex even when the images of F are no longer convex. It is also clear that V λ € R, / ΧΡάμ = λ / Ράμ Jo. Jo. Proposition 8.6.2 Let us consider measurable, integrably bounded set-valued maps F, Fi : Ω ~> X, i = 1,2 with nonempty closed images and set G{u>) := Fi{uj) + F2(u>). Then 1. - Ι^σΐμ = /n Fim + /n F2dM S- — Ιη^δΡάμ = со J^ Ράμ 3. - Vj, £ Γ, σ(/„%ρ) = /ησ(ίΉ,;ρ)μ(ώχ;) ^. — ///or some ж € /n -Fefyi. and ρ € -X"* < ρ, ж > = <т ( / Ράμ, ρ 1 ΐ/ien /or every f £ Τ satisfying χ = /n /ίΖμ, эде obtain for almost all ω € Ω, < ρ, /(ω) > = a(F(a;),p) Proof— Theorems 8.2.8 and 8.2.2 imply that G and coF are measurable. To prove the first equality it is enough to show that / βάμ С / Fia> + / F-ιάμ JQ. JO. Jn since the other inclusion is obvious. Consider a measurable selection goiG and let (fin)n>i be a dense sequence of measurable selections of Fi (which exist by Characterization Theorem 8.1.4.) Then (/in + /2m)n,m>i is a dense sequence of measurable selections of G. Set Gnm(u>) = {/ϋ(ω) + /2jH | 1 < i < n,: 1 < j < m]
328 8- Measurability and Integration of Set-Valued Maps Then Vwefi, lim d{g{u),Gnm{u)) = 0 п-юо.т-юо Thanks to Corollary 8.2.13, there exists a measurable selection gnm of Gnm satisfying HflH - flnmMH = d{g{u),Gnm{u)) Thus gnm(u) € -Fi(o;) + i^C^). Applying Theorem 8.2.9, we obtain measurable selections fnm of i; such that f^m + f%m = gnm. This yields d[ / 9ηπι<1μ, / -Fidyu + / Ρζάμ) = 0 \^Ω ./Ω Jn J By taking the limit when n, m —> oo, the first statement ensues. It is clear that со / Έάμ С / ΈδΡάμ Jn Jo. To prove the converse inclusion, fix к € L^fijR, μ) with strictly positive values and a measurable selection g of coF. Recall that Theorem 8.2.14 states that the support function ω ι-> a(coF(u>),p) = a(F(u>),p) is measurable and observe that for every ρ G X* (p, / gdμ) < / σ(οδΡ(ω),ρ)μ(άω) = / σ(Ρ(ω),ρ)μ(άω) Fix e > 0 and set ψ(χ) =< ρ, χ >. Theorem 8.2.9 on the measurability of inverse images applied with G(w) = W(F(u;),p)-ek(u>),a(F(u>),p)] yields that there exists a measurable selection / G Τ such that <P./M>> a{F{u),p)-ek{u) Consequently / σ(Ρ(ω),ρ)μ(άω)< / (< ρ, f > +ε^άμ < σ ( / Fd^p j+e \\k\\Li Jn Jci \Jn J
8.6. Integral of Set- Valued Maps 329 Since e > 0 and ρ € X* are arbitrary, using the separation theorem, we end the proof of the second and third statements. Fix x, f as in the claim 4. From the third statement σ{Ρ{ω),ρ)μ{άω) = σ{ / Ράμ, ρ) = f < p,J > άμ Consequently we deduce that j (σ{Ρ{ω),ρ)-<ρ,~]{ω)>)μ{άω) = 0 But a(F(u>),p) >< p, f(u>) >. This achieves the proof. □ The famous Lyapunov's theorem on convexity of the range of a vector measure implies that the integral of a set-valued map is convex when X is a finite dimensional vector-space. Recall that a set A £ Λ is called an atom (for the measure μ) if μ(Α) > 0 and for every measurable subset Αι с Α, μ{Α{) is equal to either 0 or μ(Α). A measure is nonatomic if Λ does not contain atoms. Dirac and discrete measures are atomic. The Lebesgue measure is nonatomic. A point ζ of a convex set К is called extremal if there is no x, у € К and 0 < λ < 1 such that ζ = λχ + (1 - X)y. We denote by ext(K) the set of all extremal points of K. When X = Rn, then the integral of any set-valued map is convex (even when the values of F are not convex) and is closed when the map is integrably bounded: Theorem 8.6.3 (Convexity of the Integral) Let F : Ω ~> Rn be a measurable set-valued map with nonempty closed images. If μ is nonatomic, then 1. — The integral Jn Ράμ is convex and extremal points of cd (Jn Ράμ) are contained in /n Ράμ. 2. — If a sequence Xk € Jn Ράμ, к = 1,2,... converges to an extremal point χ of со (J^ Ράμ), then every sequence fj. G Τ such that fkάμ = xk
330 8- Measurability and Integration of Set-Valued Maps converges in L\(Ω; Rn, μ) to some f G .F satisfying /n /<ίμ = ж. /η particular-for all χ G ext (со J^ Ράμ) there exists a unique f G Τ with χ = Jn /ίΖμ. // гп addition F is integrably bounded, then the integral of F is also compact. When the dimension of the Banach space X is no longer finite, the integral may no longer be convex, but its closure is convex: Theorem 8.6.4 (Convexity of the Closure of Integral) LetX be a separable Banach space and F : Ω ~»· X be a measurable set- valued map with nonempty closed images. If μ is nonatomic, then the closure of its integral is convex: 1. - !ηΡάμ = βδ(ΙηΡάμ). 2. — If χ G ext /n Ράμ, then f £ Τ satisfying /n }άμ = χ is unique. 3. — If F is integrably bounded, then Ι οδΡάμ = Ι Ράμ Furthermore when X is reflexive and F has convex images and is integrably bounded, then the integral Jn Ράμ is closed. Remark — It is not difficult to check that the third statement remains true if we assume that Τ Φ 0 and μ(Ω) < oo instead of the integrable boundedness of F. Π It is convenient to introduce the following Definition 8.6.5 Consiaer a map F : Ω, ~»· X. We say that f G Τ is an extremal selection of F if /n }άμ is an extremal point of со (fn Ράμ). We denote by Te the set of extremal selections: Te := { f G Τ | / ίάμ G ext (со ί Ράμ I Jo. \ Jn We deduce from Theorem 8.6.3 and the Caratheodory Theorem:
8.6. Integral of Set- Valued Maps 331 Corollary 8.6.6 Let F : Ω -^ Rn be α measurable integrably bounded set-valued map with nonempty closed images. If μ is nonatomic, then for every χ € / Ράμ there exist η + 1 extremal selections fk € Te and η + 1 measurable sets Ah £ А, к = 0,..., η, such that x = b\EXAjk)άμ where хлк is the characteristic function of A^. Proof— Fix x & fn Ράμ and let Xk > 0, Xk € ext (со /n Ράμ), к = 0,..., η be such that η η fe=0 fe=0 which exist by the Caratheodory theorem. Theorem 8.6.3 implies the existence of fk € Te such that /n fkd-l1 = £fc for every к = 0,..., п. Then the set-valued map G(u) := {Λ(ω)}^=ο,ι,...,η has closed nonempty images and is measurable. Since each Xk belongs to its integral, which is convex by Theorem 8.6.3, so does x. Hence there exists an integrable selection / := Sfe=oXAfc/fe of G such that χ = Ιηίάμ- Π Theorems 8.4.1 and 8.6.3 imply the set-valued version of the Lebesgue Dominated Convergence Theorem: Theorem 8.6.7 Let (Ω,,Α,μ) be a complete σ—finite measure space ana X a separable Banach space. Consiaer measurable set-valued maps Kn : Ω Э ω ^ Κη{ω) с X with closed images. 1. — If the function ω ι-> sup^j ά(0, Κη(ω)) is integrable, then I (Limmfn-too-Kn) άμ С Liminf^oo ( / Κηάμ
332 8- Measurability and Integration of Set-Valued Maps 2. —; -If the dimension of X is finite and ω ι-»· sup^j ||ΛΓπ(ω)|| is integrablej then ■ ·'■'- Limsup^^ ( / Κηάμ)·<Ζ I (Limsu^^^Kn) άμ \Jn J Jo. Therefore, under the last asumptions, if the subsets Κη{ώ) have a limit when η —> oo for almost every ω- G Ω, then 1лтп_юо ( / Κηάμ) = / (Limn_>oo-Kn) άμ \Jn J JQ. Proof — We begin by proving the first inclusion. An element ν of the integral of the lower limit can be written ν = fn }άμ where / is the integrable selection of the lower limit of the maps Kn{·). By- Theorem 8.4.1, it is the limit in Ll(Sl\ Χ, μ) of maps fn G JCn, whjch are integrable selections of the set-valued maps Kn(·). Hence ν is the limit of the integrals vn = / ίηύμ € / Κηάμ To prove the second inclusion, assume that ν := Km / fn (ω)μ(άω) G Limsup^oo / Κη(ω)μ(άω) where fnj G Knj. Since supn>0 ||ΛΓη(·)ΙΙ is integrable, the maps fnj(·) are integrally bounded. By the Dunford-Pettis Theorem, a subsequence (again denoted by) fnj(·) converges weakly to a map /(■), which belongs to the sequentially weak upper limit of the subsets K,n. The second statement of Theorem 8.4.1 implies that for almost all ω € Ω, /(ω) € со {Limsxxo^^^^K^u)) Therefore, by Theorem 8.6.3, we infer that ν € / со (Limsupn_tooi;!rn(a;)) μ(άω) = / Ι,ϊπιδηΡη^,χ,Κη^μ^) J9. Jn This completes the proof of the second inclusion. □
8.7. Proofs of the Convexity of the Integral 333 8.7 Proofs of the Convexity of the Integral We start with the more simple finite dimensional case. 8.7.1 Finite dimensional case We assume here that X = R" and F, μ verify all the assumptions of Theorem 8.6.3. There exist many proofs of the first statement of Theorem 8.6.3. We have chosen among them the one based on a lemma which gives an estimate of the norm ||/ — g\\Li for f,g G Τ in terms of the integrals of / and g. This lemma is intrinsically interesting since it has found several applications in control theory, for instance in the investigation of the Holder behavior of controls3. We associate with any ρ G Rn and f,g G Τ SP(f, g) ■= sup | / /άμ - у I У € / Ράμ к (ρ, J gdμ\ < (ρ, у) I Lemma 8.7.1 Let f and g belong to Τ. Then for all ρ G R" (ρ,/η9άμ) < (p>jafdv) => II/-Sb < 4n6p(/,fl) (8.4) Furthermore if (p,J ίάμ\ = <r(p,J Ράμ\ then Vs e ^, ||/-fl||bi < (2n-i)6p(/,fl) Proof — We first observe that II / II/-fllb ^ 2nsup / (f-g)dμ apaWJa AeAWJA Indeed, if fy denote the components of /, let us define Ai := {ω | /i(w) - 9ί(ω) > 0} (8.5) 3See for instance [171].
334 8- Measurability and iniegraiioB of Set-Valued Maps We observe that ii/-sIIli < Σ / ΐΛ-Λμμ = Σ / (λ-λ)^+Σ А „ (si-/i)^ Then inequalities / (fi ~ 9ί)άμ < \\f (/ - 9)άμ JAi \\JAi II / < sup / {f -ς)άμ AeA\\JA and / (βί-/ί)άμ < / (β-ί)άμ Jn\Ai Jfl\Ai II / < sup / (/ - 9)άμ imply estimate (8.5.) It remains to show that for any A e A and any ρ G X* satisfying (P. /n(/ ~ З)^) > 0, we have ί(ί-9)άμ < 26p(f,g) JA (8.6) in order to prove (8.4.) We note that we can write / (ί-9)άμ = / ίάμ-νΑ where yA := / £ίΖμ+ / ίάμ £ Ράμ JA JD. JA Jn\A J9. Therefore / /<*μ - 1/Λ + / ίάμ - yn\A = ίάμ- ^άμ (8.7) Consequently assumption (ρ, /Ω(/ — 9)άμ) > 0 and equation (ρ, Jn ίάμ - yA^i + (ρ, j^ ίάμ - yn\A^ = (ρ, j^f - g)d^j ^ (8.8) imply that either < p, Jn ίάμ - yA > or < p, Jn ίάμ - yn\A > is not larger than < ρ, /η(/ — 9)άμ >. Say for instance that p, / ίάμ-уЛ < (ρ, ίάμ- 9άμ) ι Λί / \ Jo. Jn /
8.7- Proofs of the Convexity of the Integral 335 Then and therefore p, j дdμ\ < (р,уА) U ~ 9)άμ = / ίάμ - JA II Λΐ УА < 5p(f,g) (8.9) (8.10) Otherwise P,j f^-yo.\Aj < (p,J ίΊμ- J 9άμ and interchanging the roles of A and Ω,\Α, we obtain / (/ - 9)Ίμ = / ίάμ On the other hand we always have УП\А < 5P(f,g) \\ί(ί-9)άμ < 6p(f,g) Equation (8.7) implies that ' \\JA(f - 9)άμ\\ = ΙΙ/η/φ-ϊ/ΑΐΙ < . < \\Ιηίάμ-Ιη9<Ιμ\\ + Ρηί<1μ-νίΐ\Αΐ < 26p(/,s) In summary, we proved inequality (8.6) and thus, the first statement of the lemma. To prove the second one, let ρ be such that p, J /άμ) = σ (ρ, Ι Ράμ Changing coordinates, we may assume without any loss of generality that ρ = (0,..., 0,1). Then Proposition 8.6.2 implies that /πΜ - gn(ui) > 0
336 8- Measurability and Integration of Set-Valued Maps almost everywhere in Ω. Hence in this case, proceeding as in the proof of inequality (8.5) we get ' llZ-flb < 2(n-l)supAeU||JA(/-fl)dM|| + ||/n(/- 9)άμ\\ ■ < (2n-l)supAe^||/A(/-fl)dM|| (8.11) Fix a measurable А с Ω and define уд, ym a as before. Both terms in the left hand side of (8.8) are nonnegative. Thus they are not greater than (ρ,/Ω(/— 5)^)· Hence (8.9) holds true and (8.10) follows. Consequently sup AeA fu-9№ JO. < W,g) and from (8.11), the second statement ensues. D Convexity Theorem 8.6.3 is the consequence of both the following Theorem 8.7.2 and the Lyapunov Convexity Theorem 8.7.3 we recall below. We begin by proving Theorem 8.7.2 Any extremal point of co(Jn Ράμ) belongs to Jn Έάμ. Furthermore iffy £ J7, к = 1,2,... are such that /n fkd-μ converge to an extremal point of со (fn Ράμ), then Д converges to a selection /€f ίη^(Ω;Κπ,μ). Proof— Set К :=сд (/Ω Ράμ) and let e € A" be an extremal point of K. Observe that to prove this theorem it is enough to show that every sequence /& G Τ such that lim к—юо / ίΗ<1μ = Jo. converges in ΐ}{Ω\ Rn, μ) to a selection / G Τ. We use Lemma 8.7.1 which allows derivation that it is a Cauchy sequence from the fact that the sequence of their integrals is a Cauchy sequence. Step 1. We claim that for every e > 0 therr ;> δ > 0 such that V/i,/2€.F with J ϊιάμ JO. = 1,2
8.7. Proofs of the Convexity of the Integral 337 we have ||/i - f2\\Li < e. Indeed it is enough to consider the case Κ φ {e}. Fix e > 0 and pick 0 < Ύ] < e/8n such that Q := ΚΠ(β + ηΣη~1) φ 0 where Σπ_1 denotes the unit sphere of Rn. Since Q is compact, so is coQ С В(е, η) Π Κ. The point e being extremal, it does not belong to coQ. Thus we can separate e from coQ: there exists ρ G R" with ||p|| = 1 such that (j), e) > σς>(ρ) := max{ < ρ, χ > | χ € coQ } Define the open halfspace Я := {χ € X\ <p,x>> aQ{p) } Then Hn(K\B(e,V)) = 0 (8.12) Indeed, if χ <Ξ Я П (К\В(е, ту)), then У '■= е+ и η ц(^-е) = (1-й ^ η ) е+ „ V ..χ \\х — е\\ \ \\х — е\\/ \\х — е\\ belongs to Q, so that < ρ, у >< σς>(;ρ). But <p,y>> (l- V_ A aQ(p) + ^e,,o-Q(p) = <tq(p) so that we get a contradiction. Therefore Η ПК С B(e, η). Since e belongs to the interior of Я, there exists δ > 0 such that J5(e, <5) С Я. Then КПВ(е,6) С Н ПК С Β(β,η) Consider /ι, /2 € ^ such that | fn ^άμ - e| < δ for i = 1,2. Then both /n /jifyt belong to Я" П Я so that they satisfy p, J ίίάμ\ > aQ{p)
338 8- Measurability and Integration of Set-Valued Maps Changing notations if needed, we may assume that P, J /гфЛ < (ρ,Ι ίιάμ) Therefore ' { У € In Ράμ | (p, Jn /2άμ) <<p,y>} < Οίν^ΙηΡάμΙσοίρΥ <<p,y>} С КПН By applying Lemma 8.7.1 with / = /i and д = /2 and using the inclusion Κ Π Η с В(е,η), we deduce that ' ||/ι-ΛΙΙ* < - 4nsup{||/n /ιίΖμ -y\\ | у e /n FdM, < ρ, /η /2^μ > < < ρ, у >} k <4nsup{||/n/idM-y|| | у&НпК] < 4η(2τ7) < ε and our claim follows. Step 2. We next claim that the extremal point e of К belongs to the closure of the integral of F: Vt7>0, (j Fdfi\ ηΒ(β,η) φ 0 Indeed suppose for a moment that it does not hold true for some r\ > 0. Taking r] > 0 small enough, we may assume that the set Q defined in Step 1 is nonempty. Let Η be as in Step 1. If J5(e, η) Π /η Ράμ = 0 we deduce from (8.12) that Я П / Ράμ = 0 Jo. and consequently ΗΠΚ = 0, because К is the convex hull of Jn Ράμ. This contradicts the fact that e G Κ Π Η. Step 3. Finally, we show that the extremal point e of К belongs to the integral of F: Indeed, by Step 2, it belongs to the closure of Jn Ράμ. Consider a sequence of elements Xk ■= Jn /ι^άμ of Jn Ράμ converging to e.
8.7. Proofs of the Convexity of the Integral 339 By Step 1, (fk)k>i is a Cauchy sequence in Ζ^Ω; Κπ,μ). Thus it converges to some / and has a subsequence (fk-)j>i converging almost everywhere to /. The values of F being closed, we finally obtain that for almost every ω G Ω, /(ω) G F(u>). Therefore / is an integrable selection of F satisfying /n /άμ = e, so that e belongs to the integral of F. □ We recall now Lyapunov's Convexity Theorem on the range of a vector valued measure which we need to prove the convexity of the integral in Theorem 8.6.3. Theorem 8.7.3 (Lyapunov's Convexity Theorem) We assume that μ is nonatomic and let f G Ι^(Ω; Rn, μ). Then the set u{A) := { ί ίάμ ) is convex and compact. The proof can be found in [45, Appendix C] and [8]. Proof of Theorem 8.6.3— Fix/jG.F, i = 1,2 and λ <Ξ [0, l]. We have to prove the existence of / £.!F satisfying λ / №μ + (1 - λ) / /2άμ = ί ίάμ Define the vector measure 1/:У1н Rn χ Rn by v(A) = ЦА№, jAhdv) (8-13) Then и is finite and nonatomic. Lyapunov's Theorem 8.7.3 implies that the range of и is convex. Since !/(0) = {0} & ι/(Ω) = (j^/ιΦ, jaf*dv) there exists AeA satisfying u{A) = λι/(Ω), i.e., λ / /ίάμ = / /ίάμ, λ / /2άμ = / /2£?μ Jn JA Jo. Ja
340 8- Measurability and Integration of Set-Valued Maps Therefore the map / = xaIi + Χω\α/2 is an integrable selection of F we were looking for. Prom Theorem 8.7.2, we know" that the integral /n Ράμ contains all extremal points of the convex set со$ς).Ράμ. This proves the second part of the first statement. Theorem 8.7.2 implies also the second assertion. Finally, if F is integrably bounded, then the set со Jn Ράμ is compact. Since /n Ράμ is convex and contains all the extremal points of its closed convex hull, we deduce from the Caratheodory theorem that / Ράμ = со ( / ΡάμJ = со ( j Ράμ) The proof is completed. □ 8.7.2 Infinite Dimensional Case In this section we impose on F the assumptions of the infinite dimensional Convexity Theorem 8.6.4. To prove it, we need the following extension of Lyapunov's theorem to the infinite dimensional case. Theorem 8.7.4 Let Υ be α separable Banach space. Consider an integrable map f G ^(Ω,-,Υ,μ) and define the map (vector measure) ν : Α ι—» Υ by v{A) = j /άμ J A If μ is nonatomic, then v(A) is convex and compact. Proof— Let fn G ^(Ω,-,Υ,μ), η = 1,2,... be simple maps converging to / in Ь1(0;У, μ). Define continuous linear operators Γ, Tn : L°°(fi; R, μ) ^ Υ, η = 1, 2,... by V5 € L°°(fi;R,^), Tn(g) = f 9/ηάμ, T(g) = f g/άμ which satisfy Urn ||Γη-Γ|| < lim ||/„ -/||Li = 0 П-ЮО 71—>CO "■"
8.7- Proofs of the Convexity of the Integral 341 Since fn is simple, the range of Tn is a finite dimensional space. Thus Tn is compact4 and the limit Γ of the Tn's is compact as well. To prove that u(A) is convex it is enough to check that for any x,y £ v(A) and λ € [0,1], λχ + (l - \)y € v(A). Fix such x, y, λ and let C,D £ Abe such that χ = / /d/i, у = / ίάμ Jc Jd By the Lyapunov Convexity Theorem, for every η { Тп(ха) I A € Λ} is convex Hence there exist АпёД η = 1,... such that Vn > 1, ATn(Xc) + (1-λ)Γ„(χβ) = Tn(XAJ Hence f ИЛ>) -(λζ + (1-λ)ι/)|| < \\T(XAn) - Tn(XAn)\\ + ||λΓη(χσ) + (l-\)Tn(XD) -(\T(Xc) + (l-\)T(xD))\\ k < ||ГП -ГЦ + λ ||Γ„ - ГЦ + (1 - λ) ||Γη - ГЦ =2 ||Γη - ГЦ converges to 0 since ||Γπ — ГЦ converges to 0. □ The above infinite version of Lyapunov's Convexity Theorem implies Theorem 8.6.4: Proof of Theorem 8.6.4 — The proof is similar to the one of Theorem 8.6.3. We have to show that for all /г € Τ, г = 1,2, ε > 0 and λ € [0,1], there exists / € Τ satisfying |λ / /κΖμ + (1-Х) f /2άμ - J /J < ε Define the vector measure ν : A (-► Χ χ Χ by (8.13.) By Theorem 8.7.4 the closure of the range of и is convex. Since !/(0) = {0} & !/(Ω) = (7/κΖμ, //2dM 4 A linear operator is compact if it sends the unit ball into a relatively compact set. The set of compact operators is closed in the uniform operator topology ([148, Lemma VI.5.3]) or [44, Chapter 10].
