Springer Monographs in Mathematics

Editorial Board
S. Axler
K.A. Ribet

For further volumes:

www.springer.com/series/3733
Asen L. Dontchev • R. Tyrrell Rockafellar

Implicit Functions
and Solution Mappings
A View from Variational Analysis

With 12 Illustrations

ABC
Asen L. Dontchev R. Tyrrell Rockafellar
Mathematical Reviews University of Washington
416 Fourth Street Department of Mathematics
Ann Arbor, MI 48107-8604 PO Box 354350
USA Seattle, WA 98195-4350
[email protected] USA
[email protected]

ISSN 1439-7382 e-ISSN

ISBN 978-0-387-87820-1 e-ISBN 978-0-387-87821-8
DOI 10.1007/978-0-387-87821-8
Springer Dordrecht Heidelberg London New York

Library of Congress Control Number: 2009926485

Mathematics Subject Classification (2000): 26B10, 47J07, 58C15, 49J53, 49K40, 90C31, 93C70

Springer
c Science+Business Media, LLC 2009
All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

Setting up equations and solving them has long been so important that, in popular
imagination, it has virtually come to describe what mathematical analysis and its
applications are all about. A central issue in the subject is whether the solution to
an equation involving parameters may be viewed as a function of those parameters,
and if so, what properties that function might have. This is addressed by the classical
theory of implicit functions, which began with single real variables and progressed
through multiple variables to equations in infinite dimensions, such as equations
associated with integral and differential operators.
A major aim of the book is to lay out that celebrated theory in a broader way
than usual, bringing to light many of its lesser known variants, for instance where
standard assumptions of differentiability are relaxed. However, another major aim
is to explain how the same constellation of ideas, when articulated in a suitably
expanded framework, can deal successfully with many other problems than just
solving equations.
These days, forms of modeling have evolved beyond equations, in terms, for ex-
ample, of problems of minimizing or maximizing functions subject to constraints
which may include systems of inequalities. The question comes up of whether the
solution to such a problem may be expressed as a function of the problem’s pa-
rameters, but differentiability no longer reigns. A function implicitly obtainable
this manner may only have one-sided derivatives of some sort, or merely exhibit
Lipschitz continuity or something weaker. Mathematical models resting on equa-
tions are replaced by “variational inequality” models, which are further subsumed
by “generalized equation” models.
The key concept for working at this level of generality, but with advantages even
in the context of equations, is that of the set-valued solution mapping which as-
signs to each instance of the parameter element in the model all the corresponding
solutions, if any. The central question is whether a solution mapping can be local-
ized graphically in order to achieve single-valuedness and in that sense produce a
function, the desired implicit function.
In modern variational analysis, set-valued mappings are an accepted workhorse
in problem formulation and analysis, and many tools have been developed for
v
vi Preface

handling them. There are helpful extensions of continuity, differentiability, and re-
gularity of several types, together with powerful results about how they can be ap-
plied. A corresponding further aim of this book is to bring such ideas to wider
attention by demonstrating their aptness for the fundamental topic at hand.
In line with classical themes, we concentrate primarily on local properties of so-
lution mappings that can be captured metrically, rather than on results derived from
topological considerations or involving exotic spaces. In particular, we only briefly
discuss the Nash–Moser inverse function theorem. We keep to finite dimensions in
Chapters 1 to 4, but in Chapters 5 and 6 provide bridges to infinite dimensions.
Global implicit function theorems, including the classical Hadamard theorem, are
not discussed in the book.
In Chapter 1 we consider the implicit function paradigm in the classical case of
the solution mapping associated with a parameterized equation. We give two proofs
of the classical inverse function theorem and then derive two equivalent forms of it:
the implicit function theorem and the correction function theorem. Then we grad-
ually relax the differentiability assumption in various ways and even completely
exit from it, relying instead on the Lipschitz continuity. We also discuss situations
in which an implicit function fails to exist as a graphical localization of the so-
lution mapping, but there nevertheless exists a function with desirable properties
serving locally as a selection of the set-valued solution mapping. This chapter does
not demand of the reader more than calculus and some linear algebra, and it could
therefore be used by both teachers and students in analysis courses.
Motivated by optimization problems and models of competitive equilibrium,
Chapter 2 moves into wider territory. The questions are essentially the same as in
the first chapter, namely, when a solution mapping can be localized to a function
with some continuity properties. But it is no longer an equation that is being solved.
Instead it is a condition called a generalized equation which captures a more com-
plicated dependence and covers, as a special case, variational inequality conditions
formulated in terms of the set-valued normal cone mapping associated with a con-
vex set. Although our prime focus here is variational models, the presentation is
self-contained and again could be handled by students and others without special
background. It provides an introduction to a subject of great applicability which is
hardly known to the mathematical community familiar with classical implicit func-
tions, perhaps because of inadequate accessibility.
In Chapter 3 we depart from insisting on localizations that yield implicit func-
tions and approach solution mappings from the angle of a “varying set.” We identify
continuity properties which support the paradigm of the implicit function theorem in
a set-valued sense. This chapter may be read independently from the first two. Chap-
ter 4 continues to view solution mappings from this angle but investigates substitutes
for classical differentiability. By utilizing concepts of generalized derivatives, we are
able to get implicit mapping theorems that reach far beyond the classical scope.
Chapter 5 takes a different direction. It presents extensions of the Banach open
mapping theorem which are shown to fit infinite-dimensionally into the paradigm of
the theory developed finite-dimensionally in Chapter 3. Some background in basic
functional analysis is required. Chapter 6 goes further down that road and illustrates
Preface vii

how some of the implicit function/mapping theorems from earlier in the book can
be used in the study of problems in numerical analysis.
This book is targeted at a broad audience of researchers, teachers and graduate
students, along with practitioners in mathematical sciences, engineering, economics
and beyond. In summary, it concerns one of the chief topics in all of analysis, his-
torically and now, an aid not only in theoretical developments but also in methods
for solving specific problems. It crosses through several disciplines such as real and
functional analysis, variational analysis, optimization, and numerical analysis, and
can be used in part as a graduate text as well as a reference. It starts with elemen-
tary results and with each chapter, step by step, opens wider horizons by increas-
ing the complexity of the problems and concepts that generate implicit function
phenomena.
Many exercises are included, most of them supplied with detailed guides. These
exercises complement and enrich the main results. The facts they encompass are
sometimes invoked in the subsequent sections.
Each chapter ends with a short commentary which indicates sources in the liter-
ature for the results presented (but is not a survey of all the related literature). The
commentaries to some of the chapters additionally provide historical overviews of
past developments.

Whidbey Island, Washington Asen L. Dontchev

August, 2008 R. Tyrrell Rockafellar
Acknowledgements

Special thanks are owed to our readers Marius Durea, Shu Lu, Yoshiyuki Sekiguchi
and Hristo Sendov, who gave us valuable feedback on the entire manuscript,
and to Francisco J. Aragón Artacho, who besides reviewing most of the book
helped us masterfully with all the figures. During various stages of the writing we
also benefited from discussions with Aris Daniilidis, Darinka Dentcheva, Hélène
Frankowska, Michel Geoffroy, Alexander Ioffe, Stephen Robinson, Vladimir Veliov,
and Constantin Zălinescu. We are also grateful to Mary Anglin for her help with the
final copy-editing of the book.
The authors

ix
Contents

Preface v

Acknowledgements ix

Chapter 1. Functions defined implicitly by equations 1

1A. The classical inverse function theorem 9
1B. The classical implicit function theorem 17
1C. Calmness 21
1D. Lipschitz continuity 26
1E. Lipschitz invertibility from approximations 35
1F. Selections of multi-valued inverses 47
1G. Selections from nonstrict differentiability 51

Chapter 2. Implicit function theorems for variational problems 61

2A. Generalized equations and variational problems 62
2B. Implicit function theorems for generalized equations 74
2C. Ample parameterization and parametric robustness 83
2D. Semidifferentiable functions 88
2E. Variational inequalities with polyhedral convexity 95
2F. Variational inequalities with monotonicity 106
2G. Consequences for optimization 112

Chapter 3. Regularity properties of set-valued solution mappings 131

3A. Set convergence 134
3B. Continuity of set-valued mappings 142
3C. Lipschitz continuity of set-valued mappings 148
3D. Outer Lipschitz continuity 154

xi
xii Contents

3E. Aubin property, metric regularity and linear openness 159

3F. Implicit mapping theorems with metric regularity 169
3G. Strong metric regularity 178
3H. Calmness and metric subregularity 182
3I. Strong metric subregularity 186

Chapter 4. Regularity properties through generalized derivatives 197

4A. Graphical differentiation 198
4B. Derivative criteria for the Aubin property 205
4C. Characterization of strong metric subregularity 217
4D. Applications to parameterized constraint systems 221
4E. Isolated calmness for variational inequalities 224
4F. Single-valued localizations for variational inequalities 228
4G. Special nonsmooth inverse function theorems 237
4H. Results utilizing coderivatives 245

Chapter 5. Regularity in infinite dimensions 251

5A. Openness and positively homogeneous mappings 253
5B. Mappings with closed and convex graphs 259
5C. Sublinear mappings 265
5D. The theorems of Lyusternik and Graves 274
5E. Metric regularity in metric spaces 280
5F. Strong metric regularity and implicit function theorems 292
5G. The Bartle–Graves theorem and extensions 297

Chapter 6. Applications in numerical variational analysis 311

6A. Radius theorems and conditioning 312
6B. Constraints and feasibility 320
6C. Iterative processes for generalized equations 326
6D. An implicit function theorem for Newton’s iteration 336
6E. Galerkin’s method for quadratic minimization 348
6F. Approximations in optimal control 352

References 363

Notation 371

Index 373
Chapter 1
Functions Defined Implicitly by Equations

The idea of solving an equation f (p, x) = 0 for x as a function of p, say x = s(p),

plays a huge role in classical analysis and its applications. The function obtained in
this way is said to be defined implicitly by the equation. The closely related idea of
solving an equation f (x) = y for x as a function of y concerns the inversion of f .
The circumstances in which an implicit function or an inverse function exists and
has properties like differentiability have long been studied. Still, there are features
which are not widely appreciated and variants which are essential to seeing how the
subject might be extended beyond solving only equations. For one thing, properties
other than differentiability, such as Lipschitz continuity, can come in. But funda-
mental expansions in concept, away from thinking just about functions, can serve in
interesting ways as well.
As a starter, consider for real variables x and y the extent to which the equation
x2 = y can be solved for x as a function of y. This concerns the inversion of the
function f (x) = x2 in Figure 1.1 below, as depicted through the reflection that inter-
changes the x and y axes. The reflection of the graph is not the graph of a function,
but some parts of it may have that character. For instance, a function is obtained
from a neighborhood of the point B, but not from one of the point A, no matter how
small.

y x

B
A y

A x
B

Fig. 1.1 Graphical localizations of the function y = x2 and its inverse.

A.L. Dontchev and R.T. Rockafellar, Implicit Functions and Solution Mappings: A View 1
from Variational Analysis, Springer Monographs in Mathematics,
DOI 10.1007/978-0-387-87821-8 1,  c Springer Science+Business Media, LLC 2009
2 1 Functions Defined Implicitly by Equations

Although the reflected graph in this figure is not, as a whole, the graph of a
function, it can be regarded as the graph of something more general, a “set-valued
mapping” in terminology which will be formalized shortly. The question revolves
then around the extent to which a “graphical localization” of a set-valued mapping
might be a function, and if so, what properties that function would possess. In the
case at hand, the reflected graph assigns two different x’s to y when y > 0, but no x
when y < 0, and just x = 0 when y = 0.
To formalize that framework for the general purposes of this chapter, we focus
on set-valued mappings F from IRn and IRm , signaled by the notation

F : IRn →
→ IR ,
m

by which we mean correspondences which assign to each x ∈ IRn one or more ele-
ments of IRm , or possibly none. The set of elements y ∈ IRm assigned by F to x is
denoted by F(x). However, instead of regarding F as going from IRn to a space of
subsets of IRm we identify as the graph of F the set
  
gph F = (x, y) ∈ IRn × IRm  y ∈ F(x) .

Every subset of IRn × IRm serves as gph F for a uniquely determined F : IRn → → IRm ,
so this concept is very broad indeed, but it opens up many possibilities.
When F assigns more than one element to x we say it is multi-valued at x, and
when it assigns no element at all, it is empty-valued at x. When it assigns exactly
one element y to x, it is single-valued at x, in which case we allow ourselves to write
F(x) = y instead of F(x) = {y} and thereby build a bridge to handling functions as
special cases of set-valued mappings.
Domains and ranges get flexible treatment in this way. For F : IRn → → IRm the
domain is the set   
dom F = x  F(x) = 0/ ,
while the range is   
rge F = y  y ∈ F(x) for some x ,
so that dom F and rge F are the projections of gph F on IRn and IRm respectively.
Any subset of gph F can freely be regarded then as itself the graph of a set-valued
submapping which likewise projects to some domain in IRn and range in IRm .
The functions from IRn to IRm are identified in this context with the set-valued
mappings F : IRn → → IRm such that F is single-valued at every point of dom F. When
F is a function, we can emphasize this by writing F : IRn → IRm , but the notation
F : IRn →
→ IRm doesn’t preclude F from actually being a function. Usually, though, we
use lower case letters for functions: f : IRn → IRm . Note that in this notation f can still
be empty-valued in places; it’s single-valued only on the subset dom f of IRn . Note
also that, although we employ “mapping” in a sense allowing for potential multi-
valuedness (as in a “set-valued mapping”), no multi-valuedness is ever involved
when we speak of a “function.”
A clear advantage of the framework of set-valued mappings over that of only
functions is that every set-valued mapping F : IRn → → IRm has an inverse, namely the
1 Functions Defined Implicitly by Equations 3

set-valued mapping F −1 : IRm →

→ IRn defined by
  
F −1 (y) = x  y ∈ F(x) .

The graph of F −1 is generated from the graph of F simply by reversing (x, y) to

(y, x), which in the case of m = n = 1 corresponds to the reflection in Figure 1.1.
In this manner a function f always has an inverse f −1 as a set-valued mapping.
The question of an inverse function comes down then to passing to some piece of
the graph of f −1 . For that, the notion of “localization” must come into play, as
we are about to explain after a bit more background. Traditionally, a function f :
IRn → IRm is surjective when rge f = IRm and injective when dom f = IRn and f −1
is a function; full invertibility of f corresponds to the juxtaposition of these two
properties.
In working with IRn we will, for now, keep to the Euclidean norm |x| associated
with the canonical inner product

x, x  = ∑ j=1 x j xj for x = (x1 , . . . , xn ) and x = (x1 , . . . , xn ),

n

namely
  1/2
∑ j=1 x2j
n
|x| = x, x = .

The closed ball around x̄ with radius r is then

  
IBr (x̄) = x  |x − x̄| ≤ r .

We denote the closed unit ball IB1 (0) by IB. A neighborhood of x̄ is any set U such
that IBr (x̄) ⊂ U for some r > 0. We recall for future needs that the interior of a
set C ⊂ IRn consists of all points x such that C is a neighborhood of x, whereas
the closure of C consists of all points x such that the complement of C is not a
neighborhood of x; C is open if it coincides with its interior and closed if it coincides
with its closure. The interior and closure of C will be denoted by int C and cl C.

Graphical localization. For F : IRn → → IRm and a pair (x̄, ȳ) ∈ gph F , a graphical
localization of F at x̄ for ȳ is a set-valued mapping F̃ such that

gph F̃ = (U × V ) ∩ gph F for some neighborhoods U of x̄ and V of ȳ,

so that 
F(x) ∩V when x ∈ U,
F̃ : x →
0/ otherwise.
The inverse of F̃ then has

−1 F −1 (y) ∩U when y ∈ V,
F̃ (y) =
0/ otherwise,

and is thus a graphical localization of the set-valued mapping F −1 at ȳ for x̄.

4 1 Functions Defined Implicitly by Equations

Often the neighborhoods U and V can conveniently be taken to be closed balls

IBa (x̄) and IBb (ȳ). Observe, however, that the domain of a graphical localization F̃
of F with respect to U and V may differ from U ∩ dom F and may well depend on
the choice of V .

Single-valuedness in localizations. By a single-valued localization of F at x̄ for ȳ

will be meant a graphical localization that is a function, its domain not necessarily
being a neighborhood of x̄. The case where the domain is indeed a neighborhood of
x̄ will be indicated by referring to a single-valued localization of F around x̄ for ȳ
instead of just at x̄ for ȳ.
For the function f (x) = x2 from IR to IR we started with, the set-valued inverse
mapping f −1 , which is single-valued only at 0 with f −1 (0) = 0, fails to have a
single-valued localization at 0 for 0. But as observed in Figure 1.1, it has a single-
valued localization around ȳ = 1 for x̄ = −1.
In passing from inverse functions to implicit functions more generally, we need
to pass from an equation f (x) = y to one of the form

(1) f (p, x) = 0 for a function f : IRd × IRn → IRm

in which p acts as a parameter. The question is no longer that of inverting f , but the
framework of set-valuedness is valuable nonetheless because it allows us to imme-
diately introduce the solution mapping
  
(2) S : IRd →
→ IR with S(p) = x  f (p, x) = 0 .
n

We can then look at pairs ( p̄, x̄) in gph S and ask whether S has a single-valued
localization s around p̄ for x̄. Such a localization is exactly what constitutes an
implicit function coming out of the equation. The classical implicit function theorem
deduces the existence from certain assumptions on f . A review of the form of this
theorem will help in setting the stage for later developments because of the pattern
it provides. Again, some basic background needs to be recalled, and this is also an
opportunity to fix some additional notation and terminology for subsequent use.
A function f : IRn → IR is upper semicontinuous at a point x̄ when x̄ ∈ int dom f
and for every ε > 0 there exists δ > 0 for which

f (x) − f (x̄) < ε whenever x ∈ dom f with |x − x̄| < δ .

If instead we have

−ε < f (x) − f (x̄) whenever x ∈ dom f with |x − x̄| < δ ,

then f is said to be lower semicontinuous at x̄. Such upper and lower semicontinuity
combine to continuity, meaning the existence for every ε > 0 of a δ > 0 for which

| f (x) − f (x̄)| < ε whenever x ∈ dom f with |x − x̄| < δ .

1 Functions Defined Implicitly by Equations 5

This condition, in our norm notation, carries over to defining the continuity of a
vector-valued function f : IRn → IRm at a point x̄ ∈ int dom f . However, we also
speak more generally then of f being continuous at x̄ relative to a set D when
x̄ ∈ D ⊂ dom f and this last estimate holds for x ∈ D; in that case x̄ need not belong
to int dom f . When f is continuous relative to D at every point of D, we say it is
continuous on D. The graph gph f of a function f : IRn → IRm with closed domain
dom f that is continuous on D = dom f is a closed set in IRn × IRm .
A function f : IRn → IRm is Lipschitz continuous relative to a set D, or on a set D,
if D ⊂ dom f and there is a constant κ ≥ 0 such that

| f (x ) − f (x)| ≤ κ |x − x| for all x , x ∈ D.

If f is Lipschitz continuous relative to a neighborhood of a point x̄ ∈ int dom f , f is

said to be Lipschitz continuous around x̄. A function f : IRd × IRn → IRm is Lipschitz
continuous with respect to x uniformly in p near ( p̄, x̄) ∈ int dom f if there is a
constant κ ≥ 0 along with neighborhoods U of x̄ and Q of p̄ such that

| f (p, x ) − f (p, x)| ≤ κ |x − x| for all x , x ∈ U and p ∈ Q.

Differentiability entails consideration of linear mappings. Although we generally

allow for multi-valuedness and even empty-valuedness when speaking of “map-
pings,” single-valuedness everywhere is required of a linear mapping, for which we
typically use a letter like A. A linear mapping from IRn to IRm is thus a function
A : IRn → IRm with dom A = IRn which obeys the usual rule for linearity:

A(α x + β y) = α Ax + β Ay for all x, y ∈ IRn and all scalars α , β ∈ IR.

The kernel of A is   
ker A = x  Ax = 0 .
In the finite-dimensional setting, we carefully distinguish between a linear mapping
and its matrix, but often use the same notation for both. A linear mapping A : IRn →
IRm is represented then by a matrix A with m rows, n columns, and components ai, j :
⎛ ⎞
a11 a12 ··· a1n
⎜ a21 a22 ··· a2n ⎟
A=⎜
⎝ ... .. ⎟ .
. ⎠
am1 am2 · · · amn

The inverse A−1 of a linear mapping A : IRn → IRm always exists in the set-valued
sense, but it isn’t a linear mapping unless it is actually a function with all of IRm as its
domain, in which case A is said to be invertible. From linear algebra, of course, that
requires m = n and corresponds to the matrix A being nonsingular. More generally,
if m ≤ n and the rows of the matrix A are linearly independent, then the rank of
the matrix A is m and the mapping A is surjective. In terms of the transpose of A,
6 1 Functions Defined Implicitly by Equations

denoted by AT , the matrix AAT is in this case nonsingular. On the other hand, if
m ≥ n and the columns of A are linearly independent then AT A is nonsingular.
Both the identity mapping and its matrix will be denoted by I, regardless of
dimensionality. By default, |A| is the operator norm of A induced by the Euclidean
norm,
|A| = max |Ax|.
|x|≤1

A function f : →
IRn is differentiable at a point x̄ when x̄ ∈ int dom f and
IRm
there is a linear mapping A : IRn → IRm with the property that for every ε > 0 there
exists δ > 0 with

| f (x̄ + h) − f (x̄) − Ah| ≤ ε |h| for every h ∈ IRn with |h| < δ .

If such a mapping A exists at all, it is unique; it is denoted by D f (x̄) and called the
derivative of f at x̄. A function f : IRn → IRm is said to be twice differentiable at a
point x̄ ∈ int dom f when there is a bilinear mapping N : IRn × IRn → IRm with the
property that for every ε > 0 there exists δ > 0 with

| f (x̄ + h) − f (x̄) − D f (x̄)h − N(h, h)| ≤ ε |h|2 for every h ∈ IRn with |h| < δ .

If such a mapping N exists it is unique and is called the second derivative of f at x̄,
denoted by D2 f (x̄). Higher-order derivatives can be defined accordingly.
The m × n matrix that represents the derivative D f (x̄) is called the Jacobian of f
at x̄ and is denoted by ∇ f (x̄). In the notation x = (x1 , . . . , xn ) and f = ( f1 , . . . , fm ),
the components of ∇ f (x̄) are the partial derivatives of the component functions fi :
 m,n
∂ fi
∇ f (x̄) = (x̄) .
∂xj i, j=1

In distinguishing between D f (x̄) as a linear mapping and ∇ f (x̄) as its matrix, we

can guard better against ambiguities which may arise in some situations. When the
Jacobian ∇ f (x) exists and is continuous (with respect to the matrix norms associ-
ated with the Euclidean norm) on a set D ⊂ IRn , then we say that the function f is
continuously differentiable on D; we also call such a function smooth or C 1 on D.
Accordingly, we define k times continuously differentiable (C k ) functions.
For a function f : IRd × IRn → IRm and a pair ( p̄, x̄) ∈ int dom f , the partial deriva-
tive mapping Dx f ( p̄, x̄) of f with respect to x at ( p̄, x̄) is the derivative of the function
g(x) = f ( p̄, x) at x̄. If the partial derivative mapping is continuous as a function of
the pair (p, x) in a neighborhood of ( p̄, x̄), then f is said to be continuously differen-
tiable with respect to x around ( p̄, x̄). The partial derivative Dx f ( p̄, x̄) is represented
by an m × n matrix, denoted ∇x f ( p̄, x̄) and called the partial Jacobian. Respectively,
D p f ( p̄, x̄) is represented by the m × d partial Jacobian ∇ p f ( p̄, x̄). It’s a standard fact
from calculus that if f is differentiable with respect to both p and x around ( p̄, x̄)
and the partial Jacobians ∇x f (p, x) and ∇ p f (p, x) depend continuously on p and x,
then f is continuously differentiable around ( p̄, x̄).
1 Functions Defined Implicitly by Equations 7

Fig. 1.2 The front page of Dini’s manuscript from 1877/78.

With this notation and terminology in hand, let us return to the setting of im-
plicit functions in equation (1), as traditionally addressed with tools of differen-
tiability. Most calculus books present a result going back to Dini1 , who formulated
and proved it in lecture notes of 1877/78; the cover of Dini’s manuscript is displayed
above. The version typically seen in advanced texts is what we will refer to as the
classical implicit function theorem or Dini’s theorem. In those texts the set-valued
solution mapping S in (2) never enters the picture directly, but a brief statement in
that mode will help to show where we are headed in this book.

1 Ulisse Dini (1845–1918). Many thanks to Danielle Ritelli from the University of Bologna for a
copy of Dini’s manuscript.
8 1 Functions Defined Implicitly by Equations

The statement centers on a pair ( p̄, x̄) satisfying the equation (1), or equiva-
lently such that x̄ ∈ S( p̄). It makes two assumptions: f is continuously differen-
tiable around ( p̄, x̄) and the partial Jacobian ∇x f ( p̄, x̄) is nonsingular (requiring of
course that m = n). The conclusion then is that a single-valued localization s of S
exists around p̄ for x̄ which moreover is continuously differentiable around p̄ with
Jacobian given by the formula

∇s(p) = −∇x f (p, s(p))−1 ∇ p f (p, s(p)).

The Dini classical implicit function theorem and its variants will be taken up in
detail in Section 1B after the development in Section 1A of an equivalent inverse
function theorem. Later in Chapter 1 we gradually depart from the assumption of
continuously differentiability of f to obtain far-reaching extensions of this classical
theorem. It will be illuminating, for instance, to reformulate the assumption about
the Jacobian ∇x f ( p̄, x̄) as an assumption about the function

h(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄)

giving the partial linearization of f at ( p̄, x̄) with respect to x and having h(x̄) = 0.
The condition corresponding to the invertibility of ∇x f ( p̄, x̄) can be turned into the
condition that the inverse mapping h−1 , with x̄ ∈ h−1 (0), has a single-valued local-
ization around 0 for x̄. In this way the theme of single-valued localizations can be
carried forward even into realms where f might not be differentiable and h could
be some other kind of “local approximation” of f . We will be able to operate with
a broad implicit function paradigm, extending in later chapters to much more than
solving equations. It will deal with single-valued localizations s of solution map-
pings S to “generalized equations.” These localizations s, if not differentiable, will
at least have other key properties.
1 Functions Defined Implicitly by Equations 9

1A. The Classical Inverse Function Theorem

In this section of the book, we state and prove the classical inverse function theorem
in two ways. In these proofs, and also later in the chapter, we will make use of the
following two observations from calculus.

Fact 1 (estimates for differentiable functions). If a function f : IRn → IRm is differ-

entiable at every point in a neighborhood of x̄ and the Jacobian mapping x → ∇ f (x)
is continuous at x̄, then for every ε > 0 there exists δ > 0 such that

(a) | f (x ) − f (x) − ∇ f (x)(x − x)| ≤ ε |x − x| for every x , x ∈ IBδ (x̄).

Equivalently, for every ε > 0 there exists δ > 0 such that

(b) | f (x ) − f (x) − ∇ f (x̄)(x − x)| ≤ ε |x − x| for every x , x ∈ IBδ (x̄).

Proof. For a vector h ∈ IRm with |h| = 1 and points x, x , x = x , in an open

neighborhood of x̄ where f is differentiable, define the function ϕ : IR → IR as
ϕ (t) = h, f (x+t(x − x)). Then ϕ is continuous on [0, 1] and differentiable in (0, 1)
and also ϕ  (t) = h, D f (x + t(x − x))(x − x). A basic result in calculus, the mean
value theorem, says that when a function ψ : IR → IR is continuous on an interval
[a, b] with a < b and differentiable in (a, b), then there exists a point c ∈ (a, b) such
that ψ (b) − ψ (a) = ψ  (c)(b − a); see, e.g., Bartle and Sherbert [1992], p. 197. Ap-
plying the mean value theorem to the function ϕ we obtain that there exists t¯ ∈ (0, 1)
such that
h, f (x ) − h, f (x) = h, D f (x + t¯(x − x))(x − x).
Then the triangle inequality and the assumed continuity of D f at x̄ give us (a). The
equivalence of (a) and (b) follows from the continuity of D f at x̄.

Fact 2 (stability of matrix nonsingularity). Suppose A is a matrix-valued function

from IRn to the space IRm×m of all m × m real matrices, such that the determinant
of A(x), as well as those of its minors, depends continuously on x around x̄ and the
matrix A(x̄) is nonsingular. Then there is a neighborhood U of x̄ such that A(x) is
nonsingular for every x ∈ U and, moreover, the function x → A(x)−1 is continuous
in U .
Proof. Since the nonsingularity of A(x) corresponds to its determinant being
nonzero, it is sufficient to observe that the determinant of A(x) (along with its
minors) depends continuously on x.
The classical inverse function theorem which parallels the classical implicit func-
tion theorem described in the introduction to this chapter reads as follows.
10 1 Functions Defined Implicitly by Equations

Theorem 1A.1 (classical inverse function theorem). Let f : IRn → IRn be contin-
uously differentiable in a neighborhood of a point x̄ and let ȳ := f (x̄). If ∇ f (x̄)
is nonsingular, then f −1 has a single-valued localization s around ȳ for x̄. More-
over, the function s is continuously differentiable in a neighborhood V of ȳ, and its
Jacobian satisfies

(1) ∇s(y) = ∇ f (s(y))−1 for every y ∈ V.

Examples.
1) For the function f (x) = x2 considered in the introduction, the inverse f −1 is
a set-valued mapping whose domain is [0, ∞). It has two single-valued localizations
√ √
around any ȳ > 0 for x̄ = 0, represented by either x(y) = y if x̄ > 0 or x(y) = − y
−1
if x̄ < 0. The inverse f has no single-valued localization around ȳ = 0 for x̄ = 0.
2) The inverse f −1 of the function f (x) = x3 is single-valued everywhere; it is
√ √
the function x(y) = 3 y. The inverse f −1 = 3 y is not differentiable at 0, which fits

with the observation that f (0) = 0.
3) For a higher-dimensional illustration, we look at diagonal real matrices
 
λ1 0
A=
0 λ2

and the function f : IR2 → IR2 which assigns to (λ1 , λ2 ) the trace y1 = λ1 + λ2 of
A and the determinant y2 = λ1 λ2 of A,
 
λ1 + λ2
f (λ1 , λ2 ) = .
λ 1 λ2

What can be said about the inverse of f ? The range of f consists of all y = (y1 , y2 )
such that 4y2 ≤ y21 . The Jacobian
 
1 1
∇ f (λ1 , λ2 ) =
λ 2 λ1

has determinant λ1 − λ2 , so it is nonsingular except along the line where λ1 = λ2 ,

which corresponds to 4y2 = y21 . Therefore, f −1 has a smooth single-valued local-
ization around y = (y1 , y2 ) for (λ1 , λ2 ) as long as 4y2 < y21 , in fact two such. But it
doesn’t have such a localization around other (y1 , y2 ).
It will be illuminating to look at two proofs2 of the classical inverse function
theorem. The one we lay out first requires no more background than the facts listed
at the beginning of this section, and it has the advantage of actually “calculating” a
single-valued localization of f −1 by a procedure which is well known in numerical

2 These two proofs are not really different, if we take into account that the contraction mapping
principle is proved by using a somewhat similar iterative procedure, see Section 5E.
1 Functions Defined Implicitly by Equations 11

analysis, namely Newton’s iterative method3 for solving nonlinear equations. The
second, which we include for the sake of connections with later developments, uti-
lizes a nonconstructive, but very broad, fixed-point argument.
Proof I of Theorem 1A.1. First we introduce some constants. Let a > 0 be a scalar
so small that, by appeal to Fact 2 in the beginning of this section, the Jacobian
matrix ∇ f (x) is nonsingular for every x in IBa (x̄) and the function x → ∇ f (x)−1 is
continuous in IBa (x̄). Set
c = max |∇ f (x)−1 |.
x∈IBa (x̄)

Take a > 0 smaller if necessary to obtain, on the basis of the estimate (a) in Fact 1,
that
1 
(2) | f (x ) − f (x) − ∇ f (x)(x − x)| ≤ |x − x| for every x , x ∈ IBa (x̄).
2c
Let b = a/(16c). Let s be the localization of f −1 with respect to the neighborhoods
IBb (ȳ) and IBa (x̄):
 
(3) gph s = IBb (ȳ) × IBa (x̄) ∩ gph f −1 .

We will show that s has the properties claimed. The argument is divided into three
steps.
S TEP 1: The localization s is nonempty-valued on IBb (ȳ) with x̄ ∈ s(ȳ), in par-
ticular.
The fact that x̄ ∈ s(ȳ) is immediate of course from (3), inasmuch as x̄ ∈ f −1 (ȳ).
Pick any y ∈ IBb (ȳ) and any x0 ∈ IBa/8 (x̄). We will demonstrate that the iterative
procedure

(4) xk+1 = xk − ∇ f (xk )−1 ( f (xk ) − y), k = 0, 1, . . .

produces a sequence of vectors x1 , x2 , . . . which is convergent to a point x ∈ f −1 (y)∩

IBa (x̄). The procedure (4) is the celebrated Newton’s iterative method for solving
the equation f (x) = y with a starting point x0 . By using induction we will show
that this procedure generates an infinite sequence {xk } satisfying for k = 1, 2, . . . the
following two conditions:

(5a) xk ∈ IBa (x̄)

and
3 Isaac Newton (1643–1727). In 1669 Newton wrote his paper De Analysi per Equationes Numero
Terminorum Infinitas, where, among other things, he describes an iterative procedure for approx-
imating real roots of a polynomial equation of third degree. In 1690 Joseph Raphson proposed
a similar iterative procedure for solving more general polynomial equations and attributed it to
Newton. It was Thomas Simpson who in 1740 stated the method in today’s form (using Newton’s
fluxions) for an equation not necessarily polynomial, without making connections to the works of
Newton and Raphson; he also noted that the method can be used for solving optimization problems
by setting the gradient to zero.
12 1 Functions Defined Implicitly by Equations
a
(5b) |xk − xk−1 | ≤ .
2k+1
To initialize the induction, we establish (5a) and (5b) for k = 1. Since x0 ∈
IBa/8 (x̄), the matrix ∇ f (x0 ) is indeed invertible, and (4) gives us x1 . The equality
in (4) for k = 0 can also be written as

x1 = −∇ f (x0 )−1 ( f (x0 ) − y − ∇ f (x0 )x0 ),

which we subtract from the obvious equality

x̄ = −∇ f (x0 )−1 ( f (x̄) − ȳ − ∇ f (x0 )x̄),

obtaining

x̄ − x1 = −∇ f (x0 )−1 ( f (x̄) − f (x0 ) − ȳ + y − ∇ f (x0 )(x̄ − x0 )).

Taking norms on both sides and utilizing (2) with x = x̄ and x = x0 we get
c 0
|x1 − x̄| ≤ |∇ f (x0 )−1 |(| f (x̄) − f (x0 ) − ∇ f (x0 )(x̄ − x0 )| + |y − ȳ|) ≤ |x − x̄| + cb.
2c
Inasmuch as |x0 − x̄| ≤ a/8, this yields
a a
|x1 − x̄| ≤ + cb = ≤ a.
16 8
Hence (5a) holds for k = 1. Moreover, by the triangle inequality,
a a a
|x1 − x0 | ≤ |x1 − x̄| + |x̄ − x0 | ≤ + = ,
8 8 4
which is (5b) for k = 1.
Assume now that (5a) and (5b) hold for k = 1, 2, . . . , j. Then the matrix ∇ f (xk )
is nonsingular for all such k and the iteration (4) gives us for k = j the point x j+1 :

(6) x j+1 = x j − ∇ f (x j )−1 ( f (x j ) − y).

Through the preceding iteration, for k = j − 1, we have

y = f (x j−1 ) + ∇ f (x j−1 )(x j − x j−1).

Substituting this expression for y into (6), we obtain

x j+1 − x j = −∇ f (x j )−1 ( f (x j ) − f (x j−1 ) − ∇ f (x j−1 )(x j − x j−1)).

Taking norms, we get from (2) that

1
|x j+1 − x j | ≤ c| f (x j ) − f (x j−1 ) − ∇ f (x j−1 )(x j − x j−1)| ≤ |x j − x j−1 |.
2
1 Functions Defined Implicitly by Equations 13

The induction hypothesis on (5b) for k = j then yields

1 1 a  a
|x j+1 − x j | ≤ |x j − x j−1 | ≤ = j+2 .
2 2 2 j+1 2
Hence, (5b) holds for k = j + 1. Further,
j+1 j+1
a a
|x j+1 − x̄| ≤ ∑ |xi − xi−1| + |x0 − x̄| ≤ ∑ 2i+1 + 8
i=1 i=1
a ∞ 1 a a a 5a
≤ ∑ i+ = + = ≤ a.
4 i=0 2 8 2 8 8

This gives (5a) for k = j + 1 and the induction step is complete. Thus, both (5a) and
(5b) hold for all k = 1, 2, . . ..
To verify that the sequence {xk } converges, we observe next from (5b) that, for
any k and j satisfying k > j, we have
k−1 ∞
a a
|xk − x j | ≤ ∑ |xi+1 − xi| ≤ ∑ 2 j+2 ≤ 2 j+1 .
i= j i= j

Hence, the sequence {xk } satisfies the Cauchy criterion, which is known to guaran-
tee that it is convergent.
Let x be the limit of this sequence. Clearly, from (5a), we have x ∈ IBa (x̄).
Through passing to the limit in (4), x must satisfy x = x − ∇ f (x)−1 ( f (x) − y), which
is equivalent to f (x) = y. Thus, we have proved that for any y ∈ IBb (ȳ) there exists
x ∈ IBa (x̄) such that x ∈ f −1 (y). In other words, the localization s of the inverse f −1
at ȳ for x̄ specified by (3) has nonempty values. In particular, IBb (ȳ) ⊂ dom f −1 .
S TEP 2: The localization s is single-valued on IBb (ȳ).
Let y ∈ IBb (ȳ) and suppose x and x belong to s(y). Then x, x ∈ IBa (x̄) and also
   
x = −∇ f (x)−1 f (x) − y − ∇ f (x)x and x = −∇ f (x)−1 f (x ) − y − ∇ f (x)x .

Consequently
 
x − x = −∇ f (x)−1 f (x ) − f (x) − ∇ f (x)(x − x) .

Taking norms on both sides and invoking (2), we get

1
|x − x| ≤ c| f (x ) − f (x) − ∇ f (x)(x − x)| ≤ |x − x|
2
which can only be true if x = x.
S TEP 3: The localization s is continuously differentiable in int IBb (ȳ) with ∇s(y)
expressed by (1).
An extension of the argument in Step 2 will provide a needed estimate. Consider
any y and y in IBb (ȳ) and let x = s(y) and x = s(y ). These elements satisfy
14 1 Functions Defined Implicitly by Equations
   
x = −∇ f (x)−1 f (x) − y − ∇ f (x)x and x = −∇ f (x)−1 f (x ) − y − ∇ f (x)x ,

so that
 
x − x = −∇ f (x)−1 f (x ) − f (x) − ∇ f (x)(x − x) − (y − y) .

This implies through (2) that

1
|x − x| ≤ c| f (x ) − f (x) − ∇ f (x)(x − x)| + c|y − y| ≤ |x − x| + c|y − y|,
2
hence |x − x| ≤ 2c|y − y|. Thus,

(7) |s(y ) − s(y)| ≤ 2c|y − y| for y, y ∈ IBb (ȳ).

This estimate means that the localization s is Lipschitz continuous on IBb (ȳ).
Now take any ε > 0. Then, from (a) in Fact 1, there exists δ > 0 such that
ε 
(8) | f (x ) − f (x) − ∇ f (x)(x − x)| ≤ |x − x| for every x , x ∈ IBδ (x̄).
2c2
Choose y ∈ int IBb (ȳ); then there exists τ > 0 such that τ < δ /(2c) and y+ h ∈ IBb (ȳ)
for any h ∈ IRn with |h| ≤ τ . From the estimate (7) we get that

|s(y + h) − s(y)| ≤ 2c|h| ≤ 2cτ ≤ δ .

Writing the equalities f (s(y + h)) = y + h and f (s(y)) = y as

s(y + h) = −∇ f (s(y))−1 ( f (s(y + h)) − y − h − ∇ f (s(y))s(y + h))

and
s(y) = −∇ f (s(y))−1 ( f (s(y)) − y − ∇ f (s(y))s(y))
and subtracting the second from the first, we obtain

s(y + h) − s(y) − ∇ f (s(y))−1h

= −∇ f (s(y))−1 ( f (s(y + h)) − f (s(y)) − ∇ f (s(y))(s(y + h) − s(y))).

Once again taking norms on both sides, and using (7) and (8), we get
cε
|s(y + h) − s(y) − ∇ f (s(y))−1 h| ≤ |s(y + h) − s(y)| ≤ ε |h| whenever h ∈ IBτ (0).
2c2
By definition, this says that the function s is differentiable at y and that its Jacobian
equals ∇ f (s(y))−1 , as claimed in (1). This Jacobian is continuous in int IBb (ȳ); this
comes from the continuity of ∇ f −1 in IBa (x̄) where are the values of s, and the conti-
nuity of s in int IBb (ȳ), and also taking into account that a composition of continuous
functions is continuous.
1 Functions Defined Implicitly by Equations 15

We can make a shortcut through Steps 1 and 2 of the Proof I, arriving at the
promised Proof II, if we employ a deeper result of analysis far beyond the frame-
work so far, namely the contraction mapping principle. Although we work here
in Euclidean spaces, we state this theorem in the framework of a complete metric
space, as is standard in the literature. A more general version of this principle for
set-valued mapping will be proved in Section 5E. The reader who wants to stick
with Euclidean spaces may assume that X is a closed nonempty subset of IRn with
metric ρ (x, y) = |x − y|.

Theorem 1A.2 (contraction mapping principle). Let X be a complete metric space

with metric ρ. Consider a point x̄ ∈ X and a function Φ : X → X for which there
exist scalars a > 0 and λ ∈ [0, 1) such that:
(a) ρ (Φ (x̄), x̄) ≤ a(1 − λ );
(b) ρ (Φ (x ), Φ (x)) ≤ λ ρ (x , x) for every x , x ∈ IBa (x̄).
Then there is a unique x ∈ IBa (x̄) satisfying x = Φ (x), that is, Φ has a unique fixed
point in IBa (x̄).
Most common in the literature is another formulation of the contraction mapping
principle which seems more general but is actually equivalent to 1A.2. To distin-
guish it from 1A.2, we call it basic.

Theorem 1A.3 (basic contraction mapping principle). Let X be a complete metric

space with metric ρ and let Φ : X → X . Suppose that there exists λ ∈ [0, 1) such that

ρ (Φ (x ), Φ (x)) ≤ λ ρ (x , x) for every x , x ∈ X .

Then there is a unique x ∈ X satisfying x = Φ (x).

Another equivalent version of the contraction mapping principle involves a pa-
rameter.

Theorem 1A.4 (parametric contraction mapping principle). Let P be a metric

space with metric σ and X be a complete metric space with metric ρ, and let
Φ : P × X → X . Suppose that there exist λ ∈ [0, 1) and μ ≥ 0 such that

(9) ρ (Φ (p, x ), Φ (p, x)) ≤ λ ρ (x , x) for every x , x ∈ X and p ∈ P

and

(10) ρ (Φ (p , x), Φ (p, x)) ≤ μ σ (p , p) for every p , p ∈ P and x ∈ X .

Then the mapping

  
(11) ψ : p → x ∈ X  x = Φ (p, x) for p ∈ P

is single-valued on P, which is moreover Lipschitz continuous on P with Lipschitz

constant μ /(1 − λ ).
16 1 Functions Defined Implicitly by Equations

Exercise 1A.5. Prove that theorems 1A.2, 1A.3 and 1A.4 are equivalent.
Guide. Let 1A.2 be true and let Φ satisfy the assumptions in 1A.3 with some λ ∈
[0, 1). Choose x̄ ∈ X; then Φ (x̄) ∈ X. Let a > ρ (x̄, Φ (x̄))/(1 − λ ). Then (a) and (b)
are satisfied with this a and hence there exists a unique fixed point x of Φ in IBa (x̄).
The uniqueness of this fixed point in the whole X follows from the contraction
property. To prove the converse implication first use (a)(b) to obtain that Φ maps
IBa (x̄) into itself and then use the fact that the closed ball IBa (x̄) equipped with
metric ρ is a complete metric space. Another way to have equivalence of 1A.2 and
1A.3 is to reformulate 1A.2 with a being possibly ∞.
Let 1A.3 be true and let Φ satisfy the assumptions (9) and (10) in  1A.4 with
corresponding
 λ and μ . Then, by 1A.3, for every fixed p ∈ P the set x ∈ X x =
Φ (p, x) is a singleton; that is, the mapping ψ in (11) is a function with domain
P. To complete the proof, choose p , p ∈ P and the corresponding x = Φ (p , x ),
x = Φ (p, x), and use (9), (10) and the triangle inequality to obtain

ρ (x , x) = ρ (Φ (p , x ), Φ (p, x))

≤ ρ (Φ (p , x ), Φ (p , x)) + ρ (Φ (p , x), Φ (p, x)) ≤ λ ρ (x , x) + μ σ (p , p).

Rearranging the terms gives us the desired Lipschitz continuity.

Proof II of Theorem 1A.1. Denote A = ∇ f (x̄) and let c := |A−1 |. There exists
a > 0 such that from the estimate (b) in Fact 1 (in the beginning of this section) we
have
1 
(12) | f (x ) − f (x) − ∇ f (x̄)(x − x)| ≤ |x − x| for every x , x ∈ IBa (x̄).
2c
Let b = a/(4c). The space IRn equipped with the Euclidean norm is a complete
metric space, so in this case X in Theorem 1A.2 is identified with IRn . Fix y ∈ IBb (ȳ)
and consider the function

Φy : x → x − A−1( f (x) − y) for x ∈ IBa (x̄).

We have  
ca 1
|Φy (x̄) − x̄| = | − A−1(ȳ − y)| ≤ cb = < a 1− ,
4c 2
hence condition (a) in the contraction mapping principle 1A.2 holds with the so
chosen a and λ = 1/2. Further, for any x, x ∈ IBa (x̄), from (12) we obtain that

|Φy (x) − Φy (x )| = |x − x − A−1 ( f (x) − f (x ))| ≤ |A−1 || f (x) − f (x ) − A(x − x)|
1 1
≤ c |x − x | = |x − x |.
2c 2
Thus condition (b) in 1A.2 is satisfied with the same λ . Hence, there is a unique
x ∈ IBa (x̄) such that Φy (x) = x; that is equivalent to f (x) = y.
1 Functions Defined Implicitly by Equations 17

Translated into our terminology, this tells us that f −1 has a single-valued local-
ization around ȳ for x̄ whose graph satisfies (3). The continuous differentiability is
argued once more through Step 3 of Proof I.

Exercise 1A.6. Prove Theorem 1A.1 by using, instead of iteration (4), the iteration

xk+1 = xk − ∇ f (x̄)−1 ( f (xk ) − y), k = 0, 1, . . . .

Guide. Follow the argument in Proof I with respective adjustments of the constants
involved.
In this and the following chapters we will derive the classical inverse function
theorem 1A.1 a number of times and in different ways from more general theorems
or utilizing other basic results. For instance, in Section 1F we will show how to
obtain 1A.1 from Brouwer’s invariance of domain theorem and in Section 4B we
will prove 1A.1 again with the help of the Ekeland variational principle.
There are many roads to be taken from here, by relaxing the assumptions in the
classical inverse function theorem, that lead to a variety of results. Some of them
are paved and easy to follow, others need more advanced techniques, and a few lead
to new territories which we will explore later in the book.

1B. The Classical Implicit Function Theorem

In this section we give a proof of the classical implicit function theorem stated by
Dini and described in the introduction to this chapter. We consider a function f :
IRd × IRn → IRn with values f (p, x), where p is the parameter and x is the variable
to be determined, and introduce for the equation f (p, x) = 0 the associated solution
mapping
  
(1) S : p → x ∈ IRn  f (p, x) = 0 for p ∈ IRd .

We restate the result, furnishing it with a label for reference.

Theorem 1B.1 (Dini classical implicit function theorem). Let f : IRd × IRn → IRn
be continuously differentiable in a neighborhood of ( p̄, x̄) and such that f ( p̄, x̄) = 0,
and let the partial Jacobian of f with respect to x at ( p̄, x̄), namely ∇x f ( p̄, x̄), be
nonsingular. Then the solution mapping S defined in (1) has a single-valued local-
ization s around p̄ for x̄ which is continuously differentiable in a neighborhood Q of
p̄ with Jacobian satisfying

(2) ∇s(p) = −∇x f (p, s(p))−1 ∇ p f (p, s(p)) for every p ∈ Q.

18 1 Functions Defined Implicitly by Equations

The classical inverse function theorem is the particular case of the classical im-
plicit function theorem in which f (p, x) = −p + f (x) (with a slight abuse of nota-
tion). However, it will also be seen now that the classical implicit function theorem
can be obtained from the classical inverse function theorem. For that, we first state
an easy-to-prove fact from linear algebra.

Lemma 1B.2. The square matrix

 
I 0
J= ,
B A

where I is the d × d identity matrix, 0 is the d × n zero matrix, B is an n × d matrix,

and A is an n × n nonsingular matrix, is nonsingular.
Proof. If J is singular, then there exists
 
p
y= = 0 such that Jy = 0,
x

which reduces to the equation

 
p
= 0.
Bp + Ax

Hence there exists x = 0 with Ax = 0, which contradicts the nonsingularity of A.

Proof of Theorem 1B.1. Consider the function

 
p
ϕ (p, x) =
f (p, x)

acting from IRd × IRn to itself. The inverse of this function is defined by the solutions
of the equation
   
p y1
(3) ϕ (p, x) = = ,
f (p, x) y2

where the vector (y1 , y2 ) ∈ IRd × IRn is now the parameter and (p, x) is the depen-
dent variable. The nonsingularity of the partial Jacobian ∇x f ( p̄, x̄) implies through
Lemma 1B.2 that the Jacobian of the function ϕ in (3) at the point (x̄, p̄), namely
the matrix  
I 0
J( p̄, x̄) = ,
∇ p f ( p̄, x̄) ∇x f ( p̄, x̄)
is nonsingular as well. Then, according to the classical inverse function theorem
1A.1, the inverse ϕ −1 of the function in (3) has a single-valued localization

(y1 , y2 ) → (q(y1 , y2 ), r(y1 , y2 )) around ( p̄, 0) for ( p̄, x̄)

1 Functions Defined Implicitly by Equations 19

which is continuously differentiable around ( p̄, 0). To develop formula (2), we note
that 
q(y1 , y2 ) = y1 ,
f (y1 , r(y1 , y2 )) = y2 .
Differentiating the second equality with respect to y1 by using the chain rule, we get

(4) ∇ p f (y1 , r(y1 , y2 )) + ∇x f (y1 , r(y1 , y2 )) · ∇y1 r(y1 , y2 ) = 0.

When (y1 , y2 ) is close to ( p̄, 0), the point (y1 , r(y1 , y2 )) is close to ( p̄, x̄) and then
∇x f (y1 , r(y1 , y2 )) is nonsingular (Fact 2 in Section 1A). Thus, solving (4) with re-
spect to ∇y1 r(y1 , y2 ) gives

∇y1 r(y1 , y2 ) = −∇x f (y1 , r(y1 , y2 ))−1 ∇ p f (y1 , r(y1 , y2 )).

In particular, at points (y1 , y2 ) = (p, 0) close to ( p̄, 0) we have that the mapping
p → s(p) := r(p, 0) is a single-valued localization of the solution mapping S in
(1) around p̄ for x̄ which is continuously differentiable around p̄ and its derivative
satisfies (2).
Thus, the classical implicit function theorem, as stated above, is equivalent to the
classical inverse function theorem as stated in the preceding section. We now look
at yet another equivalent result.

Theorem 1B.3 (correction function theorem). Let f : IRn → IRn be continuously

differentiable in a neighborhood of x̄. If ∇ f (x̄) is nonsingular, then the correction
mapping
  
Ξ : x → u ∈ IRn  f (x + u) = f (x̄) + ∇ f (x̄)(x − x̄) for x ∈ IRn

has a single-valued localization ξ around x̄ for 0. Moreover, ξ is continuously dif-

ferentiable in a neighborhood U of x̄ with ∇ξ (x̄) = 0.
Proof. Consider the function

ϕ : (x, u) → f (x + u) − f (x̄) − ∇ f (x̄)(x − x̄) for (x, u) ∈ IRn × IRn

in a neighborhood of (x̄, ū) for ū := 0. Since ∇u ϕ (x̄, ū) = ∇ f (x̄) is nonsingular,

we apply the classical implicit function theorem 1B.1 obtaining that the solution
mapping   
Ξ : x → u ∈ IRn  ϕ (x, u) = 0 for x ∈ IRn
has a smooth single-valued localization ξ around x̄ for 0. The chain rule gives us
∇ξ (x̄) = 0.

Exercise 1B.4. Prove that the correction function theorem implies the inverse func-
tion theorem.
Guide. Let ȳ := f (x̄) and assume A := ∇ f (x̄) is nonsingular. In these terms the
correction function theorem 1B.3 claims that the mapping
20 1 Functions Defined Implicitly by Equations
  
Ξ : z → ξ ∈ IRn  f (z + ξ ) = ȳ + A(z − x̄) for z ∈ IRn

has a single-valued localization ξ around x̄ for 0 and that ξ is continuously dif-

ferentiable around x̄ and has zero derivative at x̄. The affine function y → z(y) :=
x̄ + A−1 (y − ȳ) is the solution mapping of the linear equation ȳ + A(z − x̄) = y having
z(ȳ) = x̄. The composite function y → ξ (z(y)) hence satisfies

f (z(y) + ξ (z(y))) = ȳ + ∇ f (x̄)(z(y) − x̄) = y.

The function s(y) := z(y) + ξ (z(y)) is a single-valued localization of the inverse f −1

around ȳ for x̄. To show that ∇s(y) = ∇ f (s(y))−1 , use the chain rule.
Inasmuch as the classical inverse function theorem implies the classical implicit
function theorem, and the correction function theorem is a corollary of the classical
implicit function theorem, all three theorems — the inverse, the implicit and the
correction function theorems, stated in 1A.1, 1B.1 and 1B.3 respectively — are
equivalent.

Proposition 1B.5 (higher derivatives). In Theorem 1B.1, if f is k times continu-

ously differentiable around ( p̄, x̄) then the localization s of the solution mapping S
is k times continuously differentiable around p̄. Likewise in Theorem 1A.1, if f is k
times continuously differentiable around x̄, then the localization s of f −1 is k times
continuously differentiable around ȳ.
Proof. For the implicit function theorem 1B.1, this is an immediate consequence of
the formula in (2) by way of the chain rule for differentiation. It follows then for the
inverse function theorem 1A.1 as a special case.
If we relax the differentiability assumption for the function f , we obtain a result
of a different kind, the origins of which go back to the work of Goursat [1903].

Theorem 1B.6 (Goursat’s implicit function theorem). For the solution mapping S
defined in (1), consider a pair ( p̄, x̄) with x̄ ∈ S( p̄). Assume that:
(a) f (p, x) is differentiable with respect to x in a neighborhood of the point ( p̄, x̄),
and both f (p, x) and ∇x f (p, x) depend continuously on (p, x) in this neighborhood;
(b) ∇x f ( p̄, x̄) is nonsingular.
Then S has a single-valued localization around p̄ for x̄ which is continuous at p̄.
We will prove a far reaching generalization of this result in Section 2B, which we
supply with a detailed proof. In the following exercise we give a guide for a direct
proof.

Exercise 1B.7. Prove Theorem 1B.6.

Guide. Mimic the proof of 1A.1 by choosing a and q sufficiently small so that if

c = max |∇x f (p, x)−1 |,

x∈IBa (x̄)
p∈IBq ( p̄)
1 Functions Defined Implicitly by Equations 21

then one has, as in the estimate (a) in Fact 1, that for every x , x ∈ IBa (x̄) and p ∈
IBq ( p̄)

1 
(5) | f (p, x ) − f (p, x) − ∇x f (p, x)(x − x)| ≤ |x − x|.
2c
Then use the iteration

xk+1 = xk − ∇x f ( p̄, x̄)−1 f (p, xk )

to obtain that S has a nonempty graphical localization s around p̄ for x̄. As in Step 2
in Proof I of 1A.1, show that s is single-valued. To show continuity at p̄, for x = s(p)
subtract the equalities

x = ∇x f ( p̄, x̄)−1 ( f (p, x) − ∇x f ( p̄, x̄)x)

and
x̄ = ∇x f ( p̄, x̄)−1 ( f ( p̄, x̄) − ∇x f ( p̄, x̄)x̄),
and after adding and subtracting terms, use (5).

Exercise 1B.6. Consider a polynomial of degree n > 0,

n
p(x) = ∑ ai xi ,
i=0

where the coefficients a0 , . . . , an are real numbers. For each coefficient vector a =
(a0 , . . . , an ) ∈ IRn+1 let S(a) denote the set of all real zeros of p, so that S is a
mapping from IRn+1 to IR whose domain consists of the vectors a such that p has at
least one real zero. Let ā be a coefficient vector such that p has a simple real zero
s̄; thus p(s̄) = 0 but p (s̄) = 0. Prove that S has a smooth single-valued localization
around ā for s̄. Is such a statement correct when s̄ is a double zero?

1C. Calmness

In this section we introduce a continuity property of functions which will play an

important role in the book.

Calmness. A function f : IRn → IRm is said to be calm at x̄ relative to a set D in IRn

if x̄ ∈ D ∩ dom f and there exists a constant κ ≥ 0 such that

(1) | f (x) − f (x̄)| ≤ κ |x − x̄| for all x ∈ D ∩ dom f .

22 1 Functions Defined Implicitly by Equations

The calmness property (1) can alternatively be expressed in the form of the in-
clusion
f (x) ∈ f (x̄) + κ |x − x̄|IB for all x ∈ D ∩ dom f .
That expression connects with the generalization of the definition of calmness to
set-valued mappings, which we will discuss at length in Chapter 3.
Note that a function f which is calm at x̄ may have empty values at some points
x near x̄ when x̄ is on the boundary of dom f . If x̄ is an isolated point of D ∩ dom f ,
then trivially f is calm at x̄ relative to D with κ = 0.
We will mostly use a local version of the calmness property where the set D in
the condition (1) is a neighborhood of x̄; if such a neighborhood exists we simply
say that f is calm at x̄. Calmness of this kind can be identified with the finiteness of
the modulus which we proceed to define next.

Calmness modulus. For a function f : IRn → IRm and a point x̄ ∈ dom f , the calm-
ness modulus of f at x̄, denoted clm ( f ; x̄), is the infimum of the set of values κ ≥ 0
for which there exists a neighborhood D of x̄ such that (1) holds.
According to this, as long as x̄ is not an isolated point of dom f , the calmness
modulus satisfies
| f (x) − f (x̄)|
clm ( f ; x̄) = lim sup .
x∈dom f ,x→x̄, |x − x̄|
x=x̄

If x̄ is an isolated point we have clm ( f ; x̄) = 0. When f is not calm at x̄, from the
definition we get clm ( f ; x̄) = ∞. In this way,

f is calm at x̄ ⇐⇒ clm ( f ; x̄) < ∞.

Examples.
1) The function f (x) = x for x ≥ 0 is calm at any point of its domain [0, ∞),
always with calmness modulus
 1.
2) The function f (x) = |x|, x ∈ IR is not calm at zero but calm everywhere else.
3) The linear mapping A : x → Ax, where A is an m × n matrix, is calm at every
point x ∈ IRn and everywhere has the same modulus clm (A; x) = |A|.
Straight from the definition of the calmness modulus, we observe that
(i) clm ( f ; x̄) ≥ 0 for any x̄ ∈ dom f ;
(ii) clm (λ f ; x̄) = |λ | clm ( f ; x̄) for any λ ∈ IR and x̄ ∈ dom f ;
(iii) clm ( f + g; x̄) ≤ clm ( f ; x̄) + clm (g; x̄) for any x̄ ∈ dom f ∩ dom g.
These properties of the calmness modulus resemble those of a norm on a space of
functions f , but because clm( f ; x̄) = 0 does not imply f = 0, one could at most
contemplate a seminorm. However, even that falls short, since the modulus can take
on ∞, as can the functions themselves, which do not form a linear space because
they need not even have the same domain.

Exercise 1C.1 (properties of the calmness modulus). Prove that

(a) clm ( f ◦g; x̄)≤clm ( f ; g(x̄))·clm (g; x̄) whenever x̄ ∈ dom g and g(x̄) ∈ dom f ;
1 Functions Defined Implicitly by Equations 23

(b) clm ( f − g; x̄) = 0 ⇒ clm ( f ; x̄) = clm (g; x̄) whenever x̄ ∈ int(dom f ∩
dom g), but the converse is false.
With the concept of calmness in hand, we can interpret the differentiability of a
function f : IRn → IRm at a point x̄ ∈ int dom f as the existence of a linear mapping
A : IRn → IRm , represented by an n × m matrix, such that

(2) clm (e; x̄) = 0 for e(x) = f (x) − [ f (x̄) + A(x − x̄)].

According to property (iii) before 1C.1 there is at most one mapping M satisfying
(2). Indeed, if A1 and A2 satisfy (2) we have for the corresponding approximation
error terms e1 (x) and e2 (x) that

|A1 − A2 | = clm (e1 − e2 ; x̄) ≤ clm (e1 ; x̄) + clm(e2 ; x̄) = 0.

Thus, A is unique and the associated matrix has to be the Jacobian ∇ f (x̄). We con-
clude further from property (b) in 1C.1 that

clm ( f ; x̄) = |∇ f (x̄)|.

The following theorem complements Theorem 1A.1. It shows that the invertibil-
ity of the derivative is a necessary condition to obtain a calm single-valued localiza-
tion of the inverse.

Theorem 1C.2 (Jacobian nonsingularity from inverse calmness). Given f : IRn →

IRn and x̄ ∈ int dom f , let f be differentiable at x̄ and let ȳ := f (x̄). If f −1 has a
single-valued localization around ȳ for x̄ which is calm at ȳ, then the matrix ∇ f (x̄)
must be nonsingular.
Proof. The assumption that f −1 has a calm single-valued localization s around ȳ
for x̄ means several things: first, s is nonempty-valued around ȳ, that is, dom s is a
neighborhood of ȳ; second, s is a function; and third, s is calm at ȳ. Specifically,
there exist positive numbers a, b and κ and a function s with dom s ⊃ IBb (ȳ) and
values s(y) ∈ IBa (x̄) such that for every y ∈ IBb (ȳ) we have s(y) = f −1 (y) ∩ IBa (x̄)
and s is calm at ȳ with constant κ . Taking b smaller if necessary we have

(3) |s(y) − x̄| ≤ κ |y − ȳ| for any y ∈ IBb (ȳ).

Choose τ to satisfy 0 < τ < 1/κ . Then, since x̄ ∈ int dom f and f is differentiable
at x̄, there exists δ > 0 such that

(4) | f (x) − f (x̄) − ∇ f (x̄)(x − x̄)| ≤ τ |x − x̄| for all x ∈ IBδ (x̄).

If the matrix ∇ f (x̄) were singular, there would exist d ∈ IRn , |d| = 1, such that
∇ f (x̄)d = 0. Pursuing this possibility, let ε satisfy 0 < ε < min{a, b/τ , δ }. Then,
by applying (4) with x = x̄ + ε d, we get f (x̄ + ε d) ∈ IBb (ȳ). In terms of yε := f (x̄ +
ε d), we then have x̄ + ε d ∈ f −1 (yε ) ∩ IBa (x̄), hence s(yε ) = x̄ + ε d. The calmness
condition (3) then yields
24 1 Functions Defined Implicitly by Equations

1 1 κ κ
1 = |d| = |x̄ + ε d − x̄| = |s(yε ) − x̄| ≤ |yε − ȳ| = | f (x̄ + ε d) − f (x̄)|.
ε ε ε ε
Combining this with (4) and taking into account that ∇ f (x̄)d = 0, we arrive at 1 ≤
κτ |d| < 1 which is absurd. Hence ∇ f (x̄) is nonsingular.
Note that in the particular case of an affine function f (x) = Ax + b, where A is a
square matrix and b is a vector, calmness can be dropped from the set of assumptions
of Theorem 1C.2; the existence of a single-valued localization of f −1 around any
point is already equivalent to the nonsingularity of the Jacobian. This is not always
true even for polynomials. Indeed, the inverse of f (x) = x3 , x ∈ IR, has a single-
valued localization around the origin (which is not calm), but ∇ f (0) = 0.
The classical inverse function theorem 1A.1 combined with Theorem 1C.2 above
gives us

Theorem 1C.3 (symmetric inverse function theorem). Let f : IRn → IRn be contin-
uously differentiable around x̄. Then the following are equivalent:
(i) ∇ f (x̄) is nonsingular;
(ii) f −1 has a single-valued localization s around ȳ := f (x̄) for x̄ which is con-
tinuously differentiable around ȳ.
The formula for the Jacobian of the single-valued localization s of the inverse,

∇s(y) = ∇ f (s(y))−1 for y around ȳ,

comes as a byproduct of the statement (ii) by way of the chain rule.

A modification of the proof of Theorem 1C.2 gives us the converse to the correc-
tion function theorem.

Theorem 1C.4 (Jacobian nonsingularity from correction differentiability). Let f :

IRn → IRn be differentiable at x̄ and suppose that the correction mapping
  
Ξ : x → u ∈ IRn  f (x + u) = f (x̄) + ∇ f (x̄)(x − x̄) for x ∈ IRn

has a single-valued localization ξ around x̄ for 0 such that ξ is calm at x̄ with

clm (ξ ; x̄) = 0. Then ∇ f (x̄) is nonsingular.
Proof. If ∇ f (x̄) is singular, there must exist a vector d ∈ IRn with |d| = 1 such that
∇ f (x̄)d = 0. Then for all sufficiently small ε > 0 we have

f (x̄ + ε d + ξ (x̄ + ε d)) = f (x̄).

Thus, ε d + ξ (x̄ + ε d) ∈ Ξ (x̄) for all small ε > 0. Since Ξ has a single-valued local-
ization around x̄ we get x̄ + ε d + ξ (x̄ + ε d) = x̄. Then

1
1 = |d| = |ξ (x̄ + ε d)|.
ε
1 Functions Defined Implicitly by Equations 25

The right side of this equation goes to zero as ε → 0, and that produces a contradic-
tion.

Next, we extend the definition of calmness to its partial counterparts.

Partial calmness. A function f : IRd × IRn → IRm is said to be calm with respect to x
at ( p̄, x̄) ∈ dom f when the function ϕ with values ϕ (x) = f ( p̄, x) is calm at x̄. Such
calmness is said to be uniform in p at ( p̄, x̄) when there exists a constant κ > 0 and
neighborhoods Q of p̄ and U of x̄ such that actually

| f (p, x) − f (p, x̄)| ≤ κ |x − x̄| for all (p, x) ∈ (Q × U) ∩ dom f .

Correspondingly, the partial calmness modulus of f with respect to x at ( p̄, x̄) is

denoted as clm x ( f ; ( p̄, x̄)), while the uniform partial calmness modulus is

 x ( f ; ( p̄, x̄)) := | f (p, x) − f (p, x̄)|

clm lim sup
x→x̄, p→ p̄, |x − x̄|
(p,x)∈dom f , x=x̄

provided that every neighborhood of ( p̄, x̄) contains points (p, x) ∈ dom f with x = x̄.
Observe in this context that differentiability of f (p, x) with respect to x at ( p̄, x̄) ∈
int dom f is equivalent to the existence of a linear mapping A : IRn → IRm , the partial
derivative of f with respect to x at ( p̄, x̄), which satisfies

clm (e; x̄) = 0 for e(x) = f ( p̄, x) − [ f ( p̄, x̄) + A(x − x̄)],

and then A is the partial derivative Dx f ( p̄, x̄). In contrast, under the stronger condi-
tion that
 x (e; ( p̄, x̄)) = 0, for e(p, x) = f (p, x) − [ f ( p̄, x̄) + A(x − x̄)],
clm

we say f is differentiable with respect to x uniformly in p at ( p̄, x̄). This means that
for every ε > 0 there are neighborhoods Q of p̄ and U of x̄ such that

| f (p, x) − f (p, x̄) − Dx f ( p̄, x̄)(x − x̄)| ≤ ε |x − x̄| for p ∈ Q and x ∈ U.

Exercise 1C.5 (joint calmness criterion). Let f : IRd × IRn → IRm be calm in x uni-
formly in p and calm in p, both at ( p̄, x̄). Show that f is calm at ( p̄, x̄).

Exercise 1C.6 (nonsingularity characterization). Let f : IRn → IRn be differentiable

at x̄, let ȳ = f (x̄), and suppose that f −1 has a single-valued localization s around ȳ
for x̄ which is continuous at ȳ. Prove in this setting that s is differentiable at ȳ if and
only if the Jacobian ∇ f (x̄) is nonsingular.
Guide. The “only if” part can be obtained from Theorem 1C.2, using the fact that
if s is differentiable at x̄, it must be calm at x̄. In the other direction, starting from
26 1 Functions Defined Implicitly by Equations

the assumption that ∇ f (x̄) is nonsingular, argue in a manner parallel to the first part
of Step 3 of Proof I of Theorem 1A.1.

1D. Lipschitz Continuity

Calmness is a “one-point” version of the well-known “two-point” property of func-

tions named after Rudolf Otto Sigismund Lipschitz (1832–1903). That property has
already entered our deliberations in Section 1A in connection with the Proof II of
the classical inverse function theorem by way of the contraction mapping principle,
but we investigate it now more directly. For convenience we recall the definition:

Lipschitz continuous functions. A function f : IRn → IRm is said to be Lipschitz

continuous relative to a set D, or on a set D, if D ⊂ dom f and there exists a constant
κ ≥ 0 (Lipschitz constant) such that

(1) | f (x ) − f (x)| ≤ κ |x − x| for all x , x ∈ D.

It is said to be Lipschitz continuous around x̄ when this holds for some neighborhood
D of x̄. We say further, in the case of an open set C, that f is locally Lipschitz
continuous on C if it is a Lipschitz continuous function around every point x of C.

Lipschitz modulus. For a function f : IRn → IRm and a point x̄ ∈ int dom f , the
Lipschitz modulus of f at x̄, denoted lip ( f ; x̄), is the infimum of the set of values of
κ for which there exists a neighborhood D of x̄ such that (1) holds. Equivalently,

| f (x ) − f (x)|
(2) lip ( f ; x̄) := lim sup .
x ,x→x̄, |x − x|
x=x

Note that, by this definition, for the Lipschitz modulus we have lip ( f ; x̄) = ∞
precisely in the case where, for every κ > 0 and every neighborhood D of x̄, there
are points x , x ∈ D violating (1). Thus,

f is Lipschitz continuous around x̄ ⇐⇒ lip ( f ; x̄) < ∞.

A function f with lip ( f ; x̄) < ∞ is also called strictly continuous at x̄. For an open set
C, a function f is locally Lipschitz continuous on C exactly when lip ( f ; x) < ∞ for
every x ∈ C. Every continuously differentiable function on an open set C is locally
Lipschitz continuous on C.
1 Functions Defined Implicitly by Equations 27

Examples.
1) The function x → |x|, x ∈ IRn , is Lipschitz continuous everywhere with
lip (|x|; x) = 1; it is not differentiable at 0.
2) An affine function f : x → Ax + b, corresponding to a matrix A ∈ IRm×n and a
vector b ∈ IRm , has lip ( f ; x̄) = |A| for every x̄ ∈ IRn .
3) If f is continuously differentiable in a neighborhood of x̄, then lip ( f ; x̄) =
|∇ f (x̄)|.
Like the calmness modulus, the Lipschitz modulus has the properties of a semi-
norm, except in allowing for ∞:
(i) lip ( f ; x̄) ≥ 0 for any x̄ ∈ int dom f ;
(ii) lip (λ f ; x̄) = |λ | lip ( f ; x̄) for any λ ∈ IR and x̄ ∈ int dom f ;
(iii) lip ( f + g; x̄) ≤ lip ( f ; x̄) + lip(g; x̄) for any x̄ ∈ int dom f ∩ int dom g.

Exercise 1D.1 (properties of the Lipschitz modulus). Prove that

(a) lip ( f ◦g; x̄) ≤ lip ( f ; g(x̄)) · lip (g; x̄) when x̄ ∈ int dom g and g(x̄) ∈ int dom f ;
(b) lip ( f − g; x̄) = 0 ⇒ lip ( f ; x̄) = lip (g; x̄) when x̄ ∈ int dom f ∩ int dom g;
(c) lip ( f ; ·)is upper semicontinuous
  x̄ ∈ int dom f where it is finite;
at any
(d) the set x ∈ int dom f  lip ( f ; x) < ∞ is open.
Bounds on the Lipschitz modulus lead to Lipschitz constants relative to sets, as
long as convexity is present. First, recall that a set C ⊂ IRn is convex if

(1 − τ )x0 + τ x1 ∈ C for all τ ∈ (0, 1) when x0 , x1 ∈ C,

or in other words, if C contains for any pair of its points the entire line segment that
joins them. The most obvious convex set is the ball IB as well as its interior, while
the boundary of the ball is of course nonconvex.

Exercise 1D.2 (Lipschitz continuity on convex sets). Show that if C is a convex sub-
set of int dom f such that lip ( f ; x) ≤ κ for all x ∈ C, then f is Lipschitz continuous
relative to C with constant κ.
Guide. It is enough to demonstrate for an arbitrary choice of points x and x in C
and ε > 0 that | f (x ) − f (x)| ≤ (κ + ε )|x − x|. Argue that the line segment joining
x and x is a compact subset of int dom f which can be covered by finitely many
balls on which f is Lipschitz continuous with constant κ + ε . Moreover these balls
can be chosen in such a way that a finite sequence of points x0 , x1 , . . . , xr along the
segment, starting with x0 = x and ending with xr = x , has each consecutive pair in
one of them. Get the Lipschitz inequality for x and x from the Lipschitz inequalities
for these pairs.

Exercise 1D.3 (Lipschitz continuity from differentiability). If f is continuously dif-

ferentiable on an open set O and C is a compact convex subset of O, then f is
Lipschitz continuous relative to C with constant κ = maxx∈C |∇ f (x)|.
28 1 Functions Defined Implicitly by Equations

Convexity also provides an important class of examples of Lipschitz continu-

ous functions from IRn into itself which are not everywhere differentiable, namely
distance and projection mappings; for an illustration see Fig. 1.3.

Distance and projection. For a point x ∈ IRn and a set C ⊂ IRn , the quantity

(3) dC (x) = d(x,C) = inf |x − y|

y∈C

is called the distance from x to C. (Whether the notation dC (x) or d(x,C) is used is a
matter of convenience in a given context.) Any point y of C which is closest to x in
the sense of achieving this distance is called a projection of x on C. The set of such
projections is denoted by PC (x). Thus,

(4) PC (x) = argmin |x − y|.

y∈C

y
x

PC(x) C

Fig. 1.3 Distance and projection.

In this way, C gives rise to a distance function dC and a projection mapping PC . If

C is empty, then trivially dC (x) = ∞ for all x, whereas if C is nonempty, then dC (x)
is finite (and nonnegative) for all x. As for PC , it is, in general, a set-valued mapping
from IRn into C, but additional properties follow from particular assumptions on C,
as we explore next.

Proposition 1D.4 (properties of distance and projection).

(a) For a nonempty set C ⊂ IRn , one has dC (x) = dcl C (x) for all x. Moreover, C
is closed if and only if every x with dC (x) = 0 belongs to C.
(b) For a nonempty set C ⊂ IRn , the distance function dC is Lipschitz continuous
on IRn with Lipschitz constant κ = 1. As long as C is closed, one has

0 if x̄ ∈ int C,
lip (dC ; x̄) =
1 otherwise.
1 Functions Defined Implicitly by Equations 29

(c) For a nonempty, closed set C ⊂ IRn , the projection set PC (x) is nonempty,
closed and bounded for every x ∈ IRn .
Proof. For (a), we fix any x ∈ IRn and note that obviously dcl C (x) ≤ dC (x). This
inequality can’t be strict because for any ε > 0 we can find y ∈ cl C making |x −
y| < dcl C (x) + ε but then also find y ∈ C with |y − y | < ε , in which case we have
dC (x) ≤ |x − y| < dcl C (x) + 2ε . In particular, this argument reveals that dcl C (x) = 0
if and only  if x ∈ cl C. Having
 demonstrated that dC (x) = dcl C (x), we may conclude
that C = x  dC (x) = 0 if and only if C = cl C.
For (b), consider any points x and x along with any ε > 0. Take any point y ∈ C
such that |x − y| ≤ dC (x) + ε . We have dC (x ) ≤ |x − y| ≤ |x − x| + |x − y| ≤ |x −
x| + dC (x) + ε , and through the arbitrariness of ε therefore dC (x ) − dC (x) ≤ |x − x|.
The same thing must hold with the roles of x and x reversed, so this demonstrates
that dC is Lipschitz continuous with constant 1.
Let C be nonempty and closed. If x̄ ∈ int C, we have dC (x) = 0 for all x in a
neighborhood of x̄ and consequently lip (dC ; x̄) = 0. Suppose now that x̄ ∈ / int C. We
will show that lip (dC ; x̄) ≥ 1, in which case equality must actually hold because we
already know that dC is Lipschitz continuous on IRn with constant 1. According to
the property of the Lipschitz modulus displayed in Exercise 1D.1(c), it is sufficient
to consider x̄ ∈ / C. Let x̃ ∈ PC (x̄). Then on the line segment from x̃ to x̄ the distance
increases linearly, that is, dC (x̃ + τ (x̄ − x̃)) = τ d(x̄,C) for 0 ≤ τ ≤ 1 (prove!). Hence
for the two points x = x̃ + τ (x̄ − x̃) and x = x̃ + τ  (x̄ − x̃) we have |dC (x ) − dC (x)| =
|τ  − τ ||x̃ − x̄| = |x − x|. Note that x̄ can be approached by such pairs of points and
hence lip (dC ; x̄) ≥ 1.
Turning now to (c), we again fix any x ∈ IRn and choose a sequence of points
yk ∈ C such that |x − yk | → dC (x) as k → ∞. This sequence is bounded and therefore
has an accumulation point y in C, inasmuch as C is closed. Since |x − yk | ≥ dC (x) for
all k, it follows that |x − y| = dC (x). Thus, y ∈ PC (x), so PC (x) is not empty. Since by
definition PC (x) is the intersection of C with the closed ball with center x and radius
dC (x), it’s clear that PC (x) is furthermore closed and bounded.
It has been seen in 1D.4(c) that for any nonempty closed set C ⊂ IRn the projection
mapping PC : IRn →→ C is nonempty-compact-valued, but when might it actually be
single-valued as well? The convexity of C is the additional property that yields this
conclusion, as will be shown in the following proposition4.

Proposition 1D.5 (Lipschitz continuity of projection mappings). For a nonempty,

closed, convex set C ⊂ IRn , the projection mapping PC is single-valued (a function)
from IRn onto C which moreover is Lipschitz continuous with Lipschitz constant
κ = 1. Also,

(5) PC (x̄) = ȳ ⇐⇒ x̄ − ȳ, y − ȳ ≤ 0 for all y ∈ C.

4 A set C such that PC is single-valued is called a Chebyshev set. A nonempty, closed, convex set
is always a Chebyshev set, and in IRn the converse is also true; for proofs of this fact see Borwein
and Lewis [2006] and Deutsch [2001]. The question of whether a Chebyshev set in an arbitrary
infinite-dimensional Hilbert space must be convex is still open.
30 1 Functions Defined Implicitly by Equations

Proof. We have PC (x) = 0/ in view of 1D.4(c). Suppose ȳ ∈ PC (x̄). For any τ ∈ (0, 1),
any y ∈ IRn and yτ = (1 − τ )ȳ + τ y we have the identity

|x̄ − yτ |2 − |x̄ − ȳ|2 = |(yτ − ȳ) − (x̄ − ȳ)|2 − |x̄ − ȳ|2

(6) = |yτ − ȳ|2 − 2x̄ − ȳ, yτ − ȳ
= τ 2 |y − ȳ|2 − 2τ x̄ − ȳ, y − ȳ.

If y ∈ C, we also have yτ ∈ C by convexity, so the left side is nonnegative. This

implies that τ |y − ȳ|2 ≥ 2x̄ − ȳ, y − ȳ for all τ ∈ (0, 1). Thus, the inequality in (5)
holds. On the other hand, let x̄ − ȳ, y − ȳ ≤ 0 for all y ∈ C. If y ∈ C is such that
|x̄ − y| ≤ |x̄ − ȳ| then for yτ as in (6) we get τ 2 |y − ȳ|2 ≤ 2x̄ − ȳ, y − ȳ ≤ 0 showing
that y = ȳ. Thus (5) is fully confirmed along with the fact that PC (x̄) can’t contain
any y = ȳ.
Consider now two points x0 and x1 and their projections y0 = PC (x0 ) and y1 =
PC (x1 ). On applying (5), we see that

x0 − y0, y1 − y0  ≤ 0 and x1 − y1 , y0 − y1  ≤ 0.

When added, these inequalities give us

0 ≥ x0 − y0 − x1 + y1 , y1 − y0  = |y1 − y0|2 − x1 − x0 , y1 − y0 

and consequently

|y1 − y0 |2 ≤ x1 − x0 , y1 − y0  ≤ |x1 − x0 ||y1 − y0 |.

It follows that
|y1 − y0 | ≤ |x1 − x0 |.
Thus, PC is Lipschitz continuous with Lipschitz constant 1.
Projection mappings have many uses in numerical analysis and optimization.
Note that PC always fails to be differentiable on the boundary of C. As an example,
when C is the set of nonpositive reals IR− one has

0 for x ≥ 0,
PC (x) =
x for x < 0
and this function is not differentiable at x = 0.
It is clear from the definitions of the calmness and Lipschitz moduli that we
always have
clm ( f ; x̄) ≤ lip ( f ; x̄).
This relation is illustrated in Fig. 1.4.
In the preceding section we showed how to characterize differentiability through
calmness. Now we introduce a sharper concept of derivative which is tied up with
the Lipschitz modulus.
1 Functions Defined Implicitly by Equations 31

Fig. 1.4 Plots of calm and Lipschitz continuous functions. On the left is the plot of the func-
tion f (x) = (−1)n+1 9x + (−1)n 22n+1 /5n−2 , |x| ∈ [xn+1 , xn ] for xn = 4n−1 /5n−2 , n = 1, 2, . . . for
which clm ( f ; 0) < lip( f ; 0) < ∞. On the right is the plot of the function f (x) = (−1)n+1 (6 + n)x +
(−1)n 210(5 + n)!/(6 + n)!, |x| ∈ [xn+1 , xn ] for xn = 210(4 + n)!/(6 + n)!, n = 1, 2, . . . for which
clm ( f ; 0) < lip( f ; 0) = ∞.

Strict differentiability. A function f : IRn → IRm is said to be strictly differentiable

at a point x̄ if there is a linear mapping A : IRn → IRm such that

lip (e; x̄) = 0 for e(x) = f (x) − [ f (x̄) + A(x − x̄)].

In particular, in this case we have that clm (e; x̄) = 0 and hence f is differentiable
at x̄ with A = D f (x̄), but strictness imposes a requirement on the difference

e(x) − e(x ) = f (x) − [ f (x ) + D f (x̄)(x − x )]

also when x = x̄. Specifically, it demands the existence for each ε > 0 of a neigh-
borhood U of x̄ such that

| f (x) − [ f (x ) + D f (x̄)(x − x )]| ≤ ε |x − x | for every x, x ∈ U.

Exercise 1D.6 (strict differentiability from continuous differentiability). Prove that

any function f that is continuously differentiable in a neighborhood of x̄ is strictly
differentiable at x̄.
Guide. Adopt formula (b) in Fact 1 in the beginning of Section 1A.
The converse to the assertion in Exercise 1D.6 is false, however: f can be
strictly differentiable at x̄ without being continuously differentiable around x̄. This
is demonstrated in Fig. 1.5 showing the graphs of two functions that are both dif-
ferentiable at origin but otherwise have different properties. On the left is the graph
of the continuous function f : [−1, 1] → IR which is even, and on [0, 1] has values
f (0) = 0, f (1/n) = 1/n2, and is linear in the intervals [1/n, 1/(n + 1)]. This function
is strictly differentiable at 0, but in every neighborhood of 0 there are points where
32 1 Functions Defined Implicitly by Equations

differentiability is lacking. On the right is the graph of the function5


x/2 + x2 sin(1/x) for x = 0,
f (x) =
0 for x = 0,

which is differentiable at 0 but not strictly differentiable there.

Fig. 1.5 Plots of functions differentiable at the origin. The function on the left is strictly differen-
tiable at the origin but not continuously differentiable. The function on the right is differentiable at
the origin but not strictly differentiable there.

The second of these examples has the interesting feature that, even though
f (0) = 0 and f  (0) = 0, no single-valued localization of f −1 exists around 0 for
0. In contrast, we will see in 1D.9 that strict differentiability would ensure the avail-
ability of such a localization.

Exercise 1D.7 (strict differentiability from differentiability). Consider a function f

which is differentiable at every point in a neighborhood of x̄. Prove that f is strictly
differentiable at x̄ if and only if the Jacobian ∇ f is continuous at x̄.
Guide. Let f be strictly differentiable at x̄ and let ε > 0. Then there exists δ1 > 0
such that for any x1 , x2 ∈ IBδ1 (x̄) we have

1
(7) | f (x2 ) − f (x1 ) − ∇ f (x̄)(x2 − x1 )| ≤ ε |x1 − x2 |.
2
Fix an x1 ∈ IBδ1 /2 (x̄). For this x1 there exists δ2 > 0 such that for any x ∈ IBδ2 (u1 ),

1
(8) | f (x ) − f (x1 ) − ∇ f (x1 )(x − x1 )| ≤ ε |x − x1 |.
2
Make δ2 smaller if necessary so that IBδ2 (x1 ) ⊂ IBδ1 (x̄). By (7) with x2 replaced by
x and by (8), we have

|∇ f (x1 )(x − x1) − ∇ f (x̄)(x − x1 )| ≤ ε |x − x1 |.

5 These two examples are from Nijenhuis [1974], where the introduction of strict differentiability

is attributed to Leach [1961]. By the way, Nijenhuis dedicated his paper to Carl Allendoerfer “for
not taking the implicit function theorem for granted.” In the book we follow this advice.
1 Functions Defined Implicitly by Equations 33

This implies
|∇ f (x1 ) − ∇ f (x̄)| ≤ ε .
Since x1 is arbitrarily chosen in IBδ1 /2 (x̄), we obtain that the Jacobian is continuous
at x̄.
For the opposite direction, use Fact 1 in the beginning of Section 1A.

Exercise 1D.8 (continuous differentiability from strict differentiability). Prove that

a function f is strictly differentiable at every point of an open set O if and only if it
is continuously differentiable on O.
Guide. Apply 1D.7.
With the help of the strict derivative we can obtain a new version of the classical
inverse function theorem 1A.1.

Theorem 1D.9 (symmetric inverse function theorem under strict differentiabil-

ity). Let f : IRn → IRn be strictly differentiable at x̄. Then the following are equiva-
lent:
(i) ∇ f (x̄) is nonsingular;
(ii) f −1 has a single-valued localization s around ȳ := f (x̄) for x̄ which is strictly
differentiable at ȳ. In that case, moreover, ∇s(ȳ) = ∇ f (x̄)−1 .
Proof. The implication (i) ⇒ (ii) can be accomplished by combining various pieces
already present in the proofs of Theorem 1A.1, since strict differentiability of f at x̄
gives us by definition the estimate (b) in Fact 1 in Section 1A. Parallel to Proof II of
1A.1 we find positive constants a and b and a single-valued localization of f −1 of
the form
s : y → f −1 (y) ∩ IBa (x̄) for y ∈ IBb (ȳ).
Next, by using the equation

s(y) = −A−1 ( f (s(y)) − y − As(y)) for y ∈ IBb (ȳ),

where A = ∇ f (x̄), we demonstrate Lipschitz continuity of s around ȳ as in the begin-

ning of Step 3 of Proof I of Theorem 1A.1. Finally, to obtain strict differentiability
of s at ȳ, repeat the second part of Step 3 of Proof I with ∇ f (s(y)) replaced with A.
For the converse implication we invoke Theorem 1C.2 and the fact that strict differ-
entiability entails calmness.

Working now towards a corresponding version of the implicit function theorem,

we look at additional forms of Lipschitz continuity and strict differentiability.

Partial Lipschitz continuity. A function f : IRd × IRn → IRm is said to be Lipschitz

continuous with respect to x around ( p̄, x̄) ∈ int dom f when the function x → f ( p̄, x)
is Lipschitz continuous around x̄; the associated Lipschitz modulus of f with respect
to x is denoted by lip x ( f ; ( p̄, x̄)). We say f is Lipschitz continuous with respect to x
34 1 Functions Defined Implicitly by Equations

uniformly in p around ( p̄, x̄) ∈ int dom f when there are neighborhoods Q of p̄ and
U of x̄ along with a constant κ and such that

| f (p, x) − f (p, x )| ≤ κ |x − x | for all x, x ∈ U and p ∈ Q.

Accordingly, the partial uniform Lipschitz modulus with respect to x has the form

 ( f ; ( p̄, x̄)) := lim sup | f (p, x ) − f (p, x)| .
lip x
x,x →x̄,p→ p̄, |x − x|
x=x

Exercise 1D.10 (partial uniform Lipschitz modulus with differentiability). Show

that if the function f : IRd × IRn → IRm is differentiable with respect to x at all points
(p, x) in some neighborhood of ( p̄, x̄), then

 ( f ; ( p̄, x̄)) = lim sup |∇x f (p, x)|.

lip x
(p,x)→( p̄,x̄)

Strict partial differentiability. A function f : IRd × IRn → IRm is said to be strictly

differentiable with respect to x at ( p̄, x̄) if the function x → f ( p̄, x) is strictly differ-
entiable at x̄. It is said to be strictly differentiable with respect to x uniformly in p at
( p̄, x̄) if

 (e; ( p̄, x̄)) = 0 for e(p, x) = f (p, x) − [ f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄)],
lip x

or in other words, if for every ε > 0 there are neighborhoods Q of p̄ and U of x̄ such
that

| f (p, x) − [ f (p, x ) + Dx f ( p̄, x̄)(x − x )]| ≤ ε |x − x | for all x, x ∈ U and p ∈ Q.

Exercise 1D.11 (joint differentiability criterion). Let f : IRd × IRn → IRm be strictly
differentiable with respect to x uniformly in p and be differentiable with respect to
p, both at ( p̄, x̄). Prove that f is differentiable at ( p̄, x̄).

Exercise 1D.12 (joint strict differentiability criterion). Prove that f : IRd ×IRn → IRm
is strictly differentiable at ( p̄, x̄) if and only if it is strictly differentiable with respect
to x uniformly in p and strictly differentiable with respect to p uniformly in x, both
at ( p̄, x̄).
We state next the implicit function counterpart of Theorem 1D.9.

Theorem 1D.13 (implicit functions under strict partial differentiability). Given f :

IRd × IRn → IRn and ( p̄, x̄) with f ( p̄, x̄) = 0, suppose that f is strictly differentiable
at ( p̄, x̄) and let the partial Jacobian ∇x f ( p̄, x̄) be nonsingular. Then the solution
mapping
1 Functions Defined Implicitly by Equations 35
  
S : p → x ∈ IRn  f (p, x) = 0
has a single-valued localization s around p̄ for x̄ which is strictly differentiable at p̄
with its Jacobian expressed by

∇s( p̄) = −∇x f ( p̄, x̄)−1 ∇ p f ( p̄, x̄).

Proof. We apply Theorem 1D.9 in a manner parallel to the way that the classical
implicit function theorem 1B.1 was derived from the classical inverse function the-
orem 1A.1.

Exercise 1D.14. Let f : IR → IR be strictly differentiable at 0 and let f (0) = 0.

Consider the following equation in x with a parameter p:
 x
p f (x) = f (pt)dt.
0

Prove that the solution mapping associated with this equation has a strictly differen-
tiable single-valued localization around 0 for 0.

Guide. The function g(p, x) = p f (x)− 0x f (pt)dt satisfies (∂ g/∂ x)(0, 0) = − f (0),
which is nonzero by assumption. For any ε > 0 there exist open intervals Q and U
centered at 0 such that for every p ∈ Q and x, x ∈ U we have

∂g
|g(p, x) − g(p, x) − (0, 0)(x − x)|
∂x 
x
= |p( f (x) − f (x )) − f (pt)dt + f (0)(x − x )|
x
= |p( f (x) − f (x )) − ( f (px̃) − f (0))(x − x )|
≤ |p( f (x) − f (x ) − f  (0)(x − x ))| + |p f  (0)(x − x )|
+|( f (px̃) − f (0))(x − x )| ≤ ε |x − x |,

where the mean value theorem guarantees that xx f (pt)dt = (x − x ) f (px̃) for some
x̃ between x and x. Hence, g is strictly differentiable with respect to x uniformly in
p at (0, 0). Prove in a similar way that g is strictly differentiable with respect to p
uniformly in x at (0, 0). Then apply 1D.12 and 1D.13.

1E. Lipschitz Invertibility from Approximations

In this section we completely depart from differentiation and develop inverse
and implicit function theorems for equations in which the functions are merely
36 1 Functions Defined Implicitly by Equations

continuous. The price to pay is that the single-valued localization of the inverse that
is obtained might not be differentiable, but at least it will have a Lipschitz property.
The way to do that is found through notions of how a function f may be “approx-
imated” by another function h around a point x̄. Classical theory focuses on f being
differentiable at x̄ and approximated there by the function h giving its “linearization”
at x̄, namely h(x) = f (x̄) + ∇ f (x̄)(x − x̄). Differentiability corresponds to having
f (x) = h(x) + o(|x − x̄|) around x̄, which is the same as clm ( f − h; x̄) = 0, whereas
strict differentiability corresponds to the stronger requirement that lip ( f − h; x̄) = 0.
The key idea is that conditions like this, and others in a similar vein, can be applied
to f and h even when h is not a linearization dependent on the existence of ∇ f (x̄).
Assumptions on the nonsingularity of ∇ f (x̄), corresponding in the classical setting
to the invertibility of the linearization, might then be replaced by assumptions on
the invertibility of some other approximation h.

First-order approximations of functions. Consider a function f : IRn → IRm and

a point x̄ ∈ int dom f . A function h : IRn → IRm with x̄ ∈ int dom h is a first-order
approximation to f at x̄ if h(x̄) = f (x̄) and

clm (e; x̄) = 0 for e(x) = f (x) − h(x),

which can also be written as f (x) = h(x) + o(|x − x̄|). It is a strict first-order approx-
imation if the stronger condition holds that

lip (e; x̄) = 0 for e(x) = f (x) − h(x).

In other words, h is a first-order approximation to f at x̄ when f (x̄) = h(x̄) and

for every ε > 0 there exists δ > 0 such that

| f (x) − h(x)| ≤ ε |x − x̄| for every x ∈ IBδ (x̄),

and a strict first-order approximation when

|[ f (x) − h(x)] − [ f (x ) − h(x )]| ≤ ε |x − x | for all x, x ∈ IBδ (x̄).

Clearly, if h is a (strict) first-order approximation to f , then f is a (strict) first-

order approximation to h.
First-order approximations obey calculus rules which follow directly from the
corresponding properties of the calmness and Lipschitz moduli:
(i) If q is a (strict) first-order approximation to h at x̄ and h is a (strict) first-order
approximation to f at x̄, then q is a (strict) first-order approximation to f at x̄.
(ii) If f1 and f2 have (strict) first-order approximations h1 and h2 , respectively,
at x̄, then h1 + h2 is a (strict) first-order approximation of f1 + f2 at x̄.
(iii) If f has a (strict) first-order approximation h at x̄, then for any λ ∈ IR, λ h is
a (strict) first-order approximation of λ f at x̄.
(iv) If h is a first-order approximation of f at x̄, then clm ( f ; x̄) = clm (h; x̄).
Similarly, if h is a strict first-order approximation of f at x̄, then lip ( f ; x̄) = lip (h; x̄).
1 Functions Defined Implicitly by Equations 37

The next proposition explains how first-order approximations can be chained

together.

Proposition 1E.1 (composition of first-order approximations). Let h be a first-order

approximation of f at x̄ which is calm at x̄. Let v be a first-order approximation of u
at ȳ for ȳ := f (x̄) which is Lipschitz continuous around ȳ. Then v◦h is a first-order
approximation of u◦ f at x̄.
Proof. By the property (iv) of the first-order approximations displayed before the
statement, the function f is calm at x̄. Choose ε > 0 and let μ and λ be such that
clm ( f ; x̄) < μ and lip (v; x̄) < λ . Then there exist neighborhoods U of x̄ and V of ȳ
such that f (x) ∈ V and h(x) ∈ V for x ∈ U,

| f (x) − h(x)| ≤ ε |x − x̄| and | f (x) − f (x̄)| ≤ μ |x − x̄| for x ∈ U,

and moreover,

|u(y) − v(y)| ≤ ε |y − ȳ| and |v(y) − v(y)| ≤ λ |y − y| for y, y ∈ V.

Then for x ∈ U, taking into account that ȳ = f (x̄) = h(x̄), we have

|(u◦ f )(x) − (v◦h)(x)| = |u( f (x)) − v(h(x))|

≤ |u( f (x)) − v( f (x))| + |v( f (x)) − v(h(x))|
≤ ε | f (x) − f (x̄)| + λ | f (x) − h(x)|
≤ ε μ |x − x̄| + λ ε |x − x̄| ≤ ε (μ + λ )|x − x̄|.

Since ε can be arbitrarily small, the proof is complete.

Exercise 1E.2 (strict approximations through composition). Let the function f sat-
isfy lip ( f ; x̄) < ∞ and let the function g have a strict first-order approximation q at
ȳ, where ȳ := f (x̄). Then q◦ f is a strict first-order approximation of g◦ f at x̄.
Guide. Mimic the proof of 1E.1.
For our purposes here, and in later chapters as well, first-order approximations
offer an appealing substitute for differentiability, but an even looser notion of ap-
proximation will still lead to important conclusions.

Estimators beyond first-order approximations. Consider a function f : IRn → IRm

and a point x̄ ∈ int dom f . A function h : IRn → IRm with x̄ ∈ int dom h is an estimator
of f at x̄ with constant μ if h(x̄) = f (x̄) and

clm (e; x̄) ≤ μ < ∞ for e(x) = f (x) − h(x),

which can also be written as | f (x) − h(x)| ≤ μ |x − x̄| + o(|x − x̄|). It is a strict esti-
mator if the stronger condition holds that

lip (e; x̄) ≤ μ < ∞ for e(x) = f (x) − h(x).

38 1 Functions Defined Implicitly by Equations

In this terminology, a first-order approximation is simply an estimator with con-

stant μ = 0. Through that, any result involving estimators can immediately be spe-
cialized to a result about first-order approximations.
Estimators can be of interest even when differentiability is present. For instance,
in the case of a function f that is strictly differentiable at x̄ a strict estimator of f
at x̄ with constant μ is furnished by h(x) = f (x̄) + A(x − x̄) for any matrix A with
|∇ f (x̄) − A| ≤ μ . Such relations have a role in certain numerical procedures, as will
be seen at the end of this section and later in the book.

Theorem 1E.3 (inverse function theorem beyond differentiability). Let f : IRn →

IRn be a function with x̄ ∈ int dom f , and let h : IRn → IRn be a strict estimator of f
at x̄ with constant μ. At the point ȳ = f (x̄) = h(x̄), suppose that h−1 has a Lipschitz
continuous single-valued localization σ around ȳ for x̄ with lip (σ ; ȳ) ≤ κ for a
constant κ such that κ μ < 1. Then f −1 has a Lipschitz continuous single-valued
localization s around ȳ for x̄ with
κ
lip (s; ȳ) ≤ .
1 − κμ

This result is a particular case of the implicit function theorem 1E.13 presented
later in this section, which in turn follows from a more general result proven in
Chapter 2. The reader who does not want to wait for a proof until the next chapter
is encouraged to do the following exercise which is supplied with a detailed guide.

Exercise 1E.4. Prove Theorem 1E.3.

Guide. Below we outline a direct proof by following the steps in Proof I of The-
orem 1A.1. One may also prove this theorem in parallel to Proof II of 1A.1, by
showing that Φy (x) = r y − ( f − h)(x) has a fixed point.
First, fix any

λ ∈ (κ , ∞) and ν ∈ (μ , κ −1 ) with λ ν < 1.

Without loss of generality, suppose that x̄ = 0 and ȳ = 0 and take a small enough
that the mapping
y → h−1 (y) ∩ aIB for y ∈ aIB
is a localization of σ that is Lipschitz continuous with constant λ and also the dif-
ference e = f − h is Lipschitz continuous on aIB with constant ν . Next we choose α
satisfying
1
0 < α < a(1 − λ ν ) min{1, λ }
4
and let b := α /(4λ ). Pick any y ∈ bIB and any x0 ∈ (α /4)IB; this gives us
α α α
|y − e(x0 )| ≤ |y| + |e(x0 ) − e(0)| ≤ |y| + ν |x0 | ≤ +ν ≤ .
4λ 4 2λ
1 Functions Defined Implicitly by Equations 39

In particular |y − e(x0 )| < a, so the point y − e(x0 ) lies in the region where σ is
Lipschitz continuous. Let x1 = σ (y − e(x0 )); then
α
|x1 | = |σ (y − e(x0 ))| = |σ (y − e(x0)) − σ (0)| ≤ λ |y − e(x0 )| ≤ ,
2
so in particular x1 belongs to the ball aIB. Furthermore,

|x1 − x0 | ≤ |x1 | + |x0 | ≤ α /2 + α /4 = 3α /4.

We also have

|y − e(x1)| ≤ |y| + ν |x1 | ≤ α /(4λ ) + α /(2λ ) ≤ a,

so that y − e(x1 ) ∈ aIB.

Having started in this pattern, proceed to construct an infinite sequence of points
xk by taking
xk+1 = σ (y − e(xk ))
and prove by induction that

xk ∈ aIB, y − e(xk ) ∈ aIB and |xk − xk−1 | ≤ (λ ν )k−1 |x1 − x0 | for k = 2, 3, . . . .

Observe next that, for k > j,

k−1 k−1
a
|xk − x j | ≤ ∑ |xi+1 − xi| ≤ ∑ (λ ν )i a ≤ 1 − λ ν (λ ν ) j ,
i= j i= j

hence
lim |xk − x j | = 0.
j,k→∞
k> j

Then the sequence {xk } is Cauchy and hence convergent. Let x be its limit. Since
all xk and all y − e(xk ) are in aIB, where both e and σ are continuous, we can pass to
the limit in the equation xk+1 = σ (y − e(xk )) as k → ∞, getting

x = σ (y − e(x)), that is, x ∈ f −1 (y).

According to our construction, we have

|x| ≤ λ (|y| + |e(x) − e(0)|) ≤ λ |y| + λ ν |x|,

so that, since |y| ≤ b, we obtain

λb
|x| ≤ .
1−λν
40 1 Functions Defined Implicitly by Equations

Thus, it is established that for every y ∈ bIB there exists x ∈ f −1 (y) with |x| ≤
λ b/(1 − λ ν ). In other words, we have shown the nonempty-valuedness of the lo-
calization of f −1 given by

λb
s : y → f −1 (y) ∩ IB for y ∈ bIB.
1−λν
Next, demonstrate that this localization s is in fact single-valued and Lipschitz con-
tinuous. If for some y ∈ bIB we have two points x = x , both of them in s(y), then
subtracting x = σ (y − e(x)) from x = σ (y − e(x )) gives

0 < |x − x| = |σ (y − e(x )) − σ (y − e(x))| ≤ λ |e(x ) − e(x)| ≤ λ ν |x − x| < |x − x|,

which is absurd. Further, considering y , y ∈ bIB and recalling that s(y) = σ (y −

e(s(y))), one gets

|s(y ) − s(y)| = |σ (y − e(s(y ))) − σ (y − e(s(y)))| ≤ λ (|y − y| + ν |s(y ) − s(y)|),

and hence s is Lipschitz continuous relative to bIB with constant λ /(1 − λ ν ). This
expression is continuous and increasing as a function of λ and ν , which are greater
than κ and μ but can be chosen arbitrarily close to them, hence the Lipschitz mod-
ulus of s at ȳ satisfies the desired inequality.

When μ = 0 in Theorem 1E.3, so that h is a strict first-order approximation of

f at x̄, the conclusion about the localization s of f −1 is that lip (s; ȳ) ≤ κ . The
strict derivative version 1D.9 of the inverse function theorem corresponds to the
case where h(x) = f (x̄) + D f (x̄)(x − x̄). The assumption on h−1 in Theorem 1E.3
is tantamount then to the invertibility of D f (x̄), or equivalently the nonsingularity
of the Jacobian ∇ f (x̄); we have lip (σ ; ȳ) = |D f (x̄)−1 | = |∇ f (x̄)−1 |, and κ can be
taken to have this value. Again, though, Theorem 1E.3 does not, in general, insist
on h being a first-order approximation of f at x̄.
The following example sheds light on the sharpness of the assumptions in 1E.3
about the relative sizes of the Lipschitz modulus of the localization of h−1 and the
Lipschitz modulus of the “approximation error” f − h.
Example 1E.5 (illustration of invertibility without strict differentiability). With α ∈
(0, ∞) as a parameter, let f (x) = α x + g(x) for the function

g(x) = x sin(1/x) for x = 0,
2

0 for x = 0,

noting that f and g are differentiable with


2x sin(1/x) − cos(1/x) for x = 0,
g (x) =
0 for x = 0,
1 Functions Defined Implicitly by Equations 41

but f and g are not strictly differentiable at 0, although g is Lipschitz continuous

there with lip (g; 0) = 1.
Let h(x) = α x and consider applying Theorem 1E.3 to f and h at x̄ = 0, where
f (0) = h(0) = 0. Since f − h = g, we have for every μ > 1 that f − h is Lipschitz
continuous with constant μ on some neighborhood of 0. On the other hand, h−1
is Lipschitz continuous with constant κ = 1/α . Therefore, as long as α > 1, the
assumptions of Theorem 1E.3 are fulfilled (by taking μ in (1, α ) arbitrarily close
to 1). We are able to conclude from 1E.3 that f −1 has a single-valued localization s
around 0 for 0 such that lip (s; 0) ≤ (α − 1)−1 , despite the inapplicability of 1D.9.
When 0 < α < 1, however, f −1 has no single-valued localization around the
origin at all. This comes out of the fact that, for such α , the derivative f  (x) has
infinitely many changes of sign in any neighborhood of x̄ = 0, hence infinitely many
consecutive local maximum values and minimum values of f in any such neigh-
borhood, with both values tending to 0 as the origin is approached. Let x1 and x2 ,
0 < x1 < x2 , be two consecutive points where f  vanishes and f has a local maxi-
mum at x1 and local minimum at x2 . For a value y > 0 the equation f (x) = y must
have not only a solution in (x1 , x2 ), but also one in (0, x1 ). Hence, regardless of the
size of the neighborhood U of x̄ = 0, there will be infinitely many y values near
ȳ = 0 for which U ∩ f −1 (y) is not a singleton. Both cases are illustrated in Fig. 1.6.

Fig. 1.6 Graphs of the function f in Example 1E.5 when α = 2 on the left and α = 0.5 on the
right.

When the approximation h to f in Theorem 1E.3 is itself strictly differentiable at

the point in question, a simpler statement of the result can be made.

Corollary 1E.6 (estimators with strict differentiability). For f : IRn → IRn with
x̄ ∈ int dom f and f (x̄) = ȳ, suppose there is a strict estimator h : IRn → IRn of f at x̄
with constant μ which is strictly differentiable at x̄ with nonsingular Jacobian ∇h(x̄)
satisfying μ |∇h(x̄)−1 | < 1. Then f −1 has a Lipschitz continuous single-valued lo-
calization s around ȳ for x̄ with

|∇h(x̄)−1 |
lip (s; ȳ) ≤ .
1 − μ |∇h(x̄)−1 |
42 1 Functions Defined Implicitly by Equations

Proof. A localization σ of h−1 having lip (σ ; ȳ) = |∇h(x̄)−1 | is obtained by apply-

ing Theorem 1D.9 to h. Theorem 1E.3 can be invoked then with κ = |∇h(x̄)−1 |.

An even more special application, to the case where both f and h are linear,
yields a well-known estimate for matrices.

Corollary 1E.7 (estimation for perturbed matrix inversion). Let A and B be n × n

matrices such that A is nonsingular and |B| < |A−1 |−1 . Then A + B is nonsingular
with  −1
|(A + B)−1| ≤ |A−1 |−1 − |B| .
Proof. This comes from Corollary 1E.6 by taking f (x) = (A+B)x, h(x) = Ax, x̄ = 0,
 −1
and writing |A−1 |−1 − |B| as |A|/(1 − |A||B|).
We can state 1E.7 in other two equivalent ways, which we give next as an exer-
cise. More about the estimate in 1E.7 will be said in Chapter 5.

Exercise 1E.8 (equivalent estimation rules for matrices). Prove that the following
two statements are equivalent to Corollary 1E.7:
(a) For any n × n matrix C with |C| < 1, the matrix I + C is nonsingular and

1
(1) |(I + C)−1 | ≤ .
1 − |C|

(b) For any n × n matrices A and B with A nonsingular and |BA−1 | < 1, the matrix
A + B is nonsingular and

|A−1 |
|(A + B)−1 | ≤ .
1 − |BA−1|

Guide. Clearly, (b) implies Corollary 1E.7 which in turn implies (a) with A = I and
B = C. Let (a) hold and let the matrices A and B satisfy |BA−1 | < 1. Then, by (a) with
C = BA−1 we obtain that I + BA−1 is nonsingular, and hence A + B is nonsingular,
too. Using the equality A + B = (I + BA−1)A in (1) we have

|A−1 |
|(A + B)−1 | = |A−1 (I + BA−1 )−1 | ≤ ,
1 − |BA−1|

that is, (a) implies (b).

Corollary 1E.7 implies that, given a nonsingular matrix A,
  
(2) inf |B|  A + B is singular ≥ |A−1 |−1 .

It turns out that this inequality is actually equality, another classical result in matrix
perturbation theory.
1 Functions Defined Implicitly by Equations 43

Theorem 1E.9 (radius theorem for matrix nonsingularity). For any nonsingular ma-
trix A,   
inf |B|  A + B is singular = |A−1 |−1 .
Proof. It is sufficient to prove the inequality opposite to (2). Choose ȳ ∈ IRn with
|ȳ| = 1 and |A−1 ȳ| = |A−1 |. For x̄ = A−1 ȳ we have |x̄| = |A−1 |. The matrix

ȳ x̄T
B=−
|x̄|2

satisfies
|ȳ x̄T x| |x̄T x| |x̄T x̄| 1
|B| = max ≤ max = = = |A−1 |−1 .
|x|=1 |x̄| 2 |x|=1 |x̄| 2 |x̄| 3 |x̄|

On the other hand (A + B)x̄ = Ax̄ − ȳ = 0, and since x̄ = 0, the matrix A + B is

singular. Thus the infimum in (2) is not greater than |A−1 |−1 .

Exercise 1E.10 (radius theorem for function invertibility). Consider a function f :

IRn → IRn and a point x̄ ∈ int dom f . Call f smoothly locally invertible at x̄, for short,
when f −1 has a Lipschitz continuous single-valued localization around f (x̄) for x̄
which is strictly differentiable at f (x̄). In this terminology, prove in the case of f
being strictly differentiable at x̄ with its Jacobian ∇ f (x̄) nonsingular, that
  
inf |B|  f + B is not smoothly locally invertible at x̄ = |A−1 |−1 ,

where A = ∇ f (x̄) and the infimum is taken over all linear mappings B : IRn → IRn .
Guide. Combine 1E.9 and 1D.9.
We will extend the facts in 1E.9 and 1E.10 to the much more general context of
set-valued mappings in Chapter 6.
In the case of Theorem 1E.3 with μ = 0, an actual equivalence emerges between
the invertibility of f and that of h, as captured by the following statement. The key
is the fact that first-order approximation is a symmetric relationship between two
functions.

Theorem 1E.11 (Lipschitz invertibility with first-order approximations). Let h :

IRn → IRn be a strict first-order approximation to f : IRn → IRn at x̄ ∈ int dom f , and
let ȳ denote the common value f (x̄) = h(x̄). Then f −1 has a Lipschitz continuous
single-valued localization s around ȳ for x̄ if and only if h−1 has such a localization
σ around ȳ for x̄, in which case

(3) lip (s; ȳ) = lip (σ ; ȳ),

and moreover σ is then a first-order approximation to s at ȳ: s(y) = σ (y)+ o(|y− ȳ|).
Proof. Applying Theorem 1E.3 with μ = 0 and κ = lip (σ ; ȳ), we see that f −1 has a
single-valued localization s around ȳ for x̄ with lip (s; ȳ) ≤ lip (σ ; ȳ). In these circum-
44 1 Functions Defined Implicitly by Equations

stances, though, the symmetry in the relation of first-order approximation allows us

to conclude from the existence of this s that h−1 has a single-valued localization σ 
around ȳ for x̄ with lip (σ  ; ȳ) ≤ lip (s; ȳ). The two localizations of h have to agree
graphically in a neighborhood of (ȳ, x̄), so we can simply speak of σ and conclude
the validity of (3).
To argue that σ is a first-order approximation of s, we begin by using the iden-
tity h(σ (y)) = y to get f (σ (y)) = y + e(σ (y)) for the function e = f − h and then
transform that into

(4) σ (y) = s(y + e(σ (y))) for y near ȳ.

Let κ > lip (s; ȳ). From (3) there exists b > 0 such that
 
(5) max |σ (y) − σ (y )|, |s(y) − s(y )| ≤ κ |y − y| for y, y ∈ IBb (ȳ).

Because e(x̄) = 0 and lip (e; x̄) = 0, we know that for any ε > 0 there exists a positive
a > 0 for which

(6) |e(x) − e(x )| ≤ ε |x − x | for all x, x ∈ IBa (x̄).

Choose a 
b
0 < β ≤ min , .
κ (1 + εκ )
Then, for every y ∈ IBβ (ȳ) from (5) we have

|σ (y) − x̄| ≤ κβ ≤ a,

and
|y + e(σ (y)) − ȳ| ≤ |y − ȳ| + ε |σ (y) − x̄| ≤ β + εκβ ≤ b.
Hence, utilizing (4), (5) and (6), we obtain

|σ (y) − s(y)| = |s(y + e(σ (y))) − s(y)|

≤ κ |e(σ (y)) − e(σ (ȳ))|
≤ κε |σ (y) − σ (ȳ)| ≤ κ 2 ε |y − ȳ|.

Since ε can be arbitrarily small, we arrive at the equality clm (s − σ ; ȳ) = 0, and this
completes the proof.
Finally, we observe that these results make it possible to deduce a slightly sharper
version of the equivalence in Theorem 1D.9.

Theorem 1E.12 (extended equivalence under strict differentiability). Let f : IRn →

IRn be strictly differentiable at x̄ with f (x̄) = ȳ. Then the following are equivalent:
(i) ∇ f (x̄) is nonsingular;
(ii) f −1 has a Lipschitz continuous single-valued localization around ȳ for x̄;
1 Functions Defined Implicitly by Equations 45

(iii) f −1 has a single-valued localization around ȳ for x̄ that is strictly differen-

tiable at ȳ.
In parallel to Theorem 1E.3, it is possible also to state and prove a correspond-
ing implicit function theorem with Lipschitz continuity. For that purpose, we need
to introduce the concept of partial first-order approximations for functions of two
variables.

Partial first-order estimators and approximations. For f : IRd × IRn → IRm and a
point ( p̄, x̄) ∈ int dom f , a function h : IRn → IRm is said to be an estimator of f with
respect to x uniformly in p at ( p̄, x̄) with constant μ if h(x̄) = f (x̄, p̄) and

 x (e; ( p̄, x̄)) ≤ μ < ∞ for e(p, x) = f (p, x) − h(x).

clm

It is a strict estimator in this sense if the stronger condition holds that

 (e; ( p̄, x̄)) ≤ μ < ∞ for e(p, x) = f (p, x) − h(x).

lip x

In the case of μ = 0, such an estimator is called a partial first-order approximation.

Theorem 1E.13 (implicit function theorem beyond differentiability). Consider f :

 ( f ; ( p̄, x̄)) ≤ γ < ∞.
IRd × IRn → IRn and ( p̄, x̄) ∈ int dom f with f ( p̄, x̄) = 0 and lip p
Let h be a strict estimator of f with respect to x uniformly in p at ( p̄, x̄) with constant
μ. Suppose at the point h(x̄) = 0 that h−1 has a Lipschitz continuous single-valued
localization σ around 0 for x̄ with lip (σ ; 0) ≤ κ for a constant κ such that κ μ < 1.
Then the solution mapping
  
S : p → x ∈ IRn  f (p, x) = 0 for p ∈ IRd

has a Lipschitz continuous single-valued localization s around p̄ for x̄ with

κγ
lip (s; p̄) ≤ .
1 − κμ

Moreover, when μ = 0 the function η (p) = σ (− f (p, x̄)) is a first-order approxima-

tion to s at p̄: s(p) = η (p) + o(|p − p̄|).
This is a special instance of the combination of Theorem 2B.7 (for μ > 0) and of
Theorem 2B.8 (for μ = 0) which we will prove in Chapter 2, so there is no purpose
in giving a separate argument for it here. If f (p, x) has the form f (x) − p, in which
 ( f ; p̄, x̄) = 1, taking γ = 1 we immediately obtain Theorem 1E.3.
case lip p

Exercise 1E.14 (approximation criteria). Consider f : IRd × IRn → IRm and h : IRd ×
IRn → IRm with f ( p̄, x̄) = h( p̄, x̄), and the difference e(p, x) = f (p, x)− h(p, x). Prove
that
 x (e; ( p̄, x̄)) = 0 and clm
(a) If clm  p (e; ( p̄, x̄)) = 0, then h is a first-order approx-
imation to f at ( p̄, x̄).
46 1 Functions Defined Implicitly by Equations

 (e; ( p̄, x̄)) = 0 and clm p (e; ( p̄, x̄)) = 0, then h is a first-order approxi-
(b) If lip x
mation to f at ( p̄, x̄).
 (e; ( p̄, x̄)) = 0 and lip
(c) If lip  (e; ( p̄, x̄)) = 0, then h is a strict first-order ap-
x p
proximation to f at ( p̄, x̄).

Exercise 1E.15 (partial first-order approximation from differentiability). Let f :

IRd × IRn → IRm be differentiable with respect to x in a neighborhood of ( p̄, x̄) and
let f and ∇x f be continuous in this neighborhood. Prove that the function h(x) =
f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄) is a strict first-order approximation to f with respect to
x uniformly in p at ( p̄, x̄). Based on this, derive the Dini classical implicit function
theorem 1B.1 from 1E.13.

Exercise 1E.16 (the zero function as an approximation). Let f : IRd × IRn → IRm
 ( f ; ( p̄, x̄)) < ∞, and let u : IRd × IRm → IRk have a strict first-order ap-
satisfy lip x
proximation v with respect to y at ( p̄, ȳ), where ȳ := f ( p̄, x̄). Show that the zero
function is a strict first-order approximation with respect to x at ( p̄, x̄) to the func-
tion (p, x) → u(p, f (p, x)) − v( f (p, x)).
As another illustration of applicability of Theorem 1E.3 beyond first-order ap-
proximations, we sketch now a proof of the quadratic convergence of Newton’s
method for solving equations, a method we used in Proof I of the classical inverse
function theorem 1A.1.
Consider the equation g(x) = 0 for a continuously differentiable function g :
IRn → IRn with a solution x̄ at which the Jacobian ∇g(x̄) is nonsingular. Newton’s
method consists in choosing a starting point x0 possibly close to x̄ and generating a
sequence of points x1 , x2 , . . . according to the rule

(7) xk+1 = xk − ∇g(xk )−1 g(xk ), k = 0, 1, . . . .

According to the classical inverse function theorem 1A.1, g−1 has a smooth single-
valued localization around 0 for x̄. Consider the function f (x) = ∇g(x0 )(x − x̄) for
which f (x̄) = g(x̄) = 0 and f (x) − g(x) = −g(x) + ∇g(x0 )(x − x̄). An easy calcula-
tion shows that the Lipschitz modulus of e = f − g at x̄ can be made arbitrarily small
by making x0 close to x̄. However, this modulus must be nonzero — but less than
|∇g(x̄)−1 |−1 , the Lipschitz modulus of the single-valued localization of g−1 around
0 for x̄, if one wants to choose x0 as an arbitrary starting point from an open neigh-
borhood of x̄. Here Theorem 1E.3 comes into play with h = g and ȳ = 0, saying
that f −1 has a Lipschitz continuous single-valued localization s around 0 for x̄ with
Lipschitz constant, say, μ . (In the simple case considered this also follows directly
from the fact that if ∇g(x̄) is nonsingular at x̄, then ∇g(x) is likewise nonsingular for
all x in a neighborhood of x̄, see Fact 2 in Section 1A.) Hence, the Lipschitz constant
μ and the neighborhood V of 0 where s is Lipschitz continuous can be determined
before the choice of x0 , which is to be selected so that ∇g(x0 )(x0 − x̄) − g(x0 ) is in
V . Noting for the iteration (7) that x1 = s(∇g(x0 )(x0 − x̄) − g(x0 )) and x̄ = s(g(x̄)),
and using the smoothness of g, we obtain
1 Functions Defined Implicitly by Equations 47

|x1 − x̄| ≤ μ |g(x̄) − g(x0) − ∇g(x0 )(x̄ − x0 )| ≤ c|x0 − x̄|2

for a suitable constant c. This kind of argument works for any k, and in that way,
through induction, we obtain quadratic convergence for Newton’s iteration (7).
In Chapter 6 we will present, in a broader framework of generalized equations in
possibly infinite-dimensional spaces, a detailed proof of this quadratic convergence
of Newton’s method and study its stability properties.

