Unifies the field of optimization with

a few geometric principles

The number of books that can legitimately be called classics in their fields is
small indeed, but David Luenberger's Optimization by Vector Space Methods
certainly qualifies, Not only does Luenberger clearly demonstrate that a large
segment of the field of optimization can be effectively unified by a few
geometric principles of linear vector space theory, but his methods have
found applications quite removed from the engineering problems to which
they were first applied. Nearly 30 years after its initial publication, this book
is still among the most frequently cited sources in books and articles on
financial optimization.
The book uses functional analysisthe study of linear vector spacesto
impose simple, intuitive interpretations on complex, infinite-dimensional
problems. The early chapters offer an introduction to functional analysis,
with applications to optimization. Topics addressed include linear space,
Hilbert space, least-squares estimation, dual spaces, and linear operators and
adjoints. Later chapters deal explicitly with optimization theory, discussing

Optimization of functionals
Global theory of constrained optimization
Local theory of constrained optimization
Iterative methods of optimization

End-of-chapter problems constitute a major component of this book and

come in two basic varieties. The first consists of miscellaneous mathematical
problems and proofs that extend and supplement the theoretical material in
the text; the second, optimization problems, illustrates further areas of
application and helps the reader formulate and solve practical problems.
For professionals and graduate students in engineering, mathematics,
operations research, economics, and business and finance, Optimization by
Vector Space Methods is an indispensable source of problem-solving tools.
DAVID G. LUENBERGER is a professor in the School of Engineering at
Stanford University. He has published four textbooks and over 70 technical
papers. Professor Luenberger is a Fellow of the Institute of Electrical and
Electronics Engineers and recipient of the 1990 Bode Lecture Award. His
current research is mainly in investment science, economics, and planning.

Cover Design: Edward Smith Design, Inc.

WILEY-1NTERSCIENCE

John Wiley & Sons, Inc.

Professional, Reference andTrade Group
605 Third Avenue,NewYork,N.Y.10158-0012
New York Chichester Weinheim
Brisbane Singapore Toronto

ISBN 0 - 471- IA117 - X

90000

9 780471 181170

This text is printed on acid-free paper.

Copyright 1969 by John Wiley & Sons, Inc.
All rights reserved. Published simultaneously in Canada.
Reproduction or translation of any part of this work beyond
that permitted by Section 107 or 108 of the 1976 United
States Copyright Act without the permission of the copyright
owner is unlawful. Requests for permission or further
information should be addressed to the Permissions Department,
John Wiley & Sons, Inc., 605 Third Avenue, New York, NY
10158-0012.
This publication is designed to provide accurate and
authoritative information in regard to the subject
matter covered. it is sold with the understanding that
the publisher is not engaged in rendering legal, =counting,
or other professional services. If legal advice or other
expert assistance is required, the services of a competent
professional person should be sought.

Library of Congress Cataloging-in-Publication Data

ISBN 047I-55359-X (cloth)
ISBN 0471-18I17-X (paper)
Printed in the United States of America
/0 9 8 7 6 5

OPTIMIZATION BY
VECTOR SPACE METHODS

David G. Luenberger
Stanford University,
Stanford, California

John Wiley & Sons, Inc.

New York Chichester Weinheim Brisbane Singapore Toronto

SERIES IN DECISION AND CONTROL

Ronald A. Howard

OPTIMIZATION BY VECTOR SPACE METHODS

by David G. Luenberger
INTRODUCTION TO DYNAMIC PROGRAMMING
by George L. Nemhauser

DYNAMIC PROBABILISTIC SYSTEMS (In Preparation)

by Ronald A. Howard

OPTIMIZATION BY
VECTOR SPACE METHODS

PREFACE

This book has evolved from a course on optimization that I have taught at
Stanford University for the past five years. It is intended to be essentially
self-contained and should be suitable for classroom work or self-study.
As a text it is aimed at first- or second-year graduate students in engineering,
mathematics, operations research, or other disciplines dealing with optimization theory.
The primary objective of the book is to demonstrate that a rather large
segment of the field of optimization can be effectively unified by a few
geometric principles of linear vector space theory. By use of these principles,
important and complex infinite-dimensional problems, such as those
generated by consideration of time functions, are interpreted and solved by
methods springing from our geometric insight. Concepts such as distance,
orthogonality, and convexity play fundamental roles in this development.
Viewed in these terms, seemingly diverse problems and techniques often
are found to be intimately related.
The essential mathematical prerequisite is a familiarity with linear
algebra, preferably from the geometric viewpoint. Some familiarity with
elementary analysis including the basic notions of sets, convergence, and
continuity is assumed, but deficiencies in this area can be corrected as one
progresses through the book. More advanced concepts of analysis such as
Lebesgue measure and integration theory, although referred to in a few
isolated sections, are not required background for this book.
Imposing simple intuitive interpretations on complex infinite-dimensional problems requires a fair degree of mathematical sophistication.
The backbone of the approach taken in this book is functional analysis,
the study of linear vector spaces. In an attempt to keep the mathematical
prerequisites to a minimum while not sacrificing completeness of the development, the early chapters of the book essentially constitute an introduction
to functional analysis, with applications to optimization, for those having
the relatively modest background described above`The mathematician or
more advanced student may wish simply to scan Chapters 2, 3, 5, and 6 for
review or for sections treating applications and then concentrate on the
other chapters which deal explicitly with optimization theory.
vii

viii

PREFACE

The sequencing of the various sections is not necessarily inviolable.