342 8- Measurability and Integration of Set-Valued Maps .we have for some A £ Λ λ / /ιάμ- / f-ίάμ + λ / /2£?μ - / /2Φ Ι ./Ω ./Α II ./Ω JA < ε Setting / = хл/ι + Χω\α/2> we deduce that ^Ράμ is closed and convex. To prove the second assertion, fix an extremal point χ of /n Ράμ and let /, д G .7-" be such that ж = /n /<ίμ = /n <7<ίμ. The space X being separable, there exists a sequence of pn G Χ*, η > 1, which separates points of X. Define An = {ω € Ω Ι <ρη,/(ω) > > <pn,g(u>) >} and set /1 = XA„/ + ΧΩ\Αη5 & /2 = XA„5 + Xn\Anf Then /1, /2 € J7. Observe that if μ(Αι) > 0, then \Pn> / hdH > \Pn4 ^άμ) and therefore /n /ι<ίμ ^ Jn /2^· On the other hand - / ίιάμ + - / /2£?μ = ж Since ж is extremal, we deduce that for every η, μ{Αη) = 0. Similarly the set A'n = {ω £ П\ <Pn,g&) > > <pn,/M >} has zero measure. Since (pn)n>i separates points of X, μ({ω € Ω |/(ω) ^ sM» < X>(An^«)) = 0 π>1 The third statement follows from the first one and Proposition 8.6.2. To prove the last claim, assume that X is reflexive and that F is integrably bounded and has nonempty closed convex images. Then the set Τ is weakly compact in a{Ll(Ω; Χ, μ), L°°(fi,·Χ, μ)) and the proof follows. □
8.8. The Bang-Bang Principle 343 8.8 The Bang-Bang Principle When in addition Ω is ал interval [a, b] С R and μ is the Lebesgue measure, then we have a Caratheodory type result for measurable selections known under the name of the Bang-Bang Principle: Recall that for a closed convex set К С Rn, a convex subset Q С К is called an extremal face of К if for every x,y € К and 0< λ< 1 \x + (1 - X)y € Q =Ф ж, у € Q Every extremal point is an extremal face. The dimension of the extremal face Q is equal to the dimension of the affine space spanned by Q. The dimension of an extremal point is zero. Definition 8.8.1 Consider a map F : [a,b] ~»· Rn. A selection f G Τ is called piecewise extremal if there is a partition a = to < ti < ... < tk = b such that Jt*+1 $άμ is an extremal point of со [Jt!+1 FduA. In the next theorem we assume that Ω = [α, 6], Λ is the σ—algebra of Lebesgue measurable subsets of [a, b] and μ is the Lebesgue measure. Theorem 8.8.2 (Bang-Bang Principle) Consider a measurable, integrably bounded set-valued map F : [a, b] ~»· Rn with nonempty closed images. Then 1. — Ja Ράμ is convex and compact. 2. — If χ belongs to an extremal face of ^Ράμ of dimension k, then there exists a piecewise extremal selection f G Τ such that x = Ja }άμ and the number of subintervals on which f is extremal can be made not greater than к + 1. The above result in particular implies that for every igj' Ράμ, there exist / G Τ and a partition of [a, b] in η + 1 subintervals a = to < ... < tn < tn+i = b
344 8- Measurabihty and Integration of Set-Valued Maps such that for every 0 < i < n, the restriction f\[tuti+1] is extremal for the restriction of F to [t{, i;+i] an(^ x — L· f^f1· Although the first assertion of the Bang-Bang Principle follows from Theorem 8.6.3, we obtain it directly as a by-product of the proof without any help of the Lyapunov Theorem. In this way the proof given below is self-contained. Proof of Theorem 8.8.2 — Since μ is the Lebesgue measure we shall adopt the usual notation dt for άμ and μ(άω). Observe that convexity and closedness of the integral Ja Fdt follows from the fact that every extremal face of со Ja Fdt is a subset of Ja Fdt. Thus our first statement would result from the second one. To prove it, we proceed by induction with respect to the dimension m of the face Q. We already know, thanks to Theorem 8.7.2, that for every measurable, integrably bounded set-valued map G : [a, c] ~»· BP with nonempty closed images, every extremal point of со (/^ Gdt) belongs to J£ Gdt (where с is any number larger than a and ρ > 1 is any integer.) Thus every extremal face of со (J^ Gdt) of dimension m = 0 is a subset of J^Gdt. Assume next that the above is valid for any such map G and every extremal face of dimension smaller than k. Fix a measurable, integrably bounded map G : [a, c] ~»· Rp with nonempty closed images such that a nonempty set Q is a fc-dimensional extremal face of со (J£ Gdt). Since Q is extremal, we obtain Q Π ext fed J ΰάμ) = ext Q (8.14) Denote by Q the set of all integrable selections of G. Pick an extremal point e of Q and let ge € Q be an (extremal) selection of G such that f£ge(t)dt = e. (It exists by (8.14) and Theorem 8.7.2.) Replacing G by G — ge, we may assume that e = 0 and ge = 0. Let Ρ be the vector space spanned by Q. We claim that V g € Q such that / g(t)dt € P, g(t) € Ρ a.e. in [a, c] Ja . (8.15)
8.8. The Bang-Bang Principle 345 Indeed it is enough to show that for every measurable set A, we have JA g(t)dt € Ρ {Ρ is the intersection of a finite number of hyper- planes.) Fix a measurable A and define 51 = xAg, g2 = (х[а,с]\д)з· Then 51,52 € Q and l£gi(t)dt + \fa92{t)dt = \£g{t)dt + ~o e Q Hence, because Q is an extremal face, /%ι(ΐ)ώ = f g(t)dt € Q С Р Ja J A This ends the proof of our claim. Consider the set-valued map Gi(t) = G(t) Π Ρ with nonempty closed images. For every τ G [a, c], set π(τ) = / Gidi Jo Then \/t < τ', ВД С π(τ') С Ρ and the map τ ~»· ΈοΠ(τ) is continuous, because Gi is integrably bounded. Fix жо € Q and let ί G [a, c] be such that xq € coH(t) and for every ί < ί, жо £ WTZ{t). ■ Then жо is a boundary point of coTZ(t). Consequently it belongs to an extremal face Q' of со (ja G\dt\ of dimension smaller than k. Since со (/J Gdt) С со (J^ Gdt) and Q' с Q, (8.15) imply that Q' is also an extremal face of со (Ja Gdt]. Thus there exist a = to < ti < ... <tk = t and a selection g of G such that /^ t/(i)di = жо and for all 0 < г < к - 1, j£+1 s(i)di is an extremal point of со /t*i+1 Gdi. Set * _ J 5(0 if ΐ€[α,ΐ] 90{l) ~ \ 0 otherwise Then 50 € £ and f£go(i)dt = x0. Furthermore, since ) Γ Gdt = со I Gdt + со / Gdt Ja Ja- Jt CO
346 8- Measurability and Integration of Set-Valued Maps Ji 9o(t)dt is an extremal point of cojf Gdt. □ Remark — Note that if xq is not an extremal point of Q, then we have a choice of e in the proof above and, therefore, such xq can be reached by, at least, two different piecewise extremal elements. D 8.9 Invariant Measures L· Poincare's Recurrence Theorem 8.9.1 Linear Extension of Set-Valued Maps Let X and Υ be two compact metric spaces and B(X) and B{Y) their Borel σ-algebras. We recall that the dual C*(X) of the space of continuous functions is isomorphic to the space of Radon measures on X and that a continuous single-valued /:1нУ can be extended to a continuous linear operator Τ from C*(X) on X to C*(Y) by the formula V^C*(I), \/BGB(Y), Τ(μ)(Β) := μ(Γ\Β)) This fact can be extended to set-valued maps F : X ~»· Y. We denote by V(X) CC*(X) the (weakly compact convex set) of probability measures on X. Definition 8.9.1 Let F : Χ ~»· Υ be α set-valued map. Denote by Τ, the linear extension of F — the set-valued map from V{X) to V(Y) defined in the following way: ν G V(Y) belongs to Τ(μ) if and only if VBeSQO, v(B) < μ{Ρ-\Β)) We extend it as a set-valued map from C*(X) to C*(Y) by setting 0 if /x is nonpositive, Τ(μ) := I {0} if/A = 0 μ{Χ)Τ{μ/μ{Χ)) if μ is positive
8.9. Invariant Measures and the Poincare Recurrence Theorem 347 Proposition 8.9.2 Consider compact metric spaces Χ, Υ and a closed set-valued map with nonempty values F : X ~»· У. Then Τ is a closed convex process with nonempty values. Furthermore, for any μ G V{X), ν belongs to Τ{μ) if and only if for every open subset OcY, ν{Ό) < μ{Ρ"ι{0)) The closed convex process Τ is a (set-valued linear) extension of F in the sense that for Dirac measures, 6y G Τ(δχ) if and only if у G F(x). If G : Υ ~»· Ζ is a closed set-valued map with nonempty values from Υ to a compact metric space Z, then the extension Η of the product Η := GoF contains the product of the extensions: QoT С Ή,- Proof — Let us consider a measure μ G V{X). The image Τ'(μ) is not empty, Indeed, by the Measurable Selection Theorem 8.1.3, F, being upper semicontinuous with closed images, is measurable, so that there exists at least one measurable selection / of F. Define Vf by the formula Vf{A) := μ($~~ι(Α))^ which is a probability measure. Since f-\A) с F~l{A), we infer that μ{Γ\Α)) < μ^-^Α)) so that Uf belongs to Τ(μ). We prove now that Τ(μ) can be defined as the set of measures ν € V{X) satisfying for every open subset О с У, ν{Ό) < μ{Ρ"ι{Ό)) . We first extend this formula to compact subsets К с Υ. Since the graph of F, and thus, of F~l, is closed and X is compact, F~l is also upper semicontinuous. We then know that for any neighborhood On D F~l{K), there exists an open neighborhood Mn D К satisfying F~l{Mn) С Оп. Let us choose open subsets On such that μ{Οη) \ μ{ρ-\Κ)) Hence inequalities v{K) < v(Mn) < μ{ρ-\Μη)) < μ{Οη) imply by going to the limit that v{K) < μ{Ρ~~ι{Κ)). Take now any measurable subset В G B(Y). There exists a sequence of compact subsets Kn с В such that v[Kn) / v(B)
348 8- Measurability and Integration of Set-Valued Maps Then inequalities HKn) < μ(Ρ-\Κη)) < μ(ρ-1(Β))_ imply that v{B) < μ(Ρ~1(Β)). Assume that у G F(x). Then 6y G Τ{δχ) since, for any open subset О CY, 6y(0) < 5x{F~l{0)). This is obvious when у £ О. If not, the left-hand side is equal to 1, and so is the right-hand side, because χ G F-\y) С F~\0) Conversely, if у £ F(x), there exists an open subset О Э у such that F(x)nO = 0, i.e., χ 0 F~l{0). ТЬеп^(б>) = 1ααάδχ ((F"1^))) = 0, so that δυ £ Τ{δχ). Formula Q ο Τ С Ή. is obvious as well as convexity of the graph oiT. It remains to prove that Graph(^r) is closed when the spaces' of Radon measures are supplied with the weak-* topology. For that purpose, consider a sequence of measures (μη, νη) G Graph(^") converging to (μ, ν) in the weak-* topologies of the duals C*{X) and C*(Y) respectively. It is sufficient to prove that the graph of the restriction of Τ to V{X) is weakly closed. Indeed, when the measures μη and vn G Τ(μη) are positive and converge weakly to μ and v, subsequences of the probability measures ~μη := μη/μη(Χ) and Vn ··= νη/μη{Χ) € Ρ(μη) converge weakly to probability measures Д and V G Τ(β) respectively, because the sets of probability measures are weakly compact and the graph of the restriction of Τ to V{X) is assumed to be weakly closed. Since the measures μη{Χ) converge to μ(Χ), we deduce that μ = μ(Χ)μ~ and ν = μ{Χ)ν. Then ν = 0 when μ = 0 and otherwise, ν = μ{Χ)ν G μ(Χ)Τ(μ/μ(Χ)) Hence we consider a sequence (μη, un) G V{X) χ V(Y) of the graph of Τ converging to (μ, и). In order to prove that ν G Τ{μ), it is enough to check that for any open subset Ό С X, inequality u{0) < μ{Ρ-\θ))
8.9. Invariant Measures and the Poincare Recurrence Theorem 349 holds true thanks to the first part of the proposition. Fix an open subset О С X. Since X is compact, О is the union of an increasing sequence of open subsets Op С Ο, ρ > 1 such that for every ρ > 1, Ov С О. Let us fix ρ and observe that F~1(Op) is compact, as the image of a compact by the upper semicontinuous set-valued map F-1. Inequalities Vn>l, i/„(Op) < μη(^-1(0Ρ)) < /^(V1^)) imply, thanks to a version of Alexandroff's Theorem recalled just after the end of the proof, that j v(Op) < liminfn-^ooi/nCOp) \ < limsupn_wMn(i'-1(^)) < μ{ρ-\θρ)) < μ{Ρ-\Ό)) It remains to observe that v{0) = supp>1 v{Op) to conclude. D We prove now the version of Alexandroff's Theorem (see [148, Theorem 4.9.15]) we needed: Theorem 8.9.3 (Alexandroff) Let X be α compact metric space. Consider a sequence of Radon probability measures μη £ V(X) converging weakly to μ. Then, for any open subset О С Х, μ(0) < Hminf/infO) η—► со and for any closed subset К С X, limsup^n(ii) < μ{Κ) π-too Proof — Let О С X be an open subset. Since a Radon measure μ is regular, we can associate with any e > 0 a compact subset К С О such that μ{Ό\Κ) < ε. Let / G C(X) be a continuous function equal to 1 on the closed subset К and to 0 on the complement of O. We observe that χκ < f < Xo and thus, that μ(0) < μ{Κ)+ε < ί ίάμ + ε
350 — 8- Measurability and Integration of Set-Valued Maps Therefore, inequalities / /άμη < μη(0) imply that μ(0) < [ fdu + ε < Uminf^n(0)+e J η—►со We conclude by letting e converge to 0. The proof of the second inequality for compact subsets К is analogous, since.for any e > 0, there exists an open О D К such that μ{Ό\Κ) < ε. Define the continuous function / as above. Inequalities μη{Κ) < J ίάμη imply that limsup^n(UT) < ί ίάμ < μ{Ό) < μ(Κ) + ε □ п—юо J Remark — We deduce that for any open subset О such that μ(0) = μ(0), weak convergence of probability measures μη to μ imply that μη{0) converges to μ(Ο). The converse is also true. (See the proof of [148, Theorem 4.9.15]) for instance.) □ 8.9.2 Invariant Measures Taking X = Υ, we are able to derive the existence of invariant measures by showing that the extension Τ has fixed points on V{X). Theorem 8.9.4 Let X be α compact metric space and F:l-^1 be a closed set-valued map with nonempty values. Then there exists an invariant probability measure μ, i.e., satisfying VA£ B{X), μ{Α) < μ{Ρ-\Α)) (8.16) Proof — Let Τ be the set-valued map from V{X) to itself associating with any μ G V{X) the (nonempty) set of probability measures ν G ^Ρ(μ). Then the above proposition states that Τ is a map with closed graph and convex values from the convex compact subset V{X) С С*(Х) to itself, and thus, upper semicontinuous with convex compact values. Kakutani-Fan's Fixed-Point Theorem 3.2.35 5Recall that the proof we provided, based on the Ky Fan inequality, showed that Ky Fan's theorem remains true in locally convex Hausdorff topological vector spaces.
8.9. Invariant Measures ала the Poincare Recurrence Theorem 351 implies the existence of a fixed-point μ G V{X) of T, which is a measure satisfying VAeB(I), μ(Α) < μ(Ρ-\Α)) (8.17) Therefore, μ is invariant by F. □ Naturally, we can derive most of the properties of invariant measures enjoyed by single-valued maps. Let us show for instance that Poincare's Recurrence Theorem holds true for closed set-valued dynamical systems. Theorem 8.9.5 (Poincare's Recurrence) Let X be α compact metric space, F : X ~»· X be a closed set-valued map and μ G V(X) an invariant measure of F. For any Borelian subset В С X, let Soc := Π U Р~П(В) Ν>0 η>Ν be the subset of points χ such that for all N, there exists η > N such that F~n(x) Π Β φ 0. Then the measure of В П Дх> is equal to the measure of B. Proof— The proof is a straightforward extension of the proof in the single-valued case: We introduce the subsets BN := (J F~n(B) n>N and we observe that В С Bq, that · · · С Дл/· С Β^-ι С · · · С В0 and that BN = F-N{B0). Since μ is invariant, we deduce that μ{Β) < μ{Ρ~\Β)) < ... < μ{Ρ~Ν{Β)) < ··· and thus, that μ(Β0) < μ(Ρ-Ν(Β0)) = μ(ΒΝ) < μ(Β0) Since the sequence of Β ν is not increasing and since μ(-Β/ν) = μ(-Βο), we infer that μ(-Βοο) = μ(Βο)· Therefore, μ{ΒΠΒ00) = μ(ΒηΒο) = μ{Β) D
Chapter 9 Selections and Parametrization Introduction One of the first questions of set-valued analysis was how to relate set-valued and single-valued maps, mainly in the hope of avoiding the necessity of dealing with set-valued maps. This can be done in two ways: 1. — Find a selection / : Jf н У of a set-valued map F : X ~»· Y, i.e., a single-valued map / satisfying Vi£X, f(x) € F(x) 2. — Find a parametrization (U, /), where U is a control or parameter space and f : X xU ι-+ Υ is & single-valued map satisfying Vx € X, F(x) = {f(x,u)}ueu Naturally, we require that in both cases the selection / or the parametrization (U, f) inherits the regularity properties of F, such as continuity or Lipschitzianity. Furthermore, keeping in mind applications to control theory, we shall also study the case when both F and / are time-dependent in a measurable way. 353
354 9 - Selections and Parametrization We proved in the preceding chapter the existence of a measurable selection of a measurable map. When we ask for more regularity of a selection, we meet more technical difficulties. This is why continuous (or Lipschitz) selections are associated with continuous (respectively Lipschitz) set-valued maps with closed convex images. These selections should be obtained in a constructive way whenever this is possible. The most natural constructive way amounts to using the minimal selection, i.e., the selection / with minimal norm. Unfortunately, this procedure does not preserve semicontinuity, but only continuity for maps with closed convex images, as one can see in Section 3. Such selection is also Caratheodory for Caratheodory maps with closed convex images (see Section 5.) We shall prove the existence of continuous selections of lower semicontinuous maps with closed convex images (Michael's Theorem) in Section 1, and the existence of approximate continuous selections (CelUna's Theorem) of upper semicontinuous maps with closed convex images in Section 2. If the set-valued map is Lipschitz continuous, then, contrary to an intuitive wish, it may happen that its minimal selection is not Lipschitz. There exist different ways to get Lipschitz selections for Lipschitz set-valued maps with convex images. The barycentric selection, which also comes easily to mind, is Lipschitz, but difficult to handle. By constrast, the selection based on the Steiner point1 enjoys attractive properties. For instance, it preserves measurability, continuity and Lipschitzianity of set-valued maps. We first single out a point of a nonempty convex compact set 'The Steiner point of a convex compact set, known for over a century, has many important properties. It is defined for subsets of Euclidean space R". Jacob Steiner (1796-1863) introduced this concept in 1840. From a peasant family, he had only a village school education before coming late in life to Mathematics. He had a strong preference for geometrical proofs over analytical ones. Actually, he never provided any of them and succeeded however in keeping his audience of Berliner Universitat awake without drawing any pictures. He was overshadowed by Diersteweg, in Mors, who purposefully draws the curtains of the class room to darken it before his lectures (...die Stube audi noch ausdriick- lich verfinsterte!.) Certainly he intended to help his listeners to dream freely of geometrical (with no doubts curvaceous) shapes...
9.1. Case of Lower Semicontinuous Maps 355 К С Rn, the Steiner point sn(K), and characterize it. The Steiner selection sn(·) being Lipschitz in the Hausdorff metric, the single-valued map χ н-> sn(F(x)) is a Lipschitz selection whenever the Lipschitz set-valued map F : X ~»· Rn has convex compact images. .To be able to deal also with unbounded set-valued maps, we prove in Section 4 an Intersection Lemma, which enables us to associate with a closed convex set its convex compact subset in a Lipschitz way. Hence we obtain in a constructive way Lipschitz selections for Lipschitz set-valued maps with closed convex images. The Steiner selection and the Intersection Lemma allow us to find selections (Section 5) and parametrizations (Sections 6 and 7) of continuous, Caratheodory and Lipschitz maps with closed convex images, preserving regularity of the set-valued map under consideration. 9.1 Case of lower semicontinuous maps We shall prove the celebrated Michael's theorem stating that lower semicontinuous convex-valued maps do have continuous selections. Definition 9.1.1 Consider α set-valued map F : Χ ~»· Υ with nonempty images. A single valued map f : X н-> У is called a selection of F if for every χ G X, f(x) € F(x). Theorem 9.1.2 (Michael's Theorem) Let F be a lower semicontinuous set-valued map with closed convex values from a compact metric space X to a Banach space Y. It does have a continuous selection. Corollary 9.1.3 Let F be as in Theorem 9.1.2. Consider a subset К С X and a continuous single-valued map φ : Κ ι—» Υ such that for every χ G Κ, ψ(χ) G F(x). Then ψ can be extended to a continuous selection of F, i.e., F has a continuous selection f (defined on the whole space X) such that the restriction of f to К is equal to φ. In particular this yields that for every у G F(x) there exists a continuous selection f of F such that f(x) = y.
356 9- Selections and Parametrization Proof — It is enough to consider the lower semicontinuous set- valued map G(x) := J ρ(χ) if x i K \ {ψ{χ)} otherwise and to apply Theorem 9.1.2. D The proof of this Theorem requires the following Lemma. Lemma 9.1.4 Let F be α lower semicontinuous convex-valued map from a compact metric space X to a Banach space Υ. Then for each ε > 0, there exists a continuous function f£ satisfying Vxel, fe(x) € B(F(x),e) Proof — Since F is lower semicontinuous, we can associate with any χ £ К and any yx G F(x) an open neighborhood Ux such that Vx'£ Z4, B(yx,e)nF(x') φ 0 Since X is compact, it can be covered by η such neighborhoods UXi. Let us consider a continuous partition of unity2 {ai}i=i,...,n associated with this covering, i.e. ai : X i—> [0,1] are continuous functions such that ai vanishes outside of UXi and η V χ G X, ^ai(x) = 1 i=l Define /e by: 71 It is continuous since the functions ai are continuous. Let I{x) := {i = 1,.. .,n\ αι{χ) > 0}' It is not empty since Y2=x ai(x) = 1. When i € I{x), then χ belongs to UXi and thus, yXi€B(F(x),e) 2 See Chapter 3, Section 1.
9.1. Case of Lower Semicontinuous Maps 357 The set B(F(x),e) being convex, we infer that fe{x) belongs to B(F(x),e.) □ We are now ready to prove the Michael Theorem. Proof of Theorem 9.1.2 — We shall construct by induction a sequence of continuous maps ип:ХиУ, n=l,... satisfying the following properties: ' i) Vx € X, d{un{x),F(x)) < 2~n (9.1) ^ и) Vx € X, ||un(aO-iin-i(aO|| < 2~n~1 For η = 1, we apply the lemma with e = 1/4. Assume that we have constructed the maps un up to η and let us construct un+\. We introduce for that purpose the set-valued map Fn+\ defined by Vx€X, Fn+1{x) :=B{un{x),2-n)nF{x) By (9.1) i), Fn+i(x) is convex and nonempty. The set-valued map Fn+i is lower semicontinuous: Indeed, if xp converges to χ and if у belongs to Fn+i(x), by the lower semicontinuity of F, there exist elements yp G F(xp) converging to y. For ρ large enough, we see that ||уР-Ип(я:р)|| < 2~n because и is continuous, and therefore, that yv belongs to Fn+i{xp). This being established, we deduce from Lemma 9.1.4 the existence of a continuous map un+i satisfying: \/x € X, d{un+1{x),Fn+1{x)) < 2-(n+1) which, together with the definition of Fn+i, implies (9.1) for η + 1. We infer from inequalities (9.1) ii) that the sequence of functions un is a Cauchy sequence in the Banach space of continuous maps from the compact space X to the Banach space Y. They converge uniformly to a continuous map и : Χ ι—> Υ. Inequalities (9.1) i) imply that for all χ € X, u(x) belongs to F(x), for this subset is closed by assumption. Hence и is the continuous selection we were looking for.
358 9- Selections and Parametrization 9.2 Case of upper semicontinuous maps Upper semicontinuous set-valued maps in general do not have continuous selections even when their values are closed and convex. The following example illustrates this issue: Example Consider the upper semicontinuous set-valued map F : R ~»· R defined by F(x) := < ' { -1 } if χ < 0 [-1,1] if χ = 0 { 1 } if χ > 0 Clearly F does not have any continuous selections defined on R. D This is why it is more realistic to speak about approximate selections. We state the following result due to A. Cellina:3 Theorem 9.2.1 Let F : Χ ~»· Υ be an upper semicontinuous map from a compact metric space X to a Banach space Y. If the values of F are nonempty and convex, then for every ε > 0, there exists a locally Lipschitz single valued map fe : X н-> У such that Graph(/e) С £(Graph(F),e) and for every χ £ X, fe{x) belongs to the convex hull of the image ofF. Remark — This inclusion implies that the graphical upper limit of the approximate selections is contained in F. □ Proof— Fix e > 0. Since F is upper semicontinuous, for every χ £ X, there exists 0 < δχ < 2e such that Vy e Ή(χ,δχ), F(y) С F{x) + E-B 3We prove here the approximate selection result only for a compact metric space X, but the result is valid with a similar proof for any metric space X■ It can be proved using paracompactness of metric spaces and the fact that a locally Lipschitz partition of unity may be subordinated to any locally finite covering of a metric space. Details can be found in [33, Section 1].
9.2. Case of Upper Semicontinuous Maps 359 Figure 9.1: Approximate Selection of ал Upper Semicontinuous Map . 0 GraphCf^ —^ СгатЛС.Е}-—" <- Graph(5(F,e)) Approximate Selection in a Neighborhood of Graph(F) The family of balls {В {х,йх/4)}хех covers X. Since X is compact, there exists a finite sequence / of indices such that the family of balls {B (xiJXi/4)}iei covers X. We set δ{ = δχ; and take a locally Lipschitz partition of unity {щ} subordinated to this covering4. That is a,- : X ь-> [0,1] are locally Lipschitz functions such that for all г 6 /, α,- vanishes outside of B(xi,S{/4) and \f χ £ X, Y^ai(x) = 1 iei Let us associate with every г ζ I & point τχ £ F(B(xi,Si/4)) and define the map /e : X ь-> Υ by iei Then /e is locally Lipschitz and for every χ £ X, fe(x) £ coF(X). To show that /e is the desired approximation, fix χ £ X and let I{x) denote the subset of all г £ I such that сц{х) ф 0. Thus 4See for instance [33, Theorem 1.1.2].
360 9- Selections and Parametriz&tion χ € B{xi,8ijA) for every i €. I(x). Therefore for all i,j € I{x) d(xi,Xj) < d(x,Xi) + d(x,Xj) < (<5i + <5j)/4_ Fix к € I(x) satisfying 6k = max δ{ i€l{x) Then the latter inequality implies that \fi€l{x), Xi£B(xk,6k/2) Consequently yi € F(B(xi,6i/A)) С F(B(xk,6k)) С F(xk) + |β Since the right hand side of the above relation is convex, from the definition of /e follows that fe(x) € F(xk) + £-B Pick zk G F{xk) satisfying ^\zk — fe{x)\\ < ε/2. Thus, by the choice of δχ and zk, d((x,fe(x)),(xk,zk)) < d(x,xk) + ||^-/е(ж)|| < — + - < e which ends the proof. □ 9.3 Minimal Selection Consider a metric space X, a Banach space Υ and a set-valued map F : X ~> Y. We define the minimal map m(F(x)) := \ и € F(x) I |Ы| = min \\y\\ \ (9.2) j_ y€F{x) J When У is a Hilbert space, or, more generally, a reflexive strictly convex space5 and F has closed convex images, the minimal map is single valued. In this case it is called the minimal selection. 5A Banach space Υ is called strictly convex if for all ж, у 6 Υ which are not colinear, we have ||ж + 2/|| < ||ж|| + \\y\\.
9.3. Minimal Selection 361 A quite natural question arises: What are continuity properties of the minimal selection'? We prove below that if F is continuous and takes its values in a compact subset of a strictly convex reflexive Banach space, then the minimal selection is continuous. The upper semicontinuity (or lower semicontinuity) of F, even when it takes convex values in a compact subset, is not strong enough to imply the continuity of the minimal selection, as the following examples show: Example — We consider the set-valued map F : R ~»· R defined by F(x)-= I {2} if x*° *{h \ [1,2] if x = 0 It is upper semicontinuous with compact convex values and its minimal selection is obviously not continuous at zero. Π Example — We consider the set-valued map F : R ~»· R defined by F(x).= { [o,i] if χϊο *{X)- \ {1} if x = 0 It is lower semicontinuous with compact convex values and its minimal selection is not continuous at zero. □ First of all we investigate continuity properties of the function d(0,F(-)). Lemma 9.3.1 Let F be a set-valued map from a metric space X to a Banach space Υ with nonempty images. Set \\F(x)\\ = sup Иг/11 yeF(x) 1. — If F is lower semicontinuous (or upper semicontinuous), then the function χ *—» d(0,F(x)) is upper semicontinuous (respectively lower semicontinuous.) 2. — If F has bounded images and is lower semicontinuous (or upper semicontinuous), then the function χ н-> ||-F(a:)|| is lower semicontinuous (respectively upper semicontinuous.)
362 9— Selections and Par&metriz&tion Consequently if F is continuous, then so is d(0,F(-)) and if in addition F has bounded images, then ||-Ρ(·)|| *s also continuous. Furthermore if F is Lipschitz with a Lipschitz constant c, then d(0, F(·)) is c-Lipschitz. If moreover F has bounded images, then \\F(-)\\ is ah о c-Lipschitz. In the above c-Lipschitz means Lipschitz with the constant с Proof — If F is lower semicontinuous^hen, by Corollary 1.4.17, the function χ ι-> d(0,F(x)) is upper semicontinuous and, in the case when F has bounded images, the function χ ι-> ||F(a;)|| is lower semicontinuous by Theorem 1.4.16. If F is upper semicontinuous, then for every ~x € Χ, ε > 0 there exists a neighborhood λίε of χ such that for all χ € λίε, we have F(x) С F(x) + eB. Hence d(0,F(x) > d(0,F(x))-e and therefore d(0,F(·)) is lower semicontinuous. On the other hand if F(x) is bounded, ||F(x)|| < \\F(x)\\+e This yields that ||-Р(-)|| is upper semicontinuous whenever images of F are bounded. The latter statement follows from the very definition of Lipschitz set-valued maps. □ Proposition 9.3.2 Let F be a lower semicontinuous set-valued map from a metric space X to a Banach space Υ whose graph is closed. If the minimal map m(F(·)) has nonempty images, then its graph is closed. Consequently if Υ is strictly convex, reflexive, F has nonempty closed convex images and the range of m(F(·)) is contained in a compact subset ofY, then the minimal selection is continuous. Corollary 9.3.3 Let F be a continuous set-valued map from a metric space X to Rn with nonempty closed convex images. Then the minimal selection is continuous.