1F. Selections of Multi-valued Inverses

Consider a function f which acts between Euclidean spaces of possibly different
dimensions, say f : IRn → IRm . What can be said then about the inverse mapping
f −1 ? The case of f (x) = Ax + b with A ∈ IRm×n and b ∈ IRm gives an indication of
the differences that must be expected: when m < n, the equation Ax + b = y either
has no solution x or a continuum of solutions, so that the existence of single-valued
localizations is totally hopeless. Anyway, though, if A has full rank m, we know at
least that f −1 (y) will be nonempty for every y.
Do we really always have to assume that m = n in order to get a single-valued
localization of the inverse? Specifically, consider a function f acting from IRn to
IRm with m ≤ n and a point x̄ in the interior of the domain of f . Suppose that f is
continuous in an open neighborhood U of x̄ and the inverse f −1 has a single-valued
localization around f (x̄) for x̄. It turns out that in this case we necessarily must have
m = n. This is due to a basic result in topology, the following version of which,
stated here without proof, will serve our purposes.

Theorem 1F.1 (Brouwer invariance of domain theorem). Let O ⊂ IRn be open and
for m ≤ n let f : O → IRm be continuous and such that f −1 is single-valued on f (O).
Then f (O) is open, f −1 is continuous on f (O), and m = n.
This topological result reveals that, for continuous functions f , the dimension of
the domain space has to agree with the dimension of the range space, if there is to be
any hope of an inverse function theorem claiming the existence of a single-valued
localization of f −1 . Of course, in the theorems already viewed for differentiable
functions f , the dimensions were forced to agree because of a rank condition on
the Jacobian matrix, but we see now that this limitation has a deeper source than a
matrix condition.
Brouwer’s invariance of domain theorem helps us to obtain the following char-
acterization of the existence of a Lipschitz continuous single-valued localization of
the inverse:

Theorem 1F.2 (invertibility characterization). For a function f : IRn → IRn that is

continuous around x̄, the inverse f −1 has a Lipschitz continuous single-valued lo-
calization around f (x̄) for x̄ if and only if, in some neighborhood U of x̄, there is a
48 1 Functions Defined Implicitly by Equations

constant c > 0 such that

(1) c|x − x| ≤ | f (x ) − f (x)| for all x , x ∈ U.

Proof. Let (1) hold. There is no loss of generality  in supposing

 that U is open and
f is continuous on U. For any y ∈ f (U) := f (x)  x ∈ U , we have from (1) that
if both f (x) = y and f (x ) = y with x and x in U, then x = x ; in other words, the
mapping s which takes y ∈ f (U) to U ∩ f −1 (y) is actually a function on f (U). More-
over |s(y ) − s(y)| ≤ (1/c)|y − y| by (1), so that this function is Lipschitz continuous
relative to f (U).
But this is not yet enough to get us to the desired conclusion about f −1 . For that,
we need to know that s, or some restriction of s, is a localization of f −1 around
f (x̄) for x̄, with graph equal to (V × U) ∩ gph f −1 for some neighborhood V of
f (x̄). Brouwer’s invariance of domain theorem 1F.1 enters here: as applied to the
restriction of f to the open set U, it tells us that f (U) is open. We can therefore take
V = f (U) and be done.
Conversely, suppose that f −1 has a Lipschitz continuous single-valued localiza-
tion around f (x̄) for x̄, its domain being a neighborhood V of f (x̄). Let κ be a
Lipschitz constant for s on V . Because f is continuous around x̄, there is a neigh-
borhood U of x̄ such that f (U) ⊂ V . For x and x in U, we have s( f (x)) = x and
s( f (x )) = x , hence |x − x| ≤ κ | f (x ) − f (x)|. Thus, (1) holds for any c > 0 small
enough such that cκ ≤ 1.
We will now re-prove the classical inverse function theorem in a somewhat dif-
ferent formulation having an important extra feature, which is here derived from
Brouwer’s invariance of domain theorem 1F.1.

Theorem 1F.3 (inverse function theorem for local diffeomorphism). Let f : IRn →
IRn be continuously differentiable in a neighborhood of a point x̄ and let the Jacobian
∇ f (x̄) be nonsingular. Then for some open neighborhood U of x̄ there exists an open
neighborhood V of ȳ := f (x̄) and a continuously differentiable function s : V → U
which is one-to-one from V onto U and which satisfies s(y) = f −1 (y) ∩ U for all
y ∈ V . Moreover, the Jacobian of s is given by

∇s(y) = ∇ f (s(y))−1 for every y ∈ V.

Proof. First, we utilize a simple observation from linear algebra: for a nonsingular
n × n matrix A, one has |Ax| ≥ |x|/|A−1| for every x ∈ IRn . Thus, let c > 0 be such
that |∇ f (x̄)u| ≥ 2c|u| for every u ∈ IRn and choose a > 0 to have, on the basis of (b)
in Fact 1 of Section 1A, that

| f (x ) − f (x) − ∇ f (x̄)(x − x)| ≤ c|x − x| for every x , x ∈ IBa (x̄).

Using the triangle inequality, for any x , x ∈ IBa (x̄) we then have

| f (x ) − f (x)| ≥ |∇ f (x̄)(x − x)| − c|x − x| ≥ 2c|x − x| − c|x − x| ≥ c|x − x|.

1 Functions Defined Implicitly by Equations 49

We can therefore apply 1F.2, obtaining that there is an open neighborhood U of x̄

relative to which f is continuous and f −1 is single-valued on V := f (U). Brouwer’s
theorem 1F.1 then tells us that V is open. Then the mapping s : V → U whose graph
is gph s = gph f −1 ∩ (V × U) is the claimed single-valued localization of f −1 and
the rest is argued through Step 3 in Proof I of 1A.1.
A continuously differentiable function f acting between some open sets U and
V in IRn and having the property that the inverse mapping f −1 is continuously dif-
ferentiable, is called a diffeomorphism (or C 1 diffeomorphism) between U and V .
Theorem 1F.3 simply says that when ∇ f (x̄) is nonsingular, then f is a diffeomor-
phism6 relative to open neighborhoods U of x̄ and V of f (x̄).

Exercise 1F.4 (implicit function version). Let f : IRd × IRn → IRn be continuously
differentiable in a neighborhood of ( p̄, x̄) and such that f ( p̄, x̄) = 0, and let ∇x f ( p̄, x̄)
be nonsingular. Then for some open neighborhood U of x̄ there exists an open neigh-
borhood Q of p̄ and
  a continuously
 differentiable
 function s : Q → U such that
(p, s(p))  p ∈ Q = (p, x)  f (p, x) = 0 ∩ (Q
  × U) ; that is, s is a single-valued
localization of the solution mapping S(p) = x  f (p, x) = 0 with associated open
neighborhoods Q for p̄ and U for x̄. Moreover, the Jacobian of s is given by

∇s(p) = −∇x f (p, s(p))−1 ∇ p f (p, s(p)) for every p ∈ Q.

Guide. Apply 1F.3 in the same way as 1A.1 is used in the proof of 1B.1.

Exercise 1F.5. Derive Theorem 1D.9 from 1F.2.

Guide. By 1D.7, strict differentiability is equivalent to the assumption in Fact 1 of
Section 1A; then repeat the argument in the proof of 1F.3.
Brouwer’s invariance of domain theorem tells us that, for a function f : IRn → IRm
with m < n, the inverse f −1 fails to have a localization which is single-valued. In
this case, however, although multi-valued, f −1 may “contain” a function with the
properties of the single-valued localization for the case m = n. Such functions are
generally called selections and their formal definition is as follows.

Selections. Given a set-valued mapping F : IRn → → IRm and a set D ⊂ dom F , a func-
tion w : IR → IR is said to be a selection of F on D if dom w ⊃ D and w(x) ∈ F(x)
n m

for all x ∈ D. If (x̄, ȳ) ∈ gph F , D is a neighborhood of x̄ and w is a selection on D

which satisfies w(x̄) = ȳ, then w is said to be a local selection of F around x̄ for ȳ.
A selection of the inverse f −1 of a function f : IRn → IRm might provide a left
inverse or a right inverse to f . A left inverse to f on D is a selection l : IRm → IRn of
f −1 on f (D) such that l( f (x)) = x for all x ∈ D. Analogously, a right inverse to f on
D is a selection r : IRm → IRn of f −1 on f (D) such that f (r(y)) = y for all y ∈ f (D).
Commonly known are the right and the left inverses of the linear mapping from
IRn to IRm represented by a matrix A ∈ IRm×n that is of full rank. When m ≤ n, the
6 Theorem 1F.3 can of course be proved directly, without resorting to Brouwer’s theorem 1F.1.
50 1 Functions Defined Implicitly by Equations

right inverse of the mapping corresponds to AT (AAT )−1 , while when m ≥ n, the left
inverse7 corresponds to (AT A)−1 AT . For m = n they coincide and equal the inverse.
In general, of course, whenever a mapping f is one-to-one from a set C to f (C), any
left inverse to f on C is also a right inverse, and vice versa, and the restriction of
such an inverse to f (C) is uniquely determined.
The following result can be viewed as an extension of the classical inverse func-
tion theorem 1A.1 for selections.

Theorem 1F.6 (inverse selections when m ≤ n). Let f : IRn → IRm , where m ≤ n,
be k times continuously differentiable in a neighborhood of x̄ and suppose that its
Jacobian ∇ f (x̄) is full rank m. Then, for ȳ = f (x̄), there exists a local selection s of
f −1 around ȳ for x̄ which is k times continuously differentiable in a neighborhood
V of ȳ and whose Jacobian satisfies

(2) ∇s(ȳ) = AT (AAT )−1 , where A := ∇ f (x̄).

Proof. There are various ways to prove this; here we apply the classical inverse
function theorem. Since A has rank m, the m × m matrix AAT is nonsingular. Then
the function ϕ : IRm → IRm defined by ϕ (u) = f (AT u) is k times continuously differ-
entiable in a neighborhood of the point ū := (AT A)−1 Ax̄, its Jacobian ∇ϕ (ū) = AAT
is nonsingular, and ϕ (ū) = ȳ. Then, from Theorem 1A.1 supplemented by Proposi-
tion 1B.5, it follows that ϕ −1 has a single-valued localization σ at ȳ for ū which is k
times continuously differentiable near ȳ with Jacobian ∇σ (ȳ) = (AAT )−1 . But then
the function s(y) = AT σ (y) satisfies s(ȳ) = x̄ and f (s(y)) = y for all y near ȳ and
is k times continuously differentiable near ȳ with Jacobian satisfying (2). Thus, s(y)
is a solution of the equation f (x) = y for y close to ȳ and x close to x̄, but perhaps
not the only solution there, as it would be in the classical inverse function theorem.
Therefore, s is a local selection of f −1 around ȳ for x̄ with the desired properties.
When m = n the Jacobian becomes nonsingular and the right inverse of A in (2)
is just A−1 . The uniqueness of the localization can be obtained much as in Step 2 of
Proof I of the classical theorem 1A.1.
  
Exercise 1F.7 (parameterization of solution sets). Let M = x  f (x) = 0 for a
function f : IRn → IRm , where n − m = d > 0. Let x̄ ∈ M be a point around which f
is k times continuously differentiable, and suppose that the Jacobian ∇ f (x̄) has full
rank m. Then for some open neighborhood U of x̄ there is an open neighborhood
O of the origin in IRd and a k times continuously differentiable function s : O → U
which is one-to-one from O onto M ∩U , such that the Jacobian ∇s(0) has full rank
d and

∇ f (x̄)w = 0 if and only if there exists q ∈ IRd with ∇s(0)q = w.

7 The left inverse and the right inverse are particular cases of the Moore-Penrose pseudo-inverse

A+ of a matrix A. For more on this, including the singular-value decomposition, see Golub and
Van Loan [1996].
1 Functions Defined Implicitly by Equations 51

Guide. Choose an d × n matrix B such that the matrix

 
∇ f (x̄)
B

is nonsingular. Consider the function

 
¯f : (p, x) → f (x)
for (p, x) near (0, x̄),
B(x − x̄) − p

and apply 1F.6 (with a modification parallel to Proposition 1B.5) to the equation
f¯(p, x) = (0, 0), obtaining for the solution mapping of this equation a localization
s with f¯(p, s(p)) = (0, 0), i.e., Bs(p) = p + Bx̄ and f (s(p)) = 0. Show that this
function s has the properties claimed.

Exercise 1F.8 (strictly differentiable selections). Let f : IRn → IRm , where m ≤ n,

be strictly differentiable at x̄ with Jacobian A := ∇ f (x̄) of full rank. Then there
exists a local selection s of the inverse f −1 around ȳ := f (x̄) for x̄ which is strictly
differentiable at ȳ and with Jacobian ∇s(ȳ) satisfying (2).
Guide. Mimic the proof of 1F.6 taking into account 1D.9.

Exercise 1F.9 (implicit selections). Consider a function f : IRd × IRn → IRm , where
m ≤ n, along with the associated solution mapping
  
S : p → x ∈ IRn  f (p, x) = 0 for p ∈ IRd .

Let f ( p̄, x̄) = 0, so that x̄ ∈ S( p̄). Assume that f is strictly differentiable at ( p̄, x̄)
and suppose further that the partial Jacobian ∇x f ( p̄, x̄) is of full rank m. Then the
mapping S has a local selection s around p̄ for x̄ which is strictly differentiable at p̄
with Jacobian

(3) ∇s( p̄) = AT (AAT )−1 ∇ p f ( p̄, x̄), where A = ∇x f ( p̄, x̄).

Guide. Use 1F.8 and the argument in the proof of 1B.1.

1G. Selections from Nonstrict Differentiability

Even in the case when a function f maps IRn to itself, the inverse f −1 may fail to
have a single-valued localization at ȳ = f (x̄) for x̄ if f is not strictly differentiable
but merely differentiable at x̄ with Jacobian ∇ f (x̄) nonsingular. As when m < n, we
have to deal with just a local selection of f −1 .
52 1 Functions Defined Implicitly by Equations

Theorem 1G.1 (inverse selections from nonstrict differentiability). Let f : IRn →

IRn be continuous in a neighborhood of a point x̄ ∈ int dom f and differentiable at
x̄ with ∇ f (x̄) nonsingular. Then, for ȳ = f (x̄), there exists a local selection of f −1
around ȳ for x̄ which is continuous at ȳ. Moreover, every local selection s of f −1
around ȳ for x̄ which is continuous at ȳ has the property that

(1) s is differentiable at ȳ with Jacobian ∇s(ȳ) = ∇ f (x̄)−1 .

The verification of this claim relies on the following fixed point theorem, which
we state here without proof.

Theorem 1G.2 (Brouwer fixed point theorem). Let Q be a compact and convex set
in IRn , and let Φ : IRn → IRn be a function which is continuous on Q and maps Q into
itself. Then there exists a point x ∈ Q such that Φ (x) = x.
Proof of Theorem 1G.1. Without loss of generality, we can suppose that x̄ = 0 and
f (x̄) = 0. Let A := ∇ f (0) and choose a neighborhood U of 0 ∈ IRn . Take c ≥ |A−1 |.
Choose any α ∈ (0, c−1 ). From the assumed differentiability of f , there exists a > 0
such that x ∈ aIB implies | f (x) − Ax| ≤ α |x|. By making a smaller if necessary, we
can arrange that f is continuous in aIB and aIB ⊂ U. Let b = a(1 − cα )/c and pick
any y ∈ bIB. Consider the function
 
Φy : x → x − A−1 f (x) − y for x ∈ aIB.

This function is of course continuous on the compact and convex set aIB. Further-
more, for any x ∈ aIB we have

|Φy (x)| = |x − A−1( f (x) − y)| = |A−1 (Ax − f (x) + y)| ≤ |A−1 |(|Ax − f (x)| + |y|)
≤ c|Ax − f (x)| + c|y| ≤ cα |x| + cb ≤ cα a + ca(1 − cα )/c = a,

so Φy maps aIB into itself. Then, by Brouwer’s fixed point theorem 1G.2, there
exists a point x ∈ aIB such that Φy (x) = x. Note that, in contrast to the contraction
mapping principle 1A.2, this point may be not unique in aIB. But Φy (x) = x if and
only if f (x) = y. For each y ∈ bIB, y = 0, we pick one x ∈ aIB such that x = Φy (x);
then x ∈ f −1 (y). For y = 0 we take x = 0, which is clearly in f −1 (0). Denoting this
x by s(y), we deduce the existence of a local selection s : bIB → aIB of f −1 around 0
for 0, also having the property that for any neighborhood U of 0 there exists b > 0
such that s(y) ∈ U for y ∈ bIB, that is, s is continuous at 0.
Let s be a local selection of f −1 around 0 for 0 that is continuous at 0. Choose c,
α and a as in the beginning of the proof. Then there exists b > 0 with the property
that s(y) ∈ f −1 (y) ∩ aIB for every y ∈ b IB. This can be written as

s(y) = A−1 (As(y) − f (s(y)) + y) for every y ∈ b IB,

which gives

|s(y)| ≤ |A−1 |(|As(y) − f (s(y))| + |y|) ≤ cα |s(y)| + c|y|,

1 Functions Defined Implicitly by Equations 53

that is,
c
(2) |s(y)| ≤ |y| for all y ∈ b IB.
1 − cα
In particular, s is calm at 0. But we have even more. Choose any ε > 0. The differ-
entiability of f with ∇ f (0) = A furnishes the existence of τ ∈ (0, a) such that

(1 − cα )ε
(3) | f (x) − Ax| ≤ |x| whenever |x| ≤ τ .
c2
Let δ = min{b , τ (1 − cα )/c}. Then δ ≤ b , so that on δ IB we have our local selec-
tion s of f −1 satisfying (2) and consequently

c c (1 − cα )τ
|s(y)| ≤ δ≤ = τ when |y| ≤ δ .
1 − cα 1 − cα c
Taking norms in the identity
 
s(y) − A−1 y = −A−1 f (s(y)) − As(y) ,

and using (2) and (3), we obtain for |y| ≤ δ that

 
s(y)−A−1 y ≤ |A−1 || f (s(y))−As(y)| ≤ c(1 − cα )ε |s(y)| ≤ c(1 − cα )ε c |y| = ε |y|.
c2 c2 (1 − cα )

Having demonstrated that for any ε > 0 there exists δ > 0 such that |s(y) − A−1 y| ≤
ε |y| when |y| ≤ δ , we conclude that (1) holds, as claimed.
In order to gain more insight into what Theorem 1G.1 does or does not say,
think about the case where the assumptions of the theorem hold and f −1 has a
localization at ȳ for x̄ that avoids multi-valuedness. This localization must actually
be single-valued around ȳ, coinciding in some neighborhood with the local selection
s in the theorem. Then we have a result which appears to be fully analogous to the
classical inverse function theorem, but its shortcoming is the need to guarantee that
a localization of f −1 without multi-valuedness does exist. That, in effect, is what
strict differentiability of f at x̄, in contrast to just ordinary differentiability, is able
to provide. An illustration of how inverse multi-valuedness can indeed come up
when the differentiability is not strict has already been encountered in Example
1E.5 with α ∈ (0, 1). Observe that in this example there are infinitely many (even
uncountably many) local selections of the inverse f −1 and, as the theorem says, each
is continuous and even differentiable at 0, but also each selection is discontinuous
at infinitely many points near zero.
We can now partially extend Theorem 1G.1 to the case when m ≤ n.

Theorem 1G.3 (differentiable inverse selections). Let f : IRn → IRm be continuous

in a neighborhood of a point x̄ ∈ int dom f and differentiable at x̄ with A := ∇ f (x̄)
of full rank m. Then, for ȳ = f (x̄), there exists a local selection s of f −1 around ȳ
for x̄ which satisfies (1).
54 1 Functions Defined Implicitly by Equations

Comparing 1G.1 with 1G.3, we see that the equality m = n gives us not only the
existence of a local selection which is differentiable at ȳ but also that every local
selection which is continuous at ȳ, whose existence is assured also for m < n, is
differentiable at ȳ with the same Jacobian. Of course, if we assume in addition that
f is strictly differentiable, we obtain strict differentiability of s at ȳ. To get this last
result, however, we do not have to resort to Brouwer’s fixed point theorem 1G.2.
Theorem 1G.1 is in fact a special case of a still broader result in which f does
not need to be differentiable.

Theorem 1G.4 (inverse selections from first-order approximation). Let f : IRn →

IRm be continuous around x̄ with f (x̄) = ȳ, and let h : IRn → IRm be a first-order
approximation of f at x̄ which is continuous around x̄. Suppose h−1 has a Lipschitz
continuous local selection σ around ȳ for x̄. Then f −1 has a local selection
 s around

ȳ for x̄ for which σ is a first-order approximation at ȳ: s(y) = σ (y) + o |y − ȳ| .
Proof. We follow the proof of 1G.1 with some important modifications. Without
loss of generality, take x̄ = 0, ȳ = 0. Let U be a neighborhood of the origin in IRn .
Let γ > 0 be such that σ is Lipschitz continuous on γ IB, and let c > 0 be a constant
for this. Choose α such that 0 < α < c−1 and a > 0 with aIB ⊂ U and such that
α a ≤ γ /2, f and h are continuous on aIB, and

(4) |e(x)| ≤ α |x| for all x ∈ aIB, where e(x) = f (x) − h(x).

Let
 
a(1 − cα ) γ
(5) 0 < b ≤ min , .
c 2

For x ∈ aIB and y ∈ bIB we have

(6) |y − e(x)| ≤ α a + b ≤ γ .

Fix y ∈ bIB and consider the function

(7) Φy : x → σ (y − e(x)) for x ∈ aIB.

This function is of course continuous on aIB; moreover, from (6), the Lipschitz con-
tinuity of σ on γ IB with constant c, the fact that σ (0) = 0, and the choice of b in (5),
we obtain

|Φy (x)| = |σ (y − e(x))| = |σ (y − e(x)) − σ (0)| ≤ c(α a + b) ≤ a for all x ∈ aIB.

Hence, by Brouwer’s fixed point theorem 1G.2, there exists x ∈ aIB with x = σ (y −
e(x)). Then h(x) = h(σ (y − e(x))) = y − e(x), that is, f (x) = y. For each y ∈ bIB,
y = 0 we pick one such fixed point x of the function Φy in (7) in aIB and call it s(y);
for y = 0 we set s(0) = 0 ∈ f −1 (0). The function s is a local selection of f −1 around
0 for 0 which is, moreover, continuous at 0, since for an arbitrary neighborhood U
of 0 we found b > 0 such that s(y) ∈ U whenever |y| ≤ b. Also, for any y ∈ bIB we
1 Functions Defined Implicitly by Equations 55

have

(8) s(y) = σ (y − e(s(y))).

From the continuity of s at 0 there exists b ∈ (0, b) such that |s(y)| ≤ a for all y ∈
b IB. For y ∈ b IB, we see from (4), (8), the Lipschitz continuity of σ with constant c
and the equality σ (0) = 0 that

|s(y)| = |σ (y − e(s(y))) − σ (0)| ≤ cα |s(y)| + c|y|.

Hence, since cα < 1,

c
(9) |s(y)| ≤ |y| when |y| ≤ b .
1 − αc
Now, let ε > 0. By the assumption that h is a first-order approximation of f at 0,
there exists τ ∈ (0, a) such that

(1 − α c)ε
(10) |e(x)| ≤ |x| whenever |x| ≤ τ .
c2
Finally, taking b > 0 smaller if necessary and using (9) and (10), for any y with
|y| ≤ b we obtain

|s(y) − σ (y)| = σ (y − e(s(y))) − σ (y)|
(1 − α c)ε
≤ c|e(s(y))| ≤ c |s(y)|
c2
(1 − α c)ε c
≤c |y| = ε |y|.
c2 1 − αc
Since for any ε > 0 we found b > 0 for which this holds when |y| ≤ b , the proof is
complete.

Proof of Theorem 1G.3. Apply Theorem 1G.4 with h(x) = f (x̄) + A(x − x̄) and
σ (y) = AT (AAT )−1 y.
We state next as an exercise an implicit function counterpart of 1G.3.

Exercise 1G.5 (differentiability of a selection). Consider a function f : IRd × IRn →

IRm with m ≤ n, along with the solution mapping
  
S : p → x  f (p, x) = 0 for p ∈ IRd .

Let f ( p̄, x̄) = 0, so that x̄ ∈ S( p̄), and suppose f is continuous around ( p̄, x̄) and
differentiable at ( p̄, x̄). Assume further that ∇x f ( p̄, x̄) has full rank m. Prove that the
mapping S has a local selection s around p̄ for x̄ which is differentiable at p̄ with
Jacobian
∇s(ȳ) = AT (AAT )−1 ∇ p f ( p̄, x̄), where A = ∇x f ( p̄, x̄).
56 1 Functions Defined Implicitly by Equations

The existence of a local selection of the inverse of a function f around ȳ = f (x̄)

for x̄ implies in particular that f −1 (y) is nonempty for all y in a neighborhood of
ȳ = f (x̄). This weaker property has even deeper significance and is defined next.

Openness. A function f : IRn → IRm is said to be open at x̄ if x̄ ∈ int dom f and for
every neighborhood U of x̄ the set f (U) is a neighborhood of f (x̄).
Thus, f is open at x̄ if for every open neighborhood U of x̄ there is an open
neighborhood V of ȳ = f (x̄) such that f −1 (y) ∩U = 0/ for every y ∈ V . In particular
this corresponds to the localization of f −1 relative to V and U being nonempty-
valued on V , but goes further than referring just to a single such localization at ȳ
for x̄. It actually requires the existence of a nonempty-valued graphical localization
for every neighborhood U of x̄, no matter how small. From 1G.3 we obtain the
following basic result about openness:

Corollary 1G.6 (Jacobian criterion for openness). For a function f : IRn → IRm ,
where m ≤ n, suppose that f is continuous around x̄ and differentiable at x̄ with
∇ f (x̄) being of full rank m. Then f is open at x̄.
There is much more to say about openness of functions and set-valued mappings,
and we will explore this in detail in Chapters 3 and 5.
1 Functions Defined Implicitly by Equations 57

Commentary
Although functions given implicitly by equations had been considered earlier by
Descartes, Newton, Leibnitz, Lagrange, Bernoulli, Euler, and others, Cauchy [1831]
is credited by historians to be the first who stated and rigorously proved an im-
plicit function theorem — for analytic functions, by using his calculus of residuals
and limits. As we mentioned in the preamble to this chapter, Dini [1877/78] gave
the form of the implicit function theorem for continuously differentiable functions
which is now used in most calculus books; in his proof he relied on the mean value
theorem. More about early history of the implicit function theorem can be found
in historical notes of the paper of Hurwicz and Richter [2003] and in the book of
Krantz and Parks [2002].
Proof I of the classical inverse function theorem, 1A.1, goes back to Goursat
[1903]8. Apparently not aware of Dini’s theorem and inspired by Picard’s suc-
cessive approximation method for proving solution existence of differential equa-
tions, Goursat stated an implicit function theorem under assumptions weaker than
in Dini’s theorem, and supplied it with a new path-breaking proof. With updated
notation, Goursat’s proof employs the iterative scheme

(1) xk+1 = xk − A−1 f (p, xk ), where A = ∇x f ( p̄, x̄).

This scheme would correspond to Newton’s method for solving f (p, x) = 0 with
respect to x if A were replaced by Ak giving the partial derivative at (p, xk ) instead
of ( p̄, x̄). But Goursat proved anyway that for each p near enough to p̄ the sequence
{xk } is convergent to a unique point x(p) close to x̄, and furthermore that the func-
tion p → x(p) is continuous at p̄. Behind the scene, as in Proof I of Theorem 1A.1,
is the contraction mapping idea. An updated form of Goursat’s implicit function
theorem is given in Theorem 1B.6. In the functional analysis text by Kantorovich
and Akilov [1964], Goursat’s iteration is called a “modified Newton’s method.”
The rich potential in this proof was seen by Lamson [1920], who used the it-
erations in (1) to generalize Goursat’s theorem to what are now known as Banach
spaces9 . Especially interesting for our point of view in the present book is the fact
that Lamson was motivated by an optimization problem, namely the problem of La-
grange in the calculus of variations with equality constraints, for which he proved a
Lagrange multiplier rule by way of his implicit function theorem.
Lamson’s work was extended in a significant way by Hildebrand and Graves
[1927], who also investigated differentiability properties of the implicit function.
They first stated a contraction mapping result (their Theorem 1), in which the only
difference with the statement of Theorem 1A.2 is the presence of a superfluous pa-
rameter. The contraction mapping principle, as formulated in 1A.3, was published

8 Edouard Jean-Baptiste Goursat (1858–1936). Goursat paper from 1903 is available at

http://www.numdam.org/.
9 The name “Banach spaces” for normed linear spaces that are complete was coined by Fréchet,

according to Hildebrand and Graves [1927]; we deal with Banach spaces in Chapter 5.
58 1 Functions Defined Implicitly by Equations

five years earlier in Banach [1922] (with some easily fixed typos), but the idea be-
hind the contraction mapping was evidently known to Goursat, Picard and proba-
bly even earlier. Hildebrand and Graves [1927] cited in their paper Banach’s work
[1922], but only in the context of the definition of a Banach space. Further, based on
their parameterized formulation of 1A.2, they established an implicit function the-
orem in the classical form of 1B.1 (their Theorem 4) for functions acting in linear
metric spaces. More intriguing, however, for its surprising foresight, is their Theo-
rem 3, called by the authors a “neighborhood theorem,” where they do not assume
differentiability; they say “only an approximate differential . . . is required.” In this,
they are far ahead of their time. (We will see a similar picture with Graves’ theo-
rem later in Section 5D.) Because of the importance of this result of Hildebrand and
Graves, we provide a statement of it here in finite dimensions with some adjustments
in terminology and notation.

Theorem (Hildebrand–Graves). Let Q ⊂ IRd and consider a function f : Q × IRn →

IRn along with a point x̄ ∈ IRn . Suppose there are a positive constant a, a linear
bounded mapping A : IRn → IRn which is invertible, and a positive constant M with
M|A−1 | < 1 such that
(a) for all p ∈ Q and x, x ∈ IBa (x̄), one has | f (p, x) − f (p, x ) − A(x − x )| ≤
M|x − x |;
(b) for every p ∈ Q, one has |A−1 || f (p, x̄)| ≤ a(1 − M|A−1|).
  
Then the solution mapping S : p → x  f (p, x) = 0 is single-valued on Q [when
its values are restricted to a neighborhood of x̄].
The phrase in brackets in the last sentence is our addition: Hildebrand and Graves
apparently overlooked the fact, which is still overlooked by some writers, that the
implicit function theorem is about a graphical localization of the solution mapping.
If we assume in addition that f is Lipschitz continuous with respect to p uniformly
in x, we will obtain, according to 1E.13, that the solution mapping has a Lipschitz
continuous single-valued localization. When f is assumed strictly differentiable at
( p̄, x̄), by taking A = Dx f ( p̄, x̄) we come to 1D.13.
The classical implicit function theorem is present in many of the textbooks in
calculus and analysis written in the last hundred years. The proofs on the introduc-
tory level are mainly variations of Dini’s proof, depending on the material covered
prior to the theorem’s statement. Interestingly enough, in his text Goursat [1904]
applies the mean value theorem, as in Dini’s proof, and not the contraction mapping
iteration he introduced in his paper of 1903. Similar proofs are given as early as the
30s in Courant [1988] and most recently in Fitzpatrick [2006]. In more advanced
textbooks from the second half of the last century, such as in the popular texts of
Apostol [1962], Schwartz [1967] and Dieudonné [1969], it has become standard to
use the contraction mapping principle.
The material of Chapters 1C and 1D is mostly known, but the way it is presented
is new. We were not able to identify a calculus text in which the inverse function the-
orem is given in the symmetric form 1C.3. First-order approximations of functions
were introduced in Robinson [1991].
1 Functions Defined Implicitly by Equations 59

Theorem 1E.3 can be viewed as an extension of the Hildebrand–Graves theo-

rem (see above) where the “approximate differential” is not required to be a linear
mapping; we will get back to this result in Chapter 2 and also later in the book.
The statement of 1E.8(a) is sometimes called the Banach lemma, see, e.g., Noble
and Daniel [1977]. Kahan [1966] and many after him attribute Theorem 1E.9 to
Gastinel, without giving a reference. This result can be also found in the literature
as the “Eckart–Young theorem” with the citation of Eckart and Young [1936], which
however is a related but different kind of result, concerning the distance from a ma-
trix from another matrix with lower rank. That result in turn is much older still and
is currently referred to as the Schmidt–Eckart–Young–Mirsky theorem on singular
value decomposition. For history see Chipman [1997] and the book by Stewart and
Sun [1990].
On the other hand, 1E.9 can be derived with the help of the Schmidt–Eckart–
Young–Mirsky theorem, inasmuch as the latter implies that the distance from a non-
singular matrix to the set of rank-one matrices is equal to the smallest singular value,
which is the reciprocal to the norm of the inverse. Hence the distance from a non-
singular matrix to the set of singular matrices is not greater than the reciprocal to
the norm of the inverse. Combining this with 1E.7 gives us the radius equality. More
about stability of perturbed inversions will be presented in Chapters 5 and 6.
Brouwer’s invariance of domain theorem 1F.1 can be found, e.g., in Spanier
[1966], while Brouwer’s fixed point theorem 1G.2 is given in Dunford and Schwartz
[1958]. Theorem 1F.2 is from Kummer [1991]. Theorem 1G.3 slightly extends a re-
sult in Halkin [1974]; for extensions in other directions, see Hurwicz and Richter
[2003].
Chapter 2
Implicit Function Theorems for Variational
Problems

Solutions mappings in the classical setting of the implicit function theorem concern
problems in the form of parameterized equations. The concept can go far beyond
that, however. In any situation where some kind of problem in x depends on a pa-
rameter p, there is the mapping S that assigns to each p the corresponding set of
solutions x. The same questions then arise about the extent to which a localization
of S around a pair ( p̄, x̄) in its graph yields a function s which might be continuous
or differentiable, and so forth.
This chapter moves into that much wider territory in replacing equation-solving
problems by more complicated problems termed “generalized equations.” Such
problems arise variationally in constrained optimization, models of equilibrium, and
many other areas. An important feature, in contrast to ordinary equations, is that
functions obtained implicitly from their solution mappings typically lack differen-
tiability, but often exhibit Lipschitz continuity and sometimes combine that with the
existence of one-sided directional derivatives.
The first task is to explain “generalized equations” and their special case, some-
what confusingly termed “variational inequality” problems, which arises from the
variational geometry of sets expressing constraints. Problems of optimization and
the Lagrange multiplier conditions characterizing their solutions provide key exam-
ples. Convexity of sets and functions enters as a valuable ingredient.
From that background, the chapter proceeds to Robinson’s implicit function
theorem for parameterized variational inequalities and several of its extensions. Sub-
sequent sections introduce concepts of ample parameterization and semidifferentia-
bility, building toward major results in 2E for variational inequalities over convex
sets that are polyhedral. A follow-up in 2F looks at a type of “monotonicity” and
its consequences, after which, in 2G, a number of applications in optimization are
worked out.

A.L. Dontchev and R.T. Rockafellar, Implicit Functions and Solution Mappings: A View 61
from Variational Analysis, Springer Monographs in Mathematics,
DOI 10.1007/978-0-387-87821-8 2,  c Springer Science+Business Media, LLC 2009
62 2 Implicit Function Theorems for Variational Problems

2A. Generalized Equations and Variational Problems

By a generalized equation in IRn will be meant a condition on x of the form

(1) f (x) + F(x)  0, or equivalently − f (x) ∈ F(x),

for a function f : IRn → IRm and a (generally) set-valued mapping F : IRn →

→ IRm . The
name refers to the fact that (1) reduces to an ordinary equation f (x) = 0 when F is
the zero mapping (with F(x) containing 0 and nothing else, for every x), which we
indicate in notation by F ≡ 0. Any x satisfying (1) is a solution to (1).
Generalized equations take on importance in many situations, but an especially
common and useful type arises from normality conditions with respect to convex
sets.

Normal cones. For a convex set C ⊂ IRn and a point x ∈ C, a vector v is said to be
normal to C at x if v, x − x ≤ 0 for all x ∈ C. The set of all such vectors v is called
the normal cone to C at x and is denoted by NC (x). For x ∈ / C, NC (x) is taken to be
the empty set. The normal cone mapping is thus defined as

NC : x → NC (x) for x ∈ C,
0/ otherwise.
The term cone refers to a set of vectors which contains 0 and contains with any
of its elements v all positive multiples of v. For each x ∈ C, the normal cone NC (x)
is indeed a cone in this sense. Moreover it is closed and convex. The normal cone
mapping NC : x → NC (x) has dom NC = C. When C is closed, gph NC is a closed
subset of IRn × IRn .

Variational inequalities. For a function f : IRn → IRn and a closed convex set C ⊂
dom f , the generalized equation

(2) f (x) + NC (x)  0, or equivalently − f (x) ∈ NC (x),

is called the variational inequality for f and C.

Note that, because NC (x) = 0/ when x ∈ / C, a solution x to (2) must be a point of
C. The name of this condition originated from the fact that, through the definition of
the normal vectors to C, (2) is equivalent to having

(3) x ∈ C and  f (x), x − x ≥ 0 for all x ∈ C.

Instead of contemplating a system of infinitely many linear inequalities, however, it

is usually better to think in terms of the properties of the set-valued mapping NC ,
which the formulation in (2) helps to emphasize.
When x ∈ int C, the only normal vector at x is 0, and the condition in (2) just
becomes f (x) = 0. Indeed, the variational inequality (2) is totally the same as the
2 Implicit Function Theorems for Variational Problems 63

equation f (x) = 0 in the case of C = IRn , which makes NC ≡ 0. In general, though,

(2) imposes a relationship between f (x) and the boundary behavior of C at x.
There is a simple connection between the normal cone mapping NC and the pro-
jection mapping PC in 1D: one has

(4) v ∈ NC (x) ⇐⇒ PC (x + v) = x.

Interestingly, this projection rule for normals means for the mappings NC and PC
that

(5) NC = PC−1 − I, PC = (I + NC )−1 .

A consequence of (5) is that the variational inequality (2) can actually be written as
an equation, namely

(6) f (x) + NC (x)  0 ⇐⇒ PC (x − f (x)) − x = 0.

It should be kept in mind, though, that this doesn’t translate the solving of variatio-
nal inequalities into the classical framework of solving nonlinear equations. There,
“linearizations” are essential, but PC often fails to be differentiable, so linearizations
generally aren’t available for the equation in (6), regardless of the degree of differ-
entiability of f . Other approaches can sometimes be brought in, however, depending
on the nature of the set C. Anyway, the characterization in (6) has the advantage of
leading quickly to a criterion for the existence of a solution to a variational inequal-
ity in a basic case.

Theorem 2A.1 (solutions to variational inequalities). For a function f : IRn → IRn

and a nonempty, closed convex set C ⊂ dom f relative to which f is continuous,
the set of solutions to the variational inequality (2) is always closed. It is sure to be
nonempty when C is bounded.
Proof. Let M(x) = PC (x − f (x)). Because C is nonempty, closed and convex, the
projection mapping PC is, by 1D.5, a continuous function from IRn to C. Then M is a
continuous function from C to C under our continuity assumption on f . According
to (6), the set of solutions x to (2) is the same as the set of points x ∈ C such that
M(x) = x, which is closed. When C is bounded, we can apply Brouwer’s fixed point
theorem, 1G.2, to conclude the existence of at least one such point x.
Other existence theorems which don’t require C to be bounded can also be given,
especially for situations in which f has a property called “monotonicity.” This will
be taken up in 2F.
The examples and properties to which the rest of this section is devoted will
help to indicate the scope of the variational inequality concept. They will also lay
the foundations for the generalizations of the implicit function theorem that we are
aiming at.
64 2 Implicit Function Theorems for Variational Problems

Exercise 2A.2 (some normal cone formulas).

(a) If M is a linear subspace of IRn , then NM (x) = M ⊥ for every x ∈ M , where
⊥
M is the orthogonal complement of M .
 (b) The unit Euclidean ball IB has NIB (x) = {0} when |x| < 1, but NIB (x) =
λ x  λ ≥ 0 when |x| = 1.   
(c) The nonnegative orthant IRn+ = x = (x1 , . . . , xn )  x j ≥ 0 for j = 1, . . . , n has

(v1 , . . . , vn ) ∈ NIRn+ (x1 , . . . , xn ) ⇐⇒ v ≤ 0, v ⊥ x


v j ≤ 0 for j with x j = 0,
⇐⇒
v j = 0 for j with x j > 0.

Guide. The projection rule (4) provides an easy way of identifying the normal vec-
tors in these examples.

The formula in 2A.2(c) comes up far more frequently than might be anticipated.
A variational inequality (2) in which C = IRn+ is called a complementarity problem;
one has
− f (x) ∈ NIRn+ (x) ⇐⇒ x ≥ 0, f (x) ≥ 0, x ⊥ f (x).
Here the common notation is adopted that a vector inequality like x ≥ 0 is to be
taken componentwise, and that x ⊥ y means x, y = 0. Many variational inequali-
ties can be recast, after some manipulation, as complementarity problems, and the
numerical methodology for solving such problems has therefore received especially
much attention.
The orthogonality relation in 2A.2(a) extends to a “polarity” relation for cones
which has a major presence in our subject.

Proposition 2A.3 (polar cone). Let K be a closed, convex cone in IRn and let K ∗ be
its polar, defined by
  
(7) K ∗ = y  x, y ≤ 0 for all x ∈ K .

Then K ∗ is likewise a closed, convex cone, and its polar (K ∗ )∗ is in turn K . Further-
more, the normal vectors to K and K ∗ are related by

(8) y ∈ NK (x) ⇐⇒ x ∈ NK∗ (y) ⇐⇒ x ∈ K, y ∈ K ∗ , x, y = 0.

Proof. First consider any x ∈ K and y ∈ NK (x). From the definition of normality in
(7) we have y, x − x ≤ 0 for all x ∈ K, so the maximum of y, x  over x ∈ K is
attained at x. Because K contains all positive multiples of each of its vectors, this
comes
 down
 tohaving y, x = 0 and y, x  ≤ 0 for all x ∈ K. Therefore NK (x) =
∗ 
y∈K y⊥x .
It’s elementary that K ∗ is a cone which is closed and convex, with (K ∗ )∗ ⊃ K.
Consider any z ∈ / K. Let x = PK (z) and y = z− x. Then y = 0 and PK (x+ y) = x, hence
y ∈ NK (x), so that y ∈ K ∗ and y ⊥ x. We have y, z = y, y > 0, which confirms that
z∈/ (K ∗ )∗ . Therefore (K ∗ )∗ = K. The formula for normals to K must hold then
2 Implicit Function Theorems for Variational Problems 65
  
equally for K ∗ through symmetry: NK ∗ (y) = x ∈ K  x ⊥ y for any y ∈ K ∗ . This
establishes the normality relations that have been claimed.

TC(x)

NC(x)

Fig. 2.1 Tangent and normal cone to a convex set.

Polarity has a basic role in relating the normal vectors to a convex set to its
“tangent vectors.”

Tangent cones. For a set C ⊂ IRn (not necessarily convex) and a point x ∈ C, a
vector v is said to be tangent to C at x if
1 k
(x − x) → v for some xk → x, xk ∈ C, τ k  0.
τk
The set of all such vectors v is called the tangent cone to C at x and is denoted by
TC (x). For x ∈
/ C, TC (x) is taken to be the empty set.
Although we will mainly be occupied with normal cones to convex sets at
present, tangent cones to convex sets and even nonconvex sets will be put to se-
rious use later in the book.

Exercise 2A.4. The tangent cone TC (x) to a closed, convex set C at a point x ∈ C is
the closed, convex cone that is polar to the normal cone NC (x): one has

(9) TC (x) = NC (x)∗ , NC (x) = TC (x)∗ .

Guide. The second of the equations (9) comes immediately from the definition of
NC (x), and the first is then obtained from Proposition 2A.3.
Variational inequalities are instrumental in capturing conditions for optimality in
problems of minimization or maximization and even “equilibrium” conditions such
66 2 Implicit Function Theorems for Variational Problems

as arise in games and models of conflict. To explain this motivation, it will be helpful
to be able to appeal to the convexity of functions, at least in part.

Convex functions. A function g : IRn → IR is said to be convex relative to a convex

set C (or just convex, when C = IRn ) if

g((1 − τ )x + τ x) ≤ (1 − τ )g(x) + τ g(x) for all τ ∈ (0, 1) when x, x ∈ C.

It is strictly convex if this holds with strict inequality for x = x . It is strongly convex
with constant μ when μ > 0 and, for every x, x ∈ C,

g((1 − τ )x + τ x) ≤ (1 − τ )g(x) + τ g(x) − μτ (1 − τ )|x − x|2 for all τ ∈ (0, 1).

A function g is concave, strictly concave or strongly concave, if −g is convex,

strictly convex or strongly convex, respectively. It is affine relative to C when the
inequality is an equation, which corresponds to g being simultaneously convex and
concave relative to C.
The following are the standard criteria for convexity or strict convexity of g which
can be obtained from the definitions in terms of the gradient vectors
 n
∂g
∇g(x) = (x1 , . . . , xn )
∂xj j=1

and the Hessian matrices

 n,n
∂ 2g
∇2 g(x) = (x1 , . . . , xn ) .
∂ xi ∂ x j i, j=1

Exercise 2A.5 (characterizations of convexity).

(a) A differentiable function g : IRn → IR on an open convex set O is convex if
and only if
g(x ) ≥ g(x) + ∇g(x), x − x for all x, x ∈ O.
It is strictly convex if and only if this inequality is always strict when x = x. It is
strongly convex with constant μ, where μ > 0, if and only if
μ 
g(x ) ≥ g(x) + ∇g(x), x − x + |x − x|2 for all x, x ∈ O.
2
(b) A twice differentiable function g on an open convex set O is convex if and
only if ∇2 g(x) is positive semidefinite for every x ∈ O. It is strictly convex if ∇2 g(x)
is positive definite for every x ∈ O. (This sufficient condition for strict convexity is
not necessary, however, in general.) It is strongly convex with constant μ if and only
if μ > 0 and ∇2 g(x)w, w ≥ μ for all x ∈ O and w ∈ IRn with |w| = 1.
Guide. Because the definition of convexity revolves only around points that are
collinear, the convexity of g can be verified by showing that, for arbitrary x ∈ O and
2 Implicit Function Theorems for Variational Problems 67
  
w ∈ IRn , the function ϕ (t) = g(x +tw) is convex on the interval t  x +tw ∈ O . The
conditions for this on ϕ  (t) and ϕ  (t), known from basic calculus, can be applied
by expressing these derivatives in terms of the gradients and Hessian of g. This
approach can be used to verify all the claims.

Optimization problems. In this chapter and later, we consider optimization prob-

lems which, for a given objective function g : IRn → IR and a given constraint set
C ⊂ IRn , take the form

minimize g(x) over all x ∈ C.

The greatest lower bound of the objective function g on C, namely infx∈C g(x), is
the optimal value in the problem, which may or may not be attained, however, and
could even be infinite. If it is attained at a point x̄, then x̄ is said to furnish a global
minimum, or just a minimum, and to be a globally optimal solution; the set of such
points is denoted as argminx∈C g(x). A point x ∈ C is said to furnish a local minimum
of g relative to C and to be a locally optimal solution when, at least, g(x) ≤ g(x ) for
every x ∈ C belonging to some neighborhood of x. A global or local maximum of g
corresponds to a global or local minimum of −g.
In the context of variational inequalities, the gradient mapping ∇g : IRn → IRn
associated with a differentiable function g : IRn → IR will be a focus of attention.
Observe that
∇2 g(x) = ∇ f (x) when f (x) = ∇g(x).

Theorem 2A.6 (basic variational inequality for minimization). Let g : IRn → IR be

differentiable on an open convex set O, and let C be a closed convex subset of O. In
minimizing g over C, the variational inequality

(10) ∇g(x) + NC (x)  0, or equivalently − ∇g(x) ∈ NC (x),

is necessary for x to furnish a local minimum. It is both necessary and sufficient for
a global minimum if g is convex.
Proof. Along with x ∈ C, consider any other point x ∈ C and the function ϕ (t) =
g(x + tw) with w = x − x. From convexity we have x + tw ∈ C for all t ∈ [0, 1]. If
a local minimum of g occurs at x relative to C, then ϕ must have a local minimum
at 0 relative to [0, 1], and consequently ϕ  (0) ≥ 0. But ϕ  (0) = ∇g(x), w. Hence
∇g(x), x − x ≥ 0. This being true for arbitrary x ∈ C, we conclude through the
characterization of (2) in (3) that −∇g(x) ∈ NC (x).
In the other direction, if g is convex and −∇g(x) ∈ NC (x) we have for every x ∈ C
that ∇g(x), x − x ≥ 0, but also g(x ) − g(x) ≥ ∇g(x), x − x by the convexity
criterion in 2A.5(a). Hence g(x ) − g(x) ≥ 0 for all x ∈ C, and we have a global
minimum at x.
To illustrate the condition in Theorem 2A.6, we may use it to reconfirm the pro-
jection rule for normal vectors in (4), which can be stated equivalently as saying
that PC (z) = x if and only if z − x ∈ NC (x). Consider any nonempty, closed, convex
68 2 Implicit Function Theorems for Variational Problems

set C ⊂ IRn and any point z ∈ IRn . Let g(x) = 12 |x − z|2 , which has ∇g(x) = x − z and
∇2 g(x) ≡ I, implying strict convexity. The projection x = PC (z) is the solution to the
problem of minimizing g over C. The variational inequality (10) characterizes it by
the relation −(x − z) ∈ NC (x), which is exactly what was targeted.
According to Theorem 2A.6, minimizing a differentiable convex function g over
a closed, convex set C is equivalent to solving a type of variational inequality (2)
in which f is the gradient mapping ∇g. When C = IRn , so that we are dealing with
unconstrained minimization, this is equivalent to solving f (x) = 0 for f = ∇g. The
notion of a variational inequality thus makes it possible to pass from unconstrained
minimization to constrained minimization. Whether the problem is constrained or
unconstrained, there is no guarantee that the minimum will be attained at a unique
point (although nonuniqueness is impossible when g is strictly convex, at least), but
still, local uniqueness dominates the picture conceptually. For that reason, it does
make sense to be thinking of the task as one of “solving a generalized equation.”
When g is not convex, solving the variational inequality (2) is no longer equiva-
lent to minimization over C, but nevertheless it has a strong association with identi-
fying a local minimum. Anyway, there’s no need really to insist on a minimum. Just
as the equation ∇g(x) = 0 describes, in general, a “stationary point” of g (uncon-
strained), the variational inequality (10) can be viewed as describing a constrained
version of a stationary point, which could be of interest in itself.
The minimization rule in Theorem 2A.6 can be employed to deduce a rule for
determining normal vectors to intersections of convex sets, as in the second part of
the following proposition.

Proposition 2A.7 (normals to products and intersections).

(a) If C = C1 × C2 for closed, convex sets C1 ⊂ IRn1 and C2 ⊂ IRn2 , then for any
x = (x1 , x2 ) ∈ C one has NC (x) = NC1 (x1 ) × NC2 (x2 ).
(b) If C = C1 ∩C2 for closed, convex sets C1 and C2 in IRn , then the formula
  
NC (x) = NC1 (x) + NC2 (x) = v1 + v2  v1 ∈ NC1 (x), v2 ∈ NC2 (x)

holds for any x ∈ C such that there is no v = 0 with v ∈ NC1 (x) and −v ∈ NC2 (x). This
condition is fulfilled in particular for every x ∈ C if C1 ∩int C2 = 0/ or C2 ∩int C1 = 0/ .
Proof. To prove (a), we note that, by definition, a vector v = (v1 , v2 ) belongs to
NC (x) if and only if, for every x = (x1 , x2 ) in C1 × C2 we have 0 ≥ v, x − x =
v1 , x1 − x1 + v2 , x2 − x2 . That’s the same as having v1 , x1 − x1  ≤ 0 for all x1 ∈ C1
and v2 , x2 −x2  ≤ 0 for all x2 ∈ C2 , or in other words, v1 ∈ NC1 (x1 ) and v2 ∈ NC2 (x2 ).
In proving (b), it is elementary that if v = v1 + v2 with v1 ∈ NC1 (x) and v2 ∈
NC2 (x), then for every x in C1 ∩C2 we have both v1 , x − x ≤ 0 and v2 , x − x ≤ 0,
so that v, x − x ≤ 0. Thus, NC (x) ⊃ NC1 (x) + NC2 (x).
The opposite inclusion takes more work to establish. Fix any x ∈ C and v ∈ NC (x).
As we know from (4), this corresponds to x being the projection of x + v on C, which
we can elaborate as follows: (x, x) is the unique solution to the problem

minimize |x1 − (x + v)|2 + |x2 − (x + v)|2 over all (x1 , x2 ) ∈ C1 × C2 with x1 = x2 .

2 Implicit Function Theorems for Variational Problems 69

Consider for k = 1, 2, . . . the version of this minimization problem in which the

constraint x1 = x2 is relaxed by a penalty expression dependent on k:

minimize |x1 − (x + v)|2 + |x2 − (x + v)|2 + k|x1 − x2 |2

(11)
over all (x1 , x2 ) ∈ C1 × C2 .

The expression being minimized here is nonnegative and, as seen from the case of
x1 = x2 = x, has minimum no greater than 2|v|2 . It suffices therefore in the minimiza-
tion to consider only points x1 and x2 such that |x1 − (x+ v)|2 + |x2 − (x+ v)|2 ≤ 2|v|2
and k|x1 − x2 |2 ≤ 2|v|2 . For each k, therefore, the minimum in (11) is attained by
some (xk1 , xk2 ), and these pairs form a bounded sequence such that xk1 − xk2 → 0.
Any accumulation point of this sequence must be of the form (x̃, x̃) and satisfy
|x̃ − (x + v)|2 + |x̃ − (x + v)|2 ≤ 2|v|2 , or in other words |x̃ − (x + v)| ≤ |v|. But by the
projection rule (4), x is the unique closest point of C to x + v, the distance being |v|,
so this inequality implies x̃ = x. Therefore, (xk1 , xk2 ) → (x, x).
We investigate next the necessary condition for optimality provided for problem
(11). Invoking the formula in (a) for the normal cone to C1 × C2 at (xk1 , xk2 ), we see
that it requires
−2[xk1 − (x + v) + k(xk1 − xk2 )] ∈ NC1 (xk1 ),
−2[xk2 − (x + v) − k(xk1 − xk2 )] ∈ NC2 (xk2 ),
or equivalently, for wk = k(xk2 − xk1 ),

(12) v + (x − xk1 ) + wk ∈ NC1 (xk1 ) and v + (x − xk2) − wk ∈ NC2 (xk2 ).

Two cases have to be analyzed now separately. In the first case, we suppose that
the sequence of vectors wk is bounded and therefore has an accumulation point w.
Let vk1 = v + (x − xk1 ) + wk and vk2 = v + (x − xk2 ) − wk , so that, through (4), we have
PC1 (xk1 + vk1 ) = xk1 and PC2 (xk2 + vk2 ) = xk2 . Since xk1 → x and xk2 → x, the sequences of
vectors vk1 and vk2 have accumulation points v1 = v − w and v2 = v + w which satisfy
v1 + v2 = 2v. By the continuity of the projection mappings coming from 1D.5, we
get PC1 (x + v1 ) = x and PC2 (x + v2 ) = x. By (6), these relations mean v1 ∈ NC1 (x) and
v2 ∈ NC2 (x) and hence 2v ∈ NC1 (x) + NC2 (x). Since the sum of cones is a cone, we
get v ∈ NC1 (x) + NC2 (x). Thus NC (x) ⊂ NC1 (x) + NC2 (x), and since we have already
shown the opposite inclusion, we have equality.
In the second case, we suppose that the sequence of vectors wk is unbounded. By
passing to a subsequence if necessary, we can reduce this to having 0 < |wk | → ∞
with wk /|wk | converging to some v̄ = 0. Let

v̄k1 = [v + (x − xk1) + wk ]/|wk | and v̄k2 = [v + (x − xk2) − wk ]/|wk |.

Then v̄k1 → v̄ and v̄k2 → −v̄. By (12) we have v̄k1 ∈ NC1 (xk1 ) and v̄k2 ∈ NC1 (xk2 ), or
equivalently through (4), the projection relations PC1 (xk1 + v̄k1 ) = xk1 and PC2 (xk2 +
v̄k2 ) = xk1 . In the limit we get PC1 (x + v̄) = x and PC2 (x − v̄) = x, so that v̄ ∈ NC1 (x)
and −v̄ ∈ NC2 (x). This contradicts our assumption in (b), and we see thereby that
the second case is impossible.
70 2 Implicit Function Theorems for Variational Problems

We turn now to minimization over sets C that might not be convex and are speci-
fied by systems of constraints which have to be handled with Lagrange multipliers.
This will lead us to other valuable examples of variational inequalities, after some
elaborations.

Theorem 2A.8 (Lagrange multiplier rule). Let X ⊂ IRn and D ⊂ IRm be nonempty,
closed, convex sets, and consider the problem
  
(13) minimize g0 (x) over C = x ∈ X  g(x) ∈ D ,

for g(x) = (g1 (x), . . . , gm (x)), where the functions gi : IRn → IR, i = 0, 1, . . . , m are
continuously differentiable. Let x be a point of C at which the following constraint
qualification condition is fulfilled:

(14) there is no y ∈ ND (g(x)), y = 0, such that − y∇g(x) ∈ NX (x).

If g0 has a local minimum relative to C at x, then

(15) there exists y ∈ ND (g(x)) such that − [∇g0 (x) + y∇g(x)] ∈ NX (x).

Proof. Assume that a local minimum occurs at x. Let X  and D be compact, convex
sets which coincide with X and D in neighborhoods of x and g(x), respectively, and
are small enough that g0 (x ) ≥ g0 (x) for all x ∈ X  having g(x ) ∈ D . Consider the
auxiliary problem

minimize g0 (x ) + 12 |x − x|2

(16)
over all (x , u ) ∈ X  × D satisfying g(x ) − u = 0.

Obviously the unique solution to this is (x , u ) = (x, g(x)). Next, for k → ∞, consider
the following sequence of problems, which replace the equation in (16) by a penalty
expression:
1 k
(17) minimize g0 (x ) + |x − x|2 + |g(x ) − u|2 over all (x , u ) ∈ X  × D .
2 2
For each k let (xk , uk ) give the minimum in this relaxed problem (the minimum being
attained because the functions are continuous and the sets X  and D are compact).
The minimum value in (17) can’t be greater than the minimum value in (16), as seen
by taking (x , u ) to be the unique solution (x, g(x)) to (16). It’s apparent then that
the only possible cluster point of the bounded sequence {(xk , uk )}∞ k=1 as k → ∞ is
(x, g(x)). Thus, (xk , uk ) → (x, g(x)).
Next we apply the optimality condition in Theorem 2A.6 to problem (17) at its
solution (xk , uk ). We have NX  ×D (xk , uk ) = NX  (xk ) × ND (uk ), and on the other hand
NX  (xk ) = NX (xk ) and ND (uk ) = ND (uk ) by the choice of X  and D , at least when
k is sufficiently large. The variational inequality condition in Theorem 2A.6 comes
down in this way to
2 Implicit Function Theorems for Variational Problems 71

−[ ∇g0 (xk ) + (xk − x) + k(g(xk ) − uk )∇g(xk ) ] ∈ NX (xk ),
(18)
k(g(xk ) − uk ) ∈ ND (uk ).

By passing to subsequences if necessary, we can reduce the rest of the analysis to

distinguishing between case (A), where the norms of the vectors k(g(xk ) − uk ) ∈ IRm
stay bounded as k → ∞, and case (B), where these norms go to ∞.
In case (A) we can arrange, by passing again to a subsequence if necessary, that
the sequence of vectors k(g(xk ) − uk ) converges to some y. Then y satisfies the de-
sired relations in (18), inasmuch as (xk , uk ) → (x, g(x)) and the graphs of the map-
pings NX and ND are closed.
In case (B) we look at the vectors yk = k(g(xk )−uk )/ρk with ρk = k|g(xk )−uk | →
∞, which have |yk | = 1 and, from (18), satisfy

(19) −ρk−1 (∇g0 (xk ) + (xk − x)) − yk ∇g(xk ) ∈ NX (xk ), yk ∈ ND (uk ).

(Here we use the fact that any positive multiple of a vector in NX (xk ) or ND (uk ) is
another such vector.) By passing to a subsequence, we can arrange that the sequence
of vectors yk converges to some y, necessarily with |y| = 1. In this limit, (19) turns
into the relation in (14), which has been forbidden to hold for any y = 0. Hence case
(B) is impossible under our assumptions, and we are left with the conclusion (15)
obtained from case (A).
In the first-order optimality condition (15), y is said to be a Lagrange multiplier
vector associated with x. More can be said about this condition by connecting it with
the Lagrangian function for problem (13), which is defined by

(20) L(x, y) = g0 (x) + y, g(x) = g0 (x) + y1 g1 (x) + · · · + ym gm (x)

for y = (y1 , . . . , ym ).

Theorem 2A.9 (Lagrangian variational inequalities). In the minimization problem

(13), suppose that the set D is a cone, and let Y be the polar cone D∗ ,
  
Y = y  u, y ≤ 0 for all u ∈ D .

Then, in terms of the Lagrangian L in (20), the condition on x and y in (15) can be
written in the form

(21) −∇x L(x, y) ∈ NX (x), ∇y L(x, y) ∈ NY (y),

which furthermore can be identified with the variational inequality

(22) − f (x, y) ∈ NX×Y (x, y) for f (x, y) = (∇x L(x, y), −∇y L(x, y)).

The existence of y ∈ Y satisfying this variational inequality with x is thus necessary

for the local optimality of x in problem (13) when the constraint qualification (14)
is fulfilled. If L(·, y) is convex on X when y ∈ Y , the existence of a y satisfying this
72 2 Implicit Function Theorems for Variational Problems

variational inequality with x is moreover sufficient for x to give a global minimum

in problem (13), without any need for invoking (14).
Proof. We have ∇x L(x, y) = ∇g0 (x) + y∇g(x) and ∇y L(x, y) = g(x). The NX condi-
tion in (15) amounts therefore to the first condition in (21). The choice of Y = D∗
makes it possible through the polarity rule for normal vectors in (8) to express the
ND condition in (15) as g(x) ∈ NY (y) and identify it with the second condition in
(21), while deducing from it also that y, g(x) = 0, hence L(x, y) = g0 (x). The re-
casting of (21) as the variational inequality in (22) comes out of the product rule in
2A.7(a).
When the function L(·, y) is convex on X, the condition −∇x L(x, y) ∈ NX (x) im-
plies through Theorem 2A.6 that L(x , y) ≥ L(x, y) for all x ∈ X, where it may be
recalled that L(x, y) = g0 (x) because y, g(x) = 0. Thus, L(x , y) ≥ g0 (x) for all
x ∈ X. On the other hand, since y ∈ Y and Y = D∗ , we have y, g(x ) ≤ 0 when
g(x ) ∈ D. Therefore g0 (x ) ≥ L(x , y) ≥ g0 (x) for all x satisfying the constraints in
(13). It follows that all such x have g0 (x ) ≥ g0 (x), so x furnishes the global mini-
mum in problem (13).

Application to nonlinear programming. Theorems 2A.8 and 2A.9 cover the

case of a standard problem of nonlinear programming, where the task is to

≤ 0 for i ∈ [1, s],
(23) minimize g0 (x) over all x satisfying gi (x)
= 0 for i ∈ [s + 1, m].

This problem1 corresponds in (13) to taking X = IRn and having D be the closed,
convex cone in IRm consisting of all u = (u1 , . . . , um ) such that ui ≤ 0 for i ∈ [1, s] but
ui = 0 for i ∈ [s + 1, m]. The polar cone Y = D∗ is Y = IRs+ × IRm−s . The optimality
condition in (18) can equally well be placed then in the Lagrangian framework in
(21), corresponding to the variational inequality (22). The requirements it imposes
on x and y come out as

≤ 0 for i ∈ [1, s] with yi = 0,
y ∈ IR+ × IR , gi (x)
s m−s
(24) = 0 for all other i ∈ [1, m],
∇g0 (x) + y1 ∇g1 (x) + · · · + ym ∇gm (x) = 0.

These are the Karush–Kuhn–Tucker conditions for the nonlinear programming prob-
lem (23). According to Theorem 2A.8, the existence of y satisfying these conditions
with x is necessary for the local optimality of x under the constraint qualification
(14), which insists on the nonexistence of y = 0 satisfying the same conditions with
the term ∇g0 (x) suppressed. The existence of y satisfying (24) is sufficient for the
global optimality of x by Theorem 2A.9 as long as L(x, y) is convex as a function of
x ∈ IRn for each fixed y ∈ IRs+ × IRm−s , which is equivalent to having

g0 , g1 , . . . , gs convex, but gs+1 , . . . , gm affine.

1 In (23) and later in the book [1, s] denotes the set of integers {1, 2, . . ., s}.
2 Implicit Function Theorems for Variational Problems 73

Then (23) is a problem of convex programming. The Karush–Kuhn–Tucker condi-

tions correspond then to a saddle point property, as indicated next.

Exercise 2A.10 (variational inequality for a saddle point). Let X ⊂ IRn and Y ⊂ IRm
be any nonempty, closed, convex sets, and let L be a C 1 function on IRn × IRm such
that L(·, y) is a convex function on X for each y ∈ Y , and L(x, ·) is a concave function
on Y for each x ∈ X . The variational inequality (22) is equivalent to having (x, y) be
a saddle point of L with respect to X × Y in the sense that

x ∈ X , y ∈ Y, and L(x , y) ≥ L(x, y) ≥ L(x, y ) for all x ∈ X , y ∈ Y.

Guide. Rely on the equivalence between (21) and (22), plus Theorem 2A.6.
A saddle point as defined in Exercise 2A.10 represents an equilibrium in the two-
person zero-sum game in which Player 1 chooses x ∈ X , Player 2 chooses y ∈ Y , and
then Player 1 pays the amount L(x, y) (possibly negative) to Player 2. Other kinds
of equilibrium can likewise be captured by other variational inequalities.
For example, in an N-person game there are players 1, . . . , N, with Player k hav-
ing a nonempty strategy set Xk . Each Player k chooses some xk ∈ Xk , and is then
obliged to pay—to an abstract entity (not necessarily another player)—an amount
which depends not only on xk but also on the choices of all the other players; this
amount can conveniently be denoted by

Lk (xk , x−k ), where x−k = (x1 , . . . , xk−1 , xk+1 , . . . , xN ).

(The game is zero-sum if ∑Nk=1 Lk (xk , x−k ) = 0.) A choice of strategies xk ∈ Xk for
k = 1, . . . , N is said to furnish a Nash equilibrium if

Lk (xk , x−k ) ≥ Lk (xk , x−k ) for all xk ∈ Xk , k = 1, . . . , N.

A saddle point as in Exercise 2A.10 corresponds to the case of this where N = 2, so

x−1 and x−2 are just x2 and x1 respectively, and one has L2 (x2 , x1 ) = −L1 (x1 , x2 ).

Exercise 2A.11 (variational inequality for a Nash equilibrium). In an N -person

game as described, suppose that Xk is a closed, convex subset of IRnk and that
L(xk , x−k ) is differentiable with respect to xk for every k. Then for x = (x1 , . . . , xN )
to furnish a Nash equilibrium, it must solve the variational inequality (2) for f and
C in the case of
 
C = X1 ×· · ·×XN , f (x) = f (x1 , . . . , xN ) = ∇x1 L1 (x1 , x−1 ), . . . , ∇xN LN (xN , x−N ) .

This necessary condition is sufficient for a Nash equilibrium if, in addition, the
functions Lk (·, x−k ) on IRnk are convex.
Guide. Make use of the product rule for normals in 2A.7(a) and the optimality
condition in Theorem 2A.6.
Finally, we look at a kind of generalized equation (1) that is not a variational
inequality (2), but nonetheless has importance in many situations:
74 2 Implicit Function Theorems for Variational Problems

(25) (g1 (x), . . . , gm (x)) ∈ D,

which is (1) for f (x) = −(g1 (x), . . . , gm (x)), F(x) ≡ D. Here D is a subset of IRm ; the
format has been chosen to be that of the constraints in problem (13), or as a special
case, problem (23).
Although (25) would reduce to an equation, pure and simple, if D consists of
a single point, the applications envisioned for it lie mainly in situations where in-
equality constraints are involved, and there is little prospect or interest in a solution
being locally unique. In the study of generalized equations with parameters, to be
taken up next in 2B, our attention will at first be concentrated on issues parallel to
those in Chapter 1. Only later, in Chapter 3, will generalized equations like (25) be
brought in.
The example in (25) also brings a reminder about a feature of generalized equa-
tions which dropped out of sight in the discussion of the variational inequality case.
In (2), necessarily f had to go from IRn to IRn , whereas in (25), and in (1), f may go
from IRn to a space IRm of different dimension.

2B. Implicit Function Theorems for Generalized Equations

With the concept of a generalized equation, and in particular that of a variational
inequality problem at our disposal, we are ready to embark on a broad exploration
of implicit function theorems beyond those in Chapter 1. The object of study is now
a parameterized generalized equation

(1) f (p, x) + F(x)  0

for a function f : IRd × IRn → IRm and a set-valued mapping F : IRn →

→ IRm . Specifi-
→
cally, we consider the properties of the solution mapping S : IR → IRn defined by
p

  
(2) S : p → x  f (p, x) + F(x)  0 for p ∈ IRd .

The questions we will concentrate on answering, for now, are nevertheless the same
as in Chapter 1. To what extent might S be single-valued and possess various prop-
erties of continuity or some type of differentiability?
In a landmark paper2 , S. M. Robinson studied the solution mapping S in the
case of a parameterized variational inequality, where m = n and F is a normal cone
mapping NC : IRn → → IRn :

(3) f (p, x) + NC (x)  0, with C ⊂ IRn convex, closed and nonempty.

2 Cf. Robinson [1980].

2 Implicit Function Theorems for Variational Problems 75

His results were, from the very beginning, stated in abstract spaces, and we will
come to that in Chapter 5. Here, we confine the exposition to Euclidean spaces, but
the presentation is tailored in such a way that, for readers who are familiar with some
basic functional analysis, the expansion of the framework from Euclidean spaces to
general Banach spaces is straightforward. The original formulation of Robinson’s
theorem, up to some rewording to fit this setting, is as follows.

Theorem 2B.1 (Robinson implicit function theorem). For the solution mapping S
to a parameterized variational inequality (3), consider a pair ( p̄, x̄) with x̄ ∈ S( p̄).
Assume that:
(a) f (p, x) is differentiable with respect to x in a neighborhood of the point ( p̄, x̄),
and both f (p, x) and ∇x f (p, x) depend continuously on (p, x) in this neighborhood;
(b) the inverse G−1 of the set-valued mapping G : IRn → → IRn defined by

(4) G(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄) + NC (x), with G(x̄)  0,

has a Lipschitz continuous single-valued localization σ around 0 for x̄ with

lip (σ ; 0) ≤ κ .

Then S has a single-valued localization s around p̄ for x̄ which is continuous at

p̄, and moreover for every ε > 0 there is a neighborhood Q of p̄ such that

(5) |s(p ) − s(p)| ≤ (κ + ε )| f (p , s(p)) − f (p, s(p))| for all p , p ∈ Q.

An extended version of this result will be stated shortly as Theorem 2B.5, so

we can postpone the discussion of its proof until then. Instead, we can draw some
immediate conclusions from the estimate (5) which rely on additional assumptions
about partial calmness and Lipschitz continuity properties of f (p, x) with respect to
p and the modulus notation for such properties that was introduced in 1C and 1D.

Corollary 2B.2 (calmness of solutions). In the setting of Theorem 2B.1, if f is

calm with respect to p at ( p̄, x̄), having clm p ( f ; ( p̄, x̄)) ≤ λ , then s is calm at p̄ with
clm (s; p̄) ≤ κλ .

Corollary 2B.3 (Lipschitz continuity of solutions). In the setting of Theorem 2B.1,

if f is Lipschitz continuous with respect to p uniformly in x around ( p̄, x̄), having
 ( f ; ( p̄, x̄)) ≤ λ , then s is Lipschitz continuous around p̄ with lip (s; p̄) ≤ κλ .
lip p

Differentiability of the localization s around p̄ can’t be deduced from the estimate

in (4), not to speak of continuous differentiability around p̄, and in fact differentia-
bility may fail. Elementary one-dimensional examples of variational inequalities
exhibit solution mappings that are not differentiable, usually in connection with the
“solution trajectory” hitting or leaving the boundary of the set C. For such mappings,
weaker concepts of differentiability are available. We will touch upon this in 2D.
76 2 Implicit Function Theorems for Variational Problems

In the special case where the variational inequality treated by Robinson’s the-
orem reduces to the equation f (p, x) = 0 (namely with C = IRn , so NC ≡ 0), the
invertibility condition on the mapping G in assumption (b) of Robinson’s theorem
comes down to the nonsingularity of the Jacobian ∇x f ( p̄, x̄) in the Dini classical
implicit function theorem 1B.1. But because of the absence of an assertion about
the differentiability of s, Theorem 2B.1 falls short of yielding all the conclusions of
that theorem. It could, though, be used as an intermediate step in a proof of Theo-
rem 1B.1, which we leave to the reader as an exercise.

Exercise 2B.4. Supply a proof of the classical implicit function theorem 1B.1 based
on Robinson’s theorem 2B.1.
Guide. In the case C = IRn , so NC ≡ 0, use the Lipschitz continuity of the single-
valued localization s following from Corollary 2B.3 to show that s is continuously
differentiable around p̄ when f is continuously differentiable near ( p̄, x̄).
The invertibility property in assumption (b) of 2B.1 is what Robinson called
“strong regularity” of the generalized equation (3). A related term, “strong metric
regularity,” will be employed in Chapter 3 for set-valued mappings in reference to
the existence of Lipschitz continuous single-valued localizations of their inverses.
In the extended version of Theorem 2B.1 which we present next, the differen-
tiability assumptions on f are replaced by assumptions about an estimator h for
f ( p̄, ·), which could in particular be a first-order approximation in the x argument.
This mode of generalization was initiated in 1E.3 and 1E.13 for equations, but now
we use it for a generalized equation (1). In contrast to Theorem 2B.1, which was
concerned with the case of a variational inequality (3), the mapping F : IRn → → IRm
need not be of form NC and the dimensions n and m could in principle be different.
Remarkably, no direct assumptions need be made about F, but certain properties of
F will implicitly underlie the “invertibility” condition imposed jointly on F and the
estimator h.

Theorem 2B.5 (Robinson theorem extended beyond differentiability). For a gen-

eralized equation (1) and its solution mapping S in (2), let p̄ and x̄ be such that
x̄ ∈ S( p̄). Assume that:
(a) f (·, x̄) is continuous at p̄, and h is a strict estimator of f with respect to x
uniformly in p at ( p̄, x̄) with constant μ;
(b) the inverse G−1 of the mapping G = h + F , for which G(x̄)  0, has a Lip-
schitz continuous single-valued localization σ around 0 for x̄ with lip (σ ; 0) ≤ κ for
a constant κ such that κ μ < 1.
Then S has a single-valued localization s around p̄ for x̄ which is continuous at
p̄, and moreover for every ε > 0 there is a neighborhood Q of p̄ such that
κ +ε
(6) |s(p ) − s(p)| ≤ | f (p , s(p)) − f (p, s(p))| for all p , p ∈ Q.
1 − κμ

Theorem 2B.1 follows at once from Theorem 2B.5 by taking F to be NC and

h to be the linearization of f (·, p̄) given by h(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄), and
2 Implicit Function Theorems for Variational Problems 77

employing 1E.15. More generally, h could be a strict first-order approximation: the

case when μ = 0. That case, which has further implications, will be taken up later.
However, Theorem 2B.5 is able to extract information from much weaker relation-
ships between f and h than strict first-order approximation, and this information
can still have important consequences for the behavior of solutions to a generalized
equation, as seen in this pattern already for equations in 1E.
Our proof of Theorem 2B.5 will proceed through an intermediate stage in which
we isolate a somewhat lengthy statement as a lemma. Proving the lemma requires
an appeal to the contraction mapping principle, as formulated in Theorem 1A.2.

Lemma 2B.6. Consider a function ϕ : IRd ×IRn → IRm and a point ( p̄, x̄) ∈ int dom ϕ
and let the scalars ν ≥ 0, b ≥ 0, a > 0, and the set Q ⊂ IRd be such that p̄ ∈ Q and

|ϕ (p, x ) − ϕ (p, x)| ≤ ν |x − x | for all x , x ∈ IBa (x̄) and p ∈ Q,
(7)
|ϕ (p, x̄) − ϕ ( p̄, x̄)| ≤ b for all p ∈ Q.

Consider also a set-valued mapping M : IRm → → IRn with (ȳ, x̄) ∈ gph M where ȳ :=
ϕ ( p̄, x̄), such that for each y ∈ IBν a+b (ȳ) the set M(y) ∩ IBa (x̄) consists of exactly
one point, denoted by r(y), and suppose that the function

(8) r : y → M(y) ∩ IBa (x̄) for y ∈ IBν a+b (ȳ)

is Lipschitz continuous on IBν a+b (ȳ) with a Lipschitz constant λ . In addition, sup-
pose that
(a) λ ν < 1;
(b) λ ν a + λ b ≤ a.
Then for each p ∈ Q the set {x ∈ IBa (x̄) | x ∈ M(ϕ (p, x))} consists of exactly one
point, and the associated function

(9) s : p → {x | x = M(ϕ (p, x)) ∩ IBa (x̄)} for p ∈ Q

satisfies
λ
(10) |s(p ) − s(p)| ≤ |ϕ (p , s(p)) − ϕ (p, s(p))| for all p , p ∈ Q.
1−λν
Proof. Fix p ∈ Q and consider the function Φ p : IRn → IRn defined by

Φ p : x → r(ϕ (p, x)) for x ∈ IBa (x̄).

First, note that for x ∈ IBa (x̄) from (7) one has |ȳ − ϕ (p, x)| ≤ b + ν a, thus, by (8),
IBa (x̄) ⊂ dom Φ p . Next, if x ∈ IBa (x̄), we have from the identity x̄ = r(ϕ ( p̄, x̄)), the
Lipschitz continuity of r, and conditions (7) and (b) that

|Φ p (x̄) − x̄| = |r(ϕ (p, x̄)) − r(ϕ ( p̄, x̄))| ≤ λ |ϕ (p, x̄) − ϕ ( p̄, x̄)| ≤ λ b ≤ a(1 − λ ν ).

For any x , x ∈ IBa (x̄) we obtain

78 2 Implicit Function Theorems for Variational Problems

|Φ p (x ) − Φ p (x)| = |r(ϕ (p, x )) − r(ϕ (p, x))| ≤ λ |ϕ (p, x ) − ϕ (p, x)| ≤ λ ν |x − x|,

that is, Φ p is Lipschitz continuous in IBa (x̄) with constant λ ν < 1, from condition
(a). We are in position then to apply the contraction mapping principle 1A.2 and to
conclude from it that Φ p has a unique fixed point in IBa (x̄).
Denoting that fixed point by s(p), and doing this for every p ∈ Q, we get
a function s : Q → IBa (x̄). But having x = Φ p (x) is equivalent to having x =
r(ϕ (p, x)) = M(ϕ (p, x)) ∩ IBa (x̄). Hence s is the function in (9). Moreover, since
s(p) = r(ϕ (p, s(p))), we have from the Lipschitz continuity of r and (7) that, for
any p , p ∈ Q,

|s(p ) − s(p)| = |r(ϕ (p , s(p ))) − r(ϕ (p, s(p)))|

≤ |r(ϕ (p , s(p ))) − r(ϕ (p , s(p)))| + |r(ϕ (p , s(p))) − r(ϕ (p, s(p)))|
≤ λ |ϕ (p , s(p )) − ϕ (p , s(p))| + λ |ϕ (p , s(p)) − ϕ (p, s(p))|
≤ λ ν |s(p ) − s(p)| + λ |ϕ (p, s(p)) − ϕ (p, s(p))|.

Since λ ν < 1, we see that s satisfies (10), as needed.

It’s worth noting that the version of the contraction mapping principle utilized in
proving Lemma 2B.6 (namely, Theorem 1A.2 with X = IRn equipped with the metric
induced by the Euclidean norm | · |) can in turn be derived from Lemma 2B.6. For
that, the data in the lemma need to be specified as follows: d = m = n, ν = λ ,
a unchanged, b = |Φ (x̄) − x̄|, p̄ = 0, Q = IBb (0), ϕ (p, x) = Φ (x) + p, ȳ = Φ (x̄),
M(y) = y + x̄ − Φ (x̄), and consequently the λ in 1A.2 is 1. All the conditions of
Lemma 2B.6 hold for such data under the assumptions  of the contraction
 mapping
principle 1A.2. Hence for p = Φ (x̄) − x̄ ∈ Q the set x ∈ IB (x̄)  x = M(ϕ (p, x)) =
 a
Φ (x) consists of exactly one point; that is, Φ has a unique fixed point in IBa (x̄).
Thus, Lemma 2B.6 is actually equivalent3 to the form of the contraction mapping
principle used in its proof.
Proof of Theorem 2B.5. For an arbitrary ε > 0, choose any λ > lip (σ ; 0) and
ν > μ such that λ ν < 1 and

λ κ +ε
(11) ≤ ,
1−λν 1 − κμ

as is possible under the assumption that κ μ < 1. Let a, b and c be positive numbers
such that
|σ (y) − σ (y )| ≤ λ |y − y| for y, y ∈ IBν a+b (0),
|e(p, x ) − e(p, x)| ≤ ν |x − x | for x, x ∈ IBa (x̄) and p ∈ IBc ( p̄),

3
Lemma 2B.6 can be stated in a complete metric space X and then it will be equivalent to the
standard formulation of the contraction mapping principle in Theorem 1A.2. There is no point,
of course, in giving a fairly complicated equivalent formulation of a classical result unless, as in
our case, this formulation would bring some insights and dramatically simplify the proofs of later
results.
2 Implicit Function Theorems for Variational Problems 79

where e(p, x) = f (p, x) − h(x), and

(12) | f (p, x̄) − f ( p̄, x̄)| ≤ b for p ∈ IBc ( p̄).

Take b smaller if necessary so that bλ < a(1 − λ ν ), and accordingly adjust c to

ensure having (12). Now apply Lemma 2B.6 with r = σ , M = (h + F)−1 , ȳ = 0 and
ϕ = −e, keeping the rest of the notation the same. It’s straightforward to check that
the estimates in (7) and the conditions (a) and (b) hold for the function in (8). Then,
through the conclusion of Lemma 2B.6 and the observation that

(13) x ∈ (h + F)−1 (−e(p, x)) ⇐⇒ x ∈ S(p),

we obtain that the solution mapping S in (2) has a single-valued localization s around
p̄ for x̄. Due to (11), the inequality in (6) holds for Q = IBc ( p̄). That estimate implies
the continuity of s at p̄, in particular.
From Theorem 2B.5 we obtain a generalization of Theorem 1E.13, the result in
Chapter 1 about implicit functions without differentiability, in which the function f
is replaced now by the sum f + F for an arbitrary set-valued mapping F. The next
statement, 2B.7, covers most of this generalization; the final part of 1E.13 (giving
special consequences when μ = 0) will be addressed in the follow-up statement,
2B.8.

Theorem 2B.7 (implicit function theorem for generalized equations). Consider a

function f : IRd × IRn → IRn and a mapping F : IRn → → IRn with ( p̄, x̄) ∈ int dom f
 ( f ; ( p̄, x̄)) ≤ γ < ∞. Let h be a strict
and f ( p̄, x̄) + F(x̄)  0, and suppose that lip p
estimator of f with respect to x uniformly in p at ( p̄, x̄) with constant μ. Suppose
that (h + F)−1 has a Lipschitz continuous single-valued localization σ around 0 for
x̄ with lip (σ ; 0) ≤ κ for a constant κ such that κ μ < 1. Then the solution mapping
  
S : p → x ∈ IRn  f (p, x) + F(x)  0 for p ∈ IRd

has a Lipschitz continuous single-valued localization s around p̄ for x̄ with

κγ
lip (s; p̄) ≤ .
1 − κμ

For the case of 2B.7 with μ = 0, in which case h is a partial first-order approxima-
tion of f with respect to x at ( p̄, x̄), much more can be said about the single-valued
localization s. The details are presented in the next result, which extends the part of
1E.13 for this case, and with it, Corollaries 2B.2 and 2B.3. We see that, by adding
some relatively mild assumptions about the function f (while still allowing F to
be arbitrary!), we can develop a first-order approximation of the localized solution
mapping s in Theorem 2B.5. This opens the way to obtain differentiability proper-
ties of s, for example.

Theorem 2B.8 (extended implicit function theorem with first-order approxima-

tions). Specialize Theorem 2B.5 to the case where μ = 0 in 2B.5(a), so that h is
80 2 Implicit Function Theorems for Variational Problems

a strict first-order approximation of f with respect to x uniformly in p at ( p̄, x̄).

Then, with the localization σ in 2B.5(b) we have the following additions to the
conclusions of Theorem 2B.5:
(a) If clm p ( f ; ( p̄, x̄)) < ∞ then the single-valued localization s of the solution
mapping S in (2) is calm at p̄ with

(14) clm (s; p̄) ≤ lip (σ ; 0) · clm p ( f ; ( p̄, x̄)).

 ( f ; ( p̄, x̄)) < ∞, then the single-valued localization s of the solution

(b) If lip p
mapping S in (2) is Lipschitz continuous near p̄ with

(15)  ( f ; ( p̄, x̄)).

lip (s; p̄) ≤ lip (σ ; 0) · lip p

(c) If, along with (a), f has a first-order approximation r with respect to p at
( p̄, x̄), then, for Q as in (6), the function η : Q → IRn defined by

(16) η (p) = σ (−r(p) + f ( p̄, x̄)) for p ∈ Q

is a first-order approximation at x̄ to the single-valued localization s.

(d) If, in addition to (b)(c), σ is affine, i.e., σ (y) = x̄ + Ay for some n × m ma-
trix A, and furthermore the first-order approximation r is strict with respect to p
uniformly in x at ( p̄, x̄), then η is a strict first-order approximation of s at p̄ in the
form

(17) η (p) = x̄ + A(−r(p) + f ( p̄, x̄)) for p ∈ Q.

Proof. Let the constants a and c be as in the proof of Theorem 2B.5; then Q =
IBc ( p̄). Let U = IBa (x̄). For p ∈ Q, from (13) we have

(18) s(p) = σ (−e(p, s(p))) for e(p, x) = f (p, x) − h(x)

along with x̄ = s( p̄) = σ (0). Let κ equal lip (σ ; 0) and consider for any ε > 0 the
estimate in (6) with μ = 0. Let p ∈ Q, p = p̄ and p = p̄ in (6) and divide both sides
of (6) by |p − p̄|. Taking the limsup as p → p̄ and ε → 0 gives us (14).
Under the assumptions of (b), in a similar way, by letting p , p ∈ Q, p = p in (6),
and dividing both sides of (6) by |p − p| and passing to the limit, we obtain (15).
Observe that (15) follows directly from 2B.7.
Consider now any λ > clm (s; p̄) and ε > 0. Make the neighborhoods Q and U
smaller if necessary so that for all p ∈ Q and x ∈ U we have |s(p)− s( p̄)| ≤ λ |p − p̄|
and

(19) |e(p, x) − e(p, x̄)| ≤ ε |x − x̄|, | f (p, x̄) − r(p)| ≤ ε |p − p̄|,

and furthermore so that the points −e(p, x) and −r(p) + f ( p̄, x̄) are contained in a
neighborhood of 0 on which the function σ is Lipschitz continuous with Lipschitz
constant κ + ε = lip (σ ; 0) + ε . Then, for p ∈ Q, we get by way of (18), along with
the first inequality in (19) and the fact that e( p̄, x̄) = 0, the estimate that
2 Implicit Function Theorems for Variational Problems 81

|s(p) − η (p)| = |s(p) − σ (−r(p) + f ( p̄, x̄))|

= |σ (−e(p, s(p))) − σ (−r(p) + f ( p̄, x̄))|
≤ (κ + ε )(| − e(p, s(p)) + e(p, x̄)| + | f (p, x̄) − r(p)|)
≤ (κ + ε )ε |s(p) − x̄| + (κ + ε )ε |p − p̄| ≤ ε (κ + ε )(λ + 1)|p − p̄|.

Since ε can be arbitrarily small and also s( p̄) = x̄ = σ (0) = η ( p̄), the function η
defined in (16) is a first-order approximation of s at p̄.
Moving on to part (d) of the theorem, suppose that the assumptions in (b)(c) are
satisfied and also σ (y) = x̄ + Ay. Again, choose any ε > 0 and further adjust the
neighborhoods Q of p̄ and U of x̄ so that

|e(p, x) − e(p, x )| ≤ ε |x − x | for all x, x ∈ U and p ∈ Q,

(20)
| f (p , x) − r(p ) − f (p, x) + r(p)| ≤ ε |p − p| for all x ∈ U and p , p ∈ Q,
  

and moreover s(p) ∈ U for p ∈ Q. By part (b), the single-valued localization s is

Lipschitz continuous near p̄; let λ > lip (s; p̄) and shrink Q even more if necessary
so as to ensure that s is Lipschitz continuous with constant λ on Q. For p, p ∈ Q,
using (17), (18) and (20), we obtain

|s(p) − s(p ) − η (p) + η (p)| = |s(p) − s(p ) − A(−r(p) + r(p))|

= |A(−e(p, s(p)) + e(p , s(p )) + r(p) − r(p))|
≤ |A|| − e(p, s(p)) + e(p, s(p))|
+|A|| f (p , s(p )) − r(p ) − f (p, s(p )) + r(p)|
≤ |A|(ε |s(p) − s(p )| + ε |p − p|)
≤ |A|ε (λ + 1)|p − p|.

Since ε can be arbitrarily small, we see that the first-order approximation of s fur-
nished by η is strict, and the proof is complete.

Note that the assumption in part (d), that the localization σ of G−1 = (h + F)−1
around 0 for x̄ is affine, can be interpreted as a sort of differentiability condition on
G−1 at 0 with A giving the derivative mapping.

Corollary 2B.9 (utilization of strict differentiability). Suppose in the generalized

equation (1) with solution mapping S given by (2), that x̄ ∈ S( p̄) and f is strictly
differentiable at ( p̄, x̄). Assume that the inverse G−1 of the mapping

G(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄) + F(x), with G(x̄)  0,

has a Lipschitz continuous single-valued localization σ around 0 for x̄. Then not
only do the conclusions of Theorem 2B.5 hold for a solution localization s, but also
there is a first-order approximation η to s at p̄ given by
 
η (p) = σ − ∇ p f ( p̄, x̄)(p − p̄) .
82 2 Implicit Function Theorems for Variational Problems

Moreover, if F ≡ 0, then the first-order approximation η is strict and given by

(21) η (p) = x̄ − ∇x f ( p̄, x̄)−1 ∇ p f ( p̄, x̄)(p − p̄),

so that s is strictly differentiable at p̄.

Proof. In this case Theorem 2B.8 is applicable with h taken to be the linearization
of f ( p̄, ·) at x̄ and r taken to be the linearization of f (·, x̄) at p̄. When F ≡ 0, we
get σ (y) = x̄ + ∇x f ( p̄, x̄)−1 y, so that η as defined in (16) achieves the form in (21).
Having a strict first-order approximation by an affine function means strict differen-
tiability.

The second part of Corollary 2B.9 shows how the implicit function theorem for
equations as stated in Theorem 1D.13 is covered as a special case of Theorem 2B.7.
In the case of the generalized equation (1) where f (p, x) = g(x)− p for a function
g : IRn → IRm (d = m), so that
  
(22) S(p) = x  p ∈ g(x) + F(x) = (g + F)−1 (p),

the inverse function version of Theorem 2B.8 has the following symmetric form.

Theorem 2B.10 (inverse version). In the framework of the solution mapping (22),
consider any pair ( p̄, x̄) with x̄ ∈ S( p̄). Let h be any strict first-order approximation
to g at x̄. Then (g + F)−1 has a Lipschitz continuous single-valued localization s
around p̄ for x̄ if and only if (h + F)−1 has such a localization σ around p̄ for x̄, in
which case σ is a first-order approximation of s at p̄ and

(23) lip (s; p̄) = lip (σ ; p̄).

If, in addition, σ is affine, σ (y) = x̄ + Ay, then s is strictly differentiable at p̄ with

Ds( p̄) = A.
Proof. For the “if” part, suppose that (h + F)−1 has a localization σ as described.
Then, from (15) with f (p, x) = −p + g(x) we get lip (s; p̄) ≤ lip (σ ; 0). The “only if”
part is completely analogous because g and h play symmetric roles in the statement,
and yields lip (σ ; p̄) ≤ lip (s ; 0) for some single-valued localization s of S. The
localizations s and s have to agree graphically around (0, x̄), so we pass to a smaller
localization, again called s, and get the equality in (23). Through the observation
that r(p) = g(x̄) − p + p̄, the rest follows from Theorem 2B.8(d).
We also can modify the results presented so far in this section in the direction
indicated in Section 1F, where we considered local selections instead of single-
valued localizations. We state such a result here as an exercise.
  
Exercise 2B.11 (implicit selections). Let S = x ∈ IRn  f (p, x) + F(x)  0 for a
function f : IRd × IRn → IRm and a mapping F : IRn →
→ IRm , along with a pair ( p̄, x̄)
2 Implicit Function Theorems for Variational Problems 83

 ( f ; ( p̄, x̄)) ≤ γ < ∞. Let h be a strict first-

such that x̄ ∈ S( p̄), and suppose that lip p
order approximation of f with respect to x at ( p̄, x̄) for which (h + F)−1 has a Lip-
schitz continuous local selection σ around 0 for x̄ with lip (σ ; 0) ≤ κ. Then S has a
Lipschitz continuous local selection s around p̄ for x̄ with

lip (s; p̄) ≤ κγ .

If in addition f has a first-order approximation r with respect to p at ( p̄, x̄), then

there exists a neighborhood Q of p̄ such that the function

η : p → σ (−r(p) + f ( p̄, x̄)) for p ∈ Q

is a first-order approximation of s at p̄.

Guide. First verify the following statement, which is a simple modification of
Lemma 2B.6. For a function ϕ : IRd × IRn → IRm and a point ( p̄, x̄) ∈ int dom ϕ , let
the nonnegative scalars ν , b, the positive scalar a, and the set Q ⊂ IRd be such that
p̄ ∈ Q and the conditions (7) hold. Consider also a set-valued mapping M : IRm → → IRn
with (ȳ, x̄) ∈ gph M, where ȳ := ϕ ( p̄, x̄), and assume that there exists a Lipschitz
continuous function r on IBν a+b (ȳ) such that

r(y) ∈ M(y) ∩ IBa (x̄) for y ∈ IBν a+b (ȳ) and r(ȳ) = x̄.

In addition, suppose now that the Lipschitz constant λ for the function r is such that
the conditions (a) and
 (b) in the  statement of Lemma
 2B.6 are fulfilled. Then for
each p ∈ Q the set x ∈ IBa (x̄)  x ∈ M(ϕ (p, x)) contains a point s(p) such that the
function p → s(p) satisfies s( p̄) = x̄ and

λ
(24) |s(p ) − s(p)| ≤ |ϕ (p , s(p)) − ϕ (p, s(p))| for all p , p ∈ Q.
1−λν
  
Thus, the mapping N := p → x  x ∈ M(ϕ (p, x)) ∩ IBa (x̄) has a local selection
s around p̄ for x̄ which satisfies (24). The difference from Lemma 2B.6 is that r is
now required only to be a local selection of the mapping M with specified properties,
and then we obtain a local selection s of N at p̄ for x̄. For the rest use the proofs of
Theorems 2B.5 and 2B.8.

2C. Ample Parameterization and Parametric Robustness

The results in 2B, especially the broad generalization of Robinson’s theorem in 2B.5
and its complement in 2B.8 dealing with solution approximations, provide a sub-
stantial extension of the classical theory of implicit functions. Equations have been
replaced by generalized equations, with variational inequalities as a particular case,
84 2 Implicit Function Theorems for Variational Problems

and technical assumptions about differentiability have been greatly relaxed. Much
of the rest of this chapter will be concerned with working out the consequences in
situations where additional structure is available. Here, however, we reflect on the
ways that parameters enter the picture and the issue of whether there are “enough”
parameters, which emerges as essential in drawing good conclusions about solution
mappings.
The differences in parameterization between an inverse function theorem and an
implicit function theorem are part of a larger pattern which deserves, at this stage, a
closer look. Let’s start by considering a generalized equation without parameters,

(1) g(x) + F(x)  0,

for a function g : IRn → IRm and a set-valued mapping F : IRn →

→ IRm . We can think
of a parameterization as the choice of a function

(2) f : IRd × IRn → IRm having f ( p̄, x) ≡ g(x) for a particular p̄ ∈ IRd .

The specification of such a parameterization leads to an associated solution mapping

  
(3) S : p → x  f (p, x) + F(x)  0 ,

which we proceed to study around p̄ and a point x̄ ∈ S( p̄) for the presence of a nice
localization σ . Different parameterizations yield different solution mappings, which
may possess different properties according to the assumptions placed on f .
That’s the general framework, but the special kind of parameterization that corre-
sponds to the “inverse function” case has a fundamental role which is worth trying to
understand more fully. In that case, we simply have f (p, x) = g(x) − p in (2), so that
in (1) we are solving g(x) + F(x)  p and the solution mapping is S = (g + F)−1 .
Interestingly, this kind of parameterization comes up even in obtaining “implicit
function” results through the way that approximations are utilized. Recall that in
Theorem 2B.5, for a function h which is “close” to f ( p̄, ·) near x̄, the mapping
(h + F)−1 having x̄ ∈ (h + F)−1 (0) is required to have a Lipschitz continuous single-
valued localization around 0 for x̄. Only then are we able to deduce that the solution
mapping S in (3) has a localization of such type at p̄ for x̄. In other words, the de-
sired conclusion about S is obtained from an assumption about a simpler solution
mapping in the “inverse function” category.
When S itself already belongs to that category, because f (p, x) = g(x) − p and
S = (g + F)−1 , another feature of the situation emerges. Then, as seen in Theorem
2B.10, the assumption made about (h + F)−1 is not merely sufficient for obtain-
ing the desired localization of S but actually necessary. This distinction was already
observed in the classical setting. In the “symmetric” version of the inverse func-
tion theorem in 1C.3, the invertibility of a linearized mapping is both necessary and
sufficient for the conclusion, whereas such invertibility acts only as a sufficient con-
dition in the implicit function theorem 1B.2 (even though that theorem and the basic
version of the inverse function theorem 1A.1 are equivalent to each other).
2 Implicit Function Theorems for Variational Problems 85

An implicit function theorem for generalized equations that exhibits necessity

as well as sufficiency in its assumptions can nonetheless be derived. For this, the
parameterization must be “rich” enough.

Ample parameterization. A parameterization of the generalized equation (1) as

in (2) will be called ample at x̄ if f (p, x) is strictly differentiable with respect to p
uniformly in x at ( p̄, x̄) and the partial Jacobian ∇ p f ( p̄, x̄) is of full rank:

(4) rank ∇ p f ( p̄, x̄) = m.

The reason why the rank condition in (4) can be interpreted as ensuring the rich-
ness of the parameterization is that it can always be achieved through supplementary
parameters. Any parameterization function f having the specified strict differentia-
bility can be extended to a parameterization function f˜ with

(5) f˜(q, x) = f (p, x) − y, q = (p, y), q̄ = ( p̄, 0),

which does satisfy the ampleness condition, since trivially rank ∇q f˜(q̄, x̄) = m. The
generalized equation being solved then has solution mapping
  
(6) S : (p, y) → x  f (p, x) + F(x)  y .

Results about localizations of S can be specialized to results about S by taking y = 0.

In order to arrive at the key result about ample parameterization, asserting an
equivalence about the existence of several kinds of localizations, we need a lemma
about local selections which is related to the results presented in Section 1F.

Lemma 2C.1. Let f : IRd × IRn → IRm with ( p̄, x̄) ∈ int dom f afford an ample pa-
rameterization of the generalized equation (1) at x̄. Suppose that f has a strict first-
order approximation h : IRn → IRm with respect to x uniformly in p at ( p̄, x̄). Then
the mapping
  
(7) Ψ : (x, y) → p  e(p, x) + y = 0 for (x, y) ∈ IRn × IRm ,

where e(p, x) = f (p, x) − h(x), has a local selection ψ around (x̄, 0) for p̄ which
satisfies

(8a)  (ψ ; (x̄, 0)) = 0

lip x

and

(8b)  (ψ ; (x̄, 0)) < ∞.

lip y

Proof. Let A = ∇ p f ( p̄, x̄); then AAT is invertible. Without loss of generality, sup-
pose x̄ = 0, p̄ = 0, and f (0, 0) = 0; then h(0) = 0. Let c = |AT (AAT )−1 |. Let
0 < ε < 1/(2c) and choose a positive a such that for all x, x ∈ aIB and p, p ∈ aIB
we have
86 2 Implicit Function Theorems for Variational Problems

(9) |e(p, x ) − e(p, x)| ≤ ε |x − x |

and

(10) | f (p, x) − f (p , x) − A(p − p)| ≤ ε |p − p|.

For b = a(1 − 2cε )/c, fix x ∈ aIB and y ∈ bIB, and consider the mapping

Φx,y : p → −AT (AAT )−1 (e(p, x) + y − Ap) for p ∈ aIB.

Through (9) and (10), keeping in mind that e(0, 0) = 0, we see that

|Φx,y (0)| ≤ c|e(0, x) + y| ≤ c|e(0, x) − e(0, 0)| + c|y| ≤ cε a + cb = a(1 − cε ),

and for any p, p ∈ aIB

|Φx,y (p) − Φx,y (p )| ≤ c| f (p, x) − f (p , x) − A(p − p)| ≤ cε |p − p |.

The contraction mapping principle 1A.2 then applies, and we obtain from it the
existence of a unique p ∈ aIB such that

(11) p = −AT (AAT )−1 (e(p, x) + y − Ap).

We denote by ψ (x, y) the unique solution in aIB of this equation for x ∈ aIB and
y ∈ bIB. Multiplying both sides of (11) by A and simplifying, we get e(p, x) + y = 0.
This means that for each (x, y) ∈ aIB × bIB the equation e(p, x) + y = 0 has ψ (x, y)
as a solution. From (11), we know that

(12) ψ (x, y) = −AT(AAT )−1 ( f (ψ (x, y), x) − h(x) + y − Aψ (x, y)).

Let x, x ∈ aIB and y, y ∈ bIB. Using (9) and (10) we have

|ψ (x, y) − ψ (x , y)| ≤ c|e(ψ (x, y), x) − e(ψ (x, y), x )|

+c| f (ψ (x, y), x ) − f (ψ (x , y), x ) − A(ψ (x, y) − ψ (x, y))|
≤ cε |x − x | + cε |ψ (x, y) − ψ (x, y)|.

Hence
cε
|ψ (x, y) − ψ (x , y)| ≤
|x − x |.
1 − cε
Since ε can be arbitrarily small, we conclude that (8a) holds. Analogously, from
(12) and using again (9) and (10) we obtain

|ψ (x, y) − ψ (x, y )|
≤ c| f (ψ (x, y), x) + y − Aψ (x, y) − f (ψ (x, y ), x) − y + Aψ (x, y )|
≤ c|y − y | + cε |ψ (x, y) − ψ (x, y)|,

and then
2 Implicit Function Theorems for Variational Problems 87
c
|ψ (x, y) − ψ (x, y )| ≤ |y − y|
1 − cε
which gives us (8b).

We are now ready to present the first main result of this section:

Theorem 2C.2 (equivalences from ample parameterization). Let f parameterize

the generalized equation (1) as in (2). Suppose the parameterization is ample at
( p̄, x̄), and let h be a strict first-order approximation of f with respect to x uniformly
in p at ( p̄, x̄). Then the following properties are equivalent.
(a) S in (3) has a Lipschitz continuous single-valued localization around p̄ for x̄.
(b) (h + F)−1 has a Lipschitz continuous single-valued localization around 0
for x̄.
(c) (g + F)−1 has a Lipschitz continuous single-valued localization around 0
for x̄.
(d) S in (6) has a Lipschitz continuous single-valued localization around ( p̄, 0)
for x̄.
Proof. If the mapping (h + F)−1 has a Lipschitz continuous single-valued local-
ization around 0 for x̄, then from Theorem 2B.5 together with Theorem 2B.10 we
may conclude that the other three mappings likewise have such localizations at the
respective reference points. In other words, (b) is sufficient for (a) and (d). Also, (b)
is equivalent to (c) inasmuch as g and h are first-order approximations to each other
(Theorem 2B.10). Since (d) implies (a), the issue is whether (b) is necessary for (a).
Assume that (a) holds with a Lipschitz localization s around p̄ for x̄ and choose
λ > lip (s; p̄). Let ν > 0 be such that λ ν < 1, and consider a Lipschitz continuous
local selection ψ of the mapping Ψ in Lemma 2C.1. Then there exist positive a, b
and c such that λ ν a + λ b < a,

S(p) ∩ IBa (x̄) = s(p) for p ∈ IBν a+b ( p̄),

|s(p) − s(p )| ≤ λ |p − p| for p, p ∈ IBν a+b ( p̄),

h(x) − f (ψ (x, y), x) = y for y ∈ IBc (0), x ∈ IBa (x̄),
|ψ (x, y) − ψ (x , y)| ≤ ν |x − x | for x, x ∈ IBa (x̄) and y ∈ IBc (0),
the last from (8a), and

|ψ (y, x̄) − ψ (0, x̄)| ≤ b for y ∈ IBc (0).

We now apply Lemma 2B.6 with ϕ (p, x) = ψ (x, y) for p = y and M(p) = S(p),
thereby obtaining that the mapping
  
IBc (0)  y → x ∈ IBa (x̄)  x ∈ S(ψ (x, y))

is a function which is Lipschitz continuous on IBc (0). Noting that

(h + F)−1 (y) ∩ IBa (x̄) = {x | x = S(ψ (x, y)) ∩ IBa (x̄)},

88 2 Implicit Function Theorems for Variational Problems

we conclude that (h + F)−1 has a Lipschitz continuous single-valued localization

around 0 for x̄. Thus, (a) implies (b).
The strict differentiability property with respect to p which is assumed in the
definition of ample parameterization is satisfied of course when f is strictly differ-
entiable with respect to (p, x) at ( p̄, x̄). Then, moreover, the linearization of f ( p̄, ·)
at x̄, which is the same as the linearization of g at x̄, can be taken as the function h.
This leads to a statement about an entire class of parameterizations.

Theorem 2C.3 (parametric robustness). Consider the generalized equation (1) un-
der the assumption that x̄ is a point where g is strictly differentiable. Let h(x) =
g(x̄) + ∇g(x̄)(x − x̄). Then the following statements are equivalent.
(a) (h + F)−1 has a Lipschitz continuous single-valued localization around 0
for x̄.
(b) For every parameterization (2) in which f is strictly differentiable at (x̄, p̄),
the mapping S in (3) has a Lipschitz continuous single-valued localization around p̄
for x̄.
Proof. The implication from (a) to (b) already follows from Theorem 2B.8. The
focus is on the reverse implication. This is valid because, among the parameteriza-
tions covered by (b), there will be some that are ample. For instance, one could pass
from a given one to an ample parameterization in the mode of (5). For the solution
mapping for such a parameterization, we have the implication from (a) to (b) in
Theorem 2C.2. That specializes to what we want.
A solution x̄ to the generalized equation (1) is said to be parametrically robust
when the far-reaching property in (b) of Theorem 2C.3 holds. In that terminology,
Theorem 2C.3 gives a criterion for parametric robustness.

2D. Semidifferentiable Functions

The notion of a first-order approximation of a function at a given point has already

served us for various purposes as a substitute for differentiability, where the ap-
proximation is a linearization. We now bring in an intermediate concept in which
linearity is replaced by positive homogeneity.
A function ϕ : IRn → IRm is positively homogeneous if 0 = ϕ (0) and ϕ (λ w) =
λ ϕ (w) for all w ∈ dom ϕ and λ > 0. These conditions mean geometrically that
the graph of ϕ is a cone in IRn × IRm . A linear function is positively homogeneous
in particular, of course. The graph of a linear function ϕ : IRn → IRm is actually a
subspace of IRn × IRm .
2 Implicit Function Theorems for Variational Problems 89

Semiderivatives. A function f : IRn → IRm is said to be semidifferentiable4 at x̄

if it has a first-order approximation at x̄ of the form h(x) = f (x̄) + ϕ (x − x̄) with
ϕ continuous and positively homogeneous; when the approximation is strict, f is
strictly semidifferentiable at x̄. Either way, the function ϕ, necessarily unique, is
called the semiderivative of f at x̄ and denoted by D f (x̄), so that h(x) = f (x̄) +
D f (x̄)(x − x̄).
In a first-order approximation we have by definition that clm ( f − h)(x̄) = 0,
which in the “strict” case is replaced by lip ( f − h)(x̄) = 0. The uniqueness of the
semiderivative, when it exists, comes from the fact that any two first-order approxi-
mations f (x̄)+ ϕ (x− x̄) and f (x̄)+ ψ (x− x̄) of f at x̄ must have clm (ϕ − ψ )(0) = 0,
and under positive homogeneity that cannot hold without having ϕ = ψ . The unique-
ness can also be gleaned through comparison with directional derivatives.

One-sided directional derivatives. For f : IRn → IRm , a point x̄ ∈ dom f and a

vector w ∈ IRn , the limit
f (x̄ + tw) − f (x̄)
(1) f  (x̄; w) = lim ,
t 0 t

when it exists, is the (one-sided) directional derivative of f at x̄ for w; here t  0

means that t → 0 with t > 0. If this directional derivative exists for every w, f is said
to be directionally differentiable at x̄.
Note that f  (x̄; w) is positively homogeneous in the w argument. This comes out
of the limit definition itself. Directional differentiability is weaker than semidiffer-
entiability in general, but equivalent to it in the presence of Lipschitz continuity, as
we demonstrate next.

Proposition 2D.1. If a function f : IRn → IRm is semidifferentiable at x̄, then f is in

particular directionally differentiable at x̄ and has

(2) D f (x̄)(w) = f  (x̄; w) for all w,

so that the first-order approximation in the definition of semidifferentiability has the

form
h(x) = f (x̄) + f  (x̄; x − x̄).
When lip ( f ; x̄) < ∞, directional differentiability at x̄ in turn implies semidifferentia-
bility at x̄.
Proof. Having clm ( f − h)(x̄) = 0 for h(x) = f (x̄) + ϕ (x − x̄) as in the definition
of semidifferentiability entails having [ f (x̄ + tw) − h(x̄ + tw)]/t → 0 as t  0 with
h(x̄ + tw) = f (x̄) + t ϕ (w). Thus, ϕ (w) must be f  (x̄; w).
For the converse claim, consider λ > lip ( f ; x̄) and observe that for any u and v,

4 Also called Bouligand differentiable or B-differentiable under the additional assumption of Lip-
schitz continuity, see the book Facchinei and Pang [2003].
90 2 Implicit Function Theorems for Variational Problems

1
(3) | f  (x̄; u) − f  (x̄; v)| = lim | f (x̄ + tu) − f (x̄ + tv)| ≤ λ |u − v|.
t→0 t
t>0

Next, consider an arbitrary sequence uk → 0 and, without loss of generality, assume

that uk /|uk | → ū with |ū| = 1. Letting tk = |uk | and using the positive homogeneity
of the directional derivative, we obtain
1
0≤ | f (x̄ + uk ) − f (x̄) − f  (x̄; uk )|
|uk |

1
≤ | f (x̄ + uk ) − f (x̄ + tk ū)| + | f  (x̄;tk ū) − f  (x̄; uk )|
tk


+| f (x̄ + tk ū) − f (x̄) − f (x̄;tk ū)|

uk 1
≤ 2λ | − ū| + | ( f (x̄ + tk ū) − f (x̄)) − f  (x̄; ū)|,
tk tk
where in the final inequality we invoke (3). Since uk is arbitrarily chosen, we con-
clude by passing to the limit as k → ∞ that for h(x) = f (x̄) + f  (x̄; x − x̄) we do have
clm ( f − h; x̄) = 0.
When the semiderivative D f (x̄) : IRn → IRm is linear, semidifferentiability turns
into differentiability, and strict semidifferentiability turns into strict differentiability.
The connections known between D f (x̄) and the calmness modulus and Lipschitz
modulus of f at x̄ under differentiability can be extended to semidifferentiability by
adopting the definition that

|ϕ | = sup |ϕ (x)| for a positively homogeneous function ϕ .

|x|≤1

We then have clm (D f (x̄); 0) = |D f (x̄)| and consequently clm ( f ; x̄) = |D f (x̄)|,
which in the case of strict semidifferentiability becomes lip ( f ; x̄) = |D f (x̄)|. Thus
in particular, semidifferentiability of f at x̄ implies that clm ( f ; x̄) < ∞, while strict
semidifferentiability at x̄ implies that lip ( f ; x̄) < ∞.

Exercise 2D.2 (alternative characterization of semidifferentiability). For a function

f : IRn → IRm and a point x̄ ∈ dom f , semidifferentiability is equivalent to the exis-
tence for every w ∈ IRn of

f (x̄ + tw ) − f (x̄)

(4) lim .
t 0 t
w →w

Guide. Directional differentiability of f at x̄ corresponds to the difference quotient

functions Δt (w) = [ f (x̄ + tw) − f (x̄)]/t converging pointwise to something, namely
f  (x̄; ·), as t  0. Show that the existence of the limits in (4) means that these func-
tions converge to f  (x̄; ·) not just pointwise, but uniformly on bounded sets. Glean
2 Implicit Function Theorems for Variational Problems 91

from that the equivalence with having a first-order approximation as in the definition
of semidifferentiability.

Examples.
1) The function f (x) = e|x| for x ∈ IR is not differentiable at 0, but it is semidif-
ferentiable there and its semiderivative is given by D f (0) : w → |w|. This is actually
a case of strict semidifferentiability. Away from 0, f is of course continuously dif-
ferentiable (hence strictly differentiable).
2) The function f (x1 , x2 ) = min{x1 , x2 } on IR2 is continuously differentiable at
every point away from the line where x1 = x2 . On that line, f is strictly semidiffer-
entiable with
D f (x1 , x2 )(w1 , w2 ) = min{w1 , w2 }.
3) A function of the form f (x) = max{ f1 (x), f2 (x)}, with f1 and f2 continuously
differentiable from IRn to IR, is strictly differentiable at all points x where f1 (x) =
f2 (x) and semidifferentiable where f1 (x) = f2 (x), the semiderivative being given
there by
D f (x)(w) = max{D f1 (x)(w), D f2 (x)(w)}.
However, f might not be strictly semidifferentiable at such points; see Example
2D.5 below.
The semiderivative obeys standard calculus rules, such as semidifferentiation of
a sum, product and ratio, and, most importantly, the chain rule. We pose the verifi-
cation of these rules as exercises.

Exercise 2D.3. Let f be semidifferentiable at x̄ and let g be Lipschitz continuous

and semidifferentiable at ȳ := f (x̄). Then g◦ f is semidifferentiable at x̄ and

D(g◦ f )(x̄) = Dg(ȳ)◦D f (x̄).

Guide. Apply Proposition 1E.1 and observe that a composition of positively homo-
geneous functions is positively homogeneous.

Exercise 2D.4. Let f be strictly semidifferentiable at x̄ and g be strictly differen-

tiable at f (x̄). Then g◦ f is strictly semidifferentiable at x̄.
Guide. Apply 1E.2.

Example 2D.5. The functions f and g in Exercise 2D.4 cannot exchange places: the
composition of a strictly semidifferentiable function with a strictly differentiable
function is not always strictly semidifferentiable. For a counterexample, consider
the function f : IR2 → IR given by

f (x1 , x2 ) = min{x31 , x2 } for (x1 , x2 ) ∈ IR2 .

92 2 Implicit Function Theorems for Variational Problems

According to 2D.3, the function f is semidifferentiable at (0, 0) with semiderivative

D f (0, 0)(w1 , w2 ) = min{0, w2 }. To see that f is not strictly semidifferentiable at
(0, 0), however, observe for the function g = f − D f (0, 0) that

|g(x1 , x2 ) − g(x1, x2 )| 1

= for (x1 , x2 ) = (−ε , −ε 3 /2) and (x1 , x2 ) = (−ε , ε 4 ).
|(x1 , x2 ) − (x1 , x2 )| 1 + 2ε

As ε goes to 0 this ratio tends to 1, and therefore lip ( f − D f (0, 0); (0, 0)) ≥ 1.
Our aim now is to forge out of Theorem 2B.8 a result featuring semideriva-
tives. For this purpose, we note that if f (p, x) is (strictly) semidifferentiable at ( p̄, x̄)
jointly in its two arguments, it is also “partially (strictly) semidifferentiable” in these
arguments separately. In denoting the semiderivative of f ( p̄, ·) at x̄ by Dx f ( p̄, x̄) and
the semiderivative of f (·, x̄) at p̄ by D p f ( p̄, x̄), we have

Dx f ( p̄, x̄)(w) = D f ( p̄, x̄)(0, w), D p f ( p̄, x̄)(q) = D f ( p̄, x̄)(q, 0).

In contrast to the situation for differentiability, however, D f ( p̄, x̄)(q, w) isn’t neces-
sarily the sum of these two partial semiderivatives.

Theorem 2D.6 (implicit function theorem utilizing semiderivatives). Let x̄ ∈ S( p̄)

for the solution mapping
  
S : p → x ∈ IRn  f (p, x) + F(x)  0

associated with a choice of F : IRn → → IRm and f : IRd × IRn → IRm such that f is
strictly semidifferentiable at ( p̄, x̄). Suppose that the inverse G−1 of the mapping

G(x) = f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄) + F(x), with G(x̄)  0,

has a Lipschitz continuous single-valued localization σ around 0 for x̄ which is

semidifferentiable at 0. Then S has a Lipschitz continuous single-valued localization
s around p̄ for x̄ which is semidifferentiable at p̄ with its semiderivative given by

Ds( p̄) = Dσ (0)◦(−D p f ( p̄, x̄)).

Proof. First, note that s( p̄) = σ (0) = x̄ and the function r in Theorem 2B.8 may be
chosen as r(p) = f (x̄, p̄) + D p f ( p̄, x̄)(p − p̄). Then we have

|s(p) − s( p̄) − (Dσ (0)◦(−D p f ( p̄, x̄)))(p − p̄)| ≤ |s(p) − σ (−r(p) + r( p̄))|
+|σ (−D p f ( p̄, x̄)(p − p̄)) − σ (0) − Dσ (0)(−D p f ( p̄, x̄)(p − p̄))|.

According to Theorem 2B.8 the function p → σ (−r(p) + r( p̄)) is a first-order ap-

proximation to s at p̄, hence the first term on the right side of this inequality is of
order o(|p − p̄|) when p is close to p̄. The same is valid for the second term, since σ
is assumed to be semidifferentiable at 0. It remains to observe that the composition
of positively homogeneous mappings is positively homogeneous.
2 Implicit Function Theorems for Variational Problems 93

An important class of semidifferentiable functions will be brought in next.

Piecewise smooth functions. A function f : IRn → IRm is said to be piecewise

smooth on an open set O ⊂ dom f if it is continuous on O and for each x ∈ O there
is a finite collection { fi }i∈I of smooth (C 1 ) functions defined on a neighborhood of
x such that, for some ε > 0, one has
  
(5) f (y) ∈ fi (y)  i ∈ I when |y − x| < ε .
  