Even at the chapter level the reader may wish to alter his order of progress
through the book. The course from which this text developed is two
quarters (six months) in duration, but there is more material in the text
than can be comfortably covered in that period. By reading only the first
few sections of Chapter 3, it is possible to go directly from Chapters I and 2
to Chapters 5, 7,8, and 10 for a fairly comprehensive treatment of optimization which can be covered in about one semester. Alternatively, the
material at the end of Chapter 6 can be combined with Chapters 3 and 4 for
a unified introduction to Hilbert space problems. To help the reader make
intelligent decisions regarding his order of progress through the book,
sections of a specialized or digressive nature are indicated by an *.
The problems at the end of each chapter are of two basic varieties.
The first consists of miscellaneous mathematical problems and proofs
which extend and supplement the theoretical material in the text; the second
consists of optimization problems which illustrate further areas of application and which hopefully will help the student formulate and solve
practical problems. The problems represent a major component of the
book, and the serious student will not pass over them lightly.
I have received help and encouragement from many people during the
years of preparation of this book. Of great benefit were comments and
suggestions of Pravin Varaiya, E. Bruce Lee, and particularly Samuel
Karlin who read the entire manuscript and suggested several valuable
improvements. I wish to acknowledge the Departments of EngineeringEconomic Systems and Electrical Engineering at Stanford University for
supplying much of the financial assistance. This effort was also partially
supported by the Office of Naval Research and the National Science Fohndation. Of particular benefit, of course, have been the faces of puzzled
confusion or of elated understanding, the critical comments and the sincere
suggestions of the many students who have worked through this material
as the book evolved.
DAVID

G.

LUENBERGER

Palo Alto, California

August 1968

CONTENTS

1 INTRODUCTION

1.1
1.2
1.3
1.4

Motivation
Applications 2
The Main Principles 8
Organizatioll of the Book

2 LINEAR SPACES

10

11

2.1

Introduction.

2.2
2.3
2.4

VECTOR SPACES 11
Definition and Examples 1I
Subspaces, Linear Combinations, and Linear Varieties
Convexity and Cones 17
Linear Independence and Dimension 19

2.5
2.6
2.7
2.8
2.9
*2.10
2. I I
2.12
*2.13
*2.14
*2. I 5
2.16

NORMED LINEAR SPACES 22

Definition and Examples 22
Open and Cllosed Sets 24
Convergence: 26
Transformations and Continuity 27
The lp and Lp Spaces 29
Banach Spaces 33
Complete Subsets 38
Extreme Vahtes of Functionals, and Compactness
Quotient Spnces 41
Denseness and Separability 42
Problems 43
References 45

3 HILBERT SPACE
3.1

II

46

Introduction 46
ix

39

14

CONTENTS

3.2
3.3
3.4
3.5
3.6
3.7
*3.8
3.9

PRE-HILBERT SPACES 46
Inner Products 46
The Projection Theorem 49
Orthogonal Complements 52
The Gram-Schmidt Procedure 53
APPROXIMATION 55
The Normal Equations and Gram Matrices
Fourier Series 58
Complete Orthonormal Sequences 60
Approximation and Fourier Series 62

55

OTHER MINIMUM NORM PROBLEMS 64

3.10 The Dual Approximation Problem 64
*3.11 A Control Problem 68
3.12 Minimum Distance to a Convex Set 69
3.13 Problems 72
References 77
4 LEAST-SQUARES ESTIMATION 78

4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8

Introduction 78
Hilbert Space of Random Variables 79
The Least-Squares Estimate 82
Minimum-Variance Unbiased Estimate (Gauss-Markov Estimate) 84
Minimum-Variance Estimate 87
Additional Properties of Minimum-Variance Estimates 90
Recursive Estimation 93
Problems 97
References 102

5 DUAL SPACES

103

5.1

Introduction

103

5.2
5.3

UNEAR FUNCTIONALS 104

Basic Concepts 104
Duals of Some Common Banach Spaces

5.4
5.5

EXTENSION FORM OF THE HAHN-BANACH

THEOREM 110
Extension of Linear Functionals 110
The Dual of C[a, b] 113

5.6

The Second Dual Space

5.7

Alignment and Orthogonal Complements

106

115
116

CONTENTS

5.8 Minimum Norm Problems

5.9 Applications 122
*5.10 Weak Convergence 126

118

GEOMETRIC FORM OF THE HAHN-BANACH

THEOREM 129

5.11
5.12
*5.13
5.14

Hyperplanes and Linear Functionals 129

Hyperplanes and Convex Sets 131
Duality in Minimum Norm Problems 134
Problems 137
References 142