9.3. Minimal Selection 363 Proof— Lemma 9.3.1 implies that ||d(0, F(-))\\ is continuous. Fix x € X. Then for some δ > 0 the set {\\m(F(y))\\}yeB(F(x),6) is bounded. Proposition 9.3.2 implies that m(F(·)) is continuous at x. The element χ € X being arbitrary, the proof follows. □ Proof of Proposition 9.3.2 — Indeed by Lemma 9.3.1 the map χ ^ d{0,F{x)) = \\m{F{x))\\ is upper semicontinuous. This yields that the set-valued map χ ~> B(0,\\m(F(x))\\) has closed graph. By our assumption the graph of F is closed. Thus the intersection map χ ~> F(x)nB(0,\\m(F(x))\\) has closed graph. On the other hand m(F(x)) = F(x)nB(0,\\m(F(x))\\) and the proof of the first statement ensues. To prove the second statement it is enough to observe that a single-valued map taking its values in a compact set is continuous if and only if its graph is closed. □ When У is a Hubert space, then the compactness assumption in the above proposition may be replaced by the upper hemicontinuity of F: Theorem 9.3.4 Consider α metric space X, a Hubert space Υ and a lower semicontinuous set-valued map F : Χ ^λΥ with closed convex values. If F is upper hemicontinuous, then the minimal selection is continuous.
364 9 - Selections and Par&metriz&tion Proof — We identify Υ with its dual... The projection of 0 onto the closed convex set F(x) is the element „ и := m(F(x)) € F(x) such that IMI2 + ^(-F(x),u) = sup <u-0,u-y>-< 0 (9.3) ye-F(i) (It is actually equal to 0.) Fix e > 0 and χ € X. Since F is upper hemicontinuous there exists a neighborhood" V of χ such that for all x' € V ' (-m(F(x')), m(F(x))) < a(-F(x'),m(F(x))) < w < a(-F(x),m(F(x))) + ε < -\\m(F(x))\\2 + ε (thanks to (9.3).) On the other hand, by Lemma 9.3.1, the function χ н-► ||m(-F(a;))|| is upper semicontinuous, since F is lower semicontinuous. Then there exists another neighborhood W of χ such that \/x' €W, \\m{F{x'))\\2 < \\m{F{x))\\2 + ε These two inequalities imply that for all x' € V Π W ' ||m(iV)) - m(F(x))f « = limbs'))!!2 + \\m(F(x))\\2 + 2 <-m(F(x')),m(F(x))> k < 2\\m{F{x))\\2 + ε + 2(-||^(ζ)||2 + ε) < Зе Therefore, the minimal selection is continuous. □ 9.4 The Steiner Selection To construct Lipschitz selections of set-valued maps and to parametrize Caratheodory set-valued maps, we need a selection procedure associating with a closed convex set К С Rn a point s(K) € К in
9.4.1. Steiaer Points of Сол vex Compact Sets 365 a regular enough way. We start this section with such a selection procedure for convex compact sets K. To be able to deal also with unbounded sets, we prove an Intersection Lemma which enables us to associate with a closed convex set К a convex compact subset P{K) С К in a Lipschitz way. The last part of this section is devoted to Lipschitz selections of set-valued maps with closed convex images. 9.4.1 Steiner Points of Convex Compact Sets For a nonempty convex compact subset К of Rn, we define its Steiner point or, for the sake of simplicity, Krummungsschwerpunkte, (also called sometimes the curvature centroid) sn(K) by: ' for η = 1, Sl{K) = σ(Κ,+ΐ)/2 - σ{Κ,-ΐ)/2 < for η > 2, sn{K) = η^η~ιρσ(Κ,ρ)ω(άρ) where Σπ_1 denotes the unit sphere in Rn, σ(Κ, ·) is the support function of Κ, ω is the measure on Σπ_1 proportional to the Lebesgue measure and satisfying ω(Ση~1) = 1. ' Since σ{Κ,ρ) = σ(-Κ, -ρ), it follows that sn{K) = —sn(—K). The support function being also additive with respect to K, the map δπ(·) is linear: For all convex compact K, L С Rn and all λ, μ € R, < δπ(λΛΓ + μ^) = Asn(ur)+^sn(L) We also observe that if К is symmetric, i.e., К = —К, then sn{K) = 0. \ We show below that sn is a selection in the sense that sn(K) € К and that it is Lipschitz with respect to the Hausdorff distance: We recall that the extended Hausdorff distance between two closed subsets К-, L С Rn (which may be equal to +oo when К or L is
366 9 - Selections and Paxametrization unbounded or empty) is denned by6 H(K,L) := max< sup d(x,L), sup d(x,K) > l хек xeL J The Steiner point can be alternatively defined by using the sub- differentials of the support function of a compact convex subset K. Recall that the subdifferential да(К,р) of the support function σ(Κ, ■) is given by θσ(Κ,ρ) = {χ € K\ <p,x>= σ(Κ,ρ)} We denote by т(да(К,р)) the element of да(К,р) with the minimal norm. The function σ(Κ, ·) being continuous, Theorem 8.2.9 implies that the subdifferential θσ(Κ, ·) is measurable. Hence т(да(К,р)) being the projection of 0 onto δσ(Κ,ρ), the single-valued map т(да(К, ·)) is also measurable (Corollary 8.2.13.) Therefore, the formula stated in the following theorem has a meaning: Theorem 9-4.1 Let /C denote the family of all nonempty convex compact subsets o/Rn. Then УК € /С, sn(K) = Щщ/Вп™(да(К,р))ар where Vo\(Bn) is the measure of the η-dimensional unit ball Bn С Rn. Consequently, \/K € /C, sn{K) € К Furthermore sn(·) is Lipschitz with the constant n: VK, L €/C, \\sn(K) - sn{L)\\ < nH(K,L) 6It can be verified easily that Ή is a distance. Furthermore, a set-valued map with nonempty compact images is continuous if and only if it is continuous with respect to the Hausdorff distance.
9.4.1. Sterner Points of Couvex Compact Sets 367 Proof— Fix if, LcRn. — When η = 1, set 2/ = σ(Χ,+ΐ), ζ = -σ(Κ,-1) Then y,z Ε Κ and *i(*0 = \y+\z Ζ Κ Furthermore vohky /Bl m (δσ(χ> p)) *=5 (Γ **+£ **)=^+r and ' ||Sl(X) - Sl(L)|| < ^ < 1\σ(Κ,+ΐ) - a(L,+l)\ + \\a(K,-l) - a(L,-l)\ < H{K,L) — We assume next that η > 2. Consider the Moreau-Yosida approximation σ\(Κ, ·) of the support function σ(Κ, ■) of K. Recall that it is continuously differentiable. By integrating its gradient on the unit ball Bn whose boundary is the unit sphere Σ71"1, Stoke's formula implies that / Vax(K,p)dp = / σχ(Κ,ρ)ρμ(άρ) where μ is the Lebesgue measure on the unit sphere Σπ_1. We also recall that σ\{Κ, ·) satisfies -\\K\\ < inf ΜΚ,ρ) < σχ(Κ,ρ) < σ(Κ,ρ) < \\K\\ (where ||ΛΓ|| := sup^g^ \\x\\) and converges pointwise to σ(Κ, ·) (see Theorem 6.5.7.) Therefore lim / ρσχ(Κ,ρ)μ(άρ) = Ι ρσ(Κ,ρ)μ(άρ) λ->0+ J^n-i JT,n~i On the other hand, by Theorems 3.5.9 and 6.5.7, Vax(K, ·) is the Yosida approximation of the sub differential θσ(Κ, ·). Hence Vax(K,p) € θσ(Κ,ρχ) с К
368 9 — Selections and Paxametrization where p\ is the minimizer of the function q н-> a(K,q) + —\\q-pf and for any ρ € Bn, Va\(K,p) converges to т(да(К,р)). Therefore, the maps Va\{K, ·) being measurable and bounded by \\K\\, we infer that Urn / Vax(K,p)dp = / τη{θσ{Κ,ρ)) dp λ->Ό+ Jb-"- JBn Dividing by Vol {Bn), and setting ω (dp) := μ(άρ)/ω(Ση~1), we obtain Since ω(Ση~1) = nVo\(Bn), we derived the formula we were looking for. Proposition 8.6.2 and Theorem 8.6.3 yield that / К dp = Vo\{Bn)K JB71 Consequently ™p=)/«. "■<*'<*■'»* ε й^)//* = K To prove that sn(·) is n-Lipschitz observe that for every ρ € Σπ-1 VK, L € 1С, \a(K,p)-a{L,p)\ < H{K,L) Hence \\sn(K) - sn(L)\\ < η!^1\σ(Κ,ρ)-σ(£,ρ)\\\ρ\\ω(άρ) w < пЩК,Ь)ш{Т,п~1) = nH{K,L) The proof is complete. □
9.4.2. The Intersection Lemma 369 9.4.2 The Intersection Lemma To obtain Lipschitz selections for unbounded closed convex sets we need the following technical lemma. Lemma 9.4.2 Let /C denote the family of all nonempty closed convex sets in Rn. Then the map Ρ : Rn χ JC ^* 1С defined by P(y,K) := КпВ(у, 2d(y,K)) is Lipschitz: for all K,L € /C and x, у € Rn H(P(x,K),P(y,L)) < 5(H(K,L) + \\x - y\\) Proof — First we show that our statement holds true with χ = 2/ = 0: \/K,L € /C, H{P(0,K),P(0,L)) < 5H{K,L) (9.4) Pick K,L in /C such that H{K,L) < +oo and set e := H(K,L). If e = 0, then К = L and (9.4) holds true. Suppose next that e > 0 and let у € P(0,L) be arbitrary but fixed. To prove inequality (9.4), we need to find a point χ in P(0,K) such that \\x-y\\ < 5e (9.5) Observe first that ||y|| < 2d(0,L) < 2d(0,K) + 2e (9.6) Let x\ € К be such that ||?/-a;i|| < e If ||xi|| < 2d(0,K) then set χ := x\. Otherwise let X2 € К be such "that d{0,K) = \\x2\\ If x2 = 0 set ж = 0. Then (9.6) implies (9.5.) It remains to consider the case ||xi|| > 2d{0,K), x2 φ 0
370 9 — Selections and Paxametrization Figure 9.2: Illustration of the proof We claim that there exists χ in the interval \x\, x%\ such that ||x|| = 2d(0,K) Indeed define the continuous map ψ : [0,1] ι-> Rn by VA € [0,1], ψ{\) = \\\Xl + (1-λ)χ2|| Since the image of φ is a connected set and ψ{0) = d{0,K) L· φ(ΐ) > 2d{0,K) there exists λ € ]0,l[ which satisfies φ(λ) = 2d(0,K). So setting χ := \xi + (1 — X)x2, our claim follows. The set К being convex, x€P{0,K). We prove next that this χ is the point we are looking for. Indeed, by the choice of x\, ||яг — £Ci|| > \\x — y\\ — \\y — xi\\ > \\x — y\\ — ε (9.7) In the triangle with vertices at 0, x, X2 denote by α the angle at x
9.4.2. The Intersection Lemma 371 (see Figure 9.2.) Inequality ЦжгЦ < ||x|| yields that α € [0, f [ and ^ II^H 1 \/3 sine* < y-~ = -, cosa > — (9.8) If α = 0 (that is always the case when η = 1), then ll^ill = \\x\\ + \\x ~~ χι\\ and (9.7) yield \\y\\ > H^ill-||y _ χι\\ > \\χ\\+\\χι — χ\\~ε > 2d(0,K)+\\x — y\\—2e Thus we deduce from (9.6) that Ik-2/11 < 4e and derive (9.5.) Assume next that α > 0 and consider the triangle with vertices at 0, x, x\. Then the angle at χ is equal to π — α > ξ. Let ζ €]0, χι[ be such that ζ — χ ± x. Then ||z|| > ||x|| = 2d(0,K) (9.9) In the triangle formed by points ζ,χ,χι, the angle at χ is equal to π — α—| = | — α. Therefore (9.8) implies V3 2Г — χι Ν > \\x — xi||sm ( — — a J = \\x — xi\\ cosa > \\x — хл\\ — J-ll — II -Ml V П /II -Ml — II J-ll q (I-)- and from (9.9) follows that л/3 \\XU = \\z\\ + \\z ~ XU ^ 2d(0,.K") + —- Цж — xi\\ This and (9.7) imply л/Я 112/11 > Iki||-||y-si|| > 2d{0,K) + ~{\\x-y\\-e) - ε Therefore, thanks to (9.6), we derive л/3 2 Mi1-У|| -e)-e < 2e
372 9 — Selections and Parametriz&tion and thus (9.5.) So (9.4) is proved. Pick next any x, у ε Χ and observe that H(P(x,K),P(x,Ljf^ff(P(0, Κ - x),P(0,L - x)) Hence, using (9.4), we deduce ' H(P(x,K\P(y,L)) < H(P(x,K),P(x,L)) + H(P(x,L),P(y,L)) < m(K-x,L-x) + H(B(x,2d(x,L)),B(y,2d(y,L))) k < 5H{K,L) + \\x-y\\ + 2\\x-y\\ = 5H(K,L) + 3||s-y|| The proof is complete. □ 9.4.3 Lipschitz Selections of Lipschitz Maps An immediate application of the above construction is the existence of a Lipschitz selection for a Lipschitz set-valued map: Theorem 9.4.3 Consider α Lipschitz set-valued map F from a metric space to nonempty closed convex subsets ofRn. Then F has a Lipschitz selection f. Remark — In the next section, we prove a stronger result, which in particular implies that for any у € F(x) there exists a Lipschitz selection f οι F such that /(ж) = у. D Proof — We first observe that a map F : X ~»· Rn is Lipschitz if and only if there exists c> 0 such that Vs, у € X, 7i{F{x),F{y)) < c\\x-y\\ Define the set-valued map G : X -^ Rn with nonempty compact convex images by Vz € X, G(x) = F(x)nB(0,2\\m(F(x))\\)
9.5. Caratheodory Selections 373 By Lemma 9.4.2 it is Lipschitz. Consider the single-valued map f(x) = sn(G(x)) where sn(·) is the Steiner selection. The proof follows by the application of Theorem 9.4.1. □ Remark — Observe that if the set-valued map F is merely continuous in the Hausdorff metric·. Ve>0, 3 <5 > 0, Vie Β(χ, δ), H(F(x), F(x)) < ε then the function / defined in the above proof is a continuous selection of F. □ 9.5 Selections of Caratheodory maps Prom now on a measurable map always means a Lebesgue measurable map. Let X be a metric space and а < b be real numbers. Consider a set-valued map F : [a, b] χ Χ ~> Rn Definition 9.5.1 For every t € [a, b] let C(t) С X be a given set. 1. — The map F is called Caratheodory on {C(t)}te[au if V χ € X, F(-, x) is measurable < Vt € [a, b], F(t,·) is continuous on C(t) 2. — The map F is measurable/Lipschitz on {C(£)}ieraw if for every t € [a, 6], there exists k(t) > 0 such that V χ € X, F(-tx) is measurable < Vt € [a, b], F{t,·) is к(t)- Lipschitz on C(t) where k(t)—Lipschitz means Lipschitz with the constant k(t).
374 9- Selections and Paxametrization Let a single valued map ζ : [α, b] н-► X and a measurable single valued map w : [a, b] i-> Rn be such that- w(t) € F(t, z(t)) almost everywhere in [a, b] We refine here the selection result proved in the preceding section: Namely we show the existence of a measurable/Lipschitz selection f(t, x) € F(t, x) such that for almost every t € [a, b], w(t) = f(t,z(t)) whenever F is measurable/Lipschitz. Results of this type are sometimes needed in control theory when we investigate the local behavior of solutions for nonsmooth systems. Theorem 9.5.2 (Caratheodory Selection) Let X be a metric space, F : [a, b] x X ~»· Rn be a set-valued map with nonempty closed convex images, ζ : [a, b] ι—» X be a single valued map and w : [a, b] ι—» Rn be measurable. If F is Caratheodory on {C(£)}ie[a,6] and w(t) € F(t,z(t)) almost everywhere in [a, b] then there exists a Caratheodory selection f of F on {C(i)}ie[a,6] such that, for almost every t € [a,b], w(t) = f(t,z(t)). Proof — Define the map / by V(i,s) € [a,b]xX, f(t,x) = UF{ttX)(w(t)) where Π^(ν) is the projection map from Corollary 8.2.13, i.e., for every (t,x), HF^tiX^(w(t)) is an element of F(t,x) such that d(w(t),F(t,x)) = ||nF(ti!B)(ii;(i))-ii;(0 Corollary 8.2.13 implies that for every χ € X, the map f(-,x) is measurable. It is also clear, that w(t) = f(t, z(t)) almost everywhere in [a, b]. The last statement follows from Corollary 9.3.3. □ If F(t,·) is Lipschitz, then the selection / may be chosen to be Lipschitz with respect to x:
9.5. Caratheodory Selections 375 Theorem 9.5.3 (Measurable/Lipschitz Selection) Under all assumptions of Theorem 9.5.2 assume that the set-valued map F is measurable/Lipschitz on {С^)}щад and let k(t), t € [a, b] denote the corresponding Lipschitz constants. Then there exist с > 0 and a measurable/Lipschitz selection f of F on {C(£)}te[a,b] such that f(t,·) is ck(t)—Lipschitz on C(t) and for almost every t € [a, b], w(t) = f(t,z(t)) Furthermore с is independent of F. Proof — Let Ρ be the set-valued map defined in Lemma 9.4.2 and sn be the Steiner selection. For every t € [a, b] and χ € Rn set f(t,x) = sn(P(w(t),F(t,x))) Since w(t) € F(t,z(t)), by the very definition of the map P, we get P(w(t),F(t,z(t))) = w(t) almost everywhere in [a, b] Thus f(t,z(t)) = w(t) almost everywhere. Theorem 9.4.1 implies that f{t,x) € F(t,x). We deduce from Lemma 9.4.2 and Theorem 9.4.1 that for every t € [a,b], the map f(t; ·) is Lipschitz on C{t) with Lipschitz constant 5nk(t). It remains to show that / is measurable with respect to t. Fix χ € X. We first show that the map t ^ Φ(ί,χ) := P(w(t), F{t,x)) = F{t,x)nB(w(t), 2d(w(t),F(t,x))) is measurable. Indeed, by Theorem 8.2.4, the intersection of two measurable maps is measurable. On the other hand F(-,x) is measurable by assumption and t ~> B(w{t),2d(w(t),F(t,x))) is measurable by Corollary 8.2.13. Hence Theorem 8.2.14 implies that for every ρ € Sn~1, the support function σ(Φ(·,χ),ρ) is measurable. Thus, by the definition of
376 9- Selections and Parametrization Steiner point, f(-,x) is measurable when η = 1. When η > 2, it is enough to observe that the map ρ ι-> ρσ(Φ(ί, χ), ρ) is continuous and therefore the map t ^ / ρσ(Φ(ί,χ),ρ)ω(άρ) Js*-1 is measurable. □ 9.6 Caratheodory Parametrization We prove in this section that a Caratheodory set-valued map F with closed convex images admits a Caratheodory parametrization. Consider a metric space X, reals а < b and a set-valued map F : [a, b] χ Χ ~> Rn Definition 9.6.1 Let U be a metric space and C{t) С X, t € [a, b] be given subsets of X. We say that a single-valued map f : [a,b] xX x U ■-» Rn is a Caratheodory parametrization of F on {C(£)}ie[a,6] if ' i) V(i,s) € [a,b]xX, F(t,x)=f(t,x,U) ii) V (x,u) € Χ χ U, f(-,x,u) is measurable < Hi) V (t,u) € [a, b] χ U, f(t,-,u) is continuous on C{t) iv) V (t,x) € [a, b]xX, f(t,x,·) is continuous We recall that \\F(t, x)\\ is defined by \\F{t,x)\\ := max ||у|| y£F(t,x) Let В denote the closed unit ball in Rn. Theorem 9.6.2 (Caratheodory Parametrization) Consider a metric space X and a set-valued map F : [a,b] χ Χ ~> Rn with nonempty convex compact images.
9.6. Car&theodory Paxametrization 377 IfF is Caratheodory on {С(Ь)}щащ, then there exists a Carathe- odory parametrization f of F on {С^)}щац such that U = В and for all t € [a,b], χ € X, и,υ € В we have ||/(i,s,u)-/(i,s,t/)|| < c\\F(t,x)\\\\u-v\\ where с is independent of F. Furthermore if F is continuous, so is f. Finally, if images of F are merely closed (instead of compact), then the same statement holds true with U = Rn and \\F(t,x)\\ replaced by 1. Proof— To simplify the notations, we set Mx(t) := ||.F(i,a;)||. Let (t, x, u) € [a, b] χ Χ χ Rn. Consider the closed ball G(t, x, u) of radius 2d(Mx(t)u,F(t,x)) and center Mx(t)u, i.e., G{t,x,u) = B(Mx(t)u,2d{Mx(t)u, F(t,x))) We claim that G is measurable with.respect to t. Indeed, fix χ € X, и € Rn. Prom Theorem 8.2.11 it follows that Mx(·) — ||.Р(-,а;)|| is measurable and from Corollary 8.2.13 that the function d(Mx(-)u,F(-,x)) is measurable. Thus, using again Corollary 8.2.13, we deduce that G(-,x, u) is measurable. Let Ρ be the map defined in Lemma 9.4.2 and set $(t,x,u) := P(Mx{t)u,F{t,x)) = G(t,x,u)riF(t,x) С F(t,x) Since G(-,x, u) is measurable, Theorem 8.2.4 and our assumptions on F imply that <&(·,£, u) is measurable. Define the single-valued map / from [a,b] χ Χ χ Rn into Rn by V {t,x,u) € [a,6] x-X" x Rn, f(t,x,u) := sn($(t,x,u)) where sn is the Steiner selection. Since Φ is measurable with respect to t, using the definition of sn, we deduce that / is also measurable with respect to t. Lipschitz properties of sn and Ρ proved in Theorem 9.4.1 and Lemma 9.4.2 imply: ' \\f(t,x,u)-f(t,y,v)\ < nHmt,x,u),$(t,y,v)) (9.10) < 5n{H{F(t,x),F{t,y)) + \\Mx(t)u - My{t)v\\)
378 9- Selections and Paxametrization Since F has compact images, F(t,·) is continuous on C(t) in the Hausdorff metric. This and (9.10) imply that for every и € Rn, /(·, ·, u) is Caratheodory on {C(i)}ie[aib]. In order to prove that / is a parametrization, it remains to check г) of Definition 9.6.1 with U = B. For this aim fix t € [a, b], χ € X and у € F(t,x). Setting u ._ ί y/Mx(t) if Mx(t) φ о 1 0 otherwise we derive и € В, Mx(t)u = у, $(t,x,u) = у Hence f{t,x,u) - sn($(t,x,u)) = у This means that F(t,x) с f(t,x,B). The opposite inclusion resulting from Theorem 9.4.1, we have proved г). Assume next that F is continuous. Then, by Lemma 9.3.1, so is ||-F(·, ·)ΙΙ· Thus the set-valued map G with convex compact images is continuous. Consequently its graph, Graph (G), is closed and therefore the intersection map Φ has closed graph. Since for all (t, x), F(t, x) с Mx(t)B, we deduce that Φ is upper semicontinuous. We claim that Φ is actually continuous. Indeed in order to prove this claim, it is enough to show that it is lower semicontinuous. Fix (t, x, u) and an open set О С Rn such that Φ(ί,ι,η)ηΟ φ 0 Let z\ € Φ(ί, χ, и) П О. From the very definition of G we deduce that there exists z2 € F(t,x)nlat(G(t,x,u)) Thus the interval }zi,z2] С Int(G(i>s>u))ni,(i>s) Consequently we can find an element ζ € Rn satisfying ζ € Int(G!(i>s,u))ni,(i>s)nO
9.7. Measurable/Lipschitz paxametrization 379 Hence for some e > 0, Β(ζ,ε) С G(i,i,ti)nO This and continuity of G imply that for every (t',x',u') sufficiently close to (t,x, u), Β(ζ,ε/2) С G{t',x',u') On the other hand, F being continuous, for all (t1, x') near (t, x), B{z,e/2)nF(t\x') φ 0 Thus Φ is lower semicontinuous, and thus, is continuous. Since Φ has compact images, it is also continuous in the Haus- dorff metric. Prom the very definition of sn, we deduce that / is continuous. To prove the last statement, it is enough to apply the same proof replacing everywhere Mx(t) by 1. □ 9.7 Measurable/Lipschitz Parametrization In this section we show that the Caratheodory parametrization defined in the preceding section is measurable/Lipschitz on {C{t)}t&[a,b} whenever F is measurable/Lipschitz on {C(t)}teia,b]· Theorem 9.7.1 (Parametrization of Unbounded Maps) Consider a metric space X and a set-valued map F : [a, b] x X ~> Rn with nonempty closed convex images. Assume that F is measurable/Lipschitz on {C{t)}te[a,b] and ^ k(t), t € [a, b] denote the corresponding Lipschitz constants. Then there exists a Caratheodory parametrization f of F with U = Rn on {C(£)}ie[a,6] such that: ' V(i,u) € [a,6]xRn, f{t,-,u) is ck(t) - Lipschitz on C(i) < V (t,x) € [a,b]xX, f(t,x,·) is с - Lipschitz on Rn where с is independent of F. Furthermore if F is continuous, so is f.