The collection { fi }i∈I(x) , where I(x) = i ∈ I  f (x) = fi (x) , is said then to furnish
a local representation of f at x. A local representation in this sense is minimal if no
proper subcollection of it forms a local representation of f at x.
Note that a local representation of f at x characterizes f on a neighborhood of x,
and minimality means that this would be lost if any of the functions fi were dropped.
The piecewise smoothness terminology finds its justification in the following
observation.
Proposition 2D.7 (decomposition of piecewise smooth functions). Let f be piece-
wise smooth on an open set O with a minimal local representation { fi }i∈I at a
point x̄ ∈ O. Then for each i ∈ I(x̄) there is an open set Oi such that x̄ ∈ cl Oi and
f (x) = fi (x) on Oi .
Proof. Let ε > 0 be as in (5) with x = x̄ and  assume that IBε (x̄) ⊂ O. For each
i ∈ I(x̄), let Ui = x ∈ intIBε (x̄)  f (x) = fi (x) and Oi = intIBε (x̄) \ ∪ j=iU j . Because
f and f i are continuous, Ui is closed relative to intIBε (x̄) and therefore Oi is open.
Furthermore x̄ ∈ cl Oi , for if not, the set ∪ j=iU j would cover a neighborhood of
x̄, and then fi would be superfluous in the local representation, thus contradicting
minimality.
It’s not hard to see from this fact that a piecewise smooth function on an open
set O must be continuous on O and even locally Lipschitz continuous, since each
of the C 1 functions fi in a local representation is locally Lipschitz continuous, in
particular. Semidifferentiability in this situation takes only a little more effort to
confirm.
Proposition 2D.8 (semidifferentiability of piecewise smooth functions). If a func-
tion f : IRn → IRm is piecewise smooth on an open set O ⊂ dom f , then f is semidif-
ferentiable on O. Furthermore, the semiderivative function D f (x̄) at any point x̄ ∈ O
is itself piecewise smooth, in fact with local representation composed by the linear
functions {D fi (x̄)}i∈I(x̄) when f has local representation { fi }i∈I around x̄.
Proof. We apply the criterion in 2D.2. Consider, for any x̄ ∈ O and w ∈ IRn , se-
quences tk  0 and wk → w. Invoke a local representation as in (5). Note that be-
cause f (x̄ + tk wk ) → f (x̄), we must eventually have I(x̄ + tk wk ) ⊂ I(x̄), so that
the difference quotient Δ k = [ f (x̄ + tk wk ) − f (x̄)]/tk coincides with the difference
quotient Δki = [ fi (x̄ + tk wk ) − fi (x̄)]/tk for some i ∈ I(x̄). Since I(x̄) is finite, ev-
ery subsequence of {Δk }∞ k=1 must have a subsubsequence coinciding with a sub-
sequence of {Δki }∞ k=1 for some i ∈ I(x̄) and therefore converging to ∇ fi (x̄)·w for
94 2 Implicit Function Theorems for Variational Problems

that i. A sequence with the property that every subsequence contains a conver-
gent subsubsequence necessarily converges as a whole. Thus the sequence {Δ k }∞ k=1
converges with its limit equaling ∇ fi (x̄)·w for at least one i ∈ I(x̄). This con-
firms
 semidifferentiability
  and establishes that the semiderivative is a selection from
∇ fi (x̄)·w  i ∈ I(x̄) .
The functions in the examples given after 2D.2 are not only semidifferentiable
but also piecewise smooth. Of course, a semidifferentiable function does not have to
be piecewise smooth, e.g., when it is a selection of infinitely many, but not finitely
many, smooth functions.
A more elaborate example of a piecewise smooth function is the projection map-
ping PC on a nonempty, convex and closed set C ⊂ IRn specified by finitely many
inequalities.

Exercise 2D.9 (piecewise smoothness of special projection mappings). For a con-

vex set C of the form
  
C = x ∈ IRn  gi (x) ≤ 0, i = 1, . . . , m

for convex functions gi of class C 2 on IRn , let x̄ be a point of C at which the gra-
dients ∇gi (x̄) associated with the active constraints, i.e., the ones with gi (x̄) = 0,
are linearly independent. Then there is an open neighborhood O of x̄ such that the
projection mapping PC is piecewise smooth on O.
Guide. Since in a sufficiently small neighborhood of x̄ the inactive constraints re-
main inactive, one can assume without loss of generality that gi (x̄) = 0 for all
i = 1, . . . , m. Recall that because C is nonempty, closed and convex, PC is a Lipschitz
continuous function from IRn onto C (see 1D.5). For each u around x̄ the projection
PC (u) is the unique solution to the problem of minimizing 12 |x − u|2 in x subject to
gi (x) ≤ 0 for i = 1, . . . , m. The associated Lagrangian variational inequality (Theo-
rem 2A.9) tells us that when u belongs to a small enough neighborhood of x̄, the
point x solves the problem if and only if x is feasible and there is a subset J of the
index set {1, 2, . . ., m} and Lagrange multipliers yi ≥ 0, i ∈ J, such that

x + ∑i∈J yi ∇gi (x)T = u,
(6)
gi (x) = 0, i ∈ J.

The linear independence of the gradients of the active constraint gradients yields
that the Lagrange multiplier vector y is unique, hence it is zero for u = x = x̄. For
each fixed subset J of the index set {1, 2, . . . , m} the Jacobian of the function on the
left of (6) at (x̄, 0) is
 
In + ∑i∈J yi ∇2 gi (x̄) ∇gJ (x̄)T
Q= ,
∇gJ (x̄) 0

where
2 Implicit Function Theorems for Variational Problems 95
 
∂ gi
∇gJ (x̄) = (x̄) and In is the n × n identity matrix.
∂xj i∈J, j∈{1,...,n}

Since ∇gJ (x̄) has full rank, the matrix Q is nonsingular and then we can apply the
classical inverse function theorem (Theorem 1A.1) to the equation (6), obtaining
that its solution mapping u → (xJ (u), yJ (u)) has a smooth single-valued 
localization

around u = x̄ for (x, y) = (x̄, 0). There are finitely many subsets J of 1, . . . , m ,
and for each u close to x̄ we have PC (u) = xJ (u) for some J. Thus, the projection
mapping PC is a selection of finitely many smooth functions.

  
Exercise 2D.10. For a set of the form C = x ∈ IRn  Ax = b ∈ IRm , if the m × n
matrix A has linearly independent rows, then the projection mapping is given by

PC (x) = (I − AT (AAT )−1 A)x + AT (AAT )−1 b.

Guide. The optimality condition (6) in this case leads to the system of equations
    
x I AT PC (x)
= .
b A 0 λ

Use the identity

   −1
I AT I − AT (AAT )−1 A AT (AAT )−1
=
A 0 (AAT )−1 A −(AAT )−1

to reach the desired conclusion.

2E. Variational Inequalities with Polyhedral Convexity

In this section we apply the theory presented in the preceding sections of this chapter
to the parameterized variational inequality

(1) f (p, x) + NC (x)  0

where f : IRd × IRn → IRn and C is a nonempty, closed and convex subset of IRn . The
corresponding solution mapping S : IR p →
→ IRn , with
  
(2) S(p) = x  f (p, x) + NC (x)  0 ,

has already been the direct subject of Theorem 2B.1, the implicit function theorem
of Robinson. From there we moved on to broader results about solution mappings to
generalized equations, but now wish to summarize what those results mean back in
96 2 Implicit Function Theorems for Variational Problems

the variational inequality setting, and furthermore to explore special features which
emerge under additional assumptions on the set C.

Theorem 2E.1 (solution mappings for parameterized variational inequalities). For

a variational inequality (1) and its solution mapping (2), let p̄ and x̄ be such that
x̄ ∈ S( p̄). Assume that
(a) f is strictly differentiable at ( p̄, x̄);
(b) the inverse G−1 of the mapping

(3) G(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄) + NC (x), with G(x̄)  0,

has a Lipschitz continuous single-valued localization σ around 0 for x̄.

Then S has a Lipschitz continuous single-valued localization s around p̄ for x̄
with
lip (s; p̄) ≤ lip (σ ; 0) · |∇ p f ( p̄, x̄)|,
and this localization s has a first-order approximation η at p̄ given by

(4) η (p) = σ (−∇ p f ( p̄, x̄)(p − p̄)).

Moreover, under the ample parameterization condition

rank ∇ p f ( p̄, x̄) = n,

the existence of a Lipschitz continuous single-valued localization s of S around p̄

for p̄ not only follows from but also necessitates the existence of a localization σ of
G−1 having the properties described.
Proof. This comes from the application to S of the combination of Theorem 2B.5
and its specialization in Corollary 2B.9, together with the ample parameterization
result in Theorem 2C.2.
If the localization σ that is assumed to exist in Theorem 2E.1 is actually linear,
the stronger conclusion is obtained that s is differentiable at p̄. But that’s a circum-
stance which can hardly be guaranteed without supposing, for instance, that C is
an affine set (given by a system of linear equations). In some situations, however, s
could be at least piecewise smooth, as the projection mapping in 2D.9.
Our special goal here is trying to understand better the circumstances in which
the existence of a single-valued localization σ of G−1 around 0 for x̄ of the kind
assumed in (b) of Theorem 2E.1 is assured. It’s clear from the formula for G in (3)
that everything hinges on how a normal cone mapping NC : IRn → → IRn may relate
to an affine function x → a + Ax. The key lies in the local geometry of the graph
of NC . We will be able to make important progress in analyzing this geometry by
restricting our attention to the following class of sets C.

Polyhedral convex sets. A set C in IRn is said to be polyhedral convex when it can
be expressed as the intersection of finitely many closed half-spaces and/or hyper-
planes.
2 Implicit Function Theorems for Variational Problems 97

In other words, C is a polyhedral convex set when it can be described by a finite

set of constraints fi (x) ≤ 0 or fi (x) = 0 on affine functions fi : IRn → IR. Since an
equation fi (x) = 0 is equivalent to the pair of inequalities fi (x) ≤ 0 and − fi (x) ≤ 0,
a polyhedral convex set C is characterized by having a (nonunique) representation
of the form
  
(5) C = x  bi , x ≤ αi for i = 1, . . . , m .

Any such set must obviously be closed. The empty set 0/ and the whole space IRn are
regarded as polyhedral convex sets, in particular.
Polyhedral convex cones are characterized by having a representation (5) in
which αi = 0 for all i. A basic fact about polyhedral convex cones is that they can
equally well be represented in another way, which we recall next.

Theorem 2E.2 (Minkowski–Weyl). A set K ⊂ IRn is a polyhedral convex cone if

and only if there is a collection of vectors b1 , . . . , bm such that
  
(6) K = y1 b1 + · · · + ym bm  yi ≥ 0 for i = 1, . . . , m .

It is easy to see that the cone K ∗ that is polar to a cone K having a representation
of the kind in (6) consists of the vectors x satisfying bi , x ≤ 0 for i = 1, . . . , m. The
polar of a polyhedral convex cone having such an inequality representation must
therefore have the representation in (6), inasmuch as (K ∗ )∗ = K for any closed,
convex cone K. This fact leads to a special description of the tangent and normal
cones to a polyhedral convex set.

Theorem 2E.3 (variational geometry of polyhedral convex sets).   Let C be apoly-

hedral convex set represented as in (5). Let x ∈ C and I(x) = i  bi , x = αi , this
being the set of indices of the constraints in (5) that are active at x. Then the tangent
and normal cones to C at x are polyhedral convex, with the tangent cone having the
representation
  
(7) TC (x) = w  bi , w ≤ 0 for i ∈ I(x)

and the normal cone having the representation

  

NC (x) = v  v = ∑i=1 yi bi with yi ≥ 0 for i ∈ I(x), yi = 0 for i ∈
m
(8) / I(x) .

Furthermore, the tangent cone has the properties that

(9) W ∩ [C − x] = W ∩ TC (x) for some neighborhood W of 0

and

(10) TC (x) ⊃ TC (x̄) for all x in some neighborhood U of x̄.

98 2 Implicit Function Theorems for Variational Problems

Proof. The formula (7) for TC (x) follows from (5) just by applying the definition
of the tangent cone in 2A. Then from (7) and the preceding facts about polyhedral
cones and polarity, utilizing also the relation in 2A(8), we obtain (8). The equality
(9) is deduced simply by comparing (5) and (7). To obtain (10), observe that I(x) ⊂
I(x̄) for x close to x̄ and then the inclusion follows from (7).

TC(x)
KC(x,v)

v x

NC(x)

Fig. 2.2 Tangent, normal and critical cones to a polyhedral set.

The normal cone mapping NC associated with a polyhedral convex set C has a
special property which will be central to our analysis. It revolves around the follow-
ing notion.

Critical cone. For a convex set C, any x ∈ C and any v ∈ NC (x), the critical cone to
C at x for v is   
KC (x, v) = w ∈ TC (x)  w ⊥ v .
If C is polyhedral, then KC (x, v) is polyhedral as well, as seen immediately from
the representation in (7).

Lemma 2E.4 (reduction lemma). Let C be a polyhedral convex set in IRn , and let

x̄ ∈ C, v̄ ∈ NC (x̄), K = KC (x̄, v̄).

The graphical geometry of the normal cone mapping NC around (x̄, v̄) reduces then
to the graphical geometry of the normal cone mapping NK around (0, 0), in the sense
that

O ∩ [gph NC − (x̄, v̄)] = O ∩ gph NK for some neighborhood O of (0, 0).

In other words, one has

(11) v̄ + u ∈ NC (x̄ + w) ⇐⇒ u ∈ NK (w) for (w, u) sufficiently near to (0, 0).

2 Implicit Function Theorems for Variational Problems 99

Proof. Since we are only involved with local properties of C around one of its points
x̄, and C − x̄ agrees with the cone TC (x̄) around 0 by Theorem 2E.3, we can assume
without loss of generality that x̄ = 0 and C is a cone, and TC (x̄) = C. Then, in terms
of the polar cone C∗ (which likewise is polyhedral on the basis of Theorem 2E.2),
we have the characterization from 2A.3 that

(12) v ∈ NC (w) ⇐⇒ w ∈ NC∗ (v) ⇐⇒ w ∈ C, v ∈ C∗ , v, w = 0.

In particular for our focus on the geometry of gph NC around (0, v̄), we have from
(12) that
  
(13) NC∗ (v̄) = w ∈ C  v̄, w = 0 = K.

We know on the other hand from 2E.3 that U ∩ [C∗ − v̄] = U ∩ TC∗ (v̄) for some
neighborhood U of 0, where moreover TC∗ (v̄) is polar to NC∗ (v̄), hence equal to K ∗
by (13). Thus, there is a neighborhood O of (0, 0) such that

(14) for (w, u) ∈ O : v̄ + u ∈ NC (w) ⇐⇒ w ∈ C, u ∈ K ∗ , v̄ + u, w = 0.

This may be compared with the fact that

(15) u ∈ NK (w) ⇐⇒ w ∈ K, u ∈ K ∗ , u, w = 0.

Our goal (in the context of x̄ = 0) is to show that (14) reduces to (11), at least
when the neighborhood O in (14) is chosen still smaller, if necessary. Because of
(15), this comes down to demonstrating that v̄, w = 0 in the circumstances of (14).
We can take C to be represented by
  
(16) C = w  bi , w ≤ 0 for i = 1, . . . , m ,

in which case, as observed after 2E.2, the polar C∗ is represented by

  
(17) C∗ = y1 b1 + · · · + ym bm  yi ≥ 0 for i = 1, . . . , m .

The relations in (12) can be coordinated with these representations as follows. For
each index set I ⊂ {1, . . . , m}, consider the polyhedral convex cones
    
WI = w ∈ C  bi , w = 0 for i ∈ I , VI = ∑i∈I yi bi with yi ≥ 0 ,

with W0/ = C and V0/ = {0}. Then v ∈ NC (w) if and only if, for some I, one has w ∈ WI
and v ∈ VI . In other words, gph NC is the union of the finitely many polyhedral
convex cones GI = WI × VI in IRn × IRn .
Among these cones GI , we will only be concerned with the ones containing (0, v̄).
Let I be the collection of index sets I ⊂ {1, . . . , m} having that property. According
to (9) in 2E.3, there exists for each I ∈ I a neighborhood OI of (0, 0) such that
OI ∩ [GI − (0, v̄)] = OI ∩ TGI (0, v̄). Furthermore, TGI (0, v̄) = WI × TVI (v̄). This has
the crucial consequence that when v̄ + u ∈ NC (w) with (w, u) near enough to (0, 0),
100 2 Implicit Function Theorems for Variational Problems

we also have v̄ + τ u ∈ NC (w) for all τ ∈ [0, 1]. Since having v̄ + τ u ∈ NC (w) entails
having v̄ + τ u, w = 0 through (12), this implies that v̄, w = −τ u, w for all τ ∈
[0, 1]. Hence v̄, w = 0, as required. We merely have to shrink the neighborhood O
in (14) to lie within every OI for I ∈ I .

Example 2E.5. The nonnegative orthant IRn+ is a polyhedral convex cone in IRn ,
since it consists of the vectors x = (x1 , . . . , xn ) satisfying the linear inequalities x j ≥
0, j = 1, . . . , n. For v = (v1 , . . . , vn ), one has

v ∈ NIRn+ (x) ⇐⇒ x j ≥ 0, v j ≤ 0, x j v j = 0 for j = 1, . . . , n.

Thus, whenever v ∈ NIRn+ (x) one has in terms of the index sets
  
J1 = j  x j > 0, v j = 0 ,
  
J2 = j  x j = 0, v j = 0 ,
  
J3 = j  x j = 0, v j < 0

that the vectors w = (w1 , . . . , wn ) belonging to the critical cone to IRn+ at x for v are
characterized by
⎧
⎨ w j free for j ∈ J1 ,
w ∈ KIRn+ (x, v) ⇐⇒ w j ≥ 0 for j ∈ J2 ,
⎩ w = 0 for j ∈ J .
j 3

In the developments ahead, we will make use of not only critical cones but also
certain subspaces.

Critical subspaces. The smallest linear subspace that includes the critical cone
KC (x, v) will be denoted by KC+ (x, v), whereas the smallest linear subspace that is
included in KC (x, v) will be denoted by KC− (x, v), the formulas being
  
KC+ (x, v) = KC (x, v) − KC (x, v) = w− w  w, w ∈ K
 C (x, v) , 
(18)
KC− (x, v) = KC (x, v) ∩ [−KC (x, v)] = w ∈ KC (x, v)  − w ∈ KC (x, v) .

The formulas follow from the fact that KC (x, v) is already a convex cone. Obvi-
ously, KC (x, v) is itself a subspace if and only if KC+ (x, v) = KC− (x, v).

Theorem 2E.6 (affine-polyhedral variational inequalities). For an affine function

x → a + Ax from IRn into IRn and a polyhedral convex set C ⊂ IRn , consider the
variational inequality
a + Ax + NC(x)  0.
Let x̄ be a solution and let v̄ = −a − Ax̄, so that v̄ ∈ NC (x̄), and let K = KC (x̄, v̄) be
the associated critical cone. Then for the mappings

G(x) = a + Ax + NC(x) with G(x̄)  0,

(19)
G0 (w) = Aw + NK (w) with G0 (0)  0,
2 Implicit Function Theorems for Variational Problems 101

the following properties are equivalent:

(a) G−1 has a Lipschitz continuous single-valued localization σ around 0 for x̄,
(b) G−1 n
0 is a single-valued mapping with all of IR as its domain,
in which case G−1 0 is necessarily Lipschitz continuous globally and the function
σ (v) = x̄ + G−1
0 (v) furnishes the localization in (a). Moreover, in terms of critical
subspaces K + = KC+ (x̄, v̄) and K − = KC− (x̄, v̄), the following condition is sufficient
for (a) and (b) to hold:
+ −
(20) w∈K , Aw ⊥ K , w, Aw ≤ 0 =⇒ w = 0.

Proof. According to reduction lemma 2E.4, we have, for (w, u) in some neighbor-
hood of (0, 0), that v̄ + u ∈ NC (x̄ + w) if and only if u ∈ NK (w). In the change of
notation from u to v = u + Aw, this means that, for (w, v) in a neighborhood of
(0, 0), we have v ∈ G(x̄ + w) if and only if v ∈ G0 (w). Thus, the existence of a
Lipschitz continuous single-valued localization σ of G−1 around 0 for x̄ as in (a)
corresponds to the existence of a Lipschitz continuous single-valued localization σ0
of G−1
0 around 0 for 0; the relationship is given by σ (v) = x̄ + σ0 (v). But when-
ever v ∈ G0 (w) we have λ v ∈ G0 (λ w) for all λ > 0, i.e., the graph of G0 is a cone.
Therefore, when σ0 exists it can be scaled arbitrarily large and must correspond to
G−1 n
0 being a single-valued mapping with all of IR as its domain.
−1
We claim next that when G0 is single-valued everywhere it is necessarily Lip-
schitz continuous. This comes out of the argument pursued in the proof of 2E.4 in
analyzing the graph of NC , which applies equally well to NK , inasmuch as K is a
polyhedral convex cone. Specifically, the graph of NK is the union of finitely many
polyhedral convex cones in IRn × IRn . The same also holds then for the graphs of G0
and G−10 . It remains only to observe that if a single-valued mapping has its graph
composed of the union of finitely many polyhedral convex sets it has to be Lipschitz
continuous (prove or see 3D.6).
This leaves us with verifying that the condition in (20) is sufficient for G−1
0 to be
single-valued with all of IRn as its domain. We note in preparation for this that

(K )⊥ = K ∗ ∩ (−K ∗ ) = (K ∗ ) , (K )⊥ = K ∗ − K ∗ = (K ∗ ) .
+ − − +
(21)

We first argue that if w1 ∈ G−1 −1

0 (v) and w2 ∈ G0 (v), then v − Aw1 ∈ NK (w1 ) and
v − Aw2 ∈ NK (w2 ). This entails having

w1 ∈ K, v − Aw1 ∈ K ∗ , w1 , v − Aw1  = 0,

w2 ∈ K, v − Aw2 ∈ K ∗ , w2 , v − Aw2  = 0,

with w1 , v − Aw2  ≤ 0 and w2 , v − Aw1  ≤ 0. Then w1 − w2 ∈ K − K = K + and

−A(w1 − w2 ) ∈ K ∗ − K ∗ = (K − )⊥ , with w1 − w2 , A(w1 − w2 ) = w1 − w2 , [v −
Aw2 ]− [v− Aw2 ] ≤ 0. Under our condition (20), these relations require w1 − w2 = 0.
Thus, (20) guarantees that G−1
0 (v) can never contain more than a single w.
Working toward showing that (20) guarantees also that dom G−1 0 = IR , we next
n
−1
consider the case where dom G0 omits some point ṽ and analyze what that would
102 2 Implicit Function Theorems for Variational Problems

imply. Again we utilize the fact that the graph of G−1 0 is the union of finitely many
polyhedral convex cones in IRn × IRn . Under the mapping (v, w) → v, each of them
projects onto a cone in IRn ; the union of these cones is dom G−1 0 . Since the image
of a polyhedral convex cone under a linear transformation is another polyhedral
convex cone, in consequence of 2E.2 (since the image of a cone generated by finitely
many vectors is another such cone), and polyhedral convex cones are closed sets
in particular, this ensures that dom G−1 0 is closed. Then there is certain to exist a
point v0 ∈ dom G−1 0 that is closest to ṽ; for all τ > 0 sufficiently small, we have
/ dom G−1
v0 + τ (ṽ − v0 ) ∈ 0 . For each of the polyhedral convex cones D in the finite
−1
union making up dom G0 , if v0 ∈ D, then v0 must be the projection PD (ṽ), so
that ṽ − v0 must belong to ND (v0 ) (cf. relation (4) in 2A). It follows that, for some
neighborhood U of v0 , we have

(22) ṽ − v0 , u − v0 ≤ 0 for all u ∈ U ∩ dom G−1

0 .

Consider any w0 ∈ G−10 (v0 ); this means v0 − Aw0 ∈ NK (w0 ). Let K0 be the critical
cone to K at w0 for v0 − Aw0 :
  
(23) K0 = w ∈ TK (w0 )  w ⊥ (v0 − Aw0 ) .

In the line of argument already pursued, the geometry of the graph of NK around
(w0 , v0 − Aw0 ) can be identified with that of the graph of NK0 around (0, 0). Equiv-
alently, the geometry of the graph of G−1 0 = (A + NK )
−1 around (v , w ) can be
0 0
identified with that of (A + NK0 ) around (0, 0); for (v , w ) near enough to (0, 0),
−1 

we have w0 + w ∈ G−1   −1 
0 (v0 + v ) if and only if w ∈ (A + NK0 ) (v ). Because of
(22) holding for the neighborhood U of v0 , this implies that ṽ − v0 , v  ≤ 0 for all
v ∈ dom (A + NK0 )−1 close to 0. Thus,

(24) ṽ − v0 , Aw + u  ≤ 0 for all w ∈ K0 and u ∈ NK0 (w ).

The case of w = 0 has NK0 (w ) = K0∗ , so (24) implies in particular that ṽ − v0 , u  ≤
0 for all u ∈ K0∗ , so that ṽ − v0 ∈ (K0∗ )∗ = K0 . On the other hand, since u = 0 is
always one of the elements of NK0 (w ), we must have from (24) that ṽ− v0 , Aw  ≤ 0
for all w ∈ K0 . Here ṽ − v0 , Aw  = AT (ṽ − v0 ), w  for all w ∈ K0 , so this means
AT (ṽ − v0 ) ∈ K0∗ . In summary, (24) requires, among other things, having

ṽ − v0 ∈ K0 and AT (ṽ − v0 ) ∈ K0∗ ,

(25)
hence in particular AT (ṽ − v0 ), ṽ − v0  ≤ 0.

We observe now from the formula for K0 in (23) that K0 ⊂ TK (w0 ), where further-
more TK (w0 ) is the cone generated by the vectors w − w0 with w ∈ K and hence
lies in K − K. Therefore K0 ⊂ K + . On the other hand, because TK (w0 ) and NK (w0 )
are polar to each other by 2E.3, we have from (23) that K0 is polar to the cone
comprised by all differences v − τ (v0 − Aw0 ) with v ∈ NK (w0 ) and τ ≥ 0, which is
again polyhedral. That cone of differences must then in fact be K0∗ . Since we have
taken v0 and w0 to satisfy v0 − Aw0 ∈ NK (w0 ), and also NK (w0 ) ⊂ K ∗ , it follows that
2 Implicit Function Theorems for Variational Problems 103

K0∗ ⊂ K ∗ − K ∗ = (K − )⊥ . Thus, (25) implies that ṽ − v0 ∈ K + , AT (ṽ − v0 ) ∈ (K − )⊥ ,

with AT (ṽ − v0 ), ṽ − v0  ≤ 0. In consequence, (25) would be impossible if we knew
that

AT w ⊥ K , AT w, w ≤ 0
+ −
(26) w∈K , =⇒ w = 0.

Our endgame will be to demonstrate that (26) is actually equivalent to condition

(20).
Of course, AT w, w is the same as w, Aw. For additional comparison between
(20) and (26), we can simplify matters by expressing IRn as W1 × W2 × W3 for the
linear subspaces W1 = K − , W3 = (K + )⊥ , and W2 the orthogonal complement of K −
within K + . Any vector w ∈ IRn corresponds then to a triple (w1 , w2 , w3 ) in this prod-
uct, and there are linear transformations Ai j : W j → Wi such that

Aw ←→ (A11 w1 +A12 w2 +A13 w3 , A21 w1 +A22 w2 +A23 w3 , A31 w1 +A32 w2 +A33 w3 ).

In this schematic, (20) has the form

(27) A11 w1 + A12w2 = 0, w2 , A21 w1 + A22w2  ≤ 0 =⇒ w1 = 0, w2 = 0,

whereas (26) has the form

(28) AT T
11 w1 + A21 w2 = 0, w2 , AT T
12 w1 + A22 w2  ≤ 0 =⇒ w1 = 0, w2 = 0.

In particular, through the choice of w2 = 0, (27) insists that the only w1 with A11 w1 =
0 is w1 = 0. Thus, A11 must be nonsingular. Then the initial equation in (27) can be
solved for w1 , yielding w1 = −A−1 11 A12 w2 , and this expression can be substituted
into the inequality, thereby reducing the condition to

w2 , (A22 − A21 A−1

11 A12 )w2  ≤ 0 =⇒ w2 = 0.

In the same manner, (28) comes out as the nonsingularity of AT

11 and the property
that
w2 , (AT T T −1 T
22 − A12 (A11 ) A21 )w2  ≤ 0 =⇒ w2 = 0.

Since the nonsingularity of AT

11 is equivalent to that of A11 , and

AT T T −1 T −1 T
22 − A12(A11 ) A21 = (A22 − A21 A11 A12 ) ,

the equivalence of (27) and (28) is now evident.

Example 2E.7.
(a) When the critical cone K in Theorem 2E.6 is a subspace, the condition in
(20) reduces to the nonsingularity of the linear transformation K  w → PK (Aw),
where PK is the projection onto K.
(b) When the critical cone K in Theorem 2E.6 is pointed, in the sense that K ∩
(−K) = {0}, the condition in (20) reduces to the requirement that w, Aw > 0 for
all nonzero w ∈ K + .
104 2 Implicit Function Theorems for Variational Problems

(c) Condition (20) always holds when A is the identity matrix.

Theorem 2E.6 tells us that, for a polyhedral convex set C, the assumption in (b)
of Theorem 2E.1 is equivalent to the critical cone K = KC (x̄, − f ( p̄, x̄)) being such
that the inverse G−1
0 of the mapping G0 : w → ∇x f (x̄, p̄)w + NK (w) is single-valued
from all of IRn into itself and hence Lipschitz continuous globally. Furthermore,
Theorem 2E.6 provides a sufficient condition for this to hold. In putting these facts
together with observations about the special nature of G0 , we obtain a powerful fact
which distinguishes variational inequalities with polyhedral convexity from other
variational inequalities.

Theorem 2E.8 (localization criterion under polyhedral convexity). For a variational

inequality (1) and its solution mapping (2) under the assumption that C is polyhedral
convex and f is strictly differentiable at ( p̄, x̄), with x̄ ∈ S( p̄), let

A = ∇x f ( p̄, x̄) and K = KC (x̄, v̄) for v̄ = − f ( p̄, x̄).

Suppose that for each u ∈ IRn there is a unique solution w = s̄(u) to the auxiliary
variational inequality Aw − u + NK (w)  0, this being equivalent to saying that

(29) s̄ = (A + NK )−1 is everywhere single-valued,

in which case the mapping s̄ is Lipschitz continuous globally. (A sufficient condi-

tion for this assumption to hold is the property in (20) with respect to the critical
subspaces K + = KC+ (x̄, v̄) and K − = KC− (x̄, v̄).)
Then S has a Lipschitz continuous single-valued localization s around p̄ for x̄
which is semidifferentiable with

(30) lip (s; p̄) ≤ lip (s̄; 0) · |∇ p f ( p̄, x̄)|, Ds( p̄)(q) = s̄(−∇ p f ( p̄, x̄)q).

Moreover, under the ample parameterization condition, rank ∇ p f ( p̄, x̄) = n, condi-
tion (29) is not only sufficient but also necessary for a Lipschitz continuous single-
valued localization of S around p̄ for x̄.
Proof. We merely have to combine the observation made before this theorem’s
statement with the statement of Theorem 2E.1. According to formula (4) in that
theorem for the first-order approximation η of s at p̄, we have η ( p̄ + q) − x̄ =
s̄(−∇ p f ( p̄, x̄)q). Because K is a cone, the mapping NK is positively homogeneous,
and the same is true then for A + NK and its inverse, which is s̄. Thus, the function
q → s̄(−∇ p f ( p̄, x̄)q) gives a first-order approximation to s( p̄ + q) − s( p̄) at q = 0
that is positively homogeneous. We conclude that s is semidifferentiable at p̄ with
this function furnishing its semiderivative, as indicated in (30).
As a special case of Example 2E.7(a), if C = IRn the result in Theorem 2E.8 re-
duces once more to a version of the classical implicit function theorem. Further in-
sights into solution mappings associated with variational inequalities will be gained
in Chapter 4.
2 Implicit Function Theorems for Variational Problems 105

Exercise 2E.9. Prove that the projection mapping PC associated with a polyhedral
convex set C is Lipschitz continuous and semidifferentiable everywhere, with its
semiderivative being given by

DPC (x)(u) = PK (u) for K = KC (PC (x), x − PC (x)).

Guide. Use the relation between the projection mapping and the normal cone map-
ping given in formula 2A(4).

Additional facts about critical cones, which will be useful later, can be developed
from the special geometric structure of polyhedral convex sets.

Proposition 2E.10 (local behavior of critical cones and subspaces). Let C ⊂ IRn be
a polyhedral convex set, and let v̄ ∈ NC (x̄). Then the following properties hold:
(a) KC (x, v) ⊂ KC+ (x̄, v̄) for all (x, v) ∈ gph NC in some neighborhood of (x̄, v̄).
(b) KC (x, v) = KC+ (x̄, v̄) for some (x, v) ∈ gph NC in each neighborhood of (x̄, v̄).
Proof. By appealing to 2E.3 as in the proof of 2E.4, we can reduce to the case
where x̄ = 0 and C is a cone. Theorem 2E.2 then provides a representation in terms
of a collection of nonzero vectors b1 , . . . , bm , in which C consists of all linear combi-
nations y1 b1 + · · · + ym bm with coefficients yi ≥ 0, and the polar cone C∗ consists of
all v such that bi , v ≤ 0 for all i. We know from 2A.3 that, at any x ∈ C, the normal
cone NC (x) is formed by the vectors v ∈ C∗ such that x, v = 0, so that NC (x) is the
cone that is polar to the one comprised of all vectors w − λ x with w ∈ C and λ ≥ 0.
Since the latter cone is again polyhedral (in view of Theorem 2E.2), hence closed,
it must in turn be the cone polar to NC (x) and therefore equal to TC (x). Thus,
  
TC (x) = y1 b1 + · · · + ymbm − λ x  yi ≥ 0, λ ≥ 0 for any x ∈ C.

On the other hand, in the notation

  
I(v) = i  bi , v = 0 ,  
(31)
F(v) = y1 b1 + · · · + ym bm  yi ≥ 0 for i ∈ I(v), yi = 0 for i ∈
/ I(v) ,
  
we see that for v ∈ C∗ we have F(v) = x ∈ C  v ∈ NC (x) , i.e.,

v ∈ NC (x) ⇐⇒ x ∈ F(v).
  
Then too, for such x and v, the critical cone KC (x, v) = w ∈ TC (x)  w, v = 0 we
have
  
(32) KC (x, v) = w − λ x  w ∈ F(v), λ ≥ 0 ,

and actually KC (x̄, v̄) = F(v̄) (inasmuch as x̄ = 0). In view of the fact, evident from
(31), that

I(v) ⊂ I(v̄) and F(v) ⊂ F(v̄) for all v near enough to v̄,
106 2 Implicit Function Theorems for Variational Problems

we have, for v in some neighboorhood of v̄, that

  
x ∈ F(v̄) and KC (x, v) ⊂ w − λ x  w ∈ F(v̄), λ ≥ 0 when v ∈ NC (x).

In that case KC (x, v) ⊂ F(v̄) − F(v̄) = KC (x̄, v̄) − KC (x̄, v̄) = KC+ (x̄, v̄), so (a) is valid.
To confirm (b), it will be enough now to demonstrate that, arbitrarily close to
x̄ = 0, we can find a vector x̃ for which KC (x̃, v̄) = F(v̄) − F(v̄). Here F(v̄) consists
by definition of all nonnegative linear combinations of the vectors bi with i ∈ I(v̄),
whereas F(v̄) − F(v̄) is the subspace consisting of all linear combinations. For arbi-
trary ε > 0, let x̃ = ỹ1 b1 +
 · · · + ỹmbm with ỹi = ε fori ∈ I(v̄) but ỹi = 0 for i ∈/ I(v̄).
Then KC (x̃, v̄), equaling w − λ x̃  w ∈ F(v̄), λ ≥ 0 by (32), consists of all linear
combinations of the vectors bi for i ∈ I(v̄) in which the coefficients have the form
yi − λ ε with yi ≥ 0 and λ ≥ 0. Can any given choice of coefficients yi for i ∈ I(v̄)
be obtained in this manner? Yes, by taking λ high enough that yi + λ ε ≥ 0 for all
i ∈ I(v̄) and then setting yi = yi + λ ε . This completes the argument.

2F. Variational Inequalities with Monotonicity

Our attention shifts now from special properties of the set C in a variational inequal-
ity to special properties of the function f and their effect on solutions.

Monotone functions. A function f : IRn → IRn is said to be monotone on a set

C ⊂ dom f if C is convex and

(1)  f (x ) − f (x), x − x ≥ 0 for all x, x ∈ C.

It is strongly monotone on C if there exists μ > 0 such that

(2)  f (x ) − f (x), x − x ≥ μ |x − x|2 for all x, x ∈ C.

Specifically, then f is strongly monotone on C with constant μ.

The name “monotonicity” comes from the following characterization of the
defining property.

Exercise 2F.1 (monotonicity along line segments). Monotonicity of f on a convex

set C ⊂ dom f means that, for every x̂ ∈ C and w ∈ IRn with |w| = 1, the function
ϕ (τ ) =  f (x̂ + τ w), w is nondecreasing over the (interval of) τ values such that
x̂ + τ w ∈ C. Strong monotonicity with constant μ > 0 corresponds to the condition
that ϕ (τ  ) − ϕ (τ ) ≥ μ |τ  − τ | when τ  > τ.
An affine function f (x) = a + Ax with a ∈ IRn and A ∈ IRn×n is monotone on IRn
if and only if w, Aw ≥ 0 for all w, i.e., A is positive semidefinite. It is strongly
2 Implicit Function Theorems for Variational Problems 107

monotone if and only if w, Aw > 0 for all w = 0, i.e., A is positive definite. These
terms make no requirement of symmetry on A. It may be recalled that any square
matrix A can be written as a sum As + Aa in which As is symmetric (A∗s = As ) and
Aa is antisymmetric (A∗a = −Aa ), namely with As = 12 [A + A∗ ] and Aa = 12 [A − A∗ ];
then w, Aw = w, As w. The monotonicity of f (x) = a + Ax thus depends only on
the symmetric part As of A; the antisymmetric part Aa can be anything.
For differentiable functions f that aren’t affine, monotonicity has a similar char-
acterization with respect to the Jacobian matrices ∇ f (x).

Exercise 2F.2 (monotonicity from derivatives). For a function f : IRn → IRn that is
continuously differentiable on an open convex set O ⊂ dom f , verify the following
facts.
(a) A necessary and sufficient condition for f to be monotone on O is the positive
semidefiniteness of ∇ f (x) for all x ∈ O.
(b) If ∇ f (x) is positive definite at every point x of a closed, bounded, convex set
C ⊂ O, then f is strongly monotone on C.
(c) If C is a convex subset of O such that ∇ f (x)w, w ≥ 0 for every x ∈ C and
w ∈ C − C, then f is monotone on C.
(d) If C is a convex subset of O such that ∇ f (x)w, w ≥ μ |w|2 for every x ∈ C
and w ∈ C − C, where μ > 0, then f is strongly monotone on C with constant μ.
Guide. Derive this from the characterizations in 2F.1 by investigating the deriva-
tives of the function ϕ (τ ) introduced there. In proving (c), argue by way of the mean
value theorem that  f (x )− f (x), x − x equals ∇ f (x̃)(x − x), x − x for some point
x̃ on the line segment joining x with x .

Exercise 2F.3 (gradient connections).

(a) Let g be continuously differentiable from an open set O ⊂ IRn to IR, and let
C be a convex subset of O. Show that the function f (x) = ∇g(x) is monotone on C
if and only if g is convex on C. Show further that f is strongly monotone on C with
constant μ if and only if g is strongly convex on C with constant μ.
(b) Let h be continuously differentiable from a product O1 × O2 of open sets
O1 ⊂ IRn1 and O2 ⊂ IRn2 to IR, and let C1 ⊂ O1 and C2 ⊂ O2 be convex. Show that
the function
f (x1 , x2 ) = (∇x1 h(x1 , x2 ), −∇x2 h(x1 , x2 ))
is monotone on C1 × C2 if and only if h(x1 , x2 ) is convex with respect to x1 ∈ C1
for fixed x2 ∈ C2 , and on the other hand concave with respect to x2 ∈ C2 for fixed
x1 ∈ C1 .
Guide. Derive (a) from the characterization of the convexity and strong convexity
of g in 2A.5. Proceed similarly in (b), applying also the corresponding characteriza-
tion of concavity.
108 2 Implicit Function Theorems for Variational Problems

We are ready now to develop some special results for variational inequalities

(3) f (x) + NC (x)  0,

in which attention is devoted to the case when f is monotone. We work with the
basic perturbation scheme in which f (x) is replaced by f (x) − p for a parameter
vector p ∈ IRn . The solution mapping is then
  
(4) S(p) = x  p − f (x) ∈ NC (x) = ( f + NC )−1 (p),

with the solution set to (3) then being S(0).

Theorem 2F.4 (solution convexity for monotone variational inequalities). For a

function f : IRn → IRn and a nonempty closed convex set C ⊂ dom f relative to
which f is monotone and continuous, the solution mapping S in (4) is closed and
convex valued. In particular, therefore, the set of solutions (if any) to the variational
inequality (3) is not only closed but also convex.
Proof. It suffices to deal with S(0), since S(p) = ( f p + NC )−1 (0) for f p (x) = f (x) −
p (which is monotone and continuous like f ).
The closedness of S(0) already follows from Theorem 2A.1. To see the convexity,
consider any two points x0 and x1 in S(0). We have − f (x0 ) ∈ NC (x0 ) and − f (x1 ) ∈
NC (x1 ); this is equivalent to

(5)  f (x0 ), x − x0  ≥ 0 and  f (x1 ), x − x1  ≥ 0 for all x ∈ C.

Let x̄ = (1 − λ )x0 + λ x1 for any λ ∈ (0, 1). Then x̄ ∈ C by convexity. Consider an

arbitrary point x̃ ∈ C. The goal is to show that  f (x̄), x̃ − x̄ ≥ 0, which will confirm
that − f (x̄) ∈ NC (x̄), i.e., that x̄ ∈ S(0).
Taking t ∈ (0, 1) as a parameter, let x(t) = x̄ +t(x̃ − x̄) and note that the convexity
of C ensures x(t) ∈ C. From the monotonicity of f and the first inequality in (5) we
have

0 ≤  f (x(t)) − f (x0 ), x(t) − x0  +  f (x0 ), x(t) − x0  =  f (x(t)), x(t) − x0 .

In parallel from the second inequality in (5), we have 0 ≤  f (x(t)), x(t)− x1 . There-
fore

0 ≤ (1 − λ ) f (x(t)), x(t) − x0  + λ  f (x(t)), x(t) − x1 

=  f (x(t)), x(t) − (1 − λ )x0 − λ x1,

where the final expression equals  f (x(t)), x(t) − x̄ = t f (x(t)), x̃ − x̄, since x(t) −
x̄ = t[x̃ − x̄]. Thus, 0 ≤  f (x(t)), x̃ − x̄. Because x(t) → x̄ as t → 0, and f is contin-
uous, we conclude that  f (x̄), x̃ − x̄ ≥ 0, as required.

In order to add nonemptiness of the solution set to the conclusions of Theorem

2F.4 we need an existence theorem for the variational inequality (3). There is already
2 Implicit Function Theorems for Variational Problems 109

such a result in 2A.1, but only for bounded sets C. The following result goes beyond
that boundedness restriction, without yet imposing any monotonicity assumption
on f . When combined with monotonicity, it will have particularly powerful conse-
quences.

Theorem 2F.5 (solution existence for variational inequalities without bounded-

ness). Consider a function f : IRn → IRn and a nonempty closed convex set C ⊂
dom f relative to which f is continuous (but not necessarily monotone). Suppose
there exist x̂ ∈ C and ρ > 0 such that

(6) there is no x ∈ C with |x − x̂| ≥ ρ and  f (x), x − x̂ ≤ 0.

Then the variational inequality (3) has a solution, and every solution x of (3) satisfies
|x − x̂| < ρ.
Proof. Any solution x to (3) would have  f (x), x − x̂ ≤ 0 in particular, and then
necessarily |x − x̂| < ρ under (6). Hence it will suffice to show that (6) guarantees
the existence  of at least  x to (3) with |x − x̂| < ρ .
 one solution
Let Cρ = x ∈ C  |x − x̂| ≤ ρ and consider the modified variational inequal-
ity (3) in which C is replaced by Cρ . According to Theorem 2A.1, this modi-
fied variational inequality has a solution x̄. We have x̄ ∈ Cρ and − f (x̄) ∈ NCρ (x̄).
From 2A.7(b) we know that NCρ (x̄) = NC (x̄) + NB (x̄) for the ball B = IBρ (x̂) =
  
x  |x − x̂| ≤ ρ . Thus,

(7) − f (x̄) − w ∈ NC (x̄) for some w ∈ NB (x̄).

By demonstrating that this implies w = 0, we will be able to see that x̄ actually

satisfies (3).
The normal cone formula for the unit ball in 2A.2(b) extends in an elementary
way to the ball B and indicates that w can only be nonzero if |x̄ − x̂| = ρ and w =
λ [x̄− x̂] for some λ > 0. The normality relation in (7), requiring 0 ≥ − f (x̄)− w, x −
x̄ for all x ∈ C, can be invoked then in the case of x = x̂ to obtain 0 ≥ − f (x̄) −
λ [x̄ − x̂], x̂ − x̄, which simplifies to  f (x̄), x̄ − x̂ ≤ −λ ρ 2 . But this is impossible
under (6).
The assumption in (6) is fulfilled trivially when C is bounded, and in that way
Theorem 2A.1 is seen to be covered by Theorem 2F.5.

Corollary 2F.6 (uniform local existence). Consider a function f : IRn → IRn and a
nonempty closed convex set C ⊂ dom f relative to which f is continuous (but not
necessarily monotone). Suppose there exist x̂ ∈ C, ρ > 0 and η > 0 such that
$
(8) there is no x ∈ C with |x − x̂| ≥ ρ and  f (x), x − x̂ |x − x̂| ≤ η .

Then the solution mapping S in (4) has the property that

  
0/ = S(v) ⊂ x ∈ C  |x − x̂| < ρ when |v| ≤ η .
110 2 Implicit Function Theorems for Variational Problems

Proof. The stronger assumption here ensures that assumption (6) of Theorem 2F.5
is fulfilled by the function fv (x) = f (x) − v for every v with |v| ≤ η . Since S(v) =
( fv + NC )−1 (0), this leads to the desired conclusion.
We can proceed now to take advantage of monotonicity of f on C through the
property in 2F.1 and the observation that

 f (x), x − x̂/|x − x̂| =  f (x̂ + τ w), w when x = x̂ + τ w with τ > 0, |w| = 1.

Then, for any vector w such that x̂ + τ w ∈ C for all τ ∈ (0, ∞), the expression  f (x̂ +
τ w), w is nondecreasing as a function of τ ∈ (0, ∞) and thus has a limit (possibly
∞) as τ → ∞.

Theorem 2F.7 (solution existence for monotone variational inequalities). Consider

a function f : IRn → IRn and a nonempty closed convex set C ⊂ dom f relative to
which f is continuous and monotone. Let x̂ ∈ C and let W consist of the vectors w
with |w| = 1 such that x̂ + τ w ∈ C for all τ ∈ (0, ∞), if any.
(a) If limτ →∞  f (x̂ + τ w), w > 0 for every w ∈ W , then the solution mapping S
in (4) is nonempty-valued on a neighborhood of 0.
(b) If limτ →∞  f (x̂ + τ w), w = ∞ for every w ∈ W , then the solution mapping S
in (4) is nonempty-valued on all of IRn .
Proof. To establish (a), we aim at showing that the limit criterion it proposes is
enough to guarantee the condition (8) in Corollary 2F.6. Suppose the latter didn’t
hold. Then there would be a sequence of points xk ∈ C and a sequence of scalars
ηk > 0 such that
$
 f (xk ), xk − x̂ |xk − x̂| ≤ ηk with |xk − x̂| → ∞, ηk → 0.

Equivalently, in terms of τk = |xk − x̂| and wk = τk−1 (xk − x̂) we have  f (x̂ +
τk wk ), wk  ≤ ηk with |wk | = 1, x̂ + τk wk ∈ C and τk → ∞. Without loss of gener-
ality we can suppose that wk → w for a vector w again having |w| = 1. Then for
any τ > 0 and k high enough that τk ≥ τ , we have from the convexity of C that
x̂ + τ wk ∈ C and from the monotonicity of f that  f (x̂ + τ wk ), wk  ≤ ηk . On taking
the limit as k → ∞ and utilizing the closedness of C and the continuity of f , we get
x̂ + τ w ∈ C and  f (x̂ + τ w), w ≤ 0. This being true for any τ > 0, we see that w ∈ W
and the limit condition in (a) is violated. The validity of the claim in (a) is thereby
confirmed.
The condition in (b) not only implies the condition in (a) but also, by a slight
extension of the argument, guarantees that the criterion in Corollary 2F.6 holds for
every ε > 0.

Exercise 2F.8 (Jacobian criterion for existence and uniqueness). Let f : IRn → IRn
and C ⊂ IRn be such that f is continuously differentiable on C and monotone relative
to C. Fix x̂ ∈ C and let W consist of the vectors w with |w| = 1 such that x̂ + τ w ∈ C
for all τ ∈ (0, ∞). Suppose there exists μ > 0 such that ∇ f (x)w, w ≥ μ for every
2 Implicit Function Theorems for Variational Problems 111

w ∈ W and x ∈ C, if any, when x ∈ C. Then the solution mapping S in (4) is single-

valued on all of IRn .
Guide. Argue through the mean value theorem as applied to ϕ (τ ) =  f (x̂ + τ w), w
that ϕ (τ ) = τ ∇ f (x̂ + θ w)w, w +  f (x̂), w for some θ ∈ (0, τ ). Work toward ap-
plying the criterion in Theorem 2F.7(b).

In the perspective of 2F.2(b), the result in 2F.8 seems to come close to invoking
strong monotonicity of f in the case where f is continuously differentiable. How-
ever, it only involves special vectors w, not every nonzero w ∈ IRn . For instance, in
the affine case where f (x) = Ax + b and C = IRn+ , the criterion obtained from 2F.8 by
choosing x̂ = 0 is simply that Aw, w > 0 for every w ∈ IRn+ with |w| = 1, whereas
strong monotonicity of f would require this for w in IRn , not just IRn+ . In fact, full
strong monotonicity has bigger implications than those in 2F.8.

Theorem 2F.9 (variational inequalities with strong monotonicity). For a function

f : IRn → IRn and a nonempty closed convex set C ⊂ dom f , suppose that f is con-
tinuous relative to C and strongly monotone on C with constant μ > 0 in the sense of
(2). Then the solution mapping S in (4) is single-valued on all of IRn and moreover
Lipschitz continuous with constant μ −1 .
Proof. The strong monotonicity condition in (2) implies for an arbitrary choice of
x̂ ∈ C and w ∈ IRn with |w| = 1 that  f (x̂ + τ w) − f (x̂), τ w ≥ μτ 2 when x̂ + τ w ∈ C.
Then  f (x̂ + τ w), w ≥  f (x̂), w + μτ , from which it’s clear that the limit criterion
in Theorem 2F.7(b) is satisfied, so that S is single-valued on all of IRn .
Consider now any two vectors v0 and v1 in IRn and the corresponding solutions
x0 = S(v0 ) and x1 = S(v1 ). We have v0 − f (x0 ) ∈ NC (x0 ) and v1 − f (x1 ) ∈ NC (x1 ),
hence in particular v0 − f (x0 ), x1 −x0  ≤ 0 and v1 − f (x1 ), x0 −x1  ≤ 0. The second
of these inequalities can also be written as 0 ≤ v1 − f (x1 ), x1 − x0 , and from this
we see that v0 − f (x0 ), x1 − x0  ≤ v1 − f (x1 ), x1 − x0 , which is equivalent to

 f (x1 ) − f (x0 ), x1 − x0  ≤ v1 − v0 , x1 − x0 .

Since  f (x1 ) − f (x0 ), x1 − x0  ≥ μ |x1 − x0 |2 by our assumption of strong mono-

tonicity, while v1 − v0 , x1 − x0  ≤ |v1 − v0 ||x1 − x0 |, it follows that |x1 − x0 | ≤
μ −1 |v1 − v0 |. This verifies the claimed Lipschitz continuity with constant μ −1 .
We extend our investigations now to the more broadly parameterized variational
inequalities of the form

(9) f (p, x) + NC (x)  0

and their solution mappings

  
(10) S(p) = x  f (p, x) + NC (x)  0 ,

with the aim of drawing on the achievements in Section 2E in the presence of mono-
tonicity properties of f with respect to x.
112 2 Implicit Function Theorems for Variational Problems

Theorem 2F.10 (strong monotonicity and strict differentiability). For a variational

inequality (9) and its solution mapping (10) in the case of a nonempty, closed, con-
vex set C, let x̄ ∈ S( p̄) and assume that f is strictly differentiable at ( p̄, x̄). Suppose
for some μ > 0 that

(11) ∇x f ( p̄, x̄)w, w ≥ μ |w|2 for all w ∈ C − C.

Then S has a Lipschitz continuous single-valued localization s around p̄ for x̄ with

(12) lip (s; p̄) ≤ μ −1 |∇ p f ( p̄, x̄)|.

Proof. We apply Theorem 2E.1, observing that its assumption (b) is satisfied on
the basis of Theorem 2F.9 and the criterion in 2F.2(d) for strong monotonicity with
constant μ . Theorem 2F.9 tells us moreover that the Lipschitz constant for the lo-
calization σ in Theorem 2E.1 is no more than μ −1 , and we then obtain (12) from
Theorem 2E.1.
Theorem 2F.10 can be compared to the result in Theorem 2E.8. That result re-
quires C to be polyhedral but allows (11) to be replaced by a weaker condition in
terms of the critical cone K = KC (x̄, v̄) for v̄ = − f ( p̄, x̄). Specifically, instead of ask-
ing the inequality in (11) to hold for all w ∈ C − C, one only asks it to hold for all
w ∈ K − K such that ∇x f ( p̄, x̄)w ⊥ K ∩ (−K). The polyhedral convexity leads in this
case to the further conclusion that the localization is semidifferentiable.

2G. Consequences for Optimization

Several types of variational inequalities are closely connected with problems of op-
timization. These include the basic condition for minimization in Theorem 2A.6 and
the Lagrange condition in Theorem 2A.9, in particular. In this section we investigate
what the general results obtained for variational inequalities provide in such cases.
Recall from Theorem 2A.6 that in minimizing a continuously differentiable func-
tion g over a nonempty, closed, convex set C ⊂ IRn , the variational inequality

(1) ∇g(x) + NC (x)  0

stands as a necessary condition for x to furnish a local minimum. When g is convex

relative to C, it is sufficient for x to furnish a global minimum, but in the absence
of convexity, an x satisfying (1) might not even correspond to a local minimum.
However, there is an important case beyond convexity, which we will draw on later,
in which an x satisfying (1) can be identified through additional criteria as yielding
a local minimum.
In elucidating this case, we will appeal to the fact noted in 2A.4 that the normal
cone NC (x) and the tangent cone TC (x) are polar to each other, so that (1) can be
2 Implicit Function Theorems for Variational Problems 113

written equivalently in the form

(2) ∇g(x), w ≥ 0 for all w ∈ TC (x).

This gives a way to think about the first-order condition for a local minimum of
g in which the vectors w in (2) making the inequality hold as an equation can be
anticipated to have a special role. In fact, those vectors w comprise the critical cone
KC (x, −∇g(x)) to C at x with respect to the vector −∇g(x) in NC (x), as defined in
Section 2E:
  
(3) KC (x, −∇g(x)) = w ∈ TC (x)  ∇g(x), w = 0 .

When C is polyhedral, at least, this critical cone is able to serve in the expression of
second-order necessary and sufficient conditions for the minimization of g over C.

Theorem 2G.1 (second-order optimality on a polyhedral convex set). Let C be a

polyhedral convex set in IRn and let g : IRn → IR be twice continuously differentiable
on C. Let x̄ ∈ C and v̄ = −∇g(x̄).
(a) (necessary condition) If g has a local minimum with respect to C at x̄, then x̄
satisfies the variational inequality (1) and has w, ∇2 g(x̄)w ≥ 0 for all w ∈ KC (x̄, v̄).
(b) (sufficient condition) If x̄ satisfies the variational inequality (1) and has
w, ∇2 g(x̄)w > 0 for all nonzero w ∈ KC (x̄, v̄), then g has a local minimum rela-
tive to C at x̄, indeed a strong local minimum in the sense of there being an ε > 0
such that
ε
(4) g(x) ≥ g(x̄) + |x − x̄|2 for all x ∈ C near x̄.
2
Proof. The necessity emerges through the observation that for any x ∈ C the
function ϕ (t) = g(x̄ + tw) for w = x − x̄ has ϕ  (0) = ∇g(x̄), w and ϕ  (0) =
w, ∇2 g(x̄)w. From one-dimensional calculus, it is known that if ϕ has a local min-
imum at 0 relative to [0, 1], then ϕ  (0) ≥ 0 and, in the case of ϕ  (0) = 0, also has
ϕ  (0) ≥ 0. Having ϕ  (0) = 0 corresponds to having w ∈ KC (x̄, −∇g(x̄)).
Conversely, if ϕ  (0) ≥ 0 and in the case of ϕ  (0) = 0 also has ϕ  (0) > 0, then
ϕ has a local minimum relative to [0, 1] at 0. That one-dimensional sufficient condi-
tion is inadequate for concluding (b), however, because (b) requires a neighborhood
of x̄ relative to C, not just a separate neighborhood relative to each line segment
proceeding from x into C.
To get (b), we have to make use of the properties of the second-order Taylor
expansion of g at x̄ which are associated with g being twice differentiable there: the
error expression

e(w) = g(x̄ + w) − g(x̄) − ∇g(x̄), w − 12 w, ∇2 g(x̄)w

is of type o(|w|2 ). It will help

  to translate
 this into the notation where w = tz with
t ≥ 0 and |z| = 1. Let Z = z  |z| = 1 . To say that e(w) is of type o(|w|2 ) is to say
that the functions
114 2 Implicit Function Theorems for Variational Problems

ft (z) = e(tz)/t 2 for t > 0

converge to 0 on Z uniformly as t → 0.
We furthermore need to rely on the tangent cone property at the end of 2E.3,
which is available because C is polyhedral: there is a neighborhood W of the origin
in IRn such that, as long as w ∈ W , we have x̄ + w ∈ C if and only if w ∈ TC (x̄).
Through this, it will be enough to show, on the basis of the assumption in (b), that
the inequality in (4) holds for w ∈ TC (x̄) when |w| is sufficiently small. Equivalently,
it will be enough to produce an ε > 0 for which

(5) t −2 [g(x̄ + tz) − g(x̄)] ≥ ε for all z ∈ Z ∩ TC (x̄) when t is sufficiently small.

The assumption in (b) entails (in terms of w = tz) the existence of ε > 0 such that
z, ∇2 g(x̄)z > ε when z ∈ Z ∩ KC (x̄, v̄). This inequality also holds then for all z in
some open set containing Z ∩ KC (x̄, v̄). Let Z0 be the intersection of the complement
of that open set with Z ∩ TC (x̄). Since (1) is equivalent to (2), we have for z ∈ Z0
that ∇g(x̄), z > 0. Because Z0 is compact, we actually have an η > 0 such that
∇g(x̄), z > η for all z ∈ Z0 . We see then, in writing

t −2 [g(x̄ + tz) − g(x̄)] = ft (z) + t −1 ∇g(x̄), z + 12 z, ∇2 g(x̄)z

and referring to the uniform convergence of the functions ft to 0 on Z as t → 0,

that for t sufficiently small the left side is at least t −1 η when z ∈ Z0 and at least ε
/ Z0 . By taking t small enough that t −1 η > ε , we get (5)
when z ∈ Z ∩ TC (x̄) but z ∈
as desired.
When x̄ belongs to the interior of C, as for instance when C = IRn (an extreme case
of a polyhedral convex set), the first-order condition in (1) is simply ∇g(x̄) = 0. The
second-order conditions in 2G.1 then specify positive semidefiniteness of ∇2 g(x̄)
for necessity and positive definiteness for sufficiency. Second-order conditions for
a minimum can also be developed for convex sets that aren’t polyhedral, but not in
such a simple form. When the boundary of C is “curved,” the tangent cone property
in 2E.3 fails, and the critical cone KC (x̄, v̄) for v̄ = −∇g(x̄) no longer captures the
local geometry adequately. Second-order optimality with respect to nonpolyhedral
and even nonconvex sets specified by constraints as in nonlinear programming will
be addressed later in this section (Theorem 2G.6).

Stationary points. An x satisfying (1), or equivalently (2), will be called a station-

ary point of g with respect to minimizing over C, regardless of whether or not it
furnishes a local or global minimum.
Stationary points attract attention for their own sake, due to the role they have
in the design and analysis of minimization algorithms, for example. Our immediate
plan is to study how stationary points, as “quasi-solutions” in problems of mini-
mization over a convex set, behave under perturbations. Along with that, we will
clarify circumstances in which a stationary point giving a local minimum continues
to give a local minimum when perturbed by not too much.
2 Implicit Function Theorems for Variational Problems 115

Moving in that direction, we look now at parameterized problems of the form

(6) minimize g(p, x) over all x ∈ C,

where g : IRd × IRn → IR is twice continuously differentiable with respect to x (not

necessarily convex), and C is a nonempty, closed, convex subset of IRn . In the pattern
already seen, the variational inequality

(7) ∇x g(p, x) + NC (x)  0

provides for each p a first-order condition which x must satisfy if it furnishes a local
minimum, but only describes, in general, the stationary points x for the minimization
in (6). If C is polyhedral the question of a local minimum can be addressed through
the second-order conditions provided by Theorem 2G.1 relative to the critical cone
  
(8) KC (x, −∇x g(p, x)) = w ∈ TC (x)  w ⊥ ∇x g(p, x) .

The basic object of interest to us for now, however, is the stationary point mapping
S : IRd →
→ IRn defined by
  
(9) S(p) = x  ∇x g(p, x) + NC (x)  0 .

With respect to a choice of p̄ and x̄ such that x̄ ∈ S( p̄), it will be useful to consider
alongside of (6) an auxiliary problem with parameter v ∈ IRn in which g( p̄, ·) is
essentially replaced by its second-order expansion at x̄:

minimizeḡ(w) − v, w over all w ∈ W,

(10) x g( p̄, x̄), w + 2 w, ∇xx g( p̄, x̄)w,
= g( p̄, x̄) + ∇ 1 2
ḡ(w) 
where 
W = w x̄ + w ∈ C = C − x̄.

The subtraction of v, w “tilts” ḡ, and is referred to therefore as a tilt perturbation.
When v = 0, ḡ itself is minimized.
For this auxiliary problem the basic first-order condition comes out to be the
parameterized variational inequality

(11) ∇x g( p̄, x̄) + ∇2xx g( p̄, x̄)w − v + NW (w)  0, where NW (w) = NC (x̄ + w).

The stationary point mapping for the problem in (10) is accordingly the mapping
S̄ : IRn →
→ IRn defined by
  
(12) S̄(v) = w  ∇x g( p̄, x̄) + ∇2xx g( p̄, x̄)w + NW (w)  v .

The points w ∈ S̄(v) are sure to furnish a minimum in (10) if, for instance, the
matrix ∇2x g( p̄, x̄) is positive semidefinite, since that corresponds to the convexity of
the “tilted” function being minimized. For polyhedral C, Theorem 2G.1 could be
brought in for further analysis of a local minimum in (10). Note that 0 ∈ S̄(0).
116 2 Implicit Function Theorems for Variational Problems

Theorem 2G.2 (parameterized minimization over a convex set). Suppose in the

preceding notation, with x̄ ∈ S( p̄), that
(a) ∇x g is strictly differentiable at ( p̄, x̄), and
(b) S̄ has a Lipschitz continuous single-valued localization s̄ around 0 for 0.
Then S has a Lipschitz continuous single-valued localization s around p̄ for x̄ with

lip (s; p̄) ≤ lip (s̄; 0) · |∇2xp g( p̄, x̄)|,

and s has a first-order approximation η at p̄ given by

 
(13) η (p) = x̄ + s̄ − ∇2xp g( p̄, x̄)(p − p̄) .

On the other hand, (b) is necessary for S to have a Lipschitz continuous single-
valued localization around p̄ for x̄ when the n × d matrix ∇2xp g( p̄, x̄) has rank n.
Under the additional assumption that C is polyhedral, condition (b) is equivalent
to the condition that, for the critical cone K = KC (x̄, −∇x g( p̄, x̄)), the mapping
  
v → S̄0 (v) = w  ∇2xx g( p̄, x̄)w + NK (w)  v is everywhere single-valued.

Moreover, a sufficient condition for this can be expressed in terms of the critical
subspaces KC+ (x̄, v̄) = KC (x̄, v̄) − KC (x̄, v̄) and KC− (x̄, v̄) = KC (x̄, v̄) ∩ [−KC (x̄, v̄)] for
v̄ = −∇x g( p̄, x̄), namely

for every nonzero w ∈ KC+ (x̄, v̄)
(14) w, ∇xx g( p̄, x̄)w > 0
2
with ∇2xx g( p̄, x̄)w ⊥ KC− (x̄, v̄).

Furthermore, in this case the localization s is semidifferentiable at p̄ with semideriva-

tive given by
Ds( p̄)(q) = s̄(−∇2xp g( p̄, x̄)q).
Proof. We apply Theorem 2E.1 with f (p, x) = ∇x g(p, x). The mapping G in that
result coincides with ∇ḡ + NC , so that G−1 is S̄. Assumptions (a) and (b) cover the
corresponding assumptions in 2E.1, with σ (v) = x̄ + s̄(v), and then (13) follows
from 2E(4). In the polyhedral case we also have Theorems 2E.6 and 2E.8 at our
disposal, and this gives the rest.

Theorem 2G.3 (stability of a local minimum on a polyhedral convex set). Suppose

in the setting of the parameterized minimization problem in (6) and its stationary
point mapping S in (9) that C is polyhedral and ∇x g(p, x) is strictly differentiable
with respect to (p, x) at ( p̄, x̄), where x̄ ∈ S( p̄). With respect to the critical subspace
KC+ (x̄, v̄) for v̄ = −∇x g( p̄, x̄), assume that
+
(15) w, ∇2xx g( p̄, x̄)w > 0 for every nonzero w ∈ KC (x̄, v̄).

Then S has a localization s not only with the properties laid out in Theorem 2G.2,
but also with the property that, for every p in some neighborhood of p̄, the point
2 Implicit Function Theorems for Variational Problems 117

x = s(p) furnishes a strong local minimum in (6). Moreover, (15) is necessary for
the existence of a localization s with all these properties, when the n × d matrix
∇2xp g( p̄, x̄) has rank n.
Proof. Obviously (15) implies (14), which ensures according to Theorem 2G.2 that
S has a Lipschitz continuous single-valued localization s around p̄ for x̄. Applying
2E.10(a), we see then that
+ +
KC (x, −∇x g(p, x)) ⊂ KC (x̄, −∇x g( p̄, x̄)) = KC (x̄, v̄)

when x = s(p) and p is near enough to p̄. Since the matrix ∇2xx g(p, x) converges
to ∇2xx g( p̄, x̄) as (p, x) tends to ( p̄, x̄), it follows that w, ∇2xx g(p, x)w > 0 for all
nonzero w ∈ KC (x, −∇x g(p, x)) when x = s(p) and p is close enough to p̄. Since
having x = s(p) corresponds to having the first-order condition in (7), we conclude
that from Theorem 2G.1 that x furnishes a strong local minimum in this case.
Arguing now toward the necessity of (15) under the rank condition on ∇2xp g( p̄, x̄),
we suppose S has a Lipschitz continuous single-valued localization s around p̄ for
x̄ such that x = s(p) gives a local minimum when p is close enough to p̄. For any
x ∈ C near x̄ and v ∈ NC (x), the rank condition gives us a p such that v = −∇x g(p, x);
this follows e.g. from 1F.6. Then x = s(p) and, because we have a local minimum,
it follows that w, ∇2xx g(p, x)w ≥ 0 for every nonzero w ∈ KC (x, v). We know from
2E.10(b) that KC (x, v) = KC+ (x̄, v̄) for choices of x and v arbitrarily close to (x̄, v̄),
where v̄ = −∇x g( p̄, x̄). Through the continuous dependence of ∇2xx g(p, x) on (p, x),
we therefore have
+
(16) w, Aw ≥ 0 for all w ∈ KC (x̄, v̄), where A = ∇2xx g( p̄, x̄) is symmetric.

For this reason, we can only have w, Aw = 0 if Aw ⊥ KC+ (x̄, v̄), i.e., w , Aw = 0
for all w ∈ KC+ (x̄, v̄).
On the other hand, because the rank condition corresponds to the ample param-
eterization property, we know from Theorem 2E.8 that the existence of the single-
valued localization s requires for A and the critical cone K = KC (x̄, v̄) that the map-
ping (A + NK )−1 be single-valued. This would be impossible if there were a nonzero
w such that Aw ⊥ KC+ (x̄, v̄), because we would have w , Aw = 0 for all w ∈ K in
particular (since K ⊂ KC+ (x̄, v̄)), implying that −Aw ∈ NK (w). Then (A + NK )−1 (0)
would contain w along with 0, contrary to single-valuedness. Thus, the inequality in
(16) must be strict when w = 0.
Next we provide a complementary, global result for the special case of a tilted
strongly convex function.

Proposition 2G.4 (tilted minimization of strongly convex functions). Let g : IRn →

IR be continuously differentiable on an open set O, and let C ⊂ O be a nonempty,
closed, convex set on which g is strongly convex with constant μ > 0. Then for each
v ∈ IRn the problem

(17) minimize g(x) − v, x over x ∈ C

118 2 Implicit Function Theorems for Variational Problems

has a unique solution S(v), and the solution mapping S is Lipschitz continuous on
IRn (globally) with Lipschitz constant μ −1 .
Proof. Let gv denote the function being minimized in (17). Like g, this function
is continuously differentiable and strongly convex on C with constant μ ; we have
∇gv (x) = ∇g(x)− v. According to Theorem 2A.6, the condition ∇gv (x)+ NC (x)  0,
or equivalently x ∈ (∇g + NC )−1 (v), is both necessary and sufficient for x to furnish
the minimum in (17). The strong convexity of g makes the mapping f = ∇g strongly
monotone on C with constant μ ; see 2F.3(a). The conclusion follows now by apply-
ing Theorem 2F.9 to this mapping f .

When the function g in Proposition 2G.4 is twice continuously differentiable,

the strong monotonicity can be identified through 2A.5(b) with the inequality
∇2 g(x)w, w ≥ μ |w|2 holding for all x ∈ C and w ∈ C − C.

Exercise 2G.5. In the setting of Theorem 2G.2, condition (b) is fulfilled in particu-
lar if there exists μ > 0 such that

(18) ∇2xx g( p̄, x̄)w, w ≥ μ |w|2 for all w ∈ C − C,

and then lip (s; p̄) ≤ μ −1 . If C is polyhedral, the additional conclusion holds that,
for all p in some neighborhood of p̄, there is a strong local minimum in problem (6)
at the point x = s(p).
Guide. Apply Proposition 2G.4 to the function ḡ in the auxiliary minimization
problem (10). Get from this that s̄ coincides with S̄, which is single-valued and
Lipschitz continuous on IRn with Lipschitz constant μ −1 . In the polyhedral case,
also apply Theorem 2G.3, arguing that (18) entails (15).
Observe that because C − C is a convex set containing 0, the condition in (18)
holds for all w ∈ C − C if it holds for all w ∈ C − C with |w| sufficiently small.
We turn now to minimization over sets which need not be convex but are specified
by a system of constraints. A first-order necessary condition for a minimum in that
case was developed in a very general manner in Theorem 2A.8. Here, we restrict
ourselves to the most commonly treated problem of nonlinear programming, where
the format is to

≤ 0 for i ∈ [1, s],
(19) minimize g0 (x) over all x satisfying gi (x)
= 0 for i ∈ [s + 1, m].

In order to bring second-order conditions for optimality into the picture, we assume
that the functions g0 , g1 , . . . , gm are twice continuously differentiable on IRn .
The basic first-order condition in this case has been worked out in detail in Sec-
tion 2A as a consequence of Theorem 2A.8. It concerns the existence, relative to
x, of a multiplier vector y = (y1 , . . . , ym ) fulfilling the Karush–Kuhn–Tucker condi-
tions:
2 Implicit Function Theorems for Variational Problems 119

≤ 0 for i ∈ [1, s] with y j = 0,
y ∈ IRs+ × IRm−s , gi (x)
(20) = 0 for all other i ∈ [1, m],
∇g0 (x) + y1 ∇g1 (x) + · · · + ym ∇gm (x) = 0.

This existence is necessary for a local minimum at x as long as x satisfies the con-
straint qualification requiring that the same conditions, but with the term ∇g0 (x)
suppressed, can’t be satisfied with y = 0. It is sufficient for a global minimum at x if
g0 , g1 , . . . , gs are convex and gs+1 , . . . , gm are affine. However, we now wish to take a
second-order approach to local sufficiency, rather than rely on convexity for global
sufficiency.
The key for us will be the fact, coming from Theorem 2A.9, that (20) can be
identified in terms of the Lagrangian function

(21) L(x, y) = g0 (x) + y1 g1 (x) + · · · + ym gm (x)

with a certain variational inequality for a continuously differentiable function f :

IRn × IRm → IRn × IRm and polyhedral convex cone E ⊂ IRn × IRm , namely

f (x, y) = (∇x L(x, y), −∇y L(x, y)),
(22) f (x, y) + NE (x, y)  (0, 0), where
E = IRn × [IRs+ × IRm−s ].

Because our principal goal is to illustrate the application of the results in the preced-
ing sections, rather than push consequences for optimization theory to the limit, we
will only deal with this variational inequality under an assumption of linear inde-
pendence for the gradients of the active constraints. A constraint in (20) is inactive
at x if it is an inequality constraint with gi (x) < 0; otherwise it is active at x.

Theorem 2G.6 (second-order optimality in nonlinear programming). Let x̄ be a

point satisfying the constraints in (19). Let I(x̄) be the set of indices i of the ac-
tive constraints at x̄, and suppose that the gradients ∇gi (x̄) for i ∈ I(x̄) are linearly
independent. Let K consist of the vectors w ∈ IRn satisfying

≤ 0 for i ∈ I(x̄) with i ≤ s,
(23) ∇gi (x̄), w
= 0 for all other i ∈ I(x̄) and also for i = 0.

(a) (necessary condition) If x̄ furnishes a local minimum in problem (19), then a

multiplier vector ȳ exists such that (x̄, ȳ) not only satisfies the variational inequality
(22) but also has

(24) w, ∇2xx L(x̄, ȳ)w ≥ 0 for all w ∈ K.

(b) (sufficient condition) If a multiplier vector ȳ exists such that (x̄, ȳ) satisfies
the conditions in (20), or equivalently (22), and if (24) holds with strict inequality
when w = 0, then x̄ furnishes a local minimum in (19). Indeed, it furnishes a strong
local minimum in the sense of there being an ε > 0 such that
ε
(25) g0 (x) ≥ g0 (x̄) + |x − x̄|2 for all x near x̄ satisfying the constraints.
2
120 2 Implicit Function Theorems for Variational Problems

Proof. The linear independence of the gradients of the active constraints guaran-
tees, among other things, that x̄ satisfies the constraint qualification under which
(22) is necessary for local optimality.
In the case of a local minimum, as in (a), we do therefore have the variatio-
nal inequality (22) fulfilled by x̄ and some vector ȳ; and of course (22) holds by
assumption in (b). From this point on, therefore, we can concentrate on just the
second-order parts of (a) and (b) in the framework of having x̄ and ȳ satisfying (20).
In particular then, we have

(26) −∇g0 (x̄) = ȳ1 ∇g1 (x̄) + · · · + ȳm ∇gm (x̄),

where the multiplier vector ȳ is moreover uniquely determined by the linear inde-
pendence of the gradients of the active constraints and the stipulation in (20) that
inactive constraints get coefficient 0.
Actually, the inactive constraints play no role around x̄, so we can just as well
assume, for simplicity of exposition in our local analysis, that every constraint is
active at x̄: we have gi (x̄) = 0 for i = 1, . . . , m. Then, on the level of first-order
conditions, we just have the combination of (20), which corresponds to ∇x L(x̄, ȳ) =
0, and the requirement that ȳi ≥ 0 for i = 1, . . . , s. In this simplified context, let

≤ 0 for i = 1, . . . , s,
(27) T = set of all w ∈ IRn satisfying ∇gi (x̄), w
= 0 for i = s + 1, . . ., m,

so that the cone K described by (23) can be expressed in the form

  
(28) K = w ∈ T  ∇g0 (x̄), w = 0 .

The rest of our argument will rely heavily on the classical inverse function the-
orem, 1A.1. Our assumption that the vectors ∇gi (x̄) for i = 1, . . . , m are linearly
independent in IRn entails of course that m ≤ n. These vectors can be supplemented,
if necessary, by vectors ak for k = 1, . . . , n − m so as to form a basis for IRn . Then, by
setting gm+k (x) = ak , x − x̄, we get functions gi for i = m + 1, . . ., n such that for

g : IRn → IRn with g(x) = (g1 (x), . . . , gm (x), gm+1 (x), . . . , gn (x))

we have g(x̄) = 0 and ∇g(x̄) nonsingular. We can view this as providing, at least
locally around x̄, a change of coordinates g(x) = u = (u1 , . . . , un ), x = s(u) (for a
localization s of g−1 around 0 for x̄) in which x̄ corresponds to 0 and the constraints
in (19) correspond to linear constraints

ui ≤ 0 for i = 1, . . . , s, ui = 0 for i = s + 1, . . . , m

(with no condition on ui for i = m + 1, . . . , n), which specify a polyhedral convex set

D in IRn . Problem (19) is thereby transformed in a local sense into minimizing over
this set D the twice continuously differentiable function f (u) = g0 (s(u)), and we
are concerned with whether or not there is a local minimum at ū = 0. The necessary
2 Implicit Function Theorems for Variational Problems 121

and sufficient conditions in Theorem 2G.1 are applicable to this and entail having
− f (0) belong to ND (0). It will be useful to let ỹ stand for (ȳ, 0, . . . , 0) ∈ IRn .
The inverse function theorem reveals that the Jacobian ∇s(0) is ∇g(x̄)−1 . We
have ∇ f (0) = ∇g0 (x̄)∇s(0) by the chain rule, and on the other hand −∇g0 (x̄) =
ỹ∇g(x̄) by (26), and therefore ∇ f (0) = −ỹ. The vectors w belonging to the set T in
(27) correspond one-to-one with the vectors z ∈ D through ∇g(x̄)w = z, and under
this, through (28), the vectors w ∈ K correspond to the vectors z ∈ D such that
z, ỹ = 0, i.e., the vectors in the critical cone KD (0, ỹ) = KD (0, −∇ f (0)).
The second-order conditions in Theorem 2G.1, in the context of the transformed
version of problem (19), thus revolve around the nonnegativity or positivity of
z, ∇2 f (0)z for vectors z ∈ KD (0, ỹ). It will be useful that this is the same as the
nonnegativity or positivity or z, ∇2 h(0)z for the function

h(u) = f (u) + ỹ, u = f (u) + ȳ, Pu = L(s(u), ȳ),

where P is the projection from (u1 , . . . , um , um+1 , . . . , un ) to (u1 , . . . , um ). Further-

more,

z, ∇2 h(0)z = ϕ  (0) for the function ϕ (t) = h(tz) = L(s(tz), ȳ).

Fix any nonzero z ∈ KD (0, ỹ) and the corresponding w ∈ K, given by w =

∇s(0)z = ∇g(x̄)−1 z. Our task is to demonstrate that actually

(29) ϕ  (0) = w, ∇2xx L(x̄, ȳ)w.

Let x(t) = s(tz), so that x(0) = x̄ and x (0) = w. We have

ϕ (t) = L(x(t), ȳ),

ϕ  (t) = ∇x L(x(t), ȳ), x (t),
ϕ  (t) = w, ∇2xx L(x(t), ȳ), x (t) + ∇x L(x(t), ȳ), x (t),

hence ϕ  (0) = w, ∇2xx L(x̄, ȳ), w + ∇x L(x̄, ȳ), x (0). But ∇x L(x̄, ȳ) = 0 from the
first-order conditions. Thus, (29) holds, as claimed.
The final assertion of part (b) automatically carries over from the corresponding
assertion of part (b) of Theorem 2G.1 under the local change of coordinates that we
utilized.

Exercise 2G.7. In the context of Theorem 2G.6, let ȳ be a multiplier associated with
x̄ through the first-order condition (22). Let I0 (x̄, ȳ) be the set of indices i ∈ I(x̄) such
that i ≤ s and ȳi = 0. Then an equivalent description of the cone K in the second-
order conditions is that

≤ 0 for i ∈ I0 (x̄, ȳ),
w ∈ K ⇐⇒ ∇gi (x̄), w
= 0 for i ∈ I(x̄)\I0 (x̄, ȳ).

Guide. Utilize the fact that −∇g0 (x̄) = ȳ1 ∇g1 (x̄) + · · · + ȳm ∇gm (x̄) with ȳi ≥ 0 for
i = 1, . . . , s.
122 2 Implicit Function Theorems for Variational Problems

The alternative description in 2G.7 lends insights in some situations, but it makes
K appear to depend on ȳ, whereas in reality it doesn’t.
Next we take up the study of a parameterized version of the nonlinear program-
ming problem in the form

≤ 0 for i ∈ [1, s],
(30) minimize g0 (p, x) over all x satisfying gi (p, x)
= 0 for i ∈ [s + 1, m],

where the functions g0 , g1 , . . . , gm are twice continuously differentiable from IRd ×

IRn to IR. The Lagrangian function is now

(31) L(p, x, y) = g0 (p, x) + y1 g1 (p, x) + · · · + ym gm (p, x)

and the variational inequality capturing the associated first-order conditions is

(32) 
f (p, x, y) = (∇x L(p, x, y), −∇y L(p, x, y)),
f (p, x, y) + NE (x, y)  (0, 0), where
E = IRn × [IRs+ × IRm−s ].

The pairs (x, y) satisfying this variational inequality are the Karush–Kuhn–Tucker
pairs for the problem specified by p in (30). The x components of such pairs might
or might not give a local minimum according to the circumstances in Theorem 2G.6
(or whether certain convexity assumptions are fulfilled), and indeed we are not im-
posing a linear independence condition on the constraint gradients in (30) of the
kind on which Theorem 2G.6 was based. But these x’s serve anyway as station-
ary points and we wish to learn more about their behavior under perturbations by
studying the Karush–Kuhn–Tucker mapping S : IRd → IRn × IRm defined by
  
(33) S(p) = (x, y)  f (p, x, y) + NE (x, y)  (0, 0) .

Once more, an auxiliary problem will be important with respect to a choice of p̄

and a pair (x̄, ȳ) ∈ S( p̄). To formulate it, we let

ḡ0 (w) = L( p̄, x̄, ȳ) + ∇x L( p̄, x̄, ȳ), w + 12 w, ∇2xx L( p̄, x̄, ȳ)w,
ḡi (w) = gi ( p̄, x̄) + ∇x gi ( p̄, x̄), w for i = 1, . . . , m,

and introduce the notation

  
I = i ∈ [1, m]  gi ( p̄, x̄) = 0 ⊃ {s + 1,  . . ., m},
(34) I0 =  i ∈ [1, s]  gi ( p̄, x̄) = 0and ȳi = 0 ⊂ I,
I1 = i ∈ [1, s]  gi ( p̄, x̄) < 0 .

The auxiliary problem, depending on a tilt parameter vector v but also now an addi-
tional parameter vector u = (u1 , . . . , um ), is to
2 Implicit Function Theorems for Variational Problems 123

minimize ḡ0 (w) − v, w over all w satisfying

%
= 0 for i ∈ I\I0 ,
(35)
ḡi (w) + ui ≤ 0 for i ∈ I0 ,
free for i ∈ I1 ,
where “free” means unrestricted. (The functions ḡi for i ∈ I1 play no role in this
problem, but it will be more convenient to carry them through in this scheme than
to drop them.)
In comparison with the auxiliary problem introduced earlier in (10) with respect
to minimization over a set C, it’s apparent that a second-order expansion of L rather
than g0 has entered, but merely first-order expansions of the constraint functions
g1 , . . . , gm . In fact, only the quadratic part of the Lagrangian expansion matters,
inasmuch as ∇x L( p̄, x̄, ȳ) = 0 by the first-order conditions. The Lagrangian for the
problem in (35) depends on the parameter pair (v, u) and involves a multiplier vector
z = (z1 , . . . , zm ):

ḡ0 (w) − v, w + z1 [ḡ1 (w) + u1 ] + · · · + zm [ḡm (w) + um ] =: L̄(w, z) − v, w + z, u.

The corresponding first-order conditions are given by the variational inequality

f¯(w, z) − (v, u) + NĒ (w, z)  (0, 0), where

(36) f¯(w, z) = (∇w L̄(w, z), −∇z L̄(w,z)), Ē = IRn × W, with

≥ 0 for i ∈ I0 ,
z = (z1 , . . . , zm ) ∈ W ⇐⇒ zi
= 0 for i ∈ I1 ,

which translate into the requirements that

∇2xx L( p̄,
x̄, ȳ)w + z1 ∇x g1 ( p̄, x̄) + · · · + zm ∇x gm ( p̄, x̄) − v = 0,
(37) ≥ 0 for i ∈ I0 having ḡi (w) + ui = 0,
with zi
= 0 for i ∈ I0 having ḡi (w) + ui < 0 and for i ∈ I1 .

We need to pay heed to the auxiliary solution mapping S̄ : IRn × IRn → → IRn × IRm
defined by
     
(38) S̄(v, u) = (w, z)  f¯(w, z) + NĒ (w, z)  (v, u) = (w, z)  satisfying (37) ,

which has
(0, 0) ∈ S̄(0, 0).
The following subspaces will enter our analysis of the properties of the mapping S̄:
  
M + =  w ∈ IRn  w ⊥ ∇x gi ( p̄, x̄) for all i ∈ I\I
 0 ,
(39)
M − = w ∈ IRn  w ⊥ ∇x gi ( p̄, x̄) for all i ∈ I .
124 2 Implicit Function Theorems for Variational Problems

Theorem 2G.8 (implicit function theorem for stationary points). Let (x̄, ȳ) ∈ S( p̄)
for the mapping S in (33), constructed from functions gi that are twice continuously
differentiable. Assume for the auxiliary mapping S̄ in (38) that

S̄ has a Lipschitz continuous single-valued
(40)
localization s̄ around (0, 0) for (0, 0).

Then S has a Lipschitz continuous single-valued localization s around p̄ for (x̄, ȳ),
and this localization s is semidifferentiable at p̄ with semiderivative given by
⎡ 2 ⎤
∇xp L( p̄, x̄, ȳ)
⎢ −∇p g1 ( p̄, x̄) ⎥
⎢ ⎥
(41) Ds( p̄)(q) = s̄(−Bq), where B = ∇ p f ( p̄, x̄, ȳ) = ⎢ ⎥.
⎣ ... ⎦
−∇p gm ( p̄, x̄)

Moreover the condition in (40) is necessary for the existence of a Lipschitz continu-
ous single-valued localization of S around p̄ for (x̄, ȳ) when the (n + m) × d matrix B
has rank n + m. In particular, S̄ is sure to satisfy (40) when the following conditions
are both fulfilled:
(a) the gradients ∇x gi ( p̄, x̄) for i ∈ I are linearly independent,
(b) w, ∇2xx L( p̄, x̄, ȳ)w > 0 for every nonzero w ∈ M + with ∇2xx L( p̄, x̄, ȳ)w ⊥ M − ,
with M + and M− as in (39).
On the other hand, (40) always entails at least (a).
Proof. This is obtained by applying 2E.1 with the additions in 2E.6 and 2E.8 to the
variational inequality (32). Since ∇y L(p, x, y) = g(p, x) for g(p, x) = (g1 (p, x), . . . ,
gm (p, x)), the Jacobian in question is
 2 
∇xx L( p̄, x̄, ȳ) ∇x g( p̄, x̄)T
(42) ∇(x,y) f ( p̄, x̄, ȳ) = .
−∇x g( p̄, x̄) 0

In terms of polyhedral convex cone Y = IRs+ × IRm−s , the critical cone to the polyhe-
dral convex cone set E is

(43) KE (x̄, ȳ, − f ( p̄, x̄, ȳ)) = IRn × W

for the polyhedral cone W in (36). By taking A to be the matrix in (42) and K to be
the cone in (43), the auxiliary mapping S̄ can be identified with (A + NK )−1 in the
framework of Theorem 2E.6. (The w and v in that result have here turned into pairs
(w, z) and (v, u).)
This leads to all the conclusions except for establishing that (40) implies (a) and
working out the details of the sufficient condition provided by Theorem 2E.6. To
verify that (40) implies (a), consider any ε > 0 and let
m 
ε for i ∈ I0 ,
vε = ∑ zεi ∇x gi ( p̄, x̄), zεi =
i=1 0 otherwise.
2 Implicit Function Theorems for Variational Problems 125

Then, as seen from the conditions in (37), we have (0, zε ) ∈ S̄(vε , 0). If (a) didn’t
hold, we would also have ∑m i=1 ζi ∇x g( p̄, x̄) = 0 for some coefficient vector ζ = 0
with ζi = 0 when i ∈ I1 . Then for every δ > 0 small enough that ε + δ ζi ≥ 0 for
all i ∈ I0 , we would also have (0, zε + δ ζ ) ∈ S̄(vε , 0). Since ε and δ can be chosen
arbitrarily small, this would contradict the single-valuedness in (40). Thus, (a) is
necessary for (40).
To come next to an understanding of what the sufficient condition in 2E.6 means
here, we observe in terms of Y = IRs+ × IRm−s that

KE+ (x̄, ȳ, − f ( p̄, x̄, ȳ)) = IRn × KY+ (ȳ, g( p̄, x̄)),
KE− (x̄, ȳ, − f ( p̄, x̄, ȳ)) = IRn × KY− (ȳ, g( p̄, x̄)),

where
z ∈ KY+ (ȳ, g( p̄, x̄)) ⇐⇒ zi = 0 for i ∈ I1 ,
z ∈ KY− (ȳ, g( p̄, x̄)) ⇐⇒ zi = 0 for i ∈ I0 ∪ I1 .
In the shorthand notation
+ + − −
H = ∇2xx L( p̄, x̄, ȳ), K = KE (x̄, ȳ, − f ( p̄, x̄, ȳ)), K = KE (x̄, ȳ, − f ( p̄, x̄, ȳ)),

our concern is to have (w, z), A(w, z) > 0 for every (w, z) ∈ K + with A(w, z) ⊥ K −
and (w, z) = (0, 0). It’s clear from (41) that

(w, z), A(w, z) = w, Hw,

(w, z) ∈ K + ⇐⇒ zi = 0 for i ∈ I1 ,
A(w, z) ⊥ K − ⇐⇒ Hw + ∇x g( p̄, x̄)T z = 0 and ∇x g( p̄, x̄)w ⊥ KY− .

Having ∇x g( p̄, x̄)w ⊥ KY− corresponds to having w ⊥ ∇x gi ( p̄, x̄) for all i ∈ I \ I0 ,
which means w ∈ M + . On the other hand, having Hw+ ∇x g( p̄, x̄)T z = 0 corresponds
to having Hw = −(z1 ∇x g1 ( p̄, x̄) + · · · + zm ∇x gm ( p̄, x̄)). The sufficient condition in
2E.6 boils down, therefore, to the requirement that

w, Hw > 0 when (w, z) = (0, 0), w ∈ M , Hw = − ∑i∈I zi ∇x gi ( p̄, x̄).
+

In particular this requirement has to cover cases where w = 0 but z = 0. That’s

obviously equivalent to the linear independence in (a). Beyond that, we need only
observe that expressing Hw in the manner indicated corresponds simply to having
Hw ⊥ M− . Thus, the sufficient condition in Theorem 2E.6 turns out to be the com-
bination of (a) with (b).

Our final topic concerns the conditions under which the mapping S in (33) de-
scribes perturbations not only of stationarity, but also of local minimuma.
126 2 Implicit Function Theorems for Variational Problems

Theorem 2G.9 (implicit function theorem for local minima). Suppose in the setting
of the parameterized nonlinear programming problem (30) for twice continuously
differentiable functions gi and its Karush–Kuhn–Tucker mapping S in (33) that the
following conditions hold in the notation coming from (34):
(a) the gradients ∇x gi ( p̄, x̄) for i ∈ I are linearly independent, and
(b) w, ∇2xx L( p̄, x̄, ȳ)w > 0 for every w = 0 in the subspace M + in (39).
Then not only does S have a localization s with the properties laid out in Theorem
2G.8, but also, for every p in some neighborhood of p̄, the x component of s(p)
furnishes a strong local minimum in (30). Moreover, (a) and (b) are necessary for
this additional conclusion when n + m is the rank of the (n + m) × d matrix B in (41).
Proof. Sufficiency. Condition (b) here is a stronger assumption than (b) of Theorem
2G.8, so we can be sure that (a) and (b) guarantee the existence of a localization
s possessing the properties in that result. Moreover (b) implies satisfaction of the
sufficient condition for a local minimum at x̄ in Theorem 2G.6, inasmuch as the cone
K in that theorem is obviously contained in the set of w such that ∇gi ( p̄, x̄), w = 0
for all i ∈ I \ I0 . We need to demonstrate, however, that this local minimum property
persists in passing from p̄ to nearby p.
To proceed with that, denote the two components of s(p) by x(p) and y(p), and
let I(p), I0 (p) and I1 (p) be the index sets which correspond to x(p) as I, I0 and
I1 do to x̄, so that I(p) consists of the indices i ∈ {1, . . . , m} with gi (p, x(p)) = 0,
and I(p) \ I0 (p) consists of the indices i ∈ I(p) having yi (p) > 0 for inequality
constraints, but I1 (p) consists of the indices of the inequality constraints having
gi (p, x(p)) < 0. Consider the following conditions, which reduce to (a) and (b) when
p = p̄:
(a(p)) the gradients ∇x gi (p, x(p)) for i ∈ I(p) are linearly independent,
(b(p)) w, ∇2xx L(p, x(p), y(p))w > 0 for every w = 0 such that w ⊥ ∇x gi (p, x(p))
for all i ∈ I(p) \ I0 (p).
Since x(p) and y(p) tend toward x( p̄) = x̄ and y( p̄) = ȳ as p → p̄, the fact that
yi (p) = 0 for i ∈ I1 (p) and the continuity of the gi ’s ensure that

I(p) ⊂ I and I(p) \ I0(p) ⊃ I \ I0 for p near enough to p̄.

Through this and the fact that ∇x gi (p, x(p)) tends toward ∇x gi ( p̄, x̄) as p goes to p̄,
we see that the linear independence in (a) entails the linear independence in (a(p))
for p near enough to p̄. Indeed, not only (a(p)) but also (b(p)) must hold, in fact in
the stronger form that there exist ε > 0 and a neighborhood Q of p̄ for which

w, ∇2xx L(p, x(p), y(p))w > ε

when |w| = 1 and w ⊥ ∇x gi (p, x(p)) for all i ∈ I(p) \ I0 (p). Indeed, otherwise there
would be sequences of vectors pk → p̄ and wk → w violating this condition for
εk → 0, and this would lead to a contradiction of (b) in view of the continuous
dependence of the matrix ∇2xx L(p, x(p), y(p)) on p.
Of course, with both (a(p)) and (b(p)) holding when p is in some neighborhood
of p̄, we can conclude through Theorem 2G.6, as we did for x̄, that x(p) furnishes a
2 Implicit Function Theorems for Variational Problems 127

strong local minimum for problem (30) for such p, since the cone

∇x gi (p, x(p)), w ≤ 0 for i ∈ I(p) with i ≤ s,
K(p) = set of w satisfying
∇x gi (p, x(p)), w = 0 for i = s + 1, . . ., m and i = 0

lies in the subspace formed by the vectors w with ∇x gi (p, x(p)), w = 0 for all
i ∈ I(p) \ I0 (p).
Necessity. Suppose that S has a Lipschitz continuous single-valued localization
s around p̄ for (x̄, ȳ). We already know from Theorem 2G.8 that, under the rank
condition in question, the auxiliary mapping S̄ in (38) must have such a localization
around (0, 0) for (0, 0), and that this requires the linear independence in (a). Under
the further assumption now that x(p) gives a local minimum in problem (30) when
p is near enough to p̄, we wish to deduce that (b) must hold as well. Having a local
minimum at x(p) implies that the second-order necessary condition for optimality
in Theorem 2G.6 is satisfied with respect to the multiplier vector y(p):

(44) w, ∇2xx L(p, x(p), y(p))w ≥ 0 for all w ∈ K(p) when p is near to p̄.

We will now find a value of the parameter p close to p̄ such that (x(p), y(p)) = (x̄, ȳ)
and K(p) = M + . If I0 = 0/ there is nothing to prove. Let I0 = 0. / The rank condi-
tion on the Jacobian B = ∇ p f ( p̄, x̄, ȳ) provides through Theorem 1F.6 (for k = 1)
the existence of p(v, u), depending continuously on some (v, u) in a neighborhood
of (0, −g( p̄, x̄)), such that f (p(v, u), x̄, ȳ) = (v, u), i.e., ∇x L(p(v, u), x̄, ȳ) = v and
−g(p(v, u), x̄) = u. For an arbitrarily small ε > 0, let the vector uε have uεi = −ε
for i ∈ I0 but uεi = 0 for all other i. Let pε = p(0, uε ). Then ∇x L(pε , x̄, ȳ) = 0 with
gi (pε , x̄) = 0 for i ∈ I \ I0 but gi (pε , x̄) < 0 for i ∈ I0 as well as for i ∈ I1 . Thus,
I(pε ) = I \ I0 , I0 (pε ) = 0,
/ I1 (pε ) = I0 ∪ I1 , and (x̄, ȳ) ∈ S(pε ) and, moreover, (x̄, ȳ)
furnishes a local minimum in (30) for p = pε , moreover with K(pε ) coming out to
be the subspace
  
M (pε ) = w  w ⊥ ∇x gi (pε , x̄) for all i ∈ I \ I0 .
+

In consequence of (44) we therefore have

w, ∇2xx L(pε , x̄, ȳ)w ≥ 0 for all w ∈ M (pε ),

+

whereas we are asking in (b) for this to hold with strict inequality for w = 0 in the
case of M + = M + ( p̄).
We know that pε → p̄ as ε → 0. Owing to (a) and the continuity of the functions
gi and their derivatives, the gradients ∇x gi (pε , x̄) for i ∈ I must be linearly inde-
pendent when ε is sufficiently small. It follows from this that any w in M + can be
approximated as ε → 0 by vectors wε belonging to the subspaces M + (pε ). In the
limit therefore, we have at least that
+
(45) w, ∇2xx L( p̄, x̄, ȳ)w ≥ 0 for all w ∈ M .
128 2 Implicit Function Theorems for Variational Problems

How are we to conclude strict inequality when w = 0? It’s important that the matrix
H = ∇2xx L( p̄, x̄, ȳ) is symmetric. In line with the positive semidefiniteness in (45),
any w̄ ∈ M + with w̄, H w̄ = 0 must have H w̄ ⊥ M + . But then in particular, the
auxiliary solution mapping S̄ in (38) would have (t w̄, 0) ∈ S̄(0, 0) for all t ≥ 0, in
contradiction to the fact, coming from Theorem 2G.8, that S̄(0, 0) contains only
(0, 0) in the current circumstances.
2 Implicit Function Theorems for Variational Problems 129

Commentary
The basic facts about convexity, polyhedral sets, and tangent and normal cones
given in Section 2A are taken mainly from Rockafellar [1970]. Robinson’s implicit
function theorem was stated and proved in Robinson [1980], where the author was
clearly motivated by the problem of how the solutions of the standard nonlinear
programming problem depend on parameters, and he pursued this goal in the same
paper.
At that time it was already known from the work of Fiacco and McCormick
[1968] that under the linear independence of the constraint gradients and the stan-
dard second-order sufficient condition, together with strict complementarity slack-
ness at the reference point (which means that there are no inequality constraints
satisfied as equalities that are associated with zero Lagrange multipliers), the solu-
tion mapping for the standard nonlinear programming problem has a smooth single-
valued localization around the reference point. The proof of this result was based on
the classical implicit function theorem, inasmuch as under strict complementarity
slackness the Karush–Kuhn–Tucker system turns into a system of equations locally.
Robinson looked at the case when the strict complementarity slackness is violated,
which may happen, as already noted in 2B, when the “stationary point trajectory”
hits the constraints. Based on his implicit function theorem, which actually reached
far beyond his immediate goal, Robinson proved, still in his paper from 1980, that
under a stronger form of the second-order sufficient condition, together with linear
independence of the constraint gradients, the solution mapping of the standard non-
linear programming problem has a Lipschitz continuous single-valued localization
around the reference point; see Theorem 2G.9 for an updated statement.
This result was a stepping stone to the subsequent extensive development of sta-
bility analysis in optimization, whose maturity came with the publication of the
books Bank, Guddat, Klatte, Kummer and Tammer [1983], Levitin [1992], Bonnans
and Shapiro [2000], Klatte and Kummer [2002] and Facchinei and Pang [2003].
Robinson’s breakthrough in the stability analysis of nonlinear programming was
in fact much needed for the emerging numerical analysis of variational problems
more generally. In his paper from 1980, Robinson noted the thesis of his Ph.D.
student Josephy [1979], who proved that strong regularity yields local quadratic
convergence of Newton’s method for solving variational inequalities, a method
whose version for constrained optimization problems is well known as the sequen-
tial quadratic programming (SQP) method in nonlinear programming.
Quite a few years after Robinson’s theorem was published, it was realized that
the result could be used as a tool in the analysis of a variety of variational prob-
lems, and beyond. Alt [1990] applied it to optimal control, while in Dontchev and
Hager [1993], and further in Dontchev [1995b], the statement of Robinson’s theo-
rem was observed actually to hold for generalized equations of the form 2B(1) for
an arbitrary mapping F, not just a normal cone mapping. Variational inequalities
thus serve as an example, not a limitation. Important applications, e.g. to conver-
gence analysis of algorithms and discrete approximations to infinite-dimensional
130 2 Implicit Function Theorems for Variational Problems

variational problems, came later. In the explosion of works in this area in the 80’s
and 90’s Robinson’s contribution, if not forgotten, was sometimes taken for granted.
More about this will come out in Chapter 6.
The presentation of the material in Section 2B mainly follows Dontchev and
Rockafellar [2009a], while that in Section 2C comes from Dontchev and Rockafellar
[2001]. In Section 2D, we used some facts from the books of Facchinei and Pang
[2003] (in particular, 2D.5) and Scholtes [1994]. Theorem 2E.6 is a particular case
of a result in Dontchev and Rockafellar [1996].
Sections 2F and 2G give an introduction to the theory of monotone mappings
which for its application to optimization problems goes back to Rockafellar [1976a]
and [1976b]. Much more about this kind of monotonicity and its earlier history can
be found in Chapter 12 of the book of Rockafellar and Wets [1998]. The stability
analysis in 2G uses material from Dontchev and Rockafellar [1996,1998], but some
versions of these results could be extracted from earlier works.
Chapter 3
Regularity Properties of Set-valued Solution
Mappings

In the concept of a solution mapping for a problem dependent on parameters,

whether formulated with equations or something broader like variational inequal-
ities, we have always had to face the possibility that solutions might not exist, or
might not be unique when they do exist. This goes all the way back to the setting of
the classical implicit function theorem. In letting S(p) denote the set of all x satisfy-
ing f (p, x) = 0, where f is a given function from IRd × IRn to IRm , we cannot expect
to be defining a function S from IRd to IRn , even when m = n. In general, we only
get a set-valued mapping S. However, this mapping S could have a single-valued
localization s with properties of continuity or differentiability. The study of such
localizations, as “subfunctions” within a set-valued mapping, has been our focus so
far, but now we open up to a wider view.
There are plenty of reasons, already in the classical context, to be interested in
localizations of solution mappings   without insisting
 on single-valuedness. For in-
stance, in the case of S(p) = x  f (p, x) = 0 with f going from IRd × IRn to IRm
and m < n, it can be anticipated for a choice of p̄ and x̄ with x̄ ∈ S( p̄), under assump-
tions on ∇x f ( p̄, x̄), that a graphical localization S0 of S exists around ( p̄, x̄) such that
S0 (p) is an (n − m)-dimensional manifold which varies with p. What generalizations
of the usual notions of continuity and differentiability might help in understanding,
and perhaps quantifying, this dependence on p?
Such challenges in dealing with the dependence of a set on the parameters which
enter its definition carry over to solution mappings to variational inequalities, and
also to problems of a broader character centered on constraint systems. Just as the
vector equation f (p, x) = 0 for
f : IRd × IRn → IRm with f (p, x) = ( f1 (p, x), . . . , fm (p, x))

can be viewed as standing for a system of scalar equations

f i (p, x) = 0 for i = 1, . . . , m,

we can contemplate vector representations of mixed systems of inequalities and

equations like
A.L. Dontchev and R.T. Rockafellar, Implicit Functions and Solution Mappings: A View 131
from Variational Analysis, Springer Monographs in Mathematics,
DOI 10.1007/978-0-387-87821-8 3,  c Springer Science+Business Media, LLC 2009
132 3 Regularity Properties of Set-valued Solution Mappings

≤ 0 for i = 1, . . . , s,
(1) fi (p, x)
= 0 for i = s + 1, . . . , m,

which are important in optimization. Such a system takes the form

f (p, x) − K  0 for K = IRs− × {0}m−s.

In fact this is an instance of a parameterized generalized equation:

(2) f (p, x) + F(x)  0 with F a constant mapping, F(x) ≡ −K.

In studying the behavior of the corresponding solution mapping S : IRd →→ IRn given
by
  
(3) S(p) = x  x satisfies (2) (covering (1) as a special case),

we are therefore still, very naturally, in the realm of the “extended implicit function
theory” we have been working to build up.
Here, F is not a normal cone mapping NC , so we are not dealing with a variational
inequality. The results in Chapter 2 for solution mappings to parameterized gener-
alized equations would anyway, in principle, be applicable, but in this framework
they miss the mark. The trouble is that those results focus on the prospects of find-
ing single-valued localizations of a solution mapping, especially ones that exhibit
Lipschitz continuity. For a solution mapping S as in (3), coming from a general-
ized equation as in (2), single-valued localizations are unlikely to exist at all (apart
from the pure equation case with m = n) and aren’t even a topic of serious concern.
Rather, we are confronted with a “varying set” S(p) which cannot be reduced locally
to a “varying point.” That could be the case even if, in (2), F(x) is not a constant set
but a sort of continuously moving or deforming set. What we are looking for is not
so much a generalized implicit function theorem, but an implicit mapping theorem,
the distinction being that “mappings” truly embrace set-valuedness.
To understand the behavior of such a solution mapping S, whether qualitatively
or quantitatively, we have to turn to other concepts, beyond those in Chapter 2.
Our immediate task, in Sections 3A, 3B, 3C and 3D, is to introduce the notions
of Painlevé–Kuratowski convergence and Pompeiu–Hausdorff convergence for se-
quences of sets, and to utilize them in developing properties of continuity and
Lipschitz continuity for set-valued mappings. In tandem with this, we gain impor-
tant insights into the solution mappings (3) associated with constraint systems as in
(1) and (2), especially for cases where f is affine. We also obtain by-products con-
cerning the behavior of various mappings associated with problems of optimization.
In Section 3E, however, we open a broader investigation in which the Aubin
property, serving as a sort of localized counterpart to Lipschitz continuity for set-
valued mappings, is tied to the concept of metric regularity, which directly relates
to estimates of distances to solutions. The natural context for this is the study of
how properties of a set-valued mapping correspond to properties of its set-valued
inverse, or in other words, the paradigm of the inverse function theorem. We are
3 Regularity Properties of Set-valued Solution Mappings 133

able nevertheless to return in Section 3F to the paradigm of the implicit function

theorem, based on a stability property of metric regularity fully developed later in
Chapter 5. Powerful results, applicable to fully set-valued solution mappings (2)(3)
even when F is not just a constant mapping, are thereby obtained. Sections 3G
and 3I then take these ideas back to situations where single-valuedness is available
in a localization of a solution mapping, at least at the reference point, showing how
previous results such as those in 2B can thereby be amplified. Section 3H reveals that
a set-valued version of calmness does not similarly submit to the implicit function
theorem paradigm.
134 3 Regularity Properties of Set-valued Solution Mappings

3A. Set Convergence

Various continuity properties of set-valued mappings will be essential for the devel-
opments in this chapter. To lay the foundation for them we must first introduce two
basic concepts: convergence of sets and distance between sets.
The set of all natural numbers k = 1, 2, . . ., will be denoted by IN. The collection
of all subsets N of IN such that IN \ N is finite will be denoted by N , whereas the
collection of all infinite N ⊂ IN will be denoted by N  . This scheme is designed
for convenience in handling subsequences of a given sequence. For instance, if we
have a sequence {xk }∞ k=1 of points in IR , the notation {x }k∈N for either N ∈ N
n k

or N ∈ N designates a subsequence. In the first case it is a subsequence which
coincides with the full sequence beyond some k0 , whereas in the second case it is a
general subsequence. Limits as k → ∞ with k ∈ N will be indicated by limk∈N , or in
N
terms of arrows by →, and so forth.
For a sequence {rk }∞ k=1 in IR, the limit as k → ∞ may or may not exist—even
though we always include ∞ and −∞ as possible limit values in the obvious sense.
However, the upper limit, or “limsup,” and the lower limit, or “liminf,” do always
exist, as defined by
lim sup rk = lim sup rm ,
k→∞ k→∞ m≥k
lim inf rk = lim inf rm .
k→∞ k→∞ m≥k

An alternative description of these values is that lim supk→∞ rk is the highest r for
N
which there exists N ∈ N  such that rk → r, whereas lim infk→∞ rk is the lowest such
r. The limit itself exists if and only if these upper and lower limits coincide. For
simplicity, we often just write lim supk , lim infk and limk , with the understanding
that this refers to k → ∞.
In working with sequences of sets, a similar pattern is encountered in which
“outer” and “inner” limits always exist and give a “limit” when they agree.

Outer and inner limits. Consider a sequence {Ck }∞ n

k=1 of subsets of IR .
k
(a) The outer limit of this sequence, denoted by lim supk C , is the set of all
x ∈ IRn for which
N
there exist N ∈ N  and xk ∈ Ck for k ∈ N such that xk → x.

(b) The inner limit of this sequence, denoted by lim infk Ck , is the set of all x ∈
IRn for which
N
there exist N ∈ N and xk ∈ Ck for k ∈ N such that xk → x.

(c) When the inner and outer limits are the same set C, this set is defined to be
the limit of the sequence {Ck }∞
k=1 :

C = lim Ck = lim sup Ck = lim inf Ck .

k k k
3 Regularity Properties of Set-valued Solution Mappings 135

In this case Ck is said to converge to C in the sense of Painlevé–Kuratowski conver-

gence.
Note that although the outer and inner limit sets always exist by this definition,
they might be empty. When Ck = 0/ for all k, these sets can be described equivalently
in terms of the sequences {xk }∞k=1 that can be formed by selecting an xk ∈ Ck for each
k: the set of all cluster points of such sequences is lim supk Ck , while the set of all
limits of such sequences is lim infk Ck . Obviously lim infk Ck ⊂ lim supk Ck . When
each Ck is a singleton, lim infk Ck can at most be another singleton, but lim supk Ck
might have multiple elements.
Examples.
1) The sequence of doubletons Ck = {0, 1k } in IR has limk Ck = {0}. Indeed, every
sequence of elements xk ∈ Ck converges to 0.
2) The sequence of balls IB(xk , ρ k ) converges to IB(x, ρ ) when xk → x and ρ k → ρ .
3) A sequence of sets Ck which alternates between two different closed sets D1
and D2 , that is, Ck = D1 when k is odd and Ck = D2 when k is even, has D1 ∩ D2
as its inner limit and D1 ∪ D2 as its outer limit. Such a sequence is not convergent if
D1 = D2 .
Outer and inner limits can also be described with the help of neighborhoods:
  

(1a) lim sup Ck = x  ∀ neighborhood V of x, ∃ N ∈ N  , ∀ k ∈ N : Ck ∩V = 0/ ,
k→∞

  

(1b) lim inf Ck = x  ∀ neighborhood V of x, ∃ N ∈ N , ∀ k ∈ N : Ck ∩V = 0/ .
k→∞

Without loss of generality the neighborhoods in (1a,b) can be taken to be closed

balls; then we obtain the following more transparent definitions:
 
(2a) lim sup Ck = x  ∀ ε > 0, ∃ N ∈ N  : x ∈ Ck + ε IB (k ∈ N)},
k→∞

 
(2b) lim inf Ck = x  ∀ ε > 0, ∃ N ∈ N : x ∈ Ck + ε IB (k ∈ N)}.
k→∞

Both the outer and inner limits of a sequence {Ck }k∈IN are closed sets. Indeed, if
x∈ / lim supk Ck , then, from (2a), there exists ε > 0 such that for every N ∈ N  we
have x ∈/ Ck + ε IB, that is, IB(x, ε ) ∩Ck = 0,
/ for some k ∈ N. But then a neighborhood
of x can meet Ck for finitely many k only. Hence no points in this neighborhood can
be cluster points of sequences {xk } with xk ∈ Ck for infinitely many k. This implies
that the complement of lim supk Ck is an open set and therefore that lim supk Ck is
closed. An analogous argument works for lim infk Ck (this could also be derived by
the following Proposition 3A.1).
Recall from Section 1D that the distance from a point x ∈ IRn to a subset C of IRn
is
136 3 Regularity Properties of Set-valued Solution Mappings

dC (x) = d(x,C) = inf |x − y|.

y∈C

As long as C is closed, having d(x,C) = 0 is equivalent to having x ∈ C.

Proposition 3A.1 (distance function characterizations of limits). Outer and inner

limits of sequences of sets are described alternatively by the following formulas:
  

(3a) lim sup Ck = x  lim inf d(x,Ck ) = 0 ,
k→∞ k→∞

  

(3b) lim inf Ck = x  lim d(x,Ck ) = 0 .
k→∞ k→∞

Proof. If x ∈ lim supk Ck then, by (2a), for any ε > 0 there exists N ∈ N  such
that d(x,Ck ) ≤ ε for all k ∈ N. But then, by the definition of the lower limit for
a sequence of real numbers, as recalled in the beginning of this section, we have
lim infk→∞ d(x,Ck ) = 0. The left side of (3a) is therefore contained in the right side.
Conversely, if x is in the set on the right side of (3a), then there exists N ∈ N  and
N
xk ∈ Ck for all k ∈ N such that xk → x; then, by definition, x must belong to the left
side of (3a).
If x is not in the set on the right side of (3b), then there exist ε > 0 and N ∈ N 
such that d(x,Ck ) > ε for all k ∈ N. Then x ∈ / Ck + ε IB for all k ∈ N and hence by
(2b) x is not in lim infk C . In a similar way, from (2b) we obtain that x ∈
k / lim infk Ck
only if lim supk d(x,C ) > 0. This gives us (3b).
k

Observe that the distance to a set does not distinguish whether this set is closed
or not. Therefore, in the context of convergence, there is no difference whether the
sets in a sequence are closed or not. (But limits of all types are closed sets.)

More examples.
1) The limit of the sequence of intervals [k, ∞) as k → ∞ is the empty set, whereas
the limit of the sequence of intervals [1/k, ∞) is [0, ∞).
2) More generally for monotone sequences of subsets Ck ⊂ IRn , if Ck ⊃ Ck+1 for
all k ∈ IN, then limk Ck = ∩k cl Ck , whereas if Ck ⊂ Ck+1 for all k, then limk Ck =
cl ∪kCk .
3) The constant sequence Ck = D, where D is the set of vectors in IRn whose
coordinates are rational numbers, converges not to D, which isn’t closed, but to the
closure of D, which is IRn . More generally, if Ck = C for all k, then limk Ck = cl C.

Theorem 3A.2 (characterization of Painlevé–Kuratowski convergence). For a se-

quence Ck of sets in IRn and a closed set C ⊂ IRn one has:
(a) C ⊂ lim infk Ck if and only if for every open set O ⊂ IRn with C ∩ O = 0/ there
exists N ∈ N such that Ck ∩ O = 0/ for all k ∈ N ;
(b) C ⊃ lim supk Ck if and only if for every compact set B ⊂ IRn with C ∩ B = 0/
there exists N ∈ N such that Ck ∩ B = 0/ for all k ∈ N ;
3 Regularity Properties of Set-valued Solution Mappings 137

(c) C ⊂ lim infk Ck if and only if for every ρ > 0 and ε > 0 there is an index set
N ∈ N such that C ∩ ρ IB ⊂ Ck + ε IB for all k ∈ N ;
(d) C ⊃ lim supk Ck if and only if for every ρ > 0 and ε > 0 there is an index set
N ∈ N such that Ck ∩ ρ IB ⊂ C + ε IB for all k ∈ N ;
(e) C ⊂ lim infk Ck if and only if lim supk d(x,Ck ) ≤ d(x,C) for every x ∈ IRn ;
(f) C ⊃ lim supk Ck if and only if d(x,C) ≤ lim infk d(x,Ck ) for every x ∈ IRn .
Thus, from (c)(d) C = limk Ck if and only if for every ρ > 0 and ε > 0 there is an
index set N ∈ N such that

Ck ∩ ρ IB ⊂ C + ε IB and C ∩ ρ IB ⊂ Ck + ε IB for all k ∈ N.

Also, from (e)(f), C = limk Ck if and only if limk d(x,Ck ) = d(x,C) for every x ∈ IRn .
Proof. (a): Necessity comes directly from (1b). To show sufficiency, assume that
there exists x ∈ C \ lim infk Ck . But then, by (1b), there exists an open neighborhood
V of x such that for every N ∈ N there exists k ∈ N with V ∩ Ck = 0/ and also
V ∩C = 0./ This is the negation of the condition on the right.
(b): Let C ⊃ lim supk Ck and let there exist a compact set B with C ∩ B = 0, / such
that for any N ∈ N one has Ck ∩ B = 0/ for some k ∈ N. But then there exist N ∈ N 
and a convergent sequence xk ∈ Ck for k ∈ N whose limit is not in C, a contradiction.
Conversely, if there exists x ∈ lim supk Ck which is not in C then, from (2a), a ball
IBε (x) with sufficiently small radius ε does not meet C yet meets Ck for infinitely
many k; this contradicts the condition on the right.
Sufficiency in (c): Consider any point x ∈ C, and any ρ > |x|. For an arbitrary
ε > 0, there exists, by assumption, an index set N ∈ N such that C ∩ ρ IB ⊂ Ck + ε IB
for all k ∈ N. Then x ∈ Ck + ε IB for all k ∈ N. By (2b), this yields x ∈ lim infk Ck .
Hence, C ⊂ lim infk Ck .
Necessity of (c): It will be demonstrated that if the condition fails, there must be
a point x̄ ∈ C lying outside of lim infk Ck . To say that the condition fails is to say
that there exist ρ > 0 and ε > 0, such that, for each N ∈ N , the inclusion C ∩ ρ IB ⊂
Ck + ε IB is false for at least one k ∈ N. Then there is an index set N0 ∈ N  such that
this inclusion is false for all k ∈ N0 ; there are points xk ∈ [C ∩ ρ IB] \ [Ck + ε IB] for all
k ∈ N0 . Such points form a bounded sequence in the closed set C with the property
that d(xk ,Ck ) ≥ ε . A subsequence {xk }k∈N1 , for an index set N1 ∈ N  within N0 ,
converges in that case to a point x̄ ∈ C. Since d(xk ,Ck ) ≤ d(x̄,Ck ) + |x̄− xk |, we must
have
d(x̄,Ck ) ≥ ε /2 for all k ∈ N1 large enough.
It is impossible then for x̄ to belong to lim infk Ck , because that requires d(x̄,Ck ) to
converge to 0, cf. (3b).
Sufficiency in (d): Let x̄ ∈ lim supk Ck ; then for some N0 ∈ N  there are points
N
xk ∈ Ck such that xk →0 x̄. Fix any ρ > |x̄|, so that xk ∈ ρ IB for k ∈ N0 large enough.
By assumption, there exists for any ε > 0 an index set N ∈ N such that Ck ∩ ρ IB ⊂
C + ε IB when k ∈ N. Then for large enough k ∈ N0 ∩ N we have xk ∈ C + ε IB, hence
N
d(xk ,C) ≤ ε . Because d(x̄,C) ≤ d(xk ,C) + |xk − x̄| and xk →0 x̄, it follows from the
arbitrary choice of ε that d(x̄,C) = 0, which means x̄ ∈ C (since C is closed).
138 3 Regularity Properties of Set-valued Solution Mappings

Necessity in (d): Suppose to the contrary that one can find ρ > 0, ε > 0 and
N ∈ N  such that, for all k ∈ N, there exists xk ∈ [Ck ∩ ρ IB] \ [C + ε IB]. The sequence
{xk }k∈N is then bounded, so it has a cluster point x̄ which, by definition, belongs
to lim supk Ck . On the other hand, since each xk lies outside of C + ε IB, we have
d(xk ,C) ≥ ε and, in the limit, d(x̄,C) ≥ ε . Hence x̄ ∈
/ C, and therefore lim supk Ck is
not a subset of C.
(e): Sufficiency follows from (3b) by taking x ∈ C. To prove necessity, choose
x ∈ IRn and let y ∈ C be a projection of x on C: |x − y| = d(x,C). By the definition of
N
lim inf there exist N ∈ N and yk ∈ Ck , k ∈ N such that yk → y. For such yk we have
d(x,C ) ≤ |y − x|, k ∈ N and passing to the limit with k → ∞ we get the condition
k k

on the right.
(f): Sufficiency follows from (3a) by taking x ∈ lim supk Ck . Choose x ∈ IRn . If
x ∈ C there is nothing to prove. If not, note that for any nonnegative α the condition
d(x,C) > α is equivalent to C ∩ IBα (x) = 0. / But then from (b) there exists N ∈ N
with Ck ∩ IBα (x) = 0/ for k ∈ N, which is the same as d(x,Ck ) > α for k ∈ N. This
implies the condition on the right.
Observe that in parts (c)(d) of 3A.2 we can replace the phrase “for every ρ ” by
“there is some ρ0 ≥ 0 such that for every ρ ≥ ρ0 ”.
Set convergence can also be characterized in terms of concepts of distance be-
tween sets.

Excess and Pompeiu–Hausdorff distance. For sets C and D in IRn , the excess of
C beyond D is defined by
e(C, D) = sup d(x, D),
x∈C

where the convention is used that


/ D) =
e(0,
0 when D = 0,
/
∞ otherwise.
The Pompeiu–Hausdorff distance between C and D is the quantity

h(C, D) = max{e(C, D), e(D,C)}.

