100% found this document useful (1 vote)
2K viewsOptimization and Nonsmooth Analysis, Frank Clarke
Uploaded by
Ildikó MezeiCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
100% found this document useful (1 vote)
2K viewsOptimization and Nonsmooth Analysis, Frank Clarke
Uploaded by
Ildikó MezeiCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 321
Oyearirzte wert
Nonsmooth Analysis Optimization and
Nonsmooth Analysis
> eg SIAM's Classics in Applied Mathematics series consists of books that were previously
allowed to go out of print. These books are republished by SIAM as a professional
service because they continue to be important resources for mathematical scientists.
Editor-in-Chief
Robert E. O'Malley, Jr., University of Washington
Editorial Board
Richard A. Brualdi, University of Wisconsin-Madison
Herbert B. Keller, California Institute of Technology
Andrzej Z. Manitius, George Mason University
Ingram Olkin, Stanford University
Stanley Richardson, University of Edinburgh
Ferdinand Verbulst, Mathematisch Instituut, University of Utrecht
Classics in Applied Mathematics
C.C. Lin and L. A. Segel, Mathematics Applied to Deterministic Problems in the
Natural Sciences
Johan G. FE Belinfante and Bernard Kolman, A Survey of Lie Groups and Lie Algebras
with Applications and Computational Methods
James M. Ortega, Numerical Analysis: A Second Course
Anthony V. Fiacco and Garth P. McCormick, Nonlinear Programming: Sequential
Unconstrained Minimization Techniques
FE H. Clarke, Optimization and Nonsmooth Analysis
George FE Carrier and Carl E. Pearson, Ordinary Differential Equations
Leo Breiman, Probability
R. Bellman and G. M. Wing, An Introduction to Invariant Imbedding
Abraham Berman and Robert J. Plemmons, Nonnegative Matrices in the Mathemat-
ical Sciences
Olvi L. Mangasarian, Nonlinear Programming
*Carl Friedrich Gauss, Theory of the Combination of Observations Least Subject
to Errors: Part One, Part Two, Supplement. Translated by G. W. Stewart
Richard Bellman, Introduction to Matrix Analysis
U. M. Ascher, R. M. M. Mattheij, and R. D. Russell, Numerical Solution of Boundary
Value Problems for Ordinary Differential Equations
K. E. Brenan, S. L. Campbell, and L. R. Petzold, Numerical Solution of Initial-Value
Problems in Differential-Algebraic Equations
Charles L. Lawson and Richard J. Hanson, Solving Least Squares Problems
J. E. Dennis, Jr. and Robert B. Schnabel, Numerical Methods for Unconstrained
Optimization and Nonlinear Equations
Richard E. Barlow and Frank Proschan, Mathematical Theory of Reliability
Cornelius Lanczos, Linear Differential Operators
Richard Bellman, Introduction to Matrix Analysis, Second Edition
Beresford N. Parlett, The Symmetric Eigenvalue Problem
*First time in print. Classics in Applied Mathematics (continued)
Richard Haberman, Mathematical Models: Mechanical Vibrations, Population
Dynamics, and Traffic Flow
Peter W. M. John, Statistical Design and Analysis of Experiments
Tamer Bagar and Geert Jan Olsder, Dynamic Noncooperative Game Theory, Second
Edition.
Emanuel Parzen, Stochastic Processes
Petar Kokotovi¢, Hassan K. Khalil, and John O'Reilly, Singular Perturbation Methods
in Control: Analysis and Design
Jean Dickinson Gibbons, Ingram Olkin, and Milton Sobel, Selecting and Ordering
Populations: A New Statistical Methodology
James A. Murdock, Perturbations: Theory and Methods
Ivar Ekeland and Roger Témam, Convex Analysis and Variational Problems
Ivar Stakgold, Boundary Value Problems of Mathematical Physics, Volumes I and IL
J. M. Ortega and W. C. Rheinbolde, Iterative Solution of Nonlinear Equations in
Several Variables
David Kinderlehrer and Guido Stampacchia, An Introduction to Variational
Inequalities and Their Applications
F Natterer, The Mathematics of Computerized Tomography
Avinash C. Kak and Malcolm Slaney, Principles of Computerized Tomographic Imaging
R. Wong, Asymptotic Approximations of Integrals
©. Axelsson and V. A. Barker, Finite Element Solution of Boundary Value Problems:
Theory and Computation
David R. Brillinger, Time Series: Data Analysis and Theory
Joel N. Franklin, Methods of Mathematical Economics: Linear and Nonlinear
Programming, Fixed-Point Theorems
Philip Hartman, Ordinary Differential Equations, Second Edition
Michael D. Intriligator, Mathematical Optimization and Economic Theory
Philippe G. Ciarlet, The Finite Element Method for Elliptic Problems
Jane K. Cullum and Ralph A. Willoughby, Lanczos Algorithms for Large Symmetric
Eigenvalue Computations, Vol. I: Theory
M. Vidyasagar, Nonlinear Systems Analysis, Second Edition
Robert Mattheij and Jaap Molenaar, Ordinary Differential Equations in Theory and
Practice
Shanti S. Gupta and S. Panchapakesan, Multiple Decision Procedures: Theory and
Methodology of Selecting and Ranking Populations
Eugene L. Allgower and Kurt Georg, Introduction to Numerical Continuation Methods
Heinz-Otto Kreiss and Jens Lorenz, Initial-Beundary Value Problems and the Navier-
Stokes Equations
J. L. Hodges, Jr. and E. L. Lehmann, Basic Concepts of Probability and Statistics,
Second Edition
iii This page intentionally left blank Optimization and
Nonsmooth Analysis
Se
Frank H. Clarke
University of Montreal
Society for Industrial and Applied Mathematics
Philadelphia Library of Congress Cataloging-in-Publication Data
Clarke, Frank H.
Optimization and nonsmooth analysis / F. H. Clarke.
p- cm. ~ (Classics in applied mathematics ; 5)
Reprint. Originally published: New York : Wiley, 1983.
Includes bibliographical references.
ISBN 0-89871-256-4
1. Mathematical optimization. 2. Mathematical analysis.
I. Title. II. Title: Nonsmooth analysis. III. Series: Classics in
applied mathematics ; vol. 5.
QA402.5.C53 1990
519,3-de20 90-35113
CIP
10987654
Copyright © 1990 by the Society for Industrial and Applied Mathematics
All rights reserved. Printed in the United States of America. No part of this book may be
reproduced, stored, or transmitted without the written permission of the Publisher. For
information, write the Society for Industrial and Applied Mathematics, 3600 University
City Science Center, Philadelphia, Pennsylvania 19104-2688.
This SIAM edition is an unabridged, corrected republication of the work first published
by John Wiley & Sons, Inc., New York, 1983.
SLEUM is a registered trademark. This book is dedicated to my mother Preface
“.. nothing at all takes place in the universe in which some rule of max-
imum or minimum does not appear” So said Euler in the eighteenth cen-
tury. The statement may strike us as extreme, yet it is undeniable that
humankind’s endeavors at least are usually associated with a quest for an
optimum. This may serve to explain why there has been an enduring sym-
biosis between mathematical theories of optimization and the applications
of mathematics, even though the forms of the problems (the “paradigms”)
being considered evolve in time.