380 9- Selections and Parametrization Proof— It is enough to apply the same proof as the one of Theorem 9.6.2 replacing everywhere Mx{t) by 1 and to observe that (9.10) implies that for all (t,u) € [а,Ь] х Rn, f{t,-,u) is 5nfc(£)-Lipschitz on C(i). □ Theorem 9.7.2 (Parametrization of Bounded Maps) We posit the assumptions of Theorem 9.7.1 and we assume that the values of F are compact. Then there exists a Caratheodory parametrization f of F on the family of sets {C(t)}te[aq with U = В such that: i) V (t, u) € [a, b] x B, f(t, ·, u) is ck(t) — Lipschitz on C{t) < ii) Vi € [a,b], \/x € X, \fu,v € В \\f(t,x,u)-f(t,x,v)\\ < cllFfoaOllllu-t/H where с is independent of F. Furthermore if F is continuous, so is f. Proof— We consider the same parametrization as in the proof of Theorem 9.6.2. We already know that / is a Caratheodory parametrization of F on {C(i)}ie[ai6]. Then (9.10) and Lemma 9.3.1 imply that for all (t, u) € [a, b] x B, \\f(t,x,u)- f(t,y,u)\ < < 5n(H(F(t,x),F(t,y)) + \\Mx(t)u - My(t)u\\) < 5n (k(t) \\x - y\\ + k(t)\\x-y\\) Consequently, f(t,-,u) is 10nfc(i)-Lipschitz on C(t). □ When we allow the set of parameters to range in a Banach space, then actually the parametrization may be taken linear with respect to the parameter. Theorem 9.7.3 (Linear Parametrization) If all the assumptions of Theorem 9.7.1 are satisfied and X is compact, then there exists a map ψ : Χ χ C{X;Rn) н-> Rn
9.7. Measurable/Lipschitz paxametrization 381 and a measurable set-valued map U : [a,b] ~> C(X;~Rn) with closed convex images such that : ' i) V(i,s) € [a,b]xX, F{t,x) = <p(x,U(t)) ii) ψ{χ,·) is linear nonexpansive < Hi) Vt € [a,b], Vu € U(t), Vx,y € X, \\tp(x,u) - tp(y,u)\\ < ck(t)d(x,y) where с is independent of F. If moreover F has compact images, then the values of U are compact. Proof — Let / be a paxametrization map as in the conclusion of Theorem 9.7.1. Then V (t,u) € [a,b] χ Rn, f(t,-,u) is ck(t) - Lipschitz Hence for all t € [a,b], {f(t, -,и)}иец is a family of equicontinuous maps. Define the map U with closed convex values by : Vi € [a,b], U(t) := co({f(t,;u)}ueRn) С С(Х;Кп) If F has bounded images, then Ascoli's Theorem yields that for all t, U(t) is compact. To show that it is measurable, consider a dense sequence (?χ^)^>ι in Rn. Then for every к the map t -> f(t,-,uk) € C(X;Rn) is measurable. Set V(t) = \Jf(t,;Uk) k>l Theorem 8.2.4 implies that V is measurable and Theorem 8.2.2 yields that U is measurable. Consider the linear map ψ : Χ χ C{X;Rn) ■-» Rn, <p(x,u) := u{x)
382 9- Selections and Parametrization Clearly it is nonexpansive. Then equality г) holds true by the very definition of ψ and U(t). Notice that иг) is verified for every и € V(t). Hence, by definition of φ, it is verified for every и € coV(t). The set U(t) being equal to the closure of coV{t) the proof ensues. Π
Chapter 10 Differential Inclusions Introduction This chapter is just ал introductory survey of differential inclusions in finite dimensional vector-spaces. Many results below are stated without proofs. Their proofs and further developments can be found in [33], [50], [195]. Differential inclusions x'{t) € F{t,x{t)) (10.1) where F is a set-valued map from R χ Χ to a finite dimensional vector-space X have been studied since the thirties. Their consideration was initiated by Zaremba1 in 1934 and Marchaud in 1938. These investigators were mostly interested in the existence results2 and also examined some qualitative properties of solutions sets. We have first to agree on what we shall call a solution to such a differential inclusion. 1who kept close relations with Painleve and other French mathematicians after his study in Paris. 2While Zaremba investigated the paratingent solutions, Marchaud was mainly concerned with the contingent ones, i.e., the derivatives were taken in the contingent and paratingent sense. Since 1961, when Wazewski showed that one can use more "classical" solutions than the contingent and paratingent ones, solutions have been understood in the Caratheodory sense, i.e., absolutely continuous functions verifying (10.1) almost everywhere. 383
384 10- Differential Inclusions In the case of differential equations, there is no ambiguity since the derivative x'(·) of a solution x(-) to the differential equation x'(t) = f(t,x(t)) inherits the regularity properties of the map / and of the function x(-). This is no longer the case with differential inclusions and is one of the reasons why their study brings more difficulties than that of the ordinary differential equations. Solutions to differential inclusion (10.1) are understood in the Caratheodory sense, i.e., absolutely continuous functions verifying (10. l) almost everywhere. Interest in differential inclusions (10.1) was revived in the early sixties, when mathematicians became attracted by a new domain: control theory3. Filippov in 1959 and Wazewski in 1961 proved that under very mild assumptions the control system x' = f(t,x,u(t)), u(t) € U is measurable (10.2) may be reduced to the differential inclusion (10.1.) This considerably simplified the study of the closure of solutions to (10.2) and led to the celebrated Filippov-Wazewski relaxation theorem, which we state in Section 4. But differential inclusions also encompass much more sophisticated control systems: 1. — Closed loop control systems x'{t) = f(t,x{t),u(t)), u(t) € U(t,x(t)) (10.3) 2. — Implicit control systems f(t,x(t),x'{t),u{t)) = 0, u{t) € U(t,x(t)) (10.4) 3. — Systems with uncertainties x'{t) € f{t,x(t),u(t)) + e(t,x)B, u(t) € U{t,x(t)) (Ϊ0.5) 3Set-valued maps have been carefully — maybe unconsciously — hidden in control theory, despite the early contributions of Wazewski and Filippov. They were thought of as being useless. This is often the justification of something new one is reluctant to learn... "Everybody wants to teach, nobody wants to learn..." bitterly complained Abel.
10.0. Introduction 385 where e(t,x) is a function incorporating the disturbances of the model. Setting F(t,x) := f(t,x,U(t,x)) in the first case, F(t,x) := {υ\ Ο € /(ΐ,χ,ί/,Ε^ΐ,α:))} in the second one and F(t,x) := /(ί)χ(0,ί/(ί)χ))+ε(ί)χ)β in the third one, we replace the above control systems by differential inclusions of the type (10.1.) We shall survey three main classes of differential inclusions. First, the case when the right-hand side is upper semicontinuous with closed convex values: in this instance, we can extend Peano's Theorem on the existence of solutions to differential equations with continuous right-hand side and show that the set of solutions depends upper semicontinuously upon the initial state. This is the topic of Section 1. In Section 4, .we trade convexity of the images with more regularity on F, which is assumed to be Lipschitz with respect to the state and measurable with respect to the time. Filippov's Theorem provides not only the extension of the Cauchy-Lipschitz Theorem, but also very useful estimates ά la Gronwall, implying in particular the Lipschitz dependence on the initial state. Under this condition, we can regard Fasa kind of infinitesimal generator of a semigroup of set-valued maps, which are the reachable maps. Furthermore, it allows us to estimate the contingent, adjacent and circatangent derivatives of the solution map S := <S[o,:r] associated with the differential inclusion for almost all t € [Ο,Γ], x'{t) € F(t,x(t)) where F:[0,T]xX^X. We shall express these estimates in terms of the solution maps of adequate linearizations of differential inclusion of the form w'(t) € F'{t,x(t),x'{t)){w{t))
386 10- Differential Inclusions where for almost all t, F'(t, x, у) (u) denotes one of the (contingent, adjacent or circatangent) derivatives of the set-valued map F(t, ·) at a point (ж, у) of its graph (the set-valued map F is regarded as a family of set-valued maps χ ~> F(t,x) and the derivatives are taken with respect to the state variable only.) These linearized differential inclusions can be called the variational equations, since they extend (in various ways) the classical variational equations of ordinary differential equations. The Relaxation Theorem, stated in Section 4, connects these two classes of differential inclusions: when F is Lipschitz and compact- valued, the set of solutions to differential inclusion (10.1) is dense in the set of solutions to the relaxed or convexified differential inclusion x'(t) € coF(t,x(t)) We also mention the third case when F(t,x) := -A{x) where Л is a maximal monotone map. (This covers gradient inclusions of the form for almost all t € [Ο,Γ], x'{t) € -dV(x(t)) where the "potential" У is a lower semicontinuous convex function, describing the continuous version of the steepest descent method.) Monotonicity implies the uniqueness of the solutions. Existence is guaranteed if A is maximal monotone, so that A can be regarded (and is regarded) as the infinitesimal generator of a semigroup of (nonlinear) operators. This is elucidated in Section 3. We shall also devote Section 1 to the Viability Theorem and Section 2 to the description of some of its applications. If if is a closed subset of the domain of F (assumed to be independent of the time for simplicity), we say that a solution x(-) to a differential inclusion is viable if x(t) remains in К for all t. The purpose of the Viability Theorem is to characterize the viability property: for any initial state xq € K, there exists at least one viable solution.
10.1. The Viability Theorem 387 When F is upper semicontinuous with closed convex images, К enjoys the viability property if and only if К is a viability domain of F, i.e., if and only if Vz € K, F{x)nTK{x) φ 0 When F = f is single-valued, this theorem has been proved by Nagumo in 1942 and rediscovered 14 times since4. The above set- valued version has been proved by G. Haddad in 1981. One can also show that if К is not a viability domain of F, it still contains the largest closed viability domain, called the viability kernel. This result provides a quite useful tool. The second section is devoted to applications of the Viability Theorem. For instance, differential inequalities of the type Vi € [0.Г], V(x(t)) < w(t) can be regarded as a viability constraint of the form Vi € [Ο,Γ], (x(t),w(t)) € EV{V) so that the Lyapunov method can be adapted to differential inclusions and to functions V which are merely lower semicontinuous. This allows us to deduce the existence of the smallest lower semicontinuous Lyapunov function larger than or equal to a given function. 10.1 The Viability Theorem In all this section Χ, Υ denote finite dimensional vector-spaces, except an explicit mention to the contrary. We describe the (non deterministic) dynamics of the system by a set-valued map F from the state space X to itself. Consider initial value problems (or Cauchy problems) associated with differential inclusion for almost alii € [Ο,Γ], x'(t) € F(x(t)) (10.6) satisfying the initial condition x(0) = xq. 4This does not imply necessarily that it is true.
388 10- Differential Inclusions 10.1.1 Solutions to Differential Inclusions We have first to agree on what we shall call a solution to such a differential inclusion. Denote by L1(0,oo·, X, e~btdt) the space of integrable functions for the measure e~btdt (weighted Lebesgue spaces) and by ' i) W := W0 := π^^,Τ-,Χ) if Γ .< +oo < w ii) W-ь := W^1(0,oo;X,e-btdt) if Γ = +oo (for some b > 0) the weighted Sobolev spaces defined by ' W := {s(-) € Ρ{0,Τ·Χ) | s'(.) € Р{0,Т;Х)} < W-b ■= {x(·) € L^O.oo;X, e"btdt) \ χ'(·) € ^(0,οο;Χ, e'^dt)} We shall supply them with the topology for which a sequence xn{·) converges to x(·) if and only if г) χη(·) converges uniformly to χ(·) (on compact intervals if Τ = +oo) < ii) x'n(·) converges weakly to x'(·) in L1 (0, T; X) or in ^(Ο,οο-,Χ,ε^άί) if Γ = +oo Definition 10.1.1 (Viability and Invariance Properties) Let К be a subset ofOom(F). — A function x(-) from [Ο,Γ] to X is called viable in К if Vi € [Ο,Γ], x(t) € К — We shall say that К enjoys the local viability property or controlled invariance (for the set-valued map F) if for any initial state xq in K, there exist Г > 0 and a solution on [Ο,Γ] to differential inclusion (10.6) starting at xq and viable in K. It enjoys the global viability property (or, simply, the viability property^ if we can take Γ = oo. — We shall say that a subset К С Dom(F) is a viability domain of F if and only if \/x € K, F(x)ilTK(x) φ 0 (Ю.7)
10.1. The Viability Theorem 389 Figure 10.1: Example of a Map with Nonconvex Values 1 i i о There is no solution starting at 0 Since the contingent cone to a singleton is obviously reduced to 0, we observe that a singleton {x} is a viability domain if and only if ж is an equilibrium of F, i.e., a stationary solution to the inclusion 0 € F{x). In other words, the equilibria of a set-valued map provide the first examples of viability domains, actually, the minimal viability domains. Open subsets of the domain of F are viability domains of F, since contingent cones to open subsets are equal to the whole space. 10.1.2 Statements of the Viability Theorems Viability theorems hold true for the class of nontrivial upper semi- continuous set-valued maps with nonempty compact convex images. We observe that the only truly restrictive condition is the convexity of the images of these set-valued maps, since the continuity requirements are minimal. But we cannot dispense with it, as the following counter example shows: Let us consider X := R, К := [—1, +l] and the set-valued map
390 10- Differential Inclusions F: K^R defined by F(x) := {-1} if χ > 0 {-1,1} if χ = 0 {+1} if ж < 0 Obviously, no solution to the differential inclusion x\t) € F(x(t)) can start from 0, since 0 is not an equilibrium of this set-valued map! Theorem 10.1.2 (Haddad's Theorem) Let us assume that i) F : X ~> X is upper semicontinuous ii) the images of F are convex and compact Hi) К is locally compact Then К enjoys the heal viability property if and only if К is a viability domain of F. In other words, К is a viability domain if and only if for any initial state xq in AT, there exist Τ > 0 and a solution on [0, Γ] to differential inclusion (10.6) starting at xq and viable in K. Since open subsets of finite dimensional vector spaces are locally compact viability domains, we obtain the extension of Peano's Theorem to differential inclusions: Theorem 10.1.3 Let Ω be an open subset of a finite dimensional vector space X and F : Ω ~> X be an upper semicontinuous set- valued map with convex compact images. Then, for all xq € Ω, there exists Γ > 0 such that differential inclusion (10.6) has a solution on the interval [0, T) starting at xq- The interesting case from the global point of view is the one when the viability subset is closed. In this case, we derive from Theorem 10.1.2 a more precise statement. Theorem 10.1.4 (Local Viability Theorem) Let us consider an upper semicontinuous set-valued map F : X ~> X with compact convex images and a closed subset К С Dom(F).
10.1. The Viability Theorem 391 Then К is α viability domain if and only if it enjoys the local viability property. Actually, for any xq € К there exist a positive Τ and a solution on [О, Г] to differential inclusion (10.6) starting at xq and viable in К such that either Τ = oo or Τ < oo and lim supt_>T„ ||a;(i)|| = oo. Further adequate information —a priori estimates on the growth of F — allow exclusion of the case when limsupi_>r_ ||a;(i)|| = oo. This is the case for instance when F is bounded on K, and, in particular, when К is bounded. More generally, we can take Τ — oo when F enjoys linear growth: Definition 10.1.5 We say that a set-valued map F : X ~> X has a linear growth if there exists с > 0 such that Vi € K, \\F{x)\\ := sup ||v|| < c(||s||+l) veF(z) We shall call Peano maps nontrivial upper semicontinuous set-valued maps with nonempty compact convex images and with linear growth, or equivalently, nontrivial closed set-valued maps with convex values and linear growth. Theorem 10.1.6 (Viability Theorem) Let us consider a Peano map F from X to X and a closed subset К С Dom(-F). If К is a viability domain, then for x0 € K, there exists a solution on [0, oo[ to differential inclusion (10.6) starting at xq, which is viable in К and belongs to the space И^1'1(0, oo; X, e~btdi) for some b>0. Let us consider now a sequence of closed viability domains of a set-valued map F. We would like to address the following stability question: Is the upper limit of these closed viability domains still a closed viability domain"? Theorem 10.1.7 (Stability of Viability Domains) Let us consider a Peano map F : X ~> X. Then the upper limit of a sequence of closed viability domains of F is also a closed viability domain of F.
392 10- Differential Inclusions 10.1.3 Viability Kernels Definition 10.1.8 Denote by S(xq) the (possibly empty) set of solutions to differential inclusion (10.6) on [0,oo[ starting at xq € X. We call the set-valued map Dom(F) э χ н-> S(x) the solution map of F (or of differential inclusion (10.6).) Theorem 10.1.9 (Continuity of the Solution Map) Let us assume that F : X ~> X is a Peano map. The solution map S is upper semicontinuous with compact images from its domain to the space C(0, oo\X). Definition 10.1.10 (Viability Kernel) Let К be a subset of the domain of a set-valued map F : X ~> X. We shall say that the largest closed viability domain contained in К (which may be empty) is the viability kernel of K. Theorem 10.1.11 Let us consider a Peano map F : X ~> X and a closed subset К С Dom(.F). Then the viability kernel of К exists (possibly empty) and is the subset of initial states such that at least one solution starting from them is viable in K. The viability kernels may inherit properties of both F and K. For instance, if the graph of F and the subset К are convex, so is the viability kernel of K. If F is a closed convex process (i.e., its graph is a closed convex cone) and if К is a closed convex cone, the viability kernel is a closed convex cone. In general, viability kernels are not necessarily connected. The viability kernel contains for instance all the limit sets L(x) of the solutions x(-) to differential inclusion (10.6) defined by L(x(-)) := f|d(x([r,oo[)) T>0 Indeed, they are closed viability domains:
10.1. The Viability Theorem 393 Theorem 10.1.12 (Limit Sets are Viability Domains) Let us consider a Peano map F : X ~> X. Then the limit sets of the solutions to differential inclusion (10.6) are closed viability domains. In particular, the limit of a solution x(-) to differential inclusion (10.6) when t —> +oo, when it exists, is an equilibrium of F. The trajectories of periodic solutions to the differential inclusion (10.6) are ako closed viability domains. 10.1.4 Viability and Equilibria Equilibrium Theorem 3.2.1 states that any compact convex viability domain of a Peano map contains an equilibrium. When К is a compact viability domain, the convexity of the image F(K) implies also the existence of a viable equilibrium·. Theorem 10.1.13 Let F be a Peano map. If К С Oom(F) is a compact viability domain and if F(K) convex, then there exists an equilibrium of F in K. Proof — Assume that there is no equilibrium. This means that 0 does not belong to the closed convex subset F(K), so that the Separation Theorem implies the existence of some ρ G X* and e > 0 such that sup <v,-p>= a(F(K),-p) < -ε veF(K) Let us take any solution x(-) to differential inclusion (10.6) viable in K, which exists by the Viability Theorem. We deduce that Vi > 0, (-p,x'(t)) < -ε so that, integrating from 0 to t, we infer that Vt > 0, et < <p,x(t)-x(0)> But К being bounded, we thus derive a contradiction. Π With the same kind of techniques, we derive the following criterion on the existence of an equilibrium:
394 — 10- Differentia,! Inclusions Theorem 10.1.14 Let us assume that F is upper hemicontinuous with closed convex images and that К С Dom(F) is compact. If there exists a solution x(-) of (10.6) viable in К such that then there exists an equilibrium of F in K. inf - / \\χ'(τ)\\άτ = 0 i>0 Proof — Let us assume that therels ho viable equilibrium, i.e., that for any χ G К, О does not belong to F(x). Since these subsets are closed and convex, the Separation Theorem implies that there exists ρ in the unit sphere of X and ev > 0 such that <r{F{x),-p) < -ep In other words, we can cover the compact subset К by the subsets Vp := {x € K\ a{F{x),-p) < -ep} when ρ ranges over Σ. Since they are open, thanks to the upper hemicontinuity of F, К can be covered by q open subsets VPi. Set e := min eVi > 0 j = l,...,q ^ Consider now any solution to differential inclusion (10.6) viable in K. Hence, for any t > 0, x(t) belongs to some VPj, so that -||s'(i)|| < (-Pj,x'(t)) < a(F(x(t)),-Pj) < -ε and thus, by integrating from 0 to t, we have proved that there exists e > 0 such that for all t > 0, e < -tj\\x'{r)\\dT a contradiction of the assumption of the theorem. □ 10.2 Applications of the Viability Theorem In this section, we shall posit the linear growth conditions which guarantee the existence of solutions to the differential inclusions on the half-line [0, oo[ (Γ = +oo.)
10.2. Applications of the Viability Theorem 395 10.2.1 Linear Differential Inclusions We now take for the right-hand side of F a closed convex process and for viability domain a closed convex cone K. We recall that by- Lemma 4.2.5, Vi e K, TK{x) = K + Rx Corollary 10.2.1 Let F : X ~> X be α closed convex process and К С X be a closed convex cone. We posit the following assumptions: { i) V χ € K, R(x) := F(x) Π Τκ{χ) φ 0 ) ii) the norm ||Д|| of R is finite Then, for any initial state xq G K, there exists a solution x{·) to the "linear differential inclusion" for almost all ί > 0, x'(t) € F(x(t)) (10.8) starting at xq and viable in the cone K. 10.2.2 Lyapunov E4inctions We consider differential inclusion (10.6) and a time-dependent function w(·) defined as a solution to the differential equation w'(t) = -ip{w{t)) (10.9) where ψ : R+ ·-» R is a given continuous function with linear growth. Our problem is to characterize functions enjoying the ψ -Lyapunov property, i.e., nonnegative extended functions V : Χ ι—► R+ U {+00} (such that Dom(V) с Dom(F)) satisfying Vt > 0, V(x(t)) < w{t), to(0) = V(x0) (10.10) along at least one solution x(-) to differential inclusion (10.6) starting at xq where w is a solution to differential equation (10.9.) Since this condition amounts to saying that the epigraph of V enjoys the viability property for the differential inclusion (x?(t),w'(t)) € G{x(t),w(t)) where G(x,w) ·.= F(x) χ {-ip(w)} we can apply the Viability Theorem on the epigraph of V.
396 10- Differentia,! Inclusions This allows us to use lower semicontinuous instead of differen- tiable functions among the candidates to satisfy this Lyapunov property. We have then to translate the fact that Sp(V) is a viability domain of G, and for that purpose, to use Proposition 6.1.4 which states that the contingent cones to the epigraph are the epigraphs of the contingent epiderivatives: This leads us to extend the concept of Lyapunov functions to lower semicontinuous functions V:X ^ RU{+oo} We recall that V is said to be contingently epidifferentiable at χ if its contingent epiderivative never takes the value — oo (see Definition 6.1.2.) Definition 10.2.2 (Lyapunov Function) Let V : X н-> R+ U {+oo} be a contingently epidifferentiable function. We shall say that V is a Lyapunov function of F associated with φ if and only if V is a solution to the contingent Hamilton-Jacobi inequalities Vx 6 Dom(F), inf DfV(x)(v) + 4>(V(x)) < 0 (10.11) v£F(x) Theorem 10.2.3 Let V : Χ i-> R+ U {+00} be a contingently epidifferentiable lower semicontinuous function and F : X ~> X be a Peano map. Then V is a Lyapunov function of F associated with φ(·) if and only if for any xq G Dom(V), there exist solutions x(-) to (10.6) starting at xq and w(-) to (10.9) satisfying property (10.10.) Example W-Monotone Set-Valued Maps Let 1У:Хн R+ и {+оо} be a nonnegative extended function. We say that a set-valued map F is И^-monotone (with respect to ψ) if for every x, у £ X, we have \/u€ F(x), υ e F(y), D^W{x - y)(v - u) + φ{λ¥{χ -y))<0 (10.12) By setting V{x) — W(x — x) we obtain the following consequence of Theorem 10.2.3:
10.2. Applications of tie Viability Theorem 397 Corollary 10.2.4 Let W : X i-> R+ U {+00} be a contingently epidifferentiable, lower semicontinuous extended function and F : X ~> X be a Peano map such that —F is W-monotone with respect to ψ. Let χ be an equilibrium of F (i.e., 0 G F(1e)). Then, for any xo € Oom(F), there exist solutions x(-) and w(-) to (10.6), (10.9) satisfying s(0) = s0, ω(0) = W(xq-x) and Vi > 0, W(x(t)-x) < w(t) In particular, for W(z) := |||.г||2, we find the usual concept of monotonicity (with respect to φ): V x,y £ X,u (Ε F(x),v G F(y), (u-v,x-y) > φ (-\\x - y\\2 J D The function U : X ·-» R+ U {+00} being given, we can construct the smallest lower semicontinuous Lyapunov function larger than or equal to U, i.e., the smallest nonnegative lower semicontinuous solution Ό ψ to the contingent Hamilton-Jacobi inequalities (10.11) larger than or equal to U. Its epigraph is the viability kernel of the epigraph of U: Theorem 10.2.5 Let us consider a Peano map F : X ~> X, a continuous function ψ : R+ >—» R with linear growth and U:X^ R+U{+oo} such that Dom(!7) С Dom(-F). Then there exists a smallest nonnegative lower semicontinuous solution ΙΙφ : Dom(.F) 1-»· RU {+00} to the contingent Hamilton-Jacobi inequalities (10.11) larger than or equal to U (which can be the constant +00), which enjoys the property: V so € Dom(!7¥,) there exist solutions to (10.6), (10.9) satisfying < ^ s(0) = s0, U(xo) = w(0), Vi>0, U(x{t)) < υφ(χ(ί)) < w{t)
398 10- Differential Inclusions Let us single out the following consequence when we take for function U := ам the distance to a closed set Μ с Χ: Corollary 10.2.6 Let us consider а Реапо тар F : Χ ~> X. Then for all a > 0, there exists a smallest lower semicontinuous function dMa--X ·-> RU{+oo} larger than or equal to ам such that V xo € Oom(d,Ma), there exists a solution x(-) to (10.6) such that x(0) = so, Vi > 0, dM(x(t)) < dMa(xQ)e~at Therefore, we can regard the subsets Oom(dMa) as the basins of exponential attraction of M. 10.2.3 Tracking a Differential Inclusion Let us consider a differential inclusion (10.6) where F : X ~> X is a Peano map and an observation map Η : X ~> У from X to another finite dimensional vector-space Y. We shall in some sense "project" the differential inclusion (10.6) to a differential inclusion on the observation space Υ described by a set-valued map G for almost all t > 0, y'(t) e G(y(t)) (10.13) in order to "track" (or "filter") a solution x(-) to differential inclusion (10.6) in the following sense: V xq € Dom(-F) and yo € H(xq), there exist < solutions x(-) & y(-) to (10.6), (10.13) such that x(0) = xo, 2/(0) = г/о and V t > 0, y(t) e H{x(t)) (10.14) This property may be called the tracking property. Proposition 10.2.7 Let us consider a closed set-valued map Η from XtoY and Peano maps F : X ~> X and G :Y ~> У'. Then tracking property (10.14) holds true if and only if for every у G lm(H), we have Vx € H-\y), F(x)nDH(x,y)~1(G(y)) φ 0 (10.15)
10.3. Nonlinear Semi-Groups 399 It follows obviously from the Viability Theorem 10.1.6, because the above condition amounts to saying that V(x,y) € Graph(tf), (F(x)xG(y))nTGmmH)(x,y) φ 0 i.e., that the graph of Η is a viability domain5 of the set-valued map FxG. 10.3 Nonlinear Semi-Groups We provide now an existence and uniqueness theorem for differential inclusions the right-hand side of which is minus a maximal monotone set-valued map: Theorem 10.3.1 (Crandall-Pazy) Let A be α maximal monotone set-valued map from a Hilbert space X to X. Consider the initial value problem for the differential inclusion x'{t) € -A(x(t)), s(0) = xo where the initial state Xq is given in Oom(A). Then it has one unique solution x(-) defined on [0, oo[. Furthermore if m(A(x)) denotes the element of A(x) with the minimal norm, then x(-) is actually the solution to the differential equation for almost all t > 0, x'(t) = -m(A(x(t))) (called the slow solution). Moreover, the function t —> ||a;'(i)|| is nonincreasing. Let x(-) := S(xq) and y(-) := S(yo) be the solutions starting at xq and ?/o· Then V* > o, ||s(0 - y(0ll < lko-2/o|| Finally, ί Vi > 0, x'{t) = Yunh^Q+(x(t-\-h)-x{t))/h у and x'(-) is continuous from the right In this case, the solution map S is single-valued and is called in the literature the nonlinear semigroup of the monotone map A. 5 When G = 0, i/o = 0 and Η is single-valued, the viability kernel of ff-1(0) is closely related to the "zero dynamics" introduced by Byrnes and Isidori.