Equivalently, these quantities can be expressed by

  
e(C, D) = inf τ ≥ 0  C ⊂ D + τ IB

and   
h(C, D) = inf τ ≥ 0  C ⊂ D + τ IB, D ⊂ C + τ IB .
The excess and the Pompeiu–Hausdorff distance are illustrated in Fig. 3.1. They
are unaffected by whether C and D are closed or not, but in the case of closed sets the
infima in the alternative formulas are attained. Note that both e(C, D) and h(C, D)
can sometimes be ∞ when unbounded sets are involved. For that reason in particular,
the Pompeiu–Hausdorff distance does not furnish a metric on the space of nonempty
closed subsets of IRn , although it does on the space of nonempty closed subsets of a
3 Regularity Properties of Set-valued Solution Mappings 139

/ = ∞ for any set C, including the empty

bounded set X ⊂ IRn . Also note that e(C, 0)
set.

h(C,D) = e(D,C)

C D

e(C,D)

Fig. 3.1 Illustration of the excess and Pompeiu–Hausdorff distance.

Proposition 3A.3 (characterization of Pompeiu–Hausdorff distance). For any non-

empty sets C and D in IRn , one has

(4) h(C, D) = sup |d(x,C) − d(x, D)|.

x∈IRn

Proof. Since the distance to a set doesn’t distinguish whether the set is closed or
not, we may assume that C and D are nonempty closed sets.
According to 1D.4, for any x ∈ IRn we can pick u ∈ C such that d(x, u) = d(x,C).
For any v ∈ D, the triangle inequality tells us that d(x, v) ≤ d(x, u) + d(u, v). Taking
the infimum on both sides with respect to v ∈ D, we see that d(x, D) ≤ d(x, u) +
d(u, D), where d(u, D) ≤ e(C, D). Therefore, d(x, D) − d(x,C) ≤ e(C, D), and by
symmetry in exchanging the roles of C and D, also d(x,C) − d(x, D) ≤ e(D,C), so
that
|d(x,C) − d(x, D)| ≤ max{e(C, D), e(D,C)} = h(C, D).
Hence “≥” holds in (4).
On the other hand, since d(x,C) = 0 when x ∈ C, we have

e(C, D) = sup d(x, D) = sup |d(x, D) − d(x,C)| ≤ sup |d(x, D) − d(x,C)|

x∈C x∈C x∈IRn

and likewise e(D,C) ≤ supx∈IRn |d(x,C) − d(x, D)|, so that

max{e(C, D), e(D,C)} ≤ sup |d(x,C) − d(x, D)|.

x∈IRn
140 3 Regularity Properties of Set-valued Solution Mappings

This confirms that “≤” also holds in (4).

Pompeiu–Hausdorff convergence. A sequence of sets {Ck }∞ k=1 is said to con-

verge with respect to Pompeiu–Hausdorff distance to a set C when C is closed and
h(Ck ,C) → 0 as k → ∞.
From the definition of the Pompeiu-Hausdorff distance it follows that when a
sequence Ck converges to C, then the set C must be nonempty and only finitely
many Ck can be empty. Note that this is not the case when Painlevé–Kuratowski
convergence is considered (see the first example after 3A.1). The following theorem
exhibits the main relationship between these two types of convergence.

Theorem 3A.4 (Pompeiu–Hausdorff versus Painlevé–Kuratowski). If a sequence

of closed sets {Ck }∞k=1 converges to C with respect to Pompeiu–Hausdorff distance
then it also converges to C in the Painlevé–Kuratowski sense. The opposite implica-
tion holds if there is a bounded set X which contains C and every Ck .
Proof. By definition, Ck converges to C with respect to Pompeiu–Hausdorff dis-
tance if and only if, for any ε > 0, there exists N ∈ N with

(5) Ck ⊂ C + ε IB and C ⊂ Ck + ε IB for all k ∈ N.

Since (5) implies (2a,b), the Painlevé–Kuratowski convergence of Ck to C then fol-

lows from the convergence with respect to Pompeiu–Hausdorff distance.
Suppose now that Ck converges to C in the Painlevé–Kuratowski sense, with
C and every Ck included in a bounded set X. Then there exists ρ0 > 0 such that
Ck = Ck ∩ ρ IB and C = C ∩ ρ IB for every ρ ≥ ρ0 . We obtain that for every ρ > ρ0
and k ∈ IN,

Ck = Ck ∩ ρ IB ⊂ C + ε IB and C = C ∩ ρ IB ⊂ Ck + ε IB.

But then, for every ρ > 0 we have

Ck ∩ ρ IB ⊂ Ck ⊂ C + ε IB and C ∩ ρ IB ⊂ C ⊂ Ck + ε IB for k ∈ IN

and hence (5) holds and we have convergence of Ck to C with respect to Pompeiu–
Hausdorff distance.

Exercise 3A.5 (convergence equivalence under boundedness). For a sequence of

sets Ck in IRn and a nonempty closed set C, the following are equivalent:
(a) Ck converges to C in the Pompeiu–Hausdorff sense and C is bounded;
(b) Ck converges to C in the Painlevé–Kuratowski sense and there is a bounded
set X along with an index set N ∈ N such that Ck ⊂ X for all k ∈ N .

Theorem 3A.6 (conditions for Pompeiu–Hausdorff convergence). A sequence Ck

of sets in IRn is convergent with respect to Pompeiu–Hausdorff distance to a closed
set C ⊂ IRn if both of the following conditions hold:
3 Regularity Properties of Set-valued Solution Mappings 141

(a) for every open set O ⊂ IRn with C ∩ O = 0/ there exists N ∈ N such that
Ck ∩ O = 0/ for all k ∈ N ;
(b) for every open set O ⊂ IRn with C ⊂ O there exists N ∈ N such that Ck ⊂ O
for all k ∈ N .
Moreover, condition (a) is always necessary for Pompeiu–Hausdorff conver-
gence, while (b) is necessary when the set C is bounded.
Proof. Let (a)(b) hold and let the first inclusion in (5) be violated, that is, there exist
x ∈ C, a scalar ε > 0 and a sequence N ∈ N  such that x ∈ / Ck + ε IB for k ∈ N. Then
k
an open neighborhood of x does not meet C for infinitely many k; this contradicts
condition (a). Furthermore, (b) implies that for any ε > 0 there exists N ∈ N such
that Ck ⊂ C + ε IB for all k ∈ N, which is the second inclusion in (5). Then Pompeiu–
Hausdorff convergence follows from (5).
According to 3A.2(a), condition (a) is equivalent to C ⊂ lim infk Ck , and hence
it is necessary for Painlevé–Kuratowski convergence, and then also for Pompeiu–
Hausdorff convergence. To show necessity of (b), let C ⊂ O for some open set O ⊂
IRn . For k ∈ IN let there exist points xk ∈ C and yk in the complement of O such that
|xk − yk | → 0 as k → ∞. Since C is compact, there exists N ∈ N  and x ∈ C such
N N
that xk → x, hence yk → x as well. But then x must be also in the complement of
O, which is impossible. The contradiction so obtained shows there is an ε > 0 such
that C + ε IB ⊂ O; then, from (5), for some N ⊂ N we have Ck ⊂ O for k ∈ N.
Examples 3A.7 (unboundedness issues). As an illustration of the troubles that may
occur when we deal with unbounded sets, consider first the sequence of bounded
sets Ck ⊂ IR2 in which Ck is the segment having one end at the origin and the other
at the point (cos 1k , sin 1k ); that is,
  
 1 1
Ck = x ∈ IR2  x1 = t cos , x2 = t sin , 0 ≤ t ≤ 1 .
k k
Both the Painlevé–Kuratowski and Pompeiu–Hausdorff limits exist and are equal to
the segment having one end at the origin and the other at the point (1, 0). Also, both
conditions (a) and (b) in 3A.6 are satisfied.
Let us now modify this example by taking as Ck , instead of a segment, the whole
unbounded ray with its end at the origin. That is,
  
 1 1
Ck = x ∈ IR2  x1 = t cos , x2 = t sin , t ≥ 0 .
k k
  
The Painlevé–Kuratowski limit is the ray x ∈ IR2  x1 ≥ 0, x2 = 0 , whereas the
Pompeiu–Hausdorff limit fails to exist. In this case condition (a) in 3A.6 holds,
whereas (b) is violated.
As another example demonstrating issues with unboundedness, consider the se-
quence of sets
  1 1

Ck = x ∈ IR2  x1 > 0, x2 ≥ − ,
x1 k
142 3 Regularity Properties of Set-valued Solution Mappings

which isobviously convergent with respect to Pompeiu–Hausdorff distance to the

set C = x ∈ IR2  x1 > 0, x2 ≥ 1/x1 (choose k > 1/ε in (5)). On the other hand,
 
condition (b) in 3A.6 fails, since the open set O = x ∈ IR2  x1 > 0, x2 > 0 contains
C but does not contain Ck for any k.

3B. Continuity of Set-valued Mappings

Continuity properties of a set-valued mapping S : IRm →
→ IRn can be defined on the
basis of Painlevé–Kuratowski set convergence. Alternatively they can be defined on
the basis of Pompeiu–Hausdorff set convergence, which is the same in a context of
boundedness but otherwise is more stringent and only suited to special situations, as
explained at the end of Section 3A. Following the pattern of inner and outer limits
used in introducing Painlevé–Kuratowski convergence, we let
,
lim sup S(y) = lim sup S(yk )
y→ȳ
yk →ȳ k→∞
  

= x  ∃ yk → ȳ, ∃ xk → x with xk ∈ S(yk )

and
-
lim inf S(y) = lim inf S(yk )
y→ȳ
yk →ȳ k→∞
  
 N
= x  ∀ yk → ȳ, ∃ N ∈ N , xk → x with xk ∈ S(yk ) .

In other words, the limsup is the set of all possible limits of sequences xk ∈ S(yk )
when yk → ȳ, while the liminf is the set of points x for which there exists a sequence
xk ∈ S(yk ) when yk → ȳ such that xk → x.

Semicontinuity and continuity. A set-valued mapping S : IRm →

→ IRn is outer semi-
continuous (osc) at ȳ when
lim sup S(y) ⊂ S(ȳ)
y→ȳ

and inner semicontinuous (isc) at ȳ when

lim inf S(y) ⊃ S(ȳ).

y→ȳ

It is called Painlevé–Kuratowski continuous at ȳ when it is both osc and isc at ȳ, as

expressed by
lim S(y) = S(ȳ).
y→ȳ

On the other hand, S is called Pompeiu–Hausdorff continuous at ȳ when

3 Regularity Properties of Set-valued Solution Mappings 143

S(ȳ) is closed and lim h(S(y), S(ȳ)) = 0.

y→ȳ

These terms are invoked relative to a subset D in IRm when the properties hold
for limits taken with y → ȳ in D (but not necessarily for limits y → ȳ without this
restriction). Continuity is taken to refer to Painlevé–Kuratowski continuity, unless
otherwise specified.
For single-valued mappings both definitions of continuity reduce to the usual
definition of continuity of a function. Note that when S is isc at ȳ relative to D then
there must exist a neighborhood V of ȳ such that D ∩ V ⊂ dom S. When D = IRm ,
this means ȳ ∈ int (dom S).

Exercise 3B.1 (limit relations as equations).

(a) Show that S is osc at ȳ if and only if actually lim supy→ȳ S(y) = S(ȳ).
(b) Show that, when S(ȳ) is closed, S is isc at ȳ if and only if lim infy→ȳ S(y) =
S(ȳ).
Although the closedness of S(ȳ) is automatic from this when S is continuous at
ȳ in the Painlevé–Kuratowski sense, it needs to be assumed directly for Pompeiu–
Hausdorff continuity because the distance concept utilized for that concept is unable
to distinguish whether sets are closed or not.
Recall that a set M is closed relative to a set D when any sequence yk ∈ M ∩ D
has its cluster points in M. A set M is open relative to D if the complement of M is
closed relative to D. Also, recall that a function f : IRn → IR is lower semicontinuous

on a closed set D ⊂ IRn when the lower level set x ∈ D  f (x) ≤ α is closed for
every α ∈ IR. We defined this property at the beginning of Chapter 1 for the case
of D = IRn and for functions with values in IR, and now are merely echoing that for
D ⊂ IRn .

Theorem 3B.2 (characterization of semicontinuity). For S : IRm → → IRn , a set D ⊂ IRm

and ȳ ∈ dom S we have:
(a) S is osc at ȳ relative to D if and only if for every x ∈/ S(ȳ) there are neighbor-
hoods U of x and V of ȳ such that D ∩V ∩ S−1 (U) = 0/ ;
(b) S is isc at ȳ relative to D if and only if for every x ∈ S(ȳ) and every neighbor-
hood U of x there exists a neighborhood V of ȳ such that D ∩V ⊂ S−1 (U);
(c) S is osc at every y ∈ dom S if and only if gph S is closed;
(d) S is osc relative to a set D ⊂ IRm if and only if S−1 (B) is closed relative to D
for every compact set B ⊂ IRn ;
(e) S is isc relative to a set D ⊂ IRm if and only if S−1 (O) is open relative to D
for every open set O ⊂ IRn ;
(f) S is osc at ȳ relative to a set D ⊂ IRm if and only if the distance function
y → d(x, S(y)) is lower semicontinuous at ȳ relative to D for every x ∈ IRn ;
(g) S is isc at ȳ relative to a set D ⊂ IRm if and only if the distance function
y → d(x, S(y)) is upper semicontinuous at ȳ relative to D for every x ∈ IRn .
Thus, S is continuous relative to D at ȳ if and only if the distance function y →
d(x, S(y)) is continuous at ȳ relative to D for every x ∈ IRn .
144 3 Regularity Properties of Set-valued Solution Mappings

Proof. Necessity in (a): Suppose that there exists x ∈ / S(ȳ) such that for any neigh-
borhood U of x and neighborhood V of ȳ we have S(y) ∩U = 0/ for some y ∈ V ∩ D.
But then there exists a sequence yk → ȳ, yk ∈ D and xk ∈ S(yk ) such that xk → x.
This implies that x ∈ lim supk S(yk ), hence x ∈ S(ȳ) since S is osc, a contradiction.
Sufficiency in (a): Let x ∈ / S(ȳ). Then there exists ρ > 0 such that S(ȳ) ∩ IBρ (x) =
0;
/ the condition in the second half of (a) then gives a neighborhood V of ȳ such that
N
for every N ∈ N , every sequence yk → ȳ with yk ∈ D ∩ V has S(yk ) ∩ IBρ (x) = 0. /
But in that case d(x, S(y )) > ρ /2 for all large k, which implies, by Proposition 3A.1
k

and the definition of limsup, that x ∈ / lim supy→ȳ S(y). This means that S is osc at ȳ.
Necessity in (b): Suppose that there exists x ∈ S(ȳ) such that for some neighbor-
hood U of x and any neighborhood V of ȳ we have S(y) ∩U = 0/ for some y ∈ V ∩ D.
Then there is a sequence yk convergent to ȳ in D such that for every sequence xk → x
one has xk ∈/ S(yk ). This means that x ∈ / lim infy→ȳ S(y). But then S is not isc at ȳ.
Sufficiency in (b): If S is not isc at ȳ relative to D, then, according to 3A.2(a),
there exist an infinite sequence yk → ȳ in D, a point x ∈ S(ȳ) and an open neighbor-
hood U of x such that S(yk ) ∩ U = 0/ for infinitely many k. But then there exists a
neighborhood V of ȳ such that D ∩V is not in S−1 (U) which is the opposite of (b).
(c): S has closed graph if and only if for any (y, x) ∈ / gph S there exist open neigh-
borhoods V of y and U of x such that V ∩ S−1 (U) = 0. / From (a), this comes down
to S being osc at every y ∈ dom S.
(d): Every sequence in a compact set B has a convergent subsequence, and on
the other hand, a set consisting of a convergent sequence and its limit is a compact
set. Therefore the condition in the second part of (d) is equivalent to the condition
that if xk → x̄, yk ∈ S−1 (xk ) and yk → ȳ with yk ∈ D, one has ȳ ∈ S−1 (x̄). But this is
precisely the condition for S to be osc relative to D.
(e): Failure of the condition in (e) means the existence of an open set O and a
sequence yk → ȳ in D such that ȳ ∈ S−1(O) but yk ∈ / S−1 (O); that is, S(ȳ) ∩ O = 0/
yet S(y ) ∩ O = 0/ for all k. This last property says that lim infk S(yk ) ⊃ S(ȳ), by
k

3A.2(a). Hence the condition in (e) fails precisely when S is not isc.
The equivalences in (f) and (g) follow from 3A.2(e) and 3A.2(f).

Theorem 3B.3 (characterization of Pompeiu–Hausdorff continuity). A set-valued

mapping S : IRm → → IRn is Pompeiu–Hausdorff continuous at ȳ if S(ȳ) is closed and
both of the following conditions hold:
(a) for every open set O ⊂ IRn with S(ȳ) ∩ O = 0/ there exists a neighborhood V
of ȳ such that S(y) ∩ O = 0/ for all y ∈ V ;
(b) for every open set O ⊂ IRn with S(ȳ) ⊂ O there exists a neighborhood V of ȳ
such that S(y) ⊂ O for all y ∈ V .
Moreover, if S is Pompeiu–Hausdorff continuous at ȳ, then it is continuous at ȳ.
On the other hand, when S(ȳ) is nonempty and bounded, Pompeiu–Hausdorff con-
tinuity of S at ȳ reduces to continuity together with the existence of a neighborhood
V of ȳ such that S(V ) is bounded; in this case conditions (a) and (b) are not only
sufficient but also necessary for continuity of S at ȳ.
3 Regularity Properties of Set-valued Solution Mappings 145

Observe that we can define inner semicontinuity of a mapping S at ȳ in the

Pompeiu–Hausdorff sense by limy→ȳ e(S(ȳ), S(y)) = 0, but this is simply equiva-
lent to inner semicontinuity in the Painlevé–Kuratowski sense (compare 3B.2(b) and
3B.3(b)). In contrast, if we define outer semicontinuity in the Pompeiu–Hausdorff
sense by limy→ȳ e(S(y), S(ȳ)) = 0, we get a generally much more restrictive concept
than outer semicontinuity in the Painlevé–Kuratowski sense.
We present next two applications of these concepts to mappings that play central
roles in optimization.
Example 3B.4 (solution mapping for a system of inequalities). Consider a mapping
defined implicitly by a parameterized system of inequalities, that is,
  
S : p → x  fi (p, x) ≤ 0, i = 1, . . . , m for p ∈ IRd .

Assume that each fi is a continuous real-valued function on IRd × IRn . Then S is osc
at any point of its domain. If moreover each fi is convex in x for each p and p̄ is
such that there exists x̄ with fi ( p̄, x̄) < 0 for each i = 1, . . . , m, then S is continuous
at p̄.
  
Detail. The graph of S is the intersection of the sets (p, x)  f i (p, x) ≤ 0 , which
are closed by the continuity of fi . Then gph S is closed, and the osc property comes
from Theorem 3B.2(c). The isc part will follow from a much more general result
(Robinson-Ursescu theorem) which we present in Chapter 5.

Applications in optimization. Consider the following general problem of mini-

mization, involving a parameter p which ranges over a set P ⊂ IRd , a function
f0 : IRd × IRn → IR, and a mapping Sfeas : P →
→ IRn :

minimize f0 (p, x) over all x ∈ IRn satisfying x ∈ Sfeas (p).

Here f 0 is the objective function and Sfeas is the feasible set mapping (with Sfeas (p)
taken to be the empty set when p ∈ / P). In particular, Sfeas could be specified by
constraints in the manner of Example 3B.4, but we now allow it to be more general.
Our attention is focused now on two other mappings in this situation: the optimal
value mapping acting from IRd to IR and defined by
  
Sval : p → inf f0 (p, x)  x ∈ Sfeas (p) when the inf is finite,
x

and the optimal set mapping acting from P to IRn and defined by
  
Sopt : p → x ∈ Sfeas (p)  f0 (p, x) = Sval (p) .

Theorem 3B.5 (basic continuity properties of solution mappings in optimiza-

tion). In the preceding notation, let p̄ ∈ P be fixed with the feasible set Sfeas ( p̄)
nonempty and bounded, and suppose that:
146 3 Regularity Properties of Set-valued Solution Mappings

(a) the mapping Sfeas is Pompeiu–Hausdorff continuous at p̄ relative to P, or

equivalently, Sfeas is continuous at p̄ relative to P with Sfeas (Q ∩ P) bounded for
some neighborhood Q of p̄,
(b) the function f0 is continuous relative to P× IRn at ( p̄, x̄) for every x̄ ∈ Sfeas ( p̄).
Then the optimal value mapping Sval is continuous at p̄ relative to P, whereas the
optimal set mapping Sopt is osc at p̄ relative to P.
Proof. The equivalence in assumption (a) comes from the final statement in
Theorem 3B.3. In particular (a) implies Sfeas ( p̄) is closed, hence from bounded-
ness actually compact. Then too, since f0 ( p̄, ·) is continuous on Sfeas ( p̄) by (b), the
set Sopt ( p̄) is nonempty.
Let x̄ ∈ Sopt ( p̄). From (a) we get for any sequence pk → p̄ in P the existence of
a sequence of points xk with xk ∈ Sfeas (pk ) such that xk → x̄ as k → ∞. But then, for
any ε > 0 there exists N ∈ N such that

Sval (pk ) ≤ f0 (pk , xk ) ≤ f0 ( p̄, x̄) + ε = Sval ( p̄) + ε for k ∈ N.

This gives us

(1) lim sup Sval (p) ≤ Sval ( p̄).

p→ p̄

On the other hand, let us assume that

(2) lim inf Sval (p) < Sval ( p̄).

p→ p̄

Then there exist ε > 0 and sequences pk → p̄ in P and xk ∈ Sfeas (pk ), k ∈ IN, such
that

(3) f0 (pk , xk ) < Sval ( p̄) − ε for all k.

From (a) we see that d(xk , Sfeas ( p̄)) → 0 as k → ∞. This provides the existence of a
sequence of points x̄k ∈ Sfeas ( p̄) such that |xk − x̄k | → 0 as k → ∞. Because Sfeas ( p̄)
is compact, there must be some x̄ ∈ Sfeas ( p̄) along with an index set N ∈ N  such
N N
that x̄k → x̄, in which case xk → x̄ as well. Then, from the continuity of f0 at ( p̄, x̄),
we have f0 ( p̄, x̄) ≤ f0 (pk , xk ) + ε for k ∈ N and sufficiently large, which, together
with (3), implies for such k that

Sval ( p̄) ≤ f0 ( p̄, x̄) ≤ f0 (pk , xk ) + ε < Sval ( p̄).

The contradiction obtained proves that (2) is false. Thus,

Sval ( p̄) ≤ lim inf Sval (p),

p→ p̄

which, combined with (1), gives us the continuity of the optimal value mapping Sval
at p̄ relative to P.
3 Regularity Properties of Set-valued Solution Mappings 147

To show that Sopt is osc at p̄ relative to P, we use the equivalent condition in

3B.2(a). Suppose there exists x ∈ / Sopt ( p̄) such that for any neighborhoods U of x
and Q of p̄ there exist u ∈ U and p ∈ Q ∩ P such that u ∈ Sopt (p). This is the same
as saying that there are sequences pk → p̄ in P and uk → x as k → ∞ such that
uk ∈ Sopt (pk ). Note that since uk ∈ Sfeas (pk ) we have that x ∈ Sfeas ( p̄). But then, as
we already proved,

f0 (pk , uk ) = Sval (pk ) → Sval ( p̄) as k → ∞.

By continuity of f0 , the left side tends to f0 ( p̄, x) as k → ∞, which means that x ∈

Sopt ( p̄), a contradiction.

Example 3B.6 (minimization over a fixed set). Let X be a nonempty, compact sub-
set of IRn and let f0 be a continuous function from P× X to IR, where P is a nonempty
subset of IRd . For each p ∈ P, let

Sval (p) = min f0 (p, x), Sopt (p) = argmin f0 (p, x).
x∈X x∈X

Then the function Sval : P → IR is continuous relative to P, and the mapping Sopt :
P→
→ IRn is osc relative to P.
Detail. This exploits the case of Theorem 3B.5 where Sfeas is the constant mapping
p → X.

x2

x1

Fig. 3.2 The feasible set in Example 3B.7 for p = 0.1.

Example 3B.7 (Painlevé–Kuratowski continuity of the feasible set does not imply
continuity of the optimal value). Consider the minimization of f0 (x1 , x2 ) = ex1 + x22
on the set
  
 x1 x1
Sfeas (p) = (x1 , x2 ) ∈ IR2  − − p ≤ x 2 ≤ + p .
(1 + x21) (1 + x21)

For parameter value p = 0, the optimal value Sval (0) = 1 and occurs at Sopt (0) =
(0, 0), but for p > 0 the asymptotics of the function x1 /(1 + x21 ) open up a “phantom”
148 3 Regularity Properties of Set-valued Solution Mappings

portion of the feasible set along the negative x1 -axis, and the optimal value is 0, see
Fig. 3.2. The feasible set nonetheless does depend continuously on p in [0, ∞) in the
Painlevé–Kuratowski sense.

3C. Lipschitz Continuity of Set-valued Mappings

A quantitative notation of continuity for set-valued mappings can be formulated
with the help of the Pompeiu–Hausdorff distance between sets in the same way
that Lipschitz continuity is defined for functions. It has important uses, although it
suffers from shortcomings when the sets may be unbounded. Here we invoke the
terminology that a set-valued mapping S is closed-valued on a set D when S(y) is a
closed set for each y ∈ D.

Lipschitz continuity of set-valued mappings. A mapping S : IRm → → IRn is said

to be Lipschitz continuous relative to a (nonempty) set D in IR if D ⊂ dom S, S is
m

closed-valued on D, and there exists κ ≥ 0 (Lipschitz constant) such that

(1) h(S(y ), S(y)) ≤ κ |y − y| for all y , y ∈ D,

or equivalently, there exists κ ≥ 0 such that

(2) S(y ) ⊂ S(y) + κ |y − y|IB for all y , y ∈ D.

When S is single-valued on D, we obtain from this definition the previous notion

in Section 1D of Lipschitz continuity of a function.
One could contemplate defining Lipschitz continuity of a set-valued mapping S
without requiring S to be closed-valued, relying in that case simply on (1) or (2).
That might be workable, although κ in (2) could then be slightly larger than the κ
in (1), but a fundamental objection arises. A mapping that is continuous necessarily
does have closed values, so we would be in the position of having a concept of
Lipschitz continuity which did not entail continuity. That is a paradox we prefer
to avoid. The issue is absent for single-valued mappings, since they are trivially
closed-valued.
Lipschitz continuity of a set-valued mapping can be characterized by a property
which relates the distances to its value and the value of the inverse mapping:

Proposition 3C.1 (distance characterization of Lipschitz continuity). Consider a

closed-valued mapping S : IRm → → IRn and a nonempty subset D of dom S. Then S is
Lipschitz continuous relative to D with constant κ if and only if

(3) d(x, S(y)) ≤ κ d(y, S−1 (x) ∩ D) for all x ∈ IRn and y ∈ D.
3 Regularity Properties of Set-valued Solution Mappings 149

Proof. Let S be Lipschitz continuous relative to D with a constant κ and let

x ∈ IRn and y ∈ D. If S−1(x) ∩ D = 0, / the inequality (3) holds automatically.
Let S−1 (x) ∩ D = 0/ and choose ε > 0. Then there exists y ∈ S−1(x) ∩ D with
|y − y| ≤ d(y, S−1 (x) ∩ D) + ε . By (1),

d(x, S(y)) ≤ h(S(y ), S(y)) ≤ κ |y − y| ≤ κ d(y, S−1 (x) ∩ D) + κε .

Since the left side of this inequality does not depend on ε , passing to zero with ε we
conclude that (3) holds with the κ of (1).
Conversely, let (3) hold, let y, y ∈ D ⊂ dom S and let x ∈ S(y). Then

d(x, S(y )) ≤ κ d(y , S−1 (x) ∩ D) ≤ κ |y − y|,

since y ∈ S−1 (x) ∩ D. Taking the supremum with respect to x ∈ S(y), we obtain
e(S(y), S(y )) ≤ κ |y − y| and, by symmetry, we get (1).
For the inverse mapping F = S−1 the property described in (3) can be written as

d(x, F −1 (y)) ≤ κ d(y, F(x) ∩ D) for all x ∈ IRn , y ∈ D,

and when gph F is closed this can be interpreted in the following manner. Whenever
we pick a y ∈ D and an x ∈ dom F, the distance from x to the set of solutions u of
the inclusion y ∈ F(u) is proportional to d(y, F(x) ∩ D), which measures the extent
to which x itself fails to solve this inclusion. In Section 3E we will introduce a local
version of this property which plays a major role in variational analysis and is known
as “metric regularity.”
The difficulty with the concept of Lipschitz continuity for set-valued map-
pings S with values S(y) that may be unbounded comes from the fact that usually
h(C1 ,C2 ) = ∞ when C1 or C2 is unbounded, the only exceptions being cases where
both C1 and C2 are unbounded and “the unboundedness points in the same direc-
tion.” For instance, when C1 and C2 are lines in IR2 , one has h(C1 ,C2 ) < ∞ only
when these lines are parallel.
In the remainder of this section we consider a particular class of set-valued map-
pings, with significant applications in variational analysis, which are automatically
Lipschitz continuous even when their values are unbounded sets.

Polyhedral convex mappings. A mapping S : IRm → → IRn is said to be polyhedral

convex if its graph is a polyhedral convex set.
Here it should be recalled from Section 2E that a set is polyhedral convex if it
can be expressed as the intersection of a finite collection of closed half-spaces and/or
hyperplanes.
Example 3C.2 (polyhedral convex mappings from linear constraint systems). A
solution mapping S of the form in Example 3B.4 is polyhedral convex when the fi
there are all affine; furthermore, this continues to be true when some or all of the
constraints are equations instead of inequalities.
150 3 Regularity Properties of Set-valued Solution Mappings

In the notational context of elements x ∈ S(y) for a mapping S : IRm →→ IRn , poly-
hedral convexity of S is equivalent to the existence of a positive integer r, matrices
D ∈ IRr×n , E ∈ IRr×m , and a vector q ∈ IRr such that
  
(4) S(y) = x ∈ IRn  Dx + Ey ≤ q for all y ∈ IRm .

Note for instance that any mapping S whose graph is a linear subspace is a poly-
hedral convex mapping.

Theorem 3C.3 (Lipschitz continuity of polyhedral convex mappings). Any poly-

hedral convex mapping S : IRm →
→ IRn is Lipschitz continuous relative to its domain.
We will prove this theorem by using a fundamental result due to A. J. Hoffman
regarding approximate solutions of systems of linear inequalities. For a vector a =
(a1 , a2 , . . . , an ) ∈ IRn , we use the vector notation that

a+ = (max{0, a1 }, . . . , max{0, an}).

Also, recall that the convex hull of a set C ⊂ IRn , which will be denoted by co C,
is the smallest convex set that includes C. (It can be identified as the intersection of
all convex sets that include C, but also can be described as consisting of all linear
combinations λ0 x0 + λ1 x1 + · · · + λn xn with xi ∈ C, λi ≥ 0, and λ0 + λ1 + · · · + λn =
1; this is Carathéodory’s theorem.) The closed convex hull of C is the closure of
the convex hull of C and denoted cl co C; it is the smallest closed convex set that
contains C.

Lemma 3C.4 (Hoffman lemma). For the set-valued mapping

  
S : y → x ∈ IRn  Ax ≤ y for y ∈ IRm ,

where A is a nonzero m × n matrix, there exists a constant L such that

(5) d(x, S(y)) ≤ L|(Ax − y)+ | for every y ∈ dom S and every x ∈ IRn .

Proof. For any y ∈ dom S the set S(y) is nonempty, convex and closed, hence any
point x ∈
/ S(y) has a unique (Euclidean) projection u = PS(y) (x) on S(y) (Proposi-
tion 1D.5):

(6) u ∈ S(y), |u − x| = d(x, S(y)).

As noted in Section 2A, the projection mapping satisfies

PS(y) = (I + NS(y) )−1 ,

where NS(y) is the normal cone mapping to the convex set S(y). In these terms, the
problem of projecting x on S(y) is equivalent to that of finding the unique u = x such
that
x ∈ u + NS(y)(u).
3 Regularity Properties of Set-valued Solution Mappings 151

The formula in 2E(8) gives us a representation of the normal cone to a polyhedral

convex set specified by affine inequalities, which here comes out as
  

NS(y) (u) = v  v = ∑i=1 λi ai with λi ≥ 0, λi (ai , u − yi ) = 0, i = 1, . . . , m ,
m

where the ai ’s are the rows of the matrix A regarded as vectors in IRn . Thus, the
projection u of x on S(y), as described by (6), can be obtained by finding a pair
(u, λ ) such that

(7)
x − u − ∑m i=1 λi ai = 0,
λi ≥ 0, λi (ai , u − yi) = 0, i = 1, . . . , m.

While the projection u exists and is unique, this variational inequality might not
have a unique solution (u, λ ) because the λ component might not be unique. But
since u = x (through our assumption that x ∈ / S(y)), we can conclude from the first
relation in (7) that for any solution (u, λ ) the vector λ = (λ1 , . . . , λm ) is not the zero
vector. Consider the family J of subsets J of {1, . . . , m} for which there are real
numbers λ1 , . . . , λm with λi > 0 for i ∈ J and λi = 0 for i ∈/ J and such that (u, λ )
satisfies (7). Of course, if ai , u − yi < 0 for some i, then λi = 0 according to the
second relation (complementarity) in (7), and then this i cannot be an element of
any J. That is,

(8) J ∈ J and i ∈ J =⇒ ai , u = yi and λi > 0.

Since the set of vectors λ such that (u, λ ) solves (7) does not contain the zero vector,
we have J = 0. /
We will now prove that there is a nonempty index set J¯ ∈ J for which there are
no numbers βi , i ∈ J¯ satisfying

(9) βi ≥ 0, i ∈ J,¯ ∑ βi > 0 and ∑ βiai = 0.

j∈J¯ i∈J¯

On the contrary, suppose that for every J ∈ J this is not the case, that is, (9) holds
with J¯ = J for some βi , i ∈ J. Let J  be a set in J with a minimal number of
elements (J  might be not unique). Note that the number of elements in any J  is
greater than 1. Indeed, if there were just one element i in J  , then we would have
βi ai = 0 and βi > 0, hence ai = 0, and then, since (7) holds for (u, λ ) such that
λi = βi , i = i , λi = 0, i = i , from the first equality in (7) we would get x = u which
contradicts the assumption that x ∈ / S(y). Since J  ∈ J , there are λi > 0, i ∈ J  such
that

(10) x−u = ∑ λi ai.

i∈J

By assumption, there are also real numbers βi ≥ 0, i ∈ J  , such that

152 3 Regularity Properties of Set-valued Solution Mappings

(11) ∑ βi > 0 and ∑ βi ai = 0.

i∈J i∈J 

Multiplying both sides of the equality in (11) by a positive scalar t and adding to
(10), we obtain
x − u = ∑ (λi − t βi )ai .
i∈J 

Let    
λ 
t0 = min i  i ∈ J  with βi > 0 .
i βi 

Then for any k ∈ J  for which this minimum is attained, we have

λi − t0 βi ≥ 0 for every i ∈ J  \ k and x − u = ∑ (λi − t βi)ai .

i∈J \k

Thus, the vector λ ∈ IRm with components λi − t0 βi when i ∈ J  and λi = 0 when
i∈/ J  is such that (u, λ ) satisfies (7). Hence, we found a nonempty index set J  ∈ J
having fewer elements than J  , which contradicts the choice of J  . The contradiction
obtained proves that there is a nonempty index set J¯ ∈ J for which there are no
numbers βi , i ∈ J, satisfying (8). In particular, the zero vector in IRn is not in the
convex hull co {a j , j ∈ J}.¯
Let λ̄i > 0, i ∈ J,¯ be the corresponding vector of multipliers such that, if we set
λ̄i = 0 for i ∈
/ J,¯ we have that (u, λ̄ ) is a solution of (7). Since ∑ j∈J¯ λ̄i ai = 0, because
otherwise (9) would hold for βi = λ̄i , we have

γ := ∑ λ̄i > 0.
i∈J¯

Because (7) holds with (u, λ̄ ), using (7) and (8) we have
 λ̄ 
 i  1
d(0, co {a j , j ∈ J})|x
¯ − u| ≤  ∑ ai |x − u| = |x − u||x − u|
i∈J¯
γ γ
.1 / .1  /
=
γ
(x − u), x − u = ∑
γ i∈
λ̄ i a i , x − u
J¯
λ̄i λ̄i
= ∑ γ (ai , x − ai, u) = ∑ γ (ai , x − yi)
i∈J¯ i∈J¯
≤ max{(ai , x − yi )+ }.
i∈J¯

Hence, for some constant c independent of x and y we have

d(x, S(y)) = |x − u| ≤ c max {(ai , x − yi )+ }.

1≤i≤m

This inequality remains valid (perhaps with a different constant c) after passing from
the max vector norm to the equivalent Euclidean norm. This proves (5).
3 Regularity Properties of Set-valued Solution Mappings 153

Proof of Theorem 3C.3. Let y, y ∈ dom S and let x ∈ S(y). Since S is polyhedral,
from the representation (4) we have Dx + Ey − q ≤ 0 and then

(12) Dx + Ey − q = Dx + Ey − q − Ey + Ey ≤ −Ey + Ey.

Then from Lemma 3C.4 above we obtain the existence of a constant L such that

d(x, S(y )) ≤ L|(Dx + Ey − q)+ |,

and hence, by (12),

d(x, S(y )) ≤ L|(E(y − y))+ | ≤ L|E(y − y )|.

Since x is arbitrarily chosen in S(y), this leads to

e(S(y), S(y )) ≤ κ |y − y|

with κ = L|E|. The same must hold with the roles of y and y reversed, and in
consequence S is Lipschitz continuous on dom S.

Applications to solution mappings in linear programming. Consider the follow-

ing problem of linear programming in which y acts as a parameter:

(13) minimize c, x over all x ∈ IRn satisfying Ax ≤ y.

Here c is a fixed vector in IRn , A is a fixed matrix in IRm×n . Define the solution
mappings associated with (13) as in Section 3B, that is, the feasible set mapping
  
(14) Sfeas : y → x  Ax ≤ y ,

the optimal value mapping

  
(15) Sval : y → inf c, x  Ax ≤ y when the inf is finite,
x

and the optimal set mapping by

  
(16) Sopt : y → x ∈ Sfeas (y)  c, x = Sval (y) .

It is known from the theory of linear programming that Sopt (y) = 0/ when the infi-
mum in (15) is finite (and only then).

Exercise 3C.5 (Lipschitz continuity of mappings in linear programming). Establish

that the mappings in (14), (15) and (16) are Lipschitz continuous relative to their
domains, the domain in the case of (15) and (16) being the set D consisting of all y
for which the infimum in (15) is finite.
Guide. Derive the Lipschitz continuity of Sfeas from Theorem 3C.3, out of the con-
nection with Example 3C.2. Let κ be a Lipschitz constant for Sfeas .
154 3 Regularity Properties of Set-valued Solution Mappings

Next, for the case of Sval , consider any y, y ∈ D and any x ∈ Sopt (y), which exists
because Sopt is nonempty when y ∈ D. In particular we have x ∈ Sfeas (y). From the
Lipschitz continuity of Sfeas , there exists x ∈ Sfeas (y ) such that |x − x | ≤ κ |y − y |.
Use this along with the fact that Sval (y ) ≤ c, x  but Sval (y) = c, x to get a bound
on Sval (y ) − Sval (y) which confirms the Lipschitz continuity claimed for Sval .
For the case of Sopt , consider the set-valued mapping
  
G : (y,t) → x ∈ IRn  Ax ≤ y, c, x ≤ t for (y,t) ∈ IRm × IR.

Confirm that this mapping is polyhedral convex and apply Theorem 3C.3 to it. Ob-
serve that Sopt (y) = G(y, Sval (y)) for y ∈ D and invoke the Lipschitz continuity of
Sval .

3D. Outer Lipschitz Continuity

In this section we define a “one-point” property of set-valued mappings by fixing

one of the points y and y in the definition of Lipschitz continuity at its reference
value ȳ. Then these points no longer play symmetric roles, so we use the excess
instead of the Pompeiu–Hausdorff distance.

Outer Lipschitz continuity. A mapping S : IRm → → IRn is said to be outer Lipschitz

continuous at ȳ relative to a set D if ȳ ∈ D ⊂ dom S, S(ȳ) is a closed set, and there is
a constant κ ≥ 0 along with a neighborhood V of ȳ such that

(1) e(S(y), S(ȳ)) ≤ κ |y − ȳ| for all y ∈ V ∩ D,

or equivalently

(2) S(y) ⊂ S(ȳ) + κ |y − ȳ|IB for all y ∈ V ∩ D.

If S is outer Lipschitz continuous at every point y ∈ D relative to D with the same κ,

then S is said to be outer Lipschitz continuous relative to D.
It is clear that any mapping which is Lipschitz continuous relative to a set D
with constant κ is also outer Lipschitz continuous relative to D with constant κ , but
the converse may not be true. Also, outer Lipschitz continuity at a point ȳ implies
outer semicontinuity at ȳ. For single-valued mappings, outer Lipschitz continuity
becomes the property of calmness which we considered in Section 1C. The ex-
amples in Section 1D show how very different this property is from the Lipschitz
continuity.
The condition in the definition that the mapping is closed-valued at ȳ could be
dropped; but then the constant κ in (2) might be slightly larger than the one in
3 Regularity Properties of Set-valued Solution Mappings 155

(1), and furthermore outer Lipschitz continuity might not entail outer semicontinu-
ity (where closed-valuedness is essential). Therefore, we hold back from such an
extension.
We present next a result which historically was the main motivation for intro-
ducing the property of outer Lipschitz continuity and which complements Theo-
rem 3C.3. It uses the following concept.

Polyhedral mappings. A set-valued mapping S : IRn → → IRm will be called polyhe-

dral if gph S is the union of finitely many sets that are polyhedral convex in IRn ×IRm .
Clearly, a polyhedral mapping has closed graph, since polyhedral convex sets are
closed, and hence is osc and in particular closed-valued everywhere. Any polyhedral
convex mapping as defined in 3C is obviously a polyhedral mapping, but the graph
then is comprised of only one “piece,” whereas now we are allowing a multiplicity
of such polyhedral convex “pieces,” which furthermore could overlap.

Theorem 3D.1 (outer Lipschitz continuity of polyhedral mappings). Any polyhe-

dral mapping S : IRm →
→ IRn is outer Lipschitz continuous relative to its domain.
0
Proof. Let gph S = ki=1 Gi where the Gi ’s are polyhedral convex sets in IRm × IRn .
For each i define the mapping
 
Si : y → x  (y, x) ∈ Gi } for y ∈ IRm .

Then each Si is Lipschitz continuous on its domain, according to Theorem 3C.3. Let
ȳ ∈ dom S and let
  
J = i  there exists x ∈ IRn with (ȳ, x) ∈ Gi .

Then ȳ ∈ dom Si for each i ∈ J , and moreover,

,
(3) S(ȳ) = Si (ȳ).
i∈J

For any i ∈/ J , since the sets {ȳ} × IRn and Gi are1disjoint and polyhedral
1
convex,
there is a neighborhood Vi of ȳ such that (Vi × IRn ) Gi = 0.
/ Let V = i∈J/ V i . Then
of course V is a neighborhood of ȳ and we have

- ,
k , ,
(4) (V × IRn ) gph S ⊂ Gi \ Gi ⊂ Gi .
i=1 i∈J
/ i∈J

Let y ∈ V . If S(y) = 0,
/ then the relation (1) holds trivially. Let x be any point in S(y).
Then from (4), - ,
(y, x) ∈ (V × IRn ) gph S ⊂ Gi ,
j∈J

hence for some i ∈ J we have (y, x) ∈ Gi , that is, x ∈ Si (y). Since each Si is Lip-
schitz continuous and ȳ ∈ dom Si , with constant κi , say, we obtain by using (3) that
156 3 Regularity Properties of Set-valued Solution Mappings

d(x, S(ȳ)) ≤ maxi d(x, Si (ȳ)) ≤ maxi e(Si (y), Si (ȳ)) ≤ maxi κi |y − ȳ|.

Since x is an arbitrary point in S(y), we conclude that S is outer Lipschitz continuous

at ȳ with constant κ := maxi κi .

Exercise 3D.2 (polyhedrality of solution mappings to linear variational inequali-

ties). Given an n × n matrix A and a polyhedral convex set C in IRn , show that the
solution mapping of the linear variational inequality
  
y → S(y) = x  y ∈ Ax + NC (x) for y ∈ IRn

is polyhedral, and therefore it is outer Lipschitz continuous relative to its domain.

Guide. Any polyhedral convex set C is representable (in a non-unique manner) by
a system of affine inequalities:
  
C = x  ai , x ≤ αi for i = 1, 2, . . . , m .

We know from Section 2E that the normal cone to C at the point x ∈ C is the set
  

NC (x) = u  u = ∑i=1 yi ai , yi ≥ 0 for i ∈ I(x), yi = 0 for i ∈
m
/ I(x) ,
  
where I(x) = i  ai , x = αi is the active index set for x ∈ C. The graph of the
normal cone mapping NC is not convex, unless C is a translate of a subspace, but it
is the union, with respect to all possible subsets J of {1, . . . , m}, of the polyhedral
convex sets
  

(x, u)  u = ∑i=1 yi ai , ai , x = αi , yi ≥ 0 if i ∈ J, ai , x < αi , yi = 0 if i ∈
m
/J .

It remains to observe that the graph of the sum A+ NC is also the union of polyhedral
convex sets.
Outer Lipschitz continuity becomes automatically Lipschitz continuity when the
mapping is inner semicontinuous, a property we introduced in the preceding section.

Theorem 3D.3 (isc criterion for Lipschitz continuity). Consider a set-valued map-
ping S : IRm →
→ IRn and a convex set D ⊂ dom S such that S(y) is closed for every
y ∈ D. Then S is Lipschitz continuous relative to D with constant κ if and only if S is
both inner semicontinuous (isc) relative to D and outer Lipschitz continuous relative
to D with constant κ.
Proof. Let S be inner semicontinuous and outer Lipschitz continuous with constant
κ , both relative to D. Choose y, y ∈ D and let yt = (1 − t)y + ty. The assumed outer
Lipschitz continuity together with the closedness of the values of S implies that for
each t ∈ [0, 1] there exists a positive rt such that

S(u) ⊂ S(yt ) + κ |u − yt |IB for all u ∈ D ∩ IBrt (yt ).

3 Regularity Properties of Set-valued Solution Mappings 157

Let
  
(5) τ = sup t ∈ [0, 1]  S(ys ) ⊂ S(y) + κ |ys − y|IB for each s ∈ [0,t] .

We will show that the supremum in (5) is attained at τ = 1.

First, note that τ > 0 because r0 > 0. Since S(y) is closed, the set S(y) +
κ |yτ − y|IB is closed too, thus its complement, denoted O, is open. Suppose that
yτ ∈ S−1 (O); then, applying Theorem 3B.2(e) to the isc mapping S, we obtain that
there exists σ ∈ [0, τ ) such that yσ ∈ S−1 (O) as well. But this is impossible since
from σ < τ we have

S(yσ ) ⊂ S(y) + κ |yσ − y|IB ⊂ S(y) + κ |yτ − y|IB.

Hence, yτ ∈ / S−1 (O), that is, S(yτ ) ∩ O = 0/ and therefore S(yτ ) is a subset of S(y) +
κ |yτ − y|IB. This implies that the supremum in (5) is attained.
Let us next prove that τ = 1. If τ < 1 there must exist η ∈ (τ , 1) with |yη − yτ | <
rτ such that

(6) S(yη ) ⊂ S(y) + κ |yη − y|IB.

But then, from the definition of rτ ,

S(yη ) ⊂ S(yτ ) + κ |yη − yτ |IB ⊂ S(y) + κ (|yη − yτ | + |yτ − y|)IB = S(y) + κ |yη − y|IB,

where the final equality holds because yτ is a point in the segment [y, yη ]. This
contradicts (6), hence τ = 1. Putting τ = 1 into (5) results in S(y ) ⊂ S(y) + κ |y − y|.
By the symmetry of y and y , we obtain that S is Lipschitz continuous relative to D.
Conversely, if S is Lipschitz continuous relative to D, then S is of course outer
Lipschitz continuous. Let now y ∈ D and let O be an open set such that y ∈ S−1(O).
Then there is x ∈ S(y) and ε > 0 such that x ∈ S(y) ∩ O and x + ε IB ⊂ O. Let 0 <
ρ < ε /κ and pick a point y ∈ D ∩ IBρ (y). Then

x ∈ S(y) ⊂ S(y ) + κ |y − y|IB ⊂ S(y ) + ε IB.

Hence there exists x ∈ S(y ) with |x − x| ≤ ε and thus x ∈ S(y ) ∩ O, that is y ∈
S−1 (O). This means that S−1(O) is open relative to D, and from Theorem 3B.2(e)
we conclude that S is isc relative to D.
We obtain from Theorems 3D.1 and 3D.3 some further insights.

Corollary 3D.4 (Lipschitz continuity of polyhedral mappings). Let S : IRm → → IRn

be polyhedral and let D ⊂ dom S be convex. Then S is isc relative to D if and only
if S is actually Lipschitz continuous relative to D. Thus, for a polyhedral mapping,
continuity relative to its domain implies Lipschitz continuity.
Proof. This is immediate from 3D.3 in the light of 3D.1 and the fact that polyhedral
mappings are osc and in particular closed-valued everywhere.
158 3 Regularity Properties of Set-valued Solution Mappings

Corollary 3D.5 (single-valued polyhedral mappings). Let S : IRm →→ IRn be polyhe-

dral and let D ⊂ dom S be convex. If S is not multi-valued on D, then S must be a
Lipschitz continuous function on D.
Proof. It is sufficient to show that S is isc relative to D. Let y ∈ D and O be an open
set such that y ∈ S−1 (O); then x := S(y) ∈ O. Since S is outer Lipschitz at y, there
exists a neighborhood U of y such that if y ∈ U ∩ D then S(y ) ∈ x + κ |y − y|IB.
Taking U smaller if necessary so that x := S(y ) ∈ x + κ |y − y|IB ⊂ O for y ∈ U,
we obtain that for any y ∈ U ∩ D one has y ∈ S−1 (x ) ⊂ S−1 (O). But then S−1 (O)
must be open relative to D and, from Theorem 3B.2(e), S is isc relative to D.
In the proof of 2E.6 we used the fact that if a function f : IRn → IRm with dom f =
IRn has its graph composed by finitely many polyhedral convex sets, then it must be
Lipschitz continuous. Now this is a particular case of the preceding result.

Corollary 3D.6. If the solution mapping S of the linear variational inequality in Ex-
ercise 3D.2 is single-valued everywhere in IRn , then it must be Lipschitz continuous
globally.

Exercise 3D.7 (distance characterization of outer Lipschitz continuity). Prove that

a mapping S : IRm →
→ IRn is outer Lipschitz continuous at ȳ relative to a set D with
constant κ and neighborhood V if and only if S(ȳ) is closed and

d(x, S(ȳ)) ≤ κ d(ȳ, S−1 (x) ∩ D ∩V ) for all x ∈ IRn .

Guide. Mimic the proof of 3C.1.

In parallel with outer Lipschitz continuity we can introduce inner Lipschitz con-
tinuity of a set-valued mapping S : IRm → → IRn relative to a set D ⊂ IRm at ȳ when
ȳ ∈ D ⊂ dom S, S(ȳ) is a closed set and there exist a constant κ ≥ 0 and a neighbor-
hood V of ȳ such that

S(ȳ) ⊂ S(y) + κ |y − ȳ|IB for all y ∈ V ∩ D.

Inner Lipschitz continuity might be of interest on its own, but no significant appli-
cation of this property in variational analysis has come to light, as yet. Even very
simple polyhedral (nonconvex) mappings don’t have this property (e.g., consider the
mapping from IR to IR, which graph is the union of the axes, and choose the origin
as reference point) as opposed to the outer Lipschitz continuity which holds for ev-
ery polyhedral mapping. In addition, a local version of this property does not obey
the general implicit function theorem paradigm, as we will show in Section 3H. We
therefore drop inner Lipschitz continuity from further consideration.
3 Regularity Properties of Set-valued Solution Mappings 159

3E. Aubin Property, Metric Regularity and Linear Openness

A way to localize the concept of Lipschitz continuity of a set-valued mapping is
to focus on a neighborhood of a reference point of the graph of the mapping and
to use the Pompeiu–Hausdorff distance to a truncation of the mapping with such a
neighborhood. More instrumental turns out to be to take the excess and truncate just
one part of it. This leads us to the following definition.

Aubin property. A mapping S : IRm → → IRn is said to have the Aubin property at ȳ ∈
IR for x̄ ∈ IR if x̄ ∈ S(ȳ) and there is a constant κ ≥ 0 together with neighborhoods
m n

U of x̄ and V of ȳ such that

(1) e(S(y ) ∩U, S(y)) ≤ κ |y − y| for all y , y ∈ V,

or equivalently, there exist κ, U and V , as described, such that

(2) S(y ) ∩U ⊂ S(y) + κ |y − y|IB for all y , y ∈ V.

The infimum of κ over all such combinations of κ, U and V is called the Lipschitz
modulus of S at ȳ for x̄ and denoted by lip (S; ȳ| x̄). The absence of this property is
signaled by lip (S; ȳ| x̄) = ∞.
It is not claimed that (1) and (2) are themselves equivalent, although that is true
when S(y) is closed for every y ∈ V . Nonetheless, the infimum furnishing lip (S; ȳ| x̄)
is the same whichever formulation is adopted. When S is single-valued on a neigh-
borhood of ȳ, then the Lipschitz modulus lip (S; ȳ|S(ȳ)) equals the usual Lipschitz
modulus lip (S; ȳ) for functions.
In contrast to Lipschitz continuity, the Aubin property is tied to a particular point
in the graph of the mapping. As an example, consider the set-valued mapping S :
IR →
→ IR defined as  √
{0, 1 + y} for y ≥ 0,
S(y) =
0 for y < 0.
At 0, the value S(0) consists of two points, 0 and 1. This mapping has the Aubin
property at 0 for 0 but not at 0 for 1. Also, S is not Lipschitz continuous relative to
any interval containing 0.
Observe that when a set-valued mapping S has the Aubin property at ȳ for x̄, then,
for every point (y, x) ∈ gph S which is sufficiently close to (ȳ, x̄), it has the Aubin
property at y for x as well. It is also important to note that the Aubin property of S
at ȳ for x̄ implicitly requires ȳ be an element of int dom S; this is exhibited in the
following proposition.

Proposition 3E.1 (local nonemptiness). If S : IRm → → IRn has the Aubin property at
ȳ for x̄, then for every neighborhood U of x̄ there exists a neighborhood V of ȳ such
that S(y) ∩U = 0/ for all y ∈ V .
Proof. The inclusion (2) for y = ȳ yields
160 3 Regularity Properties of Set-valued Solution Mappings

x̄ ∈ S(y) + κ |y − ȳ|IB for every y ∈ V,

which is the same as

(x̄ + κ |y − ȳ|IB) ∩ S(y) = 0/ for every y ∈ V.

That is, S(y) intersects every neighborhood of x̄ when y is sufficiently close to ȳ.
The property displayed in Proposition 3E.1 is a local version of the inner semi-
continuity. Further, if S is Lipschitz continuous relative to an open set D, then S
has the Aubin property at any y ∈ D ∩ int dom S for any x ∈ S(y). In particular, the
inverse A−1 of a linear mapping A is Aubin continuous at any point provided that
rge A = dom A−1 has nonempty interior, that is, A is surjective. The converse is also
true, since the inverse A−1 of a surjective linear mapping A is Lipschitz continuous
on the whole space, by Theorem 3C.3, and hence A−1 has the Aubin property at any
point.

Proposition 3E.2 (single-valued localization from Aubin property). A set-valued

mapping S : IRm → → IRn has a Lipschitz continuous single-valued localization around
ȳ for x̄ with constant κ if and only if it has a localization at ȳ for x̄ that is not multi-
valued and has the Aubin property at ȳ for x̄ with constant κ.
Proof. Let s be a localization of the second type, which in general is weaker. From
Proposition 3E.1 we have ȳ ∈ int dom S, so s is a single-valued localization of S
around ȳ for x̄. Let a and b be positive constants such that y → s(y) := S(y) ∩ IBa (x̄)
is a function defined on IBb (ȳ) and let b > 0 satisfy b < min{b, a/8κ }. Then for
y, y ∈ IBb (ȳ) we have

d(s(y), S(y )) = d(S(y) ∩ IBa (x̄), S(y ))

≤ κ |y − y| ≤ κ |y − ȳ| + κ |y − ȳ| ≤ 2κ b < a/4.

Hence, there exists x ∈ S(y ) such that |x − s(y)| ≤ d(s(y), S(y ))+ a/4 ≤ a/2. Since
|s(y) − x̄| = d(x̄, S(y)) and

|x − x̄| ≤ |x − s(y)| + |s(y) − x̄| ≤ a/2 + d(x̄, S(y)) ≤ a/2 + κ |y − ȳ| < a

we obtain d(s(y), S(y ) ∩ IBa (x̄)) = d(s(y), S(y )) and therefore

κ |y − y| ≥ d(S(y) ∩ IBa (x̄), S(y ))

= d(s(y), S(y )) = d(s(y), S(y ) ∩ IBa (x̄)) = |s(y) − s(y )|.

Thus, s is a Lipschitz continuous single-valued localization of S around ȳ for x̄ with

constant κ .
Proposition 3E.2 is actually a special case of the following, more general result
in which convexity enters, inasmuch as singletons are convex sets in particular.
3 Regularity Properties of Set-valued Solution Mappings 161

Theorem 3E.3 (Lipschitz continuity under truncation). A set-valued mapping S :

IRm → → IRn whose values are convex sets has the Aubin property at ȳ for x̄ if and only
if it has a Lipschitz continuous graphical localization (not necessarily single-valued)
around ȳ for x̄, or in other words, there are neighborhoods U of x̄ and V of ȳ such
that the truncated mapping y → S(y) ∩U is Lipschitz continuous on V .
Proof. The “if” part holds even without the convexity assumption. Indeed, if y →
S(y) ∩U is Lipschitz continuous on V we have

S(y ) ∩U ⊂ S(y) ∩U + κ |y − y|IB ⊂ S(y) + κ |y − y|IB for all y , y ∈ V,

that is, S has the desired Aubin property. For the “only if” part, suppose now that S
has the Aubin property at ȳ for x̄ with constant κ , and let a > 0 and b > 0 be such
that

(3) S(y ) ∩ IBa (x̄) ⊂ S(y) + κ |y − y|IB for all y , y ∈ IBb (ȳ).

Adjust a and b so that, by 3E.1,

a
(4) S(y) ∩ IBa/2 (x̄) = 0/ for all y ∈ IBb (x̄) and b < .
4κ
Pick y, y ∈ IBb (ȳ) and let x ∈ S(y ) ∩ IBa (x̄). Then from (3) there exists x ∈ S(y) such
that

(5) |x − x | ≤ κ |y − y|.

If x ∈ IBa (x̄), there is nothing more to prove, so assume that r := |x − x̄| > a. By (4),
we can choose a point x̃ ∈ S(y) ∩ IBa/2 (x̄). Since S(y) is convex, there exists a point
z ∈ S(y) on the segment [x, x̃] such that |z − x̄| = a and then z ∈ S(y) ∩ IBa (x̄). We
will now show that

(6) |z − x | ≤ 5κ |y − y|,

which yields that the mapping y → S(y) ∩ IBa (x̄) is Lipschitz continuous on IBb (ȳ)
with constant 5κ .
By construction, there exists t ∈ (0, 1) such that z = (1 − t)x + t x̃. Then

a = |z − x̄| = |(1 − t)(x − x̄) + t(x̃ − x̄)| ≤ (1 − t)r + t|x̃ − x̄|

and in consequence t(r − |x̃ − x̄|) ≤ r − a. Since x̃ ∈ IBa/2 (x̄), we get

r−a
t≤ .
r − a/2

Using the triangle inequality |x̃ − x| ≤ |x − x̄| + |x̃ − x̄| ≤ r + a/2, we obtain
r−a
(7) |z − x| = t|x̃ − x| ≤ (r + a/2).
r − a/2
162 3 Regularity Properties of Set-valued Solution Mappings

Also, in view of (4) and (5), we have that

3a
(8) r = |x − x̄| ≤ |x − x| + |x − x̄| ≤ κ |y − y| + a ≤ κ 2b + a ≤ .
2
From (7), (8) and the inequality r > a we obtain

r + a/2 3a/2 + a/2

(9) |z − x| ≤ (r − a) ≤ (r − a) = 4(r − a).
r − a/2 a − a/2

Note that d := r − a is exactly the distance from x to the ball IBa (x̄), hence d ≤ |x− x |
because x ∈ IBa (x̄). Combining this with (9) and taking into account (6), we arrive
at
|z − x | ≤ |z − x| + |x − x| ≤ 4d + |x − x| ≤ 5|x − x| ≤ 5κ |y − y|.
But this is (6), and we are done.
The Aubin property could alternatively be defined with one variable “free,” as
shown in the next proposition.

Proposition 3E.4 (alternative description of Aubin continuity). A mapping S :

IRm →
→ IRn has the Aubin property at ȳ for x̄ with constant κ if and only if there
exist neighborhoods U of x̄ and V of ȳ such that

(10) e(S(y ) ∩U, S(y)) ≤ κ |y − y| for all y ∈ IRm and y ∈ V.

Proof. Clearly, (10) implies (1). Assume (1) with corresponding U and V and
choose positive a and b such that IBa (x̄) ⊂ U and IBb (ȳ) ⊂ V . Let 0 < a < a and
0 < b < b be such that

(11) 2κ b + a ≤ κ b.

For any y ∈ IBb (ȳ) we have from (1) that

d(x̄, S(y)) ≤ κ |y − ȳ| ≤ κ b ,

hence

(12) e(IBa (x̄), S(y)) ≤ κ b + a .

Take any y ∈ IRm . If y ∈ IBb (ȳ) the inequality in (10) comes from (1) and there is
nothing more to prove. Assume |y − ȳ| > b. Then |y − y | > b − b and from (11),
κ b + a ≤ κ (b − b) ≤ κ |y − y|. Using this in (12) we obtain

e(IBa (x̄), S(y)) ≤ κ |y − y|

and since S(y ) ∩ IBa (x̄) is obviously a subset of IBa (x̄), we come again to (10).
The Aubin property of a mapping is characterized by Lipschitz continuity of the
distance function associated with it.
3 Regularity Properties of Set-valued Solution Mappings 163

Theorem 3E.5 (distance function characterization of Aubin property). For a map-

ping S : IRm →
→ IRn with (ȳ, x̄) ∈ gph S, let s(y, x) = d(x, S(y)). Then S has the Aubin
property at ȳ for x̄ if and only if the function s is Lipschitz continuous with respect
to y uniformly in x around (ȳ, x̄), in which case one has

(13)  (s; (ȳ, x̄)).

lip (S; ȳ| x̄) = lip y

Proof. Let κ > lip (S; ȳ| x̄). Then, from 3E.1, there exist positive constants a and b
such that

(14) 0/ = S(y) ∩ IBa (x̄) ⊂ S(y ) + κ |y − y|IB for all y, y ∈ IBb (ȳ).

Without loss of generality, let a/(4κ ) ≤ b. Let y ∈ IBa/(4κ ) (ȳ) and x ∈ IBa/4 (x̄) and
let x̃ be a projection of x on cl S(y). Using 1D.4(b) and (14) with y = ȳ we have

|x − x̃| = d(x, S(y)) ≤ |x − x̄| + d(x̄, S(y))

a a a
≤ |x − x̄| + e(S(ȳ) ∩ IBa (x̄), S(y)) ≤ + κ |y − ȳ| ≤ + κ = a/2.
4 4 4κ
Hence
a a 3a
|x̄ − x̃| ≤ |x̄ − x| + |x − x̃| ≤ + = < a.
4 2 4
This gives us that

(15) |x − x̃| = d(x, S(y)) = d(x, S(y) ∩ IBa (x̄)).

Now, let y ∈ IBa/(4κ ) (ȳ). The inclusion in (14) yields

 
(16) d x, S(y ) + κ |y − y|IB ≤ d(x, S(y) ∩ IBa (x̄)).

Using the fact that for any set C and for any r ≥ 0 one has

d(x,C) − r ≤ d(x,C + rIB),

from (15) and (16) we obtain

d(x, S(y )) − κ |y − y| ≤ d(x, S(y) ∩ IBa (x̄)) = d(x, S(y)).

By the symmetry of y and y , we conclude that lip  (s; (ȳ, x̄)) ≤ κ . Since κ can be
y
arbitrarily close to lip (S; ȳ| x̄), it follows that

(17)  (s; (ȳ, x̄)) ≤ lip (S; ȳ| x̄).

lip y

Conversely, let κ > lip (s; (ȳ, x̄)). Then there exist neighborhoods U and V of
y
x̄ and ȳ, respectively, such that s(·, x) is Lipschitz continuous relative to V with a
constant κ for any given x ∈ U. Let y, y ∈ V . Since V ⊂ dom s(·, x) for any x ∈ U
we have that S(y ) ∩ U = 0./ Pick any x ∈ S(y ) ∩ U; then s(y , x) = 0 and, by the
164 3 Regularity Properties of Set-valued Solution Mappings

assumed Lipschitz continuity of s(·, x), we get

d(x, S(y)) = s(y, x) ≤ s(y , x) + κ |y − y| = κ |y − y|.

Taking supremum with respect to x ∈ S(y ) ∩ U on the left, we obtain that S has
the Aubin property at ȳ for x̄ with constant κ . Since κ can be arbitrarily close to
 (s; (ȳ, x̄)), we get
lip y
 (s; (ȳ, x̄)) ≥ lip (S; ȳ| x̄).
lip y

This, combined with (17), gives us (13).

The Aubin property of a mapping is closely tied with a property of its inverse,
called metric regularity. The concept of metric regularity goes back to the classical
Banach open mapping principle. We will devote most of Chapter 5 to studying the
metric regularity of set-valued mappings acting in infinite-dimensional spaces.

Metric regularity. A mapping F : IRn → → IRm is said to be metrically regular at x̄ for

ȳ when ȳ ∈ F(x̄) and there is a constant κ ≥ 0 together with neighborhoods U of x̄
and V of ȳ such that

(18) d(x, F −1 (y)) ≤ κ d(y, F(x)) for all (x, y) ∈ U × V.

The infimum of κ over all such combinations of κ, U and V is called the regula-
rity modulus for F at x̄ for ȳ and denoted by reg (F; x̄| ȳ). The absence of metric
regularity is signaled by reg (F; x̄| ȳ) = ∞.
Metric regularity is a valuable concept in its own right, especially for numerical
purposes. For a general set-valued mapping F and a vector y, it gives an estimate
for how far a point x is from being a solution to the generalized equation F(x)  y
in terms of the “residual” d(y, F(x)).
To be specific, let x̄ be a solution of the inclusion ȳ ∈ F(x), let F be metrically
regular at x̄ for ȳ, and let xa and ya be approximations to x̄ and ȳ, respectively. Then
from (18), the distance from xa to the set of solutions of the inclusion ya ∈ F(x)
is bounded by the constant κ times the residual d(ya , F(xa )). In applications, the
residual is typically easy to compute or estimate, whereas finding a solution might
be considerably more difficult. Metric regularity says that there exists a solution to
the inclusion ya ∈ F(x) at distance from xa proportional to the residual. In particular,
if we know the rate of convergence of the residual to zero, then we will obtain the
rate of convergence of approximate solutions to an exact one.
Proposition 3C.1 for a mapping S, when applied to F = S−1 , F −1 = S, ties the
Lipschitz continuity of F −1 relative to a set D to a condition resembling (18), but
with F(x) replaced by F(x) ∩ D on the right, and with U × V replaced by IRn × D.
We demonstrate now that metric regularity of F in the sense of (18) corresponds to
the Aubin property of F −1 for the points in question.
3 Regularity Properties of Set-valued Solution Mappings 165

Theorem 3E.6 (equivalence of metric regularity and the inverse Aubin property). A
set-valued mapping F : IRn → → IRm is metrically regular at x̄ for ȳ with a constant κ if
and only if its inverse F : IRm →
−1
→ IRn has the Aubin property at ȳ for x̄ with constant
κ, i.e. there exist neighborhoods U of x̄ and V of ȳ such that

(19) e(F −1 (y ) ∩U, F −1 (y)) ≤ κ |y − y| for all y , y ∈ V.

Thus,

(20) lip (F −1 ; ȳ| x̄) = reg (F; x̄| ȳ).

Proof. Let κ > reg (F; x̄| ȳ); then there are positive constants a and b such that (18)
holds with U = IBa (x̄), V = IBb (ȳ) and with this κ . Without loss of generality, assume
b < a/κ . Choose y, y ∈ IBb (ȳ). If F −1 (y) ∩ IBa (x̄) = 0,
/ then d(x̄, F −1 (y)) ≥ a. But
then the inequality (18) with x = x̄ yields

a ≤ d(x̄, F −1 (y)) ≤ κ d(y, F(x̄)) ≤ κ |y − ȳ| ≤ κ b < a,

a contradiction. Hence there exists x ∈ F −1 (y) ∩ IBa (x̄), and for any such x we have
from (18) that

(21) d(x, F −1 (y )) ≤ κ d(y , F(x)) ≤ κ |y − y|.

Taking the supremum with respect to x ∈ F −1 (y) ∩ IBa (x̄) we obtain (19) with U =
IBa (x̄) and V = IBb (ȳ), and therefore

(22) reg (F; x̄| ȳ) ≥ lip (F −1 ; ȳ| x̄).

Conversely, suppose there are neighborhoods U of x̄ and V of ȳ along with a

constant κ such that (19) holds. Take U and V smaller if necessary so that, according
to Proposition 3E.4, we have

(23) e(F −1 (y ) ∩U, F −1 (y)) ≤ κ |y − y| for all y ∈ IRm and y ∈ V.

/ then for any y ∈ F(x) we have x ∈ F −1 (y ) ∩U.

Let x ∈ U and y ∈ V . If F(x) = 0,
From (23), we obtain

d(x, F −1 (y)) ≤ e(F −1 (y ) ∩U, F −1 (y)) ≤ κ |y − y|.

This holds for any y ∈ F(x), hence, by taking the infimum with respect to y ∈ F(x)
in the last expression we get

d(x, F −1 (y)) ≤ κ d(y, F(x)).

(If F(x) = 0,
/ then because of the convention d(y, 0) / = ∞, this inequality holds auto-
matically.) Hence, F is metrically regular at x̄ for ȳ with a constant κ . Then we have
κ ≥ reg (F; x̄| ȳ) and hence reg (F; x̄| ȳ) ≤ lip (F −1 ; ȳ| x̄). This inequality together
with (22) results in (20).
166 3 Regularity Properties of Set-valued Solution Mappings

Observe that metric regularity of F at x̄ for ȳ does not require that x̄ ∈ int dom F.
Indeed, if x̄ is an isolated point of dom F then the right side in (18) is ∞ for all x ∈ U,
x = x̄, and then (18) holds automatically. On the other hand, for x = x̄ the right side
of (18) is always finite (since by assumption x̄ ∈ dom F), and then F −1 (y) = 0/ for
y ∈ V . This also follows from 3E.1 via 3E.6.

Exercise 3E.7 (equivalent formulation). Prove that a mapping F is metrically regu-

lar at x̄ for ȳ with constant κ if and only if there are neighborhoods U of x̄ and V of
ȳ such that

(24) d(x, F −1 (y)) ≤ κ d(y, F(x)) for all x ∈ U having F(x)∩V = 0/ and all y ∈ V.

Guide. First, note that (18) implies (24). Let (24) hold with constant κ and neigh-
borhoods IBa (x̄) and IBb (ȳ) having b < a/κ . Choose y, y ∈ IBb (ȳ). As in the proof of
3E.6 show first that F −1 (y) ∩ IBa (x̄) = 0/ by noting that F(x̄) ∩ IBb (ȳ) = 0/ and hence
the inequality in (24) holds for x̄ and y. Then for any x ∈ F −1 (y) ∩ IBa (x̄) we have
that y ∈ F(x) ∩ IBb (ȳ), that is, F(x) ∩ IBb (ȳ) = 0.
/ Thus, the inequality in (24) holds
with y and any x ∈ F −1 (y) ∩ IBa (x̄), which leads to (21) and hence to (19), in the
same way as in the proof of 3E.6. The rest follows from the equivalence of (18) and
(19) established in 3E.6.
There is a third property, which we introduced for functions in Section 1F, and
which is closely related to both metric regularity and the Aubin property.

Openness. A mapping F : IRn → → IRm is said to be open at x̄ for ȳ if ȳ ∈ F(x̄) and for
every neighborhood U of x̄, F(U) is a neighborhood of ȳ.
From the equivalence of metric regularity of F at x̄ for ȳ and the Aubin property
of F −1 at ȳ for x̄, and Proposition 3E.1, we obtain that if a mapping F is metrically
regular at x̄ for ȳ, then F is open at x̄ for ȳ. Metric regularity is actually equivalent
to the following stronger version of the openness property:

Linear openness. A mapping F : IRn → → IRm is said to be linearly open at x̄ for ȳ

when ȳ ∈ F(x̄) and there is a constant κ ≥ 0 together with neighborhoods U of x̄
and V of ȳ such that
 
(25) F(x + κ r intIB) ⊃ F(x) + r intIB ∩V for all x ∈ U and all r > 0.

Openness is a particular case of linear openness and follows from (25) for x = x̄.
Linear openness postulates openness around the reference point with balls having
proportional radii.

Theorem 3E.8 (equivalence of linear openness and metric regularity). A set-valued

mapping F : IRn → → IRm is linearly open at x̄ for ȳ if and only if F is metrically regular
at x̄ for ȳ. In this case the infimum of κ for which (25) holds is equal to reg (F; x̄| ȳ).
Proof. Let (25) hold. Choose y ∈ V and x ∈ U. Let y ∈ F(x ) (if there is no such
y there is nothing to prove). Since y = y + |y − y |w for some w ∈ IB, denoting
3 Regularity Properties of Set-valued Solution Mappings 167

r = |y− y |, for every ε > 0 we have y ∈ (F(x )+ r(1 + ε ) intIB)∩V . From (25), there
exists x ∈ F −1 (y) with |x − x | ≤ κ (1 + ε )r = κ (1 + ε )|y − y|. Then d(x , F −1 (y)) ≤
κ (1+ ε )|y −y|. Taking infimum with respect to y ∈ F(x ) on the right and passing to
zero with ε (since the left side does not depend on ε ), we obtain that F is metrically
regular at x̄ for ȳ with constant κ .
For the converse implication, we use the characterization of the Aubin property
given in Proposition 3E.4. Let x ∈ U, r > 0, and let y ∈ (F(x) + r intIB) ∩ V . Then
there exists y ∈ F(x) such that |y − y | < r. Let ε > 0 be so small that (κ + ε )|y −
y | < κ r. From (10) we obtain d(x, F −1 (y )) ≤ κ |y − y | ≤ (κ + ε )|y − y |. Then
there exists x ∈ F −1 (y ) such that |x − x | ≤ (κ + ε )|y − y |. But then y ∈ F(x ) ⊂
F(x + (κ + ε )|y − y|IB) ⊂ F(x + κ rintIB), which yields (25) with constant κ .
In the classical setting, of course, the equation f (p, x) = 0 is solved for x in
terms of p, and the goal is to determine when this reduces to x being a function of p
through a localization, moreover one with some kind of property of differentiability,
or at least Lipschitz continuity. Relinquishing single-valuedness entirely, we can
look at “solving” the relation

(26) G(p, x)  0 for a mapping G : IRd × IRn →

→ IR ,
m

or in other words studying the solution mapping S : IRd →

→ IRn defined by
  
(27) S(p) = x  G(p, x)  0 .

Fixing a pair ( p̄, x̄) such that x̄ ∈ S( p̄), we can raise questions about local behavior
of S as might be deduced from assumptions on G.
We will concentrate here on the extent to which S can be guaranteed to have the
Aubin property at p̄ for x̄. This turns out to be true when G has the Aubin property
with respect to p and a weakened metric regularity property with respect to x, but
we have to formulate exactly what we need about this in a local sense.

Partial Aubin property. The mapping G : IRd × IRn → → IRm is said to have the par-
tial Aubin property with respect to p uniformly in x at ( p̄, x̄) for ȳ if ȳ ∈ G( p̄, x̄) and
there is a constant κ ≥ 0 together with neighborhoods Q for p̄, U of x̄ and V of ȳ
such that

(28) e(G(p, x) ∩V, G(p , x)) ≤ κ |p − p| for all p, p ∈ Q and x ∈ U,

or equivalently, there exist κ, Q, U and V , as described, such that

G(p, x) ∩V ⊂ G(p , x) + κ |p − p|IB for all p, p ∈ Q and x ∈ U.

The infimum of κ over all such combinations of κ, Q, U and V is called the partial
Lipschitz modulus of G with respect to p uniformly in x at ( p̄, x̄) for ȳ and denoted
 (G; p̄, x̄| ȳ). The absence of this property is signaled by lip
by lip  (G; p̄, x̄| ȳ) = ∞.
p p
The basic result we are able now to state about the solution mapping in (27) could
be viewed as an “implicit function” complement to the “inverse function” result in
168 3 Regularity Properties of Set-valued Solution Mappings

Theorem 3E.6, rather than as a result in the pattern of the implicit function theorem
(which features approximations of one kind or another).

Theorem 3E.9 (Aubin property of general solution mappings). In (26), let G :

IRd × IRn →
→ IRm , with G( p̄, x̄)  0, have the partial Aubin property with respect to p
uniformly in x at ( p̄, x̄) for 0 with constant κ. Furthermore, in the notation (27), let
G enjoy the existence of a constant λ such that

(29) d(x, S(p)) ≤ λ d(0, G(p, x)) for all (p, x) close to ( p̄, x̄).

Then the solution mapping S in (27) has the Aubin property at p̄ for x̄ with constant
λ κ.
Proof. Take p, p ∈ Q and x ∈ S(p) ∩U so that (28) holds for a neighborhood V of
0. From (29) and then (28) we have

d(x, S(p )) ≤ λ d(0, G(p , x)) ≤ λ e(G(p, x) ∩V, G(p , x)) ≤ λ κ |p − p|.

Taking the supremum of the left side with respect to x ∈ S(p) ∩U, we obtain that S
has the Aubin property with constant λ κ .

Example 3E.10. Theorem 3E.9 cannot be extended to a two-way characterization

parallel to Theorem 3E.6. Indeed, consider the “saddle” function of two real vari-
ables f (p, x) = x2 − p2 . In this  case
 f does not 2
 satisfy (29) at the origin of IR , yet

the solution mapping S(p) = x f (p, x) = 0 = {−p, p} has the Aubin property at
0 for 0.
At the end of this section we will take a closer look at the following question. If
a mapping F is simultaneous metrically regular and has the Aubin property, both at
x̄ for ȳ for some (x̄, ȳ) ∈ gph F, then what is the relation, if any, between reg (F; x̄| ȳ)
and lip (F; x̄| ȳ)? Having in mind 3E.6, it is the same as asking what is the relation
between reg (F; x̄| ȳ) and reg (F −1 ; ȳ| x̄) or between lip (F; x̄| ȳ) and lip (F −1 ; ȳ| x̄).
When we exclude the trivial case when (x̄, ȳ) ∈ int gph F, in which case both moduli
would be zero, an answer to this question is stated in the following exercise.

Exercise 3E.11. Consider a mapping F : IRn →

→ IRm with closed graph and a point
(x̄, ȳ) ∈ gph F \ int gph F . Then

reg (F; x̄| ȳ)·lip (F; x̄| ȳ) ≥ 1,

including the limit cases when either of these moduli is 0 and then the other is ∞
under the convention 0∞ = ∞.
Guide. Let κ > reg (F; x̄| ȳ) and γ > lip (F; x̄| ȳ). Then there are neighborhoods U
of x̄ and V of ȳ corresponding to metric regularity and the Aubin property of F with
constants κ and γ , respectively. Let (x, y) ∈ U × V be such that d(x, F −1 (y)) > 0
(why does such a point exist?). Then there exists x ∈ F −1 (y) such that 0 < |x − x | =
d(x, F −1 (y)). We have
3 Regularity Properties of Set-valued Solution Mappings 169

|x − x | = d(x, F −1 (y)) ≤ κ d(y, F(x)) ≤ κ e(F(x ) ∩U, F(x)) ≤ κγ |x − x|.

Hence, κγ ≥ 1.
For a solution mapping S = F −1 of an inclusion F(x)  y with a parameter y, the
quantity lip (S; ȳ| x̄) measures how “stable” solutions near x̄ are under changes of the
parameter near ȳ. In this context, the smaller this modulus is, the “better” stability we
have. In view of 3E.11, better stability means larger values of the regularity modulus
reg (S; ȳ| x̄). In the limit case, when S is a constant function near ȳ, that is, when the
solution is not sensitive at all with respect to small changes of the parameter y near
ȳ, then lip (S; ȳ|S(ȳ)) = 0 while the metric regularity modulus of S there is infinity.
In Section 6A we will see that the “larger” the regularity modulus of a mapping is,
the “easier” it is to perturb the mapping so that it looses its metric regularity. In this
connection, E3.11 provides a good candidate for a condition number of a general
mapping, but we shall not go into this here.

3F. Implicit Mapping Theorems with Metric Regularity

In the paradigm of the implicit function theorem, as applied to a generalized equa-
tion f (p, x) + F (x)  0 with solutions x ∈ S(p), the focus is on some p̄ and x̄ ∈ S( p̄),
and on some kind of approximation of the mapping x → f ( p̄, x) + F(x). Assump-
tions about this approximation lead to conclusions about the solution mapping S
relative to p̄ and x̄. Stability properties of the approximation are the key to progress
in this direction. Our aim now is to study such stability with respect to metric re-
gularity and to show that this leads to implicit function theorem type results which
apply to set-valued solution mappings S beyond any framework of single-valued
localization.
We start this section with a particular case of a fundamental result in variational
analysis and beyond, stated below as Theorem 3F.1, which goes back to works by
Lyusternik and Graves. We will devote most of Chapter 5 to the full theory behind
this result—in an infinite-dimensional setting. In the statement of the theorem the
following concept is required.

Locally closed sets. A set C is said to be locally closed at x ∈ C if there exists a

neighborhood U of x such that the intersection C ∩U is closed.
Local closedness of a set C at x ∈ C can be equivalently defined as the existence
of a scalar a > 0 such that the set C ∩ IBa (x) is closed.

Theorem 3F.1 (inverse mapping theorem with metric regularity). Consider a map-
ping F : IRn →
→ IRm and any (x̄, ȳ) ∈ gph F at which gph F is locally closed and let κ
and μ be nonnegative constants such that

reg (F; x̄| ȳ) ≤ κ and κ μ < 1.

170 3 Regularity Properties of Set-valued Solution Mappings

Then for any function g : IRn → IRm with x̄ ∈ int dom g and lip (g; x̄) ≤ μ, one has
κ
(1) reg (g + F; x̄|g(x̄) + ȳ) ≤ .
1 − κμ

Although formally there is no inversion of a mapping in Theorem 3F.1, if this

result is stated equivalently in terms of the Aubin property of the inverse mapping
F −1 , it fits then into the pattern of the inverse function theorem paradigm. It can also
be viewed as a result concerning stability of metric regularity under perturbations
by functions with small Lipschitz constants. We can actually deduce the classical
inverse function theorem 1A.1 from 3F.1. Indeed, let f : IRn → IRn be a smooth
function around x̄ and let ∇ f (x̄) be nonsingular. Then F = D f (x̄) is metrically re-
gular everywhere and from 3F.1 for the function g(x) = f (x) − D f (x̄)(x − x̄) with
lip (g; x̄) = 0 we obtain that g + F = f is metrically regular at x̄ for f (x̄). But then f
must be open (cf. 3E.8). Establishing this fact is the main part of all proofs of 1A.1
presented so far.
We will postpone proving Theorem 3F.1 to Chapter 5, for a mapping F acting
from a complete metric space to a linear metric space. In this section we focus on
some consequences of this result and its implicit function version. Several corollar-
ies of 3F.1 will lead the way.

Corollary 3F.2 (detailed estimates). Consider a mapping F : IRn → → IRm and any
pair (x̄, ȳ) ∈ gph F at which gph F is locally closed. If reg (F; x̄| ȳ) > 0, then for any
g : IRn → IRm such that reg (F; x̄| ȳ) · lip (g; x̄) < 1, one has

(2) reg (g + F; x̄|g(x̄) + ȳ) ≤ (reg (F; x̄| ȳ)−1 − lip (g; x̄))−1 .

If reg (F; x̄| ȳ) = 0, then reg (g + F; x̄|g(x̄) + ȳ) = 0 for any g : IRn → IRm with
lip (g; x̄) < ∞. If reg (F; x̄| ȳ) = ∞, then reg (g + F; x̄|g(x̄) + ȳ) = ∞ for any g : IRn →
IRm with lip (g; x̄) = 0.
Proof. If reg (F; x̄| ȳ) < ∞, then by choosing κ and μ appropriately and pass-
ing to limits in (1) we obtain the claimed inequality (2) also for the case where
reg (F; x̄| ȳ) = 0. Let reg (F; x̄| ȳ) = ∞, and suppose that reg (g + F; x̄|g(x̄) + ȳ) < κ
for some κ and a function g with lip (g; x̄) = 0. Note that g is Lipschitz continuous
around x̄, hence the graph of g is locally closed around (x̄, g(x̄)). Then g + F has
locally closed graph at (x̄, g(x̄) + ȳ). Applying Theorem 3F.1 to the mapping g + F
with perturbation −g, and noting that lip (−g; x̄) = 0, we get reg (F; x̄| ȳ) ≤ κ , which
constitutes a contradiction.
When the perturbation g has zero Lipschitz modulus at the reference point, we
obtain another interesting fact.

Corollary 3F.3 (perturbations with Lipschitz modulus 0). Consider a mapping F :

IRn →→ IRm and any pair (x̄, ȳ) ∈ gph F at which gph F is locally closed. Then for any
g : IRn → IRm with lip (g; x̄) = 0 one has

reg (g + F; x̄|g(x̄) + ȳ) = reg (F; x̄| ȳ).

3 Regularity Properties of Set-valued Solution Mappings 171

Proof. The cases with reg (F; x̄| ȳ) = 0 or reg (F; x̄| ȳ) = ∞ are already covered by
Corollary 3F.2. If 0 < reg (F; x̄| ȳ) < ∞, we get from (2) that

reg (g + F; x̄|g(x̄) + ȳ) ≤ reg (F; x̄| ȳ).

By exchanging the roles of F and g + F, we also get

reg (F; x̄| ȳ) ≤ reg (g + F; x̄|g(x̄) + ȳ),

and in that way arrive at the claimed equality.

An elaboration of Corollary 3F.3 employs first-order approximations of a func-
tion as were introduced in Section 1E.

Corollary 3F.4 (utilization of strict first-order approximations). Consider a map-

ping F : IRn →
→ IRm and any pair (x̄, ȳ) ∈ gph F at which gph F is locally closed. Let
f : IR → IRm be continuous in a neighborhood of x̄. Then, for any h : IRn → IRm
n

which is a strict first-order approximation to f at x̄, one has

reg ( f + F; x̄| f (x̄) + ȳ) = reg (h + F; x̄|h(x̄) + ȳ).

In particular, when the strict first-order approximation is represented by the lin-

earization coming from strict differentiability, we get something even stronger.

Corollary 3F.5 (utilization of strict differentiability). Consider M = f + F for a

function f : IRn → IRm and a mapping F : IRn → → IRm , and let ȳ ∈ M(x̄). Suppose
that f : IR → IR is strictly differentiable at x̄ and that gph M is locally closed at
n m

(x̄, ȳ) (or in other words that gph F is locally closed at (x̄, ȳ − f (x̄))). Then, for the
linearization
M0 (x) = f (x̄) + ∇ f (x̄)(x − x̄) + F(x)
one has
reg (M; x̄ | ȳ) = reg (M0 ; x̄| ȳ).
In the case when m = n and the mapping F is the normal cone mapping to a
polyhedral convex set, we can likewise employ a “first-order approximation” of F.
When f is linear the corresponding result parallels 2E.6.

Corollary 3F.6 (affine-polyhedral variational inequalities). For an n × n matrix A

and a polyhedral convex set C ⊂ IRn , consider the variational inequality

Ax + NC (x)  y.

Let x̄ be a solution for ȳ, let v̄ = ȳ − Ax̄, so that v̄ ∈ NC (x̄), and let K = KC (x̄, v̄) be
the critical cone to C at x̄ for v̄. Then, for the mappings

G(x) = Ax + NC (x) with G(x̄)  ȳ,

G0 (w) = Aw + NK (w) with G0 (0)  0,
172 3 Regularity Properties of Set-valued Solution Mappings

we have
reg (G; x̄| ȳ) = reg (G0 ; 0|0).
Proof. From reduction lemma 2E.4, for (w, u) in a neighborhood of (0, 0), we have
that v̄ + u ∈ NC (x̄ + w) if and only if u ∈ NK (w). Then, for (w, v) in a neighborhood
of (0, 0), we obtain ȳ+ v ∈ G(x̄ + w) if and only if v ∈ G0 (w). Thus, metric regularity
of A + NC at x̄ for ȳ with a constant κ implies metric regularity of A + NK at 0 for 0
with the same constant κ , and conversely.
Combining 3F.5 and 3F.6 we obtain the following corollary:

Corollary 3F.7 (strict differentiability and polyhedral convexity). Consider M =

f + NC for a function f : IRn → IRn and a polyhedral convex set C ⊂ IRn , let ȳ ∈ M(x̄)
and let f be strictly differentiable at x̄. For v̄ = ȳ − f (x̄), let K = KC (v̄, x̄) be the
critical cone to the set C at x̄ for v̄. Then, for M0 (x) = ∇ f (x̄)x + NK (x) one has

reg (M; x̄| ȳ) = reg (M0 ; 0|0).

We are ready now to take up once more the study of a generalized equation
having the form

(3) f (p, x) + F(x)  0

for f : IRd × IRn → IRm and F : IRn →

→ IRm , and its solution mapping S : IRd → IRn
defined by
  
(4) S(p) = x  f (p, x) + F(x)  0 .

This time, however, we are not looking for single-valued localizations of S but aim-
ing at a better understanding of situations in which S may not have any such localiza-
tion, as in the example of parameterized constraint systems. Recall from Chapter 1
that, for f : IRd × IRn → IRm and a point ( p̄, x̄) ∈ int dom f , a function h : IRn → IRm
is said to be a strict estimator of f with respect to x uniformly in p at ( p̄, x̄) with
constant μ if h(x̄) = f (x̄, p̄) and

 (e; ( p̄, x̄)) ≤ μ < ∞ for e(p, x) = f (p, x) − h(x).

lip x

Theorem 3F.8 (implicit mapping theorem with metric regularity). For the general-
ized equation (3) and its solution mapping S in (4), and a pair ( p̄, x̄) with x̄ ∈ S( p̄),
let h : IRn → IRm be a strict estimator of f with respect to x uniformly in p at ( p̄, x̄)
with constant μ, let gph (h + F) be locally closed at (x̄, 0) and let h + F be metrically
regular at x̄ for 0 with reg (h + F; x̄|0) ≤ κ. Suppose that

(5)  ( f ; ( p̄, x̄)) ≤ λ < ∞.

κ μ < 1 and lip p

Then S has the Aubin property at p̄ for x̄, and moreover

3 Regularity Properties of Set-valued Solution Mappings 173

κλ
(6) lip (S; p̄| x̄) ≤ .
1 − κμ
This theorem will be established in an infinite-dimensional setting in Section 5E,
so we will not prove it separately here. An immediate consequence is obtained by
specializing the function h in Theorem 3F.8 to a linearization of f with respect to
x. We add to this the effect of ample parameterization, in parallel to the case of
single-valued localization in Theorem 2C.2.

Theorem 3F.9 (using strict differentiability and ample parameterization). For the
generalized equation (3) and its solution mapping S in (4), and a pair ( p̄, x̄) with
x̄ ∈ S( p̄), suppose that f is strictly differentiable at ( p̄, x̄) and that gph F is locally
closed at (x̄, − f ( p̄, x̄)). If the mapping

h + F for h(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄)

is metrically regular at x̄ for 0, then S has the Aubin property at p̄ for x̄ with

(7) lip (S; p̄| x̄) ≤ reg (h + F; x̄|0) · |∇ p f ( p̄, x̄)|.

Furthermore, when f satisfies the ample parameterization condition

(8) rank ∇ p f ( p̄, x̄) = m,

then the converse implication holds as well: the mapping h + F is metrically regular
at x̄ for 0 provided that S has the Aubin property at p̄ for x̄.
Proof of 3F.9, initial part. In these circumstances with this choice of h, the condi-
tions in (5) are satisfied because, for e = f − h,

 (e; ( p̄, x̄)) = 0 and lip

lip  ( f ; ( p̄, x̄)) = |∇ p f ( p̄, x̄)|.
x p

Thus, (7) follows from (6) with μ = 0. In the remainder of the proof, regarding
ample parameterization, we will make use of the following fact.

Proposition 3F.10 (Aubin property in composition). For a mapping M : IRd →

→ IRn
m d m →
and a function ψ : IR × IR → IR consider the composite mapping N : IR → IRn of
n

the form   
y → N(y) = x  x ∈ M(ψ (x, y)) for y ∈ IRm .
Let ψ satisfy

(9)  (ψ ; (x̄, 0)) = 0 and lip

lip  (ψ ; (x̄, 0)) < ∞,
x y

and, for p̄ = ψ (x̄, 0), let ( p̄, x̄) ∈ gph M at which gph M is locally closed. Under
these conditions, if M has the Aubin property at p̄ for x̄, then N has the Aubin
property at 0 for x̄.
174 3 Regularity Properties of Set-valued Solution Mappings

Proof. Let the mapping M have the Aubin property at p̄ for x̄ with neighborhoods
Q of p̄ and U of x̄ and constant κ > lip (M; p̄ | x̄). Choose λ > 0 with λ < 1/κ and
 (ψ ; (x̄, 0)). By (9) there exist positive constants a and b such that for
let γ > lip y
any y ∈ IBa (0) the function ψ (·, y) is Lipschitz continuous on IBb (x̄) with Lipschitz
constant λ and for every x ∈ IBb (x̄) the function ψ (x, ·) is Lipschitz continuous on
IBa (0) with Lipschitz constant γ . Pick a positive constant c and make a and b smaller
if necessary so that:
(a) IBc ( p̄) ⊂ Q and IBb (x̄) ⊂ U,
(b) the set gph M ∩ (IBc ( p̄) × IBb (x̄)) is closed, and
(c) the following inequalities are satisfied:
4κγ a
(10) ≤ b and γ a + λ b ≤ c.
1 − κλ
Let y , y ∈ IBa (0) and let x ∈ N(y ) ∩ IBb/2 (x̄). Then x ∈ M(ψ (x , y )) ∩ IBb/2 (x̄).
Further, we have

|ψ (x , y ) − p̄| ≤ |ψ (x , y ) − ψ (x , 0)| + |ψ (x , 0) − ψ (x̄, 0)| ≤ γ a + λ b/2 ≤ c

and the same for ψ (x , y). From the Aubin property of M we obtain the existence of
x1 ∈ M(ψ (x , y)) such that

|x1 − x | ≤ κ |ψ (x , y ) − ψ (x , y)| ≤ κγ |y − y|.

Thus, through the first inequality in (10),

b
|x1 − x̄| ≤ |x1 − x | + |x − x̄| ≤ κγ |y − y| + |x − x̄| ≤ κγ (2a) + ≤ b,
2
and consequently

|ψ (x1 , y) − p̄| = |ψ (x1 , y) − ψ (x̄, 0)| ≤ λ b + γ a ≤ c,

utilizing the second inequality in (10). Hence again, from the Aubin property of M
applied to x1 ∈ M(ψ (x , y)) ∩ IBb (x̄), there exists x2 ∈ M(ψ (x1 , y)) such that

|x2 − x1 | ≤ κ |ψ (x1 , y) − ψ (x , y)| ≤ κλ |x1 − x | ≤ (κλ )κγ |y − y|.

Employing induction, assume that we have a sequence {x j } with

x j ∈ M(ψ (x j−1 , y)) and |x j − x j−1 | ≤ (κλ ) j−1 κγ |y − y| for j = 1, . . . , k.

Setting x0 = x , for k = 1, 2, . . . we get

k
|xk − x̄| ≤ |x0 − x̄| + ∑ |x j − x j−1 |
j=1
3 Regularity Properties of Set-valued Solution Mappings 175

b k−1 b 2aκγ
≤ + ∑ (κλ ) j κγ |y − y| ≤ + ≤ b,
2 j=0 2 1 − κλ

where we use the first inequality in (10). Hence |ψ (xk , y) − p̄| ≤ λ b + γ a ≤ c. Then
there exists xk+1 ∈ M(ψ (xk , y)) such that

|xk+1 − xk | ≤ κ |ψ (xk , y) − ψ (xk−1 , y)| ≤ κλ |xk − xk−1 | ≤ (κλ )k κγ |y − y|,

and the induction step is complete.

The sequence {xk } is Cauchy, hence convergent to some x ∈ IBb (x̄) ⊂ U. From the
local closedness of gph M and the continuity of ψ we deduce that x ∈ M(ψ (x, y)),
hence x ∈ N(y). Furthermore, using the estimate
k k−1
κγ
|xk − x0 | ≤ ∑ |x j − x j−1 | ≤ ∑ (κλ ) j κγ |y − y| ≤ 1 − κλ |y − y|
j=1 j=0

we obtain, for any κ  ≥ (κγ )/(1 − κλ ), and on passing to the limit with respect
to k → ∞, that |x − x | ≤ κ  |y − y|. Thus, N has the Aubin property at 0 for x̄ with
constant κ  .

Proof of 3F.9, final part. Under the ample parameterization condition (8), Lemma
2C.1 guarantees the existence of neighborhoods U of x̄, V of 0, and Q of p̄, as well
as a local selection ψ : U × V → Q around (x̄, 0) for p̄ of the mapping
  
(x, y) → p  y + f (p, x) = h(x)

for h(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄) which satisfies the conditions in (9). Hence,

y + f (ψ (x, y), x) = h(x) and ψ (x, y) ∈ Q for x ∈ U, y ∈ V.

Fix y ∈ V . If x ∈ (h+F)−1 (y)∩U and p = ψ (x, y), then p ∈ Q and y+ f (p, x) = h(x),
hence x ∈ S(p)∩U. Conversely, if x ∈ S(ψ (x, y))∩U, then clearly x ∈ (h+F)−1 (y)∩
U. Thus,
  
(11) (h + F)−1 (y) ∩U = x  x ∈ S(ψ (x, y)) ∩U .

Since the Aubin property of S at p̄ for x̄ is a local property of the graph of S relative to
the point ( p̄, x̄), it holds if and only if the same holds for the truncated mapping SU :
p → S(p) ∩ U (see Exercise 3F.11 below). That equivalence is valid for (h + F)−1
as well. Thus, if the mapping SU has the Aubin property at p̄ for x̄, from Proposition
3F.10 in the context of (11), we obtain that (h + F)−1 has the Aubin property at 0
for x̄, hence, by 3E.6, h + F is metrically regular at x̄ for 0 as desired.

Exercise 3F.11. Let S : IRm → → IRn have the Aubin property at ȳ for x̄ with constant
κ. Show that for any neighborhood U of x̄ the mapping SU : y → S(y) ∩U also has
the Aubin property at ȳ for x̄ with constant κ.
176 3 Regularity Properties of Set-valued Solution Mappings

Guide. Choose sufficiently small a > 0 and b > 0 such that IBa (x̄) ⊂ U and b ≤
a/(4κ ). Then for every y, y ∈ IBb (ȳ) and every x ∈ S(y) ∩ IBa/2 (x̄) there exists x ∈
S(y ) with |x − x| ≤ κ |y − y| ≤ 2κ b ≤ a/2. Then both x and x are from U.
Let us now look at the case of 3F.9 in which F is a constant mapping, F(x) ≡ K,
which was featured at the beginning of this chapter as a motivation for investigating
real set-valuedness in solution mappings. Solving f (p, x) + F(x)  0 for a given p
then means finding an x such that − f (p, x) ∈ K. For particular choices of K this
amounts to solving some mixed system of equations and inequalities, for example.
Example 3F.12 (application to general constraint systems). For f : IRd × IRn → IRm
and a closed set K ⊂ IRm , let
  
S(p) = x  f (p, x) ∈ K .

Fix p̄ and x̄ ∈ S( p̄). Suppose that f is continuously differentiable on a neighborhood

of ( p̄, x̄), and consider the solution mapping for an associated linearized system:
  
S̄(y) = x  y − f ( p̄, x̄) − ∇x f ( p̄, x̄)(x − x̄) ∈ K .

If S̄ has the Aubin property at 0 for x̄, then S has the Aubin property at p̄ for x̄. The
converse implication holds under the ample parameterization condition (8).
The key to applying this result, of course, is being able to ascertain when the
linearized system does have the Aubin property in question. In the important case of
K = IRs− × {0}m−s, a necessary and sufficient condition will emerge in the so-called
Mangasarian–Fromovitz constraint qualification. This will be seen in Section 4D.
Example 3F.13 (application to polyhedral variational inequalities). For f : IRd ×
IRn → IRn and a convex polyhedral set C ⊂ IRn , let
  
S(p) = x  f (p, x) + NC (x)  0 .

Fix p̄ and x̄ ∈ S( p̄) and for v̄ = − f ( p̄, x̄) let K = KC (x̄, v̄) be the associated critical
cone to C. Suppose that f is continuously differentiable on a neighborhood of ( p̄, x̄),
and consider the solution mapping for an associated reduced system:
  
S̄(y) = x  ∇x f ( p̄, x̄)x + NK (x)  y .

If S̄ has the Aubin property at 0 for 0, then S has the Aubin property at p̄ for x̄. The
converse implication holds under the ample parameterization condition (8).
If a mapping F : IRn → → IRm has the Aubin property at x̄ for ȳ, then for any function
f : IR → IR with lip ( f ; x̄) < ∞, the mapping f + F has the Aubin property at x̄ for
n m

f (x̄) + ȳ as well. This is a particular case of the following observation which utilizes
ample parameterization.

Exercise 3F.14. Consider a mapping F : IRn → → IRm with (x̄, ȳ) ∈ gph F and a func-
tion f : IR × IR → IR having ȳ = − f ( p̄, x̄) and which is strictly differentiable at
d n m

( p̄, x̄) and satisfies the ample parameterization condition (8). Prove that the mapping
3 Regularity Properties of Set-valued Solution Mappings 177

x → P(x) = {p | 0 ∈ f (p, x) + F(x)}

has the Aubin property at x̄ for p̄ if and only if F has the Aubin property at x̄ for ȳ.
Guide. First, apply 3F.9 to show that, under the ample parameterization condition
(8), the mapping
  
(x, y) → Ω (x, y) = p  y + f (p, x) = 0

has the Aubin property at (x̄, ȳ) for p̄. Let F have the Aubin property at x̄ for ȳ with
neighborhoods U of x̄ of V for ȳ and constant κ . Choose a neighborhood Q of p̄ and
adjust U and V accordingly so that Ω has the Aubin property with constant λ and
neighborhoods U × V and Q. Let b > 0 be such that IBb (ȳ) ⊂ V , then choose a > 0
and adjust Q such that IBa (x̄) ⊂ U, a ≤ b/(4κ ) and also − f (p, x) ∈ IBb/2 (ȳ) for x ∈
IBa (x̄) and p ∈ Q. Let x, x ∈ IBa (x̄) and p ∈ P(x) ∩ Q. Then y = − f (p, x) ∈ F(x) ∩V
and by the Aubin property of F there exists y ∈ F(x ) such that |y − y | ≤ κ |x − x |.
But then |y − ȳ| ≤ κ (2a) + b/2 ≤ b. Thus y ∈ V and hence, by the Aubin property
of Ω , there exists p satisfying y + f (p , x ) = 0 and

|p − p| ≤ λ (|y − y| + |x − x|) ≤ λ (κ + 1)|x − x|.

Noting that p ∈ P(x ) we get that P has the Aubin property at x̄ for p̄.
Conversely, let P have the Aubin property at x̄ for p̄ with associated constant κ
and neighborhoods U and Q of x̄ and p̄, respectively. Let f be Lipschitz continuous
on Q × U with constant μ . We already know that the mapping Ω has the Aubin
property at (x̄, ȳ) for p̄; let λ be the associated constant and U ×V and Q the neigh-
borhoods of (x̄, ȳ) and of p̄, respectively. Choose c > 0 such that IBc ( p̄) ⊂ Q and let
a > 0 satisfy

IBa (x̄) ⊂ U, IBa (ȳ) ⊂ V and a max{κ , λ } ≤ c/4.

Let x, x ∈ IBa (x̄) and y ∈ F(x) ∩ IBa (ȳ). Since Ω has the Aubin property and
p̄ ∈ Ω (x̄, ȳ) ∩ IBc ( p̄), there exists p ∈ Ω (x, y) such that |p − p̄| ≤ λ (2a) ≤ c/2.
This means that p ∈ P(x) ∩ IBc/2 ( p̄) and from the Aubin property of P there ex-
ists p ∈ P(x ) so that |p − p| ≤ κ |x − x|. Thus, |p − p̄| ≤ κ (2a) + c/2 ≤ c. Let
y = − f (p , x ). Then y ∈ F(x ) because p ∈ P(x ) and the Lipschitz continuity of
f gives us

|y − y | = | f (p, x) − f (p , x )| ≤ μ (|p − p | + |x − x|) ≤ μ (κ + 1)|x − x|.

Hence, F has the Aubin property at x̄ for ȳ.

In none of the directions of the statement of 3F.14 we can replace the Aubin
property by metric regularity. Indeed, the function f (p, x) = x + p satisfies the as-
sumptions, but if we add to it the zero mapping F ≡ 0, which is not metrically
regular (anywhere), we get the mapping P(x) = −x which is metrically regular (ev-
erywhere). Taking the same f and F(x) = −x contradicts the other direction.
178 3 Regularity Properties of Set-valued Solution Mappings

3G. Strong Metric Regularity

Although our chief goal in this chapter has been the treatment of solution mappings
for which Lipschitz continuous single-valued localizations need not exist or even
be a topic of interest, the concepts and results we have built up can shed new light
on our earlier work with such localizations through their connection with metric
regularity.

Proposition 3G.1 (single-valued localizations and metric regularity). For a map-

ping F : IRn → → IRm and a pair (x̄, ȳ) ∈ gph F , the following properties are equivalent:
−1
(a) F has a Lipschitz continuous single-valued localization s around ȳ for x̄.
(b) F is metrically regular at x̄ for ȳ and F −1 has a localization at ȳ for x̄ that is
nowhere multi-valued.
Indeed, in the circumstances of (b) the localization s in (a) has lip (s; ȳ) =
reg (F; x̄| ȳ).
Proof. According to 3E.2 as applied to S = F −1 , condition (a) is equivalent to F −1
having the Aubin property at ȳ for x̄ and a localization around ȳ for x̄ that is nowhere
multi-valued. When F −1 has the Aubin property at ȳ for x̄, by 3E.1 the domain of
F −1 contains a neighborhood of ȳ, hence any localization of F −1 at ȳ for x̄ is actually
a localization around ȳ for x̄. On the other hand, we know from 3E.6 that F −1 has
the Aubin property at ȳ for x̄ if and only if F is metrically regular at x̄ for ȳ. That
result also relates the constants κ in the two properties and yields for us the final
statement.

Proposition 3G.2 (stability of single-valuedness under perturbation). For a map-

ping F : IRn →
→ IRm and a pair (x̄, ȳ) ∈ gph F , let F −1 have a Lipschitz continuous
single-valued localization s around ȳ for x̄ and let λ > lip (s; ȳ). Then for any posi-
tive ν < λ −1 and for any function g : IRn → IRm with x̄ ∈ int dom g and lip (g; x̄) < ν
the mapping (g + F)−1 has a localization at g(x̄) + ȳ for x̄ which is nowhere multi-
valued.
Proof. Our hypothesis says that there are neighborhoods U of x̄ and V of ȳ such
that for any y ∈ V the set F −1 (y) ∩U consists of exactly one point, s(y), and that the
function s : y → F −1 (y) ∩ U is Lipschitz continuous on V with Lipschitz constant
λ . Let 0 < ν < λ −1 and choose a function g : IRn → IRm and a neighborhood U  of
x̄ on which g is Lipschitz continuous with constant ν . We can find neighborhoods
U0 = IBτ (x̄) ⊂ U ∩U  and V0 = IBε (g(x̄) + ȳ) ⊂ (g(x̄) + V ) such that

(1) x ∈ U0 , y ∈ V0 =⇒ y − g(x) ∈ V.

Consider now the graphical localization of (g + F)−1 corresponding to U0 and V0 .

Each y ∈ V0 maps to the set (g + F)−1 (y) ∩U0 ; it will be demonstrated that this set
can have at most one element, and that will finish the proof.
3 Regularity Properties of Set-valued Solution Mappings 179

Suppose to the contrary that y ∈ V0 and x, x ∈ U0 , x = x , are such that both x and
xbelong to (g + F)−1 (y). Clearly x ∈ (g + F)−1 (y) ∩ U0 if and only if x ∈ U0 and
y ∈ g(x) + F(x), or equivalently y − g(x) ∈ F(x). The latter, in turn, is the same as
having x ∈ F −1 (y − g(x)) ∩U0 ⊂ F −1 (y − g(x)) ∩U = s(y − g(x)), where y − g(x) ∈
V by (1). Then

0 < |x − x | = |s(y − g(x)) − s(y − g(x))| ≤ λ |g(x) − g(x)| ≤ λ ν |x − x | < |x − x|,

which is absurd.
The observation in 3G.1 leads to a definition.

Strong metric regularity. A mapping F : IRn → → IRm having the equivalent proper-
ties in 3G.1 will be called strongly metrically regular at x̄ for ȳ.
For a linear mapping represented by an m × n matrix A, strong metric regularity
comes out as the nonsingularity of A and thus requires that m = n. Moreover, for
any single-valued function f : IRn → IRm , strong metric regularity requires m = n by
Theorem 1F.1 on the invariance of domain. This property can be seen therefore as
corresponding closely to the one in the classical implicit function theorem, except
for its focus on Lipschitz continuity instead of continuous differentiability. It was
the central property in fact, if not in name, in Robinson’s implicit function theorem
2B.1.
The terminology of strong metric regularity offers a way of gaining new per-
spectives on earlier results by translating them into the language of metric regula-
rity. Indeed, strong metric regularity is just metric regularity plus the existence of a
single-valued localization of the inverse. According to Theorem 3F.1, metric regula-
rity of a mapping F with a locally closed graph is stable under addition of a function
g with a “small” Lipschitz constant, and so too is local single-valuedness, accord-
ing to 3G.2 above. Thus, strong metric regularity must be stable under perturbation
in the same way as metric regularity. The corresponding result is a version of the
inverse function result in 2B.10 corresponding to the extended form of Robinson’s
implicit function theorem in 2B.5.

Theorem 3G.3 (inverse function theorem with strong metric regularity). Consider
a mapping F : IRn → → IRm and any (x̄, ȳ) ∈ gph F such that F is strongly metrically
regular at x̄ for ȳ. Let κ and μ be nonnegative constants such that

reg (F; x̄| ȳ) ≤ κ and κ μ < 1.

Then for any function g : IRn → IRm with x̄ ∈ int dom g and lip (g; x̄) ≤ μ, the map-
ping g + F is strongly metrically regular at x̄ for g(x̄) + ȳ. Moreover,
   κ
reg g + F; x̄g(x̄) + ȳ ≤ .
1 − κμ

Proof. Our hypothesis that F is strongly metrically regular at x̄ for ȳ implies that
a graphical localization of F −1 around (ȳ, x̄) is single-valued and continuous near
180 3 Regularity Properties of Set-valued Solution Mappings

ȳ and therefore that gph F is locally closed at (x̄, ȳ). Further, by fixing λ > κ and
using Proposition 3G.1, we can get neighborhoods U of x̄ and V of ȳ such that for
any y ∈ V the set F −1 (y) ∩ U consists of exactly one point, which we may denote
by s(y) and know that the function s : y → F −1 (y) ∩ U is Lipschitz continuous on
V with Lipschitz constant λ . Let μ < ν < λ −1 and choose a function g : IRn → IRm
and a neighborhood U  ⊂ U of x̄ on which g is Lipschitz continuous with constant
ν . Applying Proposition 3G.2 we obtain that the mapping (g + F)−1 has a local-
ization at g(x̄) + ȳ for x̄ which is nowhere multi-valued. On the other hand, since F
has locally closed graph at (x̄, ȳ), we know from Theorem 3F.1 that for such g the
mapping g + F is metrically regular at g(x̄) + ȳ for x̄. Applying Proposition 3G.1
once more, we complete the proof.
In much the same way we can state in terms of strong metric regularity an implicit
function result paralleling Theorem 2B.5.

Theorem 3G.4 (implicit function theorem with strong metric regularity). For the
generalized equation f (p, x) + F(x)  0 with f : IRn → IRm and F : IRn →
→ IRm and its
solution mapping   
S : p → x  f (p, x) + F(x)  0 ,
consider a pair ( p̄, x̄) with x̄ ∈ S( p̄). Let h : IRn → IRm be a strict estimator of f
with respect to x uniformly in p at ( p̄, x̄) with constant μ and let h + F be strongly
metrically regular at x̄ for 0 with reg (h + F; x̄|0) ≤ κ. Suppose that

 ( f ; ( p̄, x̄)) ≤ λ < ∞.

κ μ < 1 and lip p

Then S has a Lipschitz continuous single-valued localization s around p̄ for x̄, more-
over with
κλ
lip (s; p̄) ≤ .
1 − κμ
Many corollaries of this theorem could be stated in a mode similar to that in
Section 3F, but the territory has already been covered essentially in Chapter 2. We
will get back to this result in Section 5F.
In some situations, metric regularity automatically entails strong metric regula-
rity. That is the case, for instance, for a linear mapping from IRn to itself represented
by an n × n matrix A. Such a mapping is metrically regular if and only if it is sur-
jective, which means that A has full rank, but then A is nonsingular, so that we
have strong metric regularity. More generally, for any mapping which describes the
Karush-Kuhn-Tucker optimality system in a nonlinear programming problem, met-
ric regularity implies strong metric regularity. We will prove this fact in Section
4F.
We will describe now another class of mappings for which metric regularity and
strong metric regularity come out to be the same thing. This class depends on a
localized, set-valued form of the monotonicity concept which appeared in Section
2F in the context of variational inequalities.
3 Regularity Properties of Set-valued Solution Mappings 181

Locally monotone mappings. A mapping F : IRn → → IRn is said to be locally mono-

tone at x̄ for ȳ if (x̄, ȳ) ∈ gph F and for some neighborhood W of (x̄, ȳ), one has

y − y, x − x ≥ 0 whenever (x , y ), (x, y) ∈ gph F ∩W.

Theorem 3G.5 (strong metric regularity of locally monotone mappings). If a map-

ping F : IRn →
→ IRn that is locally monotone at x̄ for ȳ is metrically regular at x̄ for ȳ,
then it must be strongly metrically regular at x̄ for ȳ.
Proof. According to 3G.1, all we need to show is that a mapping F which is locally
monotone and metrically regular at x̄ for ȳ must have a localization around ȳ for x̄
which is nowhere multi-valued. Suppose to the contrary that every graphical local-
ization of F −1 at ȳ for x̄ is multi-valued. Then there are infinite sequences yk → ȳ
and xk , zk ∈ F −1 (yk ), xk → x̄, zk → x̄ such that xk = zk for all k. Let bk = |zk − xk | > 0
and hk = (zk − xk )/bk . Then we have

(2) zk , hk  = bk + xk , hk  for all k = 1, 2, . . . .

Since the metric regularity of F implies through 3E.6 the Aubin property of F −1 at
ȳ for x̄, there exist κ > 0 and a > 0 such that

F −1 (y) ∩ IBa (x̄) ⊂ F −1 (y ) + κ |y − y|IB for all y, y ∈ IBa (ȳ).

Choose a sequence of positive numbers τk satisfying

(3) τk  0 and τk < bk /2κ .

Then for k large, we have yk , yk + τk hk ∈ IBa (ȳ) and xk ∈ F −1 (yk ) ∩ IBa (x̄), and hence
there exists uk ∈ F −1 (yk + τk hk ) satisfying

(4) |uk − xk | ≤ κτk .

By the local monotonicity of F at (x̄, ȳ) we have

uk − zk , yk + τk hk − yk  ≥ 0.

This, combined with (2), yields

(5) uk , hk  ≥ zk , hk  ≥ bk + xk , hk .

We get from (3), (4) and (5) that

bk + xk , hk  ≤ uk , hk  ≤ xk , hk  + κτk < xk , hk  + (bk /2),

which is impossible. Therefore, F −1 must indeed have a localization around ȳ for x̄

which is not multi-valued.
182 3 Regularity Properties of Set-valued Solution Mappings

3H. Calmness and Metric Subregularity

A “one-point” variant of the Aubin property can be defined for set-valued mappings
in the same way as calmness of functions, and this leads to another natural topic of
investigation.

Calmness. A mapping S : IRm → → IRn is said be calm at ȳ for x̄ if (ȳ, x̄) ∈ gph S and
there is a constant κ ≥ 0 along with neighborhoods U of x̄ and V of ȳ such that

(1) e(S(y) ∩U, S(ȳ)) ≤ κ |y − ȳ| for all y ∈ V.

Equivalently, the property in (1) can be also written as

(2) S(y) ∩U ⊂ S(ȳ) + κ |y − ȳ|IB for all y ∈ V

although perhaps with larger constant κ. The infimum of κ over all such combina-
tions of κ, U and V is called the calmness modulus of S at ȳ for x̄ and denoted by
clm (S; ȳ| x̄). The absence of this property is signaled by clm (S; ȳ| x̄) = ∞.
As in the case of the Lipschitz modulus lip (S; ȳ| x̄) in 3E, it is not claimed that (1)
and (2) are themselves equivalent, although that is true when S(y) is closed for every
y ∈ V . But anyway, the infimum furnishing clm (S; ȳ| x̄) is the same with respect to
(2) as with respect to (1).
In the case when S is not multi-valued, the definition above reduces to the def-
inition of calmness of a function in Section 1C relative to a neighborhood V of ȳ;
clm (S; ȳ, S(ȳ)) = clm (S; ȳ). Indeed, for any y ∈ V \ dom S the inequality (1) holds
automatically.
Clearly, for mappings with closed values, outer Lipschitz continuity implies
calmness. In particular, we get the following fact from Theorem 3D.1.

Proposition 3H.1 (calmness of polyhedral mappings). Any mapping S : IRm → → IRn

whose graph is the union of finitely many polyhedral convex sets is calm with the
same constant κ at any ȳ for any x̄ provided that (ȳ, x̄) ∈ gph S.
In particular, any linear mapping is calm at any point of its graph, and this is also
true for its inverse. For comparison, the inverse of a linear mapping has the Aubin
property at some point if and only if the mapping is surjective.

Exercise 3H.2 (local outer Lipschitz continuity under truncation). Show that a
mapping S : IRm → → IRn with (ȳ, x̄) ∈ gph S and with S(ȳ) convex is calm at ȳ for
x̄ if and only if there is a neighborhood U of x̄ such that the truncated mapping
y → S(y) ∩U is outer Lipschitz continuous at ȳ.
Guide. Mimic the proof of 3E.3 with y = ȳ.
Is there a “one-point” variant of the metric regularity which would characterize
calmness of the inverse, in the way metric regularity characterizes the Aubin prop-
erty of the inverse? Yes, as we explore next.
3 Regularity Properties of Set-valued Solution Mappings 183

Metric subregularity. A mapping F : IRn → → IRm is called metrically subregular at

x̄ for ȳ if (x̄, ȳ) ∈ gph F and there exists κ ≥ 0 along with neighborhoods U of x̄ and
V of ȳ such that

(3) d(x, F −1 (ȳ)) ≤ κ d(ȳ, F(x) ∩V ) for all x ∈ U.

The infimum of all κ for which this holds is the modulus of metric subregula-
rity, denoted by subreg (F; x̄| ȳ). The absence of metric subregularity is signaled
by subreg(F; x̄| ȳ) = ∞.
The main difference between metric subregularity and metric regularity is that
the data input ȳ is now fixed and not perturbed to a nearby y. Since d(ȳ, F(x)) ≤
κ d(ȳ, F(x) ∩ V ), it is clear that subregularity is a weaker condition than metric re-
gularity, and
subreg (F; x̄| ȳ) ≤ reg (F; x̄| ȳ).
The following result reveals the equivalence of metric subregularity of a mapping
with calmness of its inverse:

Theorem 3H.3 (characterization by inverse calmness). For a mapping F : IRn → → IRm ,

let F(x̄)  ȳ. Then F is metrically subregular at x̄ for ȳ with a constant κ if and only
if its inverse F −1 : IRm →
→ IRn is calm at ȳ for x̄ with the same constant κ, i.e., there
exist neighborhoods U of x̄ and V of ȳ such that

(4) F −1 (y) ∩U ⊂ F −1 (ȳ) + κ |y − ȳ|IB for all y ∈ V.

Moreover, clm (F −1 ; ȳ| x̄) = subreg(F; x̄| ȳ).