The origins of analytic optimization lie in the classical calculus of
variations and are intertwined with the development of the calculus. For
this reason, perhaps, optimization theory has been slow to shed the strong
smoothness (i.e., differentiability) hypotheses that were made at its incep-
tion. The attempts to weaken these smoothness requirements have often
been ad hoc in nature and motivated by the needs of a discipline other than
mathematics (e.g., engineering). In this book a general theory of non-
smooth analysis and geometry will be developed which, with its associated
techniques, is capable of successful application to the spectrum of prob-
lems encountered in optimization. This leads not only to new results but to
powerful versions of known ones. In consequence, the approach taken
here is of interest even in the context of traditional, smooth problems on
the calculus of variations, in optima] control, or in mathematical program-
ming.
This book is meant to be useful to several types of readers. An effort is
made to identify and focus upon the central issues in optimization and
results of considerable generality concerning these issues are presented.
Because of this, although its scope is not encyclopedic, the work can serve
as a reference text for those in any of the various fields that use optimiza-
tion. An effort has been made to make the results to those who are not
expert in the book's contents. Here and elsewhere the reader will find ex-
amples drawn from economics, engineering, mathematical physics, and
various branches of analysis.
The reader who wishes not only to gain access to the main results in
the book but also to follow all the proofs will require a graduate-level
knowledge of real and functional analysis. With this prerequisite, an ad-
vance course in optimization can be based upon the book. The remaining
viii Preface
type of reader we have in mind is the expert, who will discover, we believe,
interesting tools and techniques of nonsmooth analysis and optimization.
Frank H. Clarke Acknowledgments
Writing a book is a humbling experience. One learns (among other things)
how much one depends upon others. The book could not have been writ-
ten without the much appreciated support of the Killam Foundation, which
awarded me a two-year Research Fellowship. I wish also to acknowledge
the continued support of the Natural Sciences and Engineering Research
Council of Canada. I would like to thank my colleagues at UBC and my co-
workers in the field (most of them) for stimulating me over the years (wit-
tingly or otherwise), and the Canadian Mathematical Society for choosing
my book for its new series. Foremost in my personal mentions must be
Terry Rockafellar who, for one thing, bears the blame for getting me inter-
ested in optimization in the first place. I am grateful also to the following
colleagues, all of whom have also influenced the contents of the book:
Jean-Pierre Aubin, Masako Darrough, Ivar Ekeland, Jean-Baptiste Hiriart-
Urruty, Lucien Polak, Rodrigo Restrepo, Richard Vinter, Vera Zeidan. My
thanks go to Gus Gassman and especially to Philip Loewen for cou-
rageously scouring a draft of the manuscript, and to all the staff of the UBC
Mathematics Department for their help and good cheer.
This book was first published in 1983 by John Wiley & Sons, as the initial
volume in the Canadian Mathematical Society Series of Monographs and
Advanced Texts. A Russian translation edited by V. Blagadatskih appeared
in 1988 (Nauka Press, Moscow). I wish to thank Yuri Ledyaev, who pre-
pared the Russian translation, for adding significantly to my own compila-
tion of corrections. These corrections have been incorporated into SIAM's
1990 reprinting.
F.H.C. Contents
Chapter 1 Introduction and Preview
1.1 Examples in Nonsmoocth Analysis and Optimization 2
1.2 Generalized Gradients 9
13 Three Paradigms for Dynamic Optimization 13
14 The Theory of Optimal Control and the Calculus of Variations 18
Chapter 2 Generalized Gradients
2.1 Definition and Basic Properties 25
2.2 Relation to Derivatives and Subderivatives 30
23 Basic Calculus 38
24 Associated Geometric Concepts 50
2.5 The Case in Which X is Finite-Dimensional 62
2.6 Generalized Jacobians 69
2.7 Generalized Gradients of Integral Functionals 75
28 Pointwise Maxima 85
29 Extended Calculus 95
Chapter 3 Differential Inclusions
3.1 Multifunctions and Trajectories eu
3.2 A Control Problem 120
3.3 A Problem in Resource Economics 131
3.4 Endpoint Constraints and Perturbation 142
3.5 Normality and Controllability 147
3.6 Problems with Free Time 151
3.7 Sufficient Conditions: the Hamilton—Jacobi Equation 155
Chapter 4 The Calculus of Variations
41 The Generalized Problem of Bolza 165
42 Necessary Conditions 168
43 Sufficient Conditions 177
44 Finite Lagrangians 180
45 The Multiplier Rule for Inequality Constraints 187
46 Multiple Integrals 197
xi Contents
Chapter 5 Optimal Control
5.1 Controllability 200
5.2 The Maximum Principle 210
5.3. Example: Linear Regulator with Diode 213
54 Sufficient Conditions and Existence 219
5.5 The Relaxed Control Problem 222
Chapter 6 Mathematical Programming
6.1 The Lagrange Multipler Rule 227
6.2 An Alternate Mutiplier Rule 231
63 Constraint Qualifications and Sensitivity 234
64 Calmness 238
6.5 The Value Function 241
6.6 Solvability and Surjectivity 248
Chapter 7 Topics in Analysis
7.1 Inverse and Implicit Functions 252
72 Aumann'’s Theorem 256
7.3 Sets Which are Lipschitz Epigraphs 260
74 Dependence on Initial Values 262
7.5 Ekeland’s Theorem 265
76 Directional Contractions and Fixed Points 268
7.7 Hamiltonian Trajectories and Boundary-Value Problem 270
78 Closed Trajectories of Prescribed Period 270
Comments 284
References 291
Index 305 Chapter One
Introduction and Preview
In adolescence, I hated life and was continually on the verge of suicide, from which,
however, I was restrained by the desire to know more mathematics.
BERTRAND RUSSELL, The Conquest of Happiness
Just as “nonlinear” is understood in mathematics to mean “not necessarily
linear,” we intend the term “nonsmooth” to refer to certain situations in which
smoothness (differentiability) of the data is not necessarily postulated. One of
the purposes of this book is to demonstrate that much of optimization and
analysis which have evolved under traditional smoothness assumptions can be
developed in a general nonsmooth setting; another purpose is to point out the
benefits of doing so. We shall make the following points:
1. Nonsmooth phenomena in mathematics and optimization occur natu-
rally and frequently, and there is a need to be able to deal with them.
We are thus led to study differential properties of nondifferentiable
functions,
2. There is a recent body of theory (nonsmooth analysis) and associated
techniques which are well suited to this purpose.
3. The interest and the utility of the tools and methods of nonsmooth
analysis and optimization are not confined to situations in which
nonsmoothness is present.
Our complete argument in support of these contentions is the entirety of
this book. In this chapter we get under way with a nontechnical overview of
the theory and some of its applications. The final sections are devoted to
placing in context the book’s contributions to dynamic optimization (i.e., the
1 2 Introduction and Preview
calculus of variations and optimal control), the single largest topic with which
it deals.
1.1) EXAMPLES IN NONSMOOTH ANALYSIS AND
OPTIMIZATION
For purposes of exposition, it is convenient to define five categories of
examples.
1.1.1 Existing Mathematical Constructs
The first example is familiar to anyone who has had to prepare a laboratory
report for a physics or chemistry class. Suppose that a set of observed data
points (Xo, Yo), (41, Yi)s+++5(%ys Yy) in the x-y plane is given, and consider the
problem of determining the straight line in the x-y plane that best fits the data.