400 10- Differential Inclusions 10.4 Filippov's Theorem Filippov's theorem for differential inclusions is as important as the Gronwall lemma for ordinary differential equations. Consider a finite dimensional space X, reals а < b and let F be a set-valued map from [a, b] χ Χ into closed nonempty subsets of X. We associate with it the differential inclusion for almost all t € [a, b], x'(t) £ F(t,x(t)) (10.16) Consider the solution map of (10.16), associating with each initial point xq G X the subset S[o,b](xo) :== {x I x is a solution to (10.16) on [a, 6], x(a) = xq} of solutions to differential inclusion (10.16) on [a, b]. Let у G W1,1(a, b;X) be a given absolutely continuous map. Filippov's theorem provides an estimation of the distance between у and the set S[ajb](xo) С W1'1(a,b;X) under the following assumptions on F: i) V χ € X, F(-,x) is measurable it) 3 β > 0, к G ^(α, 6; R+) such that for a.a. (10.17) ΐ € [a, 6], F(t, ·) is fc(i)-Lipschitz on y(t) + /35 иг) The map ti-^d(y'(t),F(t,y(t))) is integrable where В denotes the closed unit ball in X. Remark — Observe that under the above assumptions, the set- valued map t ~> F(t, y(t)) is measurable: Indeed, let yn be a sequence of measurable simple maps converging pointwise to y. Since the map t ^ F(t,yn(t)) is measurable, Theorem 8.2.5 and assumptions (10.17) imply the measurability of F(-, ?/(·))·
10.4. Filippov's Theorem 401 Therefore, Corollary 8.2.13 implies that the function t -> d(y'(t),F(t,y(t))) is always measurable. □ Theorem 10.4.1 (Filippov) Let us consider an absolutely continuous function у : [α, 6] н-► Χ, δ > 0 and assume that conditions (10.17) hold true. Set 7(0 = d{y'{t),F{t,y{t))), m(t) = exp (£ k(s)di and η(ί) = m(t) (s+f j{s)ds\ Ifv(b) < β, then for every xq € B(y(a), δ), there exists χ G 5[aib](a;o) such that Vie [a, b], И0-у(011 < v(t) and \W(t) - y'(t)\\ < k{t)v(t) + j(t) a.e. in [a, b] Corollary 10.4.2 (Lipschitz Dependence) Let у £ S[aib](yo) and assume that F, у satisfy assumptions (10.17) i) and ii). 'Then there exists I > 0 such that for all xq € X satisfying m(b) \\xq — 2/o|| < β we have dwi,i (y, S[ajb](xo)) < l\\xo- 2/o I We state next a "nonlinear version" of Aumann's Theorem on the convexity of the integral of a set-valued map. This theorem compares solutions to (10.16) and to the convexified (relaxed) differential inclusion: for almost all t £ [a, 6], x'{t) G со F(t,x(t)) (10.18) Theorem 10.4.3 (Filippov-Wazewski) Let у be a solution to the relaxed inclusion (10.18) on [a,b]. Assume that F and у satisfy assumptions (10.17) and that the map t ~> F(t,y(t)) is integrably bounded on [a,b]. Let η(·) be defined as in Theorem 10.4-1-
402 10- Differential Inclusions If 77(6) < β, then for every δ > 0 there exists α solution χ to (10.16) on [a,b] satisfying s(0) = ?/(0) & ||s-y||c < 5 The next theorem is a simple consequence of Theorem 10.4.3 and the Convergence Theorem 7.2.1. Theorem 10.4.4 (Relaxation Theorem) Let F :[a,b]x X *-* X be a set-valued map with closed nonempty images. Assume that F is measurable with respect to t and that there exists к G L1(a, b; R+) such that for almost every t G [a, b], F(t,-) is /c(i)-Lipschitz & V χ e X, F(t,x) С k(t)B Then the solutions to (10.16) are dense in the set of solutions to the relaxed inclusion (10.18) for the metric of uniform convergence. Corollary 10.4.5 Let S?°bAxo) denote the set of solutions to (10.18) on [a, b] with x(a) = xq. We posit all assumptions of Theorem 10.4-4· Then for every xq £ X, the closure of S[atb](xo) in the metric of uniform convergence is equal to <Su°u(£o)· Filippov's theorem allows us to investigate infinitesimal generators of reachable maps of the nonlinear dynamical system (10.16.) For all a < t < Τ < b and xq € X, define R(T,t)x0 = {x(T)\ χ € S[tiT\(xo)} It is called the reachable set of (10.16) (or of F) from (t,xo) at time T. Reachable sets enjoy the semigroup property Vi > s, R(T,s)x0 = R(T,t)R(t,s)x0 When F is sufficiently regular, the set coF(t, xq) is the infinitesimal generator of the semigroup R(-, t)xo in the sense that the "differential quotients" (R(t + h, t)xo — xo)/h converge to coF(t, xq).
10.5. Derivatives of the Solution Map 403 For that purpose, we posit the following assumptions: i) V (t, x) € [a, b] χ X, F(t, x) is nonempty and compact ii) V χ € X, F(-,x) is continuous on [a, 6] < ш) V (ίο, so) € [a, b] χ Χ, there exists a neighborhood N с[а,Ъ]хХ, L>0 such that V {t, x), (t, у) е Л/", F(i,s) С ^,у) + Ь||х-у||Б (10.19) Theorem 10.4.6 Assume that conditions (10.19) hold true. Then for every (ίο,^ο) € [a,b[xX, for all (t,x) near (ίο, so) and all small h>0 R(t + h,t)x = χ + /ι coF(i0,so) + o(t,x,h) (10.20) гу/iere Ит оЦМ = о . (ί,ι)—>(ίθ,ΐο). Ί—>0+ Д Remark — Equation (10.20) has to be understood in the following way R(t + h,t)x С x + hcoF(t0,x0) + \\o(t,x,h)\\B and x + hcoF(tQ,xQ) С R{t + h,t)x+\\o(t,x,h)\\B □ 10.5 Derivatives of the Solution Map We now provide estimates of the contingent, adjacent and circatan- gent derivatives of the solution map S := <S[o,T] associated with the differential inclusion for almost all t e [0,Г], x'(t) e F(t,x(t)) (10.21) where -F: [0,Γ] xl^l
404 10- Differential Inclusions In this section the set-valued map F is regarded as a family of set- valued maps χ ~> F(t, x) and the derivatives are taken with respect to the state variable only. Let ж be a solution of the differential inclusion (10.21.) We assume that F satisfies the following assumptions: г) V χ G X, F(-, x) is measurable ii) V t e [0, Г], VieX, F(t, x) is a closed set (10.22) Hi) 3 β > 0, k(-) € 1^(0, Γ) such that for almost all t e [0, Γ], F(t, ·) is k(t) - Lipschitz on x(t) + βΒ Consider the adjacent variational inclusion, which is the "linearized" along the trajectory χ inclusion ( w'(t) € L»bF(i,x(i),z'(i))K*)) a.e. in [0,Γ] Ί w(0) = и where и G X. In Theorems 10.5.1, 10.5.2 below we consider the solution map S as the set-valued map from Rn to the Sobolev space W^CO.TjRT1). Theorem 10.5.1 (Adjacent Variational Inclusion) If the assumptions (10.22) hold true, then for all и G X, every solution w G W1,l{Q,T;X) to the linearized inclusion (10.23) satisfies w e DbS(x(0),x)(u) In other words, {w{-)\ w'(t) € DbF{t,x{t),x'{t))(w(t)), w{0) = u}c DbS{x{0),x)(u) Proof— Let hn > 0 be a sequence converging to 0. Then, by the very definition of the adjacent derivative and Lipschitz continuity of F(t, ·), for almost all t e [0, Г], Ш. d Ut), *(«,»w+*.»(«»-*(«η _ 0 (10.24)
10.5. Derivatives of the Solution Map 405 Moreover, since x'{i) G F{t,x{t)) a.e. in [Ο,Γ] by (10.22), for all sufficiently large η and almost all t G [Ο,Γ] d(x'(t) + hnw'{t),F(t,x(t) + hnw(t))) < hn(\\w'{t)\\ + k{t)\\w{t)\\) This, (10.24) and the Lebesgue dominated convergence theorem yield L ■T d(x'{t) + hnw'(t),F(t,x(t) + hnw{t)))dt = o{hn) (10.25) where limn_>oo o{hn)/hn = 0. By the Filippov Theorem and by (10.25) there exist Μ > 0 and solutions yn G S(x(0) + hnu) satisfying for all large η Ьп-х1 - Κ-ω'\\ί1{0ιΤ,χ) < Mo{hn) Since (yn{0)-x{0))/hn = и = w(0) this implies that lim VJLZ^ = w mC(0,T;X); lim Vn~~X = w' ϊηΐΛθ,Γ;Χ) тг-юо hn 71-4CO hn Hence lim / зто)+к„)-шл=0 я-юо \ hn J Since ?x and ги are arbitrary the proof is complete. □ Consider next the circatangent variational inclusion, which is the linearization involving circatangent derivatives: ί w'(t) G CF(t,x(t)>x'(t))(w(t)) a.e. in [Ο,Γ] 1 ги(0) = u ^ ' ' where и £ X. Theorem 10.5.2 (Circatangent Variational Inclusion) Let us assume that conditions (10.22) hold true. Then for all и G X, every solution w G ^■'^(Ο,Γ-,.Χ') to the linearized inclusion (10.26) satisfies w£ С5(ж(0),ж)(?х).
406 10- Differential Inclusions In other words, {w(-) | w'{t) € CF(t,x(t),rf(t))(w(t)),'w(0) = u} C CS{x{0),x)(u)- Proof — Consider a sequence xn of solutions to (10.21) converging to ж in W1,1(0, T; X) and let hn —> 0+. Then there exists a subsequence Xj .— Хщ such that lim x'-(t) = ж'(£) a.e. in [0,T] " (10.27) i—»oo Set Xj = /in.. Then, by the definition of circatangent derivative and by (10.27), for almost all t e [0, Γ] to лU),^wHy-tf»-*^ _ 0 Moreover, using the fact that x'j(t) G F(t,Xj(t)) a.e. in [0, T],'we obtain that for almost all t € [0, Г] and all large j d (x'jit) + Xjw'itlFfaxjtt) + Xjw(t)Yj < \j (\\w'{t)\\ + k(t) \\w(t)\\) This, (10.28) and the Lebesgue dominated convergence theorem yield ( d(a;^(i) + AJV(i),F(i,a;i(i) + Ajii;(i)))di = o{Xj) (10.29) where limj_>coo(Aj)/Aj = 0. By the Filippov Theorem and (10.29), there exist Μ > 0 and solutions yj G S(xj(0) + Xju) satisfying IIyj - x'j - Xjw'\\ < Mo{hj) Since (yj(0)-Xj(0))/Xj = и = ги(0) this implies that Pi ~ ai - ... ;„ ,νη т. vv „„ ?ir^i _ Hence lim Οΐ—Ξί = ω inc(0,T;X); Urn L^ = ω' ίηΐ/(0,Τ;Χ) J—>CO Aj j—>CO Xj lim d L S(*№ + hnju)-Xj\ = o 30) J-»00 V bnj /
10.5. Derivatives of the Solution Map 407 Therefore we have proved that for every sequence of solutions xn to (10.21) converging to χ and every sequence hn —> 0+, there exists a subsequence which satisfies (10.30.) This yields that for every sequence of solutions xn converging to χ and hn —> 0+ v , / S(xn(0) + hnu) - xn\ ,hm α w, —-—i = 0 ™->°° V. hn J Since и and w are arbitrary the proof is complete. □ We consider now the contingent variational inclusion \w'{t) e coDF{t,x{t),x>{t)){w{t)) a.e. in [0,T\ (1Q 31) 1 w(0) = и V ■ ) Theorem 10.5.3 (Contingent Variational Inclusion) Consider the solution map S as a set-valued map from Rn to И^1,оо(0, Г; Rn) supplied with the weak-* topology and let x(-) be a solution to differential inclusion (10.21) starting at xq. Then the contingent derivative DS(xo,x(·)) of the solution map is contained in the solution map of the contingent variational inclusion (10.31), in the sense that DS{xq,x{-)){u) С {w{-)\ w'{t) e WDF(t,-x(t),s?(t))(w(t)),w(p) = u} Proof — Fix a direction u£R" and let w(-) be an element of DS(xo,x(-))(u). By definition of the contingent derivative, there exist sequences of elements hn —> 0+, Un —> и and wn(·) —> w(-) in the weak-* topology of W1'00^, Г; Rn) and c> 0 satisfying ' 0 IKCOII < c a-e- in [°>r] ii) x'(t) + hnw'n(t) e F(t,x(t) + hnwn{t)) a.e. (10.32) Hi) wn(0) = un Hence i) wn(·) converges pointwise to w(-) (10.33) ii) w'n{·) converges weakly in Χ^Ο,Γ-,Κ71) to w'(-)
408 10- Differential Inclusions By Mazur's Theorem and (10.33) и), a sequence of convex combinations CO vm{t) := J2 a^w'pit) p=m converges strongly to w'(·) in I'1(0, Г; X), where a^ > 0, CO p=m and for every m, a^ ^ 0 only for a finite number of p. Therefore a subsequence (again denoted) vm{-) converges to w'(-) almost everywhere. By (10.32) i), ii) for all ρ and almost all ΐ € [0, T\ w'p(t) e —{F(t,x(t)+hpwp(t))-xl{t))ncB tip Let t € [0,Γ] be a point where wm(i) converges to w'(t) and x'(i) g -F(i, x(t)). Fix an integer η > 1 and e > 0. By (10.33) г), there exists m such that hp < 1/n and ||гир(г) — w(t)\\ < 1/n for all ρ > т. Then, by setting $(y,h) := ±{F(t,x(t) + hy)-x'(t))ncB we obtain that / vm(t) & со [J Ф(у,/1) \ье]о,$[,у&оЮ+±в and therefore, by letting m go to oo, that / w'(t) £ со (J Ф(у,/1) ^е]о,^],уешМ+^в Since the subsets Ф(у, /г) are contained in the ball of radius c, we infer that u/(i) belongs to the closed convex hull of the upper limit by Lemma 1.1.9. Hence / w ''(0 e» Π ε>0, η > 1 (J Ф(у,/0 + еЯ \he]o,±],vew(t)+±B
10.5. Derivatives of the Solution Map 409 We observe now that Π U $(y,h)+eB \cDF(t,x(t)^(t))(w(t)) ε>0;π>1 ^e]0,i],yew(i)+iB J to conclude that w(-) is a solution to the contingent differential inclusion w'{t) e coDF(t,x(t),x'(t))(w(t)) a.e. in [Ο,Γ] w(0) = и Since w G DS(xo,x(-))(u) is arbitrary we end the proof. Π
411 Bibliographical Comments Chapter 1 We already have mentioned that the concepts of upper and lower limits of sets are due to Painleve in 1902 [318] as it is reported by his student Zoretti in [467]. The compactness result (Theorem 1.1.7) was obtained by Zarankie- wicz [462]. Many other proofs were provided since. The Duality Theorem 1.1.8 was proved first by Walkup & Wets [427]. Some basic calculus of upper and lower limits of sets and set- valued maps appeared before and after the second World War in the two volumes of Kuratowski's basic monograph [255]. The continuous version of Proposition 1.2.6 on the lower limits (or lower semiconti- nuity) of intersections of convex-valued maps can be found in [35] for instance. The square product of set-valued maps was introduced and used in>[40] for studying local observability of control systems under uncertainty. Upper and lower semicontinuous maps were introduced by Bouli- gand [80] and Kuratowski [254] and [255]. Berge's-book [66] played an important role in disseminating these results among mathematical economists, as one can see by reading Debreu's basic monograph [132]. The convergence issues were then reexamined by Choquet in [100] in a more general case. Pseudo-Lipschitz set-valued maps have been introduced in [47,49] in the framework of the Inverse Function Theorem in finite-dimensional spaces, and later, in [39] for Banach spaces, and have also been studied by Rockafellar in [370]. The Genericity Theorem is due to Kuratowski [254,255]. It was extended later by Choquet [100] and Shi Shuzhong [388]. - The Maximum Theorem appeared in Berge's monograph [66]. The lower semicontinuity criterion of an infinite intersection of lower semicontinuous set-valued maps was proved in the context of differential games [51]. - -
412 Bibliographical Chapter 2 Closed convex processes and their transposes have been introduced by Rockafellar [356] and further studied in [358,360]. Robinson- Ursecu's Theorem has been proved in [347,348,349,417] and the Uniform Boundedness and the Crossed Convergence Theorem in [42]. The Closed Image Theorem is taken from [35]. Closed convex process are related to Ioffe's fans ([233,238]) for which results of linear functional analysis can also be extended. Linear processes and their invariance properties were studied in [197]. Chapter 3 The literature on nonlinear analysis is too broad to give any fair account in this short review. Ky Fan's Inequality has been proved in [160] and the proof we gave is derived from Browder in finite dimensional vector-spaces and from [45,35] in the general case, where extensions are provided and further bibliographical comments can be found. The Equilibrium Theorem was derived from Ky Fan's papers [159,161] in [45]. See also recent contributions of Simons [395]. One can find in [35] further developments on the existence of solutions to inclusions. Kakutani's Theorem appeared in [243] and the set-valued version of the Leray- Schauder Theorem in [45]. Ekeland's variational principle has been proved in [154] and used very often since then. For more recent views, see [35, Chapter 5], [156] and their bibliography. The proof given here is due to Siegel [393]. One can find in [2] the history of the concept of derivative of maps between Banach spaces, from Peano, Volterra, Gateaux and Frechet to the eve of the nonsmooth era. We mention only the following [323,201,199,200]. The Inverse Function Theorem for surjective single-valued maps in Banach spaces is due to Graves [205,206]. See also the earlier paper [220] in collaboration with Hildenbrandt. Ljusternik proved in [273] that the tangent space to /_1(0) is the kernel of f'(x) when the Frechet derivative /' is continuous and surjective (See [272] for instance.) This has been extended by Leach in [265] to strong differ-
Comments 413 entiable maps. The extension of this theorem to problems with constraints and set-valued maps in finite dimensional spaces appeared in [266,46,47, 49] and [35, Chapter 7]. The use of the localization property of Ekeland's Variational Principle is due to Lebourg [266]. It has been adapted to the case of infinite dimensional spaces in [39,178,182,189, 194]. This theorem has been extended to the case when the surjectivity of the first derivative is lacking, and also, when the set-valued map is defined on metric spaces, thanks to the concept of high order variations introduced in [177,182,194,189]. Such results were motivated by control theory, in particular by local controllability of control systems in [180,190]. See [194] and the forthcoming monograph [195] for an exhaustive presentation of the theory and other applications. Monotone maps have been introduced by Zarantonello [463]. For a comprehensive account of monotone maps, see for instance the books of Brezis [83], J.-L. Lions [270] and Browder [84]. Maximal monotone maps were characterized by Minty [290] and the fact that the subdifferential is maximal monotone is due to Rockafellar [354]. See for instance [35, Chapter 6] for developments of this important domain of nonlinear analysis and further bibliography. The existence of eigenvectors of a closed convex process on cones with compact soles appeared in [31] for characterizing controllability of'closed convex processes. Existence of positive eigenvectors and extension of the Perron-Frobenius Theorem appeared in [43] and can be-found in [35, Chapter 3]. Chapter '4 It is impossible to give an exhaustive account of the many papers in which the various concepts of tangent cones appeared in the literature. They have been introduced again and again in a manifold of contexts, many of them unknown to the authors. The need to introduce contingent and paratingent directions was felt by Bouligand in the thirties in order to differentiate non differen- tiable functions (see [78,80,79].) This was taken up by Zaremba [465] and Marchaud [280] to define solutions to differential inclusions. See also the paper [100] by Choquet. The contingent cone was used under the names of tangent cone and sequential tangent cone in optimization and control theory in
414 BibJiograpbicai the fifties (see for instance the books by Neustadt [300] and Hestenes [218]) overshadowed during the sixties when convex analysis was expending, and was resurrected in this field as a parent pauvre of the tangent cone introduced by Clarke in the middle of the seventies in his thesis [106]. Clarke and Rockafellar's papers extended the main results of convex analysis to the case of locally Lipschitz functions and lower semicontinuous functions respectively. See also Clarke's book [113]. Meanwhile, Russian and Eastern European mathematicians did use many of these cones. Let us mention the books by Dubovitskij & Miljutin [147], by Ioffe & Tikhomirov [229] and by Pchenitchny [321] for instance, just to cite a few. This period saw an explosion of notions of tangent cones. We chose to emphasize only three of them, the contingent cone since it appears that it is the one which is at the origin of most theorems, the adjacent cone which is needed in Lebesgue and Sobolev spac'es (see [172,173,183,181,186,196]), and the tangent cone introduced by Clarke, because of its regularity and convexity properties. The fact that it coincides with the contingent cone for sleek subsets was first proved in [34] in finite dimensional vector-spaces. The proof that the lower limit of the contingent cones is contained in the Clarke tangent cone is due to Cornet in [118,117] in the case of Hilbert spaces (see also [325]) and the extension to any Banach space is due to Treiman [411]. We provided here a different proof based on Ekeland's Variational Principle. See also [72] on this topic. The characterization of the polar of a viability domain of a closed convex process appeared in [31,32] to study the duality relations between controllability and observability of closed convex processes. The calculus of tangent cones to intersections and inverse images appeared in [47] (finite dimensional case, generalizing formulas due to Rockafellar in [361] under stronger transversality conditions) and [39] (Banach spaces.) Theorem 4.3.9 was shown to us by Rockafellar. A discussion of results on upper limits of normal cones, which we deduce by polarity from the relations between the lower limit of the contingent cones and the Clarke tangent cones can also be found in Rockafellar and Wets' book [355]. These authors also proved that the polar cones to the contingent cones coincide with the sets of proximal normals in finite dimensional vector-spaces. See also [294,232,233,237] and [365,72,70,104] among other contributions to these issues, which were not treated in this book.
Comments 415 Convex kernels (or asymptotic cones) of adjacent (or intermediate) cones have been used in [172,173,181] in the study of optimal control of differential inclusions because the use of the mere convexity of Clarke tangent cones required too strong properties in the infinite dimensional spaces of solutions. The results on paratingent cones and hypertangent cones are taken from Shi Shuzhong's papers [391,392]. Ioffe's menagerie appeared in [236]. See also [122], [283] among other contributions. The results presented in Sections 4.6 and 4.7 appear here for the first time. Chapter 5 There were many definitions of pointwise derivatives of set-valued maps proposed in the literature. Let us mention only Banks- Jacobs' [58], Martelli-Vignoli's [282], DeBlasi's [125] and Petcherskaja's [326] to name a few. The concepts of graphical derivatives and their use in the extension of the Inverse Function Theorem have been initiated in [46] for the contingent derivative and in [47,49] for the circatangent derivatives. Adjacent derivatives have been introduced and used in control theory in [172,173,183,181,186,196]. See also Polovinkin [332] who also used the graphical approach to define derivatives of set-valued maps, as well as [378,373]. Proposition 5.2.6 was proved in [181]. Circatangent derivatives of the subdifferential of a convex function, regarded as generalized Hessians, have been introduced in [47, 49] to study the stability of solutions to convex minimization problems. More generally, derivatives of maximal monotone maps have been extensively studied by Rockafellar in [371]. The idea to use Ekeland's Variational Principle for obtaining local surjectivity criteria for single-valued maps is due to Ekeland and was taken up in [46] in the case of set-valued maps by involving the surjectivity of the contingent derivative. This result was improved by Lebourg in [266], who used the localization property of Ekeland's Variational Principle. This appoach has been further exploited in [47,49,39,177,178,182] to provide more and more general versions of this useful theorem. Analogous approaches have been used by Ioffe, in [238] for instance. For nondifferentiable single valued maps, an inverse function theorem using generalized Jacobian can be found injthe papers [109,113]
416 Bibliographical of Clarke. See also the contributions [434] of Warga, [214] of Halkin, [302], and [232,237,238] of Ioffe, who uses the concept of fans, and the references of these papers to Russian literature. The inverse mapping theorem has been extended to the case when the set-valued map is defined on a metric space, thanks to the concept of high order variations introduced in [182,178,189]. Local inverse univocity and injectivity of set-valued maps has been proved in [39,38,40] and used in the study of local observability of control systems and differential inclusions. The localization results are new. The new field of "Qualitative Physics" (see for instance [67,130]), an active branch of Artificial Intelligence, as well as of mathematical economics (see [259,260]), provide many new motivations for using the differential calculus of set-valued maps and set-valued analysis. The results on qualitative analysis appear here for the first time. Chapter 6 The characterization of the sub differential of a convex function through the normal cone to its epigraph goes back to Moreau and Rockafellar when they introduced the concept of subdifferential. Clarke used this characterization in [106] to define the generalized gradient of any lower semicontinuous function through the normal cone to the epigraph. The dual relation between subdifferential of locally Lipschitz functions and the epiderivatives (regarded as the support function of the subdifferential) was established by Clarke. Rockafellar extended it to lower semicontinuous functions in [362,363, 364,366]. Contingent epiderivatives have been introduced in [46] in the study of Lyapunov functions for differential inclusions, where they play a natural and crucial role. See also [33, Chapter 6] and" [50]. They are also fundamentally involved in the study of Hamilton- Jacobi equations and closely related to viscosity solutions: see [179, 185,188]. The proof of Theorem 6.1.6 is derived from an idea of Quincampoix. The epidifferential calculus has been developed in [39]. It has been established by Clarke [113] in the case of locally Lipschitz functions and by Rockafellar [362,363,364,366] under stronger transver- sality conditions than the ones used in [39]. Relations between the subdifferential and the local subdifferential used by Crandall, Evans and P.-L. Lions were elucidated in [191]. Lo-
Comments 417 cal sub and super differential were extensively exploited for studying viscosity solutions of Hamilton-Jacobi equations: see [119,120] and the bibliography of P.-L. Lions' book [271]. Other generalizations of the gradients or Jacobians using approximations procedures of functions or maps have been proposed. Let us mention the Warga's derivate containers [432], Halkin's [215] and [174] among other contributions using this approach. Demianov and Rubinov and their collaborators introduced pairs of generalized gradients to break down the lack of symmetry due to the epigraphical approach. See [140,141] for instance, and their bibliography. We refer to the books of Rockafellar [358], Moreau [296], Ekeland- Temam [149], [35, Chapter 4], and the forthcoming book [355] by Rockafellar and Wets for an account of convex analysis. For a short introduction to convex analysis, one can also consult [48]. Results of Section 6.5 appear here for the first time. Many suggestions have been made to define second-order approximations of functions. The one we propose fits in the unified approach of this book. Applied to the epigraph of a function, it yields Rock- afellar's extension (parabolic derivatives) given in [377,369] of analogous concept introduced by Ben-Tal and Zowe in [64]. The formula on the second order contingent epiderivative of the Moreau-Yosida approximation of a lower semicontinuous function is new. Another concept of generalized Hessian was introduced in [47, 49] by taking circatangent derivatives of subdifferentials of convex functions. In the forthcoming book [355] by Rockafellar and Wets, generalized Hessians are defined as со-differentials' of the generalized gradients. See also the contributions of Hiriart-Urruty on this topic. Chapter 7 This chapter is just a very short introduction to epigraphical (and graphical) convergence. The first results on this subject are due to Wijsman [454,454] and Walkup & Wets [427] in the convex case. It has been extended to infinite dimensional spaces and applied to the approximation of variational inequalities by Mosco [297], and then extended by Joly [242] and R. Robert [345].