Proof. Assume first that (4) holds. Let x ∈ U. If F(x) = 0, / then the right side of (3)
is ∞ and we are done. If not, having x ∈ U and y ∈ F(x) ∩ V is the same as having
x ∈ F −1 (y) ∩U and y ∈ V . For such x and y, the inclusion in (4) requires the ball x +
κ |y− ȳ|IB to have nonempty intersection with F −1 (ȳ). Then −1
 d(x, F  (ȳ)) ≤ κ |y− ȳ|.

Thus, for any x ∈ U, we must have d(x, F (ȳ)) ≤ infy κ |y − ȳ|  y ∈ F(x) ∩ V ,
−1

which is (3). This shows that (4) implies (3) and that
     
inf κ  U, V, κ satisfying (4) ≥ inf κ  U, V, κ satisfying (3) ,

the latter being by definition subreg (F; x̄| ȳ).

For the opposite direction, we have to demonstrate that if subreg (F; x̄| ȳ) < κ <
∞, then (4) holds for some choice of neighborhoods U and V . Consider any κ  with
subreg(F; x̄| ȳ) < κ  < κ . For this κ  , there exist U and V such that d(x, F −1 (ȳ)) ≤
κ  d(ȳ, F(x) ∩ V ) for all x ∈ U. Then we have d(x, F −1 (ȳ)) ≤ κ  |y − ȳ| when x ∈ U
and y ∈ F(x) ∩V , or equivalently y ∈ V and x ∈ F −1 (y) ∩U. Fix y ∈ V . If y = ȳ there
is nothing to prove; let y = ȳ. If x ∈ F −1 (y) ∩ U, then d(x, F −1 (ȳ)) ≤ κ  |y − ȳ| <
κ |y − ȳ|. Then there must be a point of x ∈ F −1 (ȳ) having |x − x| ≤ κ |y − ȳ|. Hence
we have (4), as required.
As we will see next, there is no need at all to mention a neighborhood V of ȳ in
the description of calmness and subregularity in (2) and (3).
184 3 Regularity Properties of Set-valued Solution Mappings

Exercise 3H.4 (equivalent formulations). For a mapping F : IRn → → IRm and a point
(x̄, ȳ) ∈ gph F metric subregularity of F at x̄ for ȳ with constant κ is equivalent
simply to the existence of a neighborhood U of x̄ such that

(5) d(x, F −1 (ȳ)) ≤ κ d(ȳ, F(x)) for all x ∈ U,

whereas the calmness of F −1 at ȳ for x̄ with constant κ can be identified with the
existence of a neighborhood U of x̄ such that

(6) F −1 (y) ∩U ⊂ F −1 (ȳ) + κ |y − ȳ|IB for all y ∈ IRm .

Guide. Assume that (3) holds with κ > 0 and associated neighborhoods U and V .
We can choose within V a neighborhood of the form V  = IBε (ȳ) for some ε > 0.
Let U  := U ∩ (x̄ + εκ IB) and pick x ∈ U  . If F(x) ∩ V  = 0/ then d(ȳ, F(x) ∩ V  ) =
d(ȳ, F(x)) and (3) becomes (5) for this x. Otherwise, F(x) ∩V  = 0/ and then

1 1
d(ȳ, F(x)) ≥ ε ≥ |x − x̄| ≥ d(x, F −1 (ȳ)),
κ κ
which is (5).
Similarly, (6) entails the calmness in (4), so attention can be concentrated on
showing that we can pass from (4) to (6) under an adjustment in the size of U. We
already know from 3H.3 that the calmness condition in (4) leads to the metric sub-
regularity in (3), and further, from the argument just given, that such subregularity
yields the condition in (5). But that condition can be plugged into the argument in
the proof of 3H.3, by taking V = IRm , to get the corresponding calmness property
with V = IRm but with U replaced by a smaller neighborhood of x̄.
Although we could take (5) as a redefinition of metric subregularity, we prefer
to retain the neighborhood V in (3) in order to underscore the parallel with metric
regularity; similarly for calmness.
Does metric subregularity enjoy stability properties under perturbation resem-
bling those of metric regularity and strong metric regularity? In other words, does
metric subregularity obey the general paradigm of the implicit function theorem?
The answer to this question turns out to be no even for simple functions. Indeed, the
function f (x) = x2 is clearly not metrically subregular at 0 for 0, but its derivative
D f (0), which is the zero mapping, is metrically subregular.
More generally, every linear mapping A : IRn → IRm is metrically subregular, and
hence the derivative mapping of any smooth function is metrically subregular. But of
course, not every smooth function is subregular. For this reason, there cannot be an
implicit mapping theorem in the vein of 3F.8 in which metric regularity is replaced
by metric subregularity, even for the classical case of an equation with smooth f
and no set-valued F.
An illuminating but more intricate counterexample of instability of metric sub-
regularity of set-valued mappings is as follows. In IR × IR, let gph F be the set of
all (x, y) such that x ≥ 0, y ≥ 0 and yx = 0. Then F −1 (0) = [0, ∞) ⊃ F −1 (y) for all
3 Regularity Properties of Set-valued Solution Mappings 185

y, so F is metrically subregular at ȳ = 0 for x̄ = 0, even “globally” with κ = 0. By

Theorem 3H.3, subreg(F; 0|0) = 0.
Consider, however the function f (x) = −x2 for which f (0) = 0 and ∇ f (0) =
0. The perturbed mapping f + F has ( f + F) −1
−1 −1
 single-valued everywhere: ( f +
F) (y) = 0 when y ≥ 0, and ( f + F) (y) = |y| when y ≤ 0. This mapping is not
calm at 0 for 0. Then, from Theorem 3H.3 again, f + F is not metrically subregular;
we have subreg ( f + F; 0|0) = ∞.
To conclude this section, we point out some other properties which are, in a sense,
derived from metric regularity but, like subregularity, lack such kind of stability. One
of these properties is the openness which we introduced in Section 1F: a function
f : IRn → IRm is said to be open at x̄ ∈ dom f when for any a > 0 there exists b > 0
such that

(7) f (x̄ + a intIB) ⊃ f (x̄) + b intIB.

It turns out that this property likewise fails to be preserved when f is perturbed
to f + g by a function g with lip (g; x̄) = 0. This is demonstrated by the following
example. Define f : IR2 → IR2 by taking f (0, 0) = (0, 0) and, for x = (x1 , x2 ) = (0, 0),
 2 
1 x1 − x22
f (x1 , x2 ) = 2 .
2|x1 |x2
x21 + x22

Then f satisfies (7) at x̄ = 0 with b = a, since | f (x)| = |x|. The function g(x1 , x2 ) =
(0, x32 ) has g(0, 0) = (0, 0) and lip (g; (0, 0)) = 0, but ( f + g)−1 (c, 0) = 0/ when c < 0.
A “metric regularity variant” of the openness property (7), equally failing to be
preserved under small Lipschitz continuous perturbations, as shown by this same
example, is the requirement that d(x̄, F −1 (y)) ≤ κ |y − ȳ| for y close to ȳ. The same
trouble comes up also for an “inner semicontinuity variant,” namely the condition
that there exist neighborhoods U of x̄ and V of ȳ such that F −1 (y) ∩ U = 0/ for all
y ∈ V.
If we consider calmness as a local version of the outer Lipschitz continuity, then it
might seem to be worthwhile to define a local version of inner Lipschitz continuity,
introduced in Section 3D. For a mapping S : IRm → IRn with (ȳ, x̄) ∈ gph S, this would
refer to the existence of neighborhoods U of x̄ and V of ȳ such that

(8) S(ȳ) ∩U ⊂ S(y) + κ |y − ȳ|IB for all y ∈ V.

We will not give a name to this property here, or a name to the associated property
of the inverse of a mapping satisfying (8). We will only demonstrate, by an example,
that the property of the inverse associated to (8), similar to metric subregularity, is
not stable under perturbation, in the sense we have been exploring, and hence does
not support the implicit function theorem paradigm.
→
Consider
 the mapping S : IR → IR whose values are the set of three points
{− |y|, 0, |y|} for all y ≥ 0 and the empty set for y < 0. This mapping has the
property in (8) at ȳ = 0 for x̄ = 0. Now consider the inverse S−1 and add to it the
186 3 Regularity Properties of Set-valued Solution Mappings

function g(x) = −x2 , which has zero derivative at x̄ = 0. The sum S−1 + g is the
mapping whose value at x = 0 is the interval [0, ∞) but is just zero for x = 0. The
inverse (S−1 + g)−1 has (−∞, ∞) as its value for y = 0, but 0 for y > 0 and the empty
set for y < 0. Clearly, this inverse does not have the property displayed in (8) at ȳ = 0
for x̄ = 0.
In should be noted that for special cases of mappings with particular perturba-
tions one might still obtain stability of metric subregularity, or the property associ-
ated to (8), but we shall not go into this further.

3I. Strong Metric Subregularity

The handicap of serious instability of calmness and metric subregularity can be
obviated by passing to strengthened forms of these properties.

Isolated calmness. A mapping S : IRm → → IRn is said to have the isolated calmness
property if it is calm at ȳ for x̄ and, in addition, S has a graphical localization at ȳ
for x̄ that is single-valued at ȳ itself (with value x̄). Specifically, this refers to the
existence of a constant κ ≥ 0 and neighborhoods U of x̄ and V of ȳ such that

(1) |x − x̄| ≤ κ |y − ȳ| when x ∈ S(y) ∩U and y ∈ V.

Observe that in this definition S(ȳ) ∩U is a singleton, namely the point x̄, so x̄ is
an isolated point in S(ȳ), hence the terminology. Isolated calmness can equivalently
be defined as the existence of a (possibly slightly larger) constant κ and neighbor-
hoods U of x̄ and V of ȳ such that

(2) S(y) ∩U ⊂ x̄ + κ |y − ȳ|IB when y ∈ V.

For a linear mapping A, isolated calmness holds at every point, whereas isolated
calmness of A−1 holds at some point of dom A−1 if and only if A is nonsingular.
More generally we have the following fact through Theorem 3D.1 for polyhedral
mappings, as defined there.

Proposition 3I.1 (isolated calmness of polyhedral mappings). A polyhedral map-

ping S : IRm →
→ IRn has the isolated calmness property at ȳ for x̄ if and only if x̄ is an
isolated point of S(ȳ).
Once again we can ask whether there is a property of a mapping that corresponds
to isolated calmness of its inverse. Such a property exists and, in order to unify the
terminology, we name it as follows.

Strong metric subregularity. A mapping F : IRn → → IRm is said to be strongly met-

rically subregular at x̄ for ȳ if (x̄, ȳ) ∈ gph F and there is a constant κ ≥ 0 along with
neighborhoods U of x̄ and V of ȳ such that
3 Regularity Properties of Set-valued Solution Mappings 187

(3) |x − x̄| ≤ κ d(ȳ, F(x) ∩V ) for all x ∈ U.

Equivalently, F is strongly metrically subregular at x̄ for ȳ if it is metrically sub-

regular at x̄ for ȳ and there exist neighborhoods U of x̄ and V of ȳ such that the
graphical localization V  y → F −1 (y)∩U is single-valued at ȳ with value x̄. Clearly,
(3) implies that x̄ is an isolated point of F −1 (ȳ) and that F is metrically subregular
there; hence the infimum of κ for which (3) holds is equal to subreg (F; x̄| ȳ).
Note however that, in general, the condition subreg (F; x̄| ȳ) < ∞ is not a char-
acterization of strong metric subregularity, but becomes such a criterion under the
isolatedness assumption. As an example, observe that, for a linear mapping A is al-
ways metrically subregular at 0 for 0, but it is strongly metrically subregular at 0
for 0 if and only if ker A consists of just 0, which corresponds to A being injective.
The equivalence of strong metric subregularity and isolated calmness of the inverse
is shown next:

Theorem 3I.2 (characterization by inverse isolated calmness). A mapping F :

IRn →→ IRm is strongly metrically subregular at x̄ for ȳ with constant κ if and only
if its inverse F −1 has the isolated calmness property at ȳ for x̄ with the same con-
stant κ, i.e., there exist neighborhoods U of x̄ and V of ȳ such that

(4) F −1 (y) ∩U ⊂ x̄ + κ |y − ȳ|IB when y ∈ V.

Then infimum of all κ such that the inclusion holds for some U and V equals
subreg(F; x̄| ȳ).
Proof. Assume first that F is strongly subregular at x̄ for ȳ. Let κ > subreg (F; x̄| ȳ).
Then there are neighborhoods U for x̄ and V for ȳ such that (3) holds with the
indicated κ . Consider any y ∈ V . If F −1 (y) ∩U = 0, / then (4) holds trivially. If not,
let x ∈ F −1 (y) ∩ U. This entails y ∈ F(x) ∩ V , hence d(ȳ, F(x) ∩ V ) ≤ |y − ȳ| and
consequently |x − x̄| ≤ κ |y − ȳ| by (3). Thus, x ∈ x̄ + κ |y − ȳ|IB, and we conclude
that (4) holds. Also, we see that subreg(F; x̄| ȳ) is not less than the infimum of all κ
such that (4) holds for some choice of U and V .
For the converse, suppose (4) holds for some κ and neighborhoods U and V .
Consider any x ∈ U. If F(x) ∩ V = 0/ the right side of (3) is ∞ and there is nothing
more to prove. If not, for an arbitrary y ∈ F(x) ∩ V we have x ∈ F −1 (y) ∩ U, and
therefore x ∈ x̄ + κ |y − ȳ|IB by (4), which means |x − x̄| ≤ κ |y − ȳ|. This being true
for all y ∈ F(x) ∩ V , we must have |x − x̄| ≤ κ d(ȳ, F(x) ∩ V ). Thus, (3) holds, and
in particular we have κ ≥ subreg(F; x̄| ȳ). Therefore, the infimum of κ in (4) equals
subreg(F; x̄| ȳ).
Observe also, through 3H.4, that the neighborhood V in (2) and (3) can be chosen
to be the entire space IRm , by adjusting the size of U; that is, strong metric subregu-
larity as in (3) with constant κ is equivalent to the existence of a neighborhood U 
of x̄ such that

(5) |x − x̄| ≤ κ d(ȳ, F(x)) for all x ∈ U  .

188 3 Regularity Properties of Set-valued Solution Mappings

Accordingly, the associated isolated calmness of the inverse is equivalent to the

existence of a neighborhood U  of x̄ such that

(6) F −1 (y) ∩U  ⊂ x̄ + κ |y − ȳ|IB when y ∈ IRm .

Exercise 3I.3. Provide direct proofs of the equivalence of (3) and (5), and (4) and
(6), respectively.
Guide. Use the argument in the proof of 3H.4.
Similarly to the distance function characterization in Theorem 3E.8 for the Aubin
property, the isolated calmness property is characterized by uniform calmness of the
distance function associated with the inverse mapping:

Theorem 3I.4 (distance function characterization of strong metric subregularity).

For a mapping F : IRn →
→ IRm and a point (x̄, ȳ) ∈ gph F , suppose that x̄ is an isolated
−1
point in F (ȳ) and moreover

(7) x̄ ∈ lim inf F −1 (y).

y→ȳ

Consider the function s(y, x) = d(x, F −1 (y)). Then the mapping F is strongly met-
rically subregular at x̄ for ȳ if and only if s is calm with respect to y uniformly in x
at (ȳ, x̄), in which case

 y (s; (ȳ, x̄)) = subreg(F; x̄| ȳ).

clm

Proof. Let F be strongly metrically subregular at x̄ for ȳ and let κ > subreg (F; x̄| ȳ).
Let (5) and (6) hold with U  = IBa (x̄) and also F −1 (ȳ) ∩ IBa (x̄) = x̄. Let b > 0 be
such that, according to (7), F −1 (y) ∩ IBa (x̄) = 0/ for all y ∈ IBb (ȳ). Make b smaller
if necessary so that b ≤ a/(10κ ). Choose y ∈ IBb (ȳ) and x ∈ IBa/4 (x̄); then from (6)
we have

(8) d(x̄, F −1 (y) ∩ IBa (x̄)) ≤ κ |y − ȳ|.

Since all points in F −1 (ȳ) except x̄ are at distance from x more than a/4 we obtain

(9) d(x, F −1 (ȳ)) = |x − x̄|.

Using the triangle inequality and (8), we get

d(x, F −1 (y)) ≤ |x − x̄| + d(x̄, F −1 (y))

(10)
= |x − x̄| + d(x̄, F −1 (y) ∩ IBa (x̄)) ≤ |x − x̄| + κ |y − ȳ|.

Then, taking (9) into account, we have

(11) s(y, x) − s(ȳ, x) = d(x, F −1 (y)) − d(x, F −1 (ȳ)) ≤ κ |y − ȳ|.

3 Regularity Properties of Set-valued Solution Mappings 189

Let x̃ be a projection of x on cl F −1 (y). Using (10) we obtain

|x − x̃| = d(x, F −1 (y)) ≤ |x − x̄| + κ |y − ȳ| ≤ a/4 + κ b ≤ a/4 + κ a/(10κ ) ≤ a/2,

and consequently

|x̄ − x̃| ≤ |x̄ − x| + |x − x̃| ≤ a/4 + a/2 = 3a/4 < a.

Therefore,

(12) d(x, F −1 (y) ∩ IBa (x̄)) = d(x, F −1 (y)).

According to (6),

d(x, x̄ + κ |y − ȳ|IB) ≤ d(x, F −1 (y) ∩ IBa (x̄))

and then, by (12)

(13) |x − x̄| − κ |y − ȳ| ≤ d(x, F −1 (y) ∩ IBa (x̄)) = d(x, F −1 (y)).

Plugging (9) into (13), we conclude that

(14) s(ȳ, x) − s(y, x) = d(x, F −1 (ȳ)) − d(x, F −1 (y)) ≤ κ |y − ȳ|.

Since x and y were arbitrarily chosen in dom s and close to x̄ and ȳ, respectively, we
 y (s; (ȳ, x̄)) ≤ κ , hence
obtain by combining (11) and (14) that clm

(15)  y (s; (ȳ, x̄)) ≤ subreg(F; x̄| ȳ).

clm

To show the converse inequality, let κ > clm  y (s; (ȳ, x̄)); then there exists a > 0
such that s(·, x) is calm on IBa (ȳ) with constant κ uniformly in x ∈ IBa (x̄). Adjust a so
that F −1 (ȳ) ∩ IBa (x̄) = x̄. Pick any x ∈ IBa/3 (x̄). If F(x) = 0, / (5) holds automatically.
If not, choose any y ∈ IRn such that (x, y) ∈ gph F. Since s(y, x) = 0, we have

|x − x̄| = d(x, F −1 (ȳ)) = s(ȳ, x) ≤ s(y, x) + κ |y − ȳ| = κ |y − ȳ|.

Since y is arbitrarily chosen in F(x), this gives us (5). This means that F is strongly
subregular at x̄ for ȳ with constant κ and hence

 y (s; (ȳ, x̄)) ≥ subreg(F; x̄| ȳ).

clm

Combining this with (15) brings the proof to a finish.

Exercise 3I.5 (counterexample). Show that the mapping F : IR →

→ IR given by
%
x if 0 ≤ x < 1,
F(x) = IR if x ≥ 1,
0/ if x < 0
190 3 Regularity Properties of Set-valued Solution Mappings

 y (s; (0, 0)) = ∞.

does not satisfy condition (7) and has subreg (F; 0|0) = 1 while clm

We look next at perturbations of F by single-valued mappings g in the pattern

that was followed for the other regularity properties considered in the preceding
sections.

Theorem 3I.6 (inverse mapping theorem for strong metric subregularity). Consider
a mapping F : IRn →→ IRm and a point (x̄, ȳ) ∈ gph F such that F is strongly metrically
subregular at x̄ for ȳ and let κ and μ be nonnegative constants such that

subreg(F; x̄| ȳ) ≤ κ and κ μ < 1.

Then for any function g : IRn → IRm with x̄ ∈ dom g and clm (g; x̄) ≤ μ, one has
κ
subreg(g + F; x̄|g(x̄) + ȳ) ≤ .
1 − κμ

Proof. Choose κ and μ as in the statement of the theorem and let λ > κ , ν > μ be
such that λ ν < 1. Pick g : IRn → IRm with clm (g; x̄) < ν . Without loss of generality,
let g(x̄) = 0; then there exists a > 0 such that

(16) |g(x)| ≤ ν |x − x̄| when x ∈ IBa (x̄).

Since subreg (F; x̄| ȳ) < λ , we can arrange, by taking a smaller if necessary, that

(17) |x − x̄| ≤ λ |y − ȳ| when (x, y) ∈ gph F ∩ (IBa (x̄) × IBa (ȳ)).

Let ν  = max{1, ν } and consider any

(18) z ∈ IBa/2 (ȳ) with x ∈ (g + F)−1 (z) ∩ IBa/2ν  (x̄).

These relations entail z ∈ g(x) + F(x), hence z = y + g(x) for some y ∈ F(x). From
(16) and since x ∈ IBa/2ν  (x̄), we have |g(x)| ≤ μ (a/2ν  ) ≤ a/2 (inasmuch as ν  ≥
ν ). Using the equality y − ȳ = z − g(x) − ȳ we get |y − ȳ| ≤ |z − ȳ| + |g(x)| ≤ (a/2) +
(a/2) = a. However, because (x, y) ∈ gph F ∩ (IBa (x̄) × IBa (ȳ)), through (17),

|x − x̄| ≤ λ |(z − g(x)) − ȳ| ≤ λ |z − ȳ| + λ |g(x)| ≤ λ |z − ȳ| + λ ν |x − x̄|,

hence |x − x̄| ≤ λ /(1 − λ ν )|z − ȳ|. Since x and z are chosen as in (18) and λ and ν
could be arbitrarily close to κ and μ , respectively, the proof is complete.
Corollaries that parallel those for metric regularity given in Section 3F can im-
mediately be derived.

Corollary 3I.7 (detailed estimate). Consider a mapping F : IRn → → IRm which is

strongly metrically subregular at x̄ for ȳ and a function g : IR → IRm such that
n

subreg(F; x̄| ȳ) · clm (g; x̄) < 1. Then the mapping g + F is strongly metrically sub-
regular at x̄ for ȳ, and one has
3 Regularity Properties of Set-valued Solution Mappings 191
 −1
subreg (g + F; x̄|g(x̄) + ȳ) ≤ subreg(F; x̄| ȳ)−1 − clm (g; x̄) .

If subreg(F; x̄| ȳ) = 0, then subreg(g + F; x̄| ȳ) = 0 for any g : IRn → IRm with
clm (g; x̄) < ∞. If subreg(F; x̄| ȳ) = ∞, then subreg(g + F; x̄|g(x̄) + ȳ) = ∞ for any
g : IRn → IRm with clm (g; x̄) = 0.
This result implies in particular that the property of strong metric subregularity is
preserved under perturbations with zero calmness moduli. The only difference with
the corresponding results for metric regularity in Section 3F is that now a larger
class of perturbation is allowed with first-order approximations replacing the strict
first-order approximations.

Corollary 3I.8 (utilizing first-order approximations). Consider F : IRn → → IRm , a

point (x̄, ȳ) ∈ gph F and two functions f : IR → IR and g : IR → IRm with
n m n

x̄ ∈ int dom f ∩ int dom g which are first-order approximations to each other at x̄.
Then the mapping f + F is strongly metrically subregular at x̄ for f (x̄) + ȳ if and
only if g + F is strongly metrically subregular at x̄ for g(x̄) + ȳ, in which case

subreg( f + F; x̄| f (x̄) + ȳ) = subreg (g + F; x̄|g(x̄) + ȳ).

This corollary takes a more concrete form when the first-order approximation is
represented by a linearization:

Corollary 3I.9 (linearization). Let M = f + F for mappings f : IRn → IRm and F :

IRn →
→ IRm , and let ȳ ∈ M(x̄). Suppose f is differentiable at x̄, and let

M0 = h + F for h(x) = f (x̄) + ∇ f (x̄)(x − x̄).

Then M is strongly metrically subregular at x̄ for ȳ if and only if M0 has this property.
Moreover subreg(M; x̄ | ȳ) = subreg(M0 ; x̄| ȳ).
Through 3I.1, the result in Corollary 3I.9 could equally well be stated in terms of
the isolated calmness property of M−1 in relation to that of M0−1 . We can specialize
that result in the following way.

Corollary 3I.10 (linearization with polyhedrality). Let M : IRn → → IRm with ȳ ∈ M(x̄)
be of the form M = f + F for f : IR → IR and F : IR →
n m n → IRm such that f is differen-
tiable at x̄ and F is polyhedral. Let M0 (x) = f (x̄) + ∇ f (x̄)(x − x̄) + F(x). Then M−1
has the isolated calmness property at ȳ for x̄ if and only if x̄ is an isolated point of
M0−1 (ȳ).
Proof. This applies 3I.1 in the framework of the isolated calmness restatement of
3I.9 in terms of the inverses.
Applying Corollary 3I.9 to the case where F is the zero mapping, we obtain yet
another inverse function theorem in the classical setting:

Corollary 3I.11 (an inverse function result). Let f : IRn → IRm be differentiable at
x̄ and such that ker ∇ f (x̄) = {0}. Then there exist κ > 0 and a neighborhood U of
192 3 Regularity Properties of Set-valued Solution Mappings

x̄ such that
|x − x̄| ≤ κ | f (x) − f (x̄)| for every x ∈ U.
Proof. This comes from (5).
Next, we state and prove an implicit function theorem for strong metric subregu-
larity:

Theorem 3I.12 (implicit mapping theorem with strong metric subregularity). For
the generalized equation f (p, x) + F(x)  0 and its solution mapping
  
S : p → x  f (p, x) + F(x)  0 ,

consider a pair ( p̄, x̄) with x̄ ∈ S( p̄). Let h : IRn → IRm be an estimator of f with
respect to x at ( p̄, x̄) with constant μ and let h + F be strongly metrically subregular
at x̄ for 0 with subreg(h + F; x̄|0) ≤ κ. Suppose that

(19)  p ( f ; ( p̄, x̄)) ≤ λ < ∞.

κ μ < 1 and clm

Then S has the isolated calmness property at p̄ for x̄, moreover with

κλ
clm (S; p̄| x̄) ≤ .
1 − κμ

Proof. The proof goes along the lines of the proof of Theorem 3I.6 with different
choice of constants. Let κ , μ and λ be as required and let δ > κ and ν > μ be such
that δ ν < 1. Let γ > λ . By the assumptions for the mapping h + F and the functions
f and h, there exist positive scalars a and r such that

(20) |x − x̄| ≤ δ |y| for all x ∈ (h + F)−1 (y) ∩ IBa (x̄) and y ∈ IBν a+γ r (0),

(21) | f (p, x) − f ( p̄, x)| ≤ γ |p − p̄| for all p ∈ IBr ( p̄) and x ∈ IBa (x̄),

and also, for e = f − h,

(22) |e(p, x) − e(p, x̄)| ≤ ν |x − x̄| for all x ∈ IBa (x̄) and p ∈ IBr ( p̄).

Let x ∈ S(p) ∩ IBa (x̄) for some p ∈ IBr ( p̄). Then, since h(x̄) = f ( p̄, x̄), we obtain
from (21) and (22) that

|e(p, x)| ≤ |e(p, x) − e(p, x̄)| + | f (p, x̄) − f ( p̄, x̄)|

(23)
≤ ν |x − x̄| + γ |p − p̄| ≤ ν a + γ r.

Observe that x ∈ (h + F)−1 (− f (p, x) + h(x)) ∩ IBa (x̄), and then from (20) and (23)
we have
|x − x̄| ≤ δ | − f (p, x) + h(x)| ≤ δ ν |x − x̄| + δ γ |p − p̄|.
In consequence,
3 Regularity Properties of Set-valued Solution Mappings 193

δγ
|x − x̄| ≤ |p − p̄|.
1−δν
Since δ is arbitrarily close to κ , ν is arbitrarily close to μ and γ is arbitrarily close
to λ , we arrive at the desired result.
In the theorem we state next, we can get away with a property of f at ( p̄, x̄)
which is weaker than local continuous differentiability, namely a kind of uniform
differentiability we introduced in Section 1C. We say that f (p, x) is differentiable
in x uniformly with respect to p at ( p̄, x̄) if f is differentiable with respect to (p, x)
at ( p̄, x̄) and for every ε > 0 there is a (p, x)-neighborhood of ( p̄, x̄) in which

| f (p, x) − f (p, x̄) − ∇x f ( p̄, x̄)(x − x̄)| ≤ ε |x − x̄|.

Symmetrically, we define what it means for f (p, x) to be differentiable in p uni-

formly with respect to x at ( p̄, x̄). Note that the combination of these two properties
is implied by, yet weaker than, the continuous differentiability of f at ( p̄, x̄). For
instance, the two uniformity properties hold when f (p, x) = f1 (p) + f2 (x) and we
simply have f1 differentiable at p̄ and f2 differentiable at x̄.

Theorem 3I.13 (utilizing differentiability and ample parameterization). For the

generalized equation in Theorem 3I.12 and its solution mapping S, and a pair ( p̄, x̄)
with x̄ ∈ S( p̄), suppose that f is differentiable in x uniformly with respect to p at
( p̄, x̄), and at the same time differentiable in p uniformly with respect to x at ( p̄, x̄).
If the mapping
h + F for h(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄)
is strongly metrically subregular at x̄ for 0, then S has the isolated calmness property
at p̄ for x̄ with

(24) clm (S; p̄| x̄) ≤ subreg (h + F; x̄|0) · |∇ p f ( p̄, x̄)|.

Furthermore, when f is continuously differentiable on a neighborhood of ( p̄, x̄) and

satisfies the ample parameterization condition

rank ∇ p f ( p̄, x̄) = m,

then the converse implication holds as well: the mapping h + F is strongly metrically
subregular at x̄ for 0 provided that S has the isolated calmness property at p̄ for x̄.
Proof. With this choice of h, the assumption (19) of 3I.12 holds and then (24) fol-
lows from the conclusion of this theorem. To handle the ample parameterization we
employ Lemma 2C.1 by repeating the argument in the proof of 3F.10, simply re-
placing the composition rule there with the one in the following proposition.

Proposition 3I.14 (isolated calmness in composition). For a mapping M : IRd →

→ IRn
m d m →
and a function ψ : IR × IR → IR consider the composite mapping N : IR → IRn of
n

the form
194 3 Regularity Properties of Set-valued Solution Mappings
  
y → N(y) = x  x ∈ M(ψ (x, y)) for y ∈ IRm .
Let ψ satisfy

(25)  x (ψ ; (x̄, 0)) = 0 and clm

clm  y (ψ ; (x̄, 0)) < ∞,

and let (ψ (x̄, 0), x̄) ∈ gph M . If M has the isolated calmness property at ψ (x̄, 0) for
x̄, then N has the isolated calmness property at 0 for x̄.
Proof. Let M have the isolated calmness property with neighborhoods IBb (x̄), IBc ( p̄)
and constant κ > clm (M; p̄ | x̄), where p̄ = ψ (x̄, 0). Choose λ > 0 with λ < 1/κ and
a > 0 such that for any y ∈ IBa (0) the function ψ (·, y) is calm on IBb (x̄) with calmness
 y (ψ ; (x̄, 0)) and make a and b smaller if necessary so that
constant λ . Pick γ > clm
the function ψ (x, ·) is calm on IBa (0) with constant γ and also

(26) λ b + γ a ≤ c.

Let y ∈ IBa (0) and x ∈ N(y) ∩ IBb (x̄). Then x ∈ M(ψ (x, y)) ∩ IBb (x̄). Using the as-
sumed calmness properties (25) of ψ and utilizing (26) we see that

|ψ (x, y) − p̄| = |ψ (x, y) − ψ (x̄, 0)| ≤ λ b + γ a ≤ c.

From the isolated calmness of M we then have

|x − x̄| ≤ κ |ψ (x, y) − ψ (x̄, 0)| ≤ κλ |x − x̄| + κγ |y|,

hence
κγ
|x − x̄| ≤|y|.
1 − κλ
This establishes that the mapping N has the isolated calmness property at 0 for x̄
with constant κγ /(1 − κλ ).

Example 3I.15 (application to complementarity problems). For f : IRd × IRn → IRn ,

consider the complementarity problem of finding for given p an x such that

(27) x ≥ 0, f (p, x) ≥ 0, x ⊥ f (p, x).

This corresponds to solving f (p, x) + NIRn+  0, as seen in 2A. Let x̄ be a solution

for p̄, and suppose that f is continuously differentiable in a neighborhood of ( p̄, x̄).
Consider now the linearized problem

(28) x ≥ 0, Ax + y ≥ 0, x ⊥ Ax + y, with A = ∇x f ( p̄, x̄).

Then, from 3I.13 we obtain that if the solution mapping for (28) has the isolated
calmness property at ȳ = f ( p̄, x̄) − Ax̄ for x̄, then the solution mapping for (27)
has the isolated calmness property at p̄ for x̄. Under the ample parameterization
condition, rank ∇ p f ( p̄, x̄) = n, the converse implication holds as well.
3 Regularity Properties of Set-valued Solution Mappings 195

Commentary
The inner and outer limits of sequences of sets were introduced by Painlevé in
his lecture notes as early as 1902 and later popularized by Hausdorff [1927] and
Kuratowski [1933]. The definition of excess was first given by Pompeiu [1905],
who also defined the distance between sets C and D as e(C, D) + e(D,C). Hausdorff
[1927] gave the definition we use here. These two definitions are equivalent in the
sense that they induce the same convergence of sets. The reader can find much more
about set-convergence and continuity properties of set-valued mappings together
with extended historical commentary in Rockafellar and Wets [1998]. This includes
the reason why we prefer “inner and outer” in contrast to the more common terms
“lower and upper,” so as to avoid certain conflicts in definition that unfortunately
pervade the literature.
Theorem 3B.4 is a particular case of a result sometimes referred to as the Berge
theorem; see Section 8.1 in Dontchev and Zolezzi [1993] for a general statement.
Theorem 3C.3 comes from Walkup and Wets [1969], while the Hoffman lemma,
3C.4, is due to Hoffman [1952].
The concept of outer Lipschitz continuity was introduced by Robinson [1981]
under the name “upper Lipschitz continuity” and adjusted to “outer Lipschitz con-
tinuity” later in Robinson [2007]. Theorem 3D.1 is due to Robinson [1981] while
3D.3 is a version, given in Robinson [2007], of a result due to Wu Li [1994].
The Aubin property of set-valued mappings was introduced by J.-P. Aubin
[1984], who called it “pseudo-Lipschitz continuity”; it was renamed after Aubin in
Dontchev and Rockafellar [1996]. In the literature one can also find it termed “Aubin
continuity,” but we do not use that here since the Aubin property does not imply con-
tinuity. Theorem 3E.3 is from Bessis, Ledyaev and Vinter [2001]. The name “metric
regularity” was coined by J. M. Borwein [1986a], but the origins of this concept go
back to the Banach open mapping theorem and even earlier. In the literature, metric
regularity is defined in various ways, for example in Schirotzek [2007] the property
expressed in 3E.6 is called weak metric regularity; see e.g. Mordukhovich [2006]
for other names. Theorem 3E.4 is from Rockafellar [1985]. Theorem 3E.9 comes
from Ledyaev and Zhu [1999]. For historical remarks regarding inverse and implicit
mapping theorems with metric regularity, see the commentary to Chapter 5.
As we mentioned earlier in Chapter 2, the term “strong regularity” comes from
Robinson [1980], who used it in the framework of variational inequalities. Theorem
3F.5 is a particular case of a more general result due to Kenderov [1975]; see also
Levy and Poliquin [1997].
Calmness and metric subregularity, as well as isolated calmness and metric sub-
regularity, have been considered in various contexts and under various names in
the literature; here we follow the terminology of Dontchev and Rockafellar [2004].
Isolated calmness was formally introduced in Dontchev [1995a], where its stability
(Theorem 3I.6) was first proved. The equivalent property of strong metric subregu-
larity was considered earlier, without giving it a name, by Rockafellar [1989]; see
also the commentary to Chapter 4.
Chapter 4
Regularity Properties Through Generalized
Derivatives

In the wide-ranging generalizations we have been developing of the inverse function

theorem and implicit function theorem, we have followed the idea that conclusions
about a solution mapping, concerning the Aubin property, say, or the existence of
a single-valued localization, can be drawn by confirming that some auxiliary solu-
tion mapping, obtained from a kind of approximation, has the property in question.
In the classical framework, we can appeal to a condition like the invertibility of a
Jacobian matrix and thus tie in with standard calculus. Now, though, we are far away
in another world where even a concept of differentiability seems to be lacking. How-
ever, substitutes for classical differentiability can very well be introduced and put to
work. In this chapter we show the way to that and explain numerous consequences.
First, graphical differentiation of a set-valued mapping is defined through the va-
riational geometry of the mapping’s graph. A characterization of the Aubin property
is derived and applied to the case of a solution mapping. Strong metric subregula-
rity is characterized next. Applications are made to parameterized constraint sys-
tems and special features of solution mappings for variational inequalities. There is
a review then of some other derivative concepts and the associated inverse function
theorems of Clarke and Kummer. Finally, alternative results using coderivatives are
described.

A.L. Dontchev and R.T. Rockafellar, Implicit Functions and Solution Mappings: A View 197
from Variational Analysis, Springer Monographs in Mathematics,
DOI 10.1007/978-0-387-87821-8 4,  c Springer Science+Business Media, LLC 2009
198 4 Regularity Properties Through Generalized Derivatives

4A. Graphical Differentiation

The concept of the tangent cone TC (x) to a set C in IRn at a point x ∈ C was introduced
in 2A, but it was only utilized there in the case of C being closed and convex. In 2E,
the geometry of tangent cones to polyhedral convex sets received special attention
and led to significant insights in the study of variational inequalities. Now, tangent
cones to possibly nonconvex sets will come strongly onto the stage as well, serving
as a tool for a kind of generalized differentiation. The definition of the tangent cone
is the same as before.

Tangent cones. A vector v ∈ IRn is said to be tangent to a set C ⊂ IRn at a point

x ∈ C if
1 k
(x − x) → v for some xk → x, xk ∈ C, τ k  0.
τk
The set of all such vectors v is called the tangent cone to C at x and is denoted TC (x).
The tangent cone mapping is defined as

TC : x → TC (x) for x ∈ C,
0/ otherwise.
A description equivalent to this definition is that v ∈ TC (x) if and only if there are
sequences vk → v and τ k  0 with x+ τ k vk ∈ C, or equivalently, if there are sequences
xk ∈ C, xk → x and τk  0 such that vk := (xk − x)/τk → v as k → ∞.
Note that TC (x) is indeed a cone: it contains v = 0 (as seen from taking xk ≡ x),
and contains along with any vector v all positive multiples of v. The definition can
also be recast in the notation of set convergence:

(1) TC (x) = lim sup τ −1 (C − x).

τ 0

Described as an outer limit in this way, it is clear in particular that TC (x) is always
a closed set. When C is a “smooth manifold” in IRn , TC (x) is the usual tangent
subspace, but in general, of course, TC (x) need not even be convex. The tangent
cone mapping TC has dom TC = C but gph TC is not necessarily a closed subset of
IRn × IRn even when C is closed.
As noted in 2A.4, when the set C is convex, the tangent cone TC (x) is also convex
for any x ∈ C. In this case the limsup in (1) can be replaced by lim, as shown in the
following proposition.

Proposition 4A.1 (tangent cones to convex sets). For a convex set C ⊂ IRn and a
point x ∈ C,

(2) TC (x) = lim τ −1 (C − x).

τ 0

Proof. Consider the set

4 Regularity Properties Through Generalized Derivatives 199
  
KC (x) = v  ∃ τ > 0 with x + τ v ∈ C .
Let v ∈ TC (x). Then there exist sequences τk  0 and xk ∈ C, xk → x, such that vk :=
(xk − x)/τk → v. Hence vk ∈ KC for all k and therefore v ∈ cl K. Thus, we obtain

(3) TC (x) ⊂ cl KC (x).

Now let v ∈ KC (x). Then v = (x̃ − x)/τ for some τ > 0 and x̃ ∈ C. Take an arbitrary
sequence τk  0 as k → ∞. Since C is convex, we have
τk τk
x + τk v = (1 − )x + x̃ ∈ C for all k.
τ τ
But then v ∈ (C − x)/τk for all k and hence v ∈ lim infk τk−1 (C − x). Since τk was
arbitrarily chosen, we conclude that

KC (x) ⊂ lim inf τ −1 (C − x) ⊂ lim sup τ −1 (C − x) = TC (x).

τ 0 τ 0

This combined with (3) gives us (2).

We should note that, in order to have the equality (2), the set C does not need to
be convex. Generally, sets C for which (2) is satisfied are called geometrically deriv-
able. Proposition 4A.1 simply says that all convex sets are geometrically derivable.
Starting in elementary calculus, students are taught to view differentiation in
terms of tangents to the graph of a function. This can be formulated in the notation
of tangent cones as follows. Let f : IRn → IRm be a function which is differentiable
at x with derivative mapping D f (x) : IRn → IRm . Then

(u, v) ∈ gph D f (x) ⇐⇒ (u, v) ∈ Tgph f (x, f (x)).

In other words, the derivative is completely represented geometrically by the tangent

cone to the set gph f at the point (x, f (x)). In fact, differentiability is more or less
equivalent to having Tgph f (x, f (x)) turn out to be the graph of a linear mapping.
By adopting such a geometric characterization as a definition, while not insisting
on linearity, we can introduce derivatives for an arbitrary set-valued mapping F :
IRn →
→ IRm . However, because F(x) may have more than one element y, it is essential
for the derivative mapping to depend not just on x but also on a choice of y ∈ F(x).

Graphical derivatives. For a mapping F : IRn → → IRm and a pair (x, y) with y ∈ F(x),
the graphical derivative of F at x for y is the mapping DF(x|y) : IRn → → IRm whose
graph is the tangent cone Tgph F (x, y) to gph F at (x, y):

v ∈ DF(x|y)(u) ⇐⇒ (u, v) ∈ Tgph F (x, y).

Thus, v ∈ DF(x|y)(u) if and only if there exist sequences uk → u, vk → v and τ k  0

such that y + τ k vk ∈ F(x + τ k uk ) for all k.
200 4 Regularity Properties Through Generalized Derivatives

On this level, derivative mappings may no longer even be single-valued. But

because their graphs are cones, they do always belong to the following class of
mappings, at least.

Positively homogeneous mappings. A mapping H : IRn → → IRm is called positively

homogeneous when gph H is a cone, which is equivalent to H satisfying

0 ∈ H(0) and H(λ x) = λ H(x) for λ > 0.

Clearly, the inverse of a positively homogeneous mapping is another positively

homogeneous mapping. Linear mappings are positively homogeneous as a special
case, their graphs being not just cones but linear subspaces.
Since the graphical differentiation comes from an operation on graphs, and the
graph of a mapping F can be converted to the graph of its inverse F −1 just by
interchanging variables, we immediately have the rule that

D(F −1 )(y|x) = DF(x|y)−1 .

Another useful relation is available for sums.

Proposition 4A.2 (sum rule). For a function f : IRn → IRm which is differentiable at
x, a set-valued mapping F : IRn →
→ IRm and any y ∈ F(x), one has

D( f + F)(x| f (x) + y) = D f (x) + DF(x|y).

Proof. If v ∈ D( f + F)(x| f (x) + y)(u) there exist sequences τ k  0 and uk → u and

vk → v such that

f (x) + y − f (x + τ k uk ) + τ k vk ∈ F(x + τ k uk ) for every k.

By using the definition of the derivative for f we get

y + τ k (−D f (x)u + vk ) + o(τ k ) ∈ F(x + τ k uk ).

Hence, by the definition of the graphical derivative, v ∈ D f (x)u + DF(x|y)(u).

Conversely, if v − D f (x)u ∈ DF(x|y)(u) then there exist sequences τ k  0, and
u → u and wk → v − D f (x)u such that y + τ k wk ∈ F(x + τ k uk ). Again, by the dif-
k

ferentiability of f ,

y + f (x) + τ k vk + o(τ k ) ∈ ( f + F)(x + τ k uk ) for vk = wk + D f (x)uk ,

which yields v ∈ D( f + F)(x| f (x) + y)(u).

Example 4A.3 (graphical derivative for a constraint system). Consider a general

constraint system of the form

(4) f (x) − K  y,
4 Regularity Properties Through Generalized Derivatives 201

for a function f : IRn → IRm , a set K ⊂ IRm and a parameter vector y, and let x be a
solution of (4) for y at which f is differentiable. Then for the mapping

G : x → f (x) − K, with y ∈ G(x),

one has

(5) DG(x|y)(u) = D f (x)u − TK ( f (x) − y).

Detail. This applies the sum rule to the case of a constant mapping F ≡ −K,
for which the definition of the graphical derivative gives DF(x|z) = T−K (z) =
−TK (−z).
In the special but important case of Example 4A.3 in which K = IRs− × {0}m−s
with f = ( f1 , . . . , fm ), the constraint system (4) with respect to y = (y1 , . . . , ym ) takes
the form 
≤ yi for i = 1, . . . , s,
fi (x)
= yi for i = s + 1, . . ., m.
The graphical derivative formula (5) says then that a vector v = (v1 , . . . , vm ) is in
DG(x|y)(u) if and only if

≤ vi for i ∈ [1, s] with fi (x) = yi ,
D fi (x)u
= vi for i = s + 1, . . . , m.

Example 4A.4 (graphical derivative for a variational inequality). For a function

f : IRn → IRn and a convex set C ⊂ IRn that is polyhedral, consider the variational
inequality

(6) f (x) + NC (x)  y

in which y is a parameter. Let x be a solution of (6) at which f is differentiable. Let

v = y − f (x) ∈ NC (x) and let KC (x, v) be the corresponding critical cone, this being
the polyhedral convex cone TC (x) ∩ [v]⊥ . Then for the mapping

G : x → f (x) + NC (x), with y ∈ G(x),

one has

(7) DG(x|y)(u) = D f (x)u + NKC (x,v) (u).

Detail. From the sum rule in 4A.2 we have DG(x|y)u = D f (x)u + DNC (x|v)(u).
According to Lemma 2E.4 (the reduction lemma for normal cone mappings to poly-
hedral convex sets), for any (x, v) ∈ gph NC there exists a neighborhood O of the
origin in IRn × IRn such that for (x , v ) ∈ O one has

v + v ∈ NC (x + x ) ⇐⇒ v ∈ NKC (x,v) (x ).

202 4 Regularity Properties Through Generalized Derivatives

This reveals in particular that the tangent cone to gph NC at (x, v) is just gph NKC (x,v) ,
or in other words, that DNC (x|v) is the normal cone mapping NKC (x,v) . Thus we have
(7).
Because graphical derivative mappings are positively homogeneous, general
properties of positively homogeneous mappings can be applied to them. Norm con-
cepts are available in particular for capturing quantitative characteristics.

Outer and inner norms. For any positively homogeneous mapping H : IRn →
→ IRm ,
the outer norm and the inner norm are defined, respectively, by
+ −
(8) |H| = sup sup |y| and |H| = sup inf |y|
|x|≤1 y∈H(x) |x|≤1 y∈H(x)

with the convention infy∈0/ |y| = ∞ and supy∈0/ |y| = −∞.

When H is a linear mapping, both |H|+ and |H|− reduce to the operator (matrix)
norm |H| associated with the Euclidean norm. However, it must be noted that neither
|H|+ nor |H|− satisfies the conditions in the definition of a true “norm,” inasmuch as
set-valued mappings do not even form a vector space.
The inner and outer norms have simple interpretations when H = A−1 for a linear
mapping A : IRn → IRm . Let the m × n matrix for this linear mapping be denoted
likewise by A, for simplicity. If m < n, we have A surjective (the associated matrix
being of rank m) if and only if |A−1 |− is finite, this expression being the norm of the
right inverse of A: |A−1 |− = |AT (AAT )−1 |. Then |A−1 |+ = ∞. On the other hand, if
m > n, we have |A−1 |+ < ∞ if and only if A is injective (the associated matrix has
rank n), and then |A−1 |+ = |(AT A)−1 AT | but |A−1 |− = ∞. For m = n, of course, both
norms agree with the usual matrix norm |A−1 |, and the finiteness of this quantity is
equivalent to nonsingularity of A.

Proposition 4A.5 (domains of positively homogeneous mappings). For a positively

homogeneous mapping H : IRn →→ IRm ,
+ −
(9) dom H = IRn =⇒ |H| ≥ |H| .

Moreover,
−
(10) |H| < ∞ =⇒ dom H = IRn ;

thus, if |H −1 |− < ∞ then H must be surjective.

Proof. The implications (9) and (10) are immediate from the definition (8) and its
conventions concerning the empty set.
In cases where dom H is not all of IRn , it is possible for the inequality in (9) to fail.
As an illustration, this occurs for the positively homogeneous mapping H : IR → → IR
defined by 
0 for x ≥ 0,
H(x) =
0/ for x < 0,
4 Regularity Properties Through Generalized Derivatives 203

for which |H|+ = 0 while |H|− = ∞.

Proposition 4A.6 (norm characterizations). The inner norm of a positively homo-

geneous mapping H : IRn →
→ IRm satisfies
  
− 
(11) |H| = inf κ > 0  H(x) ∩ κ IB = 0/ for all x ∈ IB .

In parallel, the outer norm satisfies

  
+  1
(12) |H| = inf κ ∈ (0, ∞)  y ∈ H(x) ⇒ |y| ≤ κ |x| = sup .
|y|=1 d(0, H −1 (y))

If H has closed graph, then furthermore

+
(13) |H| < ∞ ⇐⇒ H(0) = {0}.

If H has closed and convex graph, then the implication (10) becomes equivalence:
−
(14) |H| < ∞ ⇐⇒ dom H = IRn

and in that case |H −1 |− < ∞ if and only if H is surjective.

Proof. We get (11) and  the first part of (12) simply by rewriting the formulas in
terms of IB = x  |x| ≤ 1 and utilizing the positive homogeneity. The infimum so
obtained in (12) is unchanged when y is restricted to have |y| = 1, and in this way
it can be identified with the infimum of all κ ∈ (0, ∞) such that κ ≥ 1/|x| whenever
x ∈ H −1 (y) and |y| = 1. (It is correct in this to interpret 1/|x| = ∞ when x = 0.) This
shows that the middle expression in (12) agrees with the final one.
Moving on to (13), we observe that when |H|+ < ∞ the middle expression in (12)
implies that if (0, y) ∈ gph H then y must be 0. To prove the converse implication we
will need the assumption that gph H is closed. Suppose that H(0) = {0}. If |H|+ =
∞, there has to be a sequence of points (xk , yk ) ∈ gph H such that 0 < |yk | → ∞ but
xk is bounded. Consider then the sequence of pairs (wk , uk ) in which wk = xk /|yk |
and uk = yk /|yk |. We have wk → 0, while uk has a cluster point ū with |ū| = 1.
Moreover, (wk , uk ) ∈ gph H by the positive homogeneity and hence, through the
closedness of gph H, we must have H(0)  ū. This contradicts our assumption that
H(0) = {0} and terminates the proof of (13). The equivalence (14) follows from (9)
and a general result (Robinson-Ursescu theorem) which we will prove in Section
5B.

Corollary 4A.7 (norms of linear-constraint-type mappings). Suppose

H(x) = Ax − K

for a linear mapping A : IRn → IRm and a closed convex cone K ⊂ IRm . Then H is
positively homogeneous with closed and convex graph. Moreover
204 4 Regularity Properties Through Generalized Derivatives

|H −1 | = sup d(0, A−1 (y + K))

−
(15)
|y|≤1

and

|H −1 | < ∞ ⇐⇒ rge A − K = IRm .

−
(16)

On the other hand,

1
|H −1 | = sup
+
(17)
|x|=1 d(Ax, K)

and
 
|H −1 | < ∞ ⇐⇒
+
(18) Ax − K  0 =⇒ x = 0 .

Proof. Formula (15) follows from the definition (8) while (16) comes from (14)
applied to this case. Formula (17) follows from (12) while (18) is the specification
of (13).
We will come back to the general theory of positively homogeneous mappings
and their norms in Section 5A. In the meantime there will be applications to the case
of derivative mappings.
Some properties of the graphical derivatives of convex-valued mappings under
Lipschitz continuity are displayed in the following exercise.

Exercise 4A.8. Consider a mapping F : IRn → → IRm that is convex-valued and

Lipschitz continuous in its domain and let (x, y) ∈ gph F . Prove that in this case
DF(x|y) is convex-valued and

(19) DF(x|y)(u) = lim τ −1 (F(x + τ u) − y),

τ 0

and in particular,

(20) DF(x|y)(0) = TF(x) (y).

Guide. Observe that, by definition

DF(x|y)(u) = lim inf τ −1 (F(x + τ u ) − y).

τ  0, u →u

Since F is Lipschitz continuous, this equality reduces to

(21) DF(x|y)(u) = lim inf τ −1 (F(x + τ u) − y).

τ 0

Then use the convexity of the values of F as in the proof of Proposition 4A.1 to
show that liminf in (21) can be replaced by lim and use this to obtain convexity of
4 Regularity Properties Through Generalized Derivatives 205

DF(x|y)(u) from the convexity of F(x + τ u). Lastly, to show (20) apply 4A.1 to
(19) in the case u = 0.

Exercise 4A.9. For a positively homogeneous mapping H : IRn →

→ IRm , show that
−
|H| = 0 ⇐⇒ cl H(x)  0 for all x ∈ IRn ,
+
|H| = 0 ⇐⇒ rge H = {0}.
Guide. Apply the norm characterizations in (11) and (12).

4B. Derivative Criteria for the Aubin Property

Conditions will next be developed which characterize metric regularity and the
Aubin property in terms of graphical derivatives. From these conditions, new forms
of implicit mapping theorems will be obtained. First, we state a fundamental fact.

Theorem 4B.1 (derivative criterion for metric regularity). For a mapping F :

IRn →
→ IRm and a point (x̄, ȳ) ∈ gph F at which the gph F is locally closed, one has

reg (F; x̄| ȳ) = lim sup |DF(x|y)−1 | .

−
(1)
(x,y)→(x̄,ȳ)
(x,y)∈gph F

Thus, F is metrically regular at x̄ for ȳ if and only if the right side of (1) is finite.
The proof of Theorem 4B.1 will be furnished later in this section. Note that in the
case when m ≤ n and F is a function f which is differentiable on a neighborhood of
x̄, the representation of the regularity modulus in (1) says that f is metrically regular
precisely when the Jacobians ∇ f (x) for x near x̄ are of full rank and the inner norms
of their inverses ∇ f (x)−1 are uniformly bounded. This holds automatically when f
is continuously differentiable around x̄ with ∇ f (x̄) of full rank, in which case we get
not only metric regularity but also existence of a continuously differentiable local
selection of f −1 , as in 1F.3. When m = n this becomes nonsingularity and we come
to the classical inverse function theorem.
Also to be kept in mind here is the connection between metric regularity and the
Aubin property in 3E.6. This allows Theorem 4B.1 to be formulated equivalently as
a statement about that property of the inverse mapping.

Theorem 4B.2 (derivative criterion for the Aubin property). For a mapping F :
IRn →
→ IRm , consider the inverse mapping S = F −1 or, equivalently, the solution map-
ping S : IRm →
→ IRn for the generalized equation F(x)  y:
  
S(y) = x  F(x)  y .
206 4 Regularity Properties Through Generalized Derivatives

Fix ȳ and any x̄ ∈ S(ȳ), and suppose that gph F or, equivalently, gph S, is locally
closed at (x̄, ȳ). Then
−
(2) lip (S; ȳ| x̄) = lim sup |DS(y|x)| .
(y,x)→(ȳ,x̄)
(y,x)∈gph S

Thus, S has the Aubin property at ȳ for x̄ if and only if the right side of (2) is finite.
Note that 4B.2 can be stated for the mapping S alone, without referring to it as
an inverse or a solution mapping.
Solution mappings of much greater generality can also be handled with these
ideas. For this, we return to the framework introduced briefly at the end of Section
3E and delve into it much further. We consider the parameterized relation

(3) G(p, x)  0 for a mapping G : IRd × IRn →

→ IR
m

and its solution mapping S : IRd →

→ IRn defined by
  
(4) S(p) = x  G(p, x)  0 .

In Theorem 3E.9, a result was presented in which a partial Aubin property of G

with respect to p, combined with other assumptions, led to a conclusion that S has
the Aubin property. We are looking now toward finding derivative criteria for these
Aubin properties, so as to obtain a different type of statement about the “implicit
mapping” S.
The following theorem will be our stepping stone to progress and will have many
other interesting consequences as well. It makes use of the partial graphical deriva-
tive of G(p, x) with respect to x, which is defined as the graphical derivative of
the mapping x → G(p, x) with p fixed and denoted by Dx G(p, x|y). Of course,
D p G(p, x) has a similar meaning.

Theorem 4B.3 (solution mapping estimate). For the generalized equation (3) and
its solution mapping S in (4), let x̄ ∈ S( p̄), so that ( p̄, x̄, 0) ∈ gph G. Suppose that
gph G is locally closed at ( p̄, x̄, 0) and that the distance mapping p → d(0, G(p, x̄))
is upper semicontinuous at p̄. Then for every c ∈ (0, ∞) satisfying

lim sup |Dx G(p, x|y)−1 | < c

−
(5)
(p,x,y)→( p̄,x̄,0)
(p,x,y)∈gph G

there are neighborhoods V of p̄ and U of x̄ such that

(6) d(x, S(p)) ≤ c d(0, G(p, x)) for x ∈ U and p ∈ V.

Proof. Let c satisfy (5). Then there exists η > 0 such that
%
for every (p, x, y) ∈ gph G with |p − p̄| + max{|x − x̄|, c|y|} ≤ 2η ,
(7)
and for every v ∈ IRm , there exists u ∈ Dx G(p, x|y)−1 (v) with |u| ≤ c|v|.
4 Regularity Properties Through Generalized Derivatives 207

We can always choose η smaller so that the intersection

-  
(8) gph G (p, x, y)  |p − p̄| + max{|x − x̄|, c|y|} ≤ 2η is closed.

The next part of the proof is developed as a lemma.

Lemma 4B.4 (intermediate estimate). For c and η as above, let ε > 0 and s > 0 be
such that

(9) cε < 1 and s < εη

and let (p, ω , ν ) ∈ gph G satisfy

(10) |p − p̄| + max{|ω − x̄|, c|ν |} ≤ η .

Then for every y ∈ IBs (ν ) there exists x̂ with y ∈ G(p, x̂) such that
1
(11) |x̂ − ω | ≤ |y − ν |.
ε
In the proof of the lemma we apply a fundamental result in variational analysis,
which is stated next:

Theorem 4B.5 (Ekeland variational principle). Let (X, ρ ) be a complete metric

space and let f : X → (−∞, ∞] be a lower semicontinuous function on X which is
bounded from below. Let ū ∈ dom f . Then for every δ > 0 there exists uδ such that

f (uδ ) + δ ρ (uδ , ū) ≤ f (ū),

and
f (uδ ) < f (u) + δ ρ (u, uδ ) for every u ∈ X, u = uδ .

Proof of Lemma 4B.4. On the product space Z := IRn × IRm we introduce the norm

(x, y) := max{|x|, c|y|},

which is equivalent to the Euclidean norm. Pick ε , s and (p, ω , ν ) ∈ gph G as re-
quired in (9) and (10) and let y ∈ IBs (ν ). By (8) the set
 
E p := (x, y)  (p, x, y) ∈ gph G, |p − p̄| + (x, y) − (x̄, 0) ≤ 2η } ⊂ IRn × IRm

is closed, hence, equipped with the metric induced by the norm in question, it is a
complete metric space. The function Vp : E p → IR defined by

(12) Vp : (x, y) → |y − y| for (x, y) ∈ E p

is continuous on its domain E p . Also, (ω , ν ) ∈ dom Vp . We apply Ekeland’s variatio-

nal principle 4B.5 to Vp with ū = (ω , ν ) and the indicated ε to obtain the existence
of (x̂, ŷ) ∈ E p such that
208 4 Regularity Properties Through Generalized Derivatives

(13) Vp (x̂, ŷ) + ε (ω , ν ) − (x̂, ŷ) ≤ Vp (ω , ν )

and

(14) Vp (x̂, ŷ) ≤ Vp (x, y) + ε (x, y) − (x̂, ŷ) for every (x, y) ∈ E p .

With Vp as in (12), the inequalities (13) and (14) come down to

(15) |y − ŷ| + ε (ω , ν ) − (x̂, ŷ) ≤ |y − ν |

and

(16) |y − ŷ| ≤ |y − y| + ε (x, y) − (x̂, ŷ) for every (x, y) ∈ E p .

Through (15) we obtain in particular that

1
(17) (ω , ν ) − (x̂, ŷ) ≤ |y − ν |.
ε
Since y ∈ IBs (ν ), we then have
s
(ω , ν ) − (x̂, ŷ) ≤
ε
and consequently, from the choice of (p, ω , ν ) in (10) and s in (9),

|p − p̄| + (x̂, ŷ) − (x̄, 0)

(18) s
≤ |p − p̄| + (ω , ν ) − (x̄, 0) + (ω , ν ) − (x̂, ŷ) ≤ η + < 2η .
ε
Thus, (p, x̂, ŷ) satisfies the condition in (7), so there exists u ∈ IRn for which

(19) y − ŷ ∈ Dx G(p, x̂| ŷ)(u) and |u| ≤ c|y − ŷ|.

By the definition of the partial graphical derivative, there exist sequences τ k  0,

uk → u, and vk → y − ŷ such that

ŷ + τ k vk ∈ G(p, x̂ + τ k uk ) for all k.

Also, from (18) we know that, for sufficiently large k,

|p − p̄| + (x̂ + τ k uk , ŷ + τ k vk ) − (x̄, 0) ≤ 2η ,

implying (x̂ + τ k uk , ŷ + τ k vk ) ∈ E p . If we now plug the point (x̂ + τ k uk , ŷ + τ k vk ) into

(16) in place of (x, y), we get

|y − ŷ| ≤ |y − (ŷ + τ k vk )| + ε (x̂ + τ k uk , ŷ + τ k vk ) − (x̂, ŷ).

This gives us
4 Regularity Properties Through Generalized Derivatives 209

|y − ŷ| ≤ (1 − τ k )|y − ŷ| + τ k |vk − (y − ŷ)| + ετ k (uk , vk ),

that is,
|y − ŷ| ≤ |vk − (y − ŷ)| + ε (uk , vk ).
Passing to the limit with k → ∞ leads to |y − ŷ| ≤ ε (u, y − ŷ) and then, taking
into account the second relation in (19), we conclude that |y − ŷ| ≤ ε c|y − ŷ|. Since
ε c < 1 by (9), the only possibility here is that y = ŷ. But then y ∈ G(p, x̂) and (17)
yields (11). This proves the lemma.
We continue now with the proof of Theorem 4B.3. Let τ = η /(4c). Since the
function p → d(0, G(p, x̄)) is upper semicontinuous at p̄, there exists a positive δ ≤
cτ such that d(0, G(p, x̄)) ≤ τ /2 for all p with |p − p̄| < δ . Set V := IBδ ( p̄), U :=
IBcτ (x̄) and pick any p ∈ V and x ∈ U. We can find y such that y ∈ G(p, x̄) with
|y| ≤ d(0, G(p, x̄)) + τ /3 < τ . Note that

(20) |p − p̄| + (x̄, y) − (x̄, 0) = |p − p̄| + c|y| ≤ δ + cτ ≤ η .

Choose ε > 0 such that 1/2 < ε c < 1 and let s = εη . Then s > τ . We apply
Lemma 4B.4 with the indicated ε and s, and with (p, ω , ν ) = (p, x̄, y) which, as
seen in (20), satisfies (10), and with y = 0, since 0 ∈ IBs (y). Thus, there exists x̂
such that 0 ∈ G(p, x̂), that is, x̂ ∈ S(p), and also, from (11), |x̂ − x̄| ≤ |y|/ε . There-
fore, in view of the choice of y, we have x̂ ∈ IBτ /ε (x̄). We now consider two cases.

C ASE 1. d(0, G(p, x)) ≥ 2τ . We just proved that there exists x̂ ∈ S(p) with x̂ ∈
IBτ /ε (x̄); then

d(x, S(p)) ≤ d(x̄, S(p)) + |x − x̄|

(21) τ 2τ 1
≤ |x̄ − x̂| + |x − x̄| ≤ + cτ ≤ ≤ d(0, G(p, x)).
ε ε ε
C ASE 2. d(0, G(p, x)) < 2τ . In this case, for any y with |y| ≤ 2τ we have

|p − p̄| + max{|x − x̄|, c|y|} ≤ δ + max{cτ , 2cτ } ≤ 3cτ ≤ η

and then, by (8), the nonempty set G(p, x) ∩ 2τ IB is closed. Hence, there exists ỹ ∈
G(p, x) such that |ỹ| = d(0, G(p, x)) < 2τ and therefore
η
c|ỹ| < 2cτ = .
2
We conclude that the point (p, x, ỹ) ∈ gph G satisfies
η
|p − p̄| + max{|x − x̄|, c|ỹ|} ≤ δ + max{cτ , } ≤ η.
2
Thus, the assumptions of Lemma 4B.4 hold for (p, ω , ν ) = (p, x, ỹ), s = 2τ , and
y = 0. Hence there exists x̃ ∈ S(p) such that
210 4 Regularity Properties Through Generalized Derivatives

1
|x̃ − x| ≤ |ỹ|.
ε
Then, by the choice of ỹ,
1 1
d(x, S(p)) ≤ |x − x̃| ≤ |ỹ| = d(0, G(p, x)).
ε ε
Hence, by (21), for both cases 1 and 2, and therefore for any p in V and x ∈ U, we
have
1
d(x, S(p)) ≤ d(0, G(p, x)).
ε
Since U and V do not depend on ε , and 1/ε can be arbitrarily close to c, this gives
us (6).
With this result in hand, we can confirm the criterion for metric regularity pre-
sented at the beginning of this section.
Proof of Theorem 4B.1. For short, let dDF denote the right side of (1). We will
start by showing that reg(F; x̄| ȳ) ≤ dDF . If dDF = ∞ there is nothing to prove.
Let dDF < c < ∞. Applying Theorem 4B.3 to G(p, x) = F(x) − p and this c, let-
ting y take the place of p, we have S(y) = F −1 (y) and d(0, G(y, x)) = d(y, F(x)).
Condition (6) becomes the definition of metric regularity of F at x̄ for ȳ = p̄, and
therefore reg (F; x̄| ȳ) ≤ c. Since c can be arbitrarily close to dDF we conclude that
reg (F; x̄| ȳ) ≤ dDF .
We turn now to demonstrating the opposite inequality,

(22) reg (F; x̄| ȳ) ≥ dDF .

If reg (F; x̄| ȳ) = ∞ we are done. Suppose therefore that F is metrically regular at x̄
for ȳ with respect to a constant κ and neighborhoods U for x̄ and V for ȳ. Then

(23) d(x , F −1 (y)) ≤ κ |y − y| whenever (x , y ) ∈ gph F, x ∈ U, y ∈ V.

We know from 3E.1 that V can be chosen so small that F −1 (y) ∩ U = 0/ for every
y ∈ V . Pick any y ∈ V and x ∈ F −1 (y ) ∩U, and let v ∈ IB. Take a sequence τ k  0
such that yk := y + τ k v ∈ V for all k. By (23) and the local closedness of gph F at
(x̄, ȳ) there exists xk ∈ F −1 (y + τ k v) such that

|x − xk | = d(x , F −1 (yk )) ≤ κ |yk − y | = κτ k |v|.

For uk := (xk − x )/τ k we obtain

(24) |uk | ≤ κ |v|.

Thus, uk is bounded, so uki → u for a subsequence ki → ∞. Since (xki , y + τ ki v) ∈

gph F, we obtain (u, v) ∈ Tgph F (x , y ). Hence, by the definition of the graphical
derivative, we have u ∈ DF −1 (y |x )(v) = DF(x |y )−1 (v). The bound (24) guaran-
tees that
4 Regularity Properties Through Generalized Derivatives 211

|DF(x|y)−1 | ≤ κ .
−

Since (y, x) ∈ gph S is arbitrarily chosen near (x̄, ȳ), and κ is independent of this
choice, we conclude that (22) holds and hence we have (1).
We apply Theorem 4B.3 now to obtain for the implicit mapping result in
Theorem 3E.9 an elaboration in which graphical derivatives provide estimates. Re-
 (G; p̄, x̄| ȳ), the modulus of the partial Aubin property
call here the definition of lip p
introduced just before 3E.9.

Theorem 4B.6 (implicit mapping theorem with graphical derivatives). For the
general inclusion (3) and its solution mapping S in (4), let x̄ ∈ S( p̄), so that
( p̄, x̄, 0) ∈ gph G. Suppose that gph G is locally closed at ( p̄, x̄, 0) and that the dis-
tance d(0, G(p, x̄)) depends upper semicontinuously on p at p̄. Assume further that
G has the partial Aubin property with respect to p uniformly in x at ( p̄, x̄), and that

lim sup |Dx G(p, x|y)−1 | ≤ λ < ∞.

−
(25)
(p,x,y)→( p̄,x̄,0)
(p,x,y)∈gph G

Then S has the Aubin property at p̄ for x̄ with

(26)  (G; p̄, x̄|0).

lip (S; p̄| x̄) ≤ λ lip p

Proof. This just combines Theorem 3E.9 with the estimate now available from
Theorem 4B.3.
Note from Proposition 4A.5 that finiteness in condition (25) necessitates, in par-
ticular, having the range of Dx G(p, x|y) be all of IRm when (p, x, y) is sufficiently
close to ( p̄, x̄, 0) in gph G.
Next we specialize Theorem 4B.6 to the generalized equations we studied in
detail in Chapters 2 and 3, or in other words, to a solution mapping of the type
  
(27) S(p) = x  f (p, x) + F(x)  0 ,

where f : IRd × IRn → IRm and F : IRn →→ IRm . In the next two corollaries we take a
closer look at the Aubin property of the solution mapping (27).

Corollary 4B.7 (derivative criterion for generalized equations). For the solution
 ( f ; ( p̄, x̄)) < ∞.
mapping S in (27), and a pair ( p̄, x̄) with x̄ ∈ S( p̄), suppose that lip p
Then the mapping G(p, x) := f (p, x) + F(x) has the partial Aubin property with
respect to p uniformly in x at ( p̄, x̄) with

(28)  (G; p̄, x̄|0) ≤ lip

lip  ( f ; ( p̄, x̄)).
p p

In addition, if f is differentiable in a neighborhood of ( p̄, x̄), gph F is locally closed

at (x̄, − f ( p̄, x̄)) and
212 4 Regularity Properties Through Generalized Derivatives

lim sup |(Dx f (p, x) + DF(x|y − f (p, x)))−1 | ≤ λ < ∞,

−
(29)
(p,x,y)→( p̄,x̄,0)
y∈ f (p,x)+F (x)

then S has the Aubin property at p̄ for x̄ with

(30)  ( f ; p̄ | x̄).
lip (S; p̄| x̄) ≤ λ lip p

 ( f ; ( p̄, x̄))
Proof. By definition, the mapping G has ( p̄, x̄, 0) ∈ gph G. Let μ > lip p
and let Q and U be neighborhoods of p̄ and x̄ such that f is Lipschitz continuous
with respect to p ∈ Q uniformly in x ∈ U with Lipschitz constant μ . Let p, p ∈ Q,
x ∈ U and y ∈ G(p, x); then y − f (p, x) ∈ F(x) and we have

d(y, G(p , x)) = d(y − f (p , x), F (x)) ≤ | f (p, x) − f (p , x)| ≤ μ |p − p|.

Thus,
e(G(p, x), G(p , x)) ≤ μ |p − p|
and hence G has the partial Aubin (actually, Lipschitz) property with respect to
p uniformly in x at ( p̄, x̄) with modulus satisfying (28). The assumptions that
f is differentiable near ( p̄, x̄) and gph F is locally closed at (x̄, − f ( p̄, x̄)) yield
that gph G is locally closed at ( p̄, x̄, 0) as well. Further, observe that the function
p → d(0, G(p, x̄)) = d(− f (p, x̄), F(x̄)) is Lipschitz continuous near p̄ and therefore
upper semicontinuous at p̄. Then we can apply Theorem 4B.6 where, by using the
sum rule 4A.2, the condition (25) comes down to (29) while (26) yields (30).
From Section 3F we know that when the function f is continuously differen-
tiable, the Aubin property of the solution mapping in (27) can be obtained by pass-
ing to the linearized generalized equation, in which case we can also utilize the
ample parameterization condition. Specifically, we have the following result:

Corollary 4B.8 (derivative criterion with continuous differentiability and ample pa-
rameterization). For the solution mapping S in (27), and a pair ( p̄, x̄) with x̄ ∈ S( p̄),
suppose that f is continuously differentiable on a neighborhood of ( p̄, x̄) and that
gph F is locally closed at (x̄, − f ( p̄, x̄)). If

|(Dx f ( p̄, x̄) + DF(x|y − Dx f ( p̄, x̄)(x − x̄)))−1 | ≤ λ < ∞,

−
(31) lim sup
(x,y)→(x̄,− f ( p̄,x̄))
y∈Dx f ( p̄,x̄)(x−x̄)+F(x)

then S has the Aubin property at p̄ for x̄, moreover with

(32) lip (S; p̄| x̄) ≤ λ |∇ p f ( p̄, x̄)|.

Furthermore, when f satisfies the ample parameterization condition

(33) rank ∇ p f ( p̄, x̄) = m,

then the converse implication holds as well; that is, S has the Aubin property at p̄
for x̄ if and only if condition (31) is satisfied.
4 Regularity Properties Through Generalized Derivatives 213

Proof. According to Theorem 3F.9, the mapping S has the Aubin property at p̄ for
x̄ provided that the linearized mapping

(34) h + F for h(x) = f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄)

is metrically regular at x̄ for 0, and the converse implication holds under the ample
parameterization condition (33). Further, according to the derivative criterion for
metric regularity 4B.1, metric regularity of the mapping h + F in (34) is equivalent
to condition (31) and its regularity modulus is bounded by λ . Then the estimate (32)
follows from formula 3F(7) in the statement of 3F.9.
The purpose of the next exercise is to understand what condition (29) means in
the setting of the classical implicit function theorem.

Exercise 4B.9 (application to classical implicit functions). For a function f : IRd ×

IRn → IRm , consider the solution mapping
  
S : p → x  f (p, x) = 0

and a pair ( p̄, x̄) with x̄ ∈ S( p̄). Suppose that f is differentiable in a neighborhood
of ( p̄, x̄) with Jacobians satisfying

lim sup |Dx f (p, x)−1 | < λ

−
and lim sup |D p f (p, x)| < κ .
(p,x)→( p̄,x̄) (p,x)→( p̄,x̄)

Show that then S has the Aubin property at p̄ for x̄ with constant λ κ.
When f is continuously differentiable we can apply Corollary 4B.8, and the
assumptions in 4B.9 can in that case be captured by conditions on the Jacobian
∇ f ( p̄, x̄). Then 4B.8 goes a long way toward the classical implicit function theo-
rem, 1A.1. But Steps 2 and 3 of Proof I of that theorem would afterward need to
be carried out to reach the conclusion that S has a single-valued localization that is
smooth around p̄.
Applications of Theorem 4B.6 and its corollaries to constraint systems and va-
riational inequalities will be worked out in Sections 4D and 4E. We conclude the
present section with a variant of the graphical derivative formula for the modulus
of metric regularity in Theorem 4B.1, which will be put to use in the numerical
variational analysis of Chapter 6.
Recall that the closed convex hull of a set C ⊂ IRn , which will be denoted by
cl co C, is the smallest closed convex set that contains C.

Convexified graphical derivative. For a mapping F : IRn → → IRm and a pair (x, y)
with y ∈ F(x), the convexified graphical derivative of F at x for y is the mapping
D̃F(x|y) : IRn → → IRm whose graph is the closed convex hull of the tangent cone
Tgph F (x, y) to gph F at (x, y):

v ∈ D̃F(x|y)(u) ⇐⇒ (u, v) ∈ cl co Tgph F (x, y).

214 4 Regularity Properties Through Generalized Derivatives

Theorem 4B.10 (alternative characterization of regularity modulus). For a mapping

F : IRn →
→ IRm and a point (x̄, ȳ) ∈ gph F at which the graph of F is locally closed,
one has
reg (F; x̄| ȳ) = lim sup |D̃F(x|y)−1 | .
−

(x,y)→(x̄,ȳ)
(x,y)∈gph F

Proof. Since D̃F(x|y)−1 (v) ⊃ DF(x|y)−1 (v) for any v ∈ IRn , we have

inf |u| ≤ inf |u|,

u∈D̃F(x | y)−1 (v) u∈ DF(x | y)−1 (v)

so that |D̃F(x|y)−1 |− ≤ |DF(x|y)−1 |− . Thus,

lim sup |D̃F(x|y)−1 | ≤ reg (F; x̄| ȳ),

−

(x,y)→(x̄,ȳ)
(x,y)∈gph F

and to complete the proof we only need validate the opposite inequality. Choose λ
such that
lim sup |D̃F(x|y)−1 | ≤ λ < ∞.
−

(x,y)→(x̄,ȳ)
(x,y)∈gph F

Let r > 0 be small enough that

(35) sup inf |u| ≤ λ for all (x, y) ∈ gph F ∩ IBr (x̄, ȳ),
v∈IB u∈D̃F(x | y)−1 (v)

and that the set gph F ∩ IBr (x̄, ȳ) is closed. We will now demonstrate that

(36) sup inf |u| ≤ λ for all (x, y) ∈ gph F ∩ int IBr (x̄, ȳ),
v∈IB u∈DF(x | y)−1 (v)

which will be enough to complete the proof.   

Fix v ∈ IB. For any sets A and B let d− (A, B) := inf |a − b|  a ∈ A, b ∈ B . Fix
(x, y) ∈ gph F ∩ int IBr (x̄, ȳ), and let (u∗ , v∗ ) ∈ gph DF(x|y) and w ∈ λ IB be such that

|(w, v) − (u∗ , v∗ )| = d− (λ IB × {v}, gph DF(x|y)).

Observe that the point (u∗ , v∗ ) is the unique projection of any point in the open seg-
ment ((u∗ , v∗ ), (w, v)) on gph DF(x|y). We will show that (u∗ , v∗ ) = (w, v), thereby
confirming (36).
By the definition of the graphical derivative DF(x|y), there exist sequences
τ k  0, uk → u∗ and vk → v∗ , such that y + τ k vk ∈ F(x + τ k uk ) for all k. Let (xk , yk ) be
a point in cl gph F which is closest to (x, y)+ τ2 (u∗ + w, v∗ + v). Since (x, y) ∈ gph F
k

we have
 
 
(x, y) + τ (u∗ + w, v∗ + v) − (xk , yk ) ≤ τ |(u∗ + w, v∗ + v)|,
k k
 2  2

and consequently
4 Regularity Properties Through Generalized Derivatives 215
   
   τk 
(x, y) − (xk , yk ) ≤ (x, y) + (u∗ + w, v∗ + v) − (xk , yk )
2
τk ∗
+ |(u + w, v∗ + v)| ≤ τ k |(u∗ + w, v∗ + v)|.
2
Thus, for k sufficiently large, we have (xk , yk ) ∈ intIBr (x̄, ȳ) and hence (xk , yk ) ∈
gph F ∩ intIBr (x̄, ȳ). Setting (ūk , v̄k ) = τ1k (xk − x, yk − y), we get, from the basic prop-
erties of projections, that
1 ∗
(u + w, v∗ + v) − (ūk , v̄k ) ∈ [Tgph F (xk , yk )]∗ = gph [D̃F(xk |yk )]∗ .
2
Then, by (35), there exists wk ∈ λ IB such that v ∈ D̃F(xk |yk )(wk ) and also
3 ∗ 4 3 ∗ 4
u +w v +v
(37) − ūk , wk + − v̄k , v ≤ 0.
2 2

We will show now that (v̄k , ūk ) converges to (v∗ , u∗ ) as k → ∞. First observe that
 ∗     ∗  
 u + w v∗ + v k k 
 1  k u +w v +v
∗
k k 

 , − (ū , v̄ ) = (x, y) + τ , − (x , y )
 2 2  τ k  2 2 
  ∗  
1 u + w v∗ + v 
≤ k (x, y) + τ k , − (x, y) − τ k (uk , vk )
τ 2 2
 ∗  
 u +w v +v ∗ 
= , k k 
− (u , v ).
2 2
Therefore, since (uk , vk ) is a bounded sequence, the sequence {(ūk , v̄k )} is bounded
too and has a cluster point (ū, v̄) which, since yk = y + τ k v̄k ∈ F(xk ) = F(x + τ k ūk ),
belongs to gph DF(x|y). Moreover, by the last estimation, the limit (ū, v̄) satisfies
 ∗    ∗  
 u + w v∗ + v   u + w v∗ + v 
 ,  
− (ū, v̄) ≤  , ∗ ∗ 
− (u , v ).
 2 2 2 2

This inequality, together with the fact that (u∗ , v∗ ) is the unique closest point to
1 ∗ ∗ ∗ ∗
2 (u + w, v + v) in gph DF(x, y) implies that (ū, v̄) = (u , v ).
k
Up to a subsequence, the sequence of points w satisfying (37) converges to some
w̄ ∈ λ IB. Passing to the limit in (37) we obtain

(38) w − u∗ , w̄ + v − v∗, v ≤ 0.

Since (w, v) is the unique projection of (u∗ , v∗ ) on the closed convex set λ IB × {v},
we have

(39) w − u∗, w − w̄ ≤ 0.

216 4 Regularity Properties Through Generalized Derivatives

Finally, since (u∗ , v∗ ) is the unique projection of 12 (u∗ + w, v∗ + v) on gph DF(x|y)

which is a closed cone, we get

(40) w − u∗, u∗  + v − v∗, v∗  = 0.

In view of (38), (39) and (40), we have

|(w, v) − (u∗ , v∗ )|2 = w − u∗ , w − w̄

+ (w − u∗ , w̄ + v − v∗, v)
− (w − u∗ , u∗  + v − v∗, v∗ ) ≤ 0.

Hence w = u∗ and v = v∗ , and we are done.

Exercise 4B.11 (sum rule for convexified derivatives). For a function f : IRn → IRm
which is differentiable at x and a mapping F : IRn →
→ IRm , prove that

D̃( f + F)(x| f (x) + y)(u) = D f (x)u + D̃F(x|y)(u).

Guide. Let v ∈ D̃( f + F)(x| f (x) + y)(u). By Carathéodory’s theorem on convex

hull representation, there are sequences {uki }, {vki } and {λik } for i = 0, 1, . . . , n + m
and k = 1, 2, . . . with λik ≥ 0, ∑n+m i=0 λi = 1, such that vi ∈ D( f + F)(x| f (x) + y)(ui )
k k k

for all i and k and ∑n+m

i=0 iλ k (uk , vk ) → (u, v) as k → ∞. From 4A.2, get vk ∈ D f (x)ui +
i i i k
DF(x|y)(uki ) for all i and k. Hence ∑n+m i=0 λi (ui , vi − D f (x)ui ) ∈ cl co gph DF(x|y).
k k k k

Then pass to the limit.

We end this section with yet another proof of the classical inverse function theo-
rem 1A.1. This time it is based on the Ekeland principle given in 4B.5.
Proof of Theorem 1A.1. Without loss of generality, let x̄ = 0, f (x̄) = 0. Let A =
∇ f (0) and let δ = |A−1 |−1 . Choose a > 0 such that

δ
(41) | f (x) − f (x ) − A(x − x)| ≤ |x − x | for all x, x ∈ aIB,
2
and let b = aδ /2. We now redo Step 1 in Proof I that the localization s of f −1 with
respect to the neighborhoods bIB and aIB is nonempty-valued. The other two steps
remain the same as in Proof I.
Fix y ∈ bIB and consider the function | f (x)−y| with domain containing the closed
ball aIB, which we view as a complete metric space equipped with the Euclidean
metric. This function is continuous and bounded below, hence, by Ekeland principle
4B.5 with the indicated δ and ū = 0 there exists xδ ∈ aIB such that

δ
(42) |y − f (xδ )| < |y − f (x)| + |x − xδ | for all x ∈ aIB, x = xδ .
2
Let us assume that y = f (xδ ). Then x̃ := A−1 (y − f (xδ )) + xδ = xδ . Moreover, from
(41) with x = xδ and x = 0 and the choice of δ and b we get
4 Regularity Properties Through Generalized Derivatives 217
 
−1 −1 aδ
|x̃| ≤ |A |(|y| + | − f (xδ ) + Axδ |) ≤ |A | b + = |A−1 |aδ = a.
2

Hence we can set x = x̃ in (42), obtaining

δ
(43) |y − f (xδ )| < |y − f (x̃)| + |x̃ − xδ |.
2
Using (41), we have

|y − f (x̃)| = | f (A−1 (y − f (xδ )) + xδ ) − y|

(44) = | f (A−1 (y − f (xδ )) + xδ ) − f (xδ ) − A(A−1(y − f (xδ )))|
≤ δ2 |A−1 (y − f (xδ ))|

and also

(45) |x̃ − xδ | = |A−1 (y − f (xδ ))|.

Plugging (44) and (45) into (43), we arrive at

 
δ δ
|y − f (xδ )| < + |A−1 (y − f (xδ ))| ≤ δ |A−1 ||y − f (xδ )| = |y − f (xδ )|
2 2

which furnishes a contradiction. Thus, our assumption that y = f (xδ ) is voided, and
we have xδ ∈ f −1 (y) ∩ (aIB). This means that s is nonempty-valued, and the proof
is complete.

4C. Characterization of Strong Metric Subregularity

Strong metric subregularity of a mapping F : IRn →

→ IRm at x̄ for ȳ, where ȳ ∈ F(x̄),
was defined in Section 3I to mean the existence of a constant κ along with neigh-
borhoods U of x̄ and V of ȳ such that

(1) |x − x̄| ≤ κ d(ȳ, F(x) ∩V ) for all x ∈ U.

This property is equivalent to the combination of two other properties: that F is

metrically subregular at x̄ for ȳ, and x̄ is an isolated point of F −1 (ȳ). The associated
modulus, the infimum of all κ > 0 for which this holds for some U and V , is thus
the same as the modulus of subregularity, subreg(F; x̄| ȳ).
It was shown in 3I.2 that F is strongly metrically subregular at x̄ for ȳ if and
only if F −1 has the isolated calmness property at ȳ for x̄. As an illustration, a linear
mapping A is strongly metrically subregular at x̄ for ȳ = Ax̄ if and only if A−1 (ȳ)
consists only of x̄, i.e., A is injective. More generally, a mapping F that is polyhedral,
218 4 Regularity Properties Through Generalized Derivatives

as defined in 3D, is strongly metrically subregular at x̄ for ȳ if and only if x̄ is an

isolated point of F −1 (ȳ); this follows from 3I.1.
What makes the strong metric subregularity attractive, along the same lines as
metric regularity and strong metric regularity, is its stability with respect to approx-
imation as established in 3I.6. In particular, a function f : IRn → IRm which is differ-
entiable at x̄ is strongly metrically subregular at x̄ for f (x̄) if and only if its Jacobian
∇ f (x̄) has rank n, so that ∇ f (x̄)u = 0 implies u = 0. According to 4A.6, this is also
characterized by |D f (x̄)−1 |+ < ∞. It turns out that such an outer norm characteri-
zation can be provided also for set-valued mappings by letting graphical derivatives
take over the role of ordinary derivatives.

Theorem 4C.1 (derivative criterion for strong metric subregularity). A mapping

F : IRn →
→ IRm whose graph is locally closed at (x̄, ȳ) ∈ gph F is strongly metrically
subregular at x̄ for ȳ if and only if

(2) DF(x̄| ȳ)−1 (0) = {0},

this being equivalent to

|DF(x̄| ȳ)−1 | < ∞,

+
(3)

and in that case

subreg(F; x̄| ȳ) = |DF(x̄| ȳ)−1 | .

+
(4)

Proof. The equivalence between (2) and (3) comes from 4A.6. To get the equiv-
alence of these conditions with strong metric subregularity, suppose first that κ >
subreg(F; x̄| ȳ) so that F is strongly metrically subregular at x̄ for ȳ and (1) holds for
some neighborhoods U and V . By definition, having v ∈ DF(x̄| ȳ)(u) refers to the
existence of sequences uk → u, vk → v and τ k  0 such that ȳ + τ k vk ∈ F(x̄ + τ k uk ).
Then x̄ + τ k uk ∈ U and ȳ + τ k vk ∈ V eventually, so that (1) yields |(x̄ + τ k uk ) − x̄| ≤
κ |(ȳ + τ k vk ) − ȳ|, which is the same as |uk | ≤ κ |vk |. In the limit, this implies
|u| ≤ κ |v|. But then, by 4A.6, |DF(x̄| ȳ)−1 |+ ≤ κ and hence

subreg(F; x̄| ȳ) ≥ |DF(x̄| ȳ)−1 | .

+
(5)

In the other direction, (3) implies the existence of a κ > 0 such that

sup sup |u| < κ .

v∈IB u∈DF(x̄ | ȳ)−1 (v)

This in turn implies that |x − x̄| ≤ κ |y − ȳ| for all (x, y) ∈ gph F close to (x̄, ȳ). That
description fits with (1). Further, κ can be chosen arbitrarily close to |DF(x̄| ȳ)−1 |+ ,
and therefore |DF(x̄| ȳ)−1 |+ ≥ subreg (F; x̄| ȳ). This, combined with (5), finishes the
argument.
4 Regularity Properties Through Generalized Derivatives 219

Corollary 4C.2 (derivative criterion for isolated calmness). For a mapping F :

IRn →
→ IRm , consider the inverse S = F −1 , which can also be viewed as the solution
mapping S : IRm →→ IRn for the generalized equation F(x)  y:
  
S(y) = x  F(x)  y .

Fix ȳ and any x̄ ∈ S(ȳ), and suppose that gph F is locally closed at (x̄, ȳ), which is
the same as gph S being locally closed at (ȳ, x̄). Then
+
clm (S; ȳ| x̄) = |DS(ȳ| x̄)| .

Thus, S has the isolated calmness property at ȳ for x̄ if and only if |DS(ȳ| x̄)|+ < ∞.
Theorem 4C.1 immediately gives us the linearization result in Corollary 3I.9 by
using the sum rule in 4A.2.
Implicit function theorems could be developed for the isolated calmness of so-
lution mappings to general inclusions G(p, x)  0 in parallel to the results in 4B,
but we shall not do this here. We limit ourselves to an application of the derivative
criterion 4C.1 to the solution mapping of a generalized equation
  
(6) S(p) = x  f (p, x) + F(x)  0 ,

where f : IRd × IRn → IRm and F : IRn →

→ IRm . In the following corollary we utilize
Theorem 3I.13 and the ample parameterization condition.

Corollary 4C.3 (derivative rule for isolated calmness of solution mappings). For
the solution mapping S in (6) and a pair ( p̄, x̄) with x̄ ∈ S( p̄), suppose that f is
differentiable with respect to x uniformly in p at ( p̄, x̄) and also differentiable with
respect to p uniformly in x at ( p̄, x̄). Also, suppose that gph F is locally closed at
(x̄, − f ( p̄, x̄)). If

|(Dx f ( p̄, x̄) + DF(x̄| − f ( p̄, x̄)))−1 | ≤ λ < ∞,

+
(7)

then S has the isolated calmness property at p̄ for x̄, moreover with

(8) clm (S; p̄| x̄) ≤ λ |∇ p f ( p̄, x̄)|.

Furthermore, when f is continuously differentiable in a neighborhood of ( p̄, x̄) and

satisfies the ample parameterization condition

(9) rank ∇ p f ( p̄, x̄) = m,

then the converse implication holds as well; that is, S has isolated calmness property
at p̄ for x̄ if and only if (7) is satisfied.
Proof. We apply Theorem 3I.13 according to which the mapping S has the isolated
calmness property at p̄ for x̄ if the mapping

h + F for h(x) = f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄)

220 4 Regularity Properties Through Generalized Derivatives

is strongly metrically subregular at x̄ for 0, and the converse implication holds under
the ample parameterization condition (9). Then, it is sufficient to apply 4C.1 together
with the sum rule 4A.2 to the mapping h + F. The estimate (8) follows from formula
3I(24).

The following simple example in terms of graphical derivatives illustrates further

the distinction between metric regularity and strong metric subregularity.

x
−1 0 1

Fig. 4.1 Graph of the mapping in Example 4C.4.

Example 4C.4 (strong metric subregularity without metric regularity). Let F :

IR →
→ IR be defined by
⎧
⎨ [1 − (x − 1)2, ∞) for 0 ≤ x ≤ 1,
F(x) = [ 1 − (x + 1)2, ∞) for −1 ≤ x ≤ 0,
⎩
0/ elsewhere,
as shown in Figure 4.1. Then

[0, ∞) for u = 0,
DF(0|0)(u) =
0/ for u = 0,

and therefore
|DF(0|0)−1 | = 0, |DF(0|0)−1 | = ∞.
+ −

This fits with F being strongly metrically subregular, but not metrically regular at 0
for 0.
4 Regularity Properties Through Generalized Derivatives 221

4D. Applications to Parameterized Constraint Systems

Next we look at what the graphical derivative results in the preceding sections have
to say about a constraint system

(1) f (p, x) − K  0, or equivalently f (p, x) ∈ K,

and its solution mapping

  
(2) S : p → x  f (p, x) ∈ K

for a function f : IRd × IRn → IRm and a set K ⊂ IRm .

Theorem 4D.1 (implicit mapping theorem for a constraint system). Let x̄ ∈ S( p̄) for
solution mapping S of the constraint system (1) and suppose that f is differentiable
 ( f ; ( p̄, x̄)) < ∞, and that the set K is
in a neighborhood of ( p̄, x̄) and satisfies lip p
closed. If
 
(3) lim sup sup d 0, Dx f (p, x)−1 (v + TK ( f (p, x) − y)) ≤ λ < ∞,
(p,x,y)→( p̄,x̄,0) |v|≤1
f (p,x)−y∈K

then S has the Aubin property at p̄ for x̄, with

(4)  ( f ; ( p̄, x̄)).

lip (S; p̄, x̄) ≤ λ lip p

Proof. The assumed closedness of K and continuous differentiability of f around

( p̄, x̄) allow us to apply Corollary 4B.8 to the case of F(x) ≡ −K. Further, according
to 4A.3 we have

Dx G(p, x|y) = Dx f (p, x) − TK ( f (p, x) − y).

Next, we use the definition of inner norm in 4A(8) to write 4B(29) as (3) and apply
4B.7 to obtain that S has the Aubin property at p̄ for x̄. The estimate (4) follows
immediately from 4B(30).
A much sharper result can be obtained when f is continuously differentiable and
the set K in the system (1) is polyhedral convex.

Theorem 4D.2 (constraint systems with polyhedral convexity). Let x̄ ∈ S( p̄) for the
solution mapping S of the constraint system (1) in the case of a polyhedral convex
set K . Suppose that f is continuously differentiable in a neighborhood of ( p̄, x̄).
Then for S to have the Aubin property at x̄ for p̄, it is sufficient that

(5) rge Dx f ( p̄, x̄) − TK ( f ( p̄, x̄)) = IRm ,

in which case the corresponding modulus satisfies lip (S; p̄| x̄) ≤ λ |∇ p f ( p̄, x̄)| for
222 4 Regularity Properties Through Generalized Derivatives
 
(6) λ = sup d 0, Dx f ( p̄, x̄)−1 (v + TK ( f ( p̄, x̄))) .
|v|≤1

Moreover (5) is necessary for S to have the Aubin property at p̄ for x̄ under the
ample parameterization condition

(7) rank ∇ p f ( p̄, x̄) = m.

Proof. We invoke Theorem 4D.1 but make special use of the fact that K is polyhe-
dral. That property implies that TK (w) ⊃ TK (w̄) for all w sufficiently near to w̄, as
seen in 2E.3; we apply this to w = f (p, x) − y and w̄ = f ( p̄, x̄) in the formulas (3)
and (4) of 4D.1. The distances in question are greatest when the cone is as small as
possible; this, combined with the continuous differentiability of f , allows us to drop
the limit in (3). Further, from the equivalence relation 4A(16) in Corollary 4A.7, we
obtain that the finiteness of λ in (6) is equivalent to (5).
For the necessity, we bring in a further argument which makes use of the ample
parameterization condition (7). According to Theorem 3F.9, under (7) the Aubin
property of S at p̄ for x̄ implies metric regularity of the linearized mapping h − K
for h(x) = f ( p̄, x̄) + ∇x f ( p̄, x̄)(x − x̄). The derivative criterion for metric regularity
4B.1 tells us then that

lim sup (x,y)→(x̄,0) sup|v|≤1 d 0,
(8) f ( p̄,x̄)+Dx f ( p̄,x̄)(x−x̄)−y∈K 
Dx f ( p̄, x̄)−1 (v + TK ( f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄) − y)) < ∞.

Taking x = x̄ and y = 0 instead of limsup in (8) gives us the expression for λ in

(6) and may only decrease the left side of this inequality. We already know that the
finiteness of λ in (6) yields (5), and so we are done.
Example 4D.3 (application to systems of inequalities and equalities). For K =
IRs− × {0}m−s, the solution mapping S(p) in (2) consists, in terms of f (p, x) =
( f 1 (p, x), . . . , fm (p, x)) of all solutions x to

≤ 0 for i ∈ [1, s],
fi (p, x)
= 0 for i ∈ [s + 1, m].

Let x̄ solve this for p̄ and let each fi be continuously differentiable around ( p̄, x̄).
Then a sufficient condition for S to have the Aubin property for p̄ for x̄ is the
Mangasarian–Fromovitz condition:

∇x fi ( p̄, x̄)w < 0 for i ∈ [1, s] with fi ( p̄, x̄) = 0,
(9) ∃ w ∈ IR with
n
∇x fi ( p̄, x̄)w = 0 for i ∈ [s + 1, m],

and

(10) the vectors ∇x fi ( p̄, x̄) for i ∈ [s + 1, m] are linearly independent.

4 Regularity Properties Through Generalized Derivatives 223

Moreover, the combination of (9) and (10) is also necessary for S to have the Aubin
property under the ample parameterization condition (7). In particular, when f is
independent of p and then 0 ∈ f (x̄) − K, the Mangasarian–Fromovitz condition (9)–
(10) is a necessary and sufficient condition for metric regularity of the mapping
f − K at x̄ for 0.
Detail. According to 4D.2, it is enough to show that (5) is equivalent to the com-
bination of (9) and (10) in the case of K = IRs− × {0}m−s. Observe that the tangent
cone to the set K at f ( p̄, x̄) has the following form:

≤ 0 for i ∈ [1, s] with fi ( p̄, x̄) = 0,
(11) v ∈ TK ( f ( p̄, x̄)) ⇐⇒ vi
= 0 for i ∈ [s + 1, m].

Let (5) hold. Then, using (11), we obtain that the matrix with rows the vectors
∇x fs+1 ( p̄, x̄), . . . , ∇x fm ( p̄, x̄) must be of full rank, hence (10) holds. If (9) is violated,
then for every w ∈ IRn either ∇x fi ( p̄, x̄)w ≥ 0 for some i ∈ [1, s] with fi ( p̄, x̄) = 0, or
∇x fi ( p̄, x̄)w = 0 for some i ∈ [s + 1, m], which contradicts (5) in an obvious way.
The combination of (9) and (10) implies that for every y ∈ IRm there exist w, v ∈
IR and z ∈ IRm with zi ≤ 0 for i ∈ [1, s] with fi ( p̄, x̄) = 0 such that
n


∇x fi ( p̄, x̄)w − zi = yi for i ∈ [1, s] with fi ( p̄, x̄) = 0,
∇x fi ( p̄, x̄)(w + v) = yi for i ∈ [s + 1, m].

But then (5) follows directly from the form (11) of the tangent cone.
If f is independent of p, by 3E.6 the metric regularity of − f + K is equivalent
to the Aubin property of the inverse (− f + K)−1 , which is the same as the solution
mapping   
S(p) = x  p + f (x) ∈ K
for which the ample parameterization condition (7) holds automatically. Then, from
4D.2, for x̄ ∈ S( p̄), the Aubin property of S at p̄ for x̄ and hence metric regularity of
f − K at x̄ for p̄ is equivalent to (5) and therefore to (9)–(10).

Exercise 4D.4. Consider the constraint system in 4D.3 with f (p, x) = g(x) − p,
p̄ = 0 and g continuously differentiable near x̄. Show that the existence of a
Lipschitz continuous local selection of the solution mapping S at 0 for x̄ implies
the Mangasarian–Fromovitz condition. In other words, the existence of a Lipschitz
continuous local selection of S at 0 for x̄ implies metric regularity of the mapping
g − K at x̄ for 0.
Guide. Utilizing 2B.11, from the existence of a local selection of S at 0 for x̄ we
obtain that the inverse F0−1 of the linearization F0 (x) := g(x̄) + ∇g(x̄)(x − x̄) − K
has a Lipschitz continuous local selection at 0 for x̄. Then, in particular, for every
v ∈ IRm there exists w ∈ IRn such that

∇gi (x̄)w ≤ vi for i ∈ [1, s] with gi (x̄) = 0,
∇gi (x̄)w = vi for i ∈ [s + 1, m].
224 4 Regularity Properties Through Generalized Derivatives

This is the same as (5).

A result parallel to 4D.2 can be formulated also for isolated calmness instead of
the Aubin property.

Proposition 4D.5 (isolated calmness of constraint systems). In the setting of

Theorem 4D.1, for S to have the isolated calmness property at p̄ for x̄ it is suffi-
cient that
∇x f ( p̄, x̄)u ∈ TK ( f ( p̄, x̄)) =⇒ u = 0.
Moreover, this condition is necessary for S to have the isolated calmness property at
p̄ for x̄ under the ample parameterization condition (8).
Proof. This is a special case of Corollary 4C.3 in which we utilize 4A.2.
We note that the isolated calmness property offers little of interest in the case of
solution mappings for constraint systems beyond equations, inasmuch as it necessi-
tates x̄ being an isolated point of the solution set S( p̄); this restricts significantly the
class of constraint systems for which such a property may occur. In the following
section we will consider mappings associated with variational inequalities for which
the isolated calmness is a more natural property.

4E. Isolated Calmness for Variational Inequalities

Now we take up once more the topic of variational inequalities, to which serious at-
tention was already devoted in Chapter 2. This revolves around a generalized equa-
tion of the form

(1) f (p, x) + NC (x)  0, or − f (p, x) ∈ NC (x),

for a function f : IRd × IRn → IRn and the normal cone mapping NC associated with a
nonempty, closed, convex set C ⊂ IRn , and the solution mapping S : IRd →
→ IRn defined
by
  
(2) S(p) = x  f (p, x) + NC (x)  0 .

Especially strong results were obtained in 2E for the case in which C is a polyhedral
convex set, and that will also persist here. Of special importance in that setting is
the critical cone associated with C at a point x with respect to a vector v ∈ NC (x),
defined by

(3) KC (x, v) = TC (x) ∩ [v]⊥ ,

which is always polyhedral

 convex as well. Recall here that for any vector v ∈ IRn


we denote [v] = τ v τ ∈ IR ; then [v] is a subspace of dimension 1 if v = 0 and 0
4 Regularity Properties Through Generalized Derivatives 225

if v = 0. Accordingly, [v]⊥ is a hyperplane through the origin if v = 0 and the whole

IRn when v = 0.
In this section we examine the isolated calmness property, which is inverse to
strong metric subregularity by 3I.2.

Theorem 4E.1 (isolated calmness for variational inequalities). For the variational
inequality (1) and its solution mapping (2) under the assumption that the convex
set C is polyhedral, let x̄ ∈ S( p̄) and suppose that f is continuously differentiable
around ( p̄, x̄). Let A = ∇x f ( p̄, x̄) and let K = KC (x̄, v̄) be the corresponding critical
cone in (3) for v̄ = − f ( p̄, x̄). If

(4) (A + NK )−1 (0) = {0},

then the solution mapping S has the isolated calmness property at p̄ for x̄ with

clm (S; p̄| x̄) ≤ |(A + NK )−1 | · |∇ p f ( p̄, x̄)|.

+

Moreover, under the ample parameterization condition rank ∇ p f ( p̄, x̄) = n, the
property in (4) is not just sufficient but also necessary for S to have the isolated
calmness property at p̄ for x̄.
Proof. Utilizing the specific form of the graphical derivative established in 4A.4
and the equivalence relation 4A(13) in 4A.6, we see that (4) is equivalent to the
condition 4C(7) in Corollary 4C.3. Everything then follows from the claim of that
corollary.

Exercise 4E.2 (alternative cone condition). In terms of the cone K ∗ that is polar to
K , show that the condition in (4) is equivalent to

(5) w ∈ K, −Aw ∈ K ∗ , w ⊥ Aw =⇒ w = 0.

Guide. Make use of 2A.3.

In the important case when C = IRn+ , the variational inequality (1) turns into the
complementarity relation

(6) x ≥ 0, f (p, x) ≥ 0, x ⊥ f (p, x).

This will serve to illustrate the result in Theorem 4E.1. Using the notation introduced
in Section 2E for the analysis of a complementarity problem, we associate with the
reference point (x̄, v̄) ∈ gph NIRn+ the index sets J1 , J2 and J3 in {1, . . . , n} given by
        
J1 = j  x̄ j > 0, v̄ j = 0 , J2 = j  x̄ j = 0, v̄ j = 0 , J3 = j  x̄ j = 0, v̄ j < 0 .

Then, by 2E.5, the critical cone K = KC (x̄, v̄) = TIRn+ (x̄) ∩ [ f ( p̄, x̄)]⊥ is described by
226 4 Regularity Properties Through Generalized Derivatives
⎧
⎨ w j free for i ∈ J1 ,
(7) w ∈ K ⇐⇒ wj ≥ 0 for i ∈ J2 ,
⎩w = 0 for i ∈ J3 .
j

Example 4E.3 (isolated calmness in complementarity problems). In the case of C =

IRn+ in which the variational inequality (1) reduces to the complementarity relation
(6) and the critical cone K is given by (7), the condition (4) in Theorem 4E.1 reduces
through (5) to having the following hold for the entries ai j of the matrix A. If w j for
j ∈ J1 ∪ J2 are real numbers satisfying

= 0 for i ∈ J1 and for i ∈ J2 with wi > 0,
w j ≥ 0 for j ∈ J2 and ∑ ai j w j
j∈J ∪J
≥ 0 for i ∈ J2 with wi = 0,
1 2

then w j = 0 for all j ∈ J1 ∪ J2 .

In the particular case when J1 = J3 = 0,
/ the matrices satisfying the condition in
4E.3 are called R0 -matrices1 .
As another application of Theorem 4E.1, consider the tilted minimization prob-
lem from Section 2G:

(8) minimize g(x) − v, x over x ∈ C,

where C is a nonempty polyhedral convex subset of IRn , v ∈ IRn is a parameter, and

the function g : IRn → IR is twice continuously differentiable everywhere. We first
give a brief summary of the optimality conditions from 2G.
If x is a local optimal solution of (8) for v then x satisfies the basic first-order
necessary optimality condition

(9) ∇g(x) + NC (x)  v.

Any solution x of (9) is a stationary point for problem (8), denoted S(v), and the
associated stationary point mapping is v → S(v) = (Dg + NC )−1 (v). The set of local
minimizers of (8) for v is a subset of S(v). If the function g is convex, every station-
ary point is not only local but also a global minimizer. For the variational inequality
(9), the critical cone to C associated with a solution x for v has the form

KC (x, v − ∇g(x)) = TC (x) ∩ [v − ∇g(x)]⊥.

If x furnishes a local minimum of (8) for v, then, according to 2G.1(a), x must satisfy
the second-order necessary condition

(10) u, ∇2 g(x)u ≥ 0 for all u ∈ KC (x, v − ∇g(x)).

In addition, from 2G.1(b), when x ∈ S(v) satisfies the second-order sufficient condi-
tion

1 For a detailed description of the classes of matrices appearing in the theory of linear complemen-
tarity problems, see the book Cottle, Pang and Stone [1992].
4 Regularity Properties Through Generalized Derivatives 227

(11) u, ∇2 g(x)u > 0 for all nonzero u ∈ KC (x, v − ∇g(x)),

then x is a local optimal solution of (8) for v. Having x to satisfy (9) and (11) is
equivalent to the existence of ε > 0 and δ > 0 such that

(12) g(y) − v, y ≥ g(x) − v, x + ε |y − x|2 for all y ∈ C with |y − x| ≤ δ ,

meaning by definition that x furnishes a strong local minimum in (8).

We know from 2G.3 that the stationary point mapping S has a Lipschitz localiza-
tion s around v for x with the property that s(u) furnishes a strong local minimum
for u near v if and only if the following stronger form of the second-order sufficient
optimality condition holds:
+
w, ∇2 g(x̄)w > 0 for all nonzero w ∈ KC (x, v),

where KC+ (x, v) = KC (x, v − ∇g(x)) − KC (x, v − ∇g(x)) is the critical subspace asso-
ciated with x and v. We now complement this result with a necessary and sufficient
condition for isolated calmness of S combined with local optimality at the reference
point.

Theorem 4E.4 (role of second-order sufficiency). Consider the stationary point

mapping S for problem (8), that is, the solution mapping for (9), and let x̄ ∈ S(v̄).
Then the following are equivalent:
(i) the second-order sufficient condition (11) holds at x̄ for v̄;
(ii) the point x̄ is a local minimizer of (7) for v̄ and the mapping S has the isolated
calmness property at v̄ for x̄.
Moreover, in either case, x̄ is actually a strong local minimizer of (7) for v̄.
Proof. Denote A := ∇2 g(x̄). According to Theorem 4E.1 complemented with 4E.2,
the mapping S has the isolated calmness property at v̄ for x̄ if and only if

(13) u ∈ K, −Au ∈ K ∗ , u ⊥ Au =⇒ u = 0,

where K = KC (x̄, v̄−∇g(x̄)). Let (i) hold. Then of course x̄ is a local optimal solution
as described. If (ii) doesn’t hold, there must exist some u = 0 satisfying the condi-
tions in the left side of (13), and that would contradict the inequality u, Au > 0 in
(11).
Conversely, assume that (ii) is satisfied. Then the second-order necessary condi-
tion (10) must hold; this can be written as

u∈K =⇒ − Au ∈ K ∗ .

The isolated calmness property of S at v̄ for x̄ is identified with (13), which in turn
eliminates the possibility of there being a nonzero u ∈ K such that the inequality
in (10) fails to be strict. Thus, the necessary condition (10) turns into the sufficient
condition (11). We already know that (11) implies (12), so the proof is complete.
228 4 Regularity Properties Through Generalized Derivatives

4F. Single-valued Localizations for Variational Inequalities

We now investigate applications of graphical derivatives to the issue of whether the
solution mapping S to the variational inequality

(1) f (p, x) + NC (x)  0

has a Lipschitz continuous single-valued localization. As in the preceding section,

we have a function f : IRd × IRn → IRn and a normal cone mapping NC associated
with a polyhedral convex set C ⊂ IRn . Our starting point is Theorem 2E.8 which we
state again for completeness.

Theorem 4F.1 (localization criterion under polyhedral convexity). For the solution
mapping S of (1) under the assumption that the convex set C is polyhedral, let x̄ ∈
S( p̄) and suppose that f is continuously differentiable near ( p̄, x̄). Let
  
(2) A = ∇x f ( p̄, x̄) and K = w ∈ TC (x̄)  w ⊥ f ( p̄, x̄) ,

noting that the critical cone K is likewise polyhedral convex. Suppose the mapping
A + NK has the property that

(3) s̄ := (A + NK )−1 is everywhere single-valued,

in which case s̄ is Lipschitz continuous globally (this being equivalent to strong

metric regularity of A + NK at 0 for 0). Then the solution mapping S has a Lipschitz
continuous single-valued localization s around p̄ for x̄ which is semidifferentiable
at p̄ with
Ds( p̄)(q) = s̄(−∇ p f ( p̄, x̄)q)
and moreover

(4) lip (s; p̄) ≤ lip (s̄; 0) · |∇ p f ( p̄, x̄)|.

In addition, under the ample parameterization condition rank ∇ p f ( p̄, x̄) = n the con-
dition in (3) is not just sufficient but also necessary for S to have a Lipschitz contin-
uous single-valued localization around p̄ for x̄.
Through graphical derivatives, we will actually be able to show that strong metric
regularity of A + NK is implied simply by metric regularity, or in other words, that
the invertibility condition in (3) follows already from (A + NK )−1 having the Aubin
property at 0 for 0, due to the special structure of the mapping A + NK .
Our tactic for bringing this out will involve applying Theorem 4B.1 to A + NK .
Before that can be done, however, we put some effort into a better understanding of
the normal cone mapping NK .

Faces of a cone. For a polyhedral convex cone K , a face is a set F of the form
4 Regularity Properties Through Generalized Derivatives 229

F = K ∩ [v]⊥ for some v ∈ K ∗ .

The collection of all faces of K will be denoted by FK .
Since K is polyhedral, we obtain from the Minkowski-Weyl theorem, 2E.2, and
its surroundings that FK is a finite collection of polyhedral convex cones. This
collection contains K itself and the zero cone, in particular.

Lemma 4F.2 (critical face lemma). Let C be a convex polyhedral set, let v ∈ NC (x)
and let K = KC (x, v) be the critical cone for C at (x, v),

K = TC (x) ∩ [v]⊥ .

Then there exists a neighborhood O of (x, v) such that for every choice of (x , v ) ∈
gph NC ∩ O the corresponding critical cone KC (x , v ) has the form

KC (x , v ) = F1 − F2

for some faces F1 , F2 in FK with F2 ⊂ F1 . In particular, KC (x , v ) ⊂ K − K for

every (x , v ) ∈ gph NC ∩ O. Conversely, for every two faces F1 , F2 in FK with F2 ⊂
F1 and every neighborhood O of (x, v) there exists (x , v ) ∈ gph NC ∩ O such that
KC (x , v ) = F1 − F2 .
Proof. Because C is polyhedral, all vectors of the form x = x − x with x ∈ C close
to x are the vectors x ∈ TC (x) having sufficiently small norm. Also, for such x ,

(5) TC (x ) = TC (x) + [x ] ⊃ TC (x)

and

(6) NC (x ) = NC (x) ∩ [x ]⊥ ⊂ NC (x).

Now, let (x , v ) ∈ gph NC be close to (x, v) and let x = x − x. Then from (5) we
have  
KC (x , v ) = TC (x ) ∩ [v ] = TC (x) + [x ] ∩ [v ]⊥ .
    ⊥ 

Further, from (6) it follows that v ⊥ x and then we obtain

(7) KC (x , v ) = TC (x) ∩ [v ]⊥ + [x ] = KC (x, v ) + [x ].

We will next show that KC (x, v ) ⊂ K for v sufficiently close to v. If this were not
so, there would be a sequence vk → v and another sequence wk ∈ KC (x, vk ) such that
wk ∈/ K for all k. Each set KC (x, vk ) is a face of TC (x), but since TC (x) is polyhedral,
the set of its faces is finite, hence for some face F of TC (x) we have KC (x, vk ) = F
for infinitely many k. Note that the set gph KC (x, ·) is closed, hence for any w ∈ F,
since (vk , w) is in this graph, the limit (v, w) belongs to it as well. But then w ∈ K
and since w ∈ F is arbitrarily chosen, we have F ⊂ K. Thus the sequence wk ∈ K
for infinitely many k, which is a contradiction. Hence KC (x, v ) ⊂ K.
230 4 Regularity Properties Through Generalized Derivatives

Let (x , v ) ∈ gph NC be close to (x, v). Relation (7) tells us that KC (x , v ) =
KC (x, v ) + [x ] for x = x − x. Let F1 = TC (x) ∩ [v ]⊥ , this being a face of TC (x).
The critical cone K = KC (x, v) = TC (x) ∩ [v]⊥ is itself a face of TC (x), and any face
of TC (x) within K is also a face of K. Then F1 is a face of the polyhedral cone K.
Let F2 be the face of F1 having x in its relative interior. Then F2 is also a face of K
and therefore KC (x , v ) = F1 − F2 , furnishing the desired representation.
Conversely, let F1 be a face of K. Then there exists v ∈ K ∗ = NK (0) such that
F1 = K ∩ [v ]⊥ . The size of v does not matter; hence we may assume that v + v ∈
NC (x) by the reduction lemma 2E.4. By repeating the above argument we have F1 =
TC (x) ∩ [v ]⊥ for v := v + v . Now let F2 be a face of F1 . Let x be in the relative
interior of F2 . In particular, x ∈ TC (x), so by taking the norm of x sufficiently small
we can arrange that the point x = x + x lies in C. We have x ⊥ v and, as in (7),
 
F1 −F2 = TC (x)∩[v ]⊥ +[x ] = TC (x)+[x ] ∩[v ]⊥ = TC (x )∩[v ]⊥ = KC (x , v ).

This gives us the form required.

Our next step is to specify the derivative criterion for metric regularity in 4B.1
for the reduced mapping A + NK in 4F.1.

Lemma 4F.3 (regularity modulus from derivative criterion). For A + NK with A and
K as in (2), we have

max |(A + NF1−F2 )−1 | .

−
(8) reg (A + NK ; 0|0) =
F1 ,F2 ∈FK
F1 ⊃F2

Thus, A + NK is metrically regular at 0 for 0 if and only if |(A + NF1 −F2 )−1 |− < ∞
for every F1 , F2 ∈ FK with F1 ⊃ F2 .
Proof. From Theorem 4B.1, combined with Example 4A.4, we have that

|(A + NTK (x)∩[y−Ax]⊥ )−1 | .

−
reg (A + NK ; 0|0) = lim sup
(x,y)→(0,0)
(x,y)∈gph (A+NK )

Lemma 4F.2 with (x, v) = (0, 0) gives us the desired representation NTK (x)∩[y−Ax]⊥ =
NF1 −F2 for (x, y) near zero and hence (8).
Example 4F.4 (critical faces for complementarity problems). Consider the comple-
mentarity problem
f (p, x) + NIRn+ (x)  0,
with a solution x̄ for p̄, with K and A as in (2) (with C = IRn+ ) and index sets
        
J1 = j  x̄ j > 0, v̄ j = 0 , J2 = j  x̄ j = 0, v̄ j = 0 , J3 = j  x̄ j = 0, v̄ j < 0

for v̄ = − f ( p̄, x̄). Then the cones F1 − F2 , where F1 and F2 are closed faces of K with
F1 ⊃ F2 , are the cones K̃ of the following form: There is a partition of {1, . . . , n} into
index sets J1 , J2 , J3 with
4 Regularity Properties Through Generalized Derivatives 231

J1 ⊂ J1 ⊂ J1 ∪ J2 , J3 ⊂ J3 ⊂ J2 ∪ J3 ,

such that
⎧ 
⎨ xi free for i ∈ J1 ,

(9) x ∈ F1 − F2 ⇐⇒ x ≥ 0 for i ∈ J2 ,
⎩ i
xi = 0 for i ∈ J3 .

Detail. Each face F of K has the form K ∩[v ]⊥ for some vector v ∈ K ∗ . The vectors
v in question are those with
⎧ 
⎨ vi = 0 for i ∈ J1 ,
v ≤ 0 for i ∈ J2 ,
⎩ i
vi free for i ∈ J3 .

The closed faces F of K correspond one-to-one therefore with the subsets of J2 : the
face F corresponding to an index set J2F consists of the vectors x such that
⎧ 
⎨ xi free for i ∈ J1 ,
x ≥ 0 for i ∈ J2 \ J2F ,
⎩ i
xi = 0 for i ∈ J3 ∪ J2F .

If F1 and F2 have J2F1 ⊂ J2F2 , so that F1 ⊃ F2 , then F1 − F2 is given by (9) with

J1 = J1 ∪ [J2 \ J2F2 ], J2 = J2F2 \ J2F1 , J3 = J3 ∪ J2F1 .

Exercise 4F.5 (critical face criterion for metric regularity). For a continuously
differentiable function f : IRn → IRn and a polyhedral convex set C ⊂ IRn , let
f (x̄) + NC (x̄)  0. Show that the mapping f + NC is metrically regular at x̄ for 0
if and only if, for all choices of faces F1 and F2 of the critical cone K to the set C at
x̄ for v̄ = − f (x̄), with F1 ⊃ F2 , the following condition holds with A = ∇ f (x̄):

∀ v ∈ IRn ∃ u ∈ F1 − F2 such that (v − Au) ∈ (F1 − F2 )∗ and (v − Au) ⊥ u.

Guide. From 3F.7, metric regularity of f + NC at x̄ for 0 is equivalent to metric

regularity of A + NK at 0 for 0. Apply 4F.3 then with the characterization of the
inner norm in 4A.6 and using the fact that u ∈ NK (w) whenever w ∈ K, u ∈ K ∗ , and
u ⊥ w.

Exercise 4F.6 (variational inequality over a subspace). Show that when the critical
cone K in 4F.5 is a subspace of IRn of dimension m ≤ n, then the matrix BABT is
nonsingular, where B is the matrix whose columns form an orthonormal basis in K .
Using some of the results obtained so far in this section, we will now prove
that, for a mapping appearing in particular in the Karush–Kuhn–Tucker (KKT) op-
timality conditions in nonlinear programming, metric regularity and strong metric
regularity are equivalent properties. Specifically, consider the standard nonlinear
programming problem
232 4 Regularity Properties Through Generalized Derivatives

= 0 for i ∈ [1, r],
(10) minimize g0 (x) over all x satisfying gi (x)
≤ 0 for i ∈ [r + 1, m]

with twice continuously differentiable functions gi : IRn → IR, i = 0, 1, . . . , m. The

associated KKT optimality system has the form

(11) f (x, y) + NE (x, y)  (0, 0),

where
⎛ ∇g (x) + m ∇g (x)y ⎞
0 ∑i=1 i i
⎜ −g1 (x) ⎟
(12) f (x, y) = ⎜
⎝ . ⎟
⎠
..
−gm(x)

and

(13) E = IRn × [IRr × IRm−r

+ ].