Assuming that the given data points do not all lie on a certain line (any lab
instructor would be suspicious if they did), the notion of “best” must be
defined, and any choice is arbitrary. For a given line y = mx + 5, the error e,
at the ith data point (x;, y,) is defined to be |mx, + b — y,|. A common
definition of best approximating line requires that the slope m and the
intercept b minimize (L¥_ge?)!'” over all m and 6 (or, equivalently, DY e?). On
the face of it, it seems at least as natural to ask instead that the total error
DN oe; be minimized. The characteristics of the resulting solution certainly
differ. In Figure 1.1, for example, the dashed line represents the “least total
error” solution (see Example 2.3.17), and the solid line represents the “least
total square error” solution. Note that the former ignores the anomalous data
point which presumably corresponds to a gross measurement error. The least
squares solution, in contrast, is greatly affected by that point. One or the other
of these solutions may be preferable; the point we wish to make is that the
function £% oe; is nondifferentiable as a function of m and b. Thus the usual
methods for minimizing differentiable functions would be inapplicable to this
function, and different methods would have to be used. Of course, the reason
that the least square definition is the common one is that it leads to the
minimization of a smooth function of m and b.
The two functions being minimized above are actually special cases of the
L? and L! norms. The differentiability (or otherwise) of norms and of other
classes of functions has been and remains a central problem in functional
analysis. One of the first results in this area is due to Banach, who char-
acterized the continuous functions x on [0,1] at which the supremum norm
[lxl[:= max |x(z)|
O0
— \Vs/R_. ifV <0, 1.1 Examples in Nonsmooth Analysis and Optimization 5
Figure 1.2. The angle of incidence between the sun’s
tays and a solar panel.
where R, and R _ are different positive constants. If x is the voltage across the
capacitor, application of Kirchhoff’s Laws yields
ax a(u-x) ifx R", h: X > R” are given functions. Here p and gq are fixed vectors in
R" and R”, respectively. The value (minimum) of this prototypical mathemati-
cal programming problem clearly depends upon p and q; let us designate it
V( p,q). The function V is of fundamental interest in the analysis of the
problem, from both the theoretical and the computational points of view. Not
only does V fail to be differentiable in general, it even fails to be finite
everywhere, For example, our main interest could be in the case p = 0, g = 0,
and V(0, 0) could be finite. Yet there may easily be values of p and q arbitrarily
near 0 for which there exist no values of x satisfying Eg. (1). In that case,
V(p, q) is assigned the value +o, in accord with the usual convention that
the infimum over the empty set is +0co. We shall analyze the differential
properties of V in Chapter 6.
1.1.4 Threshold Phenomena
There are many economic and engineering situations which undergo a change
when certain thresholds are breached. The canonical example is that of a dam
holding a reserve of water whose level fluctuates with time (due, for example, 1.1 Examples in Nonsmooth Analysis and Optimization 7
to tidal or other action). If A(z) is the reservoir’s natural water level at time ¢
and if Ay is the height of the dam, the amount of water that will have to be
coped with downstream from the dam is proportional to f(t), where
f(t):= max{h(t) — hy, 0).
Of course, f is nondifferentiable in general, even if A(-) is smooth.
Let us turn to an economic example (of which there are a great number).
Suppose that a firm is faced with a demand function q( p) for its product. This
means that the quantity demanded (in some units) is g(p) when the asking
price is set at p. Let Q be the largest quantity of the product that the firm can
produce. Then, corresponding to any price p, the quantity that the firm can
actually sell is
min{q(p), Q},
so that effectively any firm with bounded output has a nonsmooth demand
function.
One of the growth laws of the classical theory of capital growth is known as
fixed proportions. It refers to situations in which inputs to production can only
be used in certain proportions (¢.g., one worker, one shovel). If, for example, L
and K measure in some units two inputs that must be used in equal propor-
tions, then the effective quantity of either input that is available to the
production process is min{Z, K }. Most dynamic problems in capital growth or
resource theory are treated by continuous models (i.¢., neither time nor input
levels are discrete) which incorporate some production law. If the fixed-propor-
tions law is used, the resulting optimal control problem has nonsmooth
dynamics. In Section 3.3 we study and solve such a problem in the theory of
nonrenewable resources.
1.1.5 Mathematical Techniques
There exist mathematical methods which involve nonsmoothness in a funda-
mental way, and which can be very useful even in situations which seem
smooth at first glance. We shall illustrate this with three such techniques, all of
which will be used subsequently. The first is called exact penalization. To
begin, let us consider the admittedly trivial problem of minimizing the function
J(x) = x over the reals, subject to the constraint x = 0. The minimum is 0,
attained at the only point x satisfying the constraint (ie. x = 0). The tech-
nique of exact penalization transforms constrained problems to unconstrained
ones by adding to the original objective function a term that penalizes
violation of the constraint. For example, as the reader may verify, the function
x + 2|x| attains its unique unconstrained minimum at x = 0 (which is the
solution to the original constrained problem). Note that the penalty term 2|x|
is nonsmooth. This is not due simply to the choice that was made; no smooth
penalty function (i.e., a nonnegative function vanishing at 0) will ever produce 8 Introduction and Preview
a problem whose minimum occurs at 0. For example, the function x + kx?
does not attain a minimum at 0, no matter how large k is.
This simple example illustrates that exact penalization hinges upon non-
smoothness. We shall use the technique in Chapter 3 to replace a constraint
x © F(x), where x is a function on [a, b], by a penalty term {p(x, x) dt,
where p is defined by
p(x, v):= Euclidean distance from v to F(x).
This is reminiscent of the distance function cited earlier. In fact, distance
functions frequently arise in connection with exact penalization (see Proposi-
tion 2.4.3).
Our second technique is the method which classically is known as the
variational principle. It consists of obtaining some desired conclusion by
constructing a functional F and finding a point u which minimizes it. Applying
the stationarity conditions (e.g., F’(u) = 0) yields the required conclusion.
Many of the initial applications of this method took place within the calculus
of variations (hence the name). (A famous example is Dirichlet’s Principle,
which seeks to find solutions of Laplace’s equation u,, + u,, = 0 by mini-
mizing the functional f{{u2 + u?} dx dy.) Now it is clear that the method can
only become more powerful as the functionals that can be considered become
more general. In particular, the admissibility of nonsmooth functionals is
useful, as we shall see.
Suppose that in applying the method of variational principles, or through
some other means, we are led to consider a functional F which does not attain
its infimum. (This is more likely to happen in infinite than in finite dimensions.)
Further suppose that u is a point which “almost minimizes” F. Ekeland’s
Theorem, which is presented in Chapter 7, is a useful tool which says
something about this situation. It asserts, roughly speaking, that there is a
functional F which is a slight perturbation of F and which attains a minimum
at a point # near u. The perturbed functional F is of the form F(v) = F(v) +
k||v — al] for a positive constant k (where || || is the norm on the space in
question), so that the theorem leads to nonsmooth functionals.
The final example of how nonsmoothness can intervene in new mathemati-
cal techniques is provided by a dual principle of least action due to the author,
and which we shall describe here only briefly. (The principle is used in Sections
7.7 and 7.8.) Hamilton’s Principle of least action in classical mechanics
amounts to the assertion that a physical system evolves in such a way as to
minimize (or, more precisely, render stationary) the action, which is the
variational functional
Q) [((p.%) ~ HG, p)) at.