418 Bibliographical Extension to the non convex case has known a considerable expansion since the resurrection of this topic with the concepts of Γ and G-convergence by de Giorgi and his collaborators (see for instance [127,126,128,128]) in the framework of calculus of variations, with the concepts of epiconvergence by R. Wets and his collaborators (see [447,379,380,381,382,448,450,451,146,25]) in the framework of convergence of random sets and functions and with the papers of Attouch motivated by the theory of homogenization in [27,28] and above all in his book [30]. We refer to this book and the forthcoming monograph by Rockafellar and Wets [355] for further references and an exhaustive presentation of this topic, for which this chapter can be used as an introduction. The results on graphical convergence of maximal monotone maps are due to Attouch [29] and the relations between lower graphical and pointwise convergence appeared in [42]. The Convergence Theorem has been used by several authors studying differential inclusions with upper semicontinuous convex- valued right-hand side (see [33, Chapter 1] for instance.) We state here a more general version of this theorem. The results on the limits of infima are taken from Wets [449]. The theorem on conjugate functions of epilimits was proved by Wijsman [453,454] in the finite dimensional case and by Mosco [297], Joly [242] and R. Robert [345] in the infinite dimensional case. The theorem on the lower limits of gradients of differentiable functions converging uniformly is new. It extends to the infinite- dimensional case a theorem of M. Crandall, C. Evans & P.-L. Lions [119] which allows one to study the stability of viscosity solutions to Hamilton-Jacobi equations. See P.-L. Lions' book [271] for further comments on sub differentials of viscosity solutions to Hamilton- Jacobi equations. Their relations with generalized gradients have been investigated in [188,89]. The theorem on convergence of subdifferentials of convex functions is due to Attouch [28,26]. See also his monograph [30]. Chapter 8 The measurable selection theorem, which, as many other important mathematical innovations of this century, is due to J. von Neumann. This result was given in a stronger form by Kuratowski and Ryll-Nardzewski [253]. In [92], Castainghas proved that measurable
Comments 419 selections are dense, the result which is often referred to now as the Castaing representation theorem. The interested reader can learn more about the history of this subject from Wagner's survey [426] of measurable selection theorems and also from Ioffe's survey [231] of Soviet literature on this subject. The book [91] by Castaing L· Valadier remains the reference for measurable set-valued maps. In particular, the proofs of Theorem 8.1.4 and of some its applications to the calculus of measurable set-valued maps are borrowed from this book. Using the Mackey topology it is possible to prove Theorem 8.2.14 without the assumption that X* is separable (see [91].) The characterization of tangent cones to subsets of Lebesgue spaces extends the results obtained in the convex case by Hess. The concept of integral has been introduced by Aumann [52] and the convexification result is due to Aumann [52], Debreu [133] and Olech [307,312,305]. The estimation Lemma and the Theorem on the convergence of selections to extremal points are taken from Olech [312]. Its more precise version was proved in [171] and used in [170] by Prankowska & Olech to study boundary trajectories of nonlinear control systems. The extension of the Convexity Theorem on the convexity of the closure of the integral to Banach spaces is due to Uhl [415] and Hiai- Umegaki [219]. • Other concepts of integral of set-valued maps have been proposed. Let us mention for instance Pontriagin's alternate integrals introduced in [333] and further studied by M. S. Nikolskii in [303,304]. The bang-bang principle is due to LaSalle and Olech. The proof we used is taken from Olech [312]. Existence of invariant measures of set-valued maps on compact spaces and the extension of Poincare Recurrence Theorem can be found in [41]. A very pedagogical account of ergodic theory is given in the monograph [261] by Lasota & Mackey. Chapter 9 The Continuous Selection Theorem is due to Michael [286,287, 288] and the Approximate Selection Theorem of an upper semicon- tinuous set-valued map to Cellina [96,97]. The idea to use the Steiner points of the images of a set-valued map as selection and parametrization procedures is due to Lojasiewicz
420 Bibliography Jr. in 1979 (О lech's seminar, see also [276]). The Steiner point was introduced by Steiner for convex polytopes of R2 in [402] and extended to convex polytopes of Rn by Griinbaum: see [207] and Grunbaum's book [208]. It has been generalized to convex bodies of finite dimensional vector spaces by Shepard in [384]. Rademacher's theorem implies that in Theorem 9.4.1 the subdif- ferential of the support function is almost everywhere single-valued. Hence the minimal selection in the formula of the Steiner point as an average of the faces on the unit ball can be replaced by any selection. Vitale proved that Steiner points cannot be extended continuously to compact convex subsets of a Hilbert space in [424]. The appendix of this paper provides a sharper estimate of the Lipschitz constant of the Steiner map than the simple one we gave. Similar Intersection Lemmas were proved by Le Donne-Marchi in [264] and by Lojasiewicz [276] and Ornelas [314]. The Parametrization Theorem for continuous maps is due to Ekeland-Valadier [151] and, for Lipschitz maps, has been proved by Lojasiewicz Jr. [276] and Ornelas [314] (with better constants, but more technical proofs.) Chapter 10 One can find an extensive bibliography on differential inclusions and their applications to control and viability theory in the books [33], [195] and [50] for instance. The Viability Theorem is due to Haddad [212,213]. For Fixed Point Theorems using properties of differential inclusions, see among other contributions Deimling's book [135] and also [136]. The Theorem on nonlinear semigroups is due to Crandall-Pazy, [164]. The original proof of Filippov's Theorem appeared in [169]. It can be also found in [33] for continuous with respect to the time set-valued maps. See also [196] for an infinite dimensional version of this theorem. The proof of Filippov-Wazewski's Theorem can be found in [169], [446] and [113] and for continuous set-valued maps in [33]. Its proof involves the Aumann and the Filippov's theorems. An infinite dimensional analogue was given in [196]. Derivatives of the solution map have been studied in [181,183,40,38,333,204].
bliography [1] ANICHINI G., CONTI G. к ZECCA P. (to appear) Approximation and selection for nonconvex multifunctions in infinite dimensional spaces, Preprint [2] ANNARATONE S. (1989) Breve storia del concetto di dif- ferenziale negli spazi vettoriali normati, Thesis Universita di Milano [3] ARTSTEIN Z. к PRIKRY K. (1987) Cartheodory Selections and the Scorza Dragoni Property, J. Math. Anal. Appl., 127,540-547 [4] ARTSTEIN Z. к WETS R. (1988) Approximating the integral of a multifunction, J. Multivariate Analysis, 24, 285-308 [5] ARTSTEIN Z. (1974) On the calculus of closed set-valued functions, Indiana Univ. Math. J., 24, 433-441 [6] ARTSTEIN Z. (1989) A Variational Convergence that Yields Chattering Systems, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier-Villars к C.R.M. Uni- versite de Montreal, 49-72 [7] ARTSTEIN Z. (1989) Piecewise Linear Approximations of Set- Valued Maps, J. Approx. Theory, 56, 41-47 [8] ARTSTEIN Z. (1989) Yet another proof of Lyapunov convexity theorem, Proceedings of the AMS [9] ARTSTEIN Z. (to appear) Parametrized Integration of Multifunctions with Applications to Control and Optimization, Preprint [10] ASEEV S.M. (1982) Approximation of semicontinuous multivalued mappings by continuous ones, Izvestiya Acad. Nauk SSSR, Vol.46. (English Translation. Math. USSR Izvestiya, [1983], Vol.20, p.435-448)
Bibliography [11] ASEEV S.M. (1983) Approximation of Semicontinuous Multivalued Mappings by Continuous Ones, Math. USSR Izvestija, 20, 435-448 [12] ASEEV S.M. (1985) Operateurs quasi-lineaires et leurs applications a la theorie des correspondances, Trudy Mat. Inst. AN SSSR, Vol.167, 25-52 [13] ASPLUND E. & ROCKAFELLAR R.T. (1969) Gradients of convex functions, Trans. Am. Math. Soc, 139, 443-467 [14] ASPLUND E. (1966) Farthest points in reflexive locally uniformly rotund Banach spaces, Isr. J. Math., 4, 213-216 [15] ASPLUND E. (1968) Frechet differentiability of convex functions, Acta Math., 121, 31-47 [16] ATTOUCH H., AUBIN J.-P., CLARKE F. H. & EKELAND I. Eds. (1989) Analyse Non Lineaire, Gauthier-Villars к C.R.M. Universite de Montreal [17] ATTOUCH H. к BREZIS H. (1986) Duality for the sum of convex functions in general Banach spaces, in Aspects of Mathematics and its applications, J. Barroso, Ed. North- Holland, 125-133 [18] ATTOUCH H. к WETS R. (1983) A convergence theory for saddle functions, Trans. A.M.S., 280, 1-41 [19] ATTOUCH H. к WETS R. (1983) Convergence des points Min-Sup et des points fixes, Comptes-Rendus de l'Academie des Sciences, PARIS, Vol.296, 657-660 [20] ATTOUCH H. к WETS R. (1987) A quantitative approach via epigraphic distance to stability of strong local minimizers, Publications AVACAM, 1 (1) - Universite de Perpignan [21] ATTOUCH H. к WETS R. (1987) Lipschitzian stability of e-approxiamate solutions in convex optimization, IIASA WP 87-25 [22] ATTOUCH H. к WETS R. (1988) Quantitative Stability of Variational Systems: I. The Epigraphical Distance, Preprint IIASA 88-8 [23] ATTOUCH H. к WETS R. (1988) Quantitative Stability of Variational Systems: II. A Framework for Nonlinear Conditioning, Preprint IIASA 88-9
Bibliography 423 [24] ATTOUCH Η. к WETS R. (1988) Another Isometry for the Legendre-Fenchel transform, J. Math. Anal. Appl.,131, 404- 411 > [25] ATTOUCH H. к WETS R. (1989) Epigraphical Analysis, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier-Villars к C.R.M. Universite de Montreal, 73- 100 [26] ATTOUCH H. (1977) Convergence de Fonctionnelles Con- vexes, In Proc. Journ, Anal. Non Lineaire, Lecture Notes in Math. # 665 Springer-Verlag, 1-40 [27] ATTOUCH H. (1977) Convergence de functions convexes, de sous-differentiels et semi-groupes associes, C.R.A.S., 284, 539- 542 [28] ATTOUCH H. (1978) Sur la Г-convergence, In Seminar on Partial Differential Equations, H. Brezis к J.-L. Lions, Eds., Pitman, [29] ATTOUCH H. (1979) Famille d'operateurs maximaux monotones et mesurabilite, Ann. Mat, Рига Appli., 120, 35-111 [30] ATTOUCH H. (1984) Variational Convergence for Functions and Operators, Research Notes in Mathematics. Pitman, London [31] AUBIN J.-R, FRANKOWSKA H. к OLECH С (1986) Controllability of convex processes, SIAM J. on Control and Optimization, 24, 1192-1211 [32] AUBIN J.-P, FRANKOWSKA H. к OLECH С (1986) Controlabilite des processus convexes, Comptes-Rendus de l'Academie des Sciences, Paris, 301, 153-156 [33] AUBIN J.-P. & CELLINA A. (1984) Differential Inclusions, Springer-Verlag, Grundlehren der math. Wiss. # 264 [34] AUBIN J.-P. & CLARKE F. H. (1977) Monotone invariant solutions to differential inclusions, J. London Math. Soc, 16, 357-366 [35] AUBIN J.-P. к EKELAND I. (1984) Applied Nonlinear Analysis, Wiley-Interscience [36] AUBIN J.-P. & FRANKOWSKA H. (1984) Trajectoires lour- des de systemes controles, Comptes-Rendus de l'Academie des Sciences, PARIS, Serie 1, 298, 521-524
424 Bibliography [37] AUBIN J.-P. & FRANKOWSKA Η. (1985) Heavy viable trajectories of controlled systems, Annales de l'lnstitut Henri- Poincare, Analyse Non Lineaire, 2, 371-395 [38] AUBIN J.-P. & FRANKOWSKA H. (1987) Observabilite des systemes sous incertitude, Comptes-Rendus de l'Academie des Sciences, PARIS, 306, 413-416 [39] AUBIN J.-P. & FRANKOWSKA H. (1987) On the Inverse Function Theorem, J. Math. Pures Appliquees, 66, 71-89 [40] AUBIN J.-P. & FRANKOWSKA H. (1989) Observability of Systems under Uncertainty, SIAM J. on Control and Optimization, Special Issue dedicated to Mac Sham, 27, [41] AUBIN J.-P., FRANKOWSKA H. & LASOTA A. (to appear) Poincare's Recurrence Theorem for Set- Valued Dynamical Systems, Bull. Polish Acad. Sciences [42] AUBIN J.-P. & WETS R. (1988) Stable approximations of set- valued maps, Ann.Inst. Henri Poincare, Analyse non lineaire, 5, 519-535 [43] AUBIN J.-P. (1978) Propriete de Perron-Frobenius pour des correspondances, C.R.A.S., 286, 911-914 [44] AUBIN J.-P. (1979) Applied Functional Analysis, Wiley- Interscience [45] AUBEN J.-P. (1979) Mathematical Methods of Game and Economic Theory, North-Holland (Studies in Mathematics and its Applications, Vol. 7, 1-619) [46] AUBIN J.-P. (1981) Contingent derivatives of set-valued maps and existence of solutions to nonlinear inclusions and differential inclusions, Advances in Mathematics, Supplementary studies, Ed. Nachbin L., 160-232 [47] AUBIN J.-P. (1982) Comportement lipschitzien des solutions de problemes de minimisation convexes, Comptes-Rendus de FAcademie des Sciences, PARIS, 295, 235-238 [48] AUBIN J.-P. (1983) L'Analyse non lineaire et ses motivations economiques, Masson [49] AUBIN J.-P. (1984) Lipchitz Behavior of Solutions to Convex Minimization Problems, Mathematics of Operations Research, 8, 87-111
Bibliography 425 [50] AUBIN J.-P. (to appear) Viability Theory [51] AUBIN J.-P. (1990) Differential Games: a Viability Approach, SIAM J. on Control and Optimization [52] AUMANN R.J. (1965) Integrals of set-valued maps, J. Math. Anal. Appl., 12, 1-12 [53] AUMANN R.J. (1967) Measurable utility and the measurable choice theorem, La Decision, Actes Coll. Int. du CNRS, AIX en Provence, 15-26 [54] AUSLENDER A. (1978) Differentiable stability in nonconvex and nondifferentiable programming, Mathematical Programming Study, 10, P. Huard (Ed.) [55] AZE D. (1988) An Inversion Theorem for Set-Valued Maps, Bull. Aust. Math. Soc, 37, 411-444 [56] BACCIOTTI A. (1988) The Local Stabilizability Problem for Nonlinear Systems, IMA Journal of Mathematical Control к Information, 5, 27-39 [57] BANACH S. (1932) Theorie des Operations Lineaires, Monographie Matematyczne, P. W. N., Warszawa [58] BANKS H.T. к JACOBS M.Q. (1970) A differential calculus for multifunctions, J. Math. Anal. Appl., 29, 246-272 [59] BARBU V. к CELLINA A. (1970) On the surjectivity of multivalued dissipative mappings, Boll. Un. Mat. Ital., 3, 817-826 [60] BEAUZAMY P. (1985) Introduction to Banach spaces and THEIR GEOMETRY, North Holland, Mathematics study, 68 [61] BEER G. (1988) On the Young-Fenchel transform for convex functions, Proc. Amer. Math. Soc, 104, 1115-1123 [62] BEN-TAL A. к ZOWE J. (1982) A Unified Theory of First and Second Order Conditions for Extremum Problems in Topological Vector Spaces, Mathematical Programming Study, 19, 39-76 [63] BEN-TAL А. к ZOWE J. (1982) Necessary and sufficient opti- mality conditions for a class of nonsmooth minimization problems, Mathematical Programming, 24, 70-91
426 Bibliography [64] BEN-TAL A. & ZOWE J. (1985) Second Order Optimality Conditions for the Ll -Minimization Problem, Appl. Math. Op- tim, 13, 45-58 [65] BERGE C. (1957) Theorie generate des jeux a n personnes, Memorial Sciences. Math. 138 [66] BERGE С (1959) Espaces Topologiques et Fonctions Multivoques, Dunod, Paris [67] BOBROW D. G. (Ed.) (1984) Qualitative Reasoning about Physical Systems, Artificial Intelligence Volume, 24 (North Holland) [68] BORISOVTCH YU. G., GELMAN B.D. & MYSHKIS A.D. (1987) Multivalued analysis and operator inclusions, Translated from Itogi Nauki i Tekhniki, Seriya Sovremennye Prqb- lemy Matematiki, Noveishie Dostizhenia, 29, 151-211 [69] BORWEIN J.M., FITZPATRICK S.P. & GILES J.R. (1987) The Differentiability of Real Functions on Normed Linear Space Using Generalized Subgradients, Journal of Mathematical Analisis and Applications, 28 (2), 512-534 [70] BORWEIN J.M & GILES J.R. (1987) The Proximal Normal Formula in Banach Space, Transactions of the Americal Mathematical Society, 302 (1), 371-381 [71] BORWEIN J.M & PREISS D. (1987) A Smooth Variational Principle with Applications to Subdifferentiability and to Differentiability of Convex Functions, Transactions of the American Mathematical Society, 303 (2), 517-527 [72] BORWEIN J.M. & STROJWAS H.M. (1986) Proximal Analysis and Boundaries of Closed Sets in Banach Space, Part I: Theory, Can. J. Math., 38, 431-452 [73] BORWEIN J.M. & ZHUANG D.M. (1989) Verifiable Necessary and Sufficient Conditions for Openness and Regularity of Set-Valued and Single-Valued Maps, J. Math. Anal. Appl. [74] BORWEIN J.M (1978) Weak tangent cones and optimization in Banach spaces, SIAM J. Cont. Opt., 16, 512-522 [75] BORWEIN J.M (1983) Adjoint Process Duality, Mathematics of Operations Research, 8, 403-434
Bibliography 427 [76] BOUCHITTE G. к VALADIER M. (1989) Multifonctions s.c.i. et regularisee s.c.i. essentielle, in Analyse Non LlNEAlRE, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier- Villars к C.R.M. Universite de Montreal, 123-150 [77] BOUCHITTE G. к VALADIER M. (1988) Integral representation of convex junctionals on a space of measures, J. Funct. An. 80, 398-420 [78] BOULIGAND G. (1930) Sur les surfaces depourvues de points hyperlimites, Ann. Soc Polon. Math., 9,32-41 [79] BOULIGAND G. (1932) Introduction a la geometrie infinitesimale directe, Gauthier-Villars [80] BOULIGAND G. (1932) Sur la semi-continuite d'inclusions et quelques sujets connexes, Enseignement Mathematique, 31, 14-22 [81] BRESSAN A, CELLINA A. к COLOMBO G. (1988) Upper semicontinuous differential inclusions without convexity, Preprint SISSA [82] BRESSAN А. к COLOMBO G. (1986) Extensions and Selections of Maps with Decomposable Values, Preprint SISSA [83] BREZIS H. (1973) Operateurs maximaux monotones ET SEMI-GROUPES DE CONTRACTION DANS LES ESPACES DE HlLBERT, North-Holland [84] BROWDER F. (1976) Nonlinear Operators and Nonlinear EQUATIONS OF EVOLUTION IN BANACH SPACES, Am. Math. Soc. Proc. Symp., 18, 2 [85] BYRNES C.I. к ISIDORI A. (to appear) Feedback Design From the Zero Dynamics Point of View, Preprint [86] BYRNES C.I. к ISIDORI A. (to appear) Output Regulation of Nonlinear Systems, Preprint [87] BYRNES C.I. к ISIDORI A. (to appear) The Analysis and Design of Nonlinear Feddback Systems. I. Zero Dynamics and Global Normal Forms, Preprint [88] BYRNES CI. к ISIDORI A. (to appear) The Analysis and Design of Nonlinear Feddback Systems. II. Global Stabilization of Minimum Phase Systems, Preprint
428 Bibliography [89] CANNARSA P.M. & FRANKOWSKA H. (to appear) Some characterizations of optimal trajectories in control theory, Preprint Scuola Normale di Pisa [90] CAPUZZO DOLCETTA I. & FALCONE M. (1989) Discrete Dynamic Programming and Viscosity Solutions of the Bellman Equation, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier-Villars & C.R.M. Universite de Montreal, 161-184 [91] CASTAING CH. & VALADIER M. (1977) Convex Analysis AND MEASURABLE MULTIFUNCTIONS, Springer-Verlag, Lecture Notes in Math. # 580 [92] CASTAING CH. (1967) Sur les multi-application mesurables, Inf. Rech. Op., 1, 91-126 [93] CASTAING CH. (1972) Integrales convexes duales, C.R.A.S., 275, 1331-1334 [94] CASTAING CH. (1988) Convergence faible et selections ex- tremales, Seminaire d'analyse convexe, Expose #2, Montpel- lier [95] CELLINA A. & COLOMBO R.M. (1989) On the representation of measurable set valued maps through selections, Preprint SISSA [96] CELLINA A. (1969) A theorem on the approximation of set- valued mappings, Rend. Ac, Na. Lincei, 47, 429-433 [97] CELLINA A. (1969) Approximation of set-valued functions and fixed point theorems, Annali Mat. Рига. Appl., 82, 17-24 [98] CESARI L. (1983) Optimization-Theory and applications, Springer-Verlag [99] CHANEYR.W. (1982) Second-order sufficiency conditions for nondifferentiable programming problems, SIAM J. Control Op- tim., 20, 20-33 [100] CHOQUET G. (1947) Convergences, Annales de Г Univ. de Grenoble, 23, 55-112 [101] CHOU C.C. (to appear) Sur I'inversion locale d'une multiap- plication, Preprint
Bibliography 429 [102] CHOW S. N. & LASOTA A. (1972) An implicit function theorem for nondifferentiable mappings, Proc. Amer. Math. Soc, 34, 141-146 [103] CLARKE F.H. &LOEWENP.D. (1989) Variational Problems with Lipschitzian Minimizers, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier-Villars & C.R.M. Universite de Montreal, 185-210 [104] CLARKE F.H. & LOEWEN P.D. (to appear) The Value Function in Optimal Control: Sensitivity, Controllability, and Time-Optimality, SIAM J. Control and Optimization, [105] CLARKE F.H. & VINTER R. (1983) Local optimality conditions and Lipschitzian solutions to the Hamilton-Jacobi equation, SIAM J. Control and Optimization, 21 (6), 856-871 [106] CLARKE F.H. (1975) Generalized gradients and applications, Trans. Am. Math. Soc, 205, 247-262 [107] CLARKE F.H. (1975) The Euler-Lagrange differential inclusion, J. Differ. Eq., 19, 80-90 [108] CLARKE F.H. (1976) A new approach to Lagrange multipliers, Math. Oper. Res., 1, 165-174 [109] CLARKE F.H. (1976) On the inverse function theorem, Рас. J. Math., 64, 97-102 [110] CLARKE F.H. (1977) Inequality constraints in the calculus of variations, Can. J. Math., 3, 528-540 [ill] CLARKE F.H. (1981) A variational proof of Aumann's theorem, Appl. Math. Optim., 7, 373-378 [112] CLARKE F.H. (1981) Generalized gradients of Lipschitz functional Adv. Math., 40, 52-67. Appeared earlier as MRC Tech. Rep., 1687, Aug. 1976 (Madison, Wisconsin) [113] CLARKE F.H. (1983) Optimization and Nonsmooth " Analysis, Wiley-Interscience [114] COLOMBO G. (1988) Non-convex problems for multifunction, PhD. Thesis, SISSA~ " [115] CORNET B. & LAROQUE G. (1986) Lipschitz properties of solutions in mathematical programming, J. Optim. Theory and Appl.