Theorem 2A.8 tells us that, under the constraint qualification condition 2A(14), for
any local minimum x of (10) there exists a Lagrange multiplier y, with yi ≥ 0 for
i = r + 1, . . . , m, such that (x, y) is a solution of (11). We will now establish an
important fact for the mapping on the left side of (11).

Theorem 4F.7 (KKT metric regularity implies strong metric regularity). Consider
the mapping F : IRn+m →
→ IRn+m defined as

(14) F : z → f (z) + NE (z)

with f as in (12) for z = (x, y) and E as in (13), and let z̄ = (x̄, ȳ) solve (11), that is,
F(z̄)  0. If F is metrically regular at z̄ for 0, then F is strongly metrically regular
there.
We already showed in Theorem 3G.5 that this kind of equivalence holds for lo-
cally monotone mappings, but here F need not be monotone even locally, although
it is a special kind of mapping in another way.
The claimed equivalence is readily apparent in a simple case of (10) when F is
an affine mapping, which corresponds to problem (10) with no constraints and with
g0 being a quadratic function, g0 (x) = 12 x, Ax + b, x for an n × n matrix A and a
vector b ∈ IRn . Then F(x, y) = Ax + b and metric regularity of F (at any point) means
that A has full rank. But then A must be nonsingular, so F is in fact strongly regular.
The general argument for F = f + NE is lengthy and proceeds through a se-
ries of reductions. First, since our analysis is local, we can assume without loss of
generality that all inequality constraints are active at x̄. Indeed, if for some index
i ∈ [r + 1, m] we have gi (x̄) < 0, then ȳi = 0. For q ∈ IRn+m consider the solution set
of the inclusion F(z)  q. Then for any q near zero and all x near x̄ we will have
gi (x) < qi , and hence any Lagrange multiplier y associated with such an x must have
4 Regularity Properties Through Generalized Derivatives 233

yi = 0; thus, for q close to zero the solution set of F(z)  q will not change if
we drop the constraint with index i. Further, if there exists an index i such that
ȳi > 0, then we can always rearrange the constraints so that ȳi > 0 for i ∈ [r + 1, s]
for some r < s ≤ m. Under these simplifying assumptions the critical cone K =
KE (z̄, v̄) to the set E in (13) at z̄ = (x̄, ȳ) for v̄ = − f (z̄) is the product IRn × IRs ×
IRm−s
+ . (Show that this form of the critical cone can be also derived by utilizing
Example 2E.5.) The normal cone mapping NK to the critical cone K has then the
form NK = {0}n × {0}s × N+m−s .
We next recall that metric regularity of F is equivalent to metric regularity of the
mapping
L : z → ∇ f (z̄)z + NK (z) for z = (x, y) ∈ IRn+m
at 0 for 0 and the same equivalence holds for strong metric regularity. This reduction
to a simpler situation has already been highlighted several times in this book, e.g. in
2E.8 for strong metric regularity and 3F.7 for metric regularity. Thus, to achieve our
goal of confirming the claimed equivalence between metric regularity and strong
regularity for F, it is enough to focus on the mapping L which, in terms of the
functions gi in (10), has the form
 
A BT
(15) L= + NK ,
−B 0

where ⎛ ⎞
m
∇x g1 (x̄)
⎜ ⎟
A = ∇2 g0 (x̄) + ∑ ∇2 gi (x̄)ȳi and B = ⎝ ..
. ⎠.
i=1
∇gm (x̄)
Taking into account the specific form of NK , the inclusion (v, w) ∈ L(x, y) becomes
⎧
⎨ v = Ax + BT y,
(16) (w + Bx)i = 0 for i ∈ [1, s],
⎩
(w + Bx)i ≤ 0, yi ≥ 0, yi (w + Bx)i = 0 for i ∈ [s + 1, m].

In further preparation for proving Theorem 4F.7, next we state and prove three
lemmas. From now on any kind of regularity is at 0 for 0, unless specified otherwise.

Lemma 4F.8 (KKT metric regularity implies strong metric subregularity). If the
mapping L in (15) is metrically regular, then it is strongly subregular.
Proof. Suppose that L is metrically regular. Then the critical face criterion dis-
played in 4F.5 with critical faces given in 4F.4 takes the following form: for ev-
ery partition J1 , J2 , J3 of {s + 1, . . . , m} and for every (v, w) ∈ IRn × IRm there exists
(x, y) ∈ IRn × IRm satisfying
234 4 Regularity Properties Through Generalized Derivatives
⎧
⎪
⎪ v = Ax + BT y,
⎪
⎪
⎨ (w + Bx)i = 0 for i ∈ [1, s],
(17) (w + Bx)i = 0 for i ∈ J1 ,
⎪
⎪
⎪
⎪ (w + Bx)i ≤ 0, yi ≥ 0, yi (w + Bx)i = 0 for i ∈ J2 ,
⎩
yi = 0 for i ∈ J3 .

In particular, denoting by B0 the submatrix of B composed by the first s rows of B,

for any index set J ⊂ {s+ 1, . . . , m}, including the empty set, if B(J) is the submatrix
of B whose rows have indices in J, then the condition involving (17) implies that
⎛ ⎞
A BT0 BT (J)
(18) the matrix N(J) = ⎝ −B0 0 0 ⎠ is nonsingular.
−B(J) 0 0

Indeed, to reach such a conclusion it is enough to take J = J1 and J2 = 0/ in (17). By
4E.1, the mapping L in (15) is strongly subregular if and only if

(19) the only solution of (17) with (v, w) = 0 is (x, y) = 0.

Now, suppose that L is not strongly subregular. Then, by (19), for some index set J ⊂
{s + 1, . . ., m}, possibly the empty set, there exists a nonzero vector (x, y) ∈ IRn × IRm
satisfying (17) for v = 0, w = 0. Note that this y has y j = 0 for j ∈ {s + 1, . . . , m} \ J.
But then the nonzero vector z = (x, y) with y having components in {1, . . . , s} × J
solves N(J)z = 0 where the matrix N(J) is defined in (18). Hence, N(J) is singular,
and then the condition involving (17) is violated; thus, the mapping L is not metri-
cally regular. This contradiction means that L must be strongly subregular.

The next two lemmas present general facts that are separate from the specific cir-
cumstances of nonlinear programming problem (10) considered. The second lemma
is a simple consequence of Brouwer’s invariance of domain theorem 1F.1:

Lemma 4F.9 (single-valued localization from continuous local selection). Let f :

IRn → IRn be continuous and let there exist an open neighborhood V of ȳ := f (x̄) and
a continuous function h : V → IRn such that h(y) ∈ f −1 (y) for y ∈ V . Then f −1 has
a single-valued graphical localization around ȳ for x̄.
Proof. Since f is a function, we have f −1 (y) ∩ f −1 (y ) = 0/ for any y, y ∈ V , y = y .
But then h is one-to-one and hence h−1 is a function defined in U := h(V ). Note
that x = h(h−1 (x)) ∈ f −1 (h−1 (x)) implies f (x) = h−1 (x) for all x ∈ U. Since h is
a function, we have that h−1 (x) = h−1 (x ) for all x, x ∈ U, x = x . Thus, f is one-
to-one on U, implying that the set f −1 (y) ∩ U consists of one point, h(y). Hence,
by Theorem 1F.1 applied to h, U is an open neighborhood of x̄. Therefore, h is a
single-valued graphical localization of f −1 around ȳ for x̄.

Lemma 4F.10 (properties of optimal solutions). Let ϕ : IRn → IR be a continuous

function and let Q : IRd →
→ IRn have the Aubin property at p̄ for x̄. Then any graphical
4 Regularity Properties Through Generalized Derivatives 235

localization around p̄ for x̄ of the solution mapping Sopt of the problem

minimize ϕ (x) subject to x ∈ Q(p)

is either multi-valued or a continuous function on a neighborhood of p̄.

Proof. Suppose that Sopt has a single-valued localization x̂(p) = Sopt (p) ∩U for all
p ∈ V for some neighborhoods U of x̄ and V of p̄ and, without loss of generality,
that Q has the Aubin property at p̄ for x̄ with the same neighborhoods U and V . Let
p ∈ V and V  pk → p as k → ∞. Since x̂(p) ∈ Q(p) ∩ U and x̂(pk ) ∈ Q(pk ) ∩ U,
there exist xk ∈ Q(pk ) and xk ∈ Q(p) such that xk → x̂(p) and also |xk − x̂(pk )| → 0
as k → ∞. From optimality,

ϕ (xk ) ≥ ϕ (x̂(pk )) and ϕ (xk ) ≥ ϕ (x̂(p)).

These two inequalities, combined with the continuity of ϕ , give us

ϕ (x̂(pk )) → ϕ (x̂(p)) as k → ∞.

Hence any limit of a sequence of minimizers x̂(pk ) is a minimizer for p, which

implies that x̂ is continuous at p.
We are now ready to complete the proof of 4F.7.
Proof of Theorem 4F.7 (final part). We already know from the argument displayed
after the statement of the theorem that metric regularity of the mapping F in (14) at
z̄ = (x̄, ȳ) for 0 is equivalent to the metric regularity of the mapping L in (15) at 0 for
0, and the same holds for the strong metric regularity. Our next step is to associate
with the mapping L the function
⎛ +
⎞
Ax + ∑si=1 bi yi + ∑m
i=s+1 bi yi
⎜ −b1 , x + y1 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
(20) H(x, y) = ⎜⎜ −bs , x + ys ⎟,
⎟
⎜ −
−bs+1 , x + ys+1 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
−bm , x + y−m

from IRn × IRm to itself, where bi are the rows of the matrix B and where we let
y+ = max{0, y} and y− = y − y+ .
For a given (v, u) ∈ IRn × IRm , let (x, y) ∈ H −1 (v, u). Then for zi = y+i , i =
s + 1, . . . , q, we have (x, z) ∈ L−1 (v, u). Indeed, for each i = s + 1, . . . , m, if yi ≤ 0,
then ui + bi , x = y− + −
i ≤ 0 and (ui + bi , x)yi = 0; otherwise ui + bi , x = yi = 0.
−1
Conversely, if (x, z) ∈ L (v, u) then for

z if zi > 0,
(21) yi = i
ui + bi , x if zi = 0,
236 4 Regularity Properties Through Generalized Derivatives

we obtain (x, y) ∈ H −1 (v, u). Thus, in order to achieve our goal for the mapping L,
we can focus on the same question for the equivalence between metric regularity
and strong metric regularity for the function H in (20).
Suppose that H is metrically regular but not strongly metrically regular. Then,
from 4F.8 and the equivalence between regularity properties of L and H, H is
strongly subregular. Consequently, its inverse H −1 has both the Aubin property and
the isolated calmness property, both at 0 for 0. In particular, since H is positively
homogeneous and has closed graph, for each w sufficiently close to 0, H −1 (w) is
a compact set contained in an arbitrarily small ball around 0. Let a > 0. For any
w ∈ aIB the problem

(22) minimize ym subject to (x, y) ∈ H −1 (w)

has a solution (x(w), y(w)) which, from the property of H −1 mentioned just above
(22), has a nonempty-valued graphical localization around 0 for 0. According to
Lemma 4F.10, this localization is either a continuous function or a multi-valued
mapping. If it is a continuous function, Lemma 4F.9 implies that H −1 has a con-
tinuous single-valued localization around 0 for 0. But then, since H −1 has the
Aubin property at that point, we conclude that H must be strongly metrically re-
gular, which contradicts the assumption made. Hence, any graphical localization
of the solution mapping of (22) is multi-valued. Thus, there exists a sequence
zk = (vk , uk ) → 0 and two sequences (xk , yk ) → 0 and (ξ k , η k ) → 0, whose k-terms
are both in H −1 (zk ), such that the m-components of yk and η k are the same, ykm = ηmk ,
but (xk , yk ) = (ξ k , η k ) for all k. Remove from yk the final component ykm and denote
the remaining vector by yk−m . Do the same for η k . Then (xk , yk−m ) and (ξ k , η−m
k ) are

both solutions of
s m−1
vk − bm ykm = Axk + ∑ bi yi + ∑ bi y+i
i=1 i=s+1
uk1 = −b1 , x + y1
..
.
uks = −bs , x + ys
..
.
uks+1 = −bs+1 , x + y−s+1
..
.
ukm−1 = −bm−1 , x + y−m−1 .

This relation concerns the reduced mapping H−m with m − 1 vectors bi , and accord-
ingly a vector y of dimension m − 1:
4 Regularity Properties Through Generalized Derivatives 237
⎛ ⎞
Ax + ∑si=1 bi yi + ∑i=s+1
m−1
bi y+i
⎜ −b1 , x + y1 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜
H−m (x, y) = ⎜ −bs , x + ys ⎟.
⎟
⎜ −
−bs+1 , x + ys+1 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ . ⎠
−bm−1 , x + y−m−1

We obtain that the mapping H−m cannot be strongly metrically regular because for
the same value zk = (vk − bmykm , uk−m ) of the parameter arbitrarily close to 0, we have
two solutions (xk , yk−m ) and (ξ k , η−m
k ). On the other hand, H
−m is metrically regular
as a submapping of H; this follows e.g. from the characterization in (17) for metric
regularity of the mapping L, which is equivalent to the metric regularity of H−m if
we choose J3 in (17) always to include the index m.
Thus, our assumption for the mapping H leads to a submapping H−m , of one less
variable y associated with the “inequality” part of L, for which the same assump-
tion is satisfied. By proceeding further with “deleting inequalities” we will end up
with no inequalities at all, and then the mapping L becomes just the linear mapping
represented by the square matrix
 
A BT 0 .
B0 0

But this linear mapping cannot be simultaneously metrically regular and not strongly
metrically regular, because a square matrix of full rank is automatically nonsingu-
lar. Hence, our assumption that the mapping H is metrically regular and not strongly
regular is void.

Exercise 4F.11. Find a formula for the metric regularity modulus of the mapping F
in (14).
Guide. The regularity modulus of F at z̄ for 0 equals the regularity modulus of the
mapping L in (15) at 0 for 0. To find a formula for the latter, utilize Lemma 4F.3.

4G. Special Nonsmooth Inverse Function Theorems

At the very beginning of this book, in Chapter 1, we were occupied with the classical
idea of inverting a function f : IRn → IRn locally, in the pattern of solving f (x) = y in
a localized sense for x in terms of y. The notion of a single-valued localization of f −1
was introduced and studied under various assumptions about differentiability. After
that, we moved on in Chapter 2 to single-valued localizations of solution mappings
238 4 Regularity Properties Through Generalized Derivatives

to generalized equations instead of just equations, and in Chapter 3 to results about

solution mappings beyond single-valued localizations. Now, though, we can return
to the original setting with new tools and describe two additional theorems in the
inverse function mode. These theorems make use of other concepts of generalized
differentiation, beyond the ones introduced in 4A.
In order to introduce the first of these alternative concepts, we have to rely on
a theorem by Rademacher, according to which a function f : IRn → IRm that is
Lipschitz continuous on an open set O is differentiable almost everywhere in O,
hence at a set of points that is dense in O. If κ is a Lipschitz constant, then from the
definition of Jacobian ∇ f (x) we have |∇ f (x)| ≤ κ at those points of differentiabil-
ity. Recall that a set C ⊂ IRn is said to be dense in the closed set D when cl C = D,
or equivalently, when for any x ∈ D any neighborhood U of x contains elements of
C. One of the simplest examples of a dense set is the set of rational numbers relative
to the set of real numbers.
Consider now any function f : IRn → IRm and any point x̄ ∈ dom f where
lip ( f ; x̄) < ∞. For any κ > lip ( f ; x̄) we have f Lipschitz continuous with constant
κ in some neighborhood U of x̄, and hence, from Rademacher’s theorem, there is a
dense set of points x in U where f is differentiable with |∇ f (x)| ≤ κ . Hence there
exist sequences xk → x̄ such that f is differentiable at xk , in which case the corre-
sponding sequence of norms |∇ f (xk )| is bounded by the Lipschitz constant κ and
hence has at least one cluster point. This leads to the following definition.

Generalized Jacobian. For f : IRn → IRm and any x̄ ∈ dom f where lip ( f ; x̄) < ∞,
denote by ∇
¯ f (x̄) the set consisting of all matrices A ∈ IRm×n for which there is a
sequence of points xk → x̄ such that f is differentiable at xk and ∇ f (xk ) → A. The
Clarke generalized Jacobian of f at p̄ is the convex hull of this set: co ∇¯ f (x̄).

Note that ∇ ¯ f (x̄) is a nonempty, closed, bounded subset of IRm×n . This ensures
that the convex set co ∇ ¯ f (x̄) is nonempty, closed, and bounded as well. Strict dif-
ferentiability of f at x̄ is known to be characterized by having ∇ ¯ f (x̄) consist of a
single matrix A (or equivalently by having co ∇ f (x̄) consist of a single matrix A),
¯
in which case A = ∇ f (x̄).
The inverse function theorem based on this notion, which we state next without
proof2 , says roughly that a Lipschitz continuous function can be inverted when all
elements of the generalized Jacobian are nonsingular. Compared with the classical
inverse function theorem, the main difference is that the single-valued graphical
localization so obtained can only be claimed to be Lipschitz continuous.

Theorem 4G.1 (Clarke’s inverse function theorem). Consider f : IRn → IRn and a
point x̄ ∈ int dom f where lip ( f ; x̄) < ∞. Let ȳ = f (x̄). If all of the matrices in the
generalized Jacobian co ∇¯ f (x̄) are nonsingular, then f −1 has a Lipschitz continuous
single-valued localization around ȳ for x̄.
For illustration, we provide some elementary cases.

2 Cf. Clarke [1976].

4 Regularity Properties Through Generalized Derivatives 239

Examples. The function f : IR → IR given by


x + x3 for x < 0,
f (x) =
2x − x2 for x ≥ 0

has generalized Jacobian co ∇ ¯ f (0) = [1, 2], which does not contain 0. According to
−1
Theorem 4G.1, f has a Lipschitz continuous single-valued localization around 0
for 0.
In contrast, the function f : IR → IR given by f (x) = |x| has co ∇¯ f (0) = [−1, 1],
which does contain 0. Although the theorem makes no claims about this case, there
is no graphical localization of f −1 around 0 for 0 that is single-valued.
A simple 2-dimensional example3 is with
 
|x1 | + x2
f (x) = ,
2x1 + |x2 |

for which
  
λ 1
co ∇
¯ f (0, 0) = | −1 ≤ λ ≤ 1, −1 ≤ τ ≤ 1 .
2 τ

This set of matrices does not contain a singular matrix, and hence 4G.1 can be
applied.
A better hold on the existence of single-valued Lipschitz continuous localizations
can be gained through a different version of graphical differentiation.

Strict graphical derivative. For a function f : IRn → IRm and any point x̄ ∈ dom f ,
the strict graphical derivative at x̄ is the set-valued mapping D∗ f (x̄) : IRn → → IRm
defined by
  
 f (xk + t k uk ) − f (xk )
D∗ f (x̄)(u) = w  ∃ (t k , xk , uk ) → (0, x̄, u) with w = lim .
k→∞ tk

When lip ( f ; x̄) < ∞, the set D∗ f (x̄)(u) is nonempty, closed and bounded in IRm
for each u ∈ IRn . Then too, the definition of D∗ f (x̄)(u) can be simplified
 by tak-
ing uk ≡ u. In this Lipschitzian setting it can be shown that D∗ f (x̄)u = Au  A ∈

co ∇¯ f (x̄) for all u if m = 1, but that fails for higher dimensions. In general, it is
known4 that for a function f : IRn → IRm with lip ( f ; x̄) < ∞, one has

(1) co ∇
¯ f (x̄)(u) ⊃ D∗ f (x̄)(u) for all u ∈ IRn .

Note that f is strictly differentiable at x̄ if and only if D∗ f (x̄) is a linear map-

ping, with the matrix for that mapping then being ∇ f (x̄). Anyway, strict graphical
derivatives can be used without having to assume even that lip ( f ; x̄) < ∞.
3 This example is from Clarke [1983].
4 Cf. Klatte and Kummer [2002], Section 6.3.
240 4 Regularity Properties Through Generalized Derivatives

The computation of strict graphical derivatives will be illustrated now in a special

case of nonsmoothness which has a basic role in various situations.
Example 4G.2. Consider the function

θ + : x → x+ := max{x, 0}, x ∈ IR.

Directly from the definition, we have, for any real u, that

%
u 
  for x̄ > 0,
(2) D∗ θ (x̄)(u) =
+
λ u  λ ∈ [0, 1] for x̄ = 0,
0 for x̄ < 0.
Similarly, the function

θ − : x → x− := min{x, 0}, x ∈ IR

satisfies
θ − (x) = x − θ + (x).
Then, just by applying the definition, we get

v ∈ D∗ θ + (x̄)(u) ⇐⇒ (u − v) ∈ D∗ θ − (x̄)(u) for any real u.

Equipped with strict graphical derivatives, we are now able to present a general-
ization of the classical inverse function theorem which furnishes a complete charac-
terization of the existence of a Lipschitz continuous localization of the inverse, and
thus sharpens the theorem of Clarke.

Theorem 4G.3 (Kummer’s inverse function theorem). Let f : IRn → IRn be contin-
uous around x̄, with f (x̄) = ȳ. Then f −1 has a Lipschitz continuous single-valued
localization around ȳ for x̄ if and only if

(3) 0 ∈ D∗ f (x̄)(u) =⇒ u = 0.

Proof. Recall Theorem 1F.2, which says that for a function f : IRn → IRn that is
continuous around x̄, the inverse f −1 has a Lipschitz continuous localization around
f (x̄) for x̄ if and only if, in some neighborhood U of x̄, there is a constant c > 0 such
that

(4) c|x − x| ≤ | f (x ) − f (x)| for all x , x ∈ O.

We will show first that (3) implies (4), from which the sufficiency of the condi-
tion will follow. With the aim of reaching a contradiction, let us assume there are
sequences ck → 0, xk → x̄ and x̃k → x̄ such that

| f (xk ) − f (x̃k )| ≤ ck |xk − x̃k |.

Then the sequence of points

4 Regularity Properties Through Generalized Derivatives 241

x̃k − xk
uk :=
|xk − x̃k |
satisfies |uk | = 1 for all k, hence a subsequence uki of it is convergent to some u = 0.
Restricting ourselves to such a subsequence, we obtain for t ki = |xki − x̃ki | that

f (xki + t ki uki ) − f (xki )

lim = 0.
ki →∞ t ki

By definition, the limit on the left side belongs to D∗ f (x̄)(u), yet u = 0, which is
contrary to (3). Hence (3) does imply (4).
For the converse, we argue that if (3) were violated, there would be sequences
t k → 0, xk → x̄, and uk → u with u = 0, for which

f (xk + t k uk ) − f (xk )
(5) lim = 0.
k→∞ tk
On the other hand, under (4) however, one has

| f (xk + t k uk ) − f (xk )|
≥ c|uk |,
tk
which combined with (5) and the assumption that uk is away from 0 leads to an
absurdity for large k. Thus (4) guarantees that (3) holds.
The property recorded in (1) indicates clearly that Clarke’s theorem follows from
that of Kummer. However, although the characterization of Lipschitz invertibility in
Kummer’s theorem looks simple, the price to be paid still lies ahead: we have to be
able to calculate the strict graphical derivative in every case of interest. This task
could be quite hard without calculus rules.
A rule that immediately follows from the definitions, at least, is the following.

Exercise 4G.4 (strict graphical derivatives of a sum). For a function f1 : IRn → IRm
that is strictly differentiable at x̄ (or, in particular, as a special case, continuously
differentiable in a neighborhood of x̄), and a function f2 : IRn → IRm that is Lipschitz
continuous around x̄, one has

D∗ ( f1 + f2 )(x̄)(u) = D f1 (x̄)u + D∗ f2 (x̄)(u).

Guide. This can be deduced right from the definitions.

Without going any further now into such rules, as developed in nonsmooth anal-
ysis, we devote the rest of this section to applying Kummer’s theorem to a specific
situation which we studied in Chapter 2.
Consider the following nonlinear programming problem with inequality con-
straints and special canonical perturbations

(6) minimize g0 (x) − v, x over all x satisfying gi (x) ≤ ui for i ∈ [1, m],
242 4 Regularity Properties Through Generalized Derivatives

where the functions gi : IRn → IR, i = 0, 1, . . . , m are twice continuously differentiable

and v ∈ IRn , u = (u1 , . . . , um )T ∈ IRm are parameters. According to the basic first-
order optimality conditions established in Section 2A, if x is a solution to (6) and
the constraint qualification condition 2A(13) holds, then there exists a multiplier
vector y = (y1 , . . . , ym ) such that the pair (x, y) satisfies the Karush–Kuhn–Tucker
conditions 2A(23), which are in the form of the variational inequality
 
−v + ∇g0 (x) + y∇g(x)
(7) ∈ NE (x, y) for E = IRn × IRm +,
−u + g(x)

where we put ⎛ ⎞ ⎛ ⎞
g1 (x) y1
⎜ ⎟ ⎜ . ⎟
g(x) = ⎝ ... ⎠ and y = ⎝ .. ⎠ .
gm (x) ym
More conveniently, for the mapping
 
∇g0 (x) + y∇g(x)
(8) G : (x, y) → + NE (x, y),
−g(x)

the solution mapping of (7) is just G−1 (here, without changing anything, we take the
negative of the first row since the normal cone to IRn is the zero mapping). We now
focus on inverting the mapping G. Choose a reference value (v̄, ū) of the parameters
and let (x̄, ȳ) solve (7) for (v̄, ū), that is, (v̄, ū) ∈ G(x̄, ȳ).
To apply Kummer’s theorem, we convert, as in the final part of the proof of 4F.7,
the variational inequality (7) into an equation involving the function H : IRn+m →
IRn+m defined as follows:
⎛ ⎞
∇g0 (x) + ∑m i=1 yi ∇gi (x)
+

⎜ −g1 (x) + y−1 ⎟

⎜ ⎟
(9) H(x, y) = ⎜ .. ⎟.
⎝ . ⎠
−gm (x) + ym −

In Section 4F we showed in particular that strong metric regularity of the mapping in

4F(15) is equivalent to strong metric regularity of the associated mapping in 4F(20).
The same argument works for the mappings in (8) and (9) where we now have non-
linear functions; for completeness, we will repeat it here. If ((x, y), (v, u)) ∈ gph H
then for zi = y+i , i = 1, . . . , m, we have that ((x, z), (v, u)) ∈ gph G. Indeed, for each
i = 1, . . . , m, if yi ≤ 0, then ui + gi (x) = y− +
i ≤ 0 and (ui + gi (x))yi = 0; otherwise
−
ui + gi (x) = yi = 0. Conversely, if ((x, z), (v, u)) ∈ gph G, then for

z if zi > 0,
(10) yi = i
ui + gi (x) if zi = 0,

we obtain (x, y) ∈ H −1 (v, u). In particular, if H −1 has a Lipschitz continuous local-

ization around (v̄, ū) for (x̄, ȳ) then G−1 has the same property at (v̄, ū) for (x̄, z̄)
4 Regularity Properties Through Generalized Derivatives 243

where z̄i = ȳ+i , and if G−1 has a Lipschitz continuous localization around (v̄, ū) for
(x̄, z̄) then H −1 has the same property at (v̄, ū) for (x̄, ȳ), where ȳ satisfies (10).
To invoke Kummer’s theorem for the function H, we need to determine the strict
graphical derivative of H. There is no trouble in differentiating the expressions
−gi (x) + y−i , inasmuch as we already know from 4G.2 the strict graphical deriva-
tive of y− . A little bit more involved is the determination of the strict graphical
derivative of ϕi (x, y) := ∇gi (x)y+i for i = 1, . . . , m. Adding and subtracting the same
expressions, passing to the limit as in the definition, and using (2), we obtain

z ∈ D∗ ϕi (x̄, ȳ)(u, v) ⇐⇒ z = ȳ+i ∇2 gi (x̄)u + λivi ∇gi (x̄), i = 1, . . . , m,

where the coefficients λi for i = 1, . . . , m satisfy

%= 1 for ȳi > 0,
(11) λi ∈ [0, 1] for ȳi = 0,
=0 for ȳi < 0.
Taking into account 4G.4, the form thereby obtained for the strict graphical deriva-
tive of the function H in (9) at (x̄, ȳ) is as follows:

ξ = Au + ∑m i=1 λi vi ∇gi (x̄),
(ξ , η ) ∈ D∗ H(x̄, ȳ)(u, v) ⇐⇒
ηi = −∇gi (x̄)u + (1 − λi)vi for i = 1, . . . , m,

where the λi ’s are as in (11), and

A = ∇2 g0 (x̄) + ∑ ȳ+i ∇2 gi (x̄).

Denoting by Λ the m × m diagonal matrix with elements λi on the diagonal, by

Im the m × m identity matrix, and setting
⎛ ⎞
∇g1 (x̄)
⎜ ⎟
B = ⎝ ... ⎠ ,
∇gm (x̄)

we obtain that
 
A BT Λ
(12) M(Λ ) ∈ D∗ H(x̄, ȳ) ⇐⇒ M(Λ ) = .
−B Im − Λ

This formula can be simplified by re-ordering the functions gi according to the sign
of ȳi . We first introduce some notation. Let I = {1, . . . , m} and, without loss of gen-
erality, suppose that
     
i ∈ I  ȳi > 0 = {1, . . . , k} and i ∈ I  ȳi = 0 = {k + 1, . . ., l}.

Let
244 4 Regularity Properties Through Generalized Derivatives
⎛ ⎞ ⎛ ⎞
∇g1 (x̄) ∇gk+1 (x̄)
⎜ ⎟ ⎜ ⎟
B+ = ⎝ ... ⎠ and B0 = ⎝ ..
. ⎠,
∇gk (x̄) ∇gl (x̄)
let Λ0 be the (l − k) × (l − k) diagonal matrix with diagonal elements λi ∈ [0, 1], let
I0 be the identity matrix for IRl−k , and let Im−l be the identity matrix for IRm−l . Then,
since λi = 1 for i = 1, . . . , k and λi = 0 for i = l + 1, . . . , m, the matrix M(Λ ) in (12)
takes the form ⎛ ⎞
A BT + BT0 Λ0 0
⎜ 0 0 0 ⎟
M(Λ0 ) = ⎝ ⎠.
−B 0 I0 − Λ0 0
0 0 Im−l
Each column of M(Λ0 ) depends on at most one λi , hence there are numbers

ak+1 , bk+1 , . . . , al , bl

such that
det M(Λ0 ) = (ak+1 + λk+1 bk+1 ) · · · (al + λl bl ).
Therefore, det M(Λ0 ) = 0 for all λi ∈ [0, 1], i = k + 1, . . . , l, if and only if the follow-
ing condition holds:

ai = 0, ai + bi = 0 and [sign ai = sign (ai + bi) or sign bi = sign (ai + bi )],

for i = k + 1, . . ., l.

Here we invoke the convention that

%
1 for a > 0,
sign a = 0 for a = 0,
−1 for a < 0.
One can immediately note that it is not possible to have simultaneously sign bi =
sign (ai + bi ) and sign ai = sign (ai + bi ) for some i. Therefore, it suffices to have

(13) ai = 0, ai + bi = 0, sign ai = sign (ai + bi ) for all i = k + 1, . . . , l.

Now, let J be a subset of {k + 1, . . ., l} and for i = k + 1, . . ., l, and let


1 for i ∈ J,
λiJ =
0 otherwise.
Let Λ J be the diagonal matrix composed by these λiJ , and let B0 (J) = Λ J B0 . Then
we can write
⎛ ⎞
A BT + B0 (J)T 0
⎜ 0 0 0⎟
M(J) := M(Λ J ) = ⎝ ⎠,
−B 0 I0J 0
0 0 Il
4 Regularity Properties Through Generalized Derivatives 245

where I0J is the diagonal matrix having 0 as (i−k)-th element if i ∈ J and 1 otherwise.
Clearly, all the matrices M(J) are obtained from M(Λ0 ) by taking each λi either 0
or 1. The condition (13) can be then written equivalently as

(14) detM(J) = 0 and sign det M(J) is the same for all J.

Let the matrix B(J) have as rows the row vectors ∇gi (x̄) for i ∈ J. Reordering the
last m − k columns and rows of M(J), if necessary, we obtain
 
A BT B(J)T 0
M(J) = +
,
−B 0 0 I

where I is now the identity for IR{k+1,...,m}\J . The particular form of the matrix M(J)
implies that M(J) fulfills (14) if and only if (14) holds for just a part of it, namely
for the matrix
⎛ ⎞
H BT+ B(J)T
(15) N(J) := ⎝ −B+ 0 0 ⎠.
−B(J) 0 0

By applying Kummer’s theorem, we arrive finally at the following result, which

sharpens the first part of the statement of Theorem 2G.8.

Theorem 4G.5 (strong regularity characterization for KKT mappings). The solu-
tion mapping of the Karush–Kuhn–Tucker variational inequality (7) has a Lipschitz
continuous single-valued localization around (v̄, ū) for (x̄, ȳ) if andonly if,
 for the

matrix N(J) in (15), det N(J) has the same nonzero sign for all J ⊂ i ∈ I  ȳi = 0 .
We should note that, in this specific case, the set of matrices D∗ H(x̄, ȳ) happens to
be convex, and then D∗ H(x̄, ȳ) coincides with its generalized Jacobian co ∇H( ¯ x̄, ȳ).

4H. Results Utilizing Coderivatives

Graphical derivatives, defined in terms of tangent cones to graphs, have been the
mainstay for most of the developments in this chapter so far, with the exception of
the variants in 4G. However, an alternative approach can be made to many of the
same issues in terms of graphical coderivatives, defined instead in terms of normal
cones to graphs. This theory is readily available in other texts in variational analysis,
so we will only lay out the principal ideas and facts here without going into their
detailed development.
Normal cones NC (x) have already been prominent, of course, in our work with
optimality conditions and variational inequalities, starting in Section 2A, but only
in the case of convex sets C. To arrive at coderivatives for a mapping F : IRn → → IRm ,
246 4 Regularity Properties Through Generalized Derivatives

we wish to make use of normal cones to gph F at points (x, y), but to keep the door
open to significant applications we need to deal with graph sets that are not convex.
The first task, therefore, is generalizing NC (x) to the case of nonconvex C.

General normal cones. For a set C ⊂ IRn and a point x ∈ C at which C is locally
closed, a vector v is said to be a regular normal if v, x − x ≤ o(|x − x|) for x ∈ C.
The set of all such vectors v is called the regular normal cone to C at x and is denoted
by N̂C (x). A vector v is said to be a general normal to C at x if there are sequences
{xk } and {vk } with xk ∈ C, such that

xk → x and vk → v with vk ∈ N̂C (xk ).

The set of all such vectors v is called the general normal cone to C at x and is
denoted by NC (x). For x ∈
/ C, NC (x) is taken to denote the empty set.
Very often, the limit process in the definition of the general normal cone NC (x)
is superfluous: no additional vectors v are produced in that manner, and one merely
has NC (x) = N̂C (x). This circumstance is termed the Clarke regularity of C at x.
When C is convex, for instance, it is Clarke regular at every one of its points x, and
the generalized normal cone NC (x) agrees with the normal cone defined earlier, in
2A. Anyway, NC (x) is always a closed cone.

Coderivatives of mappings. For a mapping F : IRn → → IRm and a pair (x, y) ∈ gph F
at which gph F is locally closed, the coderivative of F at x for y is the mapping
D∗ F(x|y) : IRm →
→ IRn defined by

w ∈ D∗ F(x|y)(z) ⇐⇒ (w, −z) ∈ Ngph F (x, y).

Obviously this is a “dual” sort of notion, but where does it fit in with classi-
cal differentiation? The answer can be seen by specializing to the case where F is
single-valued, thus reducing to a function f : IRn → IRm . Suppose f is strictly dif-
ferentiable at x; then for y = f (x), the graphical derivative D f (x|y) is of course
the linear mapping from IRn to IRm with matrix ∇ f (x). In contrast, the coderiva-
tive D∗ f (x|y) comes out as the adjoint linear mapping from IRm to IRn with matrix
∇ f (x)T .
The most striking fact about coderivatives in our context is the following simple
characterization of metric regularity.

Theorem 4H.1 (coderivative criterion for metric regularity). For a mapping F :

IRn →
→ IRm and a pair (x̄, ȳ) ∈ gph F at which gph F is locally closed, one has

reg (F; x̄| ȳ) = |D∗ F(x̄| ȳ)−1 | .

+
(1)

Thus, F is metrically regular if and only if the right side of (1) is finite, which is
equivalent to

(2) D∗ F(x̄| ȳ)(u)  0 =⇒ u = 0.

4 Regularity Properties Through Generalized Derivatives 247

If F is single-valued, that is, a function f : IRn → IRm which is strictly differen-

tiable at x̄, then the coderivative criterion means that the adjoint to the derivative
mapping D f (x̄) is injective, that is, ker ∇ f (x)T = {0}. This is equivalent to surjec-
tivity of D f (x̄) which, as we know from 3F or 4B, is equivalent to metric regularity
of f at x̄ for f (x̄).
We will now apply the coderivative criterion in 4H.1 to the variational inequality

(3) f (p, x) + NC (x)  0,

where f : IRd × IRn → IRn and C is a polyhedral convex set in IRn . Let x̄ be a solution
of (2) for p̄ and let f be continuously differentiable around ( p̄, x̄). From Theorem
4F.1 we know that a sufficient condition for the existence of a Lipschitz single-
valued localization of the solution mapping S of (2) around p̄ for x̄ is the metric
regularity at 0 for 0 of the reduced mapping

(4) A + NK ,

where the linear mapping A and the critical cone K are

  
(5) A = ∇x f ( p̄, x̄) and K = KC (x̄, v̄) = w ∈ TC (x̄)  w ⊥ f ( p̄, x̄) .

Based on earlier results in 2E we also noted there that in the case of ample parame-
terization, this sufficient condition is necessary as well. Further, in 4F.3 we obtained
a characterization of metric regularity of (3) by means of the derivative criterion in
4B.1. We will now apply the coderivative criterion for metric regularity (2) to the
mapping in (4); for that purpose we have to compute the coderivative of the mapping
in (4). The first step to do that is easy and we will give it as an exercise.

Exercise 4H.2 (reduced coderivative formula). Show that, for a linear mapping A :
IRn → IRn and a closed convex cone K ⊂ IRn one has

D∗ (A + NK )(x̄| ȳ) = A∗ + D∗ NK (x̄| ȳ − Ax̄).

Guide. Apply the definition of the general normal cone above.

Thus, everything hinges on determining the coderivative D∗ NK (0|0) of the map-
ping NK at the point (0, 0) ∈ G = gph NK . By definition, the graph of the coderivative
mapping consists of all pairs (w, −z) such that (w, z) ∈ NG (0, 0) where NG is the gen-
eral normal cone to the nonconvex set G. In these terms, for A and K as in (5), the
coderivative criterion (2) becomes

(6) (u, AT u) ∈ NG (0, 0) =⇒ u = 0.

Everything depends then on determining NG (0, 0).

We will next appeal to the known fact5 that NG (0, 0) is the limsup of polar cones
TG (x, v)∗ at (x, v) ∈ G as (x, v) → (0, 0). Because G is the union of finitely many
5 See Proposition 6.5 in Rockafellar and Wets [1998].
248 4 Regularity Properties Through Generalized Derivatives

polyhedral convex sets in IR2n (due to K being polyhedral), only finitely many cones
can be manifested as TG (x, v) at points (x, v) ∈ G near (0, 0). Thus, for a sufficiently
small neighborhood O of the origin in IR2n we have that
,
(7) NG (0, 0) = TG (x, v)∗ .
(x,v)∈O∩G

It follows from
 reduction lemma 2E.4 that TG (x, v) = gph NK(x,v) , where K(x, v) =
  
x ∈ TK (x)  x ⊥ v . Therefore,
  
TG (x, v) = (x , v )  x ∈ K(x, v), v ∈ K(x, v)∗ , x ⊥ v ,

and we have
 6 7 
TG (x, v)∗ = (r, u)  (r, u), (x , v ) ≤ 0 for all (x , v ) ∈ TG (x, v)
 
= (r, u)  r, x  + u, v  ≤ 0 for all

x ∈ K(x, v), v ∈ K(x, v)∗ with x ⊥ v .

It is evident from this (first in considering v = 0, then in considering x = 0) that

actually

(8) TG (x, v)∗ = K(x, v)∗ × K(x, v).

Hence NG (0, 0) is the union of all product sets K̂ ∗ × K̂ associated with cones K̂ such
that K̂ = K(x, v) for some (x, v) ∈ G near enough to (0, 0).
It remains to observe that the form of the critical cones K̂ = K(x, v) at points (x, v)
close to (0, 0) is already derived in Lemma 4F.2, namely, for every choice of (x, v) ∈
gph K near (0, 0) (this last requirement is actually not needed) the corresponding
critical cone K̂ = K(x, v) is given by

(9) K̂ = F1 − F2 for some faces F1 , F2 ∈ FK with F2 ⊂ F1 ,

where F is the collection of all faces of K as defined in 4F. To see this all we need
to do is to replace C by K and (x, v) by (0, 0) in the proof of 4F.2. Summarizing,
from (6), (7), (8) and (9), and the coderivative criterion in 4H.1, we come to the
following result:

Lemma 4H.3 (regularity modulus from coderivative criterion). For the mapping in
(4)(5) we have
1
(10) reg (A + NK ; 0|0) = max sup .
F1 ,F2 ∈FK u∈F −F
1 2 d(AT u, (F1 − F2 )∗ )
F1 ⊃F2 |u|=1

Thus, A + NK is metrically regular at 0 for 0 if and only if for every choice of critical
faces F1 , F2 ∈ FK with F2 ⊂ F1 ,
4 Regularity Properties Through Generalized Derivatives 249

u ∈ F1 − F2 and AT u ∈ (F1 − F2 )∗ =⇒ u = 0.

Finally, relying on 4F.1, we arrive at the following characterization of the exis-

tence of Lipschitz single-valued localization of the solution mapping S of the varia-
tional inequality (3), which parallels the one in 4F.11:

Theorem 4H.4 (Lipschitz single-valued localization from a coderivative rule). For

the solution mapping S of (3) under the assumption that the convex set C is polyhe-
dral, let x̄ ∈ S( p̄), and suppose that f is continuously differentiable near ( p̄, x̄). For
the mapping in (4)(5), suppose that
1
(11) max sup ≤ λ < ∞.
F1 ,F2 ∈FK u∈F −F
1 2 d(AT u, (F1 − F2 )∗ )
F1 ⊃F2 |u|=1

Then the solution mapping S has a Lipschitz continuous single-valued localization

s around p̄ for x̄ with
lip (s; p̄) ≤ λ |∇ p f ( p̄, x̄)|.
Moreover, under the ample parameterization condition rank ∇ p f ( p̄, x̄) = n, the con-
dition in (11) is not just sufficient but also necessary for S to have a Lipschitz con-
tinuous single-valued localization around p̄ for x̄.
Proof. We utilize 4F.1, taking into account that, because metric regularity of A+ NK
is equivalent to strong metric regularity, the Lipschitz modulus of (A + NK )−1 equals
the regularity modulus of A + NK . We then apply the formula in (10).
250 4 Regularity Properties Through Generalized Derivatives

Commentary
Graphical derivatives of set-valued mappings were introduced by Aubin [1981];
for more, see Aubin and Frankowska [1990]. The material in sections 4B and 4C
is from Dontchev, Quincampoix and Zlateva [2006], where results of Aubin and
Frankowska [1987, 1990] were used.
The statement 4B.5 of the Ekeland principle is from Ekeland [1990]. A detailed
presentation of this principle along with various forms and extensions is given in
Borwein and Zhu [2005]. The proof of the classical implicit function theorem 1A.1
given at the end of Section 4B is close, but not identical, to that in Ekeland [1990].
The derivative criterion for metric subregularity in 4C.1 was obtained by
Rockafellar [1989], but the result itself was embedded in a proof of a statement re-
quiring additional assumptions. The necessity without those assumptions was later
noted in King and Rockafellar [1992] and in the case of sufficiency by Levy [1996].
The statement and the proof of 4C.1 are from Dontchev and Rockafellar [2004].
Sections 4E and 4F collect various results scattered in the literature. Theorem
4E.4 is from Dontchev and Rockafellar [2004] while the critical face lemma 4F.2
is a particular case of Lemma 3.5 in Robinson [1984]; see also Theorem 5.6 in
Rockafellar [1989]. Theorem 4F.6 is a particular case of Theorem 3 in Dontchev
and Rockafellar [1996] which in turn is based on a deeper result in Robinson [1992],
see also Ralph [1993]. The presented proof uses a somewhat modified version of a
reduction argument from the book Klatte and Kummer [2002], Section 7.5.
Clarke’s inverse function theorem, 4G.1, was first published in Clarke [1976]; for
more information regarding the generalized Jacobian see the book of Clarke [1983].
Theorem 4G.3 is from Kummer [1991]; see also Klatte and Kummer [2002] and
Páles [1997]. It is interesting to note that a nonsmooth implicit function theorem
which is a special case of both Clarke’s theorem and Kummer’s theorem, appeared
as early as 1916 in a paper by Hedrick and Westfall [1916].
Theorem 4G.5 originates from Robinson [1980]; the proof given here uses some
ideas from Kojima [1980] and Jongen et al. [1987].
The coderivative criterion in 4H.1 goes back to the early works of Ioffe [1981,
1984], Kruger [1982] and Mordukhovich [1984]. A broad review of the role of
coderivatives in variational analysis is given in Mordukhovich [2006].
Chapter 5
Regularity in Infinite Dimensions

The theme of this chapter has origins in the early days of functional analysis and
the Banach open mapping theorem, which concerns continuous linear mappings
from one Banach space to another. The graphs of such mappings are subspaces
of the product of the two Banach spaces, but remarkably much of the classical
theory extends to set-valued mappings whose graphs are convex sets or cones in-
stead of subspaces. Openness connects up then with metric regularity and interior-
ity conditions on domains and ranges, as seen in the Robinson–Ursescu theorem.
Infinite-dimensional inverse function theorems and implicit function theorems due
to Lyusternik, Graves, and Bartle and Graves can be derived and extended. Banach
spaces can even be replaced to some degree by more general metric spaces.
Before proceeding we review some notation and terminology. Already in the first
section of Chapter 1 we stated the contraction mapping principle in metric spaces.
Given a set X, a function ρ : X × X → IR+ is said to be a metric in X when
(i) ρ (x, y) = 0 if and only if x = y;
(ii) ρ (x, y) = ρ (y, x);
(iii) ρ (x, y) ≤ ρ (x, z) + ρ (z, y) (triangle inequality).
A set X equipped with a metric ρ is called a metric space (X, ρ ). In a metric space
(X, ρ ), a sequence {xk } is called a Cauchy sequence if for every ε > 0 there ex-
ists n ∈ IN such that ρ (xk , x j ) < ε for all k, j > n. A metric space is complete if
every Cauchy sequence converges to an element of the space. Any closed set in a
Euclidean space is a complete metric space with the metric ρ (x, y) = |x − y|.
A linear (vector) space over the reals is a set X in which addition and scalar
multiplication are defined obeying the standard algebraic laws of commutativity,
associativity and distributivity. A linear space X with elements x is normed if it
is furnished with a real-valued expression x, called the norm of x, having the
properties
(i) x ≥ 0, x = 0 if and only if x = 0;
(ii) α x = |α | x for α ∈ IR;
(iii) x + y ≤ x + y.
Any normed space is a metric space with the metric ρ (x, y) = x − y. A complete
normed vector space is called a Banach space. On a finite-dimensional space, all
A.L. Dontchev and R.T. Rockafellar, Implicit Functions and Solution Mappings: A View 251
from Variational Analysis, Springer Monographs in Mathematics,
DOI 10.1007/978-0-387-87821-8 5,  c Springer Science+Business Media, LLC 2009
252 5 Regularity in Infinite Dimensions

norms are equivalent, but when we refer specifically to IRn we ordinarily have in
mind the Euclidean norm denoted by | · |. Regardless of the particular norm being
employed in a Banach space, the closed unit ball for that norm will be denoted by
IB, and the distance from a point x to a set C will be denoted by d(x,C), and so forth.
As in finite dimensions, a function A acting from a Banach space X into a Banach
space Y is called a linear mapping if dom A = X and A(α x + β y) = α Ax + β Ay for
all x, y ∈ X and all scalars α and β . The range of a linear mapping A from X to
Y is always a subspace of Y , but it might not be a closed subspace, even if A is
continuous. A linear mapping A : X → Y is surjective if rge A = Y and injective if
ker A = {0}.
Although in finite dimensions a linear mapping A : X → Y is automatically con-
tinuous, this fails in infinite dimensions; neither does surjectivity of A when X = Y
necessarily yield invertibility, in the sense that A−1 is single-valued. However, if A
is continuous at any one point of X, then it is continuous at every point of X . That,
moreover, is equivalent to A being bounded, in the sense that A carries bounded
subsets of X into bounded subsets of Y , or what amounts to the same thing due to
linearity, the image of the unit ball in X is included in some multiple of the unit ball
in Y , i.e., the value
A = sup Ax
x≤1

is finite. This expression defines the operator norm on the space L (X ,Y ), consisting
of all continuous linear mappings A : X → Y , which is then another Banach space.
Special and important in this respect is the Banach space L (X, IR), consisting
of all linear and continuous real-valued functions on X. It is the space dual to X,
symbolized by X ∗ , and its elements are typically denoted by x∗ ; the value that an
x∗ ∈ X ∗ assigns to an x ∈ X is written as x∗ , x. The dual of the Banach space X ∗ is
the bidual X ∗∗ of X; when every function x∗∗ ∈ X ∗∗ on X ∗ can be represented as x∗ →
x∗ , x for some x ∈ X , the space X is called reflexive. This holds in particular when
X is a Hilbert space with x, y as its inner product, and each x∗ ∈ X ∗ corresponds
to a function x → x, y for some y ∈ X, so that X ∗ can be identified with X itself.
Another thing to be mentioned for a pair of Banach spaces X and Y and their
duals X ∗ and Y ∗ is that any A ∈ L (X ,Y ) has an adjoint A∗ ∈ L (Y ∗ , X ∗ ) such that
Ax, y∗  = x, A∗ y∗  for all x ∈ X and y∗ ∈ Y ∗ . Furthermore, A∗  = A. A gener-
alization of this to set-valued mappings having convex cones as their graphs will be
seen later.
In fact most of the definitions, and even many of the results, in the preceding
chapters will carry over with hardly any change, the major exception being results
with proofs which truly depended on the compactness of IB. Our initial task, in Sec-
tion 5A, will be to formulate various facts in this broader setting while coordinating
them with classical theory. In the remainder of the chapter, we present inverse and
implicit mapping theorems with metric regularity in abstract spaces. Parallel results
for metric subregularity are not considered.
5 Regularity in Infinite Dimensions 253

5A. Openness and Positively Homogeneous Mappings

At the end of Chapter 1 we introduced the concept of openness of a function and
presented a Jacobian criterion for openness, but did not elaborate further. We now
return to this property in the broader context of Banach spaces X and Y . A function
f : X → Y is called open at x̄ if x̄ ∈ int dom f and, for every neighborhood U of x̄ in
X , the set f (U) is a neighborhood of f (x̄) in Y . This definition extends to set-valued
mappings F : X → → Y ; we say that F is open at x̄ for ȳ, where ȳ ∈ F(x̄), if

x̄ ∈ int dom F and for every neighborhood U of x̄,

(1) 0
the set F(U) = x∈U F(x) is a neighborhood of ȳ.

We will also be concerned with another property, introduced for mappings F :

→ IRm in 2D but likewise directly translatable to mappings F : X →
IRn → → Y , namely
that F is metrically regular at x̄ for ȳ, where ȳ ∈ F(x̄), if

there exists κ > 0 with neighborhoods U of x̄ and V of ȳ

(2)
such that d(x, F −1 (y)) ≤ κ d(y, F(x)) for all (x, y) ∈ U × V.

As before, the infimum of all such κ associated with choices of U and V is denoted
by reg (F; x̄| ȳ) and called the modulus of metric regularity of F at x̄ for ȳ.
The classical theorem about openness only addresses linear mappings. There are
numerous versions of it available in the literature; we provide the following formu-
lation:

Theorem 5A.1 (Banach open mapping theorem). For any A ∈ L (X,Y ) the follow-
ing properties are equivalent:
(a) A is surjective;
(b) A is open (at every point);
(c) 0 ∈ int A(intIB);
(d) there is a κ > 0 such that for all y ∈ Y there exists x ∈ X with Ax = y and
x ≤ κ y.
This theorem will be derived in Section 5B from a far more general result about
set-valued mappings F than just linear mappings A. Our immediate interest lies in
connecting it with the ideas in previous chapters, so as to shed light on where we
have arrived and where we are going.
The first observation to make is that (d) of Theorem 5A.1 is the same as the
existence of a κ > 0 such that d(0, A−1 (y)) ≤ κ y for all y. Clearly (d) does imply
this, but the converse holds also by passing to a slightly higher κ if need be. But
the linearity of A can also be brought in. For x ∈ X and y ∈ Y in general, we have
d(x, A−1 (y)) = d(0, A−1 (y) − x), and since z ∈ A−1 (y) − x corresponds to A(x + z) =
y, we have d(0, A−1 (y) − x) = d(0, A−1 (y − Ax)) ≤ κ y − Ax. Thus, (d) of Theorem
5A.1 is actually equivalent to:

(3) there exists κ > 0 such that d(x, A−1 (y)) ≤ κ d(y, Ax) for all x ∈ X, y ∈ Y.
254 5 Regularity in Infinite Dimensions

Obviously this is the same as the metric regularity property in (2) as specialized to
A, with the local character property becoming global through the arbitrary scaling
made available because A(λ x) = λ Ax. In fact, due to linearity, metric regularity of
A with respect to any pair (x̄, ȳ) in its graph is identical to metric regularity with
respect to (0, 0), and the same modulus of metric regularity prevails everywhere.
We can simply denote this modulus by reg A and use the formula that

(4) reg A = sup d(0, A−1 (y)) for A ∈ L (X,Y ).

y≤1

What we see then is that the condition

(e) A is metrically regular (everywhere): reg A < ∞
could be added to the equivalences in Theorem 5A.1 as a way of relating it to the
broader picture we now have of the subject of openness.

Corollary 5A.2 (invertibility of linear mappings). If a continuous linear mapping

A : X → Y is both surjective and injective, then its inverse is a continuous linear
mapping A−1 : Y → X with A−1  = reg A.
Proof. When A is both surjective and injective, then A−1 is single-valued every-
where and linear. Observe that the right side of (4) reduces to A−1 . The finiteness
of A−1  corresponds to A−1 being bounded, hence continuous.
It is worth noting also that if the range of A ∈ L (X ,Y ) is a closed subspace Y 
of Y , then Y  is a Banach space in its own right, and the facts we have recorded can
be applied to A as a surjective mapping from X to Y  .
Linear openness and the Aubin property. A result is available for set-valued
mappings F : X → → Y which has close parallels to the version of Theorem 5A.1
with (e) added, although it misses some aspects. This result corresponds in the case
of X = IRn and Y = IRn to Theorems 3E.6 and 3E.8, where an equivalence was estab-
lished between metric regularity, linear openness, and the inverse mapping having
the Aubin property. The statements and proofs of the cited theorems carry over in
the obvious manner to our present setting to yield a combined result, stated below
as Theorem 5A.3. Linear openness of F : X → Y at x̄ for ȳ, where ȳ ∈ F(x̄), means

there is a κ > 0 along

 with neighborhoods
 U of x̄ and V of ȳ such that
(5)
F(x + κ r intIB) ⊃ F(x) + r intIB ∩V for all x ∈ U, r > 0,

whereas the Aubin property of F −1 at ȳ for x̄, where x̄ ∈ F −1 (ȳ), means

there is a κ > 0 with neighborhoods U of x̄ and V of ȳ such that

(6)
e(F −1 (y) ∩U, F −1 (y )) ≤ κ y − y for all y, y ∈ V,

where the excess e(C, D) is defined in Section 3A. The infimum of all κ in (6) over
various choices of U and V is the modulus lip (F −1 ; ȳ| x̄).
5 Regularity in Infinite Dimensions 255

Theorem 5A.3 (metric regularity, linear openness and the inverse Aubin prop-
erty). For Banach spaces X and Y and a mapping F : X → → Y , the following proper-
ties with respect to a pair (x̄, ȳ) ∈ gph F are equivalent:
(a) F is linearly open at x̄ for ȳ with constant κ,
(b) F is metrically regular at x̄ for ȳ with constant κ,
(c) F −1 has the Aubin property at ȳ for x̄ with constant κ.
Moreover reg (F; x̄| ȳ) = lip (F −1 ; ȳ| x̄).
When F is taken to be a mapping A ∈ L (X ,Y ), how does the content of The-
orem 5A.3 compare with that of Theorem 5A.1? With linearity, the openness in
5A.1(b) comes out the same as the linear openness in 5A.3(a) and is easily seen to
reduce as well to the interiority condition in 5A.1(c). On the other hand, 5A.1(d)
has already been shown to be equivalent to the subsequently added property (e), to
which 5A.3(b) reduces when F = A. From 5A.3(c), though, we get yet another prop-
erty which could be added to the equivalences in Theorem 5A.1 for A ∈ L (X,Y ),
specifically that
(f) A−1 : Y →→ X has the Aubin property at every ȳ ∈ Y for every x̄ ∈ A−1 (ȳ),
where lip (A−1 ; ȳ| x̄) = reg A always. This goes farther than the observation in Corol-
lary 5A.2, which covered only single-valued A−1 . In general, of course, the Aubin
property in 5A.3(c) turns into local Lipschitz continuity when F −1 is single-valued.
An important feature of Theorem 5A.1, which is not represented at all in Theo-
rem 5A.3, is the assertion that surjectivity is sufficient, as well as necessary, for all
these properties to hold. An extension of that aspect to nonlinear F will be possible,
in a local sense, under the restriction that gph F is closed and convex. This will
emerge in the next section, in Theorem 5B.4.
Another result which we now wish to upgrade to infinite dimensions is the es-
timation for perturbed inversion which appeared in matrix form in Corollary 1E.7
with elaborations in 1E.8. It lies at the heart of the theory of implicit functions and
will eventually be generalized in more than one way. We provide it here with a direct
proof (compare with 1E.8(b)).

Lemma 5A.4 (estimation for perturbed inversion). Let A ∈ L (X,Y ) be invertible.

Then for any B ∈ L (X,Y ) with A−1 ·B < 1 one has

A−1 
(7) (A + B)−1 ≤ .
1 − A−1B

Proof. Let C = BA−1 ; then C < 1 and hence Cn  ≤ Cn → 0 as n → ∞. Also,
the elements
n
Sn = ∑ C i for n = 0, 1, . . .
i=0

form a Cauchy sequence in the Banach space L (X,Y ) which therefore converges
to some S ∈ L (X ,Y ). Observe that, for each n,

Sn (I − C) = I − Cn+1 = (I − C)Sn ,
256 5 Regularity in Infinite Dimensions

and hence, through passing to the limit, one has S = (I − C)−1 . On the other hand
n ∞
1
Sn  ≤ ∑ Ci  ≤ ∑ Ci = .
i=0 i=0 1 − C

Thus, we obtain
1
(I − C)−1  ≤ .
1 − C
All that remains is to bring in the identity (I − C)A = A − B and the inequality
C ≤ A−1 B, and to observe that the sign of B does not matter.
Note that, with the conventions ∞ · 0 = 0, 1/0 = ∞ and 1/∞ = 0, Lemma 5A.4
also covers the cases A−1  = ∞ and A−1 ·B = 1.

Exercise 5A.5. Derive Lemma 5A.4 from the contraction mapping principle 1A.2.
Guide. Setting a = A−1 , choose B ∈ L (X,Y ) with B < A−1 −1 and y ∈ Y
with y ≤ 1 − aB. Show that the mapping Φ : x → A−1 (y − Bx) satisfies the
conditions in 1A.2 with λ = aB and hence, there is a unique x ∈ aIB such that
x = A−1 (y − Bx), that is (A + B)x = y. Thus, A + B is invertible. Moreover x =
(A + B)−1(y) ≤ a for every y ∈ (1 − aB)IB, which implies that

A−1 
(A + B)−1z ≤ for every z ∈ IB.
1 − A−1B

This yields (7).

Exercise 5A.6. Let C ∈ L (X,Y ) satisfy C < 1. Prove that

C2
(I − C)−1 − I − C ≤ .
1 − C

Guide. Use the sequence of mappings Sn in the proof of 5A.4 and observe that

C2
Sn − I − C = C2 + C3 + · · · + Cn  ≤ .
1 − C

Positive homogeneity. A mapping H : X → → Y whose graph is a cone in X × Y is

called positively homogeneous. In infinite dimensions such mappings have proper-
ties similar to those developed in finite dimensions in Section 4A, but with some
complications. Outer and inner norms are defined for such mappings H as in Sec-
tion 4A, but it is necessary to take into account the possible variety of underlying
norms on X and Y in place of just the Euclidean norm earlier:
+ −
(8) H = sup sup y, H = sup inf y.
x≤1 y∈H(x) x≤1 y∈H(x)
5 Regularity in Infinite Dimensions 257

When dom H = X and H is single-valued, these two expressions agree. For H =

A ∈ L (X ,Y ), they reduce to A.
The inverse H −1 of a positively homogeneous mapping H is another positively
homogeneous mapping, and its outer and inner norms are therefore available also.
The elementary relationships in Propositions 4A.5 and 4A.6 have the following up-
date.

Proposition 5A.7 (outer and inner norms). The inner norm of a positively homo-
geneous mapping H : X →
→ Y satisfies
  
H = inf κ ∈ (0, ∞)  H(x) ∩ κ IB = 0/ for all x ∈ IB ,
−

so that, in particular,
−
(9) H < ∞ =⇒ dom H = X.

In parallel, the outer norm satisfies

   1
H = inf κ ∈ (0, ∞)  H(IB) ⊂ κ IB = sup
+
,
y=1 d(0, H −1 (y))

and we have
+
(10) H < ∞ =⇒ H(0) = {0},

with this implication becoming an equivalence when H has closed graph and
dim X < ∞.
The equivalence generally fails in (10) when dim X = ∞ because of the lack of
compactness then (with respect to the norm) of the ball IB in X.
An extension of Lemma 5A.4 to possibly set-valued mappings that are positively
homogeneous is now possible in terms of the outer norm. Recall that for a positively
homogeneous H : X → → Y and a linear B : X → Y we have (H + B)(x) = H(x) + Bx
for every x ∈ X .

Theorem 5A.8 (inversion estimate for the outer norm). Let H : X →

→ Y be positively
homogeneous with H −1 + < ∞. Then for any B ∈ L (X ,Y ) with the property that
H −1 +·B < 1, one has

H −1 +
(H + B)−1 ≤
+
(11) .
1 − H −1+ B

Proof. Having H −1 + = 0 is equivalent to having dom H = {0}; in this case,

since 0/ + y = 0/ for any y, we get (H + B)−1 + = 0 for any B ∈ L (X ,Y ) as
claimed. Suppose therefore instead that 0 < H −1 + < ∞. If the estimate (11) is
false, there is some B ∈ L (X,Y ) with B < [H −1 + ]−1 such that (H + B)−1 + >
([H −1 + ]−1 − B)−1. In particular B = 0 then, and by definition there must exist
258 5 Regularity in Infinite Dimensions

y ∈ IB and x ∈ (H + B)−1 (y) such that x > ([H −1 + ]−1 − B)−1, which is the
same as
1
> H −1  .
+
(12)
x−1 + B

But then y − Bx ∈ H(x) and

(13) y − Bx ≤ y + Bx ≤ 1 + Bx.

If y = Bx then 0 ∈ H(x), so (10) yields x = 0, a contradiction. Hence α := y −

Bx−1 > 0, and due to the positive homogeneity of H we have (α x, α (y − Bx)) ∈
gph H and α y − Bx = 1, which implies, by definition,

x
H −1  ≥
+
.
y − Bx

Combining this inequality with (12) and (13), we get

x x 1
H −1  ≥ > H −1  .
+ +
≥ =
y − Bx 1 + Bx x−1 + B

This is impossible, and the proof is at its end.

A corresponding extension of Lemma 5A.4 in terms of the inner norm will be
possible later, in Section 5C.
Normal cones and polarity. For a closed, convex cone K ⊂ X, the polar of K is
the subset K ∗ of the dual space X ∗ defined by
  
K ∗ = x∗ ∈ X ∗  x, x∗  ≤ 0 for all x ∈ K .

It is a closed convex cone in X ∗ from which K can be recovered as the polar (K ∗ )∗

of K ∗ in the sense that
  
K = x ∈ X  x, x∗  ≤ 0 for all x∗ ∈ K ∗ .

For any set C in a Banach space X and any point x ∈ C, the tangent cone TC (x) at
a point x ∈ C is defined as in 2A to consist of all limits v of sequences (1/τ k )(xk − x)
with xk → x in C and τ k  0. When C is convex, TC (x) has an equivalent description
as the closure of the convex cone consisting of all vectors λ (x − x) with x ∈ C and
λ > 0.
In infinite dimensions, the normal cone NC (x) to C at x can be introduced in
various ways that extend the general definition given for finite dimensions in 4H,
but we will only be concerned with the case of convex sets C. For that case, the
special definition in 2A suffices with only minor changes caused by the need to
work with the dual space X ∗ and the pairing x, x∗  between X and X ∗ . Namely,
NC (x) consists of all x∗ ∈ X ∗ such that
5 Regularity in Infinite Dimensions 259

x − x, x∗  ≤ 0 for all x ∈ C.

Equivalently, through the alternative description of TC (x) for convex C, the normal
cone NC (x) is the polar TC (x)∗ of the tangent cone TC (x). It follows that TC (x) is in
turn the polar cone NC (x)∗ .
As earlier, NC (x) is taken to be the empty set when x ∈/ C so as to get a set-valued
mapping NC defined for all x, but this normal cone mapping now goes from X to
X ∗ instead of from the underlying space into itself (except in the case of a Hilbert
space, where X ∗ can be identified with X as recalled above). A generalized equation
of the form
f (x) + NC (x)  0 for a function f : X → X ∗
is again a variational inequality. Such generalized equations are central, for in-
stance, to many applications involving differential or integral operators, especially
in a Hilbert space framework.

Exercise 5A.9 (normals to cones). Show that for a closed convex cone K ⊂ X and
its polar K ∗ ⊂ X ∗ , one has

x∗ ∈ NK (x) ⇐⇒ x ∈ K, x∗ ∈ K, x, x∗  = 0.

Exercise 5A.10 (linear variational inequalities on cones). Let H(x) = Ax + NK (x)

for A ∈ L (X , X ∗ ) and a closed, convex cone K ⊂ X . Show that H is positively
homogeneous with closed graph, but this graph is not convex unless K is a subspace
of X .

5B. Mappings with Closed Convex Graphs

For any mapping F : X → → Y with convex graph, the sets dom F and rge F, as the
projections of gph F on the Banach spaces X and Y , are convex sets as well. When
gph F is closed, these sets can fail to be closed (a famous example being the case
where F is a “closed” linear mapping from X to Y with domain dense in X). How-
ever, if either of them has nonempty interior, or even nonempty “core” (in the sense
about to be explained), there are highly significant consequences for the behavior
of F. This section is dedicated to developing such consequences for properties like
openness and metric regularity, but we begin with some facts that are more basic.
The core of a set C ⊂ X is defined by
  
core C = x  ∀ w ∈ X ∃ ε > 0 such that x + tw ∈ C when 0 ≤ t ≤ ε .

A set C is called absorbing if 0 ∈ core C. Obviously core C ⊃ int C always, but there
are circumstances where necessarily core C = int C. It is elementary that this holds
260 5 Regularity in Infinite Dimensions

when C is convex with int C = 0, / but more attractive is the potential of using the
purely algebraic test of whether a point x belongs to core C to confirm that x ∈ int C
without first having to establish that int C = 0.
/ Most importantly for our purposes
here,

(1) for a closed convex subset C of a Banach space, core C = int C.

An equivalent statement, corresponding to how this fact is often recorded in func-

tional analysis, is that if C is a closed convex set which is absorbing, then C must
be a neighborhood of 0. Through the observation already made about convexity, the
confirmation of this comes down to establishing that C has nonempty interior. That
can be deduced from the Baire category theorem, according to which the union
of a sequence of nowhere dense subsets of a complete metric space cannot cover
the whole space. If int C were empty, the closed sets nC for n = 1, 2, . . . would be
nowhere dense with the entire Banach space as their union, but this is impossible.
We demonstrate now that, for some of the convex sets central to the study of
closed convex graphs on which we are embarking, the core and interior coincide
even without closedness.

Theorem 5B.1 (interiority criteria for domains and ranges). For any mapping F :
X→
→ Y with closed convex graph, one has

(2) core rge F = int rge F, core dom F = int dom F.

In addition, core cl rge F = int cl rge F and core cl dom F = int cl dom F , where
moreover
int cl rge F = int rge F when dom F is bounded,
(3)
int cl dom F = int dom F when rge F is bounded.

In particular, if dom F is bounded and rge F is dense in Y , and then rge F = Y .

Likewise, if rge F is bounded and dom F is dense in X , and then dom F = X .
Proof. The equations in the line following (2) merely apply (1) to the closed convex
sets cl rge F and cl dom F. Through symmetry between F and F −1 , the two claims
in (2) are equivalent to each other, as are the two claims in (3). Also, the claims after
(3) are immediate from (3). Therefore, we only have to prove one of the claims in
(2) and one of the claims in (3).
We start with the first claim in (3). Assuming that dom F is bounded, we work
toward verifying that int rge F ⊃ int cl rge F; this gives equality, inasmuch as the
opposite inclusion is obvious. In fact, for this we only need to show that rge F ⊃
int cl rge F.
We choose ỹ ∈ int cl rge F; then there exists δ > 0 such that intIB2δ (ỹ) ⊂
int cl rge F. We will find a point x̃ such that (x̃, ỹ) ∈ gph F, so that ỹ ∈ rge F. The
point x̃ will be obtained by means of a sequence {(xk , yk )} which we now construct
by induction.
5 Regularity in Infinite Dimensions 261

Pick any (x0 , y0 ) ∈ gph F. Suppose we have already determined {(x j , y j )} ∈

gph F for j = 0, 1, . . . , k. If yk = ỹ, then take x̃ = xk and (xn , yn ) = (x̃, ỹ) for all
n = k, k + 1, . . .; that is, after the index k the sequence is constant. Otherwise,
with α k = δ /yk − ỹ, we let wk := ỹ + α k (ỹ − yk ). Then wk ∈ IBδ (ỹ) ⊂ cl rge F.
Hence there exists vk ∈ rge F such that vk − wk  ≤ yk − ỹ/2 and also uk with
(uk , vk ) ∈ gph F. Having gotten this far, we pick

αk 1
(xk+1 , yk+1 ) = (xk , yk ) + (uk , vk ).
1+α k 1 + αk

Clearly, (xk+1 , yk+1 ) ∈ gph F by its convexity. Also, the sequence {yk } satisfies

vk − wk  1 k
yk+1 − ỹ = ≤ y − ỹ.
1 + αk 2

If yk+1 = ỹ, we take x̃ = xk+1 and (xn , yn ) = (x̃, ỹ) for all n = k + 1, k + 2, . . .. If not,
we perform the induction step again. As a result, we generate an infinite sequence
{(xk , yk )}, each element of which is equal to (x̃, ỹ) after some k or has yk = ỹ for all
k and also
1 0
(4) yk − ỹ ≤ y − ỹ for all k = 1, 2, . . . .
2k
In the latter case, we have yk → ỹ. Further, for the associated sequence {xk } we
obtain
xk − uk  xk  + uk k
xk+1 − xk  = ≤ y − ỹ.
1 + αk yk − ỹ + δ
Both xk and uk are from dom F and thus are bounded. Therefore, from (4), {xk } is a
Cauchy sequence, hence (because X is a complete metric space) convergent to some
x̃. Because gph F is closed, we end up with (x̃, ỹ) ∈ gph F, as required.
Next we address the second claim in (2), where the inclusion core dom F ⊂
int dom F suffices for establishing equality. We must show that an arbitrarily cho-
sen point of core dom F belongs to int dom F, but through a translation of gph F
we can focus without loss of generality on that point in core dom F being 0,
with F(0)  0. Let F0 : X → → Y be defined by F0 (x) = F(x) ∩ IB. The graph of
F0 , being [X × IB] ∩ gph F, is closed and convex, and we have dom F0 ⊂ dom F
and rge F0 ⊂ IB (bounded). The relations already established in (3) tell us that
int cl dom F0 = int dom F0 , where cl dom F0 is a closed convex set. By demon-
strating that cl dom F0 is absorbing, we will be able to conclude from (1) that
0 ∈ int dom F0 , hence 0 ∈ int dom F. It is enough actually to show that dom F0
itself is absorbing.
Consider any x ∈ X . We have to show the existence of ε > 0 such that tx ∈ dom F0
for t ∈ [0, ε ]. We do know, because dom F is absorbing, that tx ∈ dom F for all t > 0
sufficiently small. Fix t0 as such a t, and letting y0 ∈ F(t0 x); let y = y0 /t0 , so that
t0 (x, y) ∈ gph F. The pair t(x, y) = (tx,ty) belongs then to gph F for all t ∈ [0,t0 ]
through the convexity of gph F and our arrangement that (0, 0) ∈ gph F. Take ε > 0
262 5 Regularity in Infinite Dimensions

small enough that ε y ≤ 1. Then for t ∈ [0, ε ] we have ty = ty ≤ 1, giving us
ty ∈ F(tx) ∩ IB and therefore tx ∈ dom F0 , as required.
Regularity properties will now be explored. The property of a mapping F : X → →Y
being open at x̄ for ȳ, as extended to Banach spaces X and Y in 5A(1), can be restated
equivalently in a manner that more closely resembles the linear openness property
defined in 5A(5):

(5) for any a > 0 there exists b > 0 such that F(x̄ + a intIB) ⊃ ȳ + b intIB.

Linear openness requires a linear scaling relationship between a and b. Under posi-
tive homogeneity, such scaling is automatic. On the other hand, an intermediate type
of property holds automatically without positive homogeneity when the graph of F
is convex, and it will be a stepping stone toward other, stronger, consequences of
convexity.

Proposition 5B.2 (openness of mappings with convex graph). Consider a mapping

F:X → → Y with convex graph, and let ȳ ∈ F(x̄). Then openness of F at x̄ for ȳ is
equivalent to the simpler condition that

(6) there exists c > 0 with F(x̄ + intIB) ⊃ ȳ + c intIB.

Proof. Clearly, (5) implies (6). For the converse, assume (6) and consider any a > 0.
Take b = min{1, a}c. If a ≥ 1, the left side of (6) is contained in the left side of (5),
and hence (5) holds. Suppose therefore that a < 1. Let w ∈ ȳ + b intIB. The point v =
(w/a) − (1 − a)(ȳ/a) satisfies v − ȳ = w − ȳ/a < b/a = c, hence v ∈ ȳ + c intIB.
Then from (6) there exists u ∈ x̄ + intIB with (u, v) ∈ gph F. The convexity of gph F
implies a(u, v) + (1 − a)(x̄, ȳ) ∈ gph F and yields av + (1 − a)ȳ ∈ F(au + (1 − a)x̄) ⊂
F(x̄ + a intIB). Substituting v = (w/a) − (1 − a)(ȳ/a) in this inclusion, we see that
w ∈ F(x̄ + a intIB), and since w was an arbitrary point in ȳ + b intIB, we get (5).
The following fact bridges, for set-valued mappings with convex graphs, between
condition (6) and metric regularity.

Lemma 5B.3 (metric regularity estimate). Let F : X → → Y have convex graph con-
taining (x̄, ȳ), and suppose (6) is fulfilled. Then

1 + x − x̄
(7) d(x, F −1 (y)) ≤ d(y, F(x)) for all x ∈ X, y ∈ ȳ + c intIB.
c − y − ȳ

Proof. We may assume that (x̄, ȳ) = (0, 0), since this can be arranged by translating
gph F to gph F − (x̄, ȳ). Then condition (6) has the simpler form

(8) there exists c > 0 with F(intIB) ⊃ c intIB.

Let x ∈ X and y ∈ c intIB. Observe that (7) is automatically true when x ∈

/ dom F or
y ∈ F(x), so assume that x ∈ dom F but y ∈ / F(x). Let α := c − y. Then α > 0.
Choose ε ∈ (0, α ) and find y ∈ F(x) such that y − y ≤ d(y, F(x)) + ε . The point
5 Regularity in Infinite Dimensions 263

ỹ := y + (α − ε )y − y−1 (y − y ) satisfies ỹ ≤ y + α − ε = c − ε < c, hence

ỹ ∈ c intIB. By (8) there exists x̃ ∈ intIB with ỹ ∈ F(x̃). Let β := y − y (α − ε +
y − y)−1 ; then β ∈ (0, 1). From the convexity of gph F we have

y = (1 − β )y + β ỹ ∈ (1 − β )F(x) + β F(x̃) ⊂ F((1 − β )x + β x̃).

Thus x + β (x̃ − x) ∈ F −1 (y), so d(x, F −1 (y)) ≤ β x − x̃. Noting that x − x̃ ≤

x + x̃ < x + 1 and β ≤ (α − ε )−1 y − y, we obtain

1 + x
d(x, F −1 (y)) < [d(y, F(x)) + ε ].
α −ε
Letting ε → 0, we finish the proof.
Condition (6) entails in particular having ȳ ∈ int rge F. It turns out that when
the graph of F is not only convex but also closed, the converse implication holds as
well, that is, ȳ ∈ int rge F is equivalent to (6). This is a consequence of the following
theorem, which furnishes a far-reaching generalization of the Banach open mapping
theorem.

Theorem 5B.4 (Robinson–Ursescu). Let F : X → → Y have closed convex graph and

let ȳ ∈ F(x̄). Then the following are equivalent:
(a) ȳ ∈ int rge F ,
(b) F is open at x̄ for ȳ,
(c) F is metrically regular at x̄ for ȳ.
Proof. We first demonstrate that

(9) ȳ ∈ int F(x̄ + intIB) when x̄ ∈ F −1 (ȳ) and ȳ ∈ int rge F.

By a translation, we can reduce to the case of (x̄, ȳ) = (0, 0). To conclude (9)
in this setting, where F(0)  0 and 0 ∈ int rge F, it will be enough to show, for an
arbitrary δ ∈ (0, 1), that 0 ∈ int F(δ IB). Define the mapping Fδ : X → → Y by Fδ (x) =
F(x) when x ∈ δ IB but Fδ (x) = 0/ otherwise. Then Fδ has closed convex graph given
by [δ IB×Y ]∩gph F. Also F(δ IB) = rge Fδ and dom Fδ ⊂ δ IB. We want to show that
0 ∈ int rge Fδ , but have Theorem 5B.1 at our disposal, according to which we only
need to show that rge Fδ is absorbing. For that purpose we use an argument which
closely parallels one already presented in the proof of Theorem 5B.1. Consider any
y ∈ Y . Because 0 ∈ int rge F, there exists t0 such that ty ∈ rge F when t ∈ [0,t0 ]. Then
there exists x0 such that t0 y ∈ F(x0 ). Let x = x0 /t0 , so that (t0 x,t0 y) ∈ gph F. Since
gph F is convex and contains (0, 0), it also then contains (tx,ty) for all t ∈ [0,t0 ].
Taking ε > 0 for which ε x ≤ δ , we get for all t ∈ [0, ε ] that (tx,ty) ∈ gph Fδ ,
hence ty ∈ rge Fδ , as desired.
Utilizing (9), we can put the argument for the equivalences in Theorem 5B.4 to-
gether. That (b) implies (a) is obvious. We work next on getting from (a) to (c). When
(a) holds, we have from (9) that (6) holds for some c, in which case Lemma 5B.3
provides (7). By restricting x and y to small neighborhoods of x̄ and ȳ in (7), we
deduce the metric regularity of F at x̄ for ȳ with any constant κ > 1/c. Thus, (c)
264 5 Regularity in Infinite Dimensions

holds. Finally, out of (c) and the equivalences in Theorem 5A.3 we may conclude
that F is linearly open at x̄ for ȳ, and this gets us back to (b).
The preceding argument passed through linear openness as a fourth property
which could be added to the equivalences in Theorem 5B.4, but which was left out
of the theorem’s statement for historical reasons. We now record this fact separately.

Theorem 5B.5 (linear openness from openness and convexity). For a mapping F :
X→ → Y with closed convex graph, openness at x̄ for ȳ always entails linear openness
at x̄ for ȳ.
Another fact, going beyond the original versions of Theorem 5B.4, has come up
as well.

Theorem 5B.6 (core criterion for regularity). Condition (a) of Theorem 5B.4 can
be replaced by the criterion that ȳ ∈ core rge F .
Proof. This calls up the core property in Theorem 5B.1.
We can finish tying up loose ends now by returning to the Banach open mapping
theorem at the beginning of this chapter and tracing how it fits with the Robinson–
Ursescu theorem.
Derivation of Theorem 5A.1 from Theorem 5B.4. It was already noted in the
sequel to 5A.1 that condition (d) in that result was equivalent to the metric regularity
of the linear mapping A, stated as condition (e). It remains only to observe that when
Theorem 5B.4 is applied to F = A ∈ L (X,Y ) with x̄ = 0 and ȳ = 0, the graph of
A being a closed subspace of X × Y (in particular a convex set), and the positive
homogeneity of A is brought in, we not only get (b) and (c) of Theorem 5A.1, but
also (a).
The argument for Theorem 5B.4, in obtaining metric regularity, also revealed a
relationship between that property and the openness condition in 5B.2 which can be
stated in the form
  
(10) sup c ∈ (0, ∞)  (6) holds ≤ [reg (F; x̄| ȳ)]−1 .

Exercise 5B.7 (a counterexample). Show that for the mapping F : IR →

→ IR
%
[−x, 0.5] if x > 0.25,
F(x) = [−x, 2x] if x ∈ [0, 0.25],
0/ if x < 0,
which is not positively homogeneous, and for x̄ = ȳ = 0 the inequality (10) is strict.

Exercise 5B.8 (effective domains of convex functions).

  Let g : X → (−∞, ∞] be
convex and lower semicontinuous, and let D = x  g(x) < ∞ . Show that D is a
5 Regularity in Infinite Dimensions 265

convex set which, although not necessarily closed in X , is sure to have core D =
int D. Moreover, on that interior g is locally Lipschitz continuous.
  
Guide. Look at the mapping F : X → → IR defined by F(x) = y ∈ IR  y ≥ g(x) .
Apply results in this section and also 5A.3.

5C. Sublinear Mappings

An especially interesting class of positively homogeneous mappings H : X → →Y
acting between Banach spaces X and Y consists of the ones for which gph H is not
just a cone, but a convex cone. Such mappings are called sublinear, because these
geometric properties of gph H are equivalent to the rules that

0 ∈ H(0), H(λ x) = λ H(x) for λ > 0,

(1)
H(x + x ) ⊃ H(x) + H(x ) for all x, x ,

which resemble linearity. Since the projection of a convex cone in X ×Y into X or Y

is another convex cone, it is clear for a sublinear mapping H that dom H is a convex
cone in X and rge H is a convex cone in Y . The inverse H −1 of a sublinear mapping
H is another sublinear mapping.
Although sublinearity has not been mentioned as a specific property before now,
sublinear mappings have already appeared many times. Obviously, any linear map-
ping A : X → Y is sublinear (its graph being not just a convex cone but in fact
a subspace of X × Y ). Sublinear also, though, is any mapping H : X → → Y with
H(x) = Ax−K for a convex cone K in Y . Such mappings enter the study of constraint
systems, with linear equations corresponding to K = {0}. When A is continuous and
K is closed, their graphs are closed.
Sublinear mappings with closed graph enjoy the properties laid out in 5B along
with those concerning outer and inner norms at the end of 5A. But their properties
go a lot further, as in the result stated now about metric regularity.

Theorem 5C.1 (metric regularity of sublinear mappings). For a sublinear mapping

H:X →→ Y with closed graph, and any (x, y) ∈ gph H we have
  
(2) reg (H; x|y) ≤ reg (H; 0|0) = inf κ > 0  H(κ intIB) ⊃ intIB = H −1  .
−

Moreover, reg (H; 0|0) < ∞ if and only if H is surjective, in which case H −1 is
Lipschitz continuous on Y (in the sense of Pompeiu-Hausdorff distance as defined
in 3A) and the infimum of the Lipschitz constant κ for this equals H −1 − .
Proof. Let κ > reg (H; 0|0). Then, from 5A.3, H is linearly open at 0 for 0 with
constant κ , which reduces to H(κ intIB) ⊃ intIB. On the other hand, just from know-
ing that H(κ intIB) ⊃ intIB, we obtain for arbitrary (x, y) ∈ gph H and r > 0 through
266 5 Regularity in Infinite Dimensions

the sublinearity of H that

H(x + κ r intIB) ⊃ H(x) + rH(κ intIB) ⊃ y + r intIB.

This establishes that reg (H; x|y) ≤ reg (H; 0|0) for all (x, y) ∈ gph H. Appealing
again to positive homogeneity, we get
  
(3) reg (H; 0|0) = inf κ > 0  H(κ intIB) ⊃ intIB .

The right side of (3) does not change if we replace the open balls with their closures,
hence, by 5A.7 or just by the definition of the inner norm, it equals H −1 − . This
confirms (2).
The finiteness of the right side of (3) corresponds to H being surjective, by virtue
of positive homogeneity. We are left now with showing that H −1 is Lipschitz con-
tinuous on Y with H −1− as the infimum of the available constants κ .
If H −1 is Lipschitz continuous on Y with constant κ , it must in particular have
the Aubin property at 0 for 0 with this constant, and then κ ≥ reg (H; 0|0) by 5A.3.
We already know that this regularity modulus equals H −1 − , so we are left with
proving that, for every κ > reg (H; 0|0), H −1 is Lipschitz continuous on Y with
constant κ .
Let c < [H −1− ]−1 and κ > 1/c. Taking (2) into account, we apply the inequal-
ity 5B(7) derived in 5B.3 with x = x̄ = 0 and ȳ = 0, obtaining the existence of a > 0
such that
d(0, H −1 (y)) ≤ κ d(y, H(0)) ≤ κ y for all y ∈ aIB.
(Here, without loss of generality, we replace the open ball for y by its closure.) For
any y ∈ Y , we have ay/y ∈ aIB, and from the positive homogeneity of H we get

(4) d(0, H −1 (y)) ≤ κ y for all y ∈ Y.

If H −1 − = 0 then 0 ∈ H −1 (y) for all y ∈ Y (see 4A.9), hence (4) follows automat-
ically.
Let y, y ∈ Y and x ∈ H −1 (y ). Through the surjectivity of H again, we can find
for any δ > 0 an xδ ∈ H −1 (y − y ) such that xδ  ≤ d(0, H −1 (y − y )) + δ , and then
from (4) we get

(5) xδ  ≤ κ y − y + δ .

Invoking the sublinearity of H yet once more, we obtain

x := x + xδ ∈ H −1 (y ) + H −1(y − y) ⊂ H −1 (y + y − y) = H −1 (y).

Hence x = x−xδ ∈ H −1 (y)+xδ IB. Recalling (5), we arrive finally at the existence
of x ∈ H −1 (y) such that x − x  ≤ κ y − y + δ . Since δ can be arbitrarily small,
this yields Lipschitz continuity of H −1 , and we are done.
5 Regularity in Infinite Dimensions 267

Corollary 5C.2 (finiteness of the inner norm). Let H : X →

→ Y be a sublinear map-
ping with closed graph. Then

dom H = X ⇐⇒ H− < ∞,

rge H = Y ⇐⇒ H −1− < ∞.

Proof. This comes from applying Theorem 5C.1 to both H and H −1 .

Exercise 5C.3 (regularity modulus at zero). For a sublinear mapping H : X → →Y

with closed graph, prove that
  
reg (H; 0|0) = inf κ  H(x + κ rIB) ⊃ H(x) + rIB for all x ∈ X , r > 0 .

Guide. Utilize the connections with openness properties.

Example 5C.4 (application to linear constraints). For A ∈ L (X,Y ) and a closed,

convex cone K ⊂ Y , define the solution mapping S : Y →
→ X by
  
S(y) = x ∈ X  Ax − y ∈ K .

Then S is a sublinear mapping with closed graph, and the following properties are
equivalent:
(a) S(y) = 0/ for all y ∈ Y ;
(b) there exists κ such that d(x, S(y)) ≤ κ d(Ax − y, K) for all x ∈ X, y ∈ Y ;
(c) there exists κ such that h(S(y), S(y )) ≤ κ y − y for all y, y ∈ Y ,
in which case the infimum of the constants κ that work in (b) coincides with the
infimum of the constants κ that work in (c) and equals S−.
Detail. Here S = H −1 for H(x) = Ax − K, and the assertions of Theorem 5C.1 then
translate into this form.
Additional insights into the structure of sublinear mappings will emerge from
applying a notion which comes out of the following fact.

Exercise 5C.5 (directions of unboundedness in convexity). Let C be a closed, con-

vex subset of X and let x1 and x2 belong to C. If w = 0 in X has the property that
x1 + tw ∈ C for all t ≥ 0, then it also has the property that x2 + tw ∈ C for all t ≥ 0.
Guide. Fixing any t2 > 0, show that x2 + t2 w can be approached arbitrarily closely
by points on the line segment between x2 and x1 + t1 w by taking t1 larger.
On the basis of the property in 5C.5, the recession cone rc C of a closed, convex
set C ⊂ X , defined by
  
(6) rc C = w ∈ X  ∀ x ∈ C, ∀t ≥ 0 : x + tw ∈ C ,

can equally well be described by

268 5 Regularity in Infinite Dimensions
  
(7) rc C = w ∈ X  ∃ x ∈ C, ∀t ≥ 0 : x + tw ∈ C .

It is easily seen that rc C is a closed, convex cone. In finite dimensions, C is bounded

if and only if rc C is just {0}, but in infinite dimensions there are unbounded sets
for which that holds. For a closed, convex cone K, one just has rc K = K, as seen
from the equivalence between (6) and (7) by taking x = 0 in (7).
We will apply this now to the graph of a sublinear mapping H. It should be
recalled that dom H is a convex cone, and for any convex cone K the set K ∩ [−K]
is a subspace, in fact the largest subspace within K. On the other hand, K − K is the
smallest subspace that includes K.

Proposition 5C.6 (recession cones in sublinearity). A sublinear mapping H : X →

→Y
with closed graph has

(8) rc H(x) = H(0) for all x ∈ dom H,

and on the other hand,


H(x) + H(−x) ⊂ H(0),
(9) x ∈ dom H ∩ [− dom H] =⇒
H(x) − H(x) ⊂ H(0) − H(0).

Proof. Let G = gph H, this being a closed, convex cone in X × Y , therefore having
rc G = G. For any (x, y) ∈ G, the recession cone rc H(x) consists of the vectors w
such that y + tw ∈ H(x) for all t ≥ 0, which are the same as the vectors w such that
(x, y) + t(0, w) ∈ G for all t ≥ 0, i.e., the vectors w such that (0, w) ∈ rc G = G. But
these are the vectors w ∈ H(0). That proves (8).
The first inclusion in (9) just reflects the rule that H(x + [−x]) ⊃ H(x) + H(−x)
by sublinearity. To obtain the second inclusion, let y1 and y2 belong to H(x), which
is the same as having y1 − y2 ∈ H(x) − H(x), and let y ∈ H(−x). Then by the first
inclusion we have both y1 + y and y2 + y in H(0), hence their difference lies in
H(0) − H(0).

Theorem 5C.7 (single-valuedness of sublinear mappings). For a sublinear map-

ping H : X →→ Y with closed graph and dom H = X , the following conditions are
equivalent:
(a) H is a linear mapping from L (X,Y );
(b) H is single-valued at some point {x};
(c) H+ < ∞.
Proof. Certainly (a) leads to (b). On the other hand, (c) necessitates H(0) = {0} and
hence (b) for x = 0. By 5C.6, specifically the second part of (9), if (b) holds for any
x it must hold for all x. The first part of (9) reveals then that if H(x) consists just of y,
then H(−x) consists just of −y. This property, along with the rules of sublinearity,
implies linearity. The closedness of the graph of H implies that the linear mapping
so obtained is continuous, so we have come back to (a).
5 Regularity in Infinite Dimensions 269

Note that, without the closedness of the graph of H in Theorem 5C.7, there would
be no assurance that (b) implies (a). We would still have a linear mapping, but it
might not be continuous.

Corollary 5C.8 (single-valuedness of solution mappings). In the context of Theo-

rem 5C.1, it is impossible for H −1 to be single-valued at any point without actually
turning out to be a continuous linear mapping from Y to X . The same holds for the
solution mapping S for the linear constraint system in 5C.4 when a solution exists
for every y ∈ Y .
We next state the counterpart to Lemma 5A.4 which works for the inner norm
of a positively homogeneous mapping. In contrast to the result presented in 5A.8
for the outer norm, convexity is now essential: we must limit ourselves to sublinear
mappings.

Theorem 5C.9 (inversion estimate for the inner norm). Let H : X →

→ Y be sublinear
with closed graph and have H −1 − < ∞. Then for any B ∈ L (X ,Y ) such that
H −1 −·B < 1, one has

H −1 −
(H + B)−1  ≤
−
.
1 − H −1− B

The proof of this is postponed until 5E, where it will be deduced from the connec-
tion between these properties and metric regularity in 5C.1. Perturbations of metric
regularity will be a major theme, starting in Section 5D.
Duality. A special feature of sublinear mappings, with parallels linear mappings,
is the availability of “adjoints” in the framework of the duals X ∗ and Y ∗ of the
Banach spaces X and Y . For a sublinear mapping H : X → → Y , the upper adjoint
H ∗+ : Y ∗ →
→ X ∗ is defined by

(10) (y∗ , x∗ ) ∈ gph H ∗+ ⇐⇒ x∗ , x ≤ y∗ , y for all (x, y) ∈ gph H,

whereas the lower adjoint H ∗− : Y ∗ →

→ X ∗ is defined by

(11) (y∗ , x∗ ) ∈ gph H ∗− ⇐⇒ x∗ , x ≥ y∗ , y for all (x, y) ∈ gph H.

These formulas correspond to modified polarity operations on the convex cone

gph H ⊂ X × Y with polar [gph H]∗ ⊂ X ∗ × Y ∗ . They say that gph H ∗+ consists
of the pairs (y∗ , x∗ ) such that (x∗ , −y∗ ) ∈ [gph H]∗ , while gph H ∗− consists of the
pairs (y∗ , x∗ ) such that (−x∗ , y∗ ) ∈ [gph H]∗ , and thus imply in particular that the
graphs of these adjoints are closed, convex cones — so that both of these mappings
from Y ∗ to X ∗ are sublinear with closed graph.
The switches of sign in (10) and (11) may seem a pointless distinction to make,
but they are essential in capturing rules for recovering H from its adjoints through
the fact that, when the spaces X and Y are reflexive, [gph H]∗∗ = gph H when the
convex cone gph H is closed, and then we get
270 5 Regularity in Infinite Dimensions

(12) [H ∗+ ]∗− = H and [H ∗− ]∗+ = H for sublinear H with closed graph.

When H reduces to a linear mapping A ∈ L (X,Y ), both adjoints come out as the
usual adjoint A∗ ∈ L (Y ∗ , X ∗ ). In that setting the graphs are subspaces instead of
just cones, and the difference between (10) and (11) has no effect. The fact that
A∗  = A in this case has the following generalization.

Theorem 5C.10 (duality of inner and outer norms). For any sublinear mapping
H:X →→ Y with closed graph, one has

H+ = H ∗− − = H ∗+ − ,
(13)
H− = H ∗− + = H ∗+ + .

The proof requires some additional background. First, we need to update to Ba-
nach spaces the semicontinuity properties introduced in a finite-dimensional frame-
work in Section 3B, but this only involves an extension of notation. A mapping
F :X →→ Y is inner semicontinuous at x̄ ∈ dom F if for every y ∈ F(x̄) and every
neighborhood V of y one can find a neighborhood U of x̄ with U ⊂ F −1 (V ) or,
equivalently, F(x) ∩V = 0/ for all x ∈ U (this corresponds to 3B.2). Outer semiconti-
nuity has a parallel extension. Next, we record a standard fact in functional analysis
which will be called upon.

Theorem 5C.11 (Hahn–Banach). Let M be a linear subspace of a Banach space X ,

and let p : X → IR satisfy

(14) p(x + y) ≤ p(x) + p(y) and p(tx) = t p(x) for all x, y ∈ X, t ≥ 0.

Let f : M → IR be a linear functional such that f (x) ≤ p(x) for all x ∈ M. Then
there exists a linear functional l : X → IR such that l(x) = f (x) for all x ∈ M, and
l(x) ≤ p(x) for all x ∈ X .

In this formulation of the Hahn–Banach theorem, nothing is said about continu-

ity, so X could really be any linear space — no topology is involved. But the main
applications are ones in which p is continuous and it follows that l is continuous.
Another standard fact in functional analysis, which can be derived from the Hahn–
Banach theorem in that manner, is the following separation theorem.

Theorem 5C.12 (separation theorem). Let C be a nonempty, closed, convex subset

of a Banach space X , and let x0 ∈ X . Then x0 ∈ C if and only if there exists x∗ ∈ X ∗
such that
x∗ , x0  > supx∗ , x.
x∈C

Essentially, this says geometrically that a closed convex set is the intersection of
all the “closed half-spaces” that include it.
5 Regularity in Infinite Dimensions 271

Proof of Theorem 5C.10. First, observe from (10) and (11) that

H ∗+ (y∗ ) = −H ∗−(−y∗ ) for any y∗ ∈ Y ∗ ,

so that
H ∗−  = H ∗+  H ∗−  = H ∗+  .
− − + +
and
To prove that H+ = H ∗− − we fix any y∗ ∈ Y ∗ and show that

(15) sup sup y∗ , y = inf x∗  for all y∗ ∈ IB.

x∈IB y∈H(x) x∗ ∈H ∗− (y∗ )

If infx∗ ∈H ∗− (y∗ ) x∗  < r for some r > 0, then there exist x∗ ∈ H ∗− (y∗ ) such that
x∗  < r. For any x̃ ∈ IB and ỹ ∈ H(x̃) we have

y∗ , ỹ ≤ x∗ , x̃ ≤ supx∗ , x = x∗  < r,

x∈IB

and then of course supx∈IB supy∈H(x) y∗ , y ≤ r. Hence

(16) sup sup y∗ , y ≤ inf x∗ .

x∈IB y∈H(x) x∗ ∈H ∗− (y∗ )

To prove the inequality opposite to (16) and hence the equality (15), assume that
supx∈IB supy∈H(x) y∗ , y < r for some r > 0 and pick 0 < d < r such that

(17) sup sup y∗ , y ≤ d.

x∈IB y∈H(x)

Define the mapping G : X →

→ IR by
  
G : x → z  z = y∗ , y, y ∈ H(x + IB) .

First, observe that gph G is convex. Indeed, if (x1 , z1 ), (x2 , z2 ) ∈ gph G and 0 < λ
< 1, then there exist yi ∈ Y and wi ∈ IB with zi = y∗ , yi  and yi ∈ H(xi + wi ), for
i = 1, 2. Since H is sublinear, we get λ y1 + (1 − λ )y2 ∈ H(λ (x1 + w1 )+ (1 − λ )(x2 +
w2 )). Hence, λ y1 + (1 − λ )y2 ∈ H(λ x1 + (1 − λ )x2 + IB), and thus,

λ (x1 , z1 ) + (1 − λ )(x2, z2 ) = (λ x1 + (1 − λ )x2, y∗ , λ y1 + (1 − λ )y2) ∈ gph G.

We will show next that G is inner semicontinuous at 0. Take z̃ ∈ G(0) and ε > 0. Let
z̃ = y∗ , ỹ for ỹ ∈ H(w̃) and w̃ ∈ IB. Since y∗ , · is continuous, there is some γ > 0
such that |y∗ , y − z̃| ≤ ε when y − ỹ ≤ γ . Choose δ ∈ (0, 1) such that δ ỹ ≤ γ .
If x ≤ δ , we have

(1 − δ )w̃ − x ≤ (1 − δ )w̃ + x ≤ 1,

and hence (1 − δ )w̃ − x ∈ IB. Because H is sublinear,

(1 − δ )ỹ ∈ H((1 − δ )w̃) = H(x + ((1 − δ )w̃ − x)) ⊂ H(x + IB) whenever x ≤ δ .
272 5 Regularity in Infinite Dimensions

Moreover, (1 − δ )ỹ − ỹ = δ ỹ ≤ γ , and then |y∗ , (1 − δ )ỹ − z̃| ≤ ε . Therefore,
for all x ∈ δ IB, we have y∗ , (1 − δ )ỹ ∈ G(x) ∩ IBε (z̃), and hence G is inner semi-
continuous at 0 as desired.
Let us now define a mapping K : X → → IR whose graph is the conical hull of
gph (d − G) where d is as in (17); that is, its graph is the set of points λ h for h ∈
gph (d − G) and λ ≥ 0. The conical hull of a convex set is again convex, so K is
another sublinear mapping. Since G is inner semicontinuous at 0, there is some
neighborhood U of 0 with U ⊂ dom G, and therefore dom K = X. Consider the
functional   
k : x → inf z  z ∈ K(x) for x ∈ X.
Because K is sublinear and d − H(0) ⊂ IR+ , we have

(18) K(x) + K(−x) ⊂ K(0) ⊂ IR+ .

This inclusion implies in particular that any point in −K(−x) furnishes a lower
bound in IR for the set of values K(x), for any x ∈ X. Indeed, let x ∈ X and y ∈
−K(−x). Then (18) yields K(x) − y ⊂ IR+ , and consequently y ≤ z for all z ∈ K(x).
Therefore k(x) is finite for all x ∈ X; we have dom k = X . Also, from the sublinearity
of K and the properties of the infimum, we have

k(x + y) ≤ k(x) + k(y) and k(α x) = α k(x) for all x, y ∈ X and α ≥ 0.

Consider the subspace M = {0} ⊂ X and define f : M → IR simply by f (0) :=

k(0) = 0. Applying Hahn–Banach theorem 5C.11 to f , we get a linear functional
l : X → IR such that l(0) = 0 and l(x) ≤ k(x) for all x ∈ X. We will show now that l
is continuous at 0 and hence continuous on the whole X.
Continuity at 0 means that for any ε > 0 there is δ > 0 such that (l(x) + IR+ ) ∩
ε IB = 0/ whenever x ∈ δ IB. Let z ∈ d − G(0) and take 0 < λ < 1 along with a neigh-
borhood V of z such that λ V ⊂ ε IB. Since G is inner semicontinuous at 0, there is
some δ > 0 such that

/ for all x ∈ (δ /λ )IB.

(d − G(x)) ∩V = 0,

Since d − G(x) ⊂ k(x) + IR+ and k(x) ≥ l(x), we have d − G(x) ⊂ l(x) + IR+ and
(l(x) + IR+ ) ∩V = 0/ for all x ∈ (δ /λ )IB, so that (l(λ x) + IR+ ) ∩ λ V = 0/ for all x ∈
(δ /λ )IB. This yields

(l(x) + IR+ ) ∩ ε IB = 0/ for all x ∈ δ IB,

which means that for all x ∈ δ IB there exists some z ≥ l(x) with |z| ≤ ε . The linearity
of l makes l(x) = −l(−x), and therefore |l(x)| ≤ ε for all x ∈ δ IB. This confirms the
continuity of l.
The inclusion d − G(x) − l(x) ⊂ IR+ is by definition equivalent to having d −
y∗ , y − l(x) ≥ 0 whenever x ∈ H −1 (y) − IB. Let x∗ ∈ X ∗ be such that x∗ , x = −l(x)
for all x ∈ X . Then
5 Regularity in Infinite Dimensions 273

y∗ , y − x∗ , x ≤ d for all y ∈ Y and all x ∈ H −1 (y) − IB.

Pick any y ∈ H(x) and λ > 0. Then λ y ∈ H(λ x) and y∗ , λ y − x∗ , λ x ≤ d, or
equivalently,
y∗ , y − x∗ , x ≤ d/λ .
Passing to the limit with λ → ∞, we obtain x∗ ∈ H ∗− (y∗ ). Let now x ∈ IB. Since 0 ∈
H(0), we have 0 ∈ H(−x + IB) and hence y∗ , 0 − x∗ , −x ≤ d. Therefore x∗  ≤
d < r, so that infx∗ ∈H ∗− (y∗ ) x∗  < r. This, combined with (16), gives us the equality
in (15) and hence the equalities in the first line of (13).
We will now confirm the equality in the second line of (13). Suppose H− < r
for some r > 0. Then for any x̃ ∈ IB there is some ỹ ∈ H(x̃) such that ỹ < r. Given
y∗ ∈ IB and x∗ ∈ H ∗+ (y∗ ), we have

x∗ , x̃ ≤ y∗ , ỹ ≤ ỹ < r.

This being valid for arbitrary x̃ ∈ IB, we conclude that x∗  ≤ r, and therefore
H ∗+ + ≤ r.
Suppose now that H ∗+ + < r and pick s > 0 with

x∗  = H ∗+  ≤ s < r,
+
sup
x∗ ∈H ∗+ (IB)

in which case H ∗+ (IB) ⊂ sIB. We will show that

(19) x∗ , x ≤ 1 for all x ∈ H −1 (IB) =⇒ x∗  ≤ s.

The condition on the left of (19) can be written as supy∈IB supx∈H −1 (y) x∗ , x ≤ 1,
which in turn is completely analogous to (17), with d = 1 and H replaced by H −1
and with y and y∗ replaced by x and x∗ , respectively. By repeating the argument in
the first part of the proof after (17), we obtain y∗ ∈ (H −1 )∗− (x∗ ) = (H ∗+ )−1 (x∗ ) with
y∗  ≤ 1. But then x∗ ∈ H ∗+ (IB), and since H ∗+ (IB) ⊂ sIB we have (19).
Now we will show that (19) implies

(20) s−1 IB ⊂ cl H −1 (IB).

If u ∈ cl H −1 (IB), then from 5C.12 there exists x̃∗ ∈ X ∗ with

x̃∗ , u > sup x̃∗ , x ≥ x̃∗ , 0 = 0.

x∈cl H −1 (IB)

Choose λ > 0 such that

sup x̃∗ , x < λ −1 < x̃∗ , u.

x∈cl H −1 (IB)

Then
λ x̃∗ , u > 1 > sup λ x̃∗ , x.
x∈cl H −1 (IB)
274 5 Regularity in Infinite Dimensions

According to (19) this implies that λ x̃∗ ∈ sIB. Thus,

1
s ≥ λ x̃∗  ≥ λ x̃∗ , u/u > ,
u

and therefore u ∈ s−1 IB, so (20) holds.

Our next task is to demonstrate that

(21) int s−1 IB ⊂ int H −1 (IB).

Define the mapping


x → H0 (x) = H(x) for x ∈ IB,
0/ otherwise.
Then gph H0 = gph H ∩ (X × IB) is a closed convex set and rge H0 ⊂ IB. By 5B.1 we
have
int cl H −1 (IB) = int cl dom H0 = int dom H0 = int H −1 (IB).
This equality combined with the inclusion (20) gives us (21). But then r−1 IB ⊂
int s−1 IB ⊂ H −1 (IB), ensuring H− ≤ r. This completes the proof of the second
line in (13).
The above proof can be shortened considerably in the case when X and Y are
reflexive Banach spaces, by utilizing the equality (12).

→ Y with
Exercise 5C.13 (more norm duality). For a sublinear mapping H : X →
closed graph show that
(H ∗+ )−1  = H −1  .
+ −

Exercise 5C.14 (adjoint of a sum). For a sublinear mapping G : X →

→ Y and B ∈
L (X,Y ) prove that

(H + B)∗+ = H ∗+ + B∗ and (H + B)∗− = H ∗− + B∗ .

5D. The Theorems of Lyusternik and Graves

We start with the observation that inequality 5A(7) in Lemma 5A.4, giving an esti-
mate for inverting a perturbed linear mapping A, can also be written in the form
reg A
reg A·B ≤ 1 =⇒ reg (A + B) ≤ ,
1 − reg A·B
5 Regularity in Infinite Dimensions 275

since for an invertible mapping A ∈ L (X ,Y ) one has reg A = A−1. This alter-
native formulation opens the way to extending the estimate to nonlinear and even
set-valued mappings.
First, we recall a basic definition of differentiability in infinite dimensions, which
is just an update of the definition employed in the preceding chapters in finite dimen-
sions. With differentiability as well as Lipschitz continuity and calmness, the only
difference is that the Euclidean norm is now replaced by the norms of the Banach
spaces X and Y that we work with.

Fréchet differentiability and strict differentiability. A function f : X → Y is said

to be Fréchet differentiable at x̄ if x̄ ∈ int dom f and there is a mapping M ∈ L (X,Y )
such that clm ( f − M; x̄) = 0. When such a mapping M exists, it is unique; it is called
the Fréchet derivative of f at x̄ and denoted by D f (x̄), so that

clm ( f − D f (x̄); x̄) = 0.

If actually
lip ( f − D f (x̄); x̄) = 0,
then f is said to be strictly differentiable at x̄.
Partial Fréchet differentiability and partial strict differentiability can be intro-
duced as well on the basis of the partial Lipschitz moduli, by updating the defini-
tions in Section 1D to infinite dimensions. Building on the formulas for the calmness
and Lipschitz moduli, we could alternatively express these definitions in an epsilon-
delta mode as at the beginning of Chapter 1. If a function f is Fréchet differentiable
at every point x of an open set O and the mapping x → D f (x) is continuous from
O to the Banach space L (X ,Y ), then f is said to be continuously Fréchet differ-
entiable on O. Most of the assertions in Section 1D about functions acting in finite
dimensions remain valid in Banach spaces, e.g., continuous Fréchet differentiability
around a point implies strict differentiability at this point.
The extension of the Banach open mapping theorem to nonlinear and set-
valued mappings goes back to the works of Lyusternik and Graves. In 1934,
L. A. Lyusternik published a result, saying that when a function f : X → Y is con-
tinuously Fréchet differentiable in a neighborhood of a point x̄ where f (x̄) = 0 and
its derivative mapping D f (x̄) is surjective, then the tangent manifold to f −1 (0) at x̄
is the set x̄ + ker D f (x̄). In the current setting we adopt the following statement1 of
Lyusternik theorem:

Theorem 5D.1 (Lyusternik). Consider a function f : X → Y that is continuously

Fréchet differentiable in a neighborhood of a point x̄ with the derivative mapping
D f (x̄) surjective. Then, in terms of ȳ := f (x̄), for every ε > 0 there exists δ > 0
such that

d(x, f −1 (ȳ)) ≤ ε x − x̄ whenever x ∈ (x̄ + ker D f (x̄)) and x − x̄ ≤ δ .

1In his paper of 1934 Lyusternik did not state his result as a theorem; the statement in 5D.1 is from
Dmitruk, Milyutin and Osmolovskiı̆ [1980].
276 5 Regularity in Infinite Dimensions

In 1950 L. M. Graves published a result whose formulation and proof we present

here in full, up to some minor adjustments in notation:

Theorem 5D.2 (Graves). Consider a function f : X → Y and a point x̄ ∈ int dom f

and let f be continuous in IBε (x̄) for some ε > 0. Let A ∈ L (X,Y ) be surjective and
let κ ≥ reg A. Suppose there is a nonnegative μ such that μκ < 1 and

(1)  f (x) − f (x ) − A(x − x) ≤ μ x − x whenever x, x ∈ IBε (x̄).

Then, in terms of ȳ := f (x̄) and c = κ −1 − μ, if y is such that y − ȳ ≤ cε, then the
equation y = f (x) has a solution x ∈ IBε (x̄).
Proof. Without loss of generality, let x̄ = 0 and ȳ = f (x̄) = 0. Note that κ > 0, hence
0 < c < ∞. Take y ∈ Y with y ≤ cε . Starting from x0 = 0 we use induction to
construct an infinite sequence {xk }, the elements of which satisfy for all k = 1, 2, . . .
the following three conditions:

(2a) A(xk − xk−1 ) = y − f (xk−1 ),

(2b) xk − xk−1  ≤ κ (κ μ )k−1 y

and

(2c) xk  ≤ y/c.

By (d) in the Banach open mapping theorem 5A.1 there exists x1 ∈ X such that

Ax1 = y and x1  ≤ κ y ≤ y/c.

That is, x1 satisfies all three conditions (2a), (2b) and (2c). In particular, by the
choice of y and the constant c we have x1  ≤ ε .
Suppose now that for some j ≥ 1 we have obtained points xk satisfying (2a), (2b)
and (2c) for k = 1, . . . , j. Then, since y/c ≤ ε , we have from (2c) that all the points
xk satisfy xk  ≤ ε . Again using (d) in 5A.1, we can find x j+1 such that

(3) A(x j+1 − x j ) = y − f (x j ) and x j+1 − x j  ≤ κ y − f (x j ).

If we plug y = Ax j − Ax j−1 + f (x j−1 ) into the second relation in (3) and use (1) for
x = x j and x = x j−1 which, as we already know, are from ε IB, we obtain

x j+1 − x j  ≤ κ  f (x j ) − f (x j−1 ) − A(x j − x j−1) ≤ κ μ x j − x j−1.

Then, by the induction hypothesis,

x j+1 − x j  ≤ κ (κ μ ) j y.

Furthermore,
5 Regularity in Infinite Dimensions 277

j j
κ y
x j+1  ≤ x1  + ∑ xi+1 − xi  ≤ ∑ (κ μ )i κ y ≤ = y/c.
i=1 i=0 1 − κμ

The induction step is complete: we obtain an infinite sequence of points xk satisfying

(2a), (2b) and (2c). For any k and j with k > j > 1 we have
∞
k−1 k−1
κ y
xk −x j  ≤ ∑ xi+1 −xi  ≤ ∑(κ μ )i κ y ≤ (κ μ ) j κ y ∑(κ μ )i ≤ 1 − κ μ (κ μ ) j .
i= j i= j i=0

Thus, {xk } is a Cauchy sequence, hence convergent to some x, and then, passing to
the limit with k → ∞ in (2a) and (2c), this x satisfies y = f (x) and x ≤ y/c. The
final inequality gives us x ≤ ε and the proof is finished.
Observe that in the Graves theorem no differentiability of the function f is re-
quired, but only “approximate differentiability” as in the theorem of Hildebrand and
Graves; see the commentary to Chapter 1. If we suppose that for every μ > 0 there
exists ε > 0 such that (1) holds for every x, x ∈ IBε (x̄), then A is, by definition, the
strict derivative of f at x̄, A = D f (x̄). That is, the Graves theorem encompasses the
following special case: if f is strictly differentiable at x̄ and its derivative D f (x̄) is
onto, then there exist ε > 0 and c > 0 such that for every y ∈ Y with y − ȳ ≤ cε
there is an x ∈ X such that x − x̄ ≤ ε and y = f (x).
The statement of the Graves theorem above does not reflect all the information
that can be extracted from its proof. In particular, a solution x of f (x) = y which not
only is in the ball IBε (x̄) but also satisfies x − x̄ ≤ y − ȳ/c. Taking into account
that x ∈ f −1 (y), which yields d(x̄, f −1 (y)) ≤ x − x̄, along with the form of the
constant c, we get
κ
d(x̄, f −1 (y)) ≤ y − f (x̄).
1 − κμ
Furthermore, this inequality actually holds not only at x̄ but also for all x close to x̄,
and this important extension is hidden in the proof of the theorem.
Indeed, let (1) hold for x, x ∈ IBε (x̄) and choose a positive τ < ε . Then there is
a neighborhood U of x̄ such that IBτ (x) ⊂ IBε (x̄) for all x ∈ U. Make U smaller if
necessary so that  f (x) − f (x̄) < cτ for x ∈ U. Pick x ∈ U and a neighborhood V
of ȳ such that y − f (x) ≤ cτ for y ∈ V . Then, remembering that in the proof x̄ = 0,
modify the first induction step in the following way: there exists x1 ∈ X such that

Ax1 = y − f (x) + Ax and x1 − x ≤ κ y − f (x).

Then, construct a sequence {xk } with x0 = x satisfying (3), thereby obtaining

xk − xk−1  ≤ κ (κ μ )k−1 y − f (x)

and then
k i
κ
(4) xk − x ≤ ∑ xi − xi−1  ≤ κ y − f (x) ∑ (κ μ )i−1 ≤ y − f (x).
i=1 i=1 1 − κμ
278 5 Regularity in Infinite Dimensions

Thus,
xk − x ≤ y − f (x)/c ≤ τ .
The sequence {xk } is a Cauchy sequence, and therefore convergent to some x̃. In
passing to the limit in (4) we get
κ
x̃ − x ≤ y − f (x).
1 − κμ

Since x̃ ∈ f −1 (y), we see that, under the conditions of the Graves theorem, there
exist neighborhoods U of x̄ and V of f (x̄) such that
κ
(5) d(x, f −1 (y)) ≤ y − f (x) for (x, y) ∈ U × V.
1 − κμ

The property described in (5) is something we know from Chapter 3: this is metric
regularity of the function f at x̄ for ȳ. Noting that the μ in (1) satisfies μ ≥ lip ( f −
A)(x̄) we arrive at the following result:

Theorem 5D.3 (updated Graves’ theorem). Let f : X → Y be continuous in a neigh-

borhood of x̄, let A ∈ L (X ,Y ) satisfy reg A ≤ κ < ∞, and suppose lip ( f −A)(x̄) ≤ μ
for some μ with μκ < 1. Then
κ
(6) reg ( f ; x̄ | ȳ) ≤ for ȳ = f (x̄).
1 − κμ

For f = A + B we obtain from this result the estimation for perturbed inversion
of linear mappings in 5A.4.
The version of the Lyusternik theorem2 stated as Theorem 5D.1 can be derived
from the updated Graves theorem 5D.3. Indeed, the assumptions of 5D.1 are clearly
stronger than those of Theorem 5D.3. From (5) with y = f (x̄) we get
κ
(7) d(x, f −1 (ȳ)) ≤  f (x) − f (x̄)
1 − κμ

for all x sufficiently close to x̄. Let ε > 0 and choose δ > 0 such that

(1 − κ μ )ε
(8)  f (x) − f (x̄) + D f (x̄)(x − x̄) ≤ x − x̄ whenever x ∈ IBδ (x̄).
κ
But then for any x ∈ (x̄ + ker D f (x̄)) ∩ IBδ (x̄), from (7) and (8) we obtain
κ
d(x, f −1 (y)) ≤  f (x) − f (x̄) ≤ ε x − x̄,
1 − κμ

which is the conclusion of 5D.1.

2 The iteration (3), which is a key step in the proof of Graves, is also present in the original proof
of Lyusternik [1934], see also Lyusternik and Sobolev [1965]. In the case when A is invertible, it
goes back to Goursat [1903], see the commentary to Chapter 1.
5 Regularity in Infinite Dimensions 279

Exercise 5D.4 (correction function version of Graves theorem). Show that on the
conditions of Theorem 5D.3, for ȳ = f (x̄) there exist neighborhoods U of x̄ and V
of ȳ such that for every y ∈ V and x ∈ U there exists ξ with the property
κ
f (ξ + x) = y and ξ  ≤  f (x) − y.
1 − κμ

Guide. From Theorem 5D.3 we see that there exist neighborhoods U of x̄ and V of
ȳ such that for every x ∈ U and y ∈ V
κ
d(x, f −1 (y)) ≤ y − f (x).
1 − κμ

Without loss of generality, let y = f (x); then we can slightly increase μ so that the
latter inequality becomes strict. Then there exists η ∈ f −1 (y) such that x − η  ≤
κ /(1 − κ μ )y − f (x). Next take ξ = η − x.
If the function f in 5D.3 is strictly differentiable at x̄, we can choose A = D f (x̄),
and then μ = 0. In this case (6) reduces to

(9) reg ( f ; x̄ | ȳ) ≤ reg D f (x̄) for ȳ = f (x̄).

In the following section we will show that this inequality actually holds as equality:

Theorem 5D.5. For a function f : X → Y which is strictly differentiable at x̄, one

has
reg ( f ; x̄ | ȳ) = reg D f (x̄).
We terminate this section with two major observations. For the first, assume that
the mapping A in Theorem 5D.2 is not only surjective but also invertible. Then the
iteration procedure used in the proof of the Graves theorem becomes the iteration
used first by Goursat (see the commentary to Chapter 1), namely

x j+1 = x j − A−1 ( f (x j ) − y).