This leads to the Hamiltonian equations —p = V,H, x = v,H, which are 1.2 Generalized Gradients 9
basic to classical mechanics. The functional (2) suffers from the serious
mathematical defect of being indefinite; that is, it admits no local maxima or
minima. This has limited its role, for example, in the qualitative theory of
existence of solutions to Hamilton’s equations.
Consider now the function G which is conjugate to H in the sense of convex
analysis:
(3) G(u, v):= sup {(u, x) + (0, p) — H(x, p)},
oP
and let us define the dual action to be the functional
(4) [((p.x) + G(-p, x) dt.
The function G, and hence the functional (4), are nonsmooth in general. In
Chapter 4 we shall define and derive stationarity conditions for such function-
als. It turns out that when H is convex, these stationarity conditions for the
functional (4) almost coincide with the Hamiltonian equations. (To be precise,
if (x, p) satisfies the generalized stationarity conditions for Eg. (4), then there
exist translates of x and p by constants which satisfy the Hamiltonian
equations.)
The dual action, Eq. (4), in contrast to the classical action, Eq. (2), can be
shown to attain a minimum for certain classes of Hamiltonians H. In conse-
quence, the dual action, which is a nonsmooth and nonconvex functional, has
proven to be a valuable tool in the study of classical Hamiltonian systems,
notably in the theory of periodic solutions.
1.2 GENERALIZED GRADIENTS
The generalized gradient is a replacement for the derivative. It can be defined
for a very general class of functions (and will be in Chapter 2). Our purpose
here is to give a nontechnical summary of the main definitions for those whose
primary interest lies in the results of later chapters. We begin the discussion
with the simplest setting, that of a locally Lipschitz real-valued function
defined on R” (n-dimensional Euclidean space).
Let f: R" > R be a given function, and let x be a given point in R”. The
function f is said to be Lipschitz near x if there exist a scalar K and a positive
number ¢ such that the following holds:
U(x) - f(x) < K[x” — x’| forall x”, x’in x + eB.
(Here B signifies the open unit ball in R”, so that x + eB is the open ball of
radius ¢ about x.) This is also referred to as a Lipschitz condition of rank K. 10 Introduction and Preview
We wish to study differential properties of f. Note to begin with that f need
not be differentiable at x. Indeed, it is not hard to produce a function f which
is Lipschitz near x and which fails to admit even one-sided directional
derivatives at x.
We define instead the generalized directional derivative of f which, when
evaluated at x in the direction », is given by
q) #°(x; v):= lim sup
ALO
fy + dv) -fO)
x :
The difference quotient whose upper limit is being taken is bounded above by
K|v| (for y sufficiently near x and X sufficiently near 0) in light of the
Lipschitz condition, so that f°(x; v) is a well-defined finite quantity. One
might conceive of using other expressions in defining a generalized directional
derivative. What makes f° above so useful is that, as a function of v, f°(x; v)
is positively homogeneous and subadditive. This fact allows us to define a
nonempty set df(x), the generalized gradient of f at x, as follows:
(2) Of (x)= (§ ER": f?(x; v) > (v, £) forall vin R").
By considering the properties of f°, it is possible to show that df(x) is a
nonempty convex compact subset of R”. One has, for any v,
(3) J°(x5 0) = max((f, 0): 6 © f(x),
so that knowing f° is equivalent to knowing df(x).
When f is smooth (continuously differentiable), Af(x) reduces to the
singleton set {Vf(x)}, and when f is convex, then 0f(x) coincides with what is
called the subdifferential of convex analysis; that is, the set of vectors ¢ in R”
satisfying
f(x +u)—f(x) > (uf) foralluin R".
The computation of @f(x) from this definition may appear to be a formida-
ble task. In fact, it is something that one seeks to avoid, just as in differential
calculus, where one rarely computes derivatives from the definition. Instead,
one appeals to a body of theory that characterizes generalized gradients of
certain kinds of functions, and to rules which relate the generalized gradient of
some compound function (e.g., sum, composition) to those of its simpler
components, This calculus of generalized gradients is developed in detail in
Chapter 2.
We shall see that there are several equivalent ways of defining the gen-
eralized gradient. One alternate to the route taken above hinges upon
Rademacher’s Theorem, which asserts that a locally Lipschitz function is 12 Generalized Gradients i
differentiable almost everywhere (in the sense of Lebesgue measure). Let 2, be
the set of points in x + eB at which f fails to be differentiable, and let S be any
other set of measure zero. We shall obtain in Section 2.5 the following
characterization of the generalized gradient:
(4) af (x) = cof lim V(x): x, > x, x, € S, x, EQ).
In words; f(x) is the convex hull of all points of the form lim vf(x,), where
{x,} is any sequence which converges to x while avoiding S U Q.
Normals and Tangents
Let C be a nonempty closed subset of R". An interesting, nonsmooth, Lipschitz
function related to C is its distance function d,., defined by
d(x) = min{|x — cl: ¢ € C}.
The generalized directional derivative defined earlier can be used to develop a
notion of tangency that does not require C to be smooth or convex. The
tangent cone T(x) to C at a point x in C is defined as follows:
(5) T(x): = (» © R": d2(x; v) = 0}.
It can be shown that this condition does in fact specify a closed convex cone.
Having defined a tangent cone, the likely candidate for the normal cone is
the one obtained from T.(x) by polarity. Accordingly, we define N.(x), the
normal cone to C at x, as follows:
(6) Ne(x) = (8: (8 ©) < 0 for all v in T.(x))}.
Tt follows then that N(x) is the closed convex cone generated by dd,.(x).
It is natural to ask whether 7. and N. can be defined directly without
recourse to the distance function. We shall prove in Section 2.4 that a vector »
belongs to 7.(x) if and only if (iff) it satisfies the following property:
For every sequence x, in C converging to x and every sequence f; in
(7) (0, co) converging to 0, there is a sequence v, converging to v such
that x, + ,v, belongs to C for all i.
The direct definition of N(x) requires the notion of perpendicular. A
nonzero vector v is perpendicular to C at a point x in C (this is denoted » 1 C
at x) if v =x’ — x, where the point x’ has unique closest point x in C.
(Equivalently, » = x’ — x, where there is a closed ball centered at x’ which a
Introduction and Preview
es
Figure 1.4 The vector x’ ~ x is perpendicular
to C atx.
meets C only at x; see Figure 1.4.) It will be shown that one has
an
(8) N(x) = 60 { Mimo ‘A200, L Catx, x, > x, 8, =o}.
,
(An overbar denotes closure.) This is probably the easiest way to see how the
normal cones of Figure 1.5 are obtained; tangents follow by polarity.
Everything Is One
We now have two analytic notions for functions (f° and df) and two
geometric notions for sets (7 and N,). Each of these pairs is dual: f° and df
are obtainable one from the other via Eq. (3), while 7, and Ng are each related
to the other by polarity. It turns out moreover that these analytic and
geometric concepts can be linked to one another. This is an important factor in
their utility, just as chess pieces gain in strength when used cooperatively. The
key to the relationship is the epigraph of a function f, denoted epi f, which is
€ =
Figure 1.5 Five normal (single-arrow) and tangent (double-arrow) cones to a set. 13. Three Paradigms for Dynamic Optimization 2B
the set
(x, r) © R"X Ri f(x) L(s, v)
is a given function. When L is sufficiently smooth, any (local) solution to this
problem satisfies the Euler-Lagrange equation
Q) F{ 9b (x(1), (0) = VL (x(0), ¥(0).