430 _ Bibliography [116] CORNET В. & VIAL J.-Ph. (1987) Lipschitzian solutions of programming problems, SIAM J. Control ала Optimization, -1137 [117] CORNET B. (1981) Regular Properties of Tangent and Normal Cones, Cahiers de Maths, de la Decision #8130 [118] CORNET B. (1984) Regular Properties of Open Tangent Cones, Center for Operations Research & Econometrics - Uni- versite Catholique de Louvain [119] CRANDALL M.G., EVANS L.C. & LIONS P.L. (1984) Some properties of viscosity solutions of Hamilton-Jacobi equation, Trans. Amer. Math. Soc, 282(2), 487-502 [120] CRANDALL M.G. & LIONS P.L. (1983) Viscosity solutions of Hamilton-Jaccbi equations, Trans. Amer. Math. Soc, 277, 1-42 [121] DA PRATO G. & LUNARDI A. (1988) Stability, Instability and Center Manifold Theorem for Fully Nonlinear Autonomous Parabolic Equations in Banach Space, Archive for Rational Mechanics and Analysis, 101, 2, 115-141 [122] DANCS S. (to appear) Tangent and generalized tangent cones in finite dimensional spaces. A unified approach, Preprint [123] DE BLASI F. & MIJAK J. (1983) On continuous approximations for multifunctions, Preprint Universita Degli Studi di Firenze [124] DE BLASI F. & MIJAK J. (to appear) A remark on Continuous Selections, Preprint [125] DE BLASI F.S. (1976) On differentiability of multif unctions, Рас. J. Math., 66, 67-81 [126] DE GIORGI E. & FRANZONI T. (1975) Sul un tipo di con- vergenza variazionale, Atti Ace. Naz. Lincei, 58, 842-850 [127] DE GIORGI E. (1975) Sulla convergenza di alcune successions di integrali del tipo del area, Rend. Mat. Univ. Roma, 8, 277- 294 [128] DE GIORGI E. (1977) Υ-convergenza e G-convergenza, Bol- letino di Unione Matematica Italiana, 14-A, 213-220
Bibliography 431 [129] DE GIORGI E. (1981) Generalized limits in calculus of variations, in Topics In4 Fucnctional Analysis 1980-1981, F. Strocchi, E. Zarantanello, De Giorgi, G. dal Maso & L. Modica Eds, Scuola Norm. Sup. Pisa, 117-148 [130] DE KLEER J. & BROWN J.S. (1984) A Qualitative Physics Based on Confluences, Artificial Intelligence, 24 [131] DE PASCALE D. & PAPINI P.L. (1984) Cones and Projections onto Subspaces in Banach Spaces, Bollettino U.M.I., 6, 411-420 [132] DEBREU G. (1959) Theory of Value, Wiley [133] DEBREU G. (1967) Integration of correspondances, 5th Berkeley Symposium on Math. Stat. Prob., II, Part I, 351-372 [134] DEIMLING K. & HU SHOU CHUAN (1988) Fixed Point of Weakly Inward Maps in Conical Shells, Nonlinear Analysis, Theory, Methods 1 Applications, 12, 3, 227-230 [135] DEIMLING K. (1985) Nonlinear Functional Analysis, Springer-Ver lag [136] DEIMLING K. (1986) Fixed Points of Weakly Inward Multi- functions, Nonlinear Analysis, Theory, Methods and Applications, 10, 11, 1261-1262 ' [137] DEIMLING K. (1988) Positive Fixed Points of Weakly Inward Maps, Nonlinear Analysis, Theory, Methods and Applications, 12 #3, 223-226 [138] DEIMLING K. (to appear) The old fixed point problem (dif- ferentiable case), Preprint [139] DELFOUR M. & ZOLESIO J.-P. (1989) Analyse des prob- lemes de forme par la derivation des minimal, in Analyse Non LiNEAlRE, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier-Villars & C.R.M. Universite de Montreal, 211-228 [140] DEMIANOV V.F. & RUBINOV A.M. (1980) On quasidiffer- entiable functional, Sov. Math. Dokl., 21, 14-17 [141] DEMIANOV V.F. (1982) Non Smooth Analysis (in Russian), Leningrad University Press [142] DEMIANOV V.F., LEMARECHAL С & ZOWE J. (1986) Approximation to a Set-valued mapping I: A proposal, J. Appl. _„Math. Opt., 14, 203-214
Bibliography [143] DIESTEL J. (1975) Geometry of Banach spaces, Lecture Notes in Mathematics, vol. 485, Springer Verlag, Berlin [144] DIEUDONNE J. (1981) History of Functional Analysis, Mathematical Studies # 49, North-Holland [145] DMITRUK A.V., MILYUTIN A.A' & OSMOLOVSKII N.P (1980) Ljusternik's theorem and the theory ofextrema, Uspekhi Mat. Nauk, 35:6, 11-46 / Russian Math Surveys, 35:6, 11-51 [146] DOLECKI S., SALINETTI G. & WETS R. (1983) Convergence of functions: equi-semicontinuity, Transactions of the American Mathematical Society, 276, 409-429 [147] DUBOVITSKIJ D. & MILJUTIN A.A. (1971) Necessary Conditions for a Weak Extremum in the General Problem of Optimal Control, Nauka (in Russian) [148] DUNFORD N. & SCHWARTZ J.T. (1957) Linear operators I, Wiley, N.Y. [149] EKELAND I & ТЕМАМ R. (1976) Convex Analysis and Variational Problems, North Holland, Amsterdam [150] EKELAND I. & LEBOURG G. (1976) Generic Frechet- differentiability and perturbed optimization problems in Banach spaces, Trans. Am. Math. Soc, 224, 193-216 [151] EKELAND I. & VALADIERM. (1971) Representation of set- valued mappings, J. Mat. Anal. Appl., 35, 621-629 152] EKELAND I. (1972) Remarques sur les problemes variationals 1, CRAS, 275, 1057-1059 153] EKELAND I. (1973) Remarques sur les problemes variationals 2, CRAS, 276, 1347-1348 154] EKELAND I. (1974) On the variational principle, J. Mathematical Analysis and Applications, 47, 324-353 155] EKELAND I. (1979) Nonconvex minimization problems, Bull. Amer. Math. Soc, 1, 443-474 156] EKELAND I. (to appear) The ε-variational principle revisited, CIME, Varena 157] EVANS L.C. к SOUGANIDIS P.E. (1989) A P.D.E. Approach to certain large deviation problems for systems of parabolic equations, in Analyse Non Lineaire, Attouch,
Bibliography 433 Aubin, Clarke, Ekeland Eds., Gauthier-Villars к C.R.M. Uni- versite de Montreal, 229-258 158] FAN KY (1964) Sur un theoreme minimax, C.R. Acad. Sci., 259, 3925-3928 159] FAN KY (1969) Extension of two fixed-point theorems of F.E. Browder, Math. Z., 112, 234-240 160] FAN KY (1972) A Minimax Inequality and Applications, Inequalities III, 103-113, Shisha Ed., Academic Press 161] FAN KY (1979) Fixed point theorems and related theorems for non-compact convex sets, In Game Theory and Related Topics, 151-156, North-Holland 162] FAN KY (1981) A Further Generalization of Shapley's Generalization of the Knaster-Kuratowski-Mazurkiewicz Theorem, Game Theory and Mathematical Economics, 275-279 163] FAN KY (1984) Some Properties of Convex Sets Related to Fixed Point Theorems, Math. Ann., 266, 519-537 164] FATTORINI H. & FRANKOWSKA H. (1988) Necessary conditions for infinite dimensional control problems, Proceedings of the 8th International Conference on Analysis and Optimization of Systems, Lecture Notes in Control and Information Sciences, Springer Verlag, 517-532 165] FENCHEL W. (1949) On conjugate convex functions, Mimeographed Lecture Notes, Princeton University Press 166] FENCHEL W. (1952) A remark on convex sets and polarity, Medd. Lunds Univ. Mat. Sem. (Supplement Band), 82-89 167] FILIPPOV A.F. (1958) On some problems of optimal control theory, Vestnik Moskowskogo Universiteta, Math, no.2, 25-32 168] FILIPPOV A.F. (1962) On certain questions in the theory of optimal control, SIAM J. Control, 1, 76-84 169] FILIPPOV A.F. (1967) Classical solutions of differential equations with multivalued right hand side, SIAM J. on Control, 5, 609-621 170] FRANKOWSKA H. & OLECH С (1982) Boundary solutions to differential inclusions, J. Diff. Eqs. 44, 156-165
434 Bibliography [171] FRANKOWSKA Η. & OLECH С. (1983) R-convexity of integral of set-valued function, in Contributions то Analysis and Geometry, John Hopkins University Press, Baltimore, Md., 117-129 [172] FRANKOWSKA H. (1983) Inclusions adjointes associees aux trajectoires minimales d'inclusions differentielles, Comptes- Rendus de FAcademie des Sciences, PARIS, Serie 1, 297, 461- 464 [173] FRANKOWSKA H. (1984) The maximum principle for a differential inclusion problem, in Analysis and Optimization OF Systems, Proceedings of the 6th International Conference on Analysis and Optimization of Systems, Lecture Notes in Control and Information Sciences, Springer Verlag, 517-532 [174] FRANKOWSKA H. (1984) The first order necessary conditions for nonsmooth variational and control problems, SIAM J. on Control and Optimization, 22, 1-12 [175] FRANKOWSKA H. (1984) Controlabilite locale et proprietes des semi-groupes de correspondances, Comptes-Rendus de FAcademie des Sciences, PARIS, Serie 1, 299, 165-168 [176] FRANKOWSKA H. (1986) Les cones contingents aux ensembles atteignables des systemes contrdles, Comptes-Rendus de FAcademie des Sciences, PARIS 303, Serie 1, 733-736) [177] FRANKOWSKA H. (1986) Theoreme d'application ouverte pour des correspondances, Comptes-Rendus de FAcademie des Sciences, PARIS, Serie 1, 302, 559-562 [178] FRANKOWSKA H. (1987) An open mapping principle for set- valued maps, J. of Math. Analysis and Appl., 127, 172-180) [179] FRANKOWSKA H. (1987) L'equation d'Hamilton-Jacobi con- tingente, Comptes-Rendus de FAcademie des Sciences, PARIS, Serie 1, 304, 295-298 [180] FRANKOWSKA H. (1987) Local controllability of control systems with feedbacks, Proceedings of 25th CDC Conference, IEEE, Athens, December 10-12 [181] FRANKOWSKA H. (1987) The maximum principle for an optimal solution to a differential inclusion with end point constraints SIAM J. on Control and Optimization, 25, 145-157
Bibliography 435 [182] FRANKOWSKA Η. (1987) Theor ernes d:'application ouverte et de fonction inverse, Comptes Rendus de l'Academie des Sciences, PARIS, Serie 1, 305, 773-776 [183] FRANKOWSKA H. (1987) Local controllability and infinitesimal generators of semigroups of set-valued maps, SIAM J. on Control, and Optimization, 25, 412-432 [184] FRANKOWSKA H. (1988) A general multiplier rule for infinite dimensional optimization problems with constraints, Proceedings of the First workshop on stabilization of flexible structures, Montpellier, December 11-15, 1987, J.P. Zolesiio Editor, Comcon Conference Inc [185] FRANKOWSKA H. (1988) Non smooth solutions to Hamilton- Jacobi Equations, Proceedings of the International Conference Bellman Continuum, Antibes, France, June 13-14, Lecture Notes in Control and Information Sciences, Springer Verlag [186] FRANKOWSKA H. (1989) Contingent cones to reachable sets of control systems, SIAM J. on Control and Optimization, 27, 170-198 [187] FRANKOWSKA H. (1989) Estimations a priori pour les inclusions differentielles operationelles, Comptes-Rendus de l'Academie des Sciences, PARIS, Serie 1, 308, 47-55 . [188] FRANKOWSKA H. (1989) Hamilton-Jacobi equation: viscosity solutions and generalized gradients, J. of Math. Analysis and Appl. 141, 21-26 [189] FRANKOWSKA H. (1989) High order inverse function theorems, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier-Villars & C.R.M. Universite de Montreal, 283-304 [190] FRANKOWSKA H. (1989) Local controllability of control systems with feeedbacks, Journal of Optimization Theory and Applications, 60, 277-296 [191] FRANKOWSKA H. (1989) Optimal trajectories associated to a solution of contingent Hamilton-J acobi equations, Applied Mathematics and Optimization, 19, 291-311 [192] FRANKOWSKA H. (1989) Set-valued analysis and some control problems, Proceedings of the International Conference 30 years of Modern Control Theory, Kingston, June 3-6, 1988 E. Roxin Editor, Marcel Dekker"
436 Bibliography [193] FRANKOWSKA Η. (1989) On the linearization of nonlinear control systems and exact reachability, Proceedings of International Conference on Control of Distributed Parameter Systems, Santiago di Compostella (July 1987), Lecture Notes in Control and Information Sciences, Springer Verlag [194] FRANKOWSKA H. (1990) Some inverse mapping theorems, Ann. Inst. Henri Poincare, Analyse Non Lineaire, [195] FRANKOWSKA H. (in preparation) Set-Valued Analysis and Control Theory (monograph), Birkhauser [196] FRANKOWSKA H. (1990) A priori estimates for operational differential inclusions, J. Diff. Eqs. [197] FRANKOWSKA H. (to appear) On controllability and observability of implicit systems, [198] FRECHET M. (1906) Sur quelques points du calcul fonction- nel, Rend. Circ. Mat. di Palermo, 22, 51- [199] FRECHET M. (1911) Sur la notion de differentielle, C.R. Acad. Sc, 152, 845-847 [200] FRECHET M. (1925) La notion de differentielle dans I'analyse generale, Ann. Ecole Norm. Sup., 42, 293-323 [201] GATEAUX G. (1919) Functions d'une infinite de variables in- dependantes, Bull. Soc. Math. France, 47, 163-177 [202] GAUVIN J. h JANIN R. (1989) Directional lipschitnan optimal solutions and directional derivatives for the optimal value in nonlinera mathematical programming, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier- Villars & C.R.M. Universite de Montreal, 305-324 [203] GAUVIN J. (1979) The generalized gradient of a marginal function in mathematical programming, Math. Opt. Res., 4, 458-463 [204] GINER E. (1986) Thesis, Universite de Pau [205] GRAVES L.M. (1935) Topics in Functional Calculus, Bull. Amer. Math. Soc.,41, 641-662 [206] GRAVES L.M. (1950) Some mapping theorem, Duke Math. J., 17, 111-114
Bibliography 437 [207] GRUNBAUM В. (1963) Measures of Symmetry for Convex Sets, Proc. Symposia Pure Maths. VII, 233-270 [208] GRUNBAUM B. (1967) Convex Polytopes, J. Wiley- Interscierices [209] GUSEINOV H. G., SUBBOTIN A. I. & USHAKOV V. N. (1985) Derivatives for multivalued mappings with applications to game theoretical problems of control, Problems of Control and Information Theory, Vol.14, 155-167 [210] GWINNER J. (1981) Bibliography on non-differentiable optimization and non-smooth analysis, J. Сотр. Appl. Math., 7, 277-285 [211] GWINNER J. (1981) On Fixed Points and Variational Inequalities - A Circular Tour, Nonlinear Analysis Theory, Methods and Applications, 5, 565-583 [212] HADDAD G. (1981) Monotone trajectories of differential inclusions with memory, Isr. J. Math., 39, 83-100 [213] HADDAD G. (1981) Monotone viable trajectories for functional differential inclusions, J. Diff. Eq., 42, 1-24 [214] HALKIN H. (1976) Interior mapping theorem with set-valued derivative, J. Analyse Math., 30, 200-207 [215] HALKIN H. (1976) Mathematical Programming without Differentiability, in Calculus of Variations and Control Theory, D.L. Russel, Ed., Academic Press, 279-288 [216] HAMILTON R. (1982) The inverse function theorem of Nash and Moser, Bull. Am. Math. Soc, 7, 1-64 [217] HESS С (1986) These Doctorat es-Sciences, Universite de Montpellier [218] HESTENES M. R. (1966) Calculus of Variations and Optimal Control Theory, Wiley [219] HIAI F. & UMEGAKI H. (1977) Integrals, Conditional Expectations, and Martingales of Multivalued Functions, Journal of Multivariate Analysis, 7, 149-182 [220] HILDEBRANDT & GRAVES L.M. (1927) Implicit Functions and their Differentials in General Analysis, Trans. Amer. Math. Soc.,29, 127-153
438 Bibliography [221] HIMMELBERG C.J., JACOBS M.Q. & VAN VLECK F.S. (1969) Measurable multifunction ■, selectors and Filippov's implicit function lemma, J. Math. An. Appl., 25(2), 276-284 [222] HIMMELBERG C.J. & VAN VLECK F.S. (1969) Some selection theorems for measurable functions, Can. J. Math. XXI (2), 394-399 [223] HIMMELBERG C.J. & VAN VLECK F.S. (1971) Selection and implicit function theorems for multifunction with Souslin graph, Bull. Acad. Pol. Sc; 19, 911-916 [224] HIMMELBERG C.J. (1975) Measurable relations, Fund. Math., 87, 53-72 [225] HIRIART-URRUTY J.-B. & PLAZANET Ph. 1989 (1989) Moreau 's Decomposition Theorem Revisited, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier- Villars & C.R.M. Universite de Montreal, 325-338 [226] HIRIART-URRUTY J.-B. (1978) Gradients generalises de functions marginales, SIAM J. Cont. Opt., 16, 301-316 [227] HIRIART-URRUTY J.-B. (1984) Miscellanies on nonsmooth analysis and optimization, Written version of a talk in Sopron [228] HIRIART-URRUTY J.-B. (1985) Refinements of necessary optimality conditions in nondifferentiable programming 1, Lecture Notes in Mathematics, Springer Verlag [229] IOFFE A.D. & TIHOMIROV V.M. (1974) Theory of extremal problems, Nauka Moskow (English Translation, 1878, North-Holland) [230] IOFFE A.D. (1977) On lower semicontinuity of integral functional 1, SIAM J. Cont. Opt., 15, 521-538 [231] IOFFE A.D. (1978) Survey of measurable selection theorems: Russian literature supplement, SIAM J. Contr. Opt., 16, 728- 723 [232] IOFFE A.D. (1981) Nonsmooth analysis: differential calculus of nondifferentiable mappings, Trans. Am. Math. Soc, 266, 1-56 [233] IOFFE A.D. (1982) Nonsmooth analysis and the theory of fans, in Convex Analysis and Optimization, Aubin J.-P. & Vin- ter R., edts, Pitman, 93-118
Bibliography 439 [234] IOFFE A.D. (1984) Necessary conditions in nonsmooth optimization, Mathematics of Operations Research, 9 [235] IOFFE A.D. (1983) Single-valued representation of set-valued mappings II; Application to differential inclusions, SIAM J. Control and Optimization, 21 (4), 641-651 [236] IOFFE A.D. (1986) On the theory of subdifferential, Mathematics for Optimization/ Mathematical studies series, North- Holland [237] IOFFE A.D. (1986) Approximate subdifferentials and applications II, Mathematika, 33, 111-128 [238] IOFFE A.D. (1987) On the local surjection property, J. of Nonlinear Analysis, 11, 565-592 [239] IOFFE A.D. (to appear) Approximate subdifferential of non- convex functions, Trans. Am. Math. Soc. [240] ISIDORI A. (1985) Nonlinear Control Systems: An Introduction, Springer-Verlag Lecture Notes in Control and Information Sciences, Vol.72 [241] JACOBS M.Q. (1968) Measurable multivalued mappings and Lusin's theorem, Trans. Am. Math. Soc, 134, 471-481 [242] JOLY J.-L. (1973) Une famille de topologies sur I'ensemble des functions convexes pour lesquelles la bipolatite est continue, J. Math. Pures et Appl., 52, 421-441 [243] KAKUTANI S. (1941) A generalization of Brouwer's fixed point theorem, Duke Math J., 8, 457-459 [244] KING A.J. & ROCKAFELLAR R.T. (to appear) Asymptotic Theory for Solutions in Stochastic Programming and Generalized M-estimation, Preprint [245] KING A.J. & ROCKAFELLAR R.T. (to appear) Sensitivity Analysis for Nonsmooth Generalized Equations, Preprint [246] KING A.J. & WETS R.J-B (1988) Epi-consistency of convex stochastic programs, IIASA [247] KING A.J. (1986) Asymptotic Behavior of Solutions 'in Stochastic Optimization, PhD. Thesis, University of Washington
440 Bibliography [248] KLEE V. & OLECH С (1967) Characterizations of a class of convex sets, Math. Scand., 20 (2), 290-296 [249] KLEIN F. (1936) Vorlesungen Uber Die Entewicklung Der Mathematik im 19. Jahrhundert, Springer-Verlag [250] KRASNOSELSKI M.A. & ZABREIKO P.P (1975) Geometrical METHODS OF NONLINEAR ANALYSIS (IN RUSSIAN), Nauka, Moscow [251] KRASNOSELSKI M.A. (1964) Positive solutions of Operator Equations, P. Noordhoff Ltd., The Netherlands [252] KPJVAN V. & VRKOC I. (to appear) Continuous and Lips- chitzean Selections from Measurable Set-Valued Maps. Application to Viability Theory, Preprint [253] KURATOWSKIK. & RYLL-NARDZEWSKI С (1965) A general theorem on selectors, Bull. Acad. Pol. Sc, 13, 397-403 [254] KURATOWSKI K. (1932) Les functions semi-continues dans I'espace des ensembles fermes, Fund. Math. 18, 148-159 [255] KURATOWSKI K. (1958) Topologie, vols. 1 and 2 4th. Ed. corrected, Panstowowe Wyd Nauk, Warszawa. (Academic Press, New York, 1966) [256] KURATOWSKI K. (1980) A Half Century of Polish Mathematics, Pergamon Press [257] KURZHANSKI A.B. & FILIPPOVA T.F. (1989) On the Set- Valued Calculus in Problems of Viability and Control for Dynamic Processes: the Evolution Equation, in Analyse Non Lineaire, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier- Villars & C.R.M. Universite de Montreal, 339-364 [258] KURZHANSKI A.B. (1977) Control and Observation under Conditions of Uncertainty, Nauka [259] LANCASTER K.J. (1965) The theory of Qualitative Linear Systems, Econometrica, 33, 395-408 [260] LANCASTER K.J. (1966) The Solution of Qualitative Comparative Static Problems, Quarterly Journal of Economics, 80, 278-295 [261] LASOTA A. & MACKEY M.C. (1985) Probabilistic Properties of Deterministic Systems, Cambridge University Press
Bibliography 441 [262] LASOTA A. & OPIAL Z. (1968) Fixed point theorems for multivalued mappings and optimal control problems, Bull. Acad. Pol. Sc, 16, 645-649 [263] LASOTA A. & PIANIGIANI G. (1977) Invariant Measures on Topological Spaces, Bollettino U.M.I, (5) 14B, 592-603 [264] LE DONNE A. & MARCHI M.V. (1980) Representation of Lipschitzean compact convex valued mappings, Rend. Ace Naz. Lincei, 68, 278-280 [265] LEACH S.B. (1961) A note on inverse function theorems, Proc. Am. Math. Soc, 12, 694-697 [266] LEBOURG G. (in preparation) Thesis, Universite de Paris- Dauphine [267] LECHICKI Α., LEVI S. (1986) Wijsman convergence in the hyperspace of a metric space, Preprint [268] LEESE S.J. (1974) Measurable selections in normed spaces, Proc. Edinburgh Math. Soc, 19, 147-150 [269] LEESE S.J. (1974) Multifunctions of Suslin type, Bull. Austr. Math. Soc. II, 395-411 [270] LIONS J.-L. (1969) Quelques Methodes de Resolution de Problemes Nonlineaires, Dunod, Paris [271] LIONS P.- L. (1982) Generalized Solutions of Hamilton-Jacobi Equations, Pitman, Boston [272] LIUSTERNIK L.A. & SOBOLEV V.I. (1965) Elementy funktsionalnogo analiza,, Second ed. Nauka Moskow [273] LIUSTERNIK L.A. (1934) Conditional extrema of functionals, Mat. Sb., 41, 390-401 [274] LIUSTERNIK L.A. (1947) The topology of the Calculus of Variations in the Large, (English trans. Vol. 16, American Mathematical Society, 1966) [275] LOJASIEWICZ S. Jr (1985) Some theorems of Scorza-Dragoni type for multifunctions with applications to the problem of existence of solutions for differential multivalued equations, Mathematical Control Theory, 625-643
442 Bibliography LOJASIEWTCZ S.Jr (to appear) Parametrizations of convex sets, J. Approx. Theory MAHANEY M. (1973) The Mathematical Career of Pierre de Fermat, 1601-1665, Princeton University Press MANGASARIAN O.L. (1969) Nonlinear Programming, McGraw-Hill, New York MANGASARIAN O.L., & FROMOVITZ S. (1967) The Fritz John necessary optimality conditions in the presence of equality and inequality constraints, J. of Math. Analysis and Appl., 17, 37-47 MARCHAUD H. (1934) Sur les champs de demi-cones et les equations differentielles du premier ordre, Bull. Sc. Math., 62, 1-38 MARCHAUD H. (1936) Sur les champs continus de demi- cones convexes et leurs integrates, Compos. Math. 1, 89-127 MARTELLI M. V & VIGNOLI A. (1974) On differentiability of multivalued maps, Boll. Un. Mat. Stat., 10, 701-712 MARTIN D.H., GARDNER J.R. & WATKINS G.G. (1981) Indicating Cones and the Intersection Principle for Tangential Approximants in Abstract Multiplier Rules, Journal of Optimization Theory and Applications, 33, 515-537 McLEOD R.M. (1965) Mean value theorems for vector-valued functions, Proc. Edinburgh Math Soc, 14 (2), 197-209 MICHAEL E. (1951) Topologies on spaces of subsets, Trans. AMS, 71, 151-182 MICHAEL E. (1956) Continuous selections I, Annals of Math., 63, 361-381 MICHAEL E. (1956) Continuous selections II, Annals of Math., 64, 562-580 MICHAEL E. (1957) Continuous selections III, Annals of Math., 65, 375-390 MICHAEL E. (1959) A theorem on semicontinuous set valued functions, Duke Math. J., 26, 647-651 MINTY G. (1967) A Theorem on Monotone Sets in Hubert Spaces, J. Math. Anal. Appl. 11, 434-439
Bibliography 443 [291] MIRICA S., STAICU V. & ANGELESCU N. (1987) Frechet Semi-Differentials of Real- Valued Functions, Preprint Univer- sita di Firenze [292] MIRICA S., STAICU V. & ANGELESCU N. (1988) Equivalent definition and basic properties of Frechet semi- differentials, SISSA [293] MIRICA S. (1980) A note on the generalized differentiability of mappings, Nonlinear Anal, 4, 567-575 [294] MORDUKHOVITCH B. (1988) Approximation Methods in Problems of Optimization and Control, Nauka [295] MOREAU J.-J. & VALADIER M. (1987) A Chain Rule Involving Vector Functions of Bounded Variation, Journal of Functional Analysis, 74 (2), 333-345 [296] MOREAU J.-J. (1966) Fonctionnelles convexes, Semi- naire du College de France [297] MOSCO U. (1969) Convergence of convex sets and of solutions to variational inequalitites, Advances in Math. 3, 510-585 [298] MOSCO U. (1971) On the continuity of the Young-Fenchel Transformation, J. Math. Anal. Appl., 35, 518-535 [299] NDOUTOUME J.L (1987) Analyse Variationnelle et Applications en Mecanique et Contrdle, Publication AVAMAC, 1, [300] NEUSTADT L.W. (1976) Optimization, Princeton Univ. Press, Princeton N.J. [301] NEVEU J. (1964) Bases mathematiques du cacul des probabilites, Masson [302] NG K. F. (to appear) An inverse function theorem and its application in optimization theory, Preprint, University of Hong Kong [303] NIKOLSKII M. S. (1983) On the Alternate Integral of Pon- triagin, Math. USSR Sbornik, 44, 125-132 [304] NIKOLSKII M. S. (1987) On the Lower Alternating Integral of Pontriagin in Linear Differential Games of Pusrsuit, Math. USSR Sbornik, 56, 33-47 [305] OLECH С (1966) Extremal Solutions of a Control System, J. Diff. Eq., 2, 74-101 _
444 Bibliography [3061 OLECH С. (1965) A note concerning set-valued measurable functions, Bull. Acad. Pol. Sc, 13, 317-321 [3071 OLECH C. (1967) Lexicographical order, range of integral and bang-bang principle, Mathematical Theory of Control, Academic Press, New York, 35-45 [3081 OLECH С (1968) Approximation of set-valued functions by continuous functions, Colloq. Math., 19, 285-293 [3091 OLECH С (1969) Existence theorems for optimal control problems involving multiple integrals, J. Diff. Eq., 6, 512-526 [3101 OLECH C. (1969) Existence theorems for optimal control problems with vector-valued cost functions, Trans. Am. Math. Soc, 136, 157-180 [3111 OLECH C. (1975) Existence of solutions of nonconvex orientor fields, Boll. Un. Mat. Ital., 11 (4), 189-197 [3121 OLECH С (1976) Existence Theory In Optimal Control, In Control Theory and Topics in Functional Analysis, International Atom. Energy. Agency, 291-328 [3131 OLECH C. (1987) Lectures on the integration of set-valued functions, Preprint SISSA [314] ORNELAS A. (1988) Parameterization of Caratheodory Multi- functions, SISSA [315] ORNELAS A. (1989) Thesis, SISSA [3161 OUTRATA J.V. & SCHINDLER Z. (to appear) On using of bundle methods in nondifferentiable optimal control problems, Preprint [317] OUTRATA J.V. (1983) On a Class of Nonsmooth Optimal Control Problems, Appl. Math. Optim., 10, 287-306 [318] PAINLEVE P. (1909) C.R.A.S. Paris,148, 1156 [3191 PAPAGEORGIOU N.S. (1989) Convergence and representation theorems for set valued random variables, Stochastic analysis and applications, 7(2), 187-210 [3201 PAZY A. (1983) Semi-groups of Linear Operators and Applications, Springer Verlag, Berlin [3211 PCHENITCHNY B.W. (1980) Convex analysis and extremal problems (in Russian), Nauka Moscow