In that case, one obtains the existence of a single-valued graphical localization of the
inverse f −1 around f (x̄) for x̄. If the derivative mapping D f (x̄) is merely surjective,
as assumed in the Graves theorem, the inverse f −1 may not have a single-valued
graphical localization at ȳ for x̄ but, still, this inverse, being a set-valued mapping,
has the Aubin property at ȳ for x̄.
Our second observation is that in the proof of Theorem 5D.2 we use the linearity
of the mapping A only to apply the Banach open mapping theorem. But we can em-
ploy the regularity modulus for any, even set-valued, mapping. After this somewhat
historical section, we will explore this idea further in the section which follows.
280 5 Regularity in Infinite Dimensions

5E. Metric Regularity in Metric Spaces

In this section we show that the updated Graves theorem 5D.3, in company with the
stability of metric regularity under perturbations, demonstrated in Theorem 3F.1,
can be extended to a much broader framework of set-valued mappings acting in ab-
stract spaces. Specifically, we consider a set-valued mapping F acting from a metric
space (X , ρ ) to another metric space (Y, σ ). In such spaces the standard definitions,
e.g. of the ball in X with center x and radius r and the distance from a point x to a
set C in Y , need only be adapted to metric notation:
  
IBr (x̄) = x ∈ X  ρ (x, x̄) ≤ r , d(x,C) = inf ρ (x, x ).
x ∈C

Recall that a subset C of a complete metric space is closed when d(x,C) = 0 ⇒ x ∈

C. Also recall that a set C is locally closed at a point x ∈ C if there is a neighborhood
U of x such that the intersection C ∩U is closed.
In metric spaces (X, ρ ) and (Y, σ ), the definition of the Lipschitz modulus of a
function g : X → Y is extended in a natural way, with attention paid to the metric
notation of distance:
σ (g(x), g(x ))
lip (g; x̄) = lim sup .
x,x →x̄, ρ (x, x )
x=x

For set-valued mappings F acting in such spaces, the definitions of metric regularity
and the Aubin property persist in the same manner. The equivalence of metric regu-
larity with the Aubin property of the inverse (Theorem 3E.4) with the same constant
remains valid as well.
It will be important for our efforts to take the metric space X to be complete and
to suppose that Y is a linear space equipped with a shift-invariant metric σ . Shift
invariance means that

σ (y + z, y + z) = σ (y, y ) for all y, y , z ∈ Y.

Of course, any Banach space meets these requirements.

The result stated next is just a reformulation of Theorem 3F.1 for mappings acting
in metric spaces:

Theorem 5E.1 (inverse mapping theorem for metric regularity in metric spaces).
Let (X, ρ ) be a complete metric space and let (Y, σ ) be a linear space with shift-
invariant metric σ . Consider a mapping F : X →
→ Y and any (x̄, ȳ) ∈ gph F at which
gph F is locally closed, and let κ and μ be nonnegative constants such that

reg (F; x̄| ȳ) ≤ κ and κ μ < 1.

Then for any function g : X → Y with x̄ ∈ int dom g and lip (g; x̄) ≤ μ, one has
5 Regularity in Infinite Dimensions 281

κ
(1) reg (g + F; x̄|g(x̄) + ȳ) ≤ .
1 − κμ

Before arguing this, we note that it immediately allows us to supply 5C.9 and
5D.5 with proofs.
Proof of 5C.9. We apply 5E.1 with X and Y Banach spaces, F = H, x̄ = 0 and
ȳ = 0. According to 5C.1, reg(H; 0|0) = H −1 − , so 5E.1 tells us that for any
κ > H −1 − , any B ∈ L (X,Y ) with B < 1/κ , and any μ with B < μ < 1/κ
one has from (1) that (H + B)−1 − ≤ κ /(1 − κ μ ). It remains only to pass to the
limit as κ → H −1 − and μ → B.
Proof of 5D.5. To obtain the inequality opposite to 5D(9), choose F = f and g =
D f (x̄) − f and apply 5E.1, in this case with μ = 0.
We proceed now with presenting two separate proofs of Theorem 5E.1, which
echo on a more abstract level the way we proved the classical inverse function the-
orem 1A.1 in Chapter 1. The first proof uses an iteration in line with the original
argument in the proof of the Graves theorem 5D.2, while the second proof is based
on a contraction mapping principle for set-valued mappings.

Proof I of Theorem 5E.1. Let κ and μ be as in the statement of the theorem and
choose a function g : X → Y with lip (g; x̄) ≤ μ . Without loss of generality, suppose
g(x̄) = 0. Let λ > κ and ν > μ satisfy λ ν < 1. Let α > 0 be small enough that the
set gph F ∩ (IBα (x̄) × IBα (ȳ)) is closed, g is Lipschitz continuous with constant ν on
IBα (x̄), and

(2) d(x, F −1 (y)) ≤ λ d(y, F(x)) for all (x, y) ∈ IBα (x̄) × IBα (ȳ).

From (2) with x = x̄, it follows that

(3) F −1 (y) = 0/ for all y ∈ IBα (ȳ).

Having fixed λ , α and ν , consider the following system of inequalities:

⎧
⎪
⎪ λ ν + ε < 1,
⎨ 1
[(1 + λ ν )a + λ b + ε ] + a ≤ α ,
(4) 1−(λ ν
+ε ) 
⎪
⎪
⎩ b + ν 1−(λ ν +ε ) [(1 + λ ν )a + λ b + ε ] + a ≤ α .
1

It is not difficult to see that there are positive a, b and ε that satisfy this system.
Indeed, first fix ε such that these inequalities hold strictly for a = b = 0; then pick
sufficiently small a and b so that both the second and the third inequality are not
violated.
Let x ∈ IBa (x̄) and y ∈ IBb (ȳ). We will show that

λ
(5) d(x, (g + F)−1 (y)) ≤ d(y, (g + F)(x)).
1−λν
282 5 Regularity in Infinite Dimensions

Since x and y are arbitrarily chosen in the corresponding balls around x̄ and ȳ, and
λ and ν are arbitrarily close to κ and μ , respectively, this gives us (1).
According to the choice of a and of b in (4), we have

(6) σ (y − g(x), ȳ) ≤ ν ρ (x, x̄) + σ (y, ȳ) ≤ ν a + b ≤ α .

Through (3) and (6), there exists z1 ∈ F −1 (y − g(x)) such that

(7) ρ (z1 , x) ≤ d(x, F −1 (y − g(x))) + ε ≤ λ d(y, (g + F)(x)) + ε .

If z1 = x, then x ∈ F −1 (y − g(x)), which is the same as x ∈ (g + F)−1 (y). Then (5)

holds automatically, since its left side is 0. Let z1 = x. In this case, using (2), we
obtain
ρ (z1 , x) ≤ ρ (x, x̄) + d(x̄, F −1 (y − g(x))) + ε
≤ ρ (x, x̄) + λ d(y − g(x), F(x̄)) + ε
(8) ≤ ρ (x, x̄) + λ σ (y, ȳ) + λ σ (g(x), g(x̄)) + ε
≤ ρ (x, x̄) + λ σ (y, ȳ) + λ ν ρ (x, x̄) + ε
≤ (1 + λ ν )a + λ b + ε .

Hence, by (4),

(9) ρ (z1 , x̄) ≤ ρ (z1 , x) + ρ (x, x̄) ≤ (1 + λ ν )a + λ b + ε + a ≤ α .

By induction, we construct a sequence of vectors zk ∈ IBα (x̄), with z0 = x, such

that, for k = 0, 1, . . .,

(10) zk+1 ∈ F −1 (y − g(zk )) and ρ (zk+1 , zk ) ≤ (λ ν + ε )k ρ (z1 , x).

We already found z1 which gives us (10) for k = 0. Suppose that for some n ≥ 1 we
have generated z1 , z2 , . . . , zn satisfying (10). If zn = zn−1 then zn ∈ F −1 (y − g(zn ))
and hence zn ∈ (g + F)−1 (y). Then, by using (2), (7) and (10), we get
n−1
d(x, (g + F)−1 (y)) ≤ ρ (zn , x) ≤ ∑ ρ (zi+1 , zi )
i=0
n−1
1
≤ ∑ (λ ν + ε )i ρ (z1 , x) ≤ 1 − (λ ν + ε ) ρ (z1 , x)
i=0
 
λ ε
≤ d(y, (g + F)(x)) + .
1 − (λ ν + ε ) λ

Since the left side of this inequality does not depend on the ε on the right, we are
able to obtain (5) by letting ε go to 0.
Assume zn = zn−1 . We will first show that zi ∈ IBα (x̄) for all i = 2, 3, . . . , n. Uti-
lizing (10), for such an i we have
5 Regularity in Infinite Dimensions 283

i−1 i−1
1
ρ (zi , x) ≤ ∑ ρ (z j+1, z j ) ≤ ∑ (λ ν + ε ) j ρ (z1 , x) ≤ 1 − (λ ν + ε ) ρ (z1 , x)
j=0 j=0

and therefore, through (8) and (4),

1
(11) ρ (zi , x̄) ≤ ρ (zi , x) + ρ (x, x̄) ≤ [(1 + λ ν )a + λ b + ε ] + a ≤ α .
1 − (λ ν + ε )

Thus, we have zi ∈ IBα (x̄) for all i = 1, . . . , n.

Taking into account the estimate in (11) for i = n and the third inequality in (4),
we get

σ (y − g(zn), ȳ) ≤ σ (y, ȳ) + ν ρ (zn , x̄)

 
1
≤ b+ν [(1 + λ ν )a + λ b + ε ] + a ≤ α .
1 − (λ ν + ε )

Since ρ (zn , zn−1 ) > 0, from (3) there exists zn+1 ∈ F −1 (y − g(zn )) such that
 
ρ (zn+1 , zn ) ≤ d zn , F −1 (y − g(zn )) + ερ (zn , zn−1 ),

and then (2) yields

ρ (zn+1 , zn ) ≤ λ d(y − g(zn ), F(zn )) + ερ (zn , zn−1 ).

Since zn ∈ F −1 (y − g(zn−1 )) and hence y − g(zn−1 ) ∈ F(zn ), by invoking the induc-

tion hypothesis, we obtain

ρ (zn+1 , zn ) ≤ λ σ (g(zn ), g(zn−1 )) + ε ρ (zn , zn−1 )

≤ (λ ν + ε ) ρ (zn , zn−1 ) ≤ (λ ν + ε )n ρ (z1 , x).

The induction is complete, and therefore (10) holds for all k.

Right after (10) we showed that when zk = zk−1 for some k then (5) holds. Sup-
pose now that zk+1 = zk for all k. By virtue of the second condition in (10), we see
for any natural n and m with m < n that
n−1 n−1
ρ (z1 , x)
ρ (zn , zm ) ≤ ∑ ρ (zk+1 , zk ) ≤ ∑ (λ ν + ε )k ρ (z1 , x) ≤ 1 − (λ ν + ε )
(λ ν + ε )m .
k=m k=m

We conclude that the sequence {zk } satisfies the Cauchy condition, and all its el-
ements are in IBα (x̄). Hence this sequence converges to some z ∈ IBα (x̄) which,
from (10) and the local closedness of gph F, satisfies z ∈ F −1 (y − g(z)), that is,
z ∈ (g + F)−1 (y). Moreover,
k
d(x, (g + F)−1 (y)) ≤ ρ (z, x) = lim ρ (zk , x) ≤ lim
k→∞
∑ ρ (zi+1 , zi )
k→∞ i=0
284 5 Regularity in Infinite Dimensions

k
1
≤ lim
k→∞
∑ (λ ν + ε )i ρ (z1 , x) ≤ 1 − (λ ν + ε ) ρ (z1 , x)
i=0
 
1
≤ λ d(y, (g + F)(x)) + ε ,
1 − (λ ν + ε )

the final inequality being obtained from (2) and (7). Taking the limit as ε → 0 we
obtain (5), and the proof is finished.
The second proof of Theorem 5E.1 uses the following extension of the contrac-
tion mapping principle 1A.2 for set-valued mappings, furnished with a proof the
idea of which goes back to Banach [1922], if not earlier.

Theorem 5E.2 (contraction mapping principle for set-valued mappings). Let (X, ρ )
be a complete metric space, and consider a set-valued mapping Φ : X → → X and a
point x̄ ∈ X . Suppose that there exist scalars a > 0 and λ ∈ (0, 1) such that the set
gph Φ ∩ (IBa (x̄) × IBa (x̄)) is closed and
(a) d(x̄, Φ (x̄)) < a(1 − λ );
1
(b) e(Φ (u) IBa (x̄), Φ (v)) ≤ λ ρ (u, v) for all u, v ∈ IBa (x̄).
Then Φ has a fixed point in IBa (x̄); that is, there exists x ∈ IBa (x̄) such that x ∈ Φ (x).
Proof. By assumption (a) there exists x1 ∈ Φ (x̄) such that ρ (x1 , x̄) < a(1 − λ ).
Proceeding by induction, let x0 = x̄ and suppose that there exists xk+1 ∈ Φ (xk ) ∩
IBa (x̄) for k = 0, 1, . . . , j − 1 with

ρ (xk+1 , xk ) < a(1 − λ )λ k .

By assumption (b),

d(x j , Φ (x j )) ≤ e(Φ (x j−1 ) ∩ IBa (x̄), Φ (x j )) ≤ λ ρ (x j , x j−1 ) < a(1 − λ )λ j .

This implies there is an x j+1 ∈ Φ (x j ) such that

ρ (x j+1 , x j ) < a(1 − λ )λ j .

By the triangle inequality,

j j
ρ (x j+1 , x̄) ≤ ∑ ρ (xi+1 , xi ) < a(1 − λ ) ∑ λ i < a.
i=0 i=0

Hence x j+1 ∈ Φ (x j ) ∩ IBa (x̄) and the induction step is complete.

For any k > m > 1 we then have
k−1 k−1
ρ (xk , xm ) ≤ ∑ ρ (xi+1, xi ) < a(1 − λ ) ∑ λ i < aλ m .
i=m i=m
5 Regularity in Infinite Dimensions 285

Thus, {xk } is a Cauchy sequence and consequently converges to some x ∈ IBa (x̄).
Since (xk−1 , xk ) ∈ gph Φ ∩ (IBa (x̄) × IBa (x̄)) which is a closed set, we conclude that
x ∈ Φ (x).
For completeness, we now supply with a proof the (standard) contraction map-
ping principle 1A.2.

Proof of Theorem 1A.2. Let Φ be a function which is Lipschitz continuous on

IBa (x̄) with constant λ ∈ [0, 1) and let ρ (x̄, Φ (x̄)) ≤ a(1 − λ ). By repeating the ar-
gument in the proof of 5E.2 for the sequence of points xk satisfying xk+1 = Φ (xk ),
k = 0, 1, . . . , x0 = x̄, with all strict inequalities replaced by non-strict ones, we obtain
that Φ has a fixed point in IBa (x̄). Suppose that Φ has two fixed points in IBa (x̄), that
is, there are x, x ∈ IBa (x̄), x = x , with x = Φ (x) and x = Φ (x ). Then we have

0 < ρ (x, x ) = ρ (Φ (x), Φ (x )) ≤ λ ρ (x, x ) < ρ (x, x ),

which is absurd. Hence, in this case Φ has a unique fixed point.

Proof II of Theorem 5E.1. Pick the constants κ , μ , ν and λ and a function g as in

the beginning of Proof I. Then, utilizing 5A.3, there exist positive constants a and b
such that

(12) e(F −1 (y ) ∩ IBa (x̄), F −1 (y)) ≤ λ σ (y , y) for all y , y ∈ IBb+ν a (ȳ),

(13) the set gph F ∩ (IBa (x̄) × IBb+ν a (ȳ)) is closed

and also
σ (g(x ), g(x)) ≤ ν ρ (x , x) for all x , x ∈ IBa (x̄).
Take any λ + > λ and make b > 0 smaller if necessary so that

λ +b
(14) < a/4.
1−λν
For any y ∈ IBb (g(x̄) + ȳ) and x ∈ IBa (x̄), using the shift-invariance of the metric σ
and the triangle inequality, we obtain

(15) σ (−g(x) + y, ȳ) = σ (y, g(x) + ȳ) ≤ σ (y, g(x̄) + ȳ) + σ (g(x), g(x̄)) ≤ b + ν a.

Fix y ∈ IBb (g(x̄) + ȳ) and consider the mapping

Φy : x → F −1 (−g(x) + y) for x ∈ IBa (x̄).

Let y, y ∈ IBb (g(x̄)+ ȳ), y = y , and let x ∈ (g+F)−1 (y )∩IBa/2 (x̄). We will establish
now that there is a fixed point x ∈ Φy (x) in the closed ball centered at x with radius

λ + σ (y, y )
ε := .
1−λν
286 5 Regularity in Infinite Dimensions

Since x ∈ F −1 (−g(x ) + y ) ∩ IBa (x̄) and both (x , y ) and (x , y) satisfy (15), from
(12) we get

d(x , Φy (x )) ≤ e(F −1 (−g(x ) + y) ∩ IBa (x̄), F −1 (−g(x ) + y))

≤ λ σ (y , y) < λ + σ (y , y) = ε (1 − λ ν ).

By the triangle inequality and (14), ε ≤ λ + (2b)/(1 − λ ν ) < a/2, so that IBε (x ) ⊂
IBa (x̄). Then we have that for any u, v ∈ IBε (x ),

e(Φy (u) ∩ IBε (x ), Φy (v)) ≤ e(F −1 (−g(u) + y) ∩ IBa (x̄), F −1 (−g(v) + y))
≤ λ σ (g(u), g(v)) ≤ λ ν ρ (u, v).

By (13) the set gph Φy ∩ (IBε (x̄) × IBε (x̄)) is closed; hence we can apply the contrac-
tion mapping principle in Theorem 5E.2 to the mapping Φy , with constants a = ε
and the λ taken to be the λ ν here, to obtain the existence of a fixed point x ∈ Φy (x)
within distance ε from x . Since x ∈ (g + F)−1 (y), we obtain

λ+
d(x , (g + F)−1 (y)) ≤ ε = σ (y , y).
1−λν

This tells us that (g + F)−1 has the Aubin property at g(x̄) + ȳ for x̄ with constant
λ + /(1 − λ ν ). Hence, by 5A.3, the mapping g + F is metrically regular at x̄ for
g(x̄) + ȳ with the same constant. Since x and y are arbitrarily chosen in the corre-
sponding balls around x̄ and ȳ, λ + and λ are arbitrarily close to κ , and ν is arbitrarily
close to μ , this comes down to (1).
Next, we put together an implicit function version of Theorem 5E.1.

Theorem 5E.3 (implicit mapping theorem for metric regularity in metric spaces).
Let (X , ρ ) be a complete metric space and let (Y, σ ) be a linear metric space with
→Y,
shift-invariant metric. Let (P, π ) be a metric space. For f : P× X → Y and F : X →
consider the generalized equation f (p, x) + F(x)  0 with solution mapping
  
S(p) = x  f (p, x) + F(x)  0 having x̄ ∈ S( p̄).

Let h : X → Y be a strict estimator of f with respect to x uniformly in p at ( p̄, x̄)

with constant μ, let gph (h + F) be locally closed at (x̄, 0) and suppose that h + F is
metrically regular at x̄ for 0 with reg (h + F; x̄|0) ≤ κ. Assume

 ( f ; ( p̄, x̄)) ≤ γ < ∞.

κ μ < 1 and lip p

Then S has the Aubin property at p̄ for x̄, and moreover

κγ
(16) lip (S; p̄| x̄) ≤ .
1 − κμ

Proof. Choose λ > κ and ν > μ such that λ ν < 1. Also, let β > γ . Then there exist
positive scalars α and τ such that the set gph (h + F) ∩ (IBα (x̄) × IBα (0)) is closed,
5 Regularity in Infinite Dimensions 287
 
(17) e (h + F) (y ) ∩ IBα (x̄), (h + F) (y) ≤ λ σ (y , y) for all y , y ∈ IBα (0),
−1  −1

(18) σ (r(p, x ), r(p, x)) ≤ ν ρ (x , x) for all x , x ∈ IBα (x̄) and p ∈ IBτ ( p̄)

where r(p, x) = f (p, x) − h(x), and

(19) σ ( f (p , x), f (p, x)) ≤ β π (p , p) for all p , p ∈ IBτ ( p̄) and x ∈ IBα (x̄).

Let
2λ β λβ
(20) ≥ λ+ > .
1−λν 1−λν
Now, choose positive a < α and then positive q ≤ τ such that

4λ β q
(21) νa + β q ≤ α and + a ≤ α.
1−λν
Then, from (18) and (19), for every x ∈ IBa (x̄) and p ∈ IBq ( p̄) we have

σ (r(p, x), 0) ≤ σ (r(p, x), r(p, x̄)) + σ (r(p, x̄), r( p̄, x̄))
(22)
≤ ν ρ (x, x̄) + β π (p, p̄) ≤ ν a + β q ≤ α .

Fix p ∈ IBq ( p̄) and consider the mapping

Φ p : x → (h + F)−1 (−r(p, x)) for x ∈ IBα (x̄).

Observe that for any x ∈ IBa (x̄) and p ∈ IBq ( p̄), x ∈ Φ p (x) ⇐⇒ x ∈ S(p) and also
that the set gph Φ p ∩ (IBα (x̄) × IBα (x̄)) is closed. Let p , p ∈ IBq ( p̄) with p = p and
let x ∈ S(p ) ∩ IBa (x̄). Let ε := λ + π (p , p); then ε ≤ λ + (2q). Thus, remembering
that x ∈ Φ p (x ) ∩ IBα (x̄), from (17), where we use (22), and from (18), (19) and
(20), we deduce that
 
d(x , Φ p (x )) ≤ e (h + F)−1 (−r(p , x )) ∩ IBα (x ), (h + F)−1 (−r(p, x ))
≤ λ σ ( f (p , x ), f (p, x )) ≤ λ β π (p , p) < λ + π (p , p)(1 − λ ν ) = ε (1 − λ ν ).

Since x ∈ IBa (x̄) and, by (20) and (21),

4λ β q
ε ≤ 2λ + q ≤ ,
1−λν
we get IBε (x ) ⊂ IBα (x̄). Then, for any u, v ∈ IBε (x ) using again (17) (with (22)) and
(18), we see that
288 5 Regularity in Infinite Dimensions

(x ), Φ p (v))
e(Φ p (u) ∩ IBε  
≤e (h + F)−1 (−r(u, p)) ∩ IBε (x̄), (h + F)−1 (−r(v, p))
≤ λ σ (r(p, u), r(p, v)) ≤ λ ν ρ (u, v) .

Hence the contraction mapping principle in Theorem 5E.2 applies, with the λ there
taken to be the λ ν here, and it follows that there exists x ∈ Φ p (x) ∩ IBε (x ) and hence
x ∈ S(p) ∩ IBε (x ). Thus,

d(x , S(p)) ≤ ρ (x , x) ≤ ε = λ + π (p , p).

Since this inequality holds for any x ∈ S(p ) ∩ IBa (x̄) and any λ + fulfilling (20), we
arrive at
e(S(p ) ∩ IBa (x̄), S(p)) ≤ λ + π (p , p).
That is, S has the Aubin property at p̄ for x̄ with modulus not greater than λ + . Since
λ + can be arbitrarily close to λ /(1 − λ ν ), and λ , ν and β can be arbitrarily close
to κ , μ and γ , respectively, we achieve the estimate (16).

We can also state Theorem 3F.9 in Banach spaces with only minor adjustments
in notation and terminology.

Theorem 5E.4 (using strict differentiability and ample parameterization). Let X , Y

and P be Banach spaces. For f : P × X → Y and F : X → → Y , consider the generalized
equation f (p, x) + F(x)  0 with solution mapping S and a pair ( p̄, x̄) with x̄ ∈ S( p̄).
Suppose that f is strictly differentiable at ( p̄, x̄) and that gph F is locally closed at
(x̄, − f ( p̄, x̄)). If the mapping

h + F for h(x) = f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄)

is metrically regular at x̄ for 0, then S has the Aubin property at p̄ for x̄ with

lip (S; p̄| x̄) ≤ reg (h + F; x̄|0) · D p f ( p̄, x̄).

Furthermore, when f satisfies the ample parameterization condition:

the mapping D p f ( p̄, x̄) is surjective,

then the converse implication holds as well: the mapping h + F is metrically regular
at x̄ for 0 provided that S has the Aubin property at p̄ for x̄.
The following exercise, which we supply with a detailed guide, deals with a
more general kind of perturbation and shows that not only the constant but also the
neighborhoods of metric regularity of the perturbed mapping can be independent of
the perturbation provided that its Lipschitz constant is “sufficiently small.” For better
transparency, we consider mappings with closed graphs acting in Banach spaces.

Exercise 5E.5. Let X and Y be Banach spaces and consider a continuous function
f : X → Y , a mapping F : X →
→ Y with closed graph and a point x̄ ∈ X such that
5 Regularity in Infinite Dimensions 289

0 ∈ f (x̄) + F(x̄). Let κ and μ be positive constants such that

reg (F; x̄| ȳ) < κ and κ μ < 1.

Prove that for any κ  satisfying

κ
(23) < κ
1 − κμ

there exist positive constants α and β such that for every mapping A : X × X → Y
and every x̃ ∈ IBα (x̄) and ỹ ∈ IBβ (0) with the properties that

ỹ ∈ A(x̃, x̃) + F(x̃), A(x̃, u) − f (x̃) ≤ β for every u ∈ IBα (x̃)

and

(x , u) − f (x )] − [A(x, u) − f (x)] ≤ μ x − x for every x , x, u ∈ IBα +5κ  β (x̃),

we have that for any u ∈ IBα (x̃) the mapping A(·, u) + F(·) is metrically regular at x̃
for ỹ with constant κ  and neighborhoods IBα (x̃) and IBβ (ỹ).
Guide. Let a and b be positive constants such that f + F is metrically regular with
constant κ and neighborhoods IBa (x̄) and IBb (0). Choose κ  satisfying (23) and then
positive α and β such that

α ≤ 2κ  β , 2α + 5κ β ≤ a and 6β + μα ≤ b.

Pick x, u ∈ IBα (x̃) and y ∈ IBβ (ỹ). First prove that

d(x, (A(·, u) + F(·))−1 (y)) ≤ κ  y − y

(24)
for any y ∈ (A(x, u) + F(x)) ∩ IB4β (ỹ).

Let y ∈ (A(x, u) + F(x)) ∩ IB4β (ỹ), y = y. Observe that

 − A(x, u) + f (x) + y ≤ y − ỹ + ỹ +  − A(x, u) + f (x) + A(x̃, u) − f (x̃)
+A(x̃, u) − A(x̃, x̃) + A(x̃, x̃) − f (x̃)
≤ 4β + μα + 2β ≤ 6β + μα ≤ b.

The same estimate holds of course with y replaced by y; that is, both −A(x, u) +
f (x) + y and −A(x, u) + f (x) + y are in IBb (0). Consider the mapping

Φ : v → ( f + F)−1 (−A(v, u) + f (v)) + y) for v ∈ IBα (x̃).

Since IBα (x̃) ⊂ IBa (x̄), utilizing the metric regularity of f + F we obtain

d(x, Φ (x)) = d(x, ( f + F)−1 (−A(x, u) + f (x) + y))

≤ κ d(−A(x, u) + f (x) + y, ( f + F)(x))
290 5 Regularity in Infinite Dimensions

≤ κ  − A(x, u) + f (x) + y − (y − A(x, u) + f (x))

= κ y − y < κ  y − y(1 − κ μ ) = r(1 − κ μ ),

where r := κ  y − y . Observe that r ≤ 5κ  β and IBr (x) ⊂ IBα +5κ  β (x̃) ⊂ IBa (x̄).
Again, by utilizing the metric regularity of f + F we get that for every v, w ∈ IBr (x),

e(Φ (v) ∩ IBr (x), Φ (w))

≤ sup d(z, ( f + F)−1 (−A(w, u) + f (w) + y))
z∈( f +F)−1 (−A(v,u)+ f (v)+y)∩IBa (x̄)
≤ sup κ d(−A(w, u) + f (w) + y, f (z) + F(z))
z∈( f +F)−1 (−A(v,u)+ f (v)+y)∩IBa (x̄)
≤ κ  − A(v, u) + f (v) − [−A(w, v) + f (w)] ≤ κ μ v − w.

Noting that the set gph Φ ∩ (IBr (x) × IBr (x)) is closed, Theorem 5E.2 then yields the
existence of a fixed point x̂ ∈ Φy (x̂) ∩ IBr (x); that is,

y ∈ A(x̂, u) + F(x̂) and x̂ − x ≤ κ  y − y.

Since x̂ ∈ (A(·, u) + F(·))−1 (y) we obtain (24).

Now we are ready to show the desired inequality

(25) d(x, (A(·, u) + F(·))−1 (y)) ≤ κ  d(y, A(x, u) + F(x)).

If A(x, u)+ F(x) = 0/ there is nothing to prove. If not, choose ε > 0 and w ∈ A(x, y)+
F(x) such that w− y ≤ d(y, A(x, u)+ F(x))+ ε . If w ∈ IB4β (ỹ), then from (25) with
y = w we get

d(x, (A(·, u) + F(·))−1 )(y)) ≤ κ  w − y ≤ κ  d(y, A(x, u) + F(x)) + κ  ε ,

which yields (25) since the left side does not depend on ε . Otherwise, we have

y − w ≥ w − ỹ − y − ỹ ≥ 4β − β = 3β .

But then,

d(x, (A(·, u) + F(·))−1 (y)) ≤ α + d(x̃, (A(·, u) + F(·))−1 (y))

≤ α + κ y − ỹ ≤ α + κ β ≤ 3κ  β ≤ κ  y − w
≤ κ  d(y, A(x, u) + F(x)) + κ  ε ,

which leads (25) by letting ε → 0.

5 Regularity in Infinite Dimensions 291

Example 5E.6 (counterexample for set-valued perturbations). Consider F : IR →

→ IR
and G : IR →
→ IR specified by

F(x) = {−2x, 1} and G(x) = {x2 , −1} for x ∈ IR.

Then F is metrically regular at 0 for 0 while G has the Aubin property at 0 for 0.
Moreover, both F −1 and G are Lipschitz continuous (with respect to the Pompeiu-
Hausdorff distance) on the whole of IR. We have reg (F; 0|0) = 1/2 whereas the
Lipschitz modulus of the single-valued localization of G around 0 for 0 is 0 and
serves also as the infimum of all Aubin constants. The mapping

(F + G)(x) = {x2 − 2x, x2 + 1, −2x − 1, 0} for x ∈ IR

is not metrically regular at 0 for 0. Indeed, (F −1

√ + G) has a single-valued local-
around (0, 0) of the form s(y) = 1 − 1 + y, so that d(x, (F + G)−1 (y)) =
ization s√
|x − 1 + 1 + y|, but also d(y, (F √ + G)(x)) = min{|x2 − 2x − y|, y}. Take x = ε > 0
and y = ε . Then, since (ε − 1 + 1 + ε 2 )/ε 2 → ∞, the mapping F + G is seen not
2

to be metrically regular at 0 for 0.

Exercise 5E.7. Prove that, under the conditions of Theorem 5E.1,

  
 1
inf lip (g; x̄)  F + g is not metrically regular at x̄ for ȳ + g(x̄) ≥ .
g:X→Y reg (F; x̄| ȳ)

At the end of this section we will derive from Theorem 5E.2 the following fixed
point theorem due to Nadler [1969]:

Theorem 5E.8 (Nadler). Let (X , ρ ) be a complete metric space and suppose that Φ
maps X into the set of closed subsets of X and is Lipschitz continuous in the sense
of Pompeiu-Hausdorff distance on X with Lipschitz constant λ ∈ (0, 1). Then Φ has
a fixed point.
Proof. We will first show that Φ has closed graph. Indeed, let (xk , yk ) ∈ gph Φ and
(xk , yk ) → (x, y). Then

d(y, Φ (x)) ≤ ρ (y, yk ) + d(yk , Φ (x))

≤ ρ (y, yk ) + h(Φ (xk ), Φ (x))
≤ ρ (y, yk ) + λ ρ (xk , x) → 0 as k → ∞.

Hence d(y, Φ (x)) = 0 and since Φ (x) is closed we have (x, y) ∈ gph Φ , and therefore
gph Φ is closed as claimed.
Let x̄ ∈ X and choose a > d(x̄, Φ (x̄))/(1 − λ ). Since gph Φ is closed, the set
gph Φ (x) ∩ (IBa (x̄) × IBa (x̄)) is closed as well. Furthermore, for every u, v ∈ IBa (x̄)
we obtain

e(Φ (u) ∩ IBa (x̄), Φ (v)) ≤ e(Φ (u), Φ (v)) ≤ h(Φ (u), Φ (v)) ≤ λ ρ (u, v).

Hence, by Theorem 5E.2 there exists x ∈ X such that x ∈ Φ (x).

292 5 Regularity in Infinite Dimensions

5F. Strong Metric Regularity and Implicit Function Theorems

Here we first present a strong regularity analogue of Theorem 5E.1 that provides
a sharper view of the interplay among the constants and neighborhoods of a map-
ping and its perturbation. As in the preceding section, we consider mappings act-
ing in metric spaces, to which the concept of strong regularity can be extended in
an obvious way. The following result significantly generalizes Theorem 3G.3 (and
moreover its previous version 2B.10).

Theorem 5F.1 (inverse function theorem with strong metric regularity in metric
spaces). Let (X, ρ ) be a complete metric space and let (Y, σ ) be a linear metric
space with shift-invariant metric σ . Consider a mapping F : X → → Y and any (x̄, ȳ) ∈
gph F such that, for a nonnegative constant κ and neighborhoods U of x̄ and V of ȳ,
the mapping y → F −1 (y) ∩U is a Lipschitz continuous function on V with Lipschitz
constant κ.
Then for every nonnegative constant μ with κ μ < 1 there exist neighborhoods U 
of x̄ and V  of ȳ such that, for every function g : X → Y which is Lipschitz continuous
on U with Lipschitz constant μ, the mapping y → (g + F)−1 (y) ∩U  is a Lipschitz
continuous function on g(x̄) + V  with Lipschitz constant κ /(1 − κ μ ).
Proof. We apply the standard (single-valued) version of the contracting mapping
principle, 1A.2, as in Proof II of 5E.1 but with some adjustments in the argument.
By assumption, for the function s(y) = F −1 (y) ∩U for y ∈ V we have

(1) ρ (s(y ), s(y)) ≤ κ σ (y , y) for all y , y ∈ V.

Pick μ > 0 such that κ μ < 1 and then choose positive constants a and b such that

(2) IBa (x̄) ⊂ U, IBb+μ a (ȳ) ⊂ V and κ b ≤ a(1 − κ μ ).

Choose any function g : X → Y such that

(3) σ (g(x ), g(x)) ≤ μ ρ (x , x) for all x , x ∈ U.

For any y ∈ IBb (g(x̄) + ȳ) and any x ∈ IBa (x̄) we have

σ (−g(x) + y, ȳ) = σ (y, g(x) + ȳ) ≤ σ (y, g(x̄) + ȳ) + σ (g(x), g(x̄)) ≤ b + μ a,

and hence, by (2), −g(x) + y ∈ V ⊂ dom s. Fix y ∈ IBb (g(x̄) + ȳ) and consider the
mapping
Φy : x → s(−g(x) + y) for x ∈ IBa (x̄).
Then, by using (1), (2) and (3) we get

ρ (x̄, Φy (x̄)) = ρ (s(ȳ), s(y − g(x̄))) ≤ κ σ (y, ȳ + g(x̄)) ≤ κ b ≤ a(1 − κ μ ).

Moreover, for any u, v ∈ IBa (x̄),

5 Regularity in Infinite Dimensions 293

ρ (Φy (u), Φy (v)) = ρ (s(y − g(u)), s(y − g(v))) ≤ κ σ (g(u), g(v)) ≤ κ μ ρ (u, v).

Thus, by the contraction mapping principle 1A.2, there exists a fixed point x = Φy (x)
in IBa (x̄), and there is no more than one such fixed point in IBa (x̄). The mapping from
y ∈ IBb (g(x̄) + ȳ) to the unique fixed point x(y) of Φy in IBa (x̄) is a function which
satisfies x(y) = s(y − g(x(y))); therefore, for any y, y ∈ IBb (g(x̄) + ȳ) we have

ρ (x(y), x(y )) = ρ (s(y − g(x(y))), s(y − g(x(y ))))

≤ κ (σ (y, y ) + σ (g(x(y)), g(x(y ))))
≤ κ σ (y, y ) + κ μ ρ (x(y), x(y )).

Hence,
κ
ρ (x(y), x(y )) ≤ σ (y, y ).
1 − κμ
Choosing U  = IBa (x̄) and V  = IBb (ȳ), and noting that IBb (g(x̄) + ȳ) = g(x̄) + IBb (ȳ),
we end the proof.
Compared with 3G.3, the above theorem exposes the fact that not only the Lip-
schitz constant, but also the neighborhoods associated with the Lipschitz localiza-
tion of (g + F)−1 depend on the mapping F only, and not on the perturbation g, as
long as its Lipschitz modulus is less than the reciprocal to the regularity modulus of
F. We already stated such a result for metric regularity in 5E.5.

Exercise 5F.2. Derive 5F.1 from a reformulation of 5E.5 in metric spaces.

Exercise 5F.3. In the framework of 5F.1, let F : X → → Y be strongly metrically regu-

lar at x̄ for ȳ and and let κ > reg (F; x̄| ȳ). For a metric space (P, π ) consider a func-
tion r : P × X → Y having ( p̄, x̄) ∈ int dom r and such that, for some μ ∈ (0, 1/κ ),

 (r; ( p̄, x̄)) < μ .

r(·, x̄) is continuous at p̄ and lip x

Prove that for each γ ≥ κ /(1 − κ μ ) there are neighborhoods U of x̄, V of ȳ and Q
of p̄ such that, for every p ∈ Q, the mapping y → (r(p, ·)+F)−1 (y)∩U is a Lipschitz
continuous function on r( p̄, x̄) + V with Lipschitz constant γ.
Guide. Choose neighborhoods U  of x̄ and Q of p̄ such that for each p ∈ Q the
function r(p, ·) is Lipschitz continuous on U  with Lipschitz constant μ . Applying
Theorem 5F.1, we obtain a constant γ and neighborhoods U of x̄, and V  of ȳ such
that for every p ∈ Q the mapping y → (r(p, ·) + F)−1 (y) ∩U is a Lipschitz contin-
uous function on r(p, x̄) + V  with Lipschitz constant γ . Since V  is independent of
p ∈ Q, by making Q small enough we can find a neighborhood V of ȳ such that
r( p̄, x̄) + V ⊂ r(p, x̄) + V  for every p ∈ Q.
We present next a strong regularity extension of Theorem 5E.3 such as has al-
ready appeared in various forms in the preceding chapters. We proved a weaker
version of this result in 2B.5 via Lemma 2B.6 and stated it again in Theorem 3G.4,
294 5 Regularity in Infinite Dimensions

which we left unproved. Here we treat a general case which can be deduced from
Theorem 5E.3 by taking into account that the strong metric regularity of h + F auto-
matically implies local closedness of its graph, and by adjoining to that the argument
in the proof of 3G.2. Since the result is central in this book, with the risk of repeating
ourselves, we supply it with an unabbreviated proof.

Theorem 5F.4 (implicit function theorem with strong metric regularity in metric
spaces). Let (X, ρ ) be a complete metric space and let (Y, σ ) be a linear metric
space with shift-invariant metric. Let (P, π ) be a metric space. For f : P × X → Y
and F : X →
→ Y , consider the generalized equation f (p, x) + F(x)  0 with solution
mapping   
S(p) = x  f (p, x) + F(x)  0 having x̄ ∈ S( p̄).
Let f (·, x̄) be continuous at p̄ and let h : X → Y be a strict estimator of f with
respect to x uniformly in p at ( p̄, x̄) with constant μ. Suppose that h + F is strongly
metrically regular at x̄ for 0 or, equivalently, the inverse (h + F)−1 has a Lipschitz
continuous single-valued localization ω around 0 for x̄ such that there exists κ ≥
reg (h + F; x̄|0) = lip (ω ; 0) with κ μ < 1.
Then the solution mapping S has a single-valued localization s around p̄ for x̄.
Moreover, for every ε > 0 there exists a neighborhood Q of p̄ such that
κ +ε
(4) ρ (s(p ), s(p)) ≤ σ ( f (p , s(p)), f (p, s(p))) for all p , p ∈ Q.
1 − κμ
In particular, s is continuous at p̄. In addition, if

(5) clm p ( f ; (x̄, p̄)) < ∞,

then the solution mapping S has a single-valued graphical localization s around p̄

for x̄ which is calm at p̄ with
κ
(6) clm (s; p̄) ≤ clm p ( f ; ( p̄, x̄)).
1 − κμ
If (5) is replaced by the stronger condition

(7)  ( f ; ( p̄, x̄)) < ∞,

lip p

then the graphical localization s of S around p̄ for x̄ is Lipschitz continuous near p̄

with
κ 
(8) lip (s; p̄) ≤ lip p ( f ; ( p̄, x̄)).
1 − κμ

If h : X → Y is not only a strict estimator of f , but also a strict first-order approxi-

mation of f with respect to x uniformly in p at ( p̄, x̄), then, under (5), we have

clm (s; p̄) ≤ lip (ω ; 0) clm p ( f ; ( p̄, x̄)),

5 Regularity in Infinite Dimensions 295

and under (7),

 ( f ; ( p̄, x̄)).
lip (s; p̄) ≤ lip (ω ; 0) lip p

Proof. Let ε > 0 and choose λ > κ and ν > μ such that

λ κ +ε
(9) λν < 1 and ≤ .
1−λν 1 − κμ

Then there exists a positive scalar α such that for each y ∈ IBα (ȳ) the set
(h + F)−1 (y) ∩ IBα (x̄) is a singleton, equal to the value ω (y) of the single-valued
localization of (h + F)−1 , and this localization ω is Lipschitz continuous with Lip-
schitz constant λ on IBα (0). We adjust α and choose a positive τ to also have, for
e(p, x) = f (p, x) − h(x),

(10) σ (e(p, x ), e(p, x)) ≤ ν ρ (x , x) for all x , x ∈ IBα (x̄) and p ∈ IBτ ( p̄).

Choose a positive a ≤ α satisfying

a(1 − λ ν )
(11) νa + ≤α
λ
and then a positive r ≤ τ such that

a(1 − λ ν )
(12) σ ( f (p, x̄), f ( p̄, x̄)) ≤ for all p ∈ IBr ( p̄).
λ
Then for every x ∈ IBa (x̄) and p ∈ IBr ( p̄), from (10)–(12) we have

σ (e(p, x), 0) ≤ σ (e(p, x), e(p, x̄)) + σ (e(p, x̄), e( p̄, x̄))
≤ νρ (x, x̄) + σ ( f (p, x̄), f ( p̄, x̄)) ≤ ν a + a(1 − λ ν )/λ ≤ α .

Hence, for such x and p, e(p, x) ∈ dom ω .

Fix an arbitrary p ∈ IBr ( p̄) and consider the mapping

Φ p : x → ω (−e(p, x)) for x ∈ IBa (x̄).

Observe that for any x ∈ IBa (x̄) having x = Φ p (x) implies x ∈ S(p) ∩ IBa (x̄), and
conversely. Noting that x̄ = ω (0) and using (12) we obtain

ρ (x̄, Φ p (x̄)) = ρ (ω (0), ω (−e(p, x̄))) ≤ λ σ ( f ( p̄, x̄), f (p, x̄)) ≤ a(1 − λ ν ).

Further, for any u, v ∈ IBa (x̄), using (10) we see that

ρ (Φ p (u), Φ p (v)) = ρ (ω (−e(u, p)), ω (−e(v, p)))

≤ λ σ (e(p, u), e(p, v)) ≤ λ ν ρ (u, v).

Hence the contraction mapping principle 1A.2 applies, with the λ there taken to be
the λ ν here, and it follows that for each p ∈ IBr ( p̄) there exists exactly one s(p) in
IBa (x̄) such that s(p) ∈ S(p); thus
296 5 Regularity in Infinite Dimensions

(13) s(p) = ω (−e(p, s(p))).

The function p → s(p) is therefore a single-valued localization of S around p̄ for x̄.

Moreover, from (13), for each p , p ∈ IBr ( p̄) we have

ρ (s(p ), s(p)) = ρ (ω (−e(p , s(p ))), ω (−e(p, s(p))))

≤ ρ (ω (−e(p , s(p ))), ω (−e(p , s(p))))
+ρ (ω (−e(p, s(p))), ω (−e(p, s(p))))
≤ λ σ (−e(p , s(p )), −e(p , s(p)))
+λ σ (−e(p , s(p)), −e(p, s(p)))
≤ λ ν ρ (s(p ), s(p)) + λ σ ( f (p , s(p)), f (p, s(p))).

Hence,
λ
ρ (s(p ), s(p)) ≤ σ ( f (p , s(p)), f (p, s(p))).
1−λν
Taking into account (9), we obtain (4). In particular, for p = p̄, from the continuity
of f (·, x̄) at p̄ we get that s is continuous at p̄. Under (5), the estimate (6) directly
follows from (4) by passing to zero with ε , and the same for (8) under (7). If h
is a strict first-order approximation of f , then μ could be arbitrarily small, and by
passing to lip (ω ; 0) with κ and to 0 with μ we obtain from (6) and (8) the last two
estimates in the statement.
Utilizing strict differentiability and ample parameterization we come to the fol-
lowing infinite-dimensional implicit function theorem which parallels 5E.4.

Theorem 5F.5 (using strict differentiability and ample parameterization). Let X , Y

and P be Banach spaces. For f : P × X → Y and F : X → → Y , consider the generalized
equation f (p, x) + F(x)  0 with solution mapping S and a pair ( p̄, x̄) with x̄ ∈ S( p̄)
and suppose that f is strictly differentiable at ( p̄, x̄). If the mapping

h + F for h(x) = f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄)

is strongly metrically regular at x̄ for 0, then S has a Lipschitz continuous single-

valued localization s around p̄ for x̄ with

lip (s; p̄) ≤ reg (h + F; x̄|0) · D p f ( p̄, x̄).

Furthermore, when f satisfies the ample parameterization condition:

the mapping D p f ( p̄, x̄) is surjective,

then the converse implication holds as well: the mapping h + F is strongly met-
rically regular at x̄ for 0 provided that S has a Lipschitz continuous single-valued
localization around p̄ for x̄.
5 Regularity in Infinite Dimensions 297

5G. The Bartle–Graves Theorem and Extensions

To set the stage, we begin with a Banach space version of the implication (i) ⇒ (ii)
in the symmetric inverse function theorem 1D.9.

Theorem 5G.1 (inverse function theorem in infinite dimensions). Let X be a Ba-

nach space and consider a function f : X → X and a point x̄ ∈ int dom f at which
f is strictly (Fréchet) differentiable and the derivative mapping D f (x̄) is invertible.
Then the inverse mapping f −1 has a single-valued graphical localization s around
ȳ := f (x̄) for x̄ which is strictly differentiable at ȳ, and moreover

Ds(ȳ) = [D f (x̄)]−1 .

In Section 1F we considered what may happen (in finite dimensions) when the
derivative mapping is merely surjective; by adjusting the proof of Theorem 1F.6
one obtains that when the Jacobian ∇ f (x̄) has full rank, the inverse f −1 has a local
selection which is strictly differentiable at f (x̄). The claim can be easily extended
to Hilbert (and even more general) spaces:

Exercise 5G.2 (differentiable inverse selections). Let X and Y be Hilbert spaces

and let f : X → Y be a function which is strictly differentiable at x̄ and such that
the derivative A := D f (x̄) is surjective. Then the inverse f −1 has a local selection s
around ȳ := f (x̄) for x̄ which is strictly differentiable at ȳ with derivative Ds(ȳ) =
A∗ (AA∗ )−1 , where A∗ is the adjoint of A.
Guide. Use the argument in the proof of 1F.6 with adjustments to the Hilbert space
setting. Another way of proving this result is to consider the function
 
x + A∗ u
g : (x, u) → for (x, u) ∈ X × Y,
f (x)

which satisfies g(x̄, 0) = (x̄, ȳ) and whose Jacobian is

 
I A∗
J= .
A0

In the Hilbert space context, if A is surjective then the operator J is invertible.

Hence, by Theorem 5G.1, the mapping g−1 has a single-valued graphical localiza-
tion (ξ , η ) : (v, y) → (ξ (v, y), η (v, y)) around (x̄, ȳ) for (x̄, 0). In particular, for some
neighborhoods U of x̄ and V of ȳ, the function s(y) := ξ (x̄, y) satisfies y = f (s(y))
for y ∈ V . To obtain the formula for the strict derivative, find the inverse of J.
In the particular case when the function f in 5G.2 is linear, the mapping
A∗ (AA∗ )−1 is a continuous linear selection of A−1 . A famous result by R. G. Bar-
tle and L. M. Graves [1952] yields that, for arbitrary Banach spaces X and Y , the
surjectivity of a mapping A ∈ L (X ,Y ) implies the existence of a continuous local
298 5 Regularity in Infinite Dimensions

selection of A−1 ; this selection, however, may not be linear. The original Bartle–
Graves theorem is for nonlinear mappings and says the following:

Theorem 5G.3 (Bartle–Graves). Let X and Y be Banach spaces and let f : X → Y

be a function which is strictly differentiable at x̄ and such that the derivative D f (x̄)
is surjective. Then there is a neighborhood V of ȳ := f (x̄) along with a continuous
function s : V → X and a constant γ > 0 such that

(1) f (s(y)) = y and s(y) − x̄ ≤ γ y − ȳ for every y ∈ V.

In other words, the surjectivity of the strict derivative at x̄ implies that f −1 has a
local selection s which is continuous around f (x̄) and calm at f (x̄). It is known3 that,
in contrast to the strictly differentiable local selection in 5G.2 for Hilbert spaces, the
selection in the Bartle–Graves theorem, even for a bounded linear mapping f , might
be not even Lipschitz continuous around ȳ. For this case we have:

Corollary 5G.4 (inverse selection of a surjective linear mapping in Banach spaces).

For any bounded linear mapping A from X onto Y , there is a continuous (but gener-
ally nonlinear) mapping B such that ABy = y for every y ∈ Y .
Proof. Theorem 5G.3 tells us that A−1 has a continuous local selection at 0 for 0.
Since A−1 is positively homogeneous, this selection is global.
In this section we develop a generalization of the Bartle–Graves theorem for
metrically regular set-valued mappings. First, recall that a mapping F : Y → → X is
(sequentially) inner semicontinuous on a set T ⊂ Y if for every y ∈ T , every x ∈ F(y)
and every sequence of points yk ∈ T , yk → y, there exists xk ∈ F(yk ) for k = 1, 2, . . .
such that xk → x as k → ∞. We also need a basic result which we only state here
without proof:

Theorem 5G.5 (Michael’s selection theorem). Let X and Y be Banach spaces and
consider a mapping F : Y → → X which is closed-convex-valued and inner semicon-
tinuous on dom F = 0/ . Then F has a continuous selection s : dom F → X .
We require a lemma which connects the Aubin property of a mapping with the
inner semicontinuity of a truncation of this mapping:

Lemma 5G.6 (inner semicontinuous selection from the Aubin property). Consider
a mapping S : Y → → X and any (ȳ, x̄) ∈ gph S, and suppose that S has the Aubin
property at ȳ for x̄ with constant κ. Suppose, for some c > 0, that the sets S(y) ∩
IBc (x̄) are convex and closed for all y ∈ IBc (ȳ). Then for any α > κ there exists
β > 0 such that the mapping

y → Mα (y) = S(y) ∩ IBα y−ȳ (x̄) for y ∈ IBβ (ȳ),
0/ otherwise

3 Cf. Deville, Godefroy and Zizler [1993], p. 200.

5 Regularity in Infinite Dimensions 299

is nonempty-closed-convex-valued and inner semicontinuous on IBβ (ȳ).

Proof. Let a and b be positive numbers such that the balls IBa (x̄) and IBb (ȳ) are
associated with the Aubin property of S (metric regularity of S−1 ) with a constant
κ . Without loss of generality, let max{a, b} < c. Fix α > κ and choose β such that
a c 
0 < β ≤ min , , b, c .
α 3α
For such a β the mapping Mα has nonempty closed convex values. It remains to
show that Mα is inner semicontinuous on IBβ (ȳ).
Let (y, x) ∈ gph Mα and yk → y, yk ∈ IBβ (ȳ). First, let y = ȳ. Then Mα (y) = x̄, and
from the Aubin property of S there exists a sequence of points xk ∈ S(yk ) such that
xk − x̄ ≤ κ yk − ȳ. Then xk ∈ Mα (yk ), xk → x as k → ∞ and we are done in this
case.
Now let y = ȳ. The Aubin property of S yields that there exists x̌k ∈ S(yk ) such
that
x̌k − x̄ ≤ κ yk − ȳ
and also there exists x̃k ∈ S(yk ) such that

x̃k − x ≤ κ yk − y.

Then, the choice of β above yields

c
x̌k − x̄ ≤ κβ ≤ α ≤c
3α
and

x̃k − x̄ ≤ x̃k − x + x − x̄

≤ κ yk − y + α y − ȳ
≤ 2κβ + αβ ≤ 3αβ ≤ c.

Let
(α + κ )yk − y
(2) εk = .
(α − κ )yk − ȳ + (α + κ )yk − y

Then 0 ≤ ε k < 1 and ε k  0 as k → ∞. Let xk = ε k x̌k + (1 − ε k )x̃k . Then xk ∈ S(yk ).

Moreover, we have

xk − x̄ ≤ ε k x̌k − x̄ + (1 − ε k)x̃k − x̄

≤ ε k κ yk − ȳ + (1 − ε k)(x̃k − x + x − x̄)
≤ ε k κ yk − ȳ + (1 − ε k)κ yk − y + (1 − ε k)α y − ȳ
≤ ε k κ yk − ȳ + (1 − ε k)κ yk − y
+(1 − ε k )α yk − ȳ + (1 − ε k)α yk − y
300 5 Regularity in Infinite Dimensions

≤ α yk − ȳ − ε k (α − κ )yk − ȳ

+(1 − ε k )(α + κ )yk − y ≤ α yk − ȳ,

where in the last inequality we take into account the expression (2) for ε k . Thus
xk ∈ Mα (yk ), and since xk → x, we are done.
Lemma 5G.6 allows us to apply Michael’s selection theorem to the mapping Mα ,
obtaining the following result:

Theorem 5G.7 (continuous inverse selection from metric regularity). Consider a

mapping F : X → → Y which is metrically regular at x̄ for ȳ. Let, for some c > 0,
the sets F −1 (y) ∩ IBc (x̄) be convex and closed for all y ∈ IBc (ȳ). Then for every
α > reg (F; x̄| ȳ) the mapping F −1 has a continuous local selection s around ȳ for x̄
which is calm at ȳ with

(3) clm (s; ȳ) ≤ α .

Proof. Choose α such that α > reg (F; x̄| ȳ), and apply Michael’s theorem 5G.5 to
the mapping Mα in 5G.6 for S = F −1 . By the definition of Mα , the continuous local
selection obtained in this way is calm with a constant α .
Note that the continuous local selection s in 5G.7 depends on α and therefore we
cannot replace α in (3) with reg (F; x̄| ȳ).
In the remainder of this section we show that if a mapping F satisfies the as-
sumptions of Theorem 5G.7, then for any function g : X → Y with lip (g; x̄) <
1/ reg(F; x̄| ȳ), the mapping (g + F)−1 has a continuous and calm local selection
around g(x̄) + ȳ for x̄. We will prove this generalization of the Bartle–Graves theo-
rem by repeatedly using an argument similar to the proof of Lemma 5G.6, the idea
of which goes back to (modified) Newton’s method used to prove the theorems of
Lyusternik and Graves and, in fact, to Goursat’s proof on his version of the classical
inverse function theorem. We put the theorem in the format of the general implicit
function theorem paradigm:

Theorem 5G.8 (inverse mapping theorem with continuous and calm local selec-
tions). Consider a mapping F : X → → Y and any (x̄, ȳ) ∈ gph F and suppose that for
some c > 0 the mapping IBc (ȳ)  y → F −1 (y) ∩ IBc (x̄) is closed-convex-valued. Let
κ and μ be nonnegative constants such that

reg (F; x̄| ȳ) ≤ κ and κ μ < 1.

Then for any function g : X → Y with x̄ ∈ int dom g and lip (g; x̄) ≤ μ and for every
γ with
κ
<γ
1 − κμ
the mapping (g + F)−1 has a continuous local selection s around g(x̄) + ȳ for x̄,
which moreover is calm at g(x̄) + ȳ with
5 Regularity in Infinite Dimensions 301

(4) clm (s; g(x̄) + ȳ) ≤ γ .

Proof. The proof consists of two steps. In the first step, we use induction to obtain
a Cauchy sequence of continuous functions z0 , z1 , . . ., such that zn is a continuous
and calm selection of the mapping y → F −1 (y − g(zn−1 (y))). Then we show that this
sequence has a limit in the space of continuous functions acting from a fixed ball
around ȳ to the space X and equipped with the supremum norm, and this limit is the
selection whose existence is claimed.
Choose κ and μ as in the statement of the theorem and let γ > κ /(1 − κ μ ). Let
λ , α and ν be such that κ < λ < α < 1/ν and ν > μ , and also λ /(1 − αν ) ≤ γ .
Without loss of generality, we can assume that g(x̄) = 0. Let IBa (x̄) and IBb (ȳ) be
the neighborhoods of x̄ and ȳ, respectively, that are associated with the assumed
properties of the mapping F and the function g. Specifically,
(a) For every y, y ∈ IBb (ȳ) and x ∈ F −1 (y) ∩ IBa (x̄) there exists x ∈ F −1 (y ) with

x − x ≤ λ y − y.

(b) For every y ∈ IBb (ȳ) the set F −1 (y) ∩ IBa (x̄) is nonempty, closed and convex
(that is, max{a, b} ≤ c).
(c) The function g is Lipschitz continuous on IBa (x̄) with a constant ν .
According to 5G.7, there we can find a constant β , 0 < β ≤ b, and a continuous
function z0 : IBβ (ȳ) → X such that

F(z0 (y))  y and z0 (y) − x̄ ≤ λ y − ȳ for all y ∈ IBβ (ȳ).

Choose a positive τ such that

 
a β
(5) τ ≤ (1 − αν ) min a, ,
2λ 2

and consider the mapping y → M1 (y) where

  
M1 (y) = x ∈ F −1 (y − g(z0 (y)))  x − z0 (y) ≤ αν z0 (y) − x̄

for y ∈ IBτ (ȳ) and M1 (y) = 0/ for y ∈

/ IBτ (ȳ). Clearly, (ȳ, x̄) ∈ gph M1 . Also, for any
y ∈ IBτ (ȳ), we have from the choice of λ and α , using (5), that z0 (y) ∈ IBa (x̄) and
therefore

y−g(z0 (y))− ȳ ≤ τ + ν z0 (y)− x̄ ≤ τ + νλ τ ≤ (1− αν )(1+ νλ )(β /2) ≤ β ≤ b.

Then from the Aubin property of F −1 there exists x ∈ F −1 (y − g(z0 (y))) with

x − z0 (y) ≤ λ g(z0 (y)) − g(x̄) ≤ αν z0 (y) − x̄,

which implies x ∈ M1 (y). Thus M1 is nonempty-valued. Further, if (y, x) ∈ gph M1 ,

using (5) we have that y ∈ IBb (ȳ) and also
302 5 Regularity in Infinite Dimensions
a a
x − x̄ ≤ x − z0 (y) + z0 (y) − x̄ ≤ (1 + αν )λ τ ≤ (1 − (αν )2 )λ ≤ .
2λ 2
Then, from the property (b) above, since for any y ∈ dom M the set M1 (y) is the
intersection of a closed ball with a closed convex set, the mapping M1 is closed-
convex-valued on its domain. We will show that this mapping is inner semicontinu-
ous on IBτ (ȳ).
Let y ∈ IBτ (ȳ) and x ∈ M1 (y), and let yk ∈ IBτ (ȳ), yk → y as k → ∞. If z0 (y) = x̄,
then M1 (y) = {x̄} and therefore x = x̄. Any xk ∈ M1 (yk ) satisfies

xk − z0 (yk ) ≤ αν z0 (yk ) − x̄.

Using the continuity of the function z0 , we see that xk → z0 (y) = x̄ = x; thus M1 is

inner semicontinuous.
Now let z0 (y) = x̄. Since z0 (yk ) ∈ F −1 (yk − g(x̄)) ∩ IBa (x̄), the Aubin property of
F furnishes the existence of x̌k ∈ F −1 (yk − g(z0 (yk ))) such that
−1

(6) x̌k − z0 (yk ) ≤ λ g(z0 (yk )) − g(x̄) ≤ λ ν z0 (yk ) − x̄ ≤ αν z0 (yk ) − x̄.

Then x̌k ∈ M1 (yk ), and in particular, x̌k ∈ IBa (x̄). Further, the inclusion x ∈ F −1 (y −
g(z0 (y))) ∩ IBa (x̄) combined with the Aubin property of F −1 entails the existence of
x̃k ∈ F −1 (yk − g(z0 (yk ))) such that

(7) x̃k − x ≤ λ (yk − y + ν z0(yk ) − z0 (y)) → 0 as k → ∞.

Then x̃k ∈ IBa (x̄) for large k. Let

(1 + αν )z0 (yk ) − z0 (y) + x̃k − x

ε k := .
αν z0 (y) − x̄ − λ ν z0(yk ) − x̄

Note that, for k → ∞, the numerator in the definition of ε k goes to 0 because of the
continuity of z0 and (7), while the denominator converges to (α − λ )ν z0 (y) − x̄
> 0; therefore ε k → 0 as k → ∞. Let

xk = ε k x̌k + (1 − ε k )x̃k .

Since x̃k → x and ε k → 0, we get xk → x as k → ∞ and also, since y → F −1 (y)∩IBa (x̄)

is convex-valued around (x̄, ȳ), we have xk ∈ F −1 (yk − g(z0 (yk ))) for large k. By (6),
(7), the assumption that x ∈ M1 (y), and the choice of ε k , we have

xk − z0 (yk ) ≤ ε k x̌k − z0 (yk ) + (1 − ε k)x̃k − z0 (yk )

≤ ε k λ ν z0 (yk ) − x̄ + (1 − ε k)(x̃k − x
+x − z0 (y) + z0 (y) − z0 (yk ))
≤ ε k λ ν z0 (yk ) − x̄ + x̃k − x
+(1 − ε k )αν z0 (y) − x̄ + z0(y) − z0 (yk )
≤ αν z0 (yk ) − x̄ + αν z0 (yk ) − z0 (y)
5 Regularity in Infinite Dimensions 303

+x̃k − x + z0(y) − z0 (yk )

−ε k αν z0 (y) − x̄ + ε k λ ν z0 (yk ) − x̄
≤ αν z0 (yk ) − x̄ + x̃k − x + (1 + αν )z0(y) − z0 (yk )
−ε k (αν z0 (y) − x̄ − λ ν z0(yk ) − x̄)
= αν z0 (yk ) − x̄.

We obtain that xk ∈ M1 (yk ), and since xk → x, we conclude that the mapping M1

is inner semicontinuous on its domain IBτ (ȳ). Hence, by Michael’s selection theo-
rem 5G.5, there it has a continuous selection z1 : IBτ (ȳ) → X; that is, a continuous
function z1 which satisfies

z1 (y) ∈ F −1 (y − g(z0 (y))) and z1 (y) − z0 (y) ≤ αν z0 (y) − x̄ for all y ∈ IBτ (ȳ).

Then for y ∈ IBτ (ȳ), by the choice of γ ,

z1 (y) − x̄ ≤ z1 (y) − z0 (y) + z0 (y) − x̄ ≤ (1 + αν )λ y − ȳ ≤ γ y − ȳ.

The induction step is parallel to the first step. Let z0 and z1 be as above and
suppose we have also found functions z1 , z2 , . . . , zn , such that each z j , j = 1, 2, . . . , n,
is a continuous selection of the mapping y → M j (y) where
  
M j (y) = x ∈ F −1 (y − g(z j−1(y)))  x − zn (y) ≤ αν z j−1 (y) − z j−2(y)

/ IBτ (ȳ), where we put z−1 (y) = x̄ for y ∈ IBτ (ȳ).

for y ∈ IBτ (ȳ) and M j (y) = 0/ for y ∈
Then for y ∈ IBτ (ȳ) we obtain

z j (y) − z j−1 (y) ≤ (αν ) j−1 z1 (y) − z0 (y) ≤ (αν ) j z0 (y) − x̄, j = 2, . . . , n.

Therefore,
j
z j (y) − x̄ ≤ ∑ (αν )i zi (y) − zi−1(y)
i=0
j
λ
≤ ∑ (αν )i z0(y) − x̄ ≤ 1 − αν y − ȳ ≤ γ y − ȳ.
i=0

Hence, from (5), for j = 1, 2, . . . , n,

(8) z j (y) − x̄ ≤ a

and also
λ ντ τ
(9) y − g(z j (y)) − ȳ ≤ τ + ν z j (y) − x̄ ≤ τ + ≤ ≤ β ≤ b.
1 − αν 1 − αν
Consider the mapping y → Mn+1 (y) where
304 5 Regularity in Infinite Dimensions
  
Mn+1 (y) = x ∈ F −1 (y − g(zn (y)))  x − zn (y) ≤ αν zn (y) − zn−1(y)

for y ∈ IBτ (ȳ) and Mn+1 (y) = 0/ for y ∈

/ IBτ (ȳ). As in the first step, we find that Mn+1 is
nonempty-closed-convex-valued. Let y ∈ IBτ (ȳ) and x ∈ Mn+1 (y), and let yk ∈ IBτ (ȳ),
yk → y as k → ∞. If zn−1 (y) = zn (y), then Mn+1 (y) = {zn (y)}, and consequently
x = zn (y); then from zn (yk ) ∈ F −1 (yk − g(zn−1 (yk ))) ∩ IBa (x̄) and yk − g(zn−1 (yk )) ∈
IBb (ȳ), we obtain, using the Aubin property of F −1 , that there exists xk ∈ F −1 (yk −
g(zn (yk ))) such that

xk − zn (yk ) ≤ λ g(zn (yk )) − g(zn−1(yk )) ≤ αν zn (yk ) − zn−1(yk ).

Therefore xk ∈ Mn+1 (yk ), xk → z1 (y) = x as k → ∞, and hence Mn+1 is inner semi-

continuous for the case considered.
Let zn (y) = zn−1 (y). From (8) and (9) for y = yk , since

zn (yk ) ∈ F −1 (yk − g(zn−1 (yk ))) ∩ IBa (x̄),

the Aubin property of F −1 implies the existence of x̌k ∈ F −1 (yk − g(zn (yk ))) such
that

x̌k − zn (yk ) ≤ λ g(zn (yk )) − g(zn−1 (yk )) ≤ λ ν zn (yk ) − zn−1 (yk ).

Similarly, since x ∈ F −1 (y− g(zn (y)))∩IBa (x̄), there exists x̃k ∈ F −1 (yk − g(zn (yk )))
such that

x̃k − x ≤ λ (yk − y + g(zn(yk )) − g(zn (y)))

≤ λ (yk − y + ν zn(yk ) − zn (y)) → 0 as k → ∞.

Put
αν zn−1 (y) − zn−1(yk ) + (1 + αν )zn(y) − zn (yk ) + x̃k − x
ε k := .
αν zn (y) − zn−1(y) − λ ν zn (yk ) − zn−1 (yk )

Then ε k → 0 as k → ∞. Taking

xk = ε k x̌k + (1 − ε k )x̃k ,

we obtain that xk ∈ F −1 (yk − g(zn (yk ))) for large k. Further, we estimate xk −
zn (yk ) in the same way as in the first step, that is,

xk − zn (yk ) ≤ ε k x̌k − zn (yk ) + (1 − ε k)x̃k − zn (yk )

≤ ε k λ ν zn (yk ) − zn−1(yk )
+(1 − ε k )(x̃k − x + x − zn(y) + zn(y) − zn (yk ))
≤ ε k λ ν zn (yk ) − zn−1(yk ) + x̃k − x
+(1 − ε k )αν zn (y) − zn−1 (y) + zn (y) − zn (yk )
≤ αν zn (yk ) − zn−1 (yk ) + αν zn (yk ) − zn (y)
5 Regularity in Infinite Dimensions 305

+αν zn−1 (yk ) − zn−1 (y) + x̃k − x

+zn (y) − zn (yk ) − ε k αν zn (y) − zn−1 (y)
+ε k λ ν zn (yk ) − zn−1 (yk )
≤ αν zn (yk ) − zn−1 (yk ) + x̃k − x
+(1 + αν )zn (y) − zn (yk ) + αν zn−1 (y) − zn−1 (yk )
−ε k (αν zn (y) − zn−1 (y) − λ ν zn (yk ) − zn−1 (yk ))
= αν zn (yk ) − zn−1 (yk ).

We conclude that xk ∈ Mn+1 (yk ), and since xk → x as k → ∞, the mapping Mn+1 is

inner semicontinuous on IBτ (ȳ). Hence, the mapping Mn+1 has a continuous selec-
tion zn+1 : IBτ (ȳ) → X, that is,

zn+1 (y) ∈ F −1 (y − g(zn (y))) and zn+1 (y) − zn (y) ≤ αν zn (y) − zn−1 (y).

Thus
zn+1 (y) − zn (y) ≤ (αν )(n+1) z0 (y) − x̄.
The induction step is now complete. In consequence, we have an infinite se-
quence of bounded continuous functions z0 , . . . , zn , . . . such that for all y ∈ IBτ (ȳ)
and for all n,
n
λ
zn (y) − x̄ ≤ ∑ (αν )i z0 (y) − x̄ ≤ y − ȳ ≤ γ y − ȳ
i=0 1 − αν

and moreover,

sup zn+1 (y) − zn (y) ≤ (αν )n sup z0 (y) − x̄ ≤ (αν )n λ τ for n ≥ 1.
y∈IBτ (ȳ) y∈IBτ (ȳ)

The sequence {zn } is a Cauchy sequence in the space of functions that are continu-
ous and bounded on IBτ (ȳ) equipped with the supremum norm. Then this sequence
has a limit s which is a continuous function on IBτ (ȳ) and satisfies

s(y) ∈ F −1 (y − g(s(y)))

and
λ
s(y) − x̄ ≤ y − ȳ ≤ γ y − ȳ for all y ∈ IBτ (ȳ).
1 − αν
Thus, s is a continuous local selection of (g + F)−1 which has the calmness property
(4). This brings the proof to its end.
Proof of Theorem 5G.3. Apply 5G.8 with F = D f (x̄) and g(x) = f (x) − D f (x̄)x.
Metric regularity of F is equivalent to surjectivity of D f (x̄), and F −1 is convex-
closed-valued. The mapping g has lip (g; x̄) = 0 and finally F + g = f .

Note that Theorem 5G.7 follows from 5G.8 with g the zero function.
306 5 Regularity in Infinite Dimensions

We present next an implicit mapping version of Theorem 5G.7.

Theorem 5G.9 (implicit mapping version). Let X,Y and P be Banach spaces. For
f : P × X → Y and F : X →
→ Y , consider the generalized equation f (p, x) + F(x)  0
with solution mapping
  
S(p) = x  f (p, x) + F(x)  0 having x̄ ∈ S( p̄).

Suppose that F satisfies the conditions in Theorem 5G.7 with ȳ = 0 and associate
constant κ ≥ reg (F; x̄|0) and also that f is continuous on a neighborhood of (x̄, p̄)
 ( f ; ( p̄, x̄)) ≤ μ , where μ is a nonnegative constant satisfying κ μ < 1.
and has lip x
Then for every γ satisfying
κ
(10) <γ
1 − κμ
there exist neighborhoods U of x̄ and Q of p̄ along with a continuous function s :
Q → U such that

(11) s(p) ∈ S(p) and s(p) − x̄ ≤ γ  f (p, x̄) − f ( p̄, x̄) for every p ∈ Q.

Proof. The proof is parallel to the proof of Theorem 5G.7. First we choose γ satis-
fying (10) and then λ , α and ν such that κ < λ < α < ν −1 and ν > μ , and also

λ
(12) < γ.
1 − αν
There are neighborhoodsU, V and Q of x̄, 0 and p̄, respectively, which are associated
with the metric regularity of F at x̄ for 0 with constant λ and the Lipschitz continuity
of f with respect to x with constant ν uniformly in p. By appropriately choosing a
sufficiently small radius τ of a ball around p̄, we construct an infinite sequence of
continuous and bounded functions zk : IBτ ( p̄) → X, k = 0, 1, . . ., which are uniformly
convergent on IBτ ( p̄) to a function s satisfying the conditions in (11). The initial z0
satisfies

z0 (p) ∈ F −1 (− f (p, x̄)) and z0 (p) − x̄ ≤ λ  f (p, x̄) − f ( p̄, x̄).

For k = 1, 2, . . ., the function zk is a continuous selection of the mapping

  
Mk : p → x ∈ F −1 (− f (p, zk−1 (p)))  x − zk−1 (p) ≤ αν zk−1 (p) − zk−2 (p)

for p ∈ IBτ ( p̄), where z−1 (p) = x̄. Then for all p ∈ IBτ ( p̄) we obtain

zk (p) ∈ F −1 (− f (p, zk−1 (p))) and zk (p) − zk−1 (p) ≤ (αν )k z0 (p) − x̄,

hence,
λ
(13) zk (y) − x̄ ≤  f (p, x̄) − f ( p̄, x̄).
1 − αν
5 Regularity in Infinite Dimensions 307

The sequence {zk } is a Cauchy sequence of continuous and bounded function, hence
it is convergent with respect to the supremum norm. In the limit with k → ∞, taking
into account (12) and (13), we obtain a selection s with the desired properties.

Exercise 5G.10 (specialization for closed sublinear mappings). Let F : X → → Y have

convex and closed graph, let f : X → Y be strictly differentiable at x̄ and let (x̄, ȳ) ∈
gph ( f + F). Suppose that

(14) ȳ ∈ int rge( f (x̄) + D f (x̄)(x − x̄) + F).

Prove that there exist neighborhoods U of x̄ and V of ȳ, a continuous function s :

V → U , and a constant γ, such that

( f + F)(s(y))  y and s(y) − x̄ ≤ γ y − ȳ for every y ∈ V.

Guide. Apply the Robinson-Ursescu theorem 5B.4.

308 5 Regularity in Infinite Dimensions

Commentary
The equivalence of (a) and (d) in 5A.1 was shown in Theorem 10 on p. 150 of the
original treatise of S. Banach [1932]. The statements of this theorem usually include
the equivalence of (a) and (b), which is called in Dunford and Schwartz [1958] the
“interior mapping principle.” Lemma 5A.4 is usually stated for Banach algebras,
see, e.g., Theorem 10.7 in Rudin [1991]. Theorem 5A.8 is from Robinson [1972].
The generalization of the Banach open mapping theorem to set-valued map-
pings with convex closed graphs was obtained independently by Robinson [1976]
and Ursescu [1975]; the proof of 5B.3 given here is close to the original proof in
Robinson [1976]. A particular case of this result for positively homogeneous map-
pings was shown earlier by Ng [1973]. The Baire category theorem can be found
in Dunford and Schwartz [1958], p. 20. The Robinson–Ursescu theorem is stated
in various ways in the literature, see, e.g., Theorem 3.3.1 in Aubin and Ekeland
[1984], Theorem 2.2.2 in Aubin and Frankowska [1990], Theorem 2.83 in Bonnans
and Shapiro [2000], Theorem 9.48 in Rockafellar and Wets [1998], Theorem 1.3.11
in Zălinescu [2002] and Theorem 4.21 in Mordukhovich [2006].
Sublinear mappings (under the name “convex processes”) and their adjoints were
introduced by Rockafellar [1967]; see also Rockafellar [1970]. Theorem 5C.9 first
appeared in Lewis [1999], see also Lewis [2001]. The norm duality theorem, 5C.10,
was originally proved by Borwein [1983], who later gave in Borwein [1986b] a
more detailed argument. The statement of the Hahn–Banach theorem 5C.11 is from
Dunford and Schwartz [1958], p. 62.
Theorems 5D.1 and 5D.2 are versions of results originally published in Lyusternik
[1934] and Graves [1950], with some adjustments to the current setting. Lyusternik
apparently viewed his theorem mainly as a stepping stone to obtain the Lagrange
multiplier rule for abstract minimization problems, and the title of his paper from
1934 clearly says so. It is also interesting to note that, after the statement of the
Lyusternik theorem as 8.10.2 in the functional analysis book by Lyusternik and
Sobolev [1965], the authors say that “the proof of this theorem is a modification
of the proof of the implicit function theorem, and the [Lyusternik] theorem is a
direct generalization of this [implicit function] theorem.”
It is quite likely that Graves considered his theorem as an extension of the Banach
open mapping theorem for nonlinear mappings. But there is more in its statement
and proof; namely, the Graves theorem does not involve differentiation and then,
as shown in 5D.3, can be easily extended to become a generalization of the the
basic Lemma 5A.4 for nonlinear mappings. This was mentioned already in the his-
torical remarks of Dunford and Schwartz [1958], p. 85. A further generalization in
line with the present setting was revealed in Dmitruk, Milyutin and Osmolovskiı̆
[1980], where the approximating linear mapping is replaced by a Lipschitz continu-
ous function with a sufficiently small Lipschitz constant. Estimates for the regularity
modulus of the kind given in 5D.3 are also present in Ioffe [1979].
In the second part of the last century, when the development of optimality condi-
tions was a key issue, the approach of Lyusternik was recognized for its virtues and
5 Regularity in Infinite Dimensions 309

extended to great generality. Historical remarks regarding these developments can

be found in Ioffe [2000] and Rockafellar and Wets [1998]. The statement in 5D.4
is a slightly modified version of the Lyusternik theorem as given in Section 0.2.4 of
Ioffe and Tikhomirov [1974].
Theorem 5E.1 is given as in Dontchev, Lewis and Rockafellar [2003]; earlier
results in this vein were obtained by Dontchev and Hager [1993,1994]. The con-
traction mapping theorem 5E.2 is from Dontchev and Hager [1994]. A related result
in abstract metric spaces is given in Arutyunov [2007].
Theorem 5G.3 gives the original form of the Bartle–Graves theorem as con-
tributed in Bartle and Graves [1952]. The particular form 5G.5 of Michael’s se-
lection theorem4 is Lemma 2.1 in Deimling [1992]. Lemma 5G.6 was first given
in Borwein and Dontchev [2003], while Theorem 5G.8 is from Dontchev [2004].
These two papers were largely inspired by contacts of the first author of this book
with Robert G. Bartle, who was able to read them before he passed away Sept. 18,
2002. Shortly before he died he sent to the first author a letter, where he, among
other things, wrote the following:
“Your results are, indeed, an impressive and far-reaching extension of the the-
orem that Professor Graves and I published over a half-century ago. I was a stu-
dent in a class of Graves in which he presented the theorem in the case that the
parameter domain is the interval [0, 1]. He expressed the hope that it could be gen-
eralized to a more general domain, but said that he didn’t see how to do so. By a
stroke of luck, I had attended a seminar a few months before given by André Weil,
which he titled “On a theorem by Stone.” I (mis)understood that he was referring to
M. H. Stone, rather than A. H. Stone, and attended. Fortunately, I listened carefully
enough to learn about paracompactness and continuous partition of unity5 (which
were totally new to me) and which I found to be useful in extending Graves’ proof.
So the original theorem was entirely due to Graves; I only provided an extension
of his proof, using methods that were not known to him. However, despite the fact
that I am merely a ‘middleman,’ I am pleased that this result has been found to be
useful.”
In this book we present inverse/implicit function theorems centered around va-
riational analysis, but there are many other theorems that fall into the same category
and are designed as tools in other areas of mathematics beyond variational analysis.
In particular, we do not discuss in this book the celebrated Nash–Moser theorem,
used mainly in geometric analysis and partial differential equations. A rigorous in-
troduction of this theorem and the theory around it would require a lot of space
and would tip the balance of topics and ideas away from we want to emphasize.
More importantly, we have not been able to identify (as of yet) a specific, sound
application of this theorem in variational analysis such as would have justified the
inclusion. A rigorous and nice presentation of the Nash–Moser theorem along with

4 The original statement of Michael’s selection theorem is for mappings acting from a paracompact

space to a Banach space; by a theorem of A. H. Stone every metric space is paracompact and hence
every subset of a Banach space is paracompact.
5 Michael’s theorem was not known at that time.
310 5 Regularity in Infinite Dimensions

the theory and applications behind it is given in Hamilton [1982]. In the following
lines, we only briefly point out a connection to the results in Section 5E.
The Nash–Moser theorem is about mappings acting in Fréchet spaces, which are
more general than the Banach spaces. Consider a linear (vector) space F equipped
with the collection of seminorms { · n |n ∈ IN} (a seminorm differs from a norm
in that the seminorm of a nonzero element could be zero). The topology induced
by this (countable) collection of seminorms makes the space F a locally convex
topological vector space. If x = 0 when xn = 0 for all n, the space is Hausdorff.
In a Hausdorff space, one may define a metric based the family of seminorms in the
following way:
∞
x − yn
(1) ρ (x, y) = ∑ 2−n 1 + x − yn .
n=1

It is not difficult to see that this metric is shift-invariant. A sequence {xk } is said to be
Cauchy when xk − x j n → 0 as k and j → ∞ for all n, or, equivalently, ρ (xk , x j ) → 0
as k → ∞ and j → ∞. As usual, a space is complete if every Cauchy sequence con-
verges. A Fréchet space is a complete Hausdorff metrizable locally convex topolog-
ical vector space.
Having two Fréchet spaces F and G, we can now introduce metrics ρ and σ asso-
ciated with their collections of seminorms as in (1) above, and define Lipschitz con-
tinuity and metric regularity accordingly. Then Theorem 5E.3 will apply of course
and we can obtain, e.g., a Graves-type theorem in Fréchet spaces, in terms of the
metrics ρ and σ , and also an implicit function theorem in Fréchet spaces from the
general Theorem 5E.4. To get to the Nash–Moser theorem, however, we have a long
way to go, translating the meaning of, e.g., the assumptions Theorem 5E.4, in terms
of the metrics ρ and σ for the collections of seminorms and the mappings consid-
ered. For that we will need more structure in the spaces, an ordering (grading) of the
sequence of seminorms and, moreover, a certain uniform approximation property
called the tameness condition. For the mappings, the associated tameness property
means that certain growth estimates hold. The statement of the Nash–Moser theorem
is surprisingly similar to the classical inverse function theorem, but the meaning of
the concepts used is much more involved: when a smooth tame mapping f acting be-
tween Fréchet spaces has an invertible tame derivative, then f −1 has a smooth tame
single-valued localization. The rigorous introduction of the tame spaces, mappings
and derivatives is beyond the scope of this book; we only note here that extending
the Nash–Moser theorem to set-valued mappings, e.g. in the setting of Section 5E,
is a challenging avenue for future research.
Chapter 6
Applications in Numerical Variational Analysis

The classical implicit function theorem finds a wide range of applications in numer-
ical analysis. For instance, it helps in deriving error estimates for approximations to
differential equations and is often relied on in establishing the convergence of algo-
rithms. Can the generalizations of the classical theory to which we have devoted so
much of this book have comparable applications in the numerical treatment of non-
classical problems for generalized equations and beyond? In this chapter we provide
positive answers in several directions.
We begin with a topic at the core of numerical work, the “conditioning” of a
problem and how it extends to concepts like metric regularity. We also explain how
the conditioning of a feasibility problem, like solving a system of inequalities, can
be understood. Next we take up a general iterative scheme for solving generalized
equations under metric regularity, obtaining convergence by means of our earlier
basic results. As particular cases, we get various modes of convergence of the age-
old procedure known as Newton’s method in several guises, and of the much more
recently introduced proximal point algorithm. We go a step further with Newton’s
method by showing that the mapping which assigns to an instance of a parameter the
set of all sequences generated by the method obeys, in a Banach space of sequences,
the implicit function theorem paradigm in the same pattern as the solution mapping
for the underlying generalized equation. Approximations of quadratic optimization
problems in Hilbert spaces are then studied. Finally, we apply our methodology to
discrete approximations in optimal control.

A.L. Dontchev and R.T. Rockafellar, Implicit Functions and Solution Mappings: A View 311
from Variational Analysis, Springer Monographs in Mathematics,
DOI 10.1007/978-0-387-87821-8 6,  c Springer Science+Business Media, LLC 2009
312 6 Applications in Numerical Variational Analysis

6A. Radius Theorems and Conditioning

In numerical analysis, a measure of “conditioning” of a problem is typically con-
ceived as a bound on the ratio of the size of solution (output) error to the size of
data (input) error. At its simplest, this pattern is seen when solving a linear equa-
tion Ax = y for x in terms of y when A is a nonsingular matrix in IRn×n . The data
input then is y and the solution output is A−1 y, but for computational purposes the
story cannot just be left at that. Much depends on the extent to which an input
error δ y leads to an output error δ x. The magnitudes of the errors can be mea-
sured by the Euclidean norm, say. Then, through linearity, there is the tight bound
|δ x| ≤ |A−1 ||δ y|, in which |A−1 | is the corresponding matrix (operator) norm of the
mapping y → A−1 y and in fact is the global Lipschitz constant for this mapping. In
providing such a bound on the ratio of |δ x| to |δ y|, |A−1 | is called the absolute con-
dition number for the problem of solving Ax = y. A high value of |A−1 | is a warning
flag signalling trouble in computing the solution x for a given y.
Another popular conditioning concept concerns relative errors instead of absolute
errors. In solving Ax = y, the relative error of the input is |δ y|/|y| (with y = 0),
while the relative error of the output is |δ x|/|x|. It is easy to see that the best bound
on the ratio of |δ x|/|x| to |δ y|/|y| is the product |A||A−1 |. Therefore, |A||A−1 | is
called the relative condition number for the problem of solving Ax = y. But absolute
conditioning will be the chief interest in our present context, for several reasons.
The reciprocal of the absolute condition number |A−1 | of a nonsingular matrix A
has a geometric interpretation which will serve as an important guide to our devel-
opments. It turns out to give an exact bound on how far A can be perturbed to A + B
before good behavior breaks down by A + B becoming singular, and thus has more
significance for numerical analysis than simply comparing the size of δ x to the size
of δ y. This property of the absolute condition number comes from a classical result
about matrices which was stated and proved in Chapter 1 as 1E.9:
  
 1
inf |B|  A + B is singular = , for any nonsingular matrix A.
|A−1 |

In this sense, |A−1 |−1 gives the radius of nonsingularity around A. As long as B lies
within that distance from A, the nonsingularity of A + B is assured. Clearly from this
angle as well, a large value of the condition number |A−1 | points toward numerical
difficulties.
The model provided to us by this example is that of a radius theorem, furnishing
a bound on how far perturbations of some sort in the specification of a problem can
go before some key property is lost. Radius theorems can be investigated not only
for solving equations, linear and nonlinear, but also generalized equations, systems
of constraints, etc.
We start down that track by stating the version of the cited matrix result that
works in infinite dimensions for bounded linear mappings acting in Banach spaces.
6 Applications in Numerical Variational Analysis 313

Theorem 6A.1 (radius theorem for invertibility of bounded linear mappings). Let
X and Y be Banach spaces and let A ∈ L (X ,Y ) be invertible1. Then
  
 1
(1) inf B  A + B is not invertible = .
B∈L (X,Y ) A−1 

Moreover the infimum is the same if restricted to mappings B of rank one.

Proof. The estimation of perturbed inversion in Lemma 5A.4 gives us “≥” in (1).
To obtain the opposite inequality and thereby complete the proof, we take any r >
1/A−1 and construct a mapping B of rank one such that A + B is not invertible
and B < r. There exists x̂ with Ax̂ = 1 and x̂ > 1/r. Choose an x∗ ∈ X ∗ such
that x∗ (x̂) = x̂ and x∗  = 1. The linear and bounded mapping

x∗ (x)Ax̂
(2) Bx = −
x̂

has B = 1/x̂ and (A + B)x̂ = Ax̂ − Ax̂ = 0. Then A + B is not invertible and
hence the infimum in (1) is ≤ r. It remains to note that B in (2) is of rank one.
The initial step that can be taken toward generality beyond linear mappings is
in the direction of positively homogeneous mappings H : X → → Y ; here and further
on, X and Y are Banach spaces. For such a mapping, ordinary norms can no longer
be of help in conditioning, but the outer and inner norms introduced in 4A in finite
dimensions and extended in 5A to Banach spaces can come into play:
+ −
H = sup sup y and H = sup inf y.
x≤1 y∈H(x) x≤1 y∈H(x)

Their counterparts for the inverse H −1 will have a role as well:

H −1  = sup H −1  = sup
+ −
sup x and inf x.
y≤1 x∈H −1 (y) y≤1 x∈H −1 (y)

In thinking of H −1 (y) as the set of solutions x to H(x)  y, it is clear that the outer
and inner norms of H −1 capture two different aspects of solution behavior, roughly
the distance to the farthest solution and the distance to the nearest solution (when
multi-valuedness is present). We are able to assert, for instance, that

dist(0, H −1 (y)) ≤ H −1  y for all y.

−

From that angle, H −1 − could be viewed as a sort of inner absolute condition

number—and in a similar manner, H −1 + could be viewed as a sort of outer ab-
solute condition number. This idea falls a bit short, though, because we only have
a comparison between sizes of x and y, not the size of a shift from x to x + δ x
1 This assumption can be dropped if we identify invertibility of A with A−1  < ∞ and adopt the
convention 1/∞ = 0. Similar adjustments can be made in the remaining radius theorems in this
section.
314 6 Applications in Numerical Variational Analysis

caused by a shift from y to y + δ y. Without H being linear, there seems little hope
of quantifying that aspect of error, not to speak of relative error. Nonetheless, it will
be possible to get radius theorems in which the reciprocals of H −1 + and H −1 −
are featured.
For H −1 + , we can utilize the inversion estimate for the outer norm in 5A.8. A
definition is needed first.

Extended nonsingularity. A positively homogeneous mapping H : X → → Y is said

to be nonsingular if H −1 + < ∞; it is said to be singular if H −1 + = ∞.
As shown in 5A.7, nonsingularity of H in this sense implies that H −1 (0) = {0};
moreover, when dim X < ∞ and gph H is closed the converse is true as well.

Theorem 6A.2 (radius theorem for nonsingularity of positively homogeneous map-

pings). For any H : X →→ Y that is positively homogeneous and nonsingular, one
has
  
 1
(3) inf B  H + B is singular = −1
.
B∈L (X,Y ) H +

Moreover the infimum is the same if restricted to mappings B of rank one.

Proof. The proof is parallel to that of 6A.1. From 5A.8 we get “≥” in (3), and also
“=” for the case H −1 + = 0 under the convention 1/0 = ∞. Let H −1 + > 0 and
consider any r > 1/H −1+ . There exists (x̂, ŷ) ∈ gph H with ŷ = 1 and x̂ > 1/r.
Let x∗ ∈ X ∗ , x∗ (x̂) = x̂ and x∗  = 1. The linear and bounded mapping

x∗ (x)ŷ
Bx = −
x̂

has B = 1/x̂ < r and (H + B)(x̂) = H(x̂) − ŷ  0. Then the nonzero vector x̂
belongs to (H + B)−1 (0), hence (H + B)−1 + = ∞, i.e., H + B is singular. The infi-
mum in (3) must therefore be less than r. Appealing to the choice of r we conclude
that the infimum in (3) cannot be more than 1/H −1+ , and we are done.
To develop a radius theorem about H −1 − , we have to look more narrowly
at sublinear mappings, which are characterized by having graphs that are not just
cones, as corresponds to positive homogeneity, but convex cones. For such a map-
ping H, if its graph is also closed, we have an inversion estimate for the inner norm
in 5C.9. Furthermore, we know from 5C.2 that the surjectivity of H is equivalent to
having H −1 − < ∞. We also have available the notion of the adjoint mapping as
introduced in Section 5C: the upper adjoint of H : X → → Y is the sublinear mapping
H ∗+ : Y ∗ →
→ X ∗ defined by

(y∗ , x∗ ) ∈ gph H ∗+ ⇐⇒ x∗ , x ≤ y∗ , y for all (x, y) ∈ gph H.

Recall too, from 5C.13, that for a sublinear mapping H with closed graph,

(H ∗+ )−1  = H −1  ,
+ −
(4)
6 Applications in Numerical Variational Analysis 315

and also, from 5C.14,

(5) (H + B)∗+ = H ∗+ + B∗ for any B ∈ L (X,Y ).

Theorem 6A.3 (radius theorem for surjectivity of sublinear mappings). For any
H:X →→ Y that is sublinear, surjective, and with closed graph,
  
 1
inf B  H + B is not surjective = .
B∈L (X,Y ) H −1 −

Moreover the infimum is the same if restricted to B of rank one.

Proof. For any B ∈ L (X,Y ), the mapping H + B is sublinear with closed graph, so
that (H + B)∗+ = H ∗+ + B∗ by (5). By the definition of the adjoint, H + B is surjective
if and only if H ∗+ + B∗ is nonsingular. It follows that
  

infB∈L (X,Y ) B  H + B is not surjective
(6)   

= infB∈L (X,Y ) B∗   H ∗+ + B∗ is singular .

The right side of (6) can be identified through Theorem 6A.2 with
  
 1
(7) inf C  H ∗+ + C is singular =
C∈L (Y ∗ ,X ∗ ) (H ∗+ )−1 +

by the observation that any C ∈ L (Y ∗ , X ∗ ) of rank one has the form B∗ for some
B ∈ L (X,Y ) of rank one. It remains to apply the relation in (4). In consequence of
that, the left side of (7) is 1/H −1− , and we get the desired equality.
In the case of H being a bounded linear mapping A : X → Y , Theorems 6A.2 and
6A.3 both furnish results which complement Theorem 6A.1, since nonsingularity
just comes down to A−1 being single-valued on rge A, while surjectivity corresponds
only to dom A−1 being all of Y , and neither of those properties automatically entails
the other. When X = Y = IRn , of course, all three theorems reduce to the matrix
result recalled at the beginning of this section.
The surjectivity result in 6A.3 offers more than an extended insight into equation
solving, however. It can be applied also to systems of inequalities. This is true even
in infinite dimensions, but we are not yet prepared to speak of inequality constraints
in that framework, so we limit the following illustration to solving Ax ≤ y in the
case of a matrix A ∈ IRm×n . It will be convenient to say that

Ax ≤ y is universally solvable if it has a solution x ∈ IRn for every y ∈ IRm .

We adopt for y = (y1 , . . . , ym ) ∈ IRm the maximum norm |y|∞ = max1≤k≤m |yk | but
equip IRn with any norm. The associated operator norm for linear mappings acting
from IRn to IRm is denoted by | · |∞ . Also, we use the notation for y = (y1 , . . . , ym ) that
+ + + +
y = (y1 , . . . , ym ), where yk = max{0, yk }.
316 6 Applications in Numerical Variational Analysis

Example 6A.4 (radius of universal solvability for systems of linear inequalities).

Suppose for a matrix A ∈ IRm×n that Ax ≤ y is universally solvable. Then
  
 1
inf |B|∞  (A + B)x ≤ y is not universally solvable = .
B∈IRm×n sup|x|≤1 | [Ax]+ |∞
  
Detail. We apply Theorem 6A.3 to the mapping F : x → y  Ax ≤ y . Then
|F −1 |− = sup|x|≤1 infy≥Ax |y|∞ , where the infimum equals | [Ax]+ |∞ .
More will be said about constraint systems in Section 6B.
Since surjectivity of a sublinear mapping is also equivalent to its metric regularity
at 0 for 0, we could restate Theorem 6A.3 in terms of metric regularity as well. Such
a result is actually true for any strictly differentiable function. Specifically, Corollary
5D.5 says that, for a function f : X → Y which is strictly differentiable at x̄, one has

reg ( f ; x̄ | ȳ) = reg D f (x̄) for ȳ = f (x̄).

Since any linear mapping is sublinear, this equality combined with 6A.3 gives us
yet another radius result.

Corollary 6A.5 (radius theorem for metric regularity of strictly differentiable func-
tions). Let f : X → Y be strictly differentiable at x̄, let ȳ := f (x̄), and let D f (x̄) be
surjective. Then
  
 1
inf B  f + B is not metrically regular at x̄ for ȳ + Bx̄ = .
B∈L (X,Y ) D f (x̄)−1 −

It should not escape attention here that in 6A.5 we are not focused any more on
the origins of X and Y but on a general pair (x̄, ȳ) in the graph of f . This allows us
to return to “conditioning” from a different perspective, if we are willing to think of
such a property in a local sense only.
Suppose that a y near to ȳ is perturbed to y + δ y. The solution set f −1 (y) to
the problem of solving f (x)  y is thereby shifted to f −1 (y + δ y), and we have an
interest in understanding the “error” vectors δ x such that x + δ x ∈ f −1 (y + δ y).
Since anyway x need not be the only element of f −1 (y), it is appropriate to quantify
the shift by looking for the smallest size of δ x, or in other words at dist(x, f −1 (y +
δ y)) and how it compares to δ y. This ratio, in its limit as (x, y) goes to (x̄, ȳ) and
δ y goes to 0, is precisely reg ( f ; x̄ | ȳ).
In this sense, reg ( f ; x̄ | ȳ) can be deemed the absolute condition number locally
with respect to x̄ and ȳ for the problem of solving f (x)  y for x in terms of y. We
then have a local, nonlinear analog of Theorem 6A.1, tying a condition number to
a radius. It provides something more even for linear f , of course, since in contrast
to Theorem 6A.1, it imposes no requirement of invertibility.
Corollary 6A.5 can be stated in a more general form, which we give here as an
exercise:
6 Applications in Numerical Variational Analysis 317

Exercise 6A.6. Let F : X → → Y with (x̄, ȳ) ∈ gph F being a pair at which gph F
is locally closed. Let F be metrically regular at x̄ for ȳ, and let f : X → Y satisfy
x̄ ∈ int dom f and lip ( f ; x̄) = 0. Then
  

inf B  F + B is not metrically regular at x̄ for ȳ + Bx̄
B∈L (X,Y )
  

= inf B  F + f + B is not metrically regular at x̄ for ȳ + f (x̄) + Bx̄ .
B∈L (X,Y )

Guide. Observe that, by the Banach space version of 3F.4 (which follows from
5E.1), the mapping F + B is metrically regular at ȳ + Bx̄ if and only if the mapping
F + f + B is metrically regular at ȳ + f (x̄) + Bx̄.
We will show next that, in finite dimensions at least, the radius result in 6A.5 is
valid when f is replaced by any set-valued mapping F whose graph is locally closed
around the reference pair (x̄, ȳ).

Theorem 6A.7 (radius theorem for metric regularity). Let X and Y be finite-
dimensional normed linear spaces, and for F : X → → Y and ȳ ∈ F(x̄) let gph F be
locally closed at (x̄, ȳ). Suppose F is metrically regular at x̄ for ȳ. Then
  
 1
(8) inf B F + B is not metrically regular at x̄ for ȳ + Bx̄ = .
B∈L (X,Y ) reg (F; x̄| ȳ)

Moreover, the infimum is unchanged if taken with respect to linear mappings of

rank 1, but also remains unchanged when the class of perturbations B is enlarged
to all locally Lipschitz continuous functions g, with B replaced by the Lipschitz
modulus lip (g; x̄) of g at x̄.
Proof. The general perturbation inequality derived in Theorem 5E.1, see 5E.7, pro-
duces the estimate
  
 1
inf lip (g; x̄)  F + g is not metrically regular at x̄ for ȳ + g(x̄) ≥ ,
g:X→Y reg (F; x̄| ȳ)

which becomes the equality (8) in the case when reg (F; x̄| ȳ) = 0 under the conven-
tion 1/0 = ∞. To confirm the opposite inequality when reg (F; x̄| ȳ) > 0, we apply
Theorem 4B.9, according to which

reg (F; x̄| ȳ) = lim sup D̃F(x|y)−1  ,

−
(9)
(x,y)→(x̄,ȳ)
(x,y)∈gph F

where D̃F(x|y) is the convexified graphical derivative of F at x for y. Take a se-

quence of positive real numbers εk  0. Then for k sufficiently large, say k > k̄, by
(9) there exists (xk , yk ) ∈ gph F with (xk , yk ) → (x̄, ȳ) and

reg (F; x̄| ȳ) + εk ≥ D̃F(xk |yk )−1  ≥ reg (F; x̄| ȳ) − εk > 0.
−
318 6 Applications in Numerical Variational Analysis

Let Hk := D̃F(xk |yk ) and Sk := Hk∗+ ; then norm duality gives us Hk−1 − = Sk−1+ ,
see 5C.13.
For each k > k̄ choose a positive real rk satisfying Sk−1 + − εk < 1/rk < Sk−1+ .
From the last inequality there must exist (ŷk , x̂k ) ∈ gph Sk with x̂k  = 1 and
Sk−1+ ≥ ŷk  > 1/rk . Pick y∗k ∈ Y with ŷk , y∗k  = ŷk  and y∗k  = 1, and define
the rank-one mapping Ĝk ∈ L (Y, X) by

y, y∗k 
Ĝk (y) := − x̂k .
ŷk 

Then Ĝk (ŷk ) = −x̂k and hence (Sk + Ĝk )(ŷk ) = Sk (ŷk ) + Ĝk (ŷk ) = Sk (ŷk ) − x̂k  0.
Therefore, ŷk ∈ (Sk + Ĝk )−1 (0), and since ŷk = 0 and Sk is positively homogeneous
with closed graph, we have by Proposition 5A.7, formula 5A(10), that

(Sk + Ĝk )−1  = ∞.

+
(10)

Note that Ĝk  = x̂k /ŷk  = 1/ŷk  < rk .

Since the sequences ŷk , x̂k and y∗k are bounded (and the spaces are finite-
dimensional), we can extract from them subsequences converging respectively to
ŷ, x̂ and y∗ . The limits then satisfy ŷ = reg (F; x̄| ȳ), x̂ = 1 and y∗  = 1. Define
the rank-one mapping Ĝ ∈ L (Y, X) by

y, y∗ 
Ĝ(y) := − x̂.
ŷ

Then we have Ĝ = reg (F; x̄| ȳ)−1 and Ĝk − Ĝ → 0.
Let B := (Ĝ)∗ and suppose F + B is metrically regular at x̄ for ȳ + Bx̄. Theorem
4B.9 yields that there is a finite positive constant c such that for k > k̄ sufficiently
large, we have
c > D̃(F + B)(xk |yk + Bxk )−1  .
−

Through 4B.10, this gives us

c > (D̃F(xk |yk ) + B)−1 = (Hk + B)−1  .

− −

Since B is linear, we have B∗ = ((Ĝ)∗ )∗ = Ĝ, and since Hk + B is sublinear, it follows

further by 5C.14 that

c > ([Hk + B]∗+ )−1  = (Hk∗+ + B∗ )−1  = (Sk + Ĝ)−1  .

+ + +
(11)

Take k > k̄ sufficiently large such that Ĝ − Ĝk  ≤ 1/(2c). Setting Pk := Sk + Ĝ and
Qk := Ĝk − Ĝ, we have that

[Pk−1 ]−1 ≥ 1/c > 1/(2c) ≥ Qk .

+

By using the inversion estimate for the outer norm in Theorem 5C.9, we have
6 Applications in Numerical Variational Analysis 319
 −1
(Sk + Ĝk )−1  = (Pk + Qk )−1  ≤ [Pk−1  ]−1 − Qk 
+ + +
≤ 2c < ∞.

This contradicts (10). Hence, F + B is not metrically regular at x̄ for ȳ + Bx̄. Noting
that B = Ĝ = 1/ reg(F; x̄| ȳ) and that B is of rank one, we are finished.
In a pattern just like the one laid out after Corollary 6A.5, it is appropriate to
consider reg (F; x̄| ȳ) as the local absolute condition number with respect to x̄ and ȳ
for the problem of solving F(x)  y for x in terms of y. An even grander extension
of the fact in 6A.1, that the reciprocal of the absolute condition number gives the
radius of perturbation for preserving an associated property, is thereby achieved.
Based on Theorem 6A.7, it is now easy to obtain a parallel radius result for strong
metric regularity.

Theorem 6A.8 (radius theorem for strong metric regularity). For finite-dimensional
normed linear spaces X and Y , let F : X → → Y have ȳ ∈ F(x̄). Suppose that F is
strongly metrically regular at x̄ for ȳ. Then
  
 1
(12) inf B F + B is not strongly regular at x̄ for ȳ + Bx̄ = .
B∈L (X,Y ) reg (F; x̄| ȳ)

Moreover, the infimum is unchanged if taken with respect to linear mappings of rank
1, but also remains unchanged when the class of perturbations B is enlarged to the
class of locally Lipschitz continuous functions g with B replaced by the Lipschitz
modulus lip (g; x̄).
Proof. Theorem 5F.1 reveals that “≥” holds in (12) when the linear perturbation is
replaced by a Lipschitz perturbation, and moreover that (12) is satisfied in the limit
case reg (F; x̄| ȳ) = 0 under the convention 1/0 = ∞. The inequality becomes an
equality with the observation that the assumed strong metric regularity of F implies
that F has locally closed graph at (x̄, ȳ) and is metrically regular at x̄ for ȳ. Hence
the infimum in (12) is not greater than the infimum in (8).
Next comes a radius theorem for strong subregularity to go along with the ones
for metric regularity and strong metric regularity.

Theorem 6A.9 (radius theorem for strong metric subregularity). Let X and Y be
finite-dimensional normed linear spaces X and Y , and for F : X → → Y and ȳ ∈ F(x̄)
let gph F be locally closed at (x̄, ȳ). Suppose that F is strongly metrically subregular
at x̄ for ȳ. Then
  
 1
inf B  F +B is not strongly subregular at x̄ for ȳ + Bx̄ = .
B∈L (X,Y ) subreg (F; x̄| ȳ)

Moreover, the infimum is unchanged if taken with respect to mappings B of rank 1,

but also remains unchanged when the class of perturbations is enlarged to the class
of functions g : X → Y that are calm at x̄ and continuous around x̄, with B replaced
by the calmness modulus clm (g; x̄).
320 6 Applications in Numerical Variational Analysis

Proof. From the equivalence of the strong subregularity of a mapping F at x̄ for

ȳ with the nonsingularity of its graphical derivative DF(x̄| ȳ), as shown in Theo-
rem 4C.1, we have
  

inf B F + B is not strongly subregular at x̄ for ȳ + Bx̄
B∈L (X,Y )   
(13) 
= inf B  D(F + B)(x̄| ȳ + Bx̄) is singular .
B∈L (X,Y )

We know from the sum rule for graphical differentiation (4A.1) that D(F + B)(x̄| ȳ+
Bx̄) = DF(x̄| ȳ) + B, hence
  

inf B  D(F + B)(x̄| ȳ + Bx̄) is singular
B∈L (X,Y )   
(14) 
= inf B  DF(x̄| ȳ) + B is singular .
B∈L (X,Y )

Since DF(x̄| ȳ) + B is positively homogeneous, 6A.2 translates to

  
 1
(15) inf B  DF(x̄| ȳ) + B is singular = ,
B∈L (X,Y ) DF(x̄| ȳ)−1 +

including the case DF(x̄| ȳ)−1 + = 0 with the convention 1/0 = ∞. Theorem 4C.1
tells us also that DF(x̄| ȳ)−1 + = subreg(F; x̄| ȳ) and then, putting together (13),
(14) and (15), we get the desired equality.
As with the preceding results, the modulus subreg(F; x̄| ȳ) can be regarded
as a sort of local absolute condition number. But in this case only the ratio of
dist(x̄, F −1 (ȳ + δ y)) to δ y is considered in its limsup as δ y goes to 0, not the
limsup of all the ratios dist(x, F −1 (y + δ y))/δ y with (x, y) ∈ gph F tending to
(x̄, ȳ), which gives reg (F; x̄| ȳ). Specifically, with reg (F; x̄| ȳ) appropriately termed
the absolute condition number for F locally with respect to x̄ and ȳ, subreg(F; x̄| ȳ)
is the corresponding subcondition number.
The radius-type theorems above could be rewritten in terms of the associated
equivalent properties of the inverse mappings. For example, Theorem 6A.7 could
be restated in terms of perturbations B of a mapping F whose inverse has the Aubin
property.

6B. Constraints and Feasibility

Universal solvability of systems of linear inequalities in finite dimensions was al-

ready featured in Example 6A.4 as an application of one of our radius theorems,
but now we will go more deeply into the subject of constraint systems and their
solvability. We take as our focus problems of the very general type
6 Applications in Numerical Variational Analysis 321

(1) find x such that F(x)  0

for a set-valued mapping F : X → → Y from one Banach space to another. Of course,

the set of all solutions is just F −1 (0), but we are thinking of F as representing a kind
of constraint system and are concerned with whether the set F −1 (0) might shift from
nonempty to empty under some sort of perturbation. Mostly, we will study the case
where F has convex graph.

Feasibility. Problem (1) will be called feasible if F −1 (0) = 0/ , i.e., 0 ∈ rge F , and
strictly feasible if 0 ∈ int rge F .
Two examples will point the way toward progress. Recall that any closed, convex
cone K ⊂ Y with nonempty interior induces a partial ordering “≤K ” under the rule
that y0 ≤K y1 means y1 − y0 ∈ K. Correspondingly, y0 <K y1 means y1 − y0 ∈ int K.
Example 6B.1 (convex constraint systems). Let C ⊂ X be a closed convex set, let
K ⊂ Y be a closed convex cone, and let A : C → Y be a continuous and convex
mapping with respect to the partial ordering in Y induced by K; that is,

A((1 − θ )x0 + θ x1 ) ≤K (1 − θ )A(x0 ) + θ A(x1) for x0 , x1 ∈ C when 0 < θ < 1.

Define the mapping F : X →

→ Y by

F(x) = A(x) + K if x ∈ C,
0/ if x ∈
/ C.
Then F has closed, convex graph, and feasibility of solving F(x)  0 for x refers to

∃ x̄ ∈ C such that A(x̄) ≤K 0.

On the other hand, as long as int K = 0,

/ strict feasibility refers to

∃ x̄ ∈ C such that A(x̄) <K 0.

Detail. To see this, note that, when int K = 0,

/ we have K = cl int K, so that the
convex set rge F = A(C) + K is the closure of the open set O := A(C) + int K. Also,
O is convex. It follows then that int rge F = O.

Example 6B.2 (linear-conic constraint systems). Consider the convex constraint

system in Example 6B.1 under the additional assumptions that A is linear and C is
a cone, so that the condition x̄ ∈ C can be written equivalently as x̄ ≥C 0. Then F is
sublinear, and its adjoint F ∗+ : Y ∗ →
→ X ∗ is given in terms of the adjoint A∗ of A and
the dual cones K = −K and C = −C∗ (where ∗ denotes polar) by
+ ∗ +

 ∗ ∗
∗+ ∗ A (y ) − C+ if y∗ ∈ K + ,
F (y ) =
0/ if y∗ ∈
/ K+,

so that F ∗+ (y∗ )  x∗ if and only if y∗ ≥K+ 0 and A∗ (y∗ ) ≥C+ x∗ .

322 6 Applications in Numerical Variational Analysis

Detail. In this case the graph of F is clearly a convex cone, and that means F is
sublinear. The claims about the adjoint of F follow by elementary calculation.
Along the lines of the analysis in 6A, in dealing with the feasibility problem (1)
we will be interested in perturbations in which F is replaced by F + B for some
B ∈ L (X ,Y ), and at the same time, the zero on the right is replaced by some other
b ∈ Y . Such a double perturbation, the magnitude of which can be quantified by the
norm
 
(2) (B, b) = max B, b ,

transforms the condition F(x)  0 to (F + B)(x)  b and the solution set F −1 (0) to
(F + B)−1 (b), creating infeasibility if (F + B)−1 (b) = 0,
/ i.e., if b ∈
/ rge(F + B). We
want to understand how large (B, b) can be before this happens.

Distance to infeasibility. For F : X → → Y with convex graph and 0 ∈ rge F , the

distance to infeasibility of the system F(x)  0 is defined to be the value
  

(3) inf (B, b)  b ∈
/ rge(F + B) .
B∈L (X,Y ), b∈Y

Surprisingly, perhaps, it turns out that there would be no difference if feasibility

were replaced by strict feasibility in this definition. In the next pair of lemmas, it is
assumed that F : X →→ Y has convex graph and 0 ∈ rge F.

Lemma 6B.3 (distance to infeasibility versus distance to strict infeasibility). The

distance to infeasibility is the same as the distance to strict infeasibility, namely the
value
  

(4) inf (B, b)  b ∈
/ int rge(F + B) .
B∈L (X,Y ), b∈Y

Proof. Let S1 denote the set of (B, b) over which the infimum is taken in (3) and let
S2 be the corresponding set in (4). Obviously S1 ⊂ S2 , so the first infimum cannot
be less than the second. We must show that it also cannot be greater. This amounts
to demonstrating that for any (B, b) ∈ S2 and any ε > 0 we can find (B , b ) ∈ S1
such that (B , b ) ≤ (B, b) + ε . In fact, we can get this with B = B simply by
noting that when b ∈ / int rge(F + B) there must exist b ∈ Y with b ∈ / rge(F + B)

and b − b ≤ ε .
By utilizing the Robinson–Ursescu theorem 5B.4, we can see furthermore that
the distance to infeasibility is actually the same as the distance to metric nonregu-
larity:

Lemma 6B.4 (distance to infeasibility equals radius of metric regularity). The dis-
tance to infeasibility in problem (1) coincides with the value
  

(5) inf (b, B)  F + B is not metrically regular at any x̄ for b .
(B,b)∈L (X,Y )×Y
6 Applications in Numerical Variational Analysis 323

Proof. In view of the equivalence of infeasibility with strict feasibility in 6B.3, the
Robinson–Ursescu theorem 5B.4 just says that problem (1) is feasible if and only if
F is metrically regular at x̄ for 0 for any x̄ ∈ F −1 (0), hence (5).
In order to estimate the distance to infeasibility in terms of the modulus of metric
regularity, we pass from F to a special mapping F̄ constructed as a “homogeniza-
tion” of F. We will then be able to apply to F̄ the result on distance to metric
nonregularity of sublinear mappings given in 6A.3.
We use the horizon mapping F ∞ associated with F, the graph of F ∞ in X × Y
being the recession cone of gph F in the sense of convex analysis:

(x , y ) ∈ gph F ∞ ⇐⇒ gph F + (x , y ) ⊂ gph F.

Homogenization. For F : X → → Y and 0 ∈ rge F , the homogenization of the con-

straint system F(x)  0 in (1) is the system F̄(x,t)  0, where F̄ : X × IR →
→ Y is
defined by
%
tF(t −1 x) if t > 0,
F̄(x,t) = F ∞ (x) if t = 0,
0/ if t < 0.
The solution sets to the two systems are related by

x ∈ F −1 (0) ⇐⇒ (x, 1) ∈ F̄ −1 (0).

Note that if F is positively homogeneous with closed graph, then tF(t −1 x) =

F(x) = F ∞ (x) for all t > 0, so that we simply have F̄(x,t) = F(x) for t ≥ 0, but
F̄(x,t) = 0/ for t < 0.
In what follows, we adopt the norm

(6) (x,t) = x + |t| for (x,t) ∈ X × IR.

We are now ready to state and prove a result which gives a quantitative expression
for the magnitude of the distance to infeasibility:

Theorem 6B.5 (distance to infeasibility for the homogenized mapping). Let F :

X→→ Y have closed, convex graph and let 0 ∈ rge F . Then in the homogenized sys-
tem F̄(t, x)  0 the mapping F̄ is sublinear with closed graph, and

(7) 0 ∈ int rge F ⇐⇒ 0 ∈ int rge F̄ ⇐⇒ F̄ is surjective.

Furthermore, for the given constraint system F(x)  0 one has

(8) distance to infeasibility = 1/ reg(F̄; 0, 0|0).

Proof. The definition of F̄ corresponds

 to gph F̄ being the closed convex cone in
X × IR ×Y that is generated by (x, 1, y)  (x, y) ∈ gph F . Hence F̄ is sublinear, and
also, rge F̄ is a convex cone. We have (rge F) = F(X) = F̄(X, 1). So it is obvious
324 6 Applications in Numerical Variational Analysis

that if 0 ∈ int rge F, then 0 ∈ int rge F̄. Since rge F̄ is a convex cone, the latter is
equivalent to having rge F̄ = Y , i.e., surjectivity.
Conversely now, suppose F̄ is surjective. Theorem 5B.4 (Robinson–Ursescu) in-
forms us that in this case, 0 ∈ int F̄(W ) for every neighborhood W of the origin in
IR × X. It must be verified, however, that 0 ∈ int rge F. In terms of C(t) = F̄(IB,t) ⊂
Y , it will suffice to show that 0 ∈ int C(t) for some t > 0. Note that the sublinearity
of F̄ implies that

(9) C((1 − θ )t0 + θ t1 ) ⊃ (1 − θ )C(t0 ) + θ C(t1 ) for 0 < θ < 1.

Our assumption that 0 ∈ rge F ensures having F −1 (0) = 0.

/ Choose τ ∈ (0, ∞) small
enough that 1/(2τ ) > d(0, F −1 (0)). Then

(10) 0 ∈ C(t) for all t ∈ [0, 2τ ],

whereas, because [−2τ , 2τ ] × IB is a neighborhood W of the origin in IR × X, we

have
,
(11) 0 ∈ int F̄(IB, [−2τ , 2τ ]) = int C(t).
0≤t≤2τ

We will use this to show that actually 0 ∈ int C(τ ). For y∗ ∈ Y ∗ define

σ (y∗ ,t) := sup y, y∗ , λ (t) := inf

∗
σ (y∗ ,t).
y∈C(t) y =1

The property in (9) makes σ (y∗ ,t) concave in t, and the same then follows for λ (t).
As long as 0 ≤ t ≤ 2τ , we have σ (y∗ ,t) ≥ 0 and λ (t) ≥ 0 by (10). On the other
hand, the union in (11) includes some ball around the origin. Therefore,

(12) ∃ ε > 0 such that sup σ (y∗ ,t) ≥ ε for all y∗ ∈ Y ∗ with y∗  = 1.
0≤t≤2τ

We argue next that λ (τ ) > 0. If not, then since λ is a nonnegative concave func-
tion on [0, 2τ ], we would have to have λ (t) = 0 for all t ∈ [0, 2τ ]. Supposing that
to be the case, choose δ ∈ (0, ε /2) and, in the light of the definition of λ (τ ), an
element ŷ∗ with σ (ŷ∗ , τ ) < δ . The nonnegativity and concavity of σ (ŷ∗ , ·) on [0, 2τ ]
imply then that σ (ŷ∗ ,t) ≤ (δ /τ )t when τ ≤ t ≤ 2τ and σ (ŷ∗ ,t) ≤ 2δ − (δ /τ )t when
0 ≤ t ≤ τ . But that gives us σ (ŷ∗ ,t) ≤ 2δ < ε for all t ∈ [0, 2τ ], in contradiction to
the property of ε in (12). Therefore, λ (τ ) > 0, as claimed.
We have σ (y∗ , τ ) ≥ λ (τ ) when y∗  = 1, and hence by positive homogeneity
σ (y∗ , τ ) ≥ λ (τ )y∗  for all y∗ ∈ Y ∗ . In this inequality, σ (·, τ ) is the support function
of the convex set C(τ ), or equivalently of cl C(τ ), whereas λ (τ ) ·  is the support
function of λ (τ )IB. It follows therefore that cl C(τ ) ⊃ λ (τ )IB, so that at least 0 ∈
int cl C(τ ).
Now, remembering that C(τ ) = τ F(τ −1 IB), we obtain 0 ∈ int cl F(τ −1 IB). Con-
sider the mapping
6 Applications in Numerical Variational Analysis 325

F̃(x) = F(x) if x ∈ τ −1 IB,
0/ otherwise.
Clearly rge F̃ ⊂ rge F. Applying Theorem 5B.1 to the mapping F̃ gives us

0 ∈ int cl rge F̃ = int rge F̃ ⊂ int rge F.

This completes the proof of (7).

Let us turn now to (8). The first thing to observe is that every B̄ ∈ L (X × IR,Y )
can be identified with a pair (B, b) ∈ L (X ,Y )×Y under the formula B̄(x,t) = B(x)−
tb. Moreover, under this identification we get B̄ equal to the expression in (2), due
to the choice of norm in (6). The next thing to observe is that
⎧
⎨ t(F + B)(t −1x) − tb if t > 0,
(F̄ + B̄)(x,t) = (F + B)∞ (x) if t = 0,
⎩
0/ if t < 0,

so that F̄ + B̄ gives the homogenization of the perturbed system (F + B)(x)  b.

Therefore, on the basis of what has so far been proved, we have

b ∈ int rge(F + B) ⇐⇒ F̄ + B̄ is surjective.

Hence, through Lemma 6B.4, the distance to infeasibility for the system F(x)  0
is the infimum of B̄ over all B̄ ∈ L (X × IR,Y ) such that F̄ + B̄ is not surjective.
Theorem 6A.3 then furnishes the conclusion in (8).
Passing to the adjoint mapping, we can obtain a “dual” formula for the distance
to infeasibility:

Corollary 6B.6 (distance to infeasibility for closed convex processes). Let F :

X→→ Y have closed, convex graph, and let 0 ∈ rge F . Define the convex, positively
homogeneous function h : X ∗ × Y ∗ → (−∞, ∞] by
  
h(x∗ , y∗ ) = supx,y x, x∗  − y, y∗   y ∈ F(x) .