Another condition that a solution x must satisfy is known as the Weierstrass
condition. This is the assertion that for each f, the following holds:
(3)
L(x(t), <(t) + w) — L(x(t), (2) > Cw, VL (x(t), €(1))) for all w.
The other main necessary condition is known as the Jacobi condition; it is
described in Section 4.3.
Classical mechanics makes much use of a function called the Hamiltonian. It
is the function H derived from L (the Lagrangian) via the Legendre transform,
as follows. Suppose that (at least locally) the equation p = V,, L(x, v) defines
v as a (smooth) function of (x, p). Then we define
(4) H(x, p) = (p, o(x, p)) — L(x, (x, p)).
It follows that if x satisfies the Euler-Lagrange equation (2), then the function
P(t):= Vp, L( x(t), x(1)) satisfies
(5) ~ p(t) = VHC x(t), p(t), &(4) = V,H( x(t), p(t),
a celebrated system of differential equations called Hamilton’s equations.
In order to facilitate future comparison, let us depart from classical varia-
tional theory by defining the pseudo-Hamiltonian Hp:
(6) Hp(x, p,v)= (p,v) — L(x, v).
The reader may verify that the Euler-Lagrange equation (2) and the Weierstrass 20 Introduction and Preview
condition (3) can be summarized in terms of Hp as follows:
—p = V.Hp(x, px), -& = Vp Hp(x, p, x)
a”
Hp(x, p, x) = max p(x, p, v).
°
Now suppose that side constraints of the type, say, g(x, x) < 0 are added to
the problem. Inspired by the technique of Lagrange multipliers, we may seek
to prove that for an appropriate multiplier function d, a solution x satisfies the
necessary conditions above for L replaced by L + (A, g) (see Section 4.5). The
proof of this multiplier rule (in its fullest generality) was the last great quest of
the classical school of the calculus of variations; it was completed in the work
of Bliss (1930) and McShane (1939). (See Hestenes (1966) for a thorough
treatment of the approach.) It has been eclipsed in prominence, however, by
the work of Pontryagin et al. (1962), who proved the maximum principle, which
we shall now describe.
Necessary Conditions for Po, Pp, and Py
Suppose that we now constrain the arcs x under consideration to satisfy
(8) X(t) = o(x(¢), u(t),
where u(1) belongs to U. Then L(x, x) can be expressed in the form F(x, x),
and the basic problem of minimizing (1), but subject now to (8), is seen to be
the optimal control problem P, described in the preceding section. If in the
definition (6) of the pseudo-Hamiltonian Hp we take account of Eg. (8) by
replacing v (which stands in for x) with the term $(x, u), we obtain Hp as a
function of (x, p, “):
(9) p(x, p,u) = (p(x, u)) — F(x, uv).
The maximum principle is simply the assertion that the classical necessary
conditions (7) continue to hold in this situation.
We have seen two approaches to giving necessary conditions for constrained
problems: the multiplier rule (which is a Lagrangian approach) and the
maximum principle (which employs the pseudo-Hamiltonian, and which is
therefore essentially Lagrangian in nature also). Clearly, neither of these is a
direct descendant of the Hamiltonian theory. What in fact is the true Hamilto-
nian H of a constrained problem?
To answer this question, we first note the extension of the Legendre
transform defined by Fenchel. Applied to L, it expresses the Hamiltonian H as
follows:
(10) H(x, p) = sup{(p, v) — L(x, v)). 14 The Theory of Optimal Control and the Calculus of Variations a
The advantage of this version of the transform as opposed to the classical
procedure (see Eq. (4)) is that H can be defined without referring to derivatives
of L. Now we have seen in the preceding section that the control problem P; is
equivalent to minimizing {L(x, x) dt, where L is defined by
(11) L(s,v) = inf{F(s,u):u € U,v = $(s, u)}.
If this Lagrangian is substituted in Eq. (10), a simple calculation gives
(12) A(x, p) = sup (CP. (x, 4) — F(x, u)}.
This is the true Hamiltonian for P.; observe that (in light of Eq. (9))
H(x, p) = sup H,(x, p, u).
ueU
Why has the true Hamiltonian been neglected? Because it is nondifferentiable
in general, in contrast to Hp, which is as smooth as the data of the problem.
This causes no difficulty, however, if the nonsmooth calculus described in
Section 1.2 can be brought to bear. We shall prove that if x solves the optimal
control problem, then an arc p exists such that one has
= ple)
(3) | (1)
which reduces to the classical Hamiltonian system (5) when H is smooth, We
shall confirm in discussing existence and sufficiency that H is heir to the rich
Hamiltonian theory of the calculus of variations.
The Hamiltonian H for the more general problem Py is simply the function
defined by Eq. (10); it shall figure prominently in Chapter 4. As noted earlier,
the differential inclusion problem Pp is the case of P, in which
L(x = (° if v € F(x)
, +0co otherwise.
© dH(x(1), p(t),
When this Lagrangian is substituted in Eq. (10), we obtain
H(x, p) = sup{(p,v): v € F(x)},
a function which lies at the heart of things in Chapter 3. Thus, the three main
problems P., Pp, and Py all admit necessary conditions featuring (13); one
need only interpret the Hamiltonian appropriately. (We are omitting in this
section other components of the necessary conditions that deal with boundary
terms.) The approach just outlined above is different than, for example, that of
the maximum principle; the resulting necessary conditions are not equivalent. 22 Introduction and Preview
The problem treated in Section 5.3 illustrates that (13) can in some cases
provide more precise information than the maximum principle; the opposite is
also possible.
Existence and Sufficient Conditions
The classical existence theorem of Tonelli in the calculus of variations pos-
tulates two main conditions under which the basic problem of minimizing (1)
admits a solution. These are: first, the convexity of the function v > L(s, v)
for each s, and second, certain growth conditions on L, a simple example being
(14) L(s,v) 2a + Blv}? forallsandy (B>0).
When L is convex in v, L and its Hamiltonian H (see Eq. (10)) are mutually
conjugate functions, each obtainable from the other by the Fenchel transform.
It follows that growth conditions on L can be expressed in terms of H. For
example, (14) is equivalent to
2
H(s, P)< Be @ for alls and p.
In Section 4.1 we state an extension in Hamiltonian terms due to R. T.
Rockafellar (1975) of the classical existence theory to the generalized problem
of Bolza Py. Since P, is a special case of Py, this result applies to the optimal
control problem as well (see Section 5.4).
Let us now turn to the topic of sufficient conditions. There are four main
techniques in the calculus of variations to prove that a given arc actually does
solve the basic problem. We examine these briefly with an eye to how they
extend to constrained problems.
1. Elimination. This refers to the situation in which we know that a
solution exists, and in which we apply the necessary conditions to eliminate all
other candidates. This method is common to all optimization problems. Note
the absolute necessity of an appropriate existence theorem, and of knowing
that the necessary conditions apply.