Bibliography 445 PCHENITCHNY B.W. (1986) Implicit function theorems for multi-valued mappings, Preprint IIASA PEANO G. (1892) Sur la definition de la derivee, Mathesis,II, 2, 12-14 PEANO G. (1885) Sull'integrabilita delle equazioni differen- ziali di primo ordine, Acad. Sci. Torino, 21, 677-685 PENOT J.-P. (1981) A characterization of tangential regularity, Nonlinear Anal. T.A.M., 5, 625-643 PETCHERSKAJA N.A. (1980) On differentiability of set- valued maps (in Russian), Vestnik of Leningrad University, 115-117 PIANIGIANIG. (1975) Existence of solutions for ordinary differential equations in Banach spaces, Bull. Acad. Polon. Sci., Ser. sci. math, astron. phys., 23, 853-857 PIANIGIANI G. (1977) On the Fundamental Theory of Multivalued Differential Equations, Journal of Differential Equations, 25, 30-38 PLIS A. (1961) Remark on measurable set-valued functions, Bull. Ac Pol. Sc., 9-12, 857-859 PLIS A. (1965) Measurable orientor fields, Bull. Acad. Pol. Sc, 13, 565-569 POLOVINKIN E. S. & SMIRNOV G.V. (1986) Differentiation of Multivalued Mappings and Properties of Solutions of Differential Inclusions, Soviet Math. Dokl., 33, 662-666 POLOVINKIN E. S. (1986) Thesis PONTRIAGINL. S. (1980) Linear Differential Games of Pursuit, Mat. Sb., 112, 307-380 QUIRK J. & RUPPERT R. (1965) Qualitative Economics and the Stability of Equilibrium, Review of economic studies, 32, 311-326 RADULESCU S. & RADULESCU M. (to appear) Local inversion theorems without assuming continuous differentiability, Preprint REICH S. (1972) Fixed Point in Locally Convex Spaces, Math. Z., 125, 17-31
446 Bibliography [337] REICH S. (1978) Approximate Selections, Best Approximations, Fixed Point, and Invariant Sets, Journal of Mathematical Analysis and Applications, 62, 104-113 [338] REICH S. (1979) Fixed Point Theorems for Set-Valued Mappings, Journal of Mathematical Analysis and Applications, 69, 353-358 [339] REICH S. (1983) Convergence, Resolvent Consistency, and the Fixed Point Property for Nonexpansive Mappings, Contemporary Mathematics, 18, 167-174 [340] REICH S. (1983) The almost fixed point property for nonexpansive mappings, Proceedings of the American Mathematical society, 88 (1), 44-46 [341] RICCERI B. (1984) Analisi Matematica - On multifunctions with convex graph, Rendiconti della Classe di Scienze fisiche, matematiche e naturali, 127, 64-70 [342] RICCERI B. (1985) Analyse Fonctionnelle - Un theoreme d'existence pour les inequations variationnelles, C. R. Acad. Sc; Paris, 301, 885-888 [343] RICCERI B. (1985) Fixed Points of Lower Semicontinuous Multifunctions and Applications: Alternative and Minimax Theorems, Memorie di Matematica, 9, 331-338 [344] RICCERI B. (to appear) Existence Theorems for Nonlinear Problems, Preprint [345] ROBERT R. (1974) Convergence de Fonctionnelles Convexes, Journal of Mathematical Analysis and Applications, 45 # 3, 533-555 [346] ROBINSON S.M. (1972) Normed Convex Processes, Transactions of the American Mathematical Society, 174 [347] ROBINSON S.M. (1974) Perturbed Kuhn-Tucker points and rates of convergence for a class of nonlinear-programming algorithms, Math. Programming, 7, 1-16 [348] ROBINSON S.M. (1976) Regularity and stability for convex multivalued functions, Math. Op. Res., 1, 130-143 [349] ROBINSON S.M. (1976) Stability theory for systems of inequalities, part II: differentiable nonlinear systems, SIAM J. on Numerical Analysis,13, 497-513
Bibliography 447 [350] ROBINSON S.M. (1979) Generalized equations and their solutions, Part 1, Math. Prog. Stud., 10, 128-141 [351] ROBINSON S.M. (1980) Analysis and computation of fixed points, Academic Press, New York [352] ROBINSON S.M. (1982) Generalized equations and their solutions, Part II: Applications to nonlinear programming, Mathematical Programming Study, 19, 200-221 [353] ROBINSON S.M. (1988) An implicit-function theorem for B- Differentiable functions, IIASA WP [354] ROCKAFELLAR R.T. & WETS R. (1983) Variational Systems, an Introduction, In Multifunctions and Integrands: Stochastic Analysis, Approximation and Optimization ed. Salinetti G., Springer-Verlag, Lecture Notes in Math. # 1091, 1-54 [355] ROCKAFELLAR R.T. & WETS R. (in preparation) Set-Valued Analysis and Subdifferential Calculus (monograph) [356] ROCKAFELLAR R.T. (1967) Monotone Processes of Convex and Concave Type, Mem. of AMS # 77 [357] ROCKAFELLAR R.T. (1969) Measurable dependance of convex sets and functions on parameters, J. Math. Anal. Appl., 28, 4-25 [358] ROCKAFELLAR R.T. (1970) Convex Analysis, Princeton University Press [359] ROCKAFELLAR R.T. (1970) On the maximal monotonicity of subdifferential mappings, Pacific J. Math., 33, 209-216 [360] ROCKAFELLAR R.T. (1974) Convex algebra and duality in dynamic models of production, Mathematical Models in Economics, Los (Ed), North Holland Amsterdam [361] ROCKAFELLAR R.T. (1979) Clarke's tangent cones and the boundaries of closed sets in Rn, Nonlinear Anal. Theor. Math. Appl., 3, 145-154 [362] ROCKAFELLAR R.T. (1979) Directionally Lipschitzian functions and subdifferential calculus, Proc. London Math. Soc, 39, 331-335
448 Bibliography— [363] ROCKAFELLAR R.T. (1979) La theorie des sous- gradients ET SES APPLICATIONS A L'OPTIMISATION: FUNCTIONS CONVEXEs ET NON convexes, Collection Chaire Aisen- tadt, Presses de l'Universite de Montreal, 130 [364] ROCKAFELLAR R.T. (1980) Generalized directional derivatives and subgradients of nonconvex functions, Can. J. Math., 32, 157-180 [365] ROCKAFELLAR R.T. (1981) Proximal subgradients, marginal values, and augmented Lagrangians in nonconvex optimization, Math. Oper. Res., 6, 427-437 [366] ROCKAFELLAR R.T. (1981) The theory of subgradients AND ITS APPLICATIONS TO PROBLEMS OF OPTIMIZATION OF convex AND nonconvex functions, Helderman Verlag, W. Berlin 1981 [367] ROCKAFELLAR R.T. (1982) Directional Differentiability of the Optimal Value Function in a Nonlinear Programming Problem, Differentiability of Optimal Values, 213-226 [368] ROCKAFELLAR R.T. (1982) Lagrange multipliers and sub- derivatives of optimal value functions in nonlinear programming, Mathematical Programming Study, 17, 28-66 [369] ROCKAFELLAR R.T. (1983) Marginal Values and Second- Order Necessary Conditions for Optimality, Math. Program. Studies, 26, 245-286 [370] ROCKAFELLAR R.T. (1985) Lipschitzian Properties of Multifunction, J. Nonlinear Anal. T.M.A., 9, 867-885 [371] ROCKAFELLAR R.T. (1985) Maximal Monotone Relations and the Second Derivative of Nonsmooth Functions, Ann. Inst. H. Poincare, Analyse Non Lineaire, 2, 167-184 [372] ROCKAFELLAR R.T. (1986) Lipschitzian stability in optimization: the role of nonsmooth analysis, IIASA WP-8646 [373] ROCKAFELLAR R.T. (1987) Pseudo-differentiability of set-valued mappings and its applications in optimization, (Preprint) [374] ROCKAFELLAR R.T. (1988) First and second-order pseudo- differentiability in nonlinear programming, Trans. AMS., 307, 75-108
Bibliography 449 [375] ROCKAFELLAR R.T. (1989) Proto-Differentiability of set- valued mappings and its applications in optimization, in ANALYSE NON LlNEAlRE, Attouch, Aubin, Clarke, Ekeland Eds., Gauthier-Villars & C.R.M. Universite de Montreal, 449-482 [376] ROCKAFELLAR R.T. (to appear) Extensions of subgradient calculus with applications to optimization, [377] ROCKAFELLAR R.T. (to appear) Second-order optimal- ity conditions in nonlinear programming obtained by way of pseudo-derivatives, Math. Op. Res. [378] RUBINOV A.M. (1985) The contingent derivative of a multivalued mapping and differentiability of the maximum under connected constraints, Sibirsk. Math, Zh., 26, 147-155 [379] SALINETTI G. & WETS R (1977) On the relation between two types of convergence for convex functions, J. Mathematical Analysis and Applications, 60, 211-226 [380] SALINETTI G. h WETS R. (1979) Convergence of sequences of closed sets, Topology Proceed., 4, 149-158 [381] SALINETTI G. & WETS R. (1979) On the convergence of sequences of convex sets in finite dimensions, SIAM Review, 21, 18-33 [382] SALINETTI G. & WETS R. (1981) On the convergence of closed-valued measurable multifunctions, Trans. AMS, 266, 275-289 [383] SHEPARD G. С (1964) Approximation Problems for Convex Polyhedra, Mathematika, 11, 9-18 [384] SHEPARD G. С (1966) The Steiner Point of a Convex Poly- tope, Canad. J. Math., 18, 1294-1300 [385] SHEPARD G. С (1968) A Uniqueness Theorem for the Steiner Point of a Convex Region, J. London Math. Soc, 43, 439-444 [386] SHI SHUZHONG & CHANG KUNG-CHING (1987) A local minimax theorem without compactness, in Nonlinear and Convex Analysis Proc. in Honor of Ky Fan, Ed. B. L. Lin and S. Simons, Lecture Notes in Pure and Applied Mathematics V. 107, Dekker, 211-234 [387] SHI SHUZHONG (1980) Remarques sur le gradient generalise, C.R. Acad. Serie Paris 291, 443-446
450 Bibliography [388] SHI SHUZHONG (1981) Semi-continuites generiques de multi- applications, C.R. Acad. Sci. Paris 293, Serie A, 27-29 [389] SHI SHUZHONG (1982) Remarques sur le gradient generalise, J. Math. Pures et Appl., 61, 301-310 [390] SHI SHUZHONG (1985) Ekeland's variational principle and the mountain pass lemma, Acta Mathematica Sinica, New Se- • ries, 1:4, 348-355 [391] SHI SHUZHONG (1987) Thioremes de Choquet et Analyse поп reguliere, С R. Acad. Sc. Paris, 305, Serie I, 41-44 [392] SHI SHUZHONG (1988) Choquet theorem and nonsmooth analysis, J. Math. Pures et Appl., 67, 441-432 [393] SIEGEL J. (1977) A new proof of Caristi's fixed point theorem, Proc. Am. Math. Soc, 66, 54-56 [394] SIMONS S. (1986) An Existence Theorem for Quasicon- cave Functions with Applications, Nonlinear Analysis, Theory, Methods & Applications, 10, 1133-1152 [395] SIMONS S. (1986) Cyclical coincidences of multivalued maps, J. Math. Soc. Japan, 38, 516-525 [396] SIMONS S. (1986) On a fixed point theorem of Cellina, Lincei Rend. fis. 130, 9-10 [397] SIMONS S. (1986) Two-Function Minimax Theorem and variational inequalities for functions on compact and noncompact sets, Proceedings of Symposia in Pure Mathematics, A.M.S. 45 (Part 2), 377-392 [398] SIMONS S. (1987) The continuity ofinfsup, with applications, Arch. Math., 48, 426-437 [399] SIMONS S. (to appear) Abstract Kuhn-Tucker theorems, Preprint University of California [400] SPINGARN J.E. (1983) Partial Inverse of a Monotone Operator, Applied Mathematics and Optimization, 10, 247-265 [401] STAICU V. (1988) Generalized Derivatives for Cl- parametrized Multifunctions, Preprint Universita degli studi di Firenze [402] STEINER J. (1840) Von dem Krummungsschwerpunkte ebener Curven, J. Reine Angew. Math., 21, 33-63, 101-122
Bibliography 451 [403] STUDNIARSKIM. (1985) Mean Value Theorem for Functions Possessing First Order Convex Approximations, Applications in Optimization Theory, Analysis und ihre Anwendungen, 4, 125-132 [404] STUDNIARSKI M. (1985) Mean Value Theorems and Sufficient Optimality Conditions for Nonsmooth Functions, Journal of Math. Anal, and Appl., Ill, 313-326 [405] SUBBOTIN A.I. & SUBBOTINA N. N. (1983) Differentiability properties of the value function of a differential game with integral terminal costs, Problems of Control and Information Theory, 12, 153-166 [406] TALLOS P. (to appear) Viability problems for nonautonomous differential inclusions, SIAM J. on Control and Optimization [407] THIBAULT L. (1982) Subdifferentials of nonconvex vector- valued functions, J. Math. Anal. Appl., 86, 319-344 [408] TOLAND J.F. (1978) Duality in nonconvex optimization, J. Math. Anal. Appl., 66, 399-315 [409] TOLAND J.F. (1979) A duality principle for nonconvex optimization and the calculus of variations, Arch. Rat. Mech. Anal., 71, 41-61 [410] TREIMAN J.S. (1983) A new characterization of Clarke's tangent cone and its applications to subgradient analysis and optimization, PhD. Thesis, University of Washington [411] TREIMAN J.S. (1983) Characterization of Clarke's tangent and normal cones in finite and infinite dimensions, Nonlinear Analysis, Theory, Methods et Applications, 7, 771-783 [412] TREIMAN J.S. (1987) Too Many Convex Tangent Cones, Can. J. Math. [413] TREIMAN J.S. (to appear) Finite Dimensional Optimality Condition: B-Gradients, Journal of Optimization Theory and Applications [414] TREIMAN J.S. (to appear) Shrinking Generalized Gradients, Nonlinear Analysis, Theory, Methods et Applications [415] UHL J. J. (1969) The range of a Vector-Valued Measure, Proc. A.M.S., 23, 158-163
452 Bibliography URSESCU С. (1973) Sur une generalisation de la methode de differ entiabilite, Ac. Naz. Lincei, 8, 199-204 URSESCU C. (1975) Multifunctions with closed convex graph, Czechoslovakia Mathematics J.,25, 438-441 URYSOHN P. (1927) Verhandtl. Akad. Amsterdam, 13, 29 VAINBERG S. (1956) Variational Methods for the STUDY OF NONLINEAR OPERATORS, Nauka VALADIER M. (1971) Multiaplications mesurables a valeurs convexes compactes, J. Math. Pures et Appl., 50, 265-297 VALADIER M. (1987) Functions et operateurs sur les mesures, C.R. Ac. Sc, 304, 135-137 VALADIER M. (1988) Approximation Lipschitzienne par I'interieur d'une multifonction S.C.I., Seminaire d'Analyse Convexe, Montpellier 1987, Expose η 11 VALADIER M. (1989) Entrainement unilateral, lignes de de- scente, functions lipschitziennes поп pathologiques, Comptes- Rendus de l'Academie des Sciences, PARIS, Serie 1, 308, 241- 244 VITALE R.A. (1985) The Steiner point in infinite dimensions, Israel J. Math., 52, 245-250 VLACH M. (1983) Closures and Neighborhoods Properties Induced by Tangential Approximation, Preprint Universitat Bielefeld WAGNER D.M. (1977) Survey of measurable selection theorems, SIAM J. Contr. Opt., 15, 859-903 WALKUP D.W. & WETS R. (1967) Continuity of some convex-cone-valued mapping, Proc. Amer. Math. Soc, 18, 229- 235 WARD D.E. & BORWEIN J.M. (1987) Nonsmooth Calculus in Finite Dimensions, SIAM J. Control and Optimization, 25 (5), 1312-1340 WARD D.E. (to appear) Convex Subcones of the Contingent Cones in Nonsmooth Calculus and Optimization, Preprint WARD D.E. (to appear) Isotone Tangent Cones and Nonsmooth Optimization, Preprint
Bibliography 453 [431] WARGA J. (1972) Optimal Control of Differential and Functional Equations, Academic Press, New York [432] WARGA J. (1976) Derivate containers, inverse functions and controllability, Calculus of Variations and Control Theory, Academic Press, 13-46 [433]- WARGA J. (1978) A second-Order Condition that Strengthens Pontryagin's Maximum Principle, Journal of Differential Equations, 28, 284-307 [434] WARGA J. (1978) An implicit function theorem without differentiability, Proc. Am. Math. Soc, 69, 65-69 [435] WARGA J. (1978) Controllability and a multiplier rule for nondifferentiable optimization problems, SIAM J. Control Op- tim., 16, 803-812 [436] WARGA J. (1983) Optimization and controllability without differentiability assumptions, SIAM J. Control and Optimization, 21(6), 837-855 [437] WARGA J. (1984) Iterative Optimization with Equality Constraints, Mathematics of Operations Research, 9, 592-605 [438] WARGA J. (1984) Second order necessary conditions in optimization, SIAM J. Control and Optimization, 22, 524-528 [439] WARGA J. (1986) Higher order conditions with and without Lagrange multipliers, SIAM J. Control and Optimization, 24, 715-730 [440] WARGA- J. (to appear) Higher Order Conditions for Conical Controllability^Piepniit Northeastern University [441] WARGA J. (to appear) Homeomorphisms and local C1 approximations, Preprint [442] WAZEWSKI T. (1961) Sur la semicontinuite inferieure du "Tendeur" d'un ensemble compact variant d'une facon continue, Bull. Acad. Pol. Sc, 9, 869-872 [443] WAZEWSKI T. (1961) Sur une condition equivalente a I'equation au contingent, Bull. Acad. Pol. Sc, 9, 865-867 [444] WAZEWSKI T. (1961) Systemes de commande et equations au contingent, Bull. Acad. Pol. Sc, 9, 151-155
454 Bibliography [445] WAZEWSKI Т. (1962) Sur une generalization de la notion des solutions d'une equation au contingent, Bull. Acad. Pol. Sc, 10, 11-15 [446] WAZEWSKI T. (1963) On an optimal control problem, Proc. Conference Differential equations and their applications, 229-242 -Prague (1962) [447] WETS R. (1977) On the Convergence of Random Sets, In Convex Analysis and its Applications, Aubin Ed., Springer-Verlag, Lecture Notes in Economics and Mathematical Systems # 144, 191-206 [448] WETS R. (1980) Convergence of convex functions, variational inequalities and convex optimization problems, In Variational Inequalities and Complementarity Problems, R.Cottle, F. Gianesi & J.L. Lions Ed., Springer-Verlag, 405- 419 [449] WETS R. (1982) A formula for the level sets of epi-limits and some applications, In Mathematical Theories of Optimization Eds. Cecconi P. & Zolezzi, Springer-Verlag, Lecture Notes in Mathematics #979, 256-268 [450] WETS R. (1982) Convergence of sequences of closed functions, In Proceed. Symposium on Mahtematical Programming and Data Perturbations, Eds. Fiacco Α., 16-27 [451] WETS R. (1984) On a compactness theorem for epiconvergent sequences of functions, In Mathematical Programming, Eds. COTTLE, LEMANSON & KORTE, North-Holland, 347- 351 [452] WETS R. (1986) An elementary proof of Minkowski's theorem for system of linear inequalities, Preprint [453] WIJSMAN R.A. (1964) Convergence of sequences of convex sets, cones and functions, Bulletin of the Americam Mathematical Society, 70, 186-188 [454] WIJSMAN R.A. (1966) Convergence of sequences of convex sets, cones and functions II, Bulletin of the Americam Mathematical Society, 123, 32-45 [455] WILSON W. A. (1926) Amer. Journ. Math. 48, 165 [456] WILSON W. A. (1930) On the Phragmin-Brouwer Theorem, Bull. Amer. Math. Soc. 111-114
Bibliography 455 [457] YOUNG L.C. (1981) Mathematicians and Their Times, Mathematical Studies # 48, North-Holland [458] ZAJICEKL. (1981) On Approximate Dini Derivates and Onesided Approximate Derivatives of Arbitrary Functions, Com- mentatiories Matematicae Universitatis Carolinae, 549-560 [459] ZAJICEK' L. (1981) On the Symmetry of Dini Derivates of Arbitrary Functions, Commentationes Mathematicae Universitatis Carolinae, 195-209 [460] ZAJICEK L. (1983) Differentiability of the Distance Function and Points of Multi-Valuedness of the Metric Projection in Banach Space, Czechoslovak Mathematical Journal, 33, 292- 308 [461] ZAJICEK L. (1984) A Generalization of an Ekeland-Lebourg theorem and the Differentiability of Distance Functions, Ren- diconti del Circolo Matematico di Palermo, 3, 403-410 [462] ZARANKIEWTCZ C. (1927) Fund. Math. 9, 124 [463] ZARANTONELLO E.H. (1975) Conical Spectral Theory, Topics in Functional Analysis, 37-116 [464] ZARANTONELLO E.H. (1982) Cyclical monotonicity of maximal monotone step operators, Bol. Soc. Bras. Mat., 13, 85-91 '[465] ZAREMBA S.C. (1936) Sur les equations au paratingent, Bull. Sc. Math., 60, 139-160 [466] ZEIDLERE. (1984) Nonlinear Functional Analysis and its Applications, Springer-Verlag [467] ZORETTI L. (1905) Sur les functions analytiques uniformes qui possedent un ensemble discontinu parfait de points sin- guliers, Journ. de Math. Pures et-Appl., 1, 1-51 [468] ZORETTI L. (1909) Un theoreme de la thiorie des ensembles, Bull. Soc. Math. France, 37, 116-119
Index 5-generalized gradient 249 σ—algebra 306 σ-finite measure 306 adjacent cone 126 adjacent derivative 189 adjacent epiderivative 237 approximate selection 358 asymptotic cone 159 asymptotic epiderivative 300 atom 329 Attouch's theorem 299 Baire's theorem 44, 58 ball 17 biorthogonal subspace 63 bipolar cone 63 bipolar set 63 bipolar theorem for closed convex processes 70 bipolar theorem 64 bitranspose 67 Borel σ-algebra 306 bounded family of processes 61 bounded-valued map 36 Caratheodory map 311 Caratheodory parametrization 376 Caratheodory selection theorem 374 Castaing's theorem 319 Cellina's theorem 358 characterization of monotone maps 107 Choquet's theorem 163 circatangent cone to a sequence of sets 166 circatangent cone 127 circatangent derivative 189 circatangent epiderivative 237 Clarke normal cone 157 Clarke tangent cone 127 closed convex process 56, 314 closed convex set-valued map 34, 373 closed graph criterion 73 closed graph theorem 60 closed range theorem 64 closed set-valued map 34, 41 closed-valued map 36, 56 coerciveness 88 compact-valued map 36 complementary convex 164 complete σ- finite measure space 310 complete σ-algebra 307 complete measure space 307 composition product 37 concave function 222 cone with compact sole 114 conjugate function 256 consistent sequence of functions 96 consistent 203 constrained inverse function theorem 96, 207 constraint qualification condition 140 contingent cone 32, 121 contingent derivative 181 457
458 contingent epiderivative 225 contingent Hessian 299 contingent hypoderivative 226 contingently epidifferentiable 225 continuous partition of unity 81, 386 continuous selection 355 continuous set-valued map 40, 141 convergence of gradients 295 convergence of subdifferentials 297 convergence theorem 271 convex epiderivative 241 convex function 222 convex hull of an upper limit 26, 273 convex kernel 159 convex set-valued map 34, 56 convex-valued map 36 core 36 correspondence 33 cow 119 criterion of pseudo-Lipschitzeanity 101 crossed convergence 61 curvaceous shape 354 cut set-valued map 43 dense family of selections 310 derivable cone 127 derivable set-valued map 182 derivable set 127 differential geometer 119 differential quotient 94 Dini directional derivative 94 directed set 19 distance to a set 17 domain 34 dual norm 20 duality map 125 duality pairing 19 duality relation 258 Dubovitskij-Miljutin cone 125 eigenvalue of a convex process 114 eigenvector of a convex process 114 Ekeland's rule 91, 233 epicontinuous 285 epiderivative 225 epigraph 222 epilimit 275 episleek function 225 equilibrium map 207 equilibrium theorem 84, 276 extension of a set-valued map 36, 236 extremal element 329 extremal face 88, 343 extremal selection 330 Fenchel's inequality 256 Fermat's rule 232 fixed point theorem for inward or outward maps 87, 250 Frechet differentiable single-valued map 94, 250 Gateaux differentiable single-valued map 94 generalized gradient (Clarke) 249 generalized sequence 19 generic continuity 44 generic property 44 God (spherical) 119 gradient 94 graph of the transpose of a process 67 graph of a map 34 graphical convergence of monotone maps 270 graphical convergence 267 graphical upper/lower limit of set- valued maps 267 Graves' theorem 95 Hausdorff distance 365 higher order adjacent subset 171 higher order circatangent subset 171 higher order derivatives 216
459 higher order contingent subset 171 higher order epiderivatives 260 hypertangent cone 164 hypograph 222 hypolimit 276 image of a set-valued map 34 indicator of a set 223 inf-compact 281 injective (globally) 209 integrably bounded 326 integral of a set-valued map 327 intermediate cone 126 invariant space 72 inverse image 36 inverse map 34 inverse set-valued map theorem 204 inverse stability theorem 100 inward map 87 Kakutani's fixed point theorem 87 kernel of the derivative 187 Krummungsschwerpunkte 365 Ky Fan's inequality 80 abor value 78 jax's principle 96, 99 jeray-Schauder's theorem" 90 Am sup inf 192 imit of infima 281 imit of set-valued maps 41 imit of sets 17 inear extension of a set-valued map 346 inear process 56 /ipschitz parametrization' 379 jipschitz selection 372 jpschitz set-valued map 41 jcal subdifferential 253 jcal superdifferential 253 >cal uniqueness criterion 103 tcally injective set-valued map 209 >p-sided paratingent derivative 201 /wer Dini derivative 231 wer epicontinuous 285 lower epilimit 276 lower hypolimit 276 lower limit of set-valued maps 41 lower limit of sets 17 lower semicompact 281 lower semicontinuity criterion 49 lower semicontinuous set-valued map 39 lower semicontinuously differentiable map 188 Lyapunov's Convexity theorem (in Banach spaces) 340 Lyapunov's Convexity theorem 339 marginal function 48 mathematical striptease 93 maximal monotone map 106 maximum theorem 48, 258 Mazur's theorem 271 measurable (single-valued map) 307 measurable selection 308 measurable space 306 measurable/Lipschitz selection theorem 375 measurable/Lipschitz map 373 measurable 307 Michael's theorem 355 minimum (bare) 93 minimal map 360 minimal selection theorem 363 minimum (bare) 93 Minkowski's difference of sets 159 Minty's theorem 108 Minty's trick 108 monotone map 104 Moreau's theorem 258 "Moreau-Yosida's approximation 258 Mosco's theorem 290 multifunction 33, 258 multivalued function 33, 270 negative polar cone 63 nonatomic measure 329 nontrivial extended function 222
460 nontrivial set-valued map 34 norm of a process 57 normal cone to a convex subset 139 normal cone to a set 157 open mapping theorem 57 Oresme-Fermat's rule 219 orthogonal 63 outward map 87 paratingent cone to a sequence of sets 167 paratingent cone 162 paratingent derivative 201 paratingent epiderivative 241 Po'mcare's continuation method 89, 201 point to set map 33 pointwise convergence of set-valued maps 268 pointwise inverse function theorem 101 pointwise limit of set-valued maps 313 polar cone 25, 63 polar of a viability domain 144 polar set 28, 63 Polish space 306 positive polar cone 63, 146 positively homogeneous function 222 positively homogeneous set-valued map 56 process 56 product of set-valued maps 37 projection 125 proper map 28 proper set-valued map 43 pseudo-convex function 233 pseudo-convex set-valued map 182 pseudo-convex set 145 pseudo-Lipschitz set-valued map 41 reflexive space 20 residual 44, 163 resolvent of a monotone map 106 restriction of a single-valued map 34, 182 Robinson-Ursescu's open mapping theorem 58, 194 Rocquefort 321 Sauterne 321 selection of a set-valued map 355 separable space 20 separation theorem 64 sequentially inf-compact 281 sequentially weak upper limit 20 set-valued Graves' theorem 205 set-valued map 33 Shi Shuzhong's theorem 165 simple map 307 simplex 82 sleek set-valued map 182 sleek set 101 smooth Banach space 125 sole of a cone 114 solvability theorem 85, 130 square product of set-valued maps 37 stability assumption 96 stable family of set-valued maps 203 star-shaped set 145 Steiner point 365 strict set-valued map 34 strictly convex space 360 striptease (mathematical) 93 strong convergence of a sequence in a Lebesgue space 272 strong measurability 307 subdifferential 248 subgradient 248 support function of a set 62 support function of the transpose 75
surjective closed convex process 57 tangent cone to a convex cone 143 tangent cone to a finite intersection 149 tangent cone to direct image 153 tangent cone to inverse image 148 tangent cone to the interior of a set 140 tangent cone to a convex set 83, 138 tangent cone to a sleek subset 130 tangentially regular 127 tarte Tatin 321 transpose of a process 67 transpose of a product 68 transpose of the restriction 69 transversality condition 32, 96, 147 uniform boundedness theorem 61, 152 uniformly sleek set-valued map 205 uniformly sleek set 101 uniformly smooth 133 unit ball 17, 158 upper epicontinuous 285 upper epilimit 275 upper hemicontinuous map 74, 175 upper hypo-limit 276 upper limit of a generalized sequence 19 upper limit of set-valued maps 41 upper limit of sets 17 upper semicontinuous set-valued map 38 value of a set-valued map 34 variational system 285 viability domain 84 viscosity solution 253 weak contingent cone 252 weak measurability 307 weak-* topology 19 weakened topology 19 Xenophanes of Colophon 119 461 Yosida approximation 111 Zarankiewicz's theorem 23