Then for the system F(x)  0,

 
(13) distance to infeasibility = inf max x∗ , h(x∗ , y∗ ) .
y∗ =1, x∗

Proof. By Theorem 6B.5, the distance to infeasibility is 1/ reg(F̄; 0, 0|0). On the

other hand, reg (F̄; 0, 0|0) = (F̄ ∗+ )−1 + for the adjoint mapping F̄ ∗+ : Y ∗ →
→ X∗ ×
∗ ∗+ ∗ ∗ ∗
IR. By definition, (x , s) ∈ F̄ (y ) if and only if (x , s, −y ) belongsto the polar
 cone
∗ . Because gph F̄ is the closed convex cone generated by (x, 1, y)  (x, y) ∈
(gph F̄)

gph F , this condition is the same as

s + x, x∗  − y, y∗  ≤ 0 for all (x, y) ∈ gph F

and can be expressed as s + h(x∗ , y∗ ) ≤ 0. Hence

326 6 Applications in Numerical Variational Analysis
  

(14) (F ∗+ )−1 + = sup y∗   (x∗ , s) ≤ 1, s + h(x∗ , y∗ ) ≤ 0 ,
 
where the norm on X ∗ × IR dual to the one in (6) is (x∗ , s) = max x∗ , |s| . The
distance to infeasibility, being the reciprocal of the quantity in (14), can be expressed
therefore (through the positive homogeneity of h) as
    
(15) inf max x∗ , |s|  s + h(x∗, y∗ ) ≤ 0 .
y∗ =1, x∗ , s

(In converting from (14) to an infimum restricted to y∗  = 1 in (15), we need

to be cautious about the possibility that there might be no elements (x∗ , s, y∗ ) ∈
gph (F̄ ∗+ )−1 with y∗ = 0, in which case the infimum in (15) is ∞. But then the
expression in (10) is 0, so the statement remains correct under the convention 1/∞ =
0,) Observe next that, in the infimum in (15), s will be taken to be as near to 0
∗ ∗
as possible while 
∗ ∗
 −s ≥ h(x , y ). Thus, |s| will be the max of 0 and
maintaining
h(x , y ), and max x, |s| will be the max of these two quantities and x —but
then the 0 is superfluous, and we end up with (15) equaling the expression on the
right side of (3).
We can now present our result for homogeneous systems:

Corollary 6B.7 (distance to infeasibility for sublinear mappings). Let F : X → →Y

be sublinear with closed graph and let 0 ∈ rge F . Then for the inclusion F(x)  0,
 
distance to infeasibility = inf d 0, F ∗+ (y∗ ) .
y∗ =1

Proof. In this case the function h in 6B.6 has h(x∗ , y∗ ) = 0 when x∗ ∈ F ∗+ (y∗ ), but
h(x∗ , y∗ ) = ∞ otherwise.
In particular, for a linear-conic constraint system of type x ≥C 0, A(x) ≤K 0, with
respect to a continuous linear mapping A : X → Y and closed, convex cones C ⊂ X
and K ⊂ Y , we obtain

distance to infeasibility = inf d(A∗ (y∗ ),C+ ).

y∗ ∈K + , y∗ =1

6C. Iterative Processes for Generalized Equations

Our occupation with numerical matters turns even more serious in this section,
where we consider computational methods for solving generalized equations. The
problem is to

(1) find x such that f (x) + F(x)  0,

6 Applications in Numerical Variational Analysis 327

where f : X → Y is a continuous function and F : X → → Y is a set-valued mapping

with closed graph; both X and Y are Banach spaces. As we already know, the model
of a generalized equation covers huge territory. The classical case of nonlinear equa-
tions corresponds to having F = 0, whereas by taking F ≡ −K for a fixed set K one
gets various constraint systems. When F is the normal cone mapping NC associated
with a closed, convex set C ⊂ X , and Y = X ∗ , we have a variational inequality.
With the aim of approximating a solution to the generalized equation (1), we
consider the following general iterative process: Choose a sequence of functions
Ak : X × X → Y and an initial point x0 , and generate a sequence {xk }∞ k=0 iteratively
by taking xk+1 to be a solution to the auxiliary generalized equation

(2) Ak (xk+1 , xk ) + F(xk+1 )  0, for k = 0, 1, . . . .

Our goal is to specify conditions on the sequence of functions Ak in relation to the

function f under which the process (2) is convergent, at least in a certain sense. We
don’t take on the task of explaining how the subproblems in (2) might themselves be
solved. That is a separate issue, but of course those subproblems ought to be chosen
to be simpler and easier to solve, depending on the form of F. Our concern here
lies only with the process defined by (2) and what our earlier results are able to say
about it.
Specific choices of the sequence of mappings Ak in the general iterative process
(2) lead to known computational methods for solving (1). Under the assumption that
f is differentiable, if we take Ak (x, u) = f (u) + D f (u)(x − u) for all k, the iteration
(2) becomes the following version of Newton’s method for solving the generalized
equation (1):

(3) f (xk ) + D f (xk )(xk+1 − xk ) + F(xk+1 )  0, for k = 0, 1, . . . .

This approach uses “partial linearization,” in which we linearize f at the current

point but leave F intact. It reduces to the standard version of Newton’s method for
solving the nonlinear equation f (x) = 0 when F is the zero mapping. We used this
method to prove the classical inverse function theorem 1A.1.
In the case when (1) represents the optimality systems for a nonlinear program-
ming problem, the iteration (3) becomes the popular sequential quadratic program-
ming (SQP) algorithm for optimization. We will briefly describe the SQP algorithm
later in the section.
Although one might imagine that a “true” Newton-type method for (1) ought to
involve some kind of approximation to F as well as f , such an extension runs into
technical difficulties, in particular for infinite-dimensional variational problems.
If we choose Ak (x, u) = λk (x − u) + f (x) in (2) for some sequence of positive
numbers λk , we obtain the basic form of the proximal point method:

(4) λk (xk+1 − xk ) + f (xk+1 ) + F(xk+1 )  0, for k = 0, 1, . . . .

At the end of the section we will provide more details about this method, putting it
in the perspective of monotone mappings and optimization problems.
328 6 Applications in Numerical Variational Analysis

Our main result, which follows, concerns convergence of the iterative process (2)
under the assumption of metric regularity of the mapping f + F.

Theorem 6C.1 (convergence under metric regularity). Let x̄ be a solution to (1),

let the mapping f + F be metrically regular at x̄ for 0 and let κ > reg ( f + F; x̄|0).
Consider a sequence of mappings Ak : X × X → Y with the following property: there
exist sequences of nonnegative numbers {εk } and { μk } satisfying
κ μk
(5) sup κεk < 1 and sup < 1,
k k 1 − κεk

and a neighborhood U of x̄ so that

(6)  f (x) − Ak (x, u) − [ f (x ) − Ak (x , u)] ≤ εk x − x  for every x, x , u ∈ U,

(7) Ak (x̄, u) − f (x̄) ≤ μk u − x̄ for every u ∈ U,

for all k = 0, 1, . . . .
Then there is a neighborhood O of x̄ such that, for any starting point x0 ∈ O and
any sequence δk  0 satisfying

κ μk + δk
(8) γk := < 1 for k = 0, 1, . . . ,
1 − κεk

there exists a sequence {xk } generated by the procedure (2) which converges to x̄
with

(9) xk+1 − x̄ ≤ γk xk − x̄ for all k = 0, 1, . . . .

Proof. Let constants a > 0 and b > 0 be such that the mapping f + F is metrically
regular at x̄ for 0 with constant κ and neighborhoods IBa (x̄) and IBb (0). Make a
smaller if necessary so that IBa (x̄) ⊂ U and

(10) (εk + μk )a ≤ b for all k.

From the second inequality in (5) there exists a sequence δk  0 satisfying (8);
choose such a sequence and determine γk from (8). Pick x0 ∈ IBa (x̄). If x0 = x̄ then
take xk = x̄ for all k and there is nothing more to prove. If not, consider the function
x → g0 (x) := f (x) − A0 (x, x0 ). For any x ∈ IBa (x̄), using (6), (7) and (10), and noting
that Ak (x̄, x̄) = f (x̄) from (7), we have

g0 (x) =  f (x) − A0 (x, x0 ) − f (x̄) + A0 (x̄, x̄)

(11) ≤  f (x) − A0 (x, x0 ) − f (x̄) + A0 (x̄, x0 ) + A0(x̄, x0 ) − A0 (x̄, x̄)
≤ ε0 x − x̄ + μ0 x0 − x̄ ≤ ε0 a + μ0a ≤ b.
6 Applications in Numerical Variational Analysis 329

We will demonstrate that the mapping Φ0 : x → ( f + F)−1 (g0 (x)) satisfies the as-
sumptions of the contraction mapping principle for set-valued mappings (Theorem
5E.2).
By virtue of the metric regularity of f + F, the form of g0 , the fact that − f (x̄) =
−A(x̄, x̄) ∈ F(x̄), and (7), we have

d(x̄, Φ0 (x̄)) = d(x̄, ( f + F)−1 (g0 (x̄))) ≤ κ d(g0 (x̄), ( f + F)(x̄))

= κ d(−A0 (x̄, x0 ), F(x̄)) ≤ κ A0 (x̄, x0 ) − A(x̄, x̄)
≤ κ μ0 x0 − x̄ < γ0 x0 − x̄(1 − κε0),

where γ0 is defined in (8) and hence, γ0 x0 − x̄ ≤ a. Let u, v ∈ IBγ0 x0 −x̄ (x̄). In-
voking again the metric regularity of f + F as well as the estimate (11), for any
u, v ∈ IBγ0 x0 −x̄ (x̄) we obtain

e(Φ0 (u) ∩ IBγ0 x0 −x̄ (x̄), Φ0 (v)) ≤ e(Φ0 (u) ∩ IBa (x̄), Φ0 (v))
  
= sup d(x, ( f + F)−1 (g0 (v)))  x ∈ ( f + F)−1 (g0 (u)) ∩ IBa (x̄)
  
≤ sup κ d(g0 (v), f (x) + F(x))  x ∈ ( f + F)−1 (g0 (u)) ∩ IBa (x̄)
≤ κ  f (v) − A0 (v, x0 ) − [ f (u) − A0(u, x0 )] ≤ κε0 u − v.

Hence, by the contraction mapping principle 5E.2 there exists a fixed point x1 ∈
Φ0 (x1 ) ∩ IBγ0 x0 −x̄ (x̄). This translates to g0 (x1 ) = f (x1 ) − A0 (x1 , x0 ) ∈ ( f + F)(x1 ),
meaning that x1 is obtained from x0 by iteration (2) and satisfies (9) for k = 0.
The induction step is now clear. If xk ∈ IBa (x̄) and xk = x̄, by defining gk (x) =
f (x) − Ak (x, xk ), we obtain as in (11) that gk (x) ≤ b for all x ∈ IBak (x̄). Then
Theorem 5E.2 applies to Φk : x → ( f + F)−1 (gk (x)) on the ball IBγk xk −x̄ (x̄) and
yields the existence of an iterate xk+1 satisfying (9). The condition in (8) ensures
that the sequence {xk } is convergent and its limit is x̄.
It is a standard concept in numerical analysis that a sequence {xk } is linearly
convergent to x̄ when
xk+1 − x̄
lim sup < 1.
k→∞ xk − x̄
Thus, the sequence {xk } whose existence is claimed in 6C.1, is linearly convergent
to x̄. If the stronger condition

xk+1 − x̄
lim =0
k→∞ xk − x̄

holds, then xk is said to be superlinearly convergent to x̄. If an iterative method, like

(2), produces a sequence which is linearly (superlinearly) convergent, one says that
the method itself is linearly (superlinearly) convergent.

Corollary 6C.2 (superlinear convergence). Under the conditions of Theorem 6C.1,

assume in addition, that the sequence { μk } can be chosen to satisfy
330 6 Applications in Numerical Variational Analysis

(12) lim μk = 0.
k→∞

Then the sequence of iterates xk according to (2), whose existence is claimed in

6C.1, is superlinearly convergent to x̄.
Proof. In this case the sequence {γk } in (8) converges to zero and then the claimed
mode of convergence follows from (9).

Exercise 6C.3 (convergence under strong metric subregularity). Let x̄ be a solution

of (1). Let f + F be strongly subregular at x̄ for 0, and consider the iteration process
(2) under the conditions (5), (6) and (7) in 6C.1, where κ > subreg( f + F; x̄|0).
Prove there exists a neighborhood O of x̄ such that, if a sequence {xk } is generated
by (2) and has all its elements xk ∈ O, then {xk } is linearly convergent to x̄. Under
(12), this sequence is superlinearly convergent.
Guide. Utilizing the strong metric subregularity of f + F and in particular from
formula 3I(5), show the existence of a positive constant a such that

(13) x − x̄ ≤ κ d(0, ( f + F)(x)) for all x ∈ IBa (x̄).

Let xk ∈ IBa (x̄) for k = 0, 1, . . . be generated by (2), and let gk (x) = f (x) − Ak (x, xk ).
From (6) and (7),

gk (xk+1 ) ≤  f (xk+1 ) − Ak (xk+1 , xk ) − [ f (x̄) − Ak (x̄, xk )]

+ Ak (x̄, xk ) − f (x̄) ≤ εk xk+1 − x̄ + μk xk − x̄.

Since gk (xk+1 ) ∈ ( f + F)(xk+1 ), (13) for x = xk+1 gives us

xk+1 − x̄ ≤ κεk xk+1 − x̄ + κ μk xk − x̄,

which leads to the desired conclusion.

Note that 6C.1 claims the existence of a convergent sequence which of course
may be not unique, whereas 6C.3 doesn’t guarantee the existence of any sequence
generated by (2), inasmuch as strong subregularity doesn’t guarantee local solvabil-
ity of (2). We do obtain existence and uniqueness of a sequence generated by (2)
when we assume strong metric regularity of the mapping in (1).

Exercise 6C.4 (convergence under strong metric regularity). Under the conditions
of Theorem 6C.1, assume in addition that the mapping f + F is strongly metrically
regular at x̄ for 0. Then there exists a neighborhood O of x̄ such that, for any x0 ∈ O,
there is a unique sequence {xk } generated by the iterative process (2). This sequence
is linearly convergent to x̄. If (12) holds, the sequence is superlinearly convergent.
Guide. This could be verified in several ways, one of which is to repeat the proof
of 6C.1 using the standard contraction mapping principle, 1A.2, instead of 5E.2.
We will see next what the assumptions in 6C.1 mean in the specific cases of
Newton’s method (3) and proximal point method (4).
6 Applications in Numerical Variational Analysis 331

Corollary 6C.5 (convergence of Newton’s method). Consider Newton’s method

(3) as a specific case of the iteration (2) under the assumptions that x̄ is a solution to
(1) and the function f is continuously differentiable near x̄. Then we have:
(i) if f + F is metrically regular at x̄ for 0, then there exists a neighborhood O
of x̄ such that, for any x0 ∈ O, there is a sequence {xk } generated by the method
starting at x0 which is linearly convergent to x̄;
(ii) if f + F is strongly metrically subregular at x̄ for 0, then there exists a neigh-
borhood O of x̄ such that any sequence {xk } generated by the method which is
contained in O is linearly convergent to x̄;
(iii) if f + F is strongly metrically regular at x̄ for 0, then for the neighborhood O
in (i) and any x0 ∈ O the method generates a unique sequence {xk }. This sequence,
according to (i), is linearly convergent to x̄.
Proof. In this case the sequence of mapping is A(x, u) = f (u) + D f (u)(x − u), the
same for all k. It is straightforward to check that the assumed smoothness of f near
 ( f − A; (x̄, x̄)) = 0 and clm (A(x̄, ·); x̄) = 0. Thus, we can choose
x̄ implies both lip x
constant sequences, εk = ε and μk = μ , for sufficiently small ε and μ , obtaining
through 6C.1–4 the claimed modes of convergence.
Under a stronger condition on the function f , Newton’s method (3) converges
quadratically. The precise result is as follows.

Theorem 6C.6 (quadratic convergence of Newton’s method). Consider Newton’s

method (3) for a function f which is continuously differentiable near x̄ and such
that lip (D f ; x̄) < ∞. Assume that the mapping f + F is metrically regular at x̄ for 0.
Then for any γ satisfying
1
(14) γ> reg ( f + F; x̄|0)·lip (D f ; x̄)
2
there exists a neighborhood O of x̄ such that, for any x0 ∈ O, there is a sequence
{xk } generated by the method which is quadratically convergent to x̄ in the sense
that

(15) xk+1 − x̄ ≤ γ xk − x̄2 for k = 0, 1, . . . .

Proof. The proof follows the fixed point argument in the proof of 6C.1, but with
some modifications that require attention. Choose γ as in (14) and let

(16) κ > reg ( f + F; x̄|0) and μ > lip (D f ; x̄) be such that γ > κ μ /2.

Further, choose a > 0 and b > 0 so that f + F is metrically regular at x̄ for 0 with
constant κ and neighborhoods IBa (x̄) and IBb (0). Make a > 0 smaller if necessary
so that

(17) D f (x ) − D f (x) ≤ μ x − x for x , x ∈ IBa (x̄)

and also, taking into account (16),

332 6 Applications in Numerical Variational Analysis

5 2 1
(18) μ a ≤ b, κ μ a < 1, κ μ < γ (1 − κ μ a) and γ a < 1.
2 2
Here and in the following section we use an estimate for smooth functions obtained
by elementary calculus. From the standard equality
 1
f (u) − f (v) = D f (v + t(u − v))(u − v)dt,
0

we have through (17) that, for all u, v ∈ IBa (x̄),

 f (u) − f (v) − D f (v)(u − v)

 
=  01 Dx f (v + t(u − v))(u − v)dt − D f (v)(u − v) ≤ μ 01 tdtu − v2.

This yields
1
(19)  f (u) − f (v) − D f (v)(u − v) ≤ μ u − v2.
2
Fix w ∈ IBa (x̄) and consider the function

x → g(w, x) = f (x̄) + D f (x̄)(x − x̄) − f (w) − D f (w)(x − w).

Using (19), and then the first condition in (18), we get

g(w, x) ≤  f (x̄) + D f (x̄)(x − x̄) − f (x)

(20) + f (x) − f (w) − D f (w)(x − w)
≤ 12 μ x − x̄2 + 12 μ x − w2 ≤ 12 μ a2 + 12 μ 4a2 ≤ b.

Pick x0 ∈ IBa (x̄), x0 = x̄, and consider the mapping Φ0 (x) = ( f + F)−1 (g(x0 , x)).
Noting that 0 ∈ ( f + F)(x̄) and using the metric regularity of f + F together with
(20), and also (18) and (19), we obtain

d(x̄, Φ0 (x̄)) = d(x̄, ( f + F)−1 (g(x0 , x̄))) ≤ κ d(g(x0 , x̄), f (x̄) + F(x̄))
≤ κ  f (x̄) − f (x0 ) − D f (x0 )(x̄ − x0 )
1
≤ κ μ x0 − x̄2 < r0 (1 − κ μ a),
2
where r0 = γ x0 − x̄2 ≤ a. Moreover, for any u, v ∈ IBr0 (x̄),

e(Φ0 (u) ∩ IBr0 (x̄), Φ0 (v)) ≤ κ g(x0 , u) − g(x0, v)

≤ κ (D f (x̄) − D f (x0 ))(u − v) ≤ κ μ au − v.

Hence, by 5E.2 there exists x1 ∈ Φ (x1 ) ∩ IBr0 (x̄), which translates to having x1 ob-
tained from x0 as a first iterate of Newton’s method (3) and satisfying (15) for k = 0.
The induction step is completely analogous, giving us a sequence {xk } which satis-
fies (15). Since γ a < 1 as required in (18), this sequence is convergent.
6 Applications in Numerical Variational Analysis 333

In the following section we will present a more elaborate parameterized version

of 6C.6 under strong metric regularity, in which case, the sequence {xk }, whose
existence is claimed in 6C.6, is unique in O.

Corollary 6C.7 (convergence of proximal point method). Consider the proximal

point method (4) as a specific case of the iteration (2) under the assumptions that x̄
is a solution to (1) and the function f is continuous at x̄.
(i) If f + F is metrically regular at x̄ for 0 and
1
(21) supk λk < ,
2 reg ( f + F; x̄|0)

then there exists a neighborhood O of x̄ such that for any x0 ∈ O there is a sequence
{xk } generated by the method starting at x0 which is linearly convergent to x̄.
(ii) If f + F is strongly metrically subregular at x̄ for 0 and
1
(22) supk λk < ,
2 subreg( f + F; x̄|0)

then there exists a neighborhood O of x̄ such that any sequence {xk } generated by
the method which is contained in O is linearly convergent to x̄.
(iii) If f + F is strongly metrically regular at x̄ for 0, then for the neighborhood O
in (i) and any x0 ∈ O the method generates a unique sequence {xk }. This sequence,
according to (i), is linearly convergent to x̄.
If the sequence of numbers λk in (4) is chosen such that limk→∞ λk = 0, the
convergence claimed in (i), (ii) and (iii) is superlinear.
Proof. With Ak (x, u) = λk (x − u) + f (x), in (6)–(8) we can take εk = μk = λk for
all k = 0, 1, . . . . Then, with a particular choice of κ , from (21) or (22) we get (6)
and also (7)(8) for any neighborhood U of x̄. If λk → 0, then (12) holds, implying
superlinear convergence.
Let us go back to Newton’s method in the general form (3) and apply it to the
nonlinear programming problem considered in sections 2A and 2G:

≤ 0 for i ∈ [1, s],
(23) minimize g0 (x) over all x satisfying gi (x)
= 0 for i ∈ [s + 1, m],

where the functions gi : IRn → IR are twice continuously differentiable. In terms of

the Lagrangian function

L(x, y) = g0 (x) + y1 g1 (x) + · · · + ym gm (x),

the first-order optimality (Karush–Kuhn–Tucker) condition takes the form


∇x L(x, y) = 0,
(24) g(x) ∈ N s m−s (y),
IR+ ×IR
334 6 Applications in Numerical Variational Analysis

where we denote by g(x) the vector with components g1 (x), . . . , gm (x). Let x̄ be
a local minimum for (23) satisfying the constraint qualification, and let ȳ be an
associated Lagrange multiplier vector. As applied to the variational inequality (24),
Newton’s method (3) consists in generating a sequence {(xk , yk )} starting from a
point (x0 , y0 ), close enough to (x̄, ȳ), according to the iteration

∇x L(xk , yk ) + ∇2xx L(xk , yk )(xk+1 − xk ) + ∇g(xk )T (yk+1 − yk ) = 0,
(25)
g(xk ) + ∇g(xk )(xk+1 − xk ) ∈ NIRs+ ×IRm−s (yk+1 ).

Theorem 2G.9 (with suppressed dependence on the parameter p in its statement)

provides conditions under which the variational inequality (24) is strongly metri-
cally regular at the reference point: linear independence of the gradients of the active
constraints and a strong form of the second-order sufficient optimality condition; we
recall these conditions below in Example 6C.8. Under these conditions, we can find
(xk+1 , yk+1 ) which satisfies (25) by solving the quadratic programming problem

1
minimize x − xk , ∇2xx L(xk , yk )(x − xk )
2 
T
+∇x L(xk , yk ) − ∇g(xk ) yk , (x − xk )
(26)

≤ 0 for i ∈ [1, s],
subject to gi (xk ) + ∇gi (xk )(x − xk )
= 0 for i ∈ [s + 1, m].

Thus, in the circumstances of (23) under strong metric regularity of the mapping in
(24), Newton’s method (3) comes down to sequentially solving quadratic programs
of the form (26). This specific application of Newton’s method is therefore called
the sequential quadratic programming (SQP) method.
We summarize the conclusions obtained so far about the SQP method as an il-
lustration of the power of the theory developed in this section.
Example 6C.8 (quadratic convergence of SQP). Consider the nonlinear program-
ming problem (23) with the associated Karush–Kuhn–Tucker condition (24) and let
x̄ be a solution with an associated Lagrange multiplier vector ȳ. In the notation
  
I = i ∈ [1, m]  gi (x̄) = 0 ⊃ {s + 1,  . . ., m},
I0 = i ∈ [1, s]  gi (x̄) = 0 and ȳi = 0 ⊂ I

and   
M + =  w ∈ IRn  w ⊥ ∇x gi (x̄) for all i ∈ I\I
 0 ,
M − = w ∈ IRn  w ⊥ ∇x gi (x̄) for all i ∈ I ,
suppose that the following conditions are both fulfilled:
(a) the gradients ∇x gi (x̄) for i ∈ I are linearly independent,
(b) w, ∇2xx L(x̄, ȳ)w > 0 for every nonzero w ∈ M + with ∇2xx L( p̄, x̄, ȳ)w ⊥ M − .
6 Applications in Numerical Variational Analysis 335

Then there exists a neighborhood O of (x̄, ȳ) such that, for any starting point
(x0 , y0 ) ∈ O, the SQP method (26) generates a unique sequence which converges
quadratically to (x̄, ȳ).
There are various numerical issues related to implementation of the SQP method
that have been investigated in the last several decades, and various enhancements
are available as commercial software, but we shall not go into this further.
Lastly, we will discuss a bit more the proximal point method in the context of
monotone mappings. First, note that the iterative process (4) can be equally well
written as

(27) xk+1 ∈ [I + λk−1 T ]−1 (xk ) for i = 1, 2, . . . , where T = f + F.

It has been extensively studied under the additional assumption that X is a Hilbert
space (e.g., consider IRn under the Euclidean norm) and T is a maximal monotone
mapping from X to X . Monotonicity, which we considered in 2F only for functions
from IRn to IRn , refers in the case of a potentially set-valued mapping T to the prop-
erty of having

(28) y − y, x − x ≥ 0 for all (x, y), (x , y ) ∈ gph T.

It is called maximal when no more points can be added to gph T without running
into a violation of (28). (A localized monotonicity for set-valued mappings was
introduced at the end of 3G, but again only in finite dimensions.)
The following fact about maximal monotone mappings, recalled here without its
proof, underlies much of the literature on the proximal point method in basic form
and indicates its fixed-point motivation.

Theorem 6C.9 (resolvents of maximal monotone mappings). Let X be a Hilbert

space, and let T : X →
→ X be maximal monotone. Then for any c > 0 the mapping
Pc = (I + cT )−1 is single-valued with all of X as its domain and moreover is non-
expansive; in other words, it is globally Lipschitz continuous from X into X with
Lipschitz constant 1. The fixed points of Pc are the points x̄ such that T (x̄)  0 (if
any), and they form a closed, convex set.
According to this, xk+1 always exists and is uniquely determined from xk in the
proximal point iterations (4) as expressed in (27), when f + F is maximal monotone.
Here is an important example of that circumstance, which we again state without
bringing out its proof:

Theorem 6C.10 (maximal monotonicity in a variational inequality). Let X be a

Hilbert space, and let F = NC for a nonempty, closed, convex set in C. Let f : C → X
be continuous and monotone. Then f + F is maximal monotone.
The “proximal point” terminology comes out of this framework through an ap-
plication to optimization, as now explained. For a real-valued function g on a Hilbert
space X with derivative Dg(x), we denote by ∇g(x), as in the case of X = IRn , the
unique element of X such that Dg(x)w = ∇g(x), w for all w ∈ X.
336 6 Applications in Numerical Variational Analysis

Example 6C.11 (connections with minimization). Let X be a Hilbert space, let C

be a nonempty, closed, convex subset of X, and let h : X → IR be convex and contin-
uously (Fréchet) differentiable. Let f (x) = ∇h(x). Then f is continuous and mono-
tone, and the variational inequality

f (x) + NC (x)  0,

as an instance of the generalized equation (1), describes the points x (if any) which
minimize h over C. In comparison, in the iterations for this case of the proximal
point method in the basic form (4), the point xk+1 determined from xk is the unique
minimizer of h(x) + (λk /2)||x − xk ||2 over C.
Detail. This invokes the gradient monotonicity property associated with convexity
in 2F.3(a) (which is equally valid in infinite dimensions), along with the optimality
condition in 2A.6. The addition of the quadratic expression (λk /2)||x − xk ||2 to h
creates a function hk which is strongly convex with constant λk and thus attains its
minimum, moreover uniquely.
The expression (λk /2)||x − xk ||2 in 6C.11 is called a proximal term because it
helps to keep x near to the current point xk . Its effect is to stabilize the procedure
while inducing technically desirable properties like strong convexity in place of
plain convexity. It’s from this that the algorithm got its name.
Instead of adding a quadratic term to h, the strategy in Example 6C.11 could be
generalized to adding a term rk (x − xk ) for some other convex function rk having its
minimum at the origin, and adjusting the algorithm accordingly.

Exercise 6C.12. Prove Theorem 6C.1 by using the result stated in Exercise 5E.5.

6D. An Implicit Function Theorem for Newton’s Iteration

In this section we get back to the parameterized generalized equation

(1) f (p, x) + F(x)  0, or equivalently − f (p, x) ∈ F(x),

for a function f : P × X → Y and a (generally set-valued) mapping F, where p ∈ P

is a parameter and P, X and Y are Banach spaces. We assume that the function f
is Fréchet differentiable with respect to x and continuous together with its partial
derivative Dx f (p, x) everywhere, and that the mapping F has closed graph2 . Asso-
ciated with the generalized equation (1) as usual is its solution mapping
  
(2) S : p → x  f (p, x) + F(x)  0 for p ∈ P.
2 Here, as well as in the preceding section, we could of course use local versions of these assump-
tions since our analysis is local.
6 Applications in Numerical Variational Analysis 337

We will focus on the version of Newton’s method treated in the previous section,
the only difference being that now we utilize the partial derivative of the function f
with respect to x:

(3) f (p, xk ) + Dx f (p, xk )(xk+1 − xk ) + F(xk+1 )  0, for k = 0, 1, . . . ,

with a given starting point x0 . We will consider the method (3) more broadly by
reconceiving Newton’s iteration as an inclusion, the solution of which gives a whole
sequence instead of just an element in X. Let l∞ (X) be the Banach space consisting
of all infinite sequences ξ = {x1 , x2 , . . . , xk , . . .} with elements xk ∈ X, k = 1, 2, . . . ,
equipped with the supremum norm

ξ ∞ = sup xk .
k≥1

Define a mapping Ξ : X × P → → l∞ (X) as

 

Ξ : (u, p) → ξ ∈ l∞ (X) 
8
(4) ∞
-
( f (p, xk ) + Dx f (p, xk )(xk+1 − xk ) + F(xk+1 ))  0 with x0 = u ,
k=0

whose value for a given (u, p) is the set of all sequences {xk }∞ k=1 generated by New-
ton’s iteration (3) for p that start from u. If x̄ is a solution to (1) for p̄, then the
constant sequence ξ̄ = {x̄, x̄, . . . , x̄, . . .} satisfies ξ̄ ∈ Ξ (x̄, p̄).
Our first result reveals uniform quadratic convergence under strong metric regu-
larity.

Theorem 6D.1 (uniform convergence of Newton’s iteration). For the generalized

equation (1) with solution mapping S in (2), let x̄ ∈ S( p̄). Let

 ( f ; ( p̄, x̄)) + lip

lip  (Dx f ; ( p̄, x̄)) < ∞,
p x

and let the mapping

(5) G(x) = f ( p̄, x̄) + Dx f ( p̄, x̄)(x − x̄) + F(x) for which G(x̄)  0

be strongly metrically regular at x̄ for 0 with associated Lipschitz continuous single-

valued localization σ around 0 for x̄ of the inverse G−1 . Then for every
1  (Dx f ; ( p̄, x̄))
(6) γ> lip (σ ; 0)·lip x
2
there exist neighborhoods Q of p̄ and U of x̄ such that, for every p ∈ Q and u ∈ U ,
there is exactly one sequence ξ (u, p) with components x1 , . . . , xk , . . . all belonging
to U and generated by Newton’s iteration (3) starting from u for the value p of the
parameter. This sequence converges to the value s(p) of the Lipschitz continuous
localization s of the solution mapping S around p̄ for x̄ whose existence is claimed
338 6 Applications in Numerical Variational Analysis

in Theorem 5F.4. Moreover the convergence is quadratic with constant γ, that is,

(7) xk+1 − s(p) ≤ γ xk − s(p)2 for k = 0, 1, . . . .

Thus, the mapping Ξ in (4) has a single-valued graphical localization ξ around

(x̄, p̄) for ξ̄ . In addition, for u close to x̄ and p close to p̄ the value ξ (u, p) of this
localization is a sequence which converges quadratically to the associated solution
s(p) for p as described in (7).
Proof. Choose γ as in (6) and then κ > lip (σ ; 0) and μ > lip (Dx f ; ( p̄, x̄)) such
x
that κ μ < 2γ . Next, choose ε > 0 so that κε < 1 and furthermore
κμ
(8) ≤ γ.
2(1 − κε )

The assumed strong regularity of the mapping G in (6) at x̄ for 0 and the choice of
κ guarantee the existence of positive constants α  and b such that the mapping y →
σ (y) = G−1 (y) ∩ IBα  (x̄) is a Lipschitz continuous function on IBb (0) with Lipschitz
constant κ . Along with the mapping G consider the parameterized mapping

x → G p,w (x) = f (p, w) + Dx f (p, w)(x − w) + F(x).

Note that G p,w (x) = G(x) + r(p, w; x), where the function

r(p, w; x) = f (p, w) + Dx f (p, w)(x − w) − f ( p̄, x̄) − Dx f ( p̄, x̄)(x − x̄)

is affine, and hence Lipschitz continuous, with Lipschitz constant

η (p, w) = Dx f (p, w) − Dx f ( p̄, x̄).

Now, let κ  be such that κ > κ  > lip (σ ; 0), and let χ > 0 satisfy

κ
χκ  < 1 and < κ.
1 − χκ 

Applying 5F.3, which is a special case of 5F.1, and taking into account that
r( p̄, x̄; x̄) = 0, we obtain the existence of positive constants α ≤ α  and b ≤ b such
that, for p and w satisfying η (p, w) ≤ χ , the mapping y → G−1 p,w (y) ∩ IBα (x̄) is a
Lipschitz continuous function on IBb (0) with Lipschitz constant κ . We denote this
function by Θ (p, w; ·).
Since Dx f is continuous, there are positive constants c and a such that η (p, w) ≤
χ as long as p ∈ IBc ( p̄) and w ∈ IBa (x̄). Make a and c smaller if necessary so that
a ≤ α and moreover

(9) Dx f (p, x) − Dx f (p, x ) ≤ μ x − x  for x, x ∈ IBa (x̄) and p ∈ IBc ( p̄).

By Theorem 5F.4, we can further adjust a and c so that the truncation S(p) ∩ IBa (x̄)
of the solution mapping S in (2) is a function s which is Lipschitz continuous on
6 Applications in Numerical Variational Analysis 339

 ( f ; ( p̄, x̄)). Next, take a even

IBc ( p̄) with Lipschitz constant some λ > lip (σ ; 0)·lip p
smaller if necessary so that
27 2 3 1 9
(10) μ a ≤ b, μa ≤ ε, κ μ a < 1 − κε and γ a ≤ 1.
8 2 2 2
The first and the third inequality in (10) allow us to choose δ > 0 satisfying
1 1
(11) δ + μ a2 ≤ b and κδ + κ μ a2 ≤ a(1 − κε ).
8 2
Then make c even smaller if necessary so that

(12) s(p) − x̄ ≤ a/2 and  f (p, x̄) − f ( p̄, x̄) ≤ δ for p ∈ IBc ( p̄).

Summarizing to this point, we have determined constants a, b and c such that, for
each p ∈ IBc ( p̄) and w ∈ IBa (x̄), the function Θ (p, w, ·) is Lipschitz continuous on
IBb (0) with constant κ , and also, the conditions (9)–(12) are satisfied.
From 6C(19) applied now to the function f (p, x) we have through (9) that, for all
u, v ∈ IBa (x̄) and p ∈ IBc ( p̄),

1
(13)  f (p, u) − f (p, v) − Dx f (p, v)(u − v) ≤ μ u − v2.
2
Fix p ∈ IBc ( p̄) and w ∈ IBa (x̄), and consider the function

x → g(p, w; x) = − f (p, w) − Dx f (p, w)(x − w)

(14)
+ f (p, s(p)) + Dx f (p, s(p))(x − s(p)).

Recall that here s(p) = S(p) ∩ IBa/2 (x̄) for all p ∈ IBc ( p̄). For any x ∈ IBa (x̄), we have
using (9) and (13) that

g(p, w; x) ≤  f (p, s(p)) − f (p, w) − Dx f (p, w)(s(p) − w)

+(Dx f (p, w) − Dx f (p, s(p)))(x − s(p))
≤ 12 μ w − s(p)2 + μ w − s(p)x − s(p) ≤ 27 8 μa .
2

Then, from the first inequality in (10),

(15) g(p, w; x) ≤ b.

By way of (12), (13) and the first inequality in (11), we come to

 f ( p̄, x̄) − f (p, s(p)) − Dx f (p, s(p))(x̄ − s(p))

(16) ≤  f ( p̄, x̄) − f (p, x̄) +  f (p, x̄) − f (p, s(p)) − Dx f (p, s(p))(x̄ − s(p))
≤ δ + 12 μ s(p) − x̄2 ≤ δ + 18 μ a2 ≤ b.

Hence, remembering that p ∈ IBc ( p̄) and s(p) ∈ IBa (x̄), we see that both g(p, w; x)
and f ( p̄, x̄) − f (p, s(p)) − Dx f (p, s(p))(x̄ − s(p)) are in the domain of Θ (p, s(p); ·)
where this function is Lipschitz continuous with Lipschitz constant κ .
340 6 Applications in Numerical Variational Analysis

We now choose p ∈ IBc ( p̄) and u ∈ IBa (x̄), and construct a sequence ξ (u, p) gen-
erated by Newton’s iteration (3) starting from u for the value p of the parameter,
whose existence, uniqueness and quadratic convergence is claimed in the statement
of the theorem.
If u = s(p) there is nothing to prove, so assume u = s(p). Our first step is to show
that, for the function g defined in (14), the mapping

Φ0 : x → Θ (p, s(p); g(p, u; x))

has a unique fixed point in IBa (x̄). Utilizing the equality

x̄ = Θ (p, s(p); − f ( p̄, x̄) + f (p, s(p)) + Dx f (p, s(p))(x̄ − s(p))),

plus (15), (16) and the Lipschitz continuity of Θ (p, s(p); ·) in IBb (0) with constant
κ , and then the second inequality in (12), (13) and the second inequality in (11), we
get

x̄ − Φ0 (x̄)
= Θ (p, s(p); − f ( p̄, x̄) + f (p, s(p)) + Dx f (p, s(p))(x̄ − s(p)))
−Θ (p, s(p); g(p, u; x̄))
≤ κ  − f ( p̄, x̄) + f (p, s(p)) + Dx f (p, s(p))(x̄ − s(p))
(17) −[− f (p, u) − Dx f (p, u)(x̄ − u)
+ f (p, s(p)) + Dx f (p, s(p))(x̄ − s(p))]
= κ  − f ( p̄, x̄) + f (p, u) + Dx f (p, u)(x̄ − u)
≤ κ  − f ( p̄, x̄) + f (p, x̄) + κ  f (p, u) − f (p, x̄) − Dx f (p, u)(u − x̄)
≤ κδ + 12 κ μ u − x̄2 ≤ κδ + 12 κ μ a2 ≤ a(1 − κε ).

Further, for any v, v ∈ IBa (x̄), by (15), the Lipschitz continuity of Θ (p, s(p); ·), (9),
and the second inequality in (10), we obtain

Φ0 (v) − Φ0 (v ) = Θ (p, s(p); g(p, u; v)) − Θ (p, s(p); g(p, u; v ))
(18) ≤ κ g(p, u; v) − g(p, u; v) = κ (−Dx f (p, u) + Dx f (p, s(p)))(v − v )
≤ κ μ u − s(p)v − v ≤ 32 aκ μ v − v ≤ κε v − v.

Hence, by 1A.2, there is a fixed point x1 ∈ Φ0 (x1 ) ∩ IBa (x̄). This translates to
g(p, u; x1 ) ∈ G p,s(p) (x1 ) or, equivalently,

0 ∈ f (p, u) + Dx f (p, u)(x1 − u) + F(x1 ).

This means that x1 is obtained by Newton’s iteration (3) from u for p, and there is
no more than just one such iterate in IBa (x̄).
Now we will demonstrate that x1 satisfies a tighter estimate. Let

ω0 = γ u − s(p)2.

Then ω0 > 0 and, by the last inequality in (10), ω0 ≤ γ (a + a/2)2 ≤ a/2. We apply
again the basic contraction mapping principle 1A.2 to the mapping Φ0 but now on
6 Applications in Numerical Variational Analysis 341

IBω0 (s(p)). Noting that s(p) = Θ (p, s(p); 0) and using (8), (13) and (15), we have

s(p) − Φ0 (s(p)) = Θ (p, s(p); 0) − Θ (p, s(p); g(p, u; s(p)))

(19) ≤ κ g(p, u; s(p)) = κ  − f (p, u) − Dx f (p, u)(s(p) − u) + f (p, s(p))
≤ 12 κ μ u − s(p)2 ≤ γ (1 − κε )u − s(p)2 = ω0 (1 − κε ).

Since IBω0 (s(p)) ⊂ IBa (x̄), we immediately get from (18) that

(20) Φ0 (v) − Φ0 (v ) ≤ κε v − v for any v, v ∈ IBω0 (s(p)).

Thus, the contraction mapping principle applied to the function Φ0 on the ball
IBω1 (s(p)) yields the existence of x1 in this ball such that x1 = Φ0 (x1 ). But the
fixed point x1 of Φ0 in IBω0 (s(p)) must then coincide with the unique fixed point x1
of Φ0 in the larger set IBa (x̄). Hence the fixed point x1 of Φ0 on IBa (x̄) satisfies

x1 − s(p) ≤ γ u − s(p)2 ,

which means that (7) holds for k = 0 with x0 = u.

The induction step is now clear: if the claim holds for k = 1, 2, . . . , n, then
by defining Φn : x → Θ (p, s(p); g(p, xn ; x)) and replacing u by xn in (17) and
(18), we obtain that Φn has a unique fixed point xn+1 in IBa (x̄). This tells us that
g(p, xn ; xn+1 ) ∈ G p,s(p)(xn+1 ) and in consequence that xn+1 is the unique Newton’s
iterate from xn for p which is in IBa (x̄).
Next, by employing again the contraction mapping principle as in (19) and (20)
to Φn , but now on the ball IBωn (s(p)) for ωn = γ xn − s(p)2 , we obtain that xn+1 is
at distance ωn from s(p). Invoking the first inequality in (12) and then the last one
in (10) we have
a
θ := γ x0 − s(p) ≤ γ (x0 − x̄ + s(p) − x̄) ≤ γ (a + ) < 1.
2
Therefore
k −1
(21) xk − s(p) ≤ θ 2 x0 − s(p)

so the sequence {xk } is convergent to s(p) and moreover the convergence is

quadratic as described in (7). This completes the proof of the theorem.
We can go a step further and arrive at an implicit function theorem for Newton’s
iteration which is strikingly similar to the implicit function theorem 5F.4.

Theorem 6D.2 (implicit function theorem for Newton’s iteration). In addition to

the assumptions of Theorem 6D.1, suppose that

lip (Dx f ; ( p̄, x̄)) < ∞.

Then the single-valued localization ξ of the mapping Ξ in (4) around (x̄, p̄) for ξ̄
described in Theorem 6D.1 is Lipschitz continuous near (x̄, p̄), moreover with
342 6 Applications in Numerical Variational Analysis

(22)  (ξ ; (x̄, p̄)) = 0

lip and lip  ( f ; ( p̄, x̄)).
 (ξ ; (x̄, p̄)) ≤ lip (σ ; 0)·lip
u p p

Proof. First, recall some notation and facts established in Theorem 6D.1 and its
proof. We know that for any κ > lip (σ ; 0), there exist positive constants a, α , b
and c such that a ≤ α and, for every p ∈ IBc ( p̄) and w ∈ IBa (x̄), the mapping y →
G−1
p,w (y) ∩ IBα (x̄) is a function, with values Θ (p, w; y), which is Lipschitz continuous
on IBb (0) with Lipschitz constant κ ; moreover, the truncation S(p) ∩ IBa (x̄) of the
solution mapping in (2) is a Lipschitz continuous function on IBc ( p̄) and its values
are in IBa/2 (x̄); also, for any starting point u ∈ IBa (x̄) and any p ∈ IBc ( p̄), there is a
unique sequence ξ (u, p) starting from u and generated by Newton’s method (3) for
p whose components are contained in IBa (x̄), with this sequence being quadratically
convergent to s(p) as described in (7).
Our starting observation is that, for any positive a ≤ a, by adjusting the size of
the constant c and taking as a starting point u ∈ IBa (x̄), we can arrange that, for any
p ∈ IBc ( p̄), all elements xk of the sequence ξ (u, p) are actually in IBa (x̄). Indeed, by
taking δ > 0 to satisfy (11) with a replaced by a and then choosing c so that (12)
holds for the new δ and for a , then all requirements for a will hold for a as well
and hence all Newton’s iterates xk will be at distance a from x̄.
Choose
η > lip (Dx f ; ( p̄, x̄)) and ν > lip ( f ; ( p̄, x̄)).
p

Pick a positive constant d ≤ a/2 and make c smaller if necessary, so that for every
p, p ∈ IBc ( p̄) and every w, w ∈ IBd (x̄), we have

(23) Dx f (p , w ) − Dx f (p, w) ≤ η (p − p + w − w),

(24)  f (p , w) − f (p, w) ≤ ν p − p,

and, in addition, for every x ∈ IBd (x̄), every p, p ∈ IBc ( p̄) and every w, w ∈ IBd (x̄),
we have

(25)  f (p , w ) + Dx f (p , w )(x − w ) − f (p, w) − Dx f (p, w)(x − w) ≤ b.

Choose a positive τ such that τκ < 1/3. Make d and c smaller if necessary so that

(26) 3η (d + c) < τ .

Since
κτ 1
<
1 − κτ 2
we can take c still smaller in order to have
κτ (2d) + 3κ (τ + ν )(2c)
(27) ≤ d.
1 − κτ
Let p, p ∈ IBc ( p̄), u, u ∈ IBd (x̄), (p, u) = (p , u ). In accordance with Theorem
6D.1 and the observation above, let ξ (p, u) = (x1 , . . . , xk , . . .) be the unique sequence
6 Applications in Numerical Variational Analysis 343

generated by Newton’s iteration (3) starting from u whose components xk are all in
IBd (x̄) and hence in IBa/2 (x̄). For this sequence, denoting x0 = u, we know that for
all k ≥ 0

(28) xk+1 = Θ (p, xk ; 0) := ( f (p, xk ) + Dx f (p, xk )(· − xk ) + F(·))−1 (0) ∩ IBα (x̄).

Let
κτ u − u + κ (τ + ν )p − p
γ0 = .
1 − κτ
By using (27) we get that γ0 ≤ d and then IBγ0 (x1 ) ⊂ IBa (x̄). Consider the function

Φ0 : x → Θ (p, u; − f (p , u ) − Dx f (p , u )(x − u ) + f (p, u) + Dx f (p, u)(x − u)).

Employing (25) and then the Lipschitz continuity of Θ (p, u; ·) on IBb (0), and apply-
ing (13), (23), (24), (26) and (28), we obtain
(29)
x1 − Φ0 (x1 ) = Θ (p, u; 0)−
Θ (p, u; − f (p , u ) − Dx f (p , u )(x1 − u) + f (p, u) + Dx f (p, u)(x1 − u))
≤ κ  f (p , u) − f (p , u ) − Dx f (p , u )(u − u)
−Dx f (p , u )(x1 − u) − f (p, u) + f (p, u) + Dx f (p, u)(x1 − u)
≤ κ  f (p , u) − f (p , u ) − Dx f (p , u )(u − u)
+κ (Dx f (p, u) − Dx f (p , u ))(x1 − u) + κ  − f (p, u) + f (p, u)
≤ 12 κη u − u2 + κη u − ux1 − u
+κη p − px1 − u + κν p − p
≤ 3κη du − u + κ (2η d + ν )p − p
≤ κτ u − u + κ (τ + ν )p − p = γ0 (1 − κτ ).

For v, v ∈ IBγ0 (x1 ), we have by way of (23), (24) and (26) that

Φ0 (v) − Φ0 (v ) ≤ κ (−Dx f (p , u ) + Dx f (p, u))(v − v )

(30)
≤ 2κη (d + c)v − v ≤ κτ v − v.

Hence, by the contraction mapping principle 1A.2, there is a unique x1 in IBγ0 (x1 )
such that

x1 = Θ (p, u; − f (p , u ) − Dx f (p , u )(x1 − u ) + f (p, u) + Dx f (p, u)(x1 − u)).

But then
f (p , u ) + D f (p , u )(x1 − u ) + F(x1 )  0,
that is, x1 is the unique Newton’s iterate from u for p which satisfies

x1 − x1  ≤ γ0 .

Since γ0 ≤ d, we obtain that x1 ∈ IBa (x̄) and then x1 is the unique Newton’s iteration
from u for p which is in IBa (x̄).
344 6 Applications in Numerical Variational Analysis

By induction, we construct a sequence ξ  = {x1 , x2 , . . . , xk , . . .} ∈ Ξ (p , u ) such

that the distance from xk to the corresponding component xk of ξ satisfies the esti-
mate
κτ xk−1 − xk−1  + κ (τ + ν )p − p
(31) xk − xk  ≤ γk−1 := for k = 2, 3, . . . .
1 − κτ
Suppose that for some n > 1 we have found x2 , x3 , . . . , xn with this property. First,
observe that
  k  
κτ k+1 κ (τ + ν ) κτ i
γk ≤ 
u − u  + p − p  ∑ 
,
1 − κτ 1 − κτ i=0 1 − κτ

from which we get the estimate that, for all k = 0, 1, . . . , n − 1,

κτ κ (τ + ν )
(32) γk ≤ u − u + p − p.
1 − κτ 1 − 2κτ
In particular, we obtain through (27) that γk ≤ d for all k, and consequently xk ∈
IBd (xk ) ⊂ IBa (x̄).
To show that xn+1 is a Newton’s iterate from xn for p , we proceed in the same
way as in obtaining x1 from u for p . Consider the function

Φk : x → Θ (p, xk ; − f (p , xk ) − Dx f (p , xk )(x − xk ) + f (p, xk ) + Dx f (p, xk )(x − xk )).

By replacing Φ0 by Φk , u by xk , u by xk , and x1 by xk+1 in (29) and (30), we get

xk+1 − Φk (xk+1 ) ≤ κτ xk − xk  + κ (τ + ν )p − p = γk (1 − κτ )

and
Φk (v) − Φk (v ) ≤ κτ v − v for any v, v ∈ IBγk (xk+1 ).
Then, by the contraction mapping principle 1A.2 there is a unique xk+1 in IBγk (xk+1 )
with xk+1 = Φk (xk+1 ), which gives us

f (p , xk ) + Dx f (p , xk )(xk+1 − xk ) + F(xk+1 )  0.

Moreover, since γk ≤ d, we have that xk+1 ∈ IBa (x̄).

We have constructed a sequence x1 , . . . , xk , . . ., generated by Newton’s iteration
for p starting from u , and whose components are in IBa (x̄). According to Theorem
6D.1, this sequence must be the value ξ (u , p ) of the single-valued localization ξ
whose value ξ (u, p) is the sequence x1 , . . . , xk , . . . . Taking into account (31) and
(32), we come to the estimate

ξ (u, p) − ξ (u, p )∞ ≤ O(τ )u − u + (κν + O(τ ))p − p.

Since τ can be chosen arbitrarily small, this yields (22).

6 Applications in Numerical Variational Analysis 345

As in the case of the classical implicit function theorem, the inverse function
version of Theorem 6D.2 turns into an “if and only if” result.
Consider the generalized equation (1) with f (p, x) = g(x) − p whose solution
mapping S = (g + F)−1 , and let x̄ ∈ S(0). In order to apply 6D.2 suppose that g
is differentiable near x̄ with lip (Dg; x̄) < ∞. The corresponding Newton’s iteration
mapping in (3) then has the form
 

ϒ : (u, p) → ξ ∈ l∞ (X) 
8
(33) ∞
-
(g(xk ) + Dg(xk )(xk+1 − xk ) + F(xk+1 ))  p with x0 = u .
k=0

Theorem 6D.3 (inverse function theorem for Newton’s iteration). The mapping g +
F is strongly regular at x̄ for 0 if and only if the mapping ϒ in (33) has a Lipschitz
continuous single-valued localization ξ around (x̄, 0) for ξ̄ with

(34)  (ξ ; (x̄, 0)) < 1

lip u

and is such that, for each (u, p) close to (x̄, 0), the sequence ξ (u, p) is convergent.
Moreover, in this case

(35)  (ξ ; (x̄, 0)) = reg (g + F; x̄|0).

lip p

Proof. The “only if” part follows from the combination of 6D.1 and 6D.2. Noting
that the Lipschitz modulus of the single-valued localization σ in 6D.1 equals the
regularity modulus of g + F, from (22) we get

(36)  (ξ ; (x̄, 0)) ≤ lip (σ ; 0) = reg (g + F; x̄|0).

lip p

 (ξ ; (x̄, 0)), a positive ε < 1 and corresponding

To prove the “if” part, choose κ > lip p
neighborhoods U of x̄ and Q of 0 such that the sequence ξ (u, p) is the only element
of ϒ (u, p) whose components x1 , . . . , xk , . . . are in U, and moreover the function
ξ acting from X × Y to l∞ (X) is Lipschitz continuous with Lipschitz constants κ
in p ∈ Q uniformly in u ∈ U and still more, from (34), ξ is Lipschitz continuous
with Lipschitz constants ε in u ∈ U uniformly in p ∈ Q. From the assumed local
closedness of gph F it is possible to make Q and U smaller if necessary so that for
any p ∈ Q and any sequence with components vk ∈ U convergent to v and satisfying

(37) g(vk ) + Dg(vk )(vk+1 − vk ) + F(vk+1 )  p for all k = 1, 2, . . . ,

one has g(v) + F(v)  p.

Let p, p ∈ Q and let x ∈ (g + F)−1 (p) ∩ U. The constant sequence all elements
of which are equal x, namely, χ = (x, x, . . . , x, . . .), is obviously convergent to the
solution x of the inclusion g(x) + F(x)  p. Then χ ∈ ϒ (x, p) and all its components
are in U, hence χ = ξ (x, p). By assumption,
346 6 Applications in Numerical Variational Analysis

(38) χ − ξ (x, p )∞ ≤ κ p − p,

and moreover ξ (x, p ) = {x1 , . . . , xk , . . .} is convergent. By definition, ξ (x, p ) satis-

fies
g(xk ) + Dg(xk )(xk+1 − xk ) + F(xk+1 )  p for all k = 1, 2, . . . .
From the property described in (37) we obtain that the sequence ξ (x, p ) is conver-
gent to a solution x ∈ (g + F)−1 (p ) ∩U. Hence, using (38), we have

x − x ≤ x − xk  + xk − x 

≤ χ − ξ (x, p )∞ + xk − x  ≤ κ p − p + xk − x .

Since xk → x as k → ∞, we conclude by passing to the limit in this last inequality

that

(39) x − x ≤ κ p − p.

This means that the mapping g + F is metrically regular at x̄ for 0. We will demon-
strate that the mapping (g + F)−1 has a single-valued localization around 0 for
x̄. We know that dom(g + F)−1 contains a neighborhood of 0. Assume that for
any neighborhoods U of x̄ and Q of 0 there exist p ∈ Q and w, w ∈ U such that
w = w and both w and w are in (g + F)−1 (p). Then the constant sequences
{w, w, . . . , w, . . .} ∈ ϒ (w, p) and {w , w , . . . , w , . . .} ∈ ϒ (w , p) and all their compo-
nents are in U, hence {w, w, . . . , w, . . .} = ξ (w, p) and {w , w , . . . , w , . . .} = ξ (w , p).
In the beginning of the proof we have chosen the neighborhoods U and Q such that
for a fixed p ∈ V the mapping u → ξ (u, p) is a Lipschitz continuous function from
X to l∞ (X) with Lipschitz constant ε < 1, and hence this condition holds for all of
its components. This yields

w − w ≤ ε w − w  < w − w ,

which is absurd. Hence, (g + F)−1 has a single-valued localization s around 0 for x̄.
But then from (39) this localization is Lipschitz continuous around 0 with lip (s; 0) ≤
κ . The Banach space versions of Theorems 2B.10 and 3G.1 say that lip (σ ; 0) =
lip (s; 0) = reg (g + F; x̄|0) and hence lip (σ ; 0) ≤ κ . Since κ could be arbitrarily
 p (ξ ; (x̄, 0)), we get the inequality opposite to (36), and hence the equality
close to lip
(35) holds.
As an illustration of possible applications of the results in Theorems 6D.1 and
6D.2 in studying complexity of Newton’s iteration, we will produce an estimate
for the number of iterations needed to achieve a particular accuracy of the method,
which is the same for all values of the parameter p in some neighborhood of the
reference point p̄. Given an accuracy measure ρ , suppose that Newton’s method (3)
is to be terminated at the k-th step if

(40) d(0, f (p, xk ) + F(xk )) ≤ ρ .

6 Applications in Numerical Variational Analysis 347

Also suppose that the constant μ and the constants a and c are chosen to satisfy
(9). For p ∈ IBc ( p̄) consider the unique sequence {xk } generated by (3) for p, all
elements of which are in IBa (x̄). Since xk is a Newton’s iterate from xk−1 , we have
that

f (p, xk ) − f (p, xk−1 ) − Dx f (p, xk−1 )(xk − xk−1 ) ∈ f (p, xk ) + F(xk ).

Using (13), we have

d(0, f (p, xk ) + F(xk ))

(41) ≤  f (p, xk ) − f (p, xk−1 ) − Dx f (p, xk−1 )(xk − xk−1 )
≤ 12 μ xk − xk−1 2 .

Let kρ be the first iteration at which (40) holds; then for k < kρ from (41) we obtain

1
(42) ρ < μ xk − xk−1 2 .
2
Further, utilizing (21) we get
k −2
xk − xk−1  ≤ xk − s(p) + xk−1 − s(p) ≤ θ 2 (1 + θ )(x0 − x̄ + s(p) − x̄),

and from the choice of x0 and the first inequality in (12) we have
k −2 3a
xk − xk−1  ≤ θ 2 (1 + θ ) .
2
But then, taking into account (42), we obtain

k+1 9a (1 + θ )
1 2 2
ρ < μθ 2 .
2 4θ 4
Therefore kρ satisfies
  
8θ 4 ρ
kρ ≤ log2 logθ − 1.
9a μ (1 + θ )2
2

Thus, we have obtained an upper bound of the number of iterations needed to

achieve a particular accuracy, which, most importantly, is the same for all values
of the parameter p in some neighborhood of the reference value p̄. This tells us, for
example, that, under the assumptions of Theorem 6D.2, small changes of parame-
ters in a problem don’t affect the performance of Newton’s method as applied to this
problem.

Exercise 6D.4 (using ample parameterization). In addition to the conditions in

6D.2, let f be strictly differentiable at ( p̄, x̄) and the derivative D p f ( p̄, x̄) be surjec-
tive. Show that the metric regularity of the mapping G in (5) is not only sufficient,
348 6 Applications in Numerical Variational Analysis

but also necessary for the existence of a single-valued localization of the mapping
Ξ in (4) whose values are convergent, as in the statement of 6D.1.

6E. Galerkin’s Method for Quadratic Minimization

The topic of this section is likewise a traditional scheme in numerical analysis and
its properties of convergence, again placed in a broader setting than the classical
one. The problem at which this scheme will be directed is quadratic optimization in
a Hilbert space setting:
1
(1) minimize x, Ax − v, x over x ∈ C,
2
where C is a nonempty, closed and convex set in a Hilbert space X, and v ∈ X
is a parameter.
 Here ·, · denotes the inner product in X ; the associated norm is
x = x, x. We take A : X → X to be a linear and bounded mapping, entailing
dom A = X ; furthermore, we take A to be self-adjoint, x, Ay = y, Ax for all x, y ∈ X
and require that

(2) x, Ax ≥ μ x2 for all x ∈ C − C, for a constant μ > 0.

This property of A, sometimes called coercivity (a term which can have conflict-
ing manifestations), corresponds to A being strongly monotone relative to C in the
sense defined in 2F, as well as to the quadratic function in (1) being strongly convex
relative to C. For X = IRn , (2) is equivalent to positive definiteness of A relative to
the subspace generated by C −C. For any Hilbert space X in which that subspace is
dense, it entails A being invertible with A−1  ≤ μ −1 .
In the usual framework for Galerkin’s method, C would be all of X , so the tar-
geted problem would be unconstrained. The idea is to consider an increasing se-
quence of finite-dimensional subspaces Xk of X, and by iteratively minimizing over
Xk , to get a solution point x̂k , generate a sequence which, in the limit, solves the
problem for X .
This approach has proven valuable in circumstances where X is a standard func-
tion space and the special functions making up the subspaces Xk are familiar tools
of approximation, such as trigonometric expansions. Here, we will work more gen-
erally with convex sets Ck furnishing “inner approximations” to C, with the eventual
possibility of taking Ck = C ∩ Xk for a subspace Xk .
In Section 2G with X = IRn , we looked at a problem like (1) in which the function
was not necessarily quadratic, and we studied the dependence of its solution on the
parameter v. Before proceeding with anything else, we must update to our Hilbert
space context with a quadratic function the particular facts from that development
which will be called upon.
6 Applications in Numerical Variational Analysis 349

Theorem 6E.1 (optimality and its characterization). For problem (1) under condi-
tion (2), there exists for each v a unique solution x. The solution mapping S : v → x
is thus single-valued with dom S = X . Moreover, this mapping S is Lipschitz con-
tinuous with constant μ −1 , and it is characterized by a variational inequality:

(3) x = S(v) ⇐⇒ − v + Ax + NC(x)  0.

Proof. The existence of a solution x for a fixed v comes from the fact that, for each
sufficiently large α ∈ IR the set Cα of x ∈ C for which the function being minimized
in (1) has value ≤ α is nonempty, convex, closed and bounded, with the bound
coming from (2). Such a subset of X is weakly compact; the intersection of the
Cα that are nonempty is therefore nonempty. That intersection is comprised of all
possible solutions x. The uniqueness of such x follows however from the strong
convexity of the function in question. The characterization of x in (3) is proved
exactly as in the case of X = IRn in 2A.6. The Lipschitz property of S comes out of
the same argument that was used in the second half of the proof of 2F.9, utilizing
the strong monotonicity of A.
As an important consequence of Theorem 6E.1, we get a Hilbert space version
of the projection result in 1D.5 for convex sets in IRn .

Corollary 6E.2 (projections onto convex sets). For a nonempty, closed, convex set
C in a Hilbert space X , there exists for each v ∈ X a unique nearest point x of
C, called the projection of v on C and denoted by PC (v). The projection mapping
PC : X → C is Lipschitz continuous with constant 1.
Proof. Take A = I in (1), noting that (2) holds then with μ = 1. Problem (1) is
equivalent then to minimizing ||x − v|| over x ∈ C, because the expression being
minimized differs from 12 ||x − v||2 only by the constant term 12 ||v||2 .
In Galerkin’s method, when we get to it, there will be need of comparing solu-
tions to (1) with solutions to other problems for the same v but sets different from
C. In effect, we have to be able to handle the choice of C as another sort of param-
eter. For a start, consider just two different sets, D1 and D2 . How might solutions
to the versions of (1) with D1 and D2 in place of C, but with fixed v, relate to each
other? To get anywhere with this we require a joint strong monotonicity condition
extending (2):

(4) x, Ax ≥ μ x2 for all x ∈ Di − D j and i, j ∈ {1, 2}, i = j, where μ > 0.

Obviously (4) holds without any fuss over different sets if we simply have A strongly
monotone with constant μ on all of X.

Proposition 6E.3 (solution estimation for varying sets). Consider any nonempty,
closed, convex sets D1 and D2 in X satisfying (4). If x1 and x2 are the solutions of
problem (1) with constraint sets D1 and D2 , respectively, in place of C, then

(5) μ x1 − x2 2 ≤ Ax1 − v, u1 − x2  + Ax2 − v, u2 − x1  for all u1 ∈ D1 , u2 ∈ D2 .

350 6 Applications in Numerical Variational Analysis

Proof. From (4) we have

(6) μ x1 − x2 2 ≤ A(x1 − x2 ), x1 − x2,

whereas for any u1 ∈ D1 and u2 ∈ D2 , (3) gives us

(7) 0 ≤ Ax1 − v, u1 − x1 , 0 ≤ Ax2 − v, u2 − x2 .

Adding the inequalities in (7) to the one in (6) and rearranging the sum, we obtain

μ x1 − x22 ≤ A(x1 − x2 ), x1 − x2  + Ax1 − v, u1 − x1  + Ax2 − v, u2 − x2 

= Ax1 − v, u1 − x2  + Ax2 − v, u2 − x1 ,

as claimed in (5).
Having this background at our disposal, we are ready to make progress with our
generalized version of Galerkin’s method. We consider along with C a sequence of
sets Ck ⊂ X for k = 1, 2, . . . which, like C, are nonempty, closed and convex. We
suppose that

(8) Ck ⊂ Ck+1 ⊂ · · · ⊂ C, with cl [C1 ∪C2 ∪ · · ·] = C,

and let

(9) Sk = the solution mapping for (1) with Ck in place of C,

as provided by Theorem 6E.1 through the observation that (2) carries over to any
subset of C. By generalized Galerkin’s sequence associated with (8) for a given v,
we will mean the sequence of solutions x̂k = Sk (v), k = 1, 2, . . . .

Theorem 6E.4 (general rate of convergence). Let S be the solution mapping to

(1) as provided by Theorem 6E.1 under condition (2), and let {Ck } be a sequence
of nonempty, closed, convex sets satisfying (8). Then for any v the associated
Galerkin’s sequence {x̂k } = {Sk (v)} converges to x̂ = S(v). In fact, there is a con-
stant c such that

(10) x̂k − x̂ ≤ c d(x̂,Ck )1/2 for all k.

Proof. On the basis of (8), we have dist(x̂,Ck ) → 0. The sequence of projections

x̄k = PCk (x̂) with ||x̄k − x̂|| = d(x̂,Ck ), whose existence is guaranteed by 6E.2, con-
verges then to x̂. From 6E.3 applied to D1 = Ck and D2 = C, with x̂k and x̂ in the
place of the x1 and x2 there, and on the other hand u1 = x̄k and u2 = x̂k , we get
μ x̂k − x̂2 ≤ Ax̂k − v, x̄k − x̂ and therefore

μ x̂k − x̂2 ≤ A(x̂k − x̂) + Ax̂ − v, x̄k − x̂

(11)
≤ (Ax̂k − x̂ + Ax̂ + v) dist(x̂,Ck ).
6 Applications in Numerical Variational Analysis 351

This quadratic inequality in dk = x̂k − x̂ implies that the sequence {dk } is bounded,
say by b. Putting this b in place of x̂k − x̂ on the right side of (11), we get a bound
of the form in (10).
Is the square root describing the rate of convergence through the estimate in (10)
exact? The following example shows that this is indeed the case, and no improve-
ment is possible, in general.
Example 6E.5 (counterexample to improving
  the general
 estimate). Consider prob-
lem (1) in the case of X = IR2 , C = (x1 , x2 )  x2 ≤ 0 (lower half-plane), v = (0, 1)
and A = I, so that the issue revolves around projecting v on C and the solution is
x̂ = (0, 0). For each k = 1, 2, . . . let ak = (1/k, 0) and let Ck consist of the points
x ∈ C such that x − ak , v − ak  ≤ 0. Then the projection x̂k of v on Ck is ak , and

1
|x̂k − x̂| = 1/k, d(x̂,Ck ) = √ .
k 1 + k2
In this case the ratio |x̂k − x̂|/d(x̂,Ck ) p is unbounded in k for any p > 1/2.
Detail. The fact that the projection of v on Ck is ak comes from the observation
that v − ak ∈ NCk (ak ). A similar observation confirms that the specified x̄k is the
projection of x̂ on Ck . The ratio |x̂k − x̂|/d(x̂,Ck ) p can be calculated as k2p−1 (1 +
1/(k2 ) p/2 , and from that the conclusion is clear that it is bounded with respect to k
if and only if 2 − (1/p) ≤ 0, or in other words, p ≤ 1/2.

(0,1)

x^ x^k

Ck

Fig. 6.1 Illustration to Example 6E.5.

There is, nevertheless, an important case in which the exponent 1/2 in (10) can
be replaced by 1. This case is featured in the following result:

Theorem 6E.6 (improved rate of convergence for subspaces). Under the conditions
of Theorem 6E.4, if the sets C and Ck are subspaces of X , then there is a constant c
such that

(12) x̂k − x̂ ≤ cd(x̂,Ck ) for all k.

352 6 Applications in Numerical Variational Analysis

Proof. In this situation the variational inequality in (3) reduces to the requirement
that Ax − v ⊥ C. We then have Ax̂ − v ∈ C⊥ ⊂ Ck⊥ and Ax̂k − v ∈ Ck⊥ , so that A(x̂k −
x̂) ∈ Ck⊥ . Consider now an arbitrary x ∈ Ck , noting that since x̂k ∈ Ck we also have
x̂k − x ∈ Ck . We calculate from (2) that

μ x̂k − x̂2 ≤ A(x̂k − x̂), x̂k − x̂

= A(x̂k − x̂), x̂k − x + A(x̂k − x̂), x − x̂
= A(x̂k − x̂), x − x̂ ≤ Ax̂k − x̂x − x̂.

This gives us the estimate (12).

The result in 6E.6 corresponds to the classical Galerkin method, at least if C is
all of X . We can combine it with the one in 6E.4 as follows.

Corollary 6E.7 (application to intersections with subspaces). Let S be the solution

mapping to (1) as provided by Theorem 6E.1 under condition (2). Let {Xk } be an
increasing sequence of closed subspaces of X such that (8) holds for the sets Ck =
C ∩ Xk . Then for any v the associated Galerkin’s sequence {x̂k } = {Sk (v)} converges
to x̂ = S(v) at the rate indicated in (10), but if C itself is a subspace, it converges at
the rate indicated in (12).
The closure condition in (8), in the case of Ck = C ∩Xk , says that dist(x,C ∩Xk ) →
0 as k → ∞ for every x ∈ C. When the Hilbert space X is separable we may choose the
subspaces Xk by taking a countable dense subset x1 , x2 , . . . of X and letting Xk be the
span of x1 , . . . , xk . Because the subspaces are finite-dimensional, Galerkin’s method
in this case can be viewed as a discretization scheme. The property that dist(x,C ∩
Xk ) → 0 as k → ∞ for every x ∈ C is called the consistency of the discretization
scheme. In the following section we will look at the discretization of a specific
variational problem.

6F. Approximations in Optimal Control

For an example which illustrates how the theory of solution mappings can be applied
in infinite dimensions with an eye toward numerical approximations, we turn to a
basic problem in optimal control, the so-called linear-quadratic regulator problem.
That problem takes the form:
 1 
1 T T T T
(1) minimize [x(t) Qx(t) + u(t) Ru(t)] + s(t) x(t) − r(t) u(t) dt
0 2

subject to

(2) ẋ(t) = Ax(t) + Bu(t) + p(t) for a.e. t ∈ [0, 1], x(0) = a,
6 Applications in Numerical Variational Analysis 353

and the constraint that

(3) u(t) ∈ U for a.e. t ∈ [0, 1].

This concerns the control system governed by (2) in which x(t) ∈ IRn is the state at
time t and u(t) is the control exercised at time t. The choice of the control function
u : [0, 1] → IRm yields from the initial state a ∈ IRn and the dynamical equation in (2)
a corresponding state trajectory x : [0, 1] → IRn with derivative ẋ. The matrices A, B,
Q and R have dimensions fitting these circumstances, with Q and R symmetric and
positive semidefinite so as to ensure (as will be seen below) that the function being
minimized in (1) is convex.
The set U ⊂ IRm from which the values of the control have to be selected from in
(3) is nonempty, convex and compact3 . We also assume that the matrix R is positive
definite relative to U; in other words, there exists μ > 0 such that

(4) uT Ru ≥ μ |u|2 for all u ∈ U − U.

Any control function u : [0, 1] → U is required to be measurable (“a.e.” refers as

usual to “almost everywhere” with respect to Lebesgue measure), and since it takes
values in the bounded set U for a.e. t ∈ [0, 1], it is essentially bounded. But we take
the set of feasible control functions to be a larger subset of L2 (IRm , [0, 1]) functions,
in which space the inner product and the norm are
 1 
u, v = u(t)T v(t) dt, u2 = u, u.
0

We follow that Hilbert space pattern throughout, assuming that the function r in (1)
belongs to L2 (IRm , [0, 1]) while p and s belong to L2 (IRn , [0, 1]). This is a convenient
compromise which will put us in the framework of quadratic optimization in 6E.
There are two ways of looking at problem (1). We can think of it in terms of min-
imizing over function pairs (u, x) constrained by both (2) and (3), or we can regard
x as a “dependent variable” produced from u through (2) and standard facts about
differential equations, so as to think of the minimization revolving only around the
choice of u. For any u satisfying (3) (and therefore essentially bounded), there is
a unique state trajectory x specified by (2) in the sense of x being an absolutely
continuous function of t and therefore differentiable a.e. Due to the assumption
that p ∈ L2 (IRn , [0, 1]), the derivative ẋ can then be interpreted as an element of
L2 (IRn , [0, 1]) as well. Indeed, x is given by the Cauchy formula
t
(5) x(t) = e a +
At
eA(t−τ ) (Bu(τ ) + p(τ ))d τ for all t ∈ [0, 1].
0

In particular, we can view it as belonging to the Banach space C(IRn , [0, 1]) of con-
tinuous functions from [0, 1] to IRn equipped with the norm

3 We do not really need U to be bounded, but this assumption simplifies the analysis.
354 6 Applications in Numerical Variational Analysis

x∞ = max |x(t)|.

0≤t≤1

The relation between u and x can be cast in a frame of inputs and outputs. Define
the mapping T : L2 (IRn , [0, 1]) → L2 (IRn , [0, 1]) as
 t
(6) (Tw)(t) = eA(t−τ ) w(τ )d τ for a.e. t ∈ [0, 1],
0

and, on the other hand, let W : L2 (IRn , [0, 1]) → L2 (IRn , [0, 1]) be the mapping defined
by

(7) for p ∈ L2 (IRn , [0, 1]), W (p) is the solution to Ẇ = AW + p, W (0) = a.

Finally, with a slight abuse of notation, denote by B the mapping from L2 (IRm , [0, 1])
to L2 (IRn , [0, 1]) associated with the matrix B, that is (Bu)(t) = Bu(t); later we do
the same for the mappings Q and R. Then the formula for x in (5) comes out as

(8) x = (T B)(u) + W (p),

where u is the input, x is the output, and p is a parameter. Note that in this case we
are treating x as an element of L2 (IRn , [0, 1]) instead of C(IRn , [0, 1]). This makes no
real difference but will aid in the analysis.

Exercise 6F.1 (adjoint in the Cauchy formula). Prove that the mapping T defined by
(6) is linear and bounded. Also show that the adjoint (dual) mapping T ∗ , satisfying
x, Tu = T ∗ x, u, is given by
 1
T (τ −t)
(T ∗ x)(t) = eA x(τ )d τ for a.e. t ∈ [0, 1].
t

Also show (T B)∗ = B∗ T ∗ , where B∗ is the mapping L2 (IRn , [0, 1]) to L2 (IRm , [0, 1])
associated with the transposed matrix BT ; that is
 1
∗ T (τ −t)
((T B) x)(t) = BT eA x(τ )d τ for a.e. t ∈ [0, 1].
t

Guide. Apply the rule for changing the order of integration

 1  t  1 1
x(t)T eA(t−τ ) w(τ )d τ dt = x(t)T eA(t−τ ) w(τ )dtd τ ,
0 0 0 τ

and interpret what it says.

To shorten notation in what follows, we will just write L2 for both L2 (IRm , [0, 1])
and L2 (IRn , [0, 1]), leaving it to the reader to keep in mind which elements lie in IRm
and which lie in IRn .
The change of variables z = x − w with w = W (p) as in (7) gives the following
reformulation of (1)–(3), where the parameter p is transferred to the problem of
6 Applications in Numerical Variational Analysis 355

minimizing the objective function

 1 
1 T T T T
(9) [z(t) Qz(t) + u(t) Ru(t)] + (s(t) + QW(p)(t)) z(t) − r(t) u(t) dt
0 2

subject to

(10) ż = Az + Bu, z(0) = 0, u(t) ∈ U for a.e. t ∈ [0, 1].

In (9) we have dropped the constant terms that do not affect the solution. Noting
that z = (T B)(u) and utilizing the adjoint T ∗ of the mapping T , let

(11) V (y) = −r + B∗T ∗ (s + QW (p)) for y = (p, s, r),

and define the self-adjoint bounded linear mapping A : L2 → L2 by

(12) A = B∗ T ∗ QT B + R.

Here, as for the mapping B, we regard Q and R as linear bounded mappings acting
between L2 spaces: for (Ru)(t) = Ru(t), and so forth. Let
  
(13) C = u ∈ L2  u(t) ∈ U for a.e. t ∈ [0, 1] .

With this notation, problem (9)–(10) can be written in the form treated in 6E:
1
(14) minimize u, A u + V (y), u subject to u ∈ C.
2

Exercise 6F.2 (coercivity in control). Prove that the set C in (13) is a closed and
convex subset of L2 and that the mapping A ∈ L (L2 , L2 ) in (12) satisfies the con-
dition
u, A u ≥ μ u22 for all u ∈ C − C,
where μ is the constant in (4).
Applying Theorem 6E.1 in the presence of 6F.2, we obtain a necessary and suf-
ficient condition for the optimality of u in problem (14), namely the variational
inequality

(15) V (y) + A u + NC (u)  0.

For (15), or equivalently for (14) or (1)–(3), we arrive then at the following result of
implicit-function type:

Theorem 6F.3 (implicit function theorem for optimal control in L2 ). Under (4), the
solution mapping S which goes from parameter elements y = (p, s, r) to pairs (u, x)
solving (1)–(3) is single-valued and globally Lipschitz continuous from the space
L2 (IRn × IRn × IRm , [0, 1]) to the space L2 (IRm , [0, 1]) × C(IRn , [0, 1]).
356 6 Applications in Numerical Variational Analysis

Proof. Because V in (11) is an affine function of y = (p, s, r), we obtain from 6E.1
that for each y problem (14) has a unique solution u(y) and, moreover, the function
y → u(y) is globally Lipschitz continuous in the respective norms. The value u(y)
is the unique optimal control in problem (1)–(3) for y. Taking norms in the Cauchy
formula (5), we see further that for any y = (p, s, r) and y = (p , s , r ), if x and x are
the corresponding solutions of (2) for u(y), p, and u(y ), p , then, for some constants
c1 and c2 , we get
 t
|x(t) − x (t)| ≤ c1 (|B||u(y)(τ ) − u(y )(τ )| + |p(τ ) − p (τ )|)d τ
0
≤ c2 (u(y) − u(y)2 + p − p2 ).

Taking the supremum on the left and having in mind that y → u(y) is Lipschitz
continuous, we obtain that the optimal trajectory mapping y → x(y) is Lipschitz
continuous from the L2 space of y to C(IRn , [0, 1]). Putting these facts together, we
confirm the claim in the theorem.
The optimal control u whose existence and uniqueness for a given y is asserted in
6F.3 is actually, as an element of L2 , an equivalence class of functions differing from
each other only on sets of measure zero in [0, 1]. Thus, having specified an optimal
control function u, we may change its values u(t) on a t-set of measure zero without
altering the value of the expression being minimized or affecting optimality. We will
go on to show now that one can pick a particular function from the equivalence class
which has better continuity properties with respect to both time and the parameter
dependence.
For a given control u and parameter y = (p, s, r), let

ψ = T ∗ (Qx + s),

where x solves (8). Then, through the Cauchy formula and 6F.1, ψ is given by
 1
T (τ −t)
ψ (t) = eA (Qx(τ ) + s(τ )) d τ .
t

Hence, ψ is a continuous function which is differentiable almost everywhere in [0, 1]

and its derivative ψ̇ is in L2 . Further, taking into account (6) and (8), ψ̇ satisfies

(16) ψ̇ (t) = −AT ψ (t) − Qx(t) − s(t) for a.e. t ∈ [0, 1], ψ (1) = 0,

where x is the solution of (2) for the given u. The function ψ is called the adjoint or
dual trajectory associated with a given control u and its corresponding state trajec-
tory x, and (16) is called the adjoint equation. Bearing in mind the particular form
of V (y) in (11) and that, by definition,

ψ = T ∗ (Qx + s) = T ∗ [QT Bu + s + QW(p)],

we can re-express the variational inequality (15) in terms of ψ as

6 Applications in Numerical Variational Analysis 357

(17) −r + Ru + B∗ψ , v − u ≥ 0 for all v ∈ C,

where B∗ stands for the linear mapping associated with the transpose of the matrix
B. The boundary value problem combining (2) and (16), coupled with the variational
inequality (17), fully characterizes the solution to problem (1)–(3).
We need next a standard fact from Lebesgue integration. For a function ϕ on
[0, 1], a point tˆ ∈ (0, 1) is said to be a Lebesgue point of ϕ when
 tˆ+ε
1
lim ϕ (τ )d τ = ϕ (t).
ε →0 2ε tˆ−ε

It is known that when ϕ is integrable on [0, 1], its set of Lebesgue points is of full
measure 1.
Now, let u be the optimal control for a particular parameter value y, and let x and
ψ be the associated optimal trajectory and adjoint trajectory, respectively. Let tˆ be a
Lebesgue point of both u and r (the set of such tˆ is of full measure). Pick any w ∈ U,
and for 0 < ε < min{tˆ, 1 − t̂} consider the function

w for t ∈ (tˆ − ε , tˆ + ε ),
ûε (t) =
u(t) otherwise.

Then for every sufficiently small ε the function ûε is a feasible control, i.e., belongs
to the set C in (13), and from (17) we obtain
 tˆ+ε
(−r(τ ) + Ru(τ ) + BTψ (τ ))T (w − u(τ ))d τ ≥ 0.
tˆ−ε

Since tˆ is a Lebesgue point of the function under the integral (we know that ψ is
continuous and hence its set of Lebesgue points is the entire interval [0, 1]), we can
pass to zero with ε and by taking into account that tˆ is an arbitrary point from a set
of full measure in [0, 1] and that w can be any element of U, come to the following
pointwise variational inequality which is required to hold for a.e. t ∈ [0, 1]:

(18) (−r(t) + Ru(t) + BTψ (t))T (w − u(t)) ≥ 0 for every w ∈ U.

As is easily seen, (18) implies (17) as well, and hence these two variational inequal-
ities are equivalent.
Summarizing, we can now say that a feasible control u is the solution of (1)–(3)
for a given y = (p, s, r) with corresponding optimal trajectory x and adjoint trajectory
ψ if and only if the triple (u, x, ψ ) solves the following boundary value problem
coupled with a pointwise variational inequality:

ẋ(t) = Ax(t) + Bu(t) + p(t), x(0) = a,
(19a)
ψ̇ (t) = −ATψ (t) − Qx(t) − s(t), ψ (1) = 0,

(19b) r(t) ∈ Ru(t) + BTψ (t) + NU (u(t)) for a.e. t ∈ [0, 1].
358 6 Applications in Numerical Variational Analysis

That is, for an optimal control u and associated optimal state and adjoint trajectories
x and ψ , there exists a set of full measure in [0, 1] such that (19b) holds for every t
in this set. Under an additional condition on the function r we obtain the following
result:

Theorem 6F.4 (implicit function theorem for continuous optimal controls). Let the
parameter y = (p, s, r) in (1)–(3) be such that the function r is Lipschitz continuous
on [0, 1]. Then, from the equivalence class of optimal control functions for this y,
there exists an optimal control u(y) for which (19b) holds for all t ∈ [0, 1] and which
is Lipschitz continuous with respect to t on [0, 1]. Moreover, the solution mapping
y → u(y) is Lipschitz continuous from the space L2 (IRn × IRn , [0, 1]) × C(IRm , [0, 1])
to the space C(IRm , [0, 1]).
Proof. It is clear that the adjoint trajectory ψ is Lipschitz continuous in t on [0, 1]
for any feasible control; indeed, it is the solution of the linear differential equation
(16), the right side of which is a function in L2 . Let x and ψ be the optimal state and
adjoint trajectories and let u be a function satisfying (19b) for all t ∈ σ where σ is a
/ σ we define u(t) to be the unique solution of the
set of full measure in [0, 1]. For t ∈
following strongly monotone variational inequality in IRn :

(20) q(t) ∈ Ru + NU (u), where q(t) = r(t) − BT ψ (t).

Then this u is within the equivalence class of optimal controls, and, moreover, the
vector u(t) satisfies (20) for all t ∈ [0, 1]. Noting that q is a Lipschitz continuous
function in t on [0, 1], we get from 2F.10 that for each fixed t ∈ [0, 1] the solution
mapping of (20) is Lipschitz continuous with respect to q(t). Since the composition
of Lipschitz continuous functions is Lipschitz continuous, the particular optimal
control function u which satisfies (19b) for all t ∈ [0, 1] is Lipschitz continuous in t
on [0, 1].
We already know from 6F.3 that the optimal trajectory mapping y → x(y) is
Lipschitz continuous into C(IRn , [0, 1]). By the same argument, the associated ad-
joint mapping y → ψ (y) is Lipschitz continuous from L2 to C(IRn , [0, 1]). But then,
according to 2F.10 again, for every y, y with r, r Lipschitz continuous and every
t ∈ [0, 1], with the optimal control values u(y)(t) and u(y )(t) at t being the unique
solutions of (20), we have
 
|u(y)(t) − u(y)(t)| ≤ μ −1 |r(t) − r (t)| + |B||ψ (y)(t) − ψ (y)(t)| .

This holds for every t ∈ [0, 1], so by invoking the maximum norm we get the desired
result.
We focus next on the issue of solving problem (1)–(3) numerically. By this we
mean determining the optimal control function u. This is a matter of recovering
a function on [0, 1] which is only specified implicitly, in this case by a variatio-
nal problem. Aside from very special cases, it means producing numerically an ac-
ceptable approximation of the desired function u. For simplicity, let us assume that
y = (p, s, r) = 0.
6 Applications in Numerical Variational Analysis 359

For such an approximation we may focus on a finite-dimensional space of func-

tions on [0, 1] within L2 . Suppose that the interval [0, 1] is divided into N pieces
[ti ,ti+1 ] by equally spaced nodes ti , i = 0, 1, . . . , N, with t0 = 0 and tN = 1, the fixed
mesh size being h = ti − ti−1 = 1/N. To approximate the optimal control function
u that is Lipschitz continuous on [0, 1] according to 6F.4, we will employ piecewise
constant functions across the grid {ti } that are continuous from the right at each ti ,
i = 0, 1, . . . , N − 1 and from the left at tN = 1. Specifically, for a given N, we consider
the subset of L2 given by
  

u  u(t) = u(ti ) for t ∈ [ti ,ti+1 ), i = 0, 1, . . . , N −2, u(t) = u(tN−1 ) for t ∈ [tN−1 ,tN ] .

In order to fully discretize problem (1)–(3) and transform it into a finite-dimensional

optimization problem, we also need to use finite-dimensional approximations of the
operators of integration and differentiation involved.
Rather than (1)–(3), we now invoke a discretization of the optimality system
(19ab). For solving the differential equations in (19a) we use the simplest Euler
scheme over the mesh {ti }. The Euler scheme applied to (19a), forward for the state
equation and backward for the adjoint equation, combined with restricting the func-
tional variational inequality (19b) to the nodes of the scheme, results in the following
discrete-time boundary value problem coupled with a finite-dimensional variational
inequality:
⎧
⎨ xi+1 = (I + hA)xi + hBui, x0 = a,
(21) ψi = (I + hAT )ψi+1 + hQxi+1 , ψN = 0,
⎩
0 ∈ Rui + BT ψi + NU (ui ) for i = 0, 1, . . . , N − 1.

There are various numerical techniques for solving problems of this form; here we
shall not discuss this issue.
We will now derive an estimate for the error in approximating the solution of
problem (1)–(3) by use of discretization (21) of the optimality system (19ab).
We suppose that for each given N we can solve (21) exactly, obtaining vectors
uNi ∈ U, i = 0, . . . , N − 1, and xNi ∈ IRn , ψiN ∈ IRn , i = 0, . . . , N. For a given N, the
solution (uN , xN , ψ N ) of (21) is identified with a function on [0, 1], where xN and
ψ N are the piecewise linear and continuous interpolations across the grid {ti } over
[0, 1] of (a, xN1 , . . . , xNN ) and (ψ0N , ψ1N , . . . , ψN−1
N , 0), respectively, and uN is the piece-

wise constant interpolation of (u0 , u1 , . . . , uNN−1 ) which is continuous from the right
N N

across the grid points ti = ih, i = 0, 1, . . . , N − 1 and from the left at tN = 1. The
functions xN and ψ N are piecewise differentiable and their derivatives ẋN and ψ̇ N
are piecewise constant functions which are assumed to have the same continuity
properties in t as the control uN . Thus, (uN , xN , ψ N ) is a function defined in the
whole interval [0, 1], and it belongs to L2 .

Theorem 6F.5 (error estimate for discrete approximation). Consider problem (1)–
(3) with r = 0, s = 0 and p = 0 under condition (4) and let, according to 6F.4,
(u, x, ψ ) be the solution of the equivalent optimality system (19ab) for all t ∈ [0, 1],
360 6 Applications in Numerical Variational Analysis

with u Lipschitz continuous in t on [0, 1]. Consider also the discretization (21) and,
for N = 1, 2, . . . and mesh size h = 1/N , denote by (uN , xN , ψ N ) its solution ex-
tended by interpolation to the interval [0, 1] in the manner described above. Then
the following estimate holds:

(22) uN − u∞ + xN − x∞ + ψ N − ψ ∞ = O(h).

Proof. For t ∈ [ti ,ti+1 ), i = 0, 1, . . . , N − 1, let

pN (t) = A(xN (ti ) − xN (t)),

sN (t) = AT (ψ N (ti+1 ) − ψ N (t)) + Q(xN (ti+1 ) − xN (t)),
rN (t) = −BT (ψ N (ti ) − ψ N (t)).

By virtue of the control uN being piecewise constant and

xN (ti+1 ) − xN (ti ) ψ N (ti+1 ) − ψ N (ti )

ẋN (t) = , ψ̇ N (t) = for t ∈ [ti ,ti+1 ),
h h
for i = 0, 1, . . . , N − 1, the discretized optimality system (21) can be written as fol-
lows: for all t ∈ [ti ,ti+1 ), i = 0, 1, . . . , N − 1, and t = 1,
⎧ N
⎨ ẋ (t) = AxN (t) + BuN (t) + pN (t), xN (0) = a,
T
ψ̇ (t) = −A ψ (t) − Qx (t) − s (t), ψ N (1) = 0,
(23) N N N N
⎩ N
r (t) ∈ RuN (t) + BT ψ N (t) + NU (uN (t)).

Observe that system (23) has the same form as (19ab) with a particular choice of
the parameters. Specifically, (uN , xN , ψ N ) is the solution of (19ab) for the parameter
value yN := (pN , sN , rN ), while (u, x, ψ ) is the solution of (19ab) for y = (p, s, r) =
(0, 0, 0). Then, by the implicit function theorem 6F.4, the solution mapping of (19)
is Lipschitz continuous in the maximum norms, so there exists a constant c such that

(24) uN − u∞ + xN − x∞ + ψ N − ψ ∞ ≤ cyN ∞ .

To finish the proof, we need to show that

(25) yN ∞ = max{pN ∞ , sN ∞ , rN ∞ } = O(h).

For that purpose we employ the following standard result in the theory of difference
equations which we state here without proof:

Lemma 6F.6 (discrete Gronwall lemma). Consider reals αi , i = 0, . . . , N , which

satisfy
i
0 ≤ α0 ≤ a and 0 ≤ αi+1 ≤ a + b ∑ α j for i = 0, . . . , N.
j=0

Then 0 ≤ αi ≤ a(1 + b) for i = 0, . . . , N. Similarly, if

i
6 Applications in Numerical Variational Analysis 361

N
0 ≤ αN ≤ a and 0 ≤ αi+1 ≤ a + b ∑ α j for i = 0, . . . , N,
j=i+1

then 0 ≤ αi ≤ a(1 + b)N−i for i = 0, . . . , N.

Continuing on this basis with the proof of (25), we observe that xN is piecewise
linear across the grid {ti }; clearly

|xN (t) − xN (ti )| ≤ |xN (ti+1 ) − xN (ti )| for t ∈ [ti ,ti+1 ], i = 0, 1, . . . , N − 1.

Then, since all ui are from the compact set U, from the first equation in (21) we get

|pN (t)| ≤ h(c1 |xN (ti )| + c2 ) for t ∈ [ti ,ti+1 ], i = 0, 1, . . . , N − 1

with some constants c1 , c2 independent of N. On the other hand, the first equation
in (21) can be written equivalently as
i
xN (ti+1 ) = a + ∑ h(AxN (t j ) + BuN (t j )),
j=1

and then, by taking norms and applying the direct part of discrete Gronwall lemma
6F.6, we obtain that sup0≤i≤N |xN (ti )| is bounded by a constant which does not de-
pend on N. This gives us error of order O(h) for pN in the maximum norm. By
repeating this argument for the discrete adjoint equation (the second equation in
(21)), but now applying the backward part of 6F.7, we get the same order of magni-
tude for sN and rN . This proves (25) and hence also (22).

Note that the order of the discretization error is O(h), which is sharp for the Euler
scheme. Using higher-order schemes may improve the order of approximation, but
this may require better continuity properties of the optimal control.
In the proof of 6F.5 we used the combination of the implicit function theo-
rem 6F.4 for the variational system involved and the estimate (25) for the resid-
ual yN = (pN , sN , rN ) of the approximation scheme. The convergence to zero of the
residual comes out of the approximation scheme and the continuity properties of the
solution of the original problem with respect to time t; in numerical analysis this is
called the consistency of the problem and its approximation. The property emerg-
ing from the implicit function theorem 6F.4, that is, the Lipschitz continuity of the
solution with respect to the residual, is sometimes called stability. Theorem 6F.5
furnishes an illustration of a well-known paradigm in numerical analysis: stability
plus consistency yields convergence.
Having the analysis of the linear-quadratic problem as a basis, we could pro-
ceed to more general nonlinear and nonconvex optimal control problems and obtain
convergence of approximations and error estimates by applying more advanced im-
plicit function theorems using, e.g., linearization of the associated nonlinear opti-
mality systems. However, this would involve more sophisticated techniques which
go beyond the scope of this book, so here is where we stop.
362 6 Applications in Numerical Variational Analysis

Commentary
Theorem 6A.2 is from Dontchev, Lewis and Rockafellar [2003], where it is supplied
with a direct proof. Theorem 6A.3 was first shown by Lewis [1999]; see also Lewis
[2001]. Theorem 6A.7 was initially proved in Dontchev, Lewis and Rockafellar
[2003] by using the characterization of the metric regularity of a mapping in terms
of the nonsingularity of its coderivative (see Section 4H) and applying the radius
theorem for nonsingularity in 6A.2. The proof given here is from Dontchev, Quin-
campoix and Zlateva [2006]. For extensions to infinite-dimensional spaces see Ioffe
[2003a,b]. Theorems 6A.8 and 6A.9 are from Dontchev and Rockafellar [2004].
The material in Section 6B is basically from Dontchev, Lewis and Rockafellar
[2003]. The results in Sections 6C and 6D have roots in several papers; see Rock-
afellar [1976a,b], Robinson [1994], Dontchev [2000], Aragón Artacho, Dontchev,
and Geoffroy [2007] and Dontchev and Rockafellar [2009b].
Most of the results in Section 6E can be found in basic texts on variational meth-
ods; for a recent such book see Attouch, Buttazzo and Michaille [2006]. Section
6F presents a very simplified version of a result in Dontchev [1996]; for advanced
studies in this area see Malanowski [2001] and Veliov [2006].
References

Alt, W. [1990], The LagrangeNewton method for infinite-dimensional optimization problems,

Numerical Functional Analysis and Optimization, 11, 201–224.
Apostol, T. [1962], Calculus, Blaisdell Publ. Co., Waltham, MA.
Aragón Artacho, F. J., A. L. Dontchev and M. H. Geoffroy [2007], Convergence of the proximal
point method for metrically regular mappings, ESAIM Proceedings, 17, 1–8.
Arutyunov, A. [2007], Covering mappings in metric spaces and fixed points, Doklady
Mathematics, 76, 665–668.
Attouch, H., G. Buttazzo and G. Michaille [2006], Variational analysis in Sobolev and BV spaces,
Applications to PDEs and optimization, MPS/SIAM Series on Optimization, 6,
Society for Industrial and Applied Mathematics, Philadelphia, PA; Mathematical
Programming Society, Philadelphia, PA.
Aubin, J-P. [1981], Contingent derivatives of set-valued maps and existence of solutions to
nonlinear inclusions and differential inclusions, in Mathematical analysis and
applications, Part A, edited by L. Nachbin, Advances in Mathematics: Supplementary
Studies, 7A, pp. 159–229, Academic Press, New York–London.
Aubin, J-P. [1984], Lipschitz behavior of solutions to convex minimization problems, Mathematics
of Operations Research, 9, 87–111.
Aubin, J-P. and I. Ekeland [1984], Applied nonlinear analysis, John Wiley & Sons, Inc., New York.
Aubin, J-P. and H. Frankowska [1987], On inverse function theorems for set-valued maps,
Journal de Mathématiques Pures et Appliquées, 66, 71–89.
Aubin, J-P. and H. Frankowska [1990], Set-valued analysis, Systems and Control: Foundations &
Applications, 2, Birkhäuser, Boston, MA.
Banach, S. [1922], Sur les opérations dans les ensembles abstraits et leur application aux équations
intégrales, Fundamenta Mathematicae, 3, 133–181.
Banach, S. [1932], Théorie des Operations Linéaires, Monografje Matematyczne, Warszawa,
English translation by North Holland, Amsterdam, 1987.
Bank, B., J. Guddat, D. Klatte, B. Kummer and K. Tammer [1983], Nonlinear parametric
optimization, Birkhäuser, Basel–Boston, MA.
Bartle, R. G. and L. M. Graves [1952], Mappings between function spaces, Transactions of
the American Mathematical Society, 72, 400–413.

363
364 References

Bartle, R. G. and D. R. Sherbert [1992], Introduction to real analysis, Second edition, John Wiley,
New York.
Bessis, D. N., Yu. S. Ledyaev and R. B. Vinter [2001], Dualization of the Euler and Hamiltonian
inclusions, Nonlinear Analysis, 43, 861–882.
Bonnans, J. F. and A. Shapiro [2000], Perturbation analysis of optimization problems, Springer
Series in Operations Research, Springer, New York.
Borwein, J. M. [1983], Adjoint process duality, Mathematics of Operations Research, 8, 403–434.
Borwein, J. M. [1986a], Stability and regular points of inequality systems, Journal of Optimization
Theory and Applications, 48, 9–52.
Borwein, J. M. [1986b], Norm duality for convex processes and applications, Journal of
Optimization Theory and Applications, 48, 53–64.
Borwein, J. M. and A. L. Dontchev [2003], On the Bartle-Graves theorem, Proceedings of the
American Mathematical Society, 131, 2553–2560.
Borwein, J. M. and A. S. Lewis [2006], Convex analysis and nonlinear optimization. Theory and
examples, Second edition, CMS Books in Mathematics/Ouvrages de Mathématiques
de la SMC, 3. Springer, New York.
Borwein, J. M. and Q. J. Zhu [2005], Techniques of variational analysis, CMS Books in
Mathematics/Ouvrages de Mathématiques de la SMC, 20, Springer, New York.
Cauchy, A-L. [1831], Résumé d’un mémoir sur la mécanique céleste et sur un nouveau calcul
appelé calcul des limites, in Oeuvres Complétes d’Augustun Cauchy, volume 12,
pp. 48–112, Gauthier-Villars, Paris 1916. Edited by L’Académie des Sciences.
The part pertaining to series expansions was read at a meeting of the Academie de
Turin, 11 October, 1831.
Chipman, J. S. [1997], “Proofs” and proofs of the Eckart-Young theorem, in Stochastic processes
and functional analysis, 71–83, Lecture Notes in Pure and Applied Mathematics, 186,
Dekker, New York.
Clarke, F. H. [1976], On the inverse function theorem, Pacific Journal of Mathematics, 64, 97–102.
Clarke, F. H. [1983], Optimization and nonsmooth analysis, Canadian Mathematical Society Series
of Monographs and Advanced Texts. A Wiley-Interscience Publication.
John Wiley & Sons, Inc., New York.
Cottle, R. W., J.-S. Pang and R. E. Stone [1992], The linear complementarity problem, Academic
Press, Inc., Boston, MA.
Courant, R. [1988], Differential and integral calculus, Vol. II. Translated from the German by
E. J. McShane. Reprint of the 1936 original. Wiley Classics Library.
A Wiley-Interscience Publication. John Wiley & Sons, Inc., New York.
Deimling, K. [1992], Multivalued differential equations, Walter de Gruyter, Berlin.
Deville, R., G. Godefroy and V. Zizler [1993], Smoothness and renormings in Banach spaces,
Pitman Monographs and Surveys in Pure and Applied Mathematics, 64. Longman
Scientific & Technical, Harlow; copublished in the United States with
John Wiley & Sons, Inc., New York.
Deutsch, F. [2001], Best approximation in inner product spaces, CMS Books in
Mathematics/Ouvrages de Mathématiques de la SMC, 7. Springer, New York.
Dieudonné, J. [1969], Foundations of modern analysis, Enlarged and corrected printing.
Pure and Applied Mathematics, Vol. 10-I, Academic Press, New York.
References 365

Dini, U. [1877/78], Analisi infinitesimale, Lezioni dettate nella R. Universitá di Pisa.

Dmitruk, A. V., A. A. Milyutin and N. P. Osmolovskiı̆ [1980], Lyusternik’s theorem and the theory
of extremum, Uspekhi Mat. Nauk, 35, no. 6 (216), 11–46 (Russian).
Dontchev, A. L. [1995a], Characterizations of Lipschitz stability in optimization, in Recent
developments in well-posed variational problems, 95–115, Kluwer, Dordrecht.
Dontchev, A. L. [1995b], Implicit function theorems for generalized equations, Mathematical
Programming, 70, Ser. A, 91–106.
Dontchev, A. L. [1996], An a priori estimate for discrete approximations in nonlinear optimal
control, SIAM Journal of Control and Optimization, 34, 1315–1328.
Dontchev, A. L. [2000], Lipschitzian stability of Newton’s method for variational inclusions, in
System modelling and optimization (Cambridge, 1999), 119–147, Kluwer Acad.
Publ., Boston, MA.
Dontchev, A. L. [2004], A local selection theorem for metrically regular mappings. Journal of
Convex Analysis, 11, 81–94.
Dontchev, A. L. and W. W. Hager [1993], Lipschitzian stability in nonlinear control and
optimization, SIAM Journal of Control and Optimization, 31, 569–603.
Dontchev, A. L. and W. W. Hager [1994], An inverse mapping theorem for set-valued maps,
Proceedings of the American Mathematical Society, 121, 481–489.
Dontchev, A. L. and R. T. Rockafellar [1996], Characterizations of strong regularity for variational
inequalities over polyhedral convex sets, SIAM Journal on Optimization, 6,
1087–1105.
Dontchev A. L. and R. T. Rockafellar [1998], Characterizations of Lipschitzian stability in
nonlinear programming, in Mathematical programming with data perturbations,
65–82, Lecture Notes in Pure and Appl. Math., 195, Dekker, New York.
Dontchev, A. L. and R. T. Rockafellar [2001], Ample parameterization of variational inclusions,
SIAM Journal on Optimization, 12, 170–187.
Dontchev, A. L. and R. T. Rockafellar [2004], Regularity and conditioning of solution mappings
in variational analysis, Set-Valued Analysis, 12, 79–109.
Dontchev, A. L. and R. T. Rockafellar [2009a], Robinson’s implicit function theorem and its
extensions, Mathematical Programming, 117, 129–147.
Dontchev, A. L. and R. T. Rockafellar [2009b], Newton’s method for generalized equations:
a sequential implicit function theorem, to appear in Mathematical Programming.
Dontchev, A. L. and T. Zolezzi [1993], Well-posed optimization problems, Lecture Notes in
Mathematics 1543, Springer, Berlin.
Dontchev, A. L., A. S. Lewis and R. T. Rockafellar [2003], The radius of metric regularity,
Transactions of the American Mathematical Society, 355, 493–517.
Dontchev, A. L., M. Quincampoix and N. Zlateva [2006], Aubin criterion for metric regularity,
Journal of Convex Analysis, 13, 281–297.
Dunford, N. and J. T. Schwartz [1958], Linear Operators I. General Theory, with the assistance of
W. G. Bade and R. G. Bartle. Pure and Applied Mathematics, Vol. 7 Interscience,
Publishers, Inc., New York; Interscience Publishers, Ltd., London.
Eckart, C. and G. Young [1936], The approximation of one matrix by another of lower rank,
Psychometrica, 1, 211–218.
366 References

Ekeland, I. [1990], The ε -variational principle revisited, with notes by S. Terracini, in Methods of
nonconvex analysis (Varenna 1989), 1–15, Lecture Notes in Mathematics, 1446,
Springer, Berlin.
Facchinei, F. and J-S. Pang [2003], Finite-dimensional variational inequalities and
complementarity problems, Springer, New York.
Fiacco, A. V. and G. P. McCormick [1968], Nonlinear programming: Sequential unconstrained
minimization techniques, John Wiley & Sons, Inc., New York–London–Sydney.
Fitzpatrick, P. [2006], Advanced calculus, Thomson Brooks/Cole.
Golub, G. H. and C. F. Van Loan [1996], Matrix computations, The John Hopkins University Press,
Baltimore, MD, 3d edition.
Goursat, Ed. [1903], Sur la théorie des fonctions implicites, Bulletin de la Société Mathématiques
de France, 31, 184–192.
Goursat, Ed. [1904], A course in mathematical analysis, English translation by E. R. Hedrick,
Ginn Co., Boston.
Graves, L. M. [1950], Some mapping theorems, Duke Mathematical Journal, 17, 111–114.
Halkin, H. [1974], Implicit functions and optimization problems without continuous
differentiability of the data, SIAM Journal on Control, 12, 229–236.
Hamilton, R. S. [1982], The inverse function theorem of Nash and Moser, Bulletin of the
American Mathematical Society (N.S.), 7, 65–222.
Hausdorff, F. [1927], Mengenlehre, Walter de Gruyter and Co., Berlin.
Hedrick, E. R. and W. D. A. Westfall [1916], Sur l’existence des fonctions implicites, Bulletin de
la Société Mathématiques de France, 44, 1–13.
Hildebrand, H. and L. M. Graves [1927], Implicit functions and their differentials in general
analysis, Transactions of the American Mathematical Society, 29, 127–153.
Hoffman, A. J. [1952], On approximate solutions of systems of linear inequalities, Journal of
Research of the National Bureau of Standards, 49, 263–265.
Hurwicz, L. and M. K. Richter [2003], Implicit functions and diffeomorphisms without C 1 ,
Advances in mathematical economics, 5, 65–96.
Ioffe, A. D. [1979], Regular points of Lipschitz functions, Transactions of the American
Mathematical Society, 251, 61–69.
Ioffe, A. D. [1981], Nonsmooth analysis: differential calculus of nondifferentiable mappings,
Transactions of the American Mathematical Society, 266, 1–56.
Ioffe, A. D. [1984], Approximate subdifferentials and applications, I. The finite-dimensional
theory, Transactions of the American Mathematical Society, 281, 389–416.
Ioffe, A. D. [2000], Metric regularity and subdifferential calculus, Uspekhi Mat. Nauk, 55, no. 3
(333), 103–162 (Russian), English translation in Russian Mathematical Surveys,
55, 501–558.
Ioffe, A. D. [2003a], On robustness of the regularity property of maps, Control and Cybernetics,
32, 543–554.
Ioffe, A. D. [2003b], On stability estimates for the regularity property of maps, in H. Brezis,
K.C. Chang, S.J. Li, and P. Rabinowitz, editors, Topological methods, variational
methods and their applications, pp. 133–142, World Science Publishing, NJ.
References 367

Ioffe, A. D. and V. M. Tikhomirov [1974], Theory of extremal problems, Nauka, Moscow

(Russian).
Jongen, H. Th., T. Möbert, J. Rückmann and K. Tammer [1987], On inertia and Schur complement
in optimization, Linear Algebra and Applications, 95, 97–109.
Josephy, N. H. [1979], Newton’s method for generalized equations and the PIES energy model,
Ph.D. Dissertation, Department of Industrial Engineering, University of
Wisconsin-Madison.
Kahan, W. [1966], Numerical linear algebra, Canadian Mathematics Bulletin, 9, 757–801.
Kantorovich, L. V. and G. P. Akilov [1964], Functional analysis in normed spaces, Macmillan,
New York. Translation by D. E. Brown from Russian original, Fizmatgiz,
Moscow, 1959.
Kenderov, P. [1975], Semi-continuity of set-valued monotone mappings, Fundamenta
Mathematicae, 88, 61–69.
King, A. J. and R. T. Rockafellar [1992], Sensitivity analysis for nonsmooth generalized equations,
Mathematical Programming, Ser. A, 55(2), 193–212.
Klatte, D. and B. Kummer [2002], Nonsmooth equations in optimization. Regularity, calculus,
methods and applications. Kluwer, Dordrecht.
Kojima, M. [1980], Strongly stable stationary solutions in nonlinear programs, in Analysis and
computation of fixed points (Madison, Wisconsin, 1979), pp. 93–138, Academic
Press, New York–London.
Krantz, S. G. and H. R. Parks [2002], The implicit function theorem. History, theory, and
applications, Birkhäuser, Boston.
Kruger, A. [1982], On characterizing the covering property of nonsmooth operators, in Proceedings
of the School on Theory of Operators in Function Spaces, Minsk, 94–95 (Russian).
Kummer, B. [1991], An implicit-function theorem for C0,1 -equations and parametric
C1,1 -optimization, Journal of Mathematical Analysis and Applications, 158, 35–46.
Kuratowski, K. [1933], Topologie, I & II, Panstwowe Wydawnictwo Naukowe, Warszawa.
Lamson, K. W. [1920], A general implicit function theorem with an application to problems of
relative minima, American Journal of Mathematics, 42, 243–256.
Leach, E. B. [1961], A note on inverse function theorems, Proceedings of the American
Mathematical Society, 12, 694–697.
Ledyaev, Yu. S. and Q. J. Zhu [1999], Implicit multifunction theorems, Set-Valued Analysis,
7, 209–238.
Levitin, E. S. [1992], Perturbation theory in mathematical programming and its applications,
Nauka, Moscow (Russian).
Levy, A. B. [1996], Implicit multifunction theorems for the sensitivity analysis of variational
conditions, Mathematical Programming, Ser. A, 74(3), 333–350.
Levy, A. B. and R. A. Poliquin [1997], Characterizing the single-valuedness of multifunctions.
Set-Valued Analysis, 5, 351–364.
Lewis, A. S. [1999], Ill-conditioned convex processes and conic linear systems, Mathematics of
Operations Research, 24, 829–834.
Lewis A. S. [2001], Ill-conditioned inclusions, Set-Valued Analysis, 9, 375–381.
368 References

Li, Wu [1994], Sharp Lipschitz constants for basic optimal solutions and basic feasible solutions
of linear programs, SIAM Journal on Control and Optimization, 32, 140–153.
Lyusternik, L. A. [1934], On the conditional extrema of functionals, Mat. Sbornik, 41, 390–401
(Russian).
Lyusternik, L. A. and V. I. Sobolev [1965], Elements of functional analysis, Nauka, Moscow
(Russian).
Malanowski, K. [2001], Stability and sensitivity analysis for optimal control problems with
control-state constraints, Dissertationes Mathematicae, 394, 55pp.
Mordukhovich, B. S. [1984], Nonsmooth analysis with nonconvex generalized differentials and
conjugate mappings, Doklady Akad. Nauk BSSR, 28, 976–979, (Russian).
Mordukhovich, B. S. [2006], Variational analysis and generalized differentiation, I. Basic theory,
Springer, Berlin.
Nadler, Sam B., Jr. [1969], Multi-valued contraction mappings, Pacific Journal of Mathematics,
30, 475–488.
Ng, Kung Fu [1973], An open mapping theorem, Proceedings of Cambridge Philosophical
Society, 74, 61–66.
Nijenhuis, A. [1974], Strong derivatives and inverse mappings, The American Mathematical
Monthly, 81, 969–980.
Noble, B. and J. W. Daniel [1977], Applied linear algebra, Second edition, Prentice-Hall, Inc.,
Englewood Cliffs, N.J.
Páles, Z. [1997], Inverse and implicit function theorems for nonsmooth maps in Banach spaces,
Journal of Mathematical Analysis and Applications, 209, 202–220.
Pompeiu, D. [1905], Fonctions de variable complexes, Annales de la Faculté des Sciences
de l´ Université de Toulouse, 7, 265–345.
Ralph, D. [1993], A new proof of Robinson’s homeomorphism theorem for PL-normal maps,
Linear Algebra and Applications, 178, 249–260.
Robinson, S. M. [1972], Normed convex processes, Transactions of the American Mathematical
Society, 174, 127–140.
Robinson, S. M. [1976], Regularity and stability for convex multivalued functions, Mathematics of
Operations Research, 1, 130–143.
Robinson, S. M. [1980], Strongly regular generalized equations, Mathematics of Operations
Research, 5, 43–62.
Robinson, S. M. [1981], Some continuity properties of polyhedral multifunctions, Mathematical
Programming Study, 14, 206–214.
Robinson, S. M. [1984], Local structure of feasible sets in nonlinear programming, Part II:
Nondegeneracy, Mathematical Programming Study, 22, 217–230.
Robinson, S. M. [1991], An implicit-function theorem for a class of nonsmooth functions,
Mathematics of Operations Research, 16, 292–309.
Robinson, S. M. [1992], Normal maps induced by linear transformations, Mathematics of
Operations Research, 17, 691–714.
Robinson, S. M. [1994], Newton’s method for a class of nonsmooth functions, Set-valued
Analysis, 2, 291–305.
References 369

Robinson, S. M. [2007], Solution continuity in monotone affine variational inequalities, SIAM

Journal in Optimization, 18, 1046–1060.
Rockafellar, R. T. [1967], Monotone processes of convex and concave type, Memoirs of
the American Mathematical Society, 77, Providence, RI.
Rockafellar, R. T. [1970], Convex analysis, Princeton University Press, Princeton, NJ.
Rockafellar, R. T. [1976a], Monotone operators and the proximal point algorithm, SIAM Journal
on Control and Optimization, 14, 877–898.
Rockafellar, R. T. [1976b], Augmented Lagrangians and applications of the proximal point
algorithm in convex programming, Mathematics of Operations Research, 1, 97–116.
Rockafellar, R. T. [1985], Lipschitzian properties of multifunctions, Nonlinear Analysis, 9,
867–885.
Rockafellar, R. T. [1989], Proto-differentiability of set-valued mappings and its applications
in optimization, in Analyse non linéaire (Perpignan 1987), Ann. Inst. H. Poincaré,
Anal. Non Linéaire, 6, suppl., 448–482.
Rockafellar, R. T. and R. J-B. Wets [1998], Variational Analysis, Springer, Berlin.
Rudin, W. [1991], Functional analysis, Second edition, McGraw-Hill, Inc., New York.
Schirotzek, W. [2007], Nonsmooth analysis, Springer, Berlin.
Scholtes, S. [1994], Introduction to piecewise differentiable equations, Preprint 53, Institut für
Statistik und Mathematische Wirtschaftstheorie, Universität Karlsruhe, Karlsruhe.
Spanier, E. H. [1966], Algebraic topology, McGraw-Hill, New York.
Stewart, G. W. and J. G. Sun [1990], Matrix perturbation theory, Academic Press, Boston.
Schwartz, L. [1967], Analyse mathématique, I, Hermann, Paris.
Ursescu, C. [1975], Multifunctions with convex closed graph, Czechoslovak Mathematical Journal,
25 (100), 438–441.
Veliov, V. M. [2006], Approximations with error estimates for optimal control problems for
linear systems, in Large-scale scientific computing, 263–270, Lecture Notes in
Computer Science, 3743, Springer, Berlin.
Walkup, D. W. and R. J-B. Wets [1969], A Lipschitzian characterization of convex polyhedra,
Proceedings of the American Mathematical Society, 23, 167–173.
Zălinescu, C. [2002], Convex analysis in general vector spaces, World Scientific Publishing, NJ.
Notation

2C(4): formula (4) in Section 2C

IR: the real numbers
IN: the natural numbers
N : the collection of all subsets N of IN such that IN \ N is finite
N  : the collection of all infinite subsets of IN
{xk }: a sequence with elements xk
εk  0: a sequence of positive numbers εk tending to 0
lim supk Ck : outer limit
lim infk Ck : inner limit
|x|: Euclidean norm
x: any norm
x, y: canonical inner product, bilinear form
|H|+ : outer norm
|H|− : inner norm
IBa (x): closed ball with center x and radius r
IB: closed unit ball
cl C: closure
int C: interior
core C: core
rc C: recession cone
PC : projection mapping
TC (x): tangent cone
NC (x): normal cone
K ∗ : polar to cone K, mapping adjoint to K, space dual to K
KC (x, v): critical cone
AT : transposition
rank A: rank
ker A: kernel
det A: determinant
dC (x), d(x,C): distance from x to C
e(C, D): the excess of C beyond D
h(C, D): Pompeiu-Hausforff distance
dom F: domain
rge F: range
gph F: graph
∇ f (x): Jacobian
D f (x): derivative

371
372 Notation

C k : the space of k-times continuously differentiable functions

DF(x | y): graphical derivative
D∗ F(x | y): coderivative
clm ( f ; x), clm(S; y| x): calmness modulus
lip( f ; x), lip(S; y| x): Lipschitz modulus
 p ( f ; (p, x)): partial calmness modulus
clm

lip p ( f ; (p, x)): partial Lipschitz modulus
reg (F; x | y): regularity modulus
subreg (F; x | y): subregularity modulus
Index

adjoint strict partial, 34

equation, 356 discretization, 359
upper and lower, 269 distance, 28
ample parameterization, 85 Pompeiu–Hausdorff, 138
Aubin property, 159 to infeasibility, 322
partial, 167
Ekeland variational principle, 207
calmness estimator, 38
isolated, 186 partial, 45
partial, 25 excess, 138
Cauchy formula, 353
complementarity problem, 64 first-order approximation, 36
cone, 62 partial, 45
critical, 98 function
normal, 62 calm, 22
polar, 64 convex, 66
recession, 267 Lipschitz continuous, 5
tangent, 65 monotone, 106
constraint qualification, 70 piecewise smooth, 93
contraction mapping principle, 15 positively homogeneous, 88
for set-valued mappings, 284 semidifferentiable, 89
convergence strictly differentiable, 31
linear, 329 upper semicontinuous, 4
of iterations under metric regularity, 328
Painlevé–Kuratowski, 135 Galerkin method, 348
Pompeiu–Hausdorff, 140 generalized equation, 62
quadratic, 331 generalized Jacobian, 238
set, 134
superlinear, 329 homogenization, 323
convex programming, 73
implicit function theorem
derivative classical (Dini), 17
convexified graphical, 213 for generalized equations, 79
Fréchet, 275 for local minima, 125
graphical, 199 for Newton iteration, 341
one-sided directional, 89 for optimal control, 355
strict graphical, 239 for stationary points, 123

373
374 Index

Goursat, 20 outer Lipschitz continuous, 154

Robinson, 75 outer semicontinuous, 142
utilizing semiderivatives, 92 Painlevé–Kuratowski continuous, 142
with first-order approximations, 79 polyhedral, 155
with strong metric regularity, 180 polyhedral convex, 150
with strong metric regularity in metric Pompeiu–Hausdorff continuous, 142
spaces, 294 positively homogeneous, 200
implicit mapping theorem stationary point, 115
for a constraint system, 221 sublinear, 265
with graphical derivatives, 211 with closed convex graph, 259
with metric regularity, 172 metric regularity, 164
with metric regularity in metric spaces, 286 coderivative criterion, 246
with strong metric subregularity, 192 derivative criterion, 205
inner and outer limits, 134 of sublinear mappings, 265
inverse function theorem strong, 179
beyond differentiability, 38 metric subregularity, 182
Clarke, 238 derivative criterion for strong, 218
classical, 10 strong, 186
for local diffeomorphism, 48 modulus
Kummer, 240 calmness, 22
symmetric, 24 Lipschitz, 26
with strong metric regularity in metric metric regularity, 164
spaces, 292 metric subregularity, 183
with strong metric regularity, 179 partial calmness, 25
inverse mapping theorem partial uniform Lipschitz, 34
with continuous and calm local selections,
300 Nash equilibrium, 73
with metric regularity, 169 necessary condition for optimality, 69
with metric regularity in metric spaces, 280 Newton method, 11
with strong metric subregularity, 190 uniform convergence, 337
for generalized equations, 327
Karush–Kuhn–Tucker conditions, 72 nonlinear programming, 72
norm
Lagrange multiplier rule, 70 duality, 270
lemma outer and inner, 202
critical face, 229 operator, 6
discrete Gronwall, 360
Hoffman, 150 openness, 56
reduction, 98 linear, 166
optimal control, 352
Mangasarian–Fromovitz constraint qualifica- optimization, 67
tion, 176
mapping parametric robustness, 88
adjoint, 252 projection, 28
calm, 182 proximal point method, 327
feasible set, 145
horizon, 323 saddle point, 73
inner semicontinuous, 142 second-order optimality, 113
linear, 5 selection, 49
Lipschitz continuous, 148 seminorm, 22
locally monotone, 181 set
maximal monotone, 335 adsorbing, 259
optimal set, 145 convex, 27
optimal value, 145 locally closed, 169
Index 375

polyhedral convex, 97 Lyusternik, 275

space Michael selection, 298
dual, 252 Minkowski–Weyl, 97
metric, 251 Nadler, 291
SQP method, 334 Nash–Moser, 310
radius for metric regularity, 317
theorem radius for strong metric regularity, 319
Baire category, 260 radius for strong metric subregularity, 319
Banach open mapping, 253 Robinson–Ursescu, 263
Bartle–Graves, 298 two-person zero-sum game, 73
Brouwer fixed point, 52
Brouwer invariance of domain, 47 variational inequality, 62
correction function, 19 for a Nash equilibrium, 73
Graves, 276 affine polyhedral, 100
Hahn–Banach, 270 Lagrangian, 71
Hildebrand–Graves, 58 monotone, 110