2. Convexity. Every optimization problem has its version of the general
tule that for “convex problems,” the necessary conditions are also sufficient. It
is easy to prove that if L(s, v) is convex (jointly in (s, v)), then any (admissi-
ble) solution x of the Euler-Lagrange equation (2) is a solution of the basic
problem. (As simple and useful as it is, this fact is not known in the classical
theory.) As for P,, we may use the fact that convexity of L is equivalent to the
Hamiltonian H(s, p) being concave in s (it is always convex in p) in order to
state sufficient conditions in Hamiltonian terms (see Section 4.3). This ap-
proach can be extended to derive sufficient conditions for the optimal control
problem (see Section 5.4). 1.4 The Theory of Optimal Control and the Calculus of Variations B
3. Conjugacy and Fields. There is a beautiful chapter in the calculus of
variations that ties together certain families of solutions to the Euler-Lagrange
equation (fields), the zeros of a certain second-order differential equation
(conjugate points), and solutions of a certain partial differential equation (the
Hamilton-Jacobi equation), and uses these ingredients to derive sufficient
conditions. There has been only partial success in extending this theory to
constrained problems (e.g., P-; see the method of geodesic coverings (Young,
1969), which is based on fields). Recently Zeidan (1982) has used the Hamilto-
nian approach and the classical technique of canonical transformations to
develop a conjugacy approach to sufficient conditions for P, (see Section 4.3).
4. Hamilton-Jacobi Methods. The classical Hamilton-Jacobi equation is
the following partial differential equation for a function W(t, x):
W(t, x) + H(x,W,(t, x)) = 0.
It has long been known in classical settings that this equation is closely linked
to optimality. A discretization of the equation leads to the useful numerical
technique known as dynamic programming. Our main interest in it stems from
its role as a verification technique. Working in the context of the differential
inclusion problem P,, we shall use generalized gradients and the true Hamilto-
nian to define a generalized Hamilton—Jacobi equation (see Section 3.7). To a
surprising extent, it turns out that the existence of a certain solution to this
extended Hamilton-Jacobi equation is both a necessary and a sufficient
condition for optimality. Chapter Two
Generalized Gradients
What is now proved was once only imagined.
WILLIAM BLAKE, The Marriage of Heaven and Hell
1 was losing my distrust of generalizations.
THORNTON WILDER, Theophilus North
In this chapter we gather a basic toolkit that will be used throughout the rest
of the book. The theory and the calculus of generalized gradients are developed
in detail, beginning with the case of a real-valued “locally Lipschitz” function
defined on a Banach space. We also develop an associated geometric theory of
normal and tangent cones, and explore the relationship between all of these
concepts and their counterparts in smooth and in convex analysis, Later in the
chapter special attention is paid to the case in which the underlying space is
finite-dimensional, and also to the generalized gradients of certain important
kinds of functionals. Examples are given, and extensions of the theory to
non-Lipschitz and vector-valued functions are also carried out.
At the risk of losing some readers for one of our favorite chapters, we will
concede that the reader who wishes only to have access to the statements of
certain results of later chapters may find the introduction to generalized
gradients given in Chapter 1 adequate for this purpose. Those who wish to
follow all the details of the proofs (in this and later chapters) will require a
background in real and functional analysis. In particular, certain standard
constructs and results of the theory of Banach spaces are freely invoked. (Both
the prerequisites and the difficulty of many proofs are greatly reduced if it is
assumed that the Banach space is finite-dimensional.)
24 2.1 Definition and Basic Properties 5
2.1 DEFINITION AND BASIC PROPERTIES
We shall be working in a Banach space X whose elements x we shall refer to as
vectors or points, whose norm we shall denote ||x||, and whose open unit ball is
denoted B; the closed unit ball is denoted B.
The Lipschitz) Condition
Let Y be a subset of X. A function f: Y > R is said to satisfy a Lipschitz
condition (on Y) provided that, for some nonnegative scalar K, one has
Q) f(y) — £0) < Kily = yl
for all points y, y’ in Y; this is also referred to as a Lipschitz condition of rank
K. We shall say that f is Lipschitz (of rank K) near x if, for some e > 0, f
satisfies a Lipschitz condition (of rank K) on the set x + eB (i.¢., within an
e-neighborhood of x).
For functions of a real variable, a Lipschitz condition is a requirement that
the graph of f not be “too steep.” It is easy to see that a function having this
property near a point need not be differentiable there, nor need it admit
directional derivatives in the classical sense.
The Generalized Directional Derivative
Let f be Lipschitz near a given point x, and let v be any other vector in X. The
generalized directional derivative of f at x in the direction v, denoted {°(x; v),
is defined as follows:
f(x; 0) = lim sup Hy +e) fly) ) =f) ,
“t40
where of course y is a vector in X and f¢ is a positive scalar. Note that this
definition does not presuppose the existence of any limit (since it involves an
upper limit only), that it involves only the behavior of f near x, and that it
differs from traditional definitions of directional derivatives in that the base
point (y) of the difference quotient varies. The utility of f° stems from the
following basic properties.
2.1.1 Proposition
Let f be Lipschitz of rank K near x. Then
(a) The function v — {°(x; v) is finite, positively homogeneous, and subaddi-
tive on X, and satisfies
IP (x5 v)| < Kile). 6 Generalized Gradients
(b) f°(x; v) is upper semicontinuous as a function of (x, v) and, as a function of
v alone, is Lipschitz of rank K on X.
(©) fP0u —e) = (-f)°Os 0).
Proof. In view of the Lipschitz condition, the absolute value of the
difference quotient in the definition of f°(x; v) is bounded by K||v|| when y is
sufficiently near x and ¢ sufficiently near 0. It follows that |f°(x; v)| admits
the same upper bound. The fact that f°(x; Av) = Af°(x; v) for any A > 0 is
immediate, so let us now turn to the subadditivity. With all the upper limits
below understood to be taken as y > x and ¢ | 0, we calculate:
f(y + w + w) - f(y)
t
< lim supLi2 te + 0) SU # 0) inn sup fOr + ) =f)
t
f°(x; v + w) = lim sup
(since the upper limit of a sum is bounded above by the sum of the upper
limits). The first upper limit in this last expression is f°(x; v), since the term
y + tw represents in essence just a dummy variable converging to x. We
conclude
fo(xs 0+ w) < f(x; 0) + f°(x5 w),
which establishes (a).
Now let {x;} and {v,} be arbitrary sequences converging to x and »v,
respectively. For each i, by definition of upper limit, ‘there exist y, in X and
1, > 0 such that
1
In allt<
£°(x;3 8) -+ < £Oy + 4.0) ~ FO) “a = fy)
~ fOr + 40) £01), (Oi, + He) = Uy + 42)
1 t ,
i i
Note that the last term is bounded in magnitude by K ||v, — vj (in view of the
Lipschitz condition). Upon taking upper limits (as i > 00), we derive
lim sup f°(x,; »,) < f°(x; ©),
imo
which establishes the upper semicontinuity.
Finally, let any v and w in X be given. We have
f(y + 0) - f(y) < f(y + tw) - f(y) + Kile — wilt 2.1 Definition and Basic Properties 27
for y near x, ¢ near 0. Dividing by ¢ and taking upper limits as y > x, ¢ 10,
gives
FC; v) < fox w) + Kilo — wif.
Since this also holds with v and w switched, (b) follows. To prove (c), we
calculate:
f(x! = ) - f(x)
f(x; —v):= lim sup 4+ +
x! ox t
110
= lim sup
uox
tLo
where u:= x’ — tv
(=f)(u + ) = (su)
7 :
=(-/)°(x;v), asstated. O
The Generalized Gradient
The Hahn-Banach Theorem asserts that any positively homogeneous and
subadditive functional on X majorizes some linear functional on X. Under the
conditions of Proposition 2.1.1, therefore, there is at least one linear functional
$: X — R such that, for all v in X, one has f°(x; v) > §(v). It follows also that
§ is bounded, and hence belongs to the dual space X* of continuous linear
functionals on X, for which we adopt the convention of using (§, ») or (v, §)
for §(v). We are led to the following definition: the generalized gradient of f at
x, denoted df(x), is the subset of X* given by
(§ @X*: f?(x; v) > fv) for all vin X}.
We denote by ||f||4 the norm in X*:
[Sle = sup((f, 0): 0 © X,|Ivl| < 1},
and B,, denotes the open unit ball in X*. The following summarizes some basic
properties of the generalized gradient.
2.1.2 Proposition
Let f be Lipschitz of rank K near x. Then
(a) f(x) is a nonempty, convex, weak*-compact subset of X* and \|§ lle < K
for every § in df(x).
(b) For every v in X, one has
F°(x5 v) = max( (%, 0) = f°(x; v), a contradiction which establishes
(b). a
2.1.3 Example—The Absolute-Value Function
As in the case of the classical calculus, one of our goals is to have to resort
rarely to the definition in order to calculate generalized gradients in practice.
For illustrative purposes, however, let us nonetheless so calculate the gener-
alized gradient of the absolute-value function on the reals. In this case, X = R
and f(x) = |x| (f is Lipschitz by the triangle inequality). If x is strictly
positive, we calculate
f°(x; v) = lim yttwry_ ;
yox ot
110
so that 0f(x), the set of numbers § satisfying v > {v for all v, reduces to the
singleton {1}. Similarly, 0f(x) = {— 1} if x < 0. The remaining case is x = 0.
We find
. _ v ifv>0
£e0;v)={ _® ifv <0;
that is, {°(0; v) = [v|. Thus 9/(0) consists of those § satisfying |v| > ¢v for all
v; that is, 0f(0) = [—1, 1].
Support Functions
As Proposition 2.1.2 makes clear, it is equivalent to know the set 0f(x) or the
function f°(x; +); each is obtainable from the other. This is an instance of a
general fact: closed convex sets are characterized by their support functions.
Recall that the support function of a nonempty subset C of X is the function o¢:
X* + RU {+0} defined by
oc(§):= sup{(f, x): x © C}.
If = is a subset of X*, its support function is defined on X**. If we view X as a
subset of X**, then for x € X, one has
oy(x) = sup{(f, x): ©).
The following facts are known (see Hérmander, 1954). 2.1 Definition and Basic Properties 2»
2.1.4 Proposition
Let C, D be nonempty closed convex subsets of X, and let 2, A be nonempty
weak*-closed convex subsets of X*. Then
(a) CC Diff oc($) < op(b) forall § < x*.
(b) 2 CA iff oy(x) < a(x) forallx € X.
(c) 2 is weak*-compact iff 03(-) is finite-valued on X.
(d) A given function 0: X > R U {+ 00} is positively homogeneous, subaddi-
tive, lower semicontinuous (strong or weak) and not identically +00 iff
there is a nonempty weak*-closed convex subset = of X* such that 6 = oy.
Any such & is unique.
A multifunction T: X > Y is a mapping from X to the subsets of Y. Its graph is
the set
Grl:= ((x, y)i x © X, y ST(x)).
T is closed if Gr I is closed in X X Y (relative to a given topology). When X
and Y are Banach spaces, we define upper semicontinuity of T at x to be the
following property: for all e > 0, there exists 6 > 0 such that
T(x’) C T(x) + eBy forall x’ € x + SBy.
We continue to assume that f is Lipschitz near x. The first assertion below
teiterates that f°(x; +) is the support function of df(x).
21.5 Proposition
(a) £ € Of(x) iff fox; v) > (Sv) for all v in X.
(b) Let x, and §, be sequences in X and X* such that §, © Of(x,). Suppose that
x; converges to x, and that § is a cluster point of §; in the weak* topology.
Then one has § € Of(x). (That is, the multifunction Of is weak*-closed.)
(c)
af(x)= AM U_ af(y).
8>0 yextsB
(d) If X is finite-dimensional, then Of is upper semicontinuous at x.
Proof. Assertion (a) is evident from Propositions 2.1.2 and 2.1.4. In order
to prove (b), let any v in X be given. There is a subsequence of the numerical
sequence ({,,v) which converges to ¢({,v) (we do not relabel), One has
°(x; v) 2 ¢§, 0) by property (a), which implies f°(x; v) > ¢{,0) by the
upper semicontinuity of f° (see Proposition 2.1.1(b)). Since v is arbitrary, £
belongs to 0 f(x) by assertion (a).
Part (c) of the proposition is an immediate consequence of (b), so we turn
now to (d). If @f fails to be upper semicontinuous at x, then we can construct 2 30 Generalized Gradients
sequence x, converging to x and a sequence {, converging to ¢ such that ¢,
belongs to df(x,) for each i, yet § does not belong to 0f(x). This contradicts
assertion (b). O
22 RELATION TO DERIVATIVES AND SUBDERIVA TIVES
We pause to recall some terminology before pursuing the study of generalized
gradients, The main results of this section will be that af reduces to the
derivative if fis C', and to the subdifferential of convex analysis if f is convex.
Classical Derivatives
Let F map X to another Banach space Y. The usual (one-sided) directional
derivative of F at x in the direction v is
F’(x; »):= lim Flat to) = F(x) + to) = F(x)
110 t
when this limit exists. F is said to admit a G@reaux derivative at x, an element
in the space £(X, Y) of continuous linear functionals from X to Y denoted
DF(x), provided that for every v in X, F’(x; v) exists and equals (DF(x), v).
Let us note that this is equivalent to saying that the difference quotient
converges for each v, that one has
lim F(x + & = F(x) _ (DF(x), 6),
«10
and that the convergence is uniform with respect to v in finite sets (the last is
automatically true). If the word “finite” in the preceding sentence is replaced
by “compact,” the derivative is known as Hadamard; for “bounded” we
obtain the Fréchet derivative. In general, these are progressively more demand-
ing requirements. When X = R”, Hadamard and Fréchet differentiability are
equivalent; when F is Lipschitz near x (i.e., when for some constant K one has
F(x’) — F(xly < K||x’ — xy for all x’, x” near x) then Hadamard and
Gateaux differentiabilities coincide.
Strict Differentiability
It turns out that the differential concept most naturally linked to the theory of
this chapter is that of strict differentiability (Bourbaki: “strictement dérivable”).
We shall say that F admits a strict derivative at x, an element of £(X, Y)
denoted D, F(x), provided that for each v, the following holds:
lim P(x! + to) — F(x’) = (D,F(x), v),
xx t
elo 2.2 Relation to Derivatives and Subderivatives 31
and provided the convergence is uniform for v in compact sets. (This last
condition is automatic if F is Lipschitz near x). Note that ours is a “Hada-
mard-type strict derivative.”
2.2.1 Proposition
Let F map a neighborhood of x to Y, and let § be an element of &( X,Y). The
following are equivalent:
(a) F is strictly differentiable at x and D,F(x) = §.
(bo) F is Lipschitz near x, and for each v in X one has
lim F(x! + tw) — F(x’) ©
xox t
tlo
(f, 0).
Proof. Assume (a). The equality in (b) holds by assumption, so to prove (b)
we need only show that F is Lipschitz near x. If this is not the case, there exist
sequences {x,} and {x;} converging to x such that x,, x/ lie in x + (1/i)B and
WFC) — Fx Sly > ey — rll.
Let us define ¢, > 0 and », via x/ = x, + #,0, and |}v,|| = 17/7. It follows that
1,70.
Let V consist of the points in the sequence {v,} together with 0. Note that V
is compact, so that by definition of D, F(x) for any e > 0 there exists 2, such
that, for all i > n,, for all » in V, one has
<é
Y
: ~ (DF(x),0)
'
| F(x, + t;v) — F(x,)
But this is impossible since when v = v,, the term[F(x, + t;v) — F(x;)]/t; has
norm exceeding i'/? by construction. Thus (b) holds.
We now posit (b). Let V be any compact subset of X and e any positive
number. In view of (b), there exists for each v in V a number 8(v) > 0 such
that
()
; - (0)
0) converging to x, v, 0, respectively, one has
lim sup 8 F(x, + t,0,) — FOu)
142 [Olly 4
— (DF(x), v,)} = 0.
By the mean-value theorem, there is x* between x, and x, + f,v, such that
the last expression equals
(@DF(x?), 0.) ~ (@DF(x), 2.) = (@[ DF(x*) - DF(x)], 0).
This goes to 0 (uniforinly in 0) as i > 00 by the continuity of DF. O
We are now ready to explore the relationship between the various deriva-
tives defined above and the generalized gradient.
2.2.2 Proposition
Let f be Lipschitz near x and admit a Gdteaux (or Hadamard, or strict, or
Fréchet) derivative Df(x). Then Df(x) € df(x).
Proof. By definition, f’(x; v) exists for each v and equals (Df(x), v).
Clearly one has f’ < f° from the definition of the latter, so one has f°(x; v) > 2.2. Relation to Derivatives and Subderivatives 33
(Df(x), v) for all v in X. The required conclusion now follows from Proposi-
tion 2.1.5(a). O
2.2.3 Example
That df(x) can contain points other than Df(x) is illustrated by the familiar
example on R: f(x):= x?sin(1/x). This function is Lipschitz near 0, and it is
easy to show that f°(0; v) = jv]. It follows that @f(0) = [—1, 1], a set which
contains the (nonstrict) derivative Df(0) = 0.
2.2.4 Proposition
If f is strictly differentiable at x, then f is Lipschitz near x and 0f(x) = (D,f(x))}.
Conversely, if f is Lipschitz near x and 0f(x) reduces to a singleton {&), then f is
strictly differentiable at x and D, f(x) = §.
Proof. Suppose first that D,f(x) exists (so that f is Lipschitz near x by
Proposition 2.2.1). Then, by definition of f°, one has f°(x; v) = (D,f(x), v)
for all v, and it follows from Proposition 2.1.5(a) that df(x) reduces to
{D,f(x)}. To prove the converse, it suffices to show that the condition of
Proposition 2.2.1(b) holds for each v in X. We begin by showing that
f°(x; 0) = (fv) for each v. (Note that f°(x; v0) > ¢f,v) by Proposition
2.1.2(b).) By the Hahn—Banach Theorem there exists {’ € X* majorized by
°(x; -) and agreeing with f°(x; -) at v. It follows that {’ € df(x), and we
have f°(x; v) = ({’, v) > ¢f, v). If (f, v) were less than f°(x; v), then ¢, ¢”
would be distinct elements of df(x), contrary to hypothesis, Thus f°(x; v) =
<§ v) for all v.
We now calculate:
tim ing LF) “S09 iy gyp LSU" +)
xox t ox t
110 110
, ~ wb) — Fx’
= —1im sup f(x' + w — Ww) —f(x' + wv)
xX t
to
a
—f°(x; —v) = 8-0) = 0)
= f°(x; v) = lim sup Lat + te) ~ f(x) :
xix t
110
This establishes the limit condition of Proposition 2.2.1(b) and completes the
proof. O
Corollary
If f is Lipschitz near x and X is finite-dimensional, then 0f(x') reduces to a
singleton for every x’ in x + eB iff f is continuously differentiable on x + eB. 34 Generalized Gradients
Proof. Invoke the corollary to Proposition 2.2.1 together with Proposition
2.1.5(d); note that for a point-valued map, continuity and upper semicontinu-
ity coincide. O
2.2.5. Example (Indefinite Integrals)
Let ¢: [0,1] + R belong to L™{0,1], and define a (Lipschitz) function f:
[0,1] > R via f(x) = fe'o(2) dt. Let us calculate 0f(x). We know that f is
differentiable for almost all x, with f’(x) = ¢(x); for any such x, we have
(x) © @f(x) by Proposition 2.2.2. It follows from this and from upper
semicontinuity (see Proposition 2.1.5) that all essential cluster points of $ at x
(ie., those that persist upon the removal of any set of measure zero) belong to
af(x).
Let $*(x) and ¢ (x) denote the essential supremum and essential infimum
of $ at x. The remarks above and the fact that @f(x) is convex imply that
Of(x) contains the interval [¢"(x), ¢*(x)]. From the equality
f(y +t) ~10) = P96) ds
one easily deduces that f°(x; 1) is bounded above by $*(x). Any ¢ in df(x)
satisfies f°(x; 1) > 1§ (by Proposition 2.1.5(a)), so we derive § < $*(x). Simi-
larly, ¢ > @7(x). We arrive at the conclusion
Af(x) = [o (x), o* (x).
Convex Functions
Let U be an open convex subset of X. Recall that a function f: U > R is said
to be convex provided that, for all u, uw’ in U and d in [0, 1], one has
f(Aut+ (L-A)u’) | so that y = px is in U.
If A = 1 /p, then the set
V={v:0 =(1-A)x’ +Ay, x’ © eB} 2.2 Relation to Derivatives and Subderivatives 35
is a neighborhood of x = Ay with radius (1 — A)e. For all » in V, one has by
convexity
f(v) < (1A) f(x’) +Af(y) < M+ ALY),
so that f is bounded above on a neighborhood of x. If z is any point in
x +(1—A)eB, there is another such point z’ such that x = (z+ z’)/2,
whence
F(x) < af (2) + af(2’).
It follows that
(2) > 2f(x) ~ f(z") > 2f(x) —M-AL(y),
so that now f is also seen to be bounded below near x; we have established that
f is bounded near x.
Let N be a bound on |/| on the set x + 25B, where 6 > 0. For distinct
X\,X_ in x + 6B, set x; = x, + (6/a)(x, — x,) where & =||x, — x,|, and
note that x, is in x + 26B. Solving for x, gives
__ 6 48
ee at Ot
and so by convexity
Hx) < 2g ila) + SG lloa)
Then
fx) ~ fle) < 2G L165) — fed] < FU) - fd
which combined with |f| < N and a = ||x. — x,|| yields
2N
f (x2) -f (m1) < lee — xl.
Since the roles of x, and x, may be interchanged, we conclude that / is
Lipschitz near x. O
The proof showed:
Corollary
Let f be convex with |f|