6 LINEAR OPERATORS AND ADJOINTS 143

6.1
6.2

Introduction 143
Fundamentals 143
INVERSE OPERATORS 147

6.3
6.4

Linearity ofInverses 147

The Banach Inverse Theorem
ADJOINTS

6.5
6.6
6.7
*6.8

148

150

Definition and Examples 15('

Relations between Range and Nullspace 155
Duality Relations for Convex Cones 157
Geometric Interpretation of Adjoints 159
OPTIMIZATION IN HILBERT SPACE 160

6.9
6.10
6.11
6.12

The Normal Equations 160

The Dual Problem 161
Pseudoinverse Operators 163
Problems 165
References 168

7 OPTIMIZATION OF FUNCTIONALS 169

7.1

Introduction 169
LOCAL THEORY 171

7.2
7.3
7.4
*7.5
*7.6
7.7

Gateaux and Frechet Differentials 171

Frechet Derivatives 175
Extrema 177
Euler-Lagrange Equations 179
Problems with Variable End Points 183
Problems with Constraints 185

xi

X11

CONTENTS

GLOBAL THEORY 190

7.8
*7.9
7.10
7.11
7.12
*7.13
7.14

Convex and Concave Functionals 190

Properties of the Set [f, C] 192
Conjugate Convex Functionals 195
Conjugate Concave FUl1ctionals 199
Dual Optimization Problems 200
Min- Max Theorem of Game Theory 206
Problems 209
References 212

8 GLOBAL THEORY OF CONSTRAINED OPTIMIZATION 213

8.1
8.2
8.3
8.4
8.5
8.6
8.7
8.8

Introduction 213
Positive Cones and Convex Mappings
Lagrange Multipliers 216
Sufficiency 219
Sensitivity 221
Duality 223
Applications 226
Problems 236
References 238

214

9 LOCAL THEORY OF CONSTRAINED OPTIMIZATION 239

9.1
9.2
. 9.3
9.4
9.5
*9.6
9.7

Introduction

239

LAGRANGE MULTIPLIER THEOREMS 240

Inverse Function Theorem 240
Equality Constraints 242
Inequality Constraints (Kuhn-Tucker Theorem) 247
OPTIMAL CONTROL THEORY 254
Basic Necessary Conditions 254
Pontryagin Maximum Principle 261
Problems 266
References 269

10 ITERATIVE METHODS OF OPTIMIZATION 271

10.1

Introduction

271

10.2

METHODS FOR SOLVING EQUATIONS 272

Successive Approximation 272

10.3 Newton's Method 277

CONTENTS

DESCENT METHODS 283

10.4 General Philosophy 283
10.5 Steepest Descent 285
CONJUGATE DIRECTION METHODS 290

10.6 Fourier Series 290

*10.7 Orthogonalization of Moments 293
10.8 The Conjugate Gradient Method 294
METHODS FOR SOLVING CONSTRAINED
PROBLEMS 297

10.9
10.10
10.11
10.12

Projection Methods 297

The Primal..Dual Method 299
Penalty Functions 302
Problems 308
References 311

SYMBOL INDEX

321

SUBJECT INDEX

323

xiii

NOTATION

Sets

If x is a member of the set S, we write xeS. The notation y S means y is

not a member of S.
A set may be specified by listing its elements between braces such as
S = {l, 2, 3} for the slet consisting of the first three positive integers.
Alternatively, a set S may be specified as consisting of all elements of the
set X which have the .property P. This is written S = {x e X: P(x}} or, if
X is understood, S = {x: P(x}}.
The union of two sets Sand T is denoted S u T and consists of those
elements that are in either S or T.
The intersection of two sets Sand T is denoted S ('\ T and consists of
those elements that are in both Sand T. Two sets are disjoint if their
intersection is empty.
If S is defined as a subset of elements of X, the complement of S, denoted

S, consists of those elements of X that are not in S.

A set S is a subset of the set T if every element of S is also an element of
T. In this case we write: SeT or T::> S. If SeT and S is not equal to T
then S is said to be a proper subset of T.

Sets of Real Numbers

If a and b are real numbers, [a, b] denotes the set of real numbers x satisfying a:::;; X :::;; h. A rounded instead of square bracket denotes strict
inequality in the definition. Thus (a, bJ denotes all x with a < x :::;; h.
If S is a set of real numbers bounded ahove, then there is a smallest real
number y such that x y for all XES. 1 he number y is called the least
upper bound or supremum of S and is denoted sup (x) or sup {x: XES}
If S is not bounded above we write sup (x)
xeS

xeS

= 00.

Similarly, the greatest

lower bound or infimum of a set S is denoted inf (x) or inf {x:

xeS

xv

XES}.

xvi

NOTATION

Sequences
A sequence Xl, X 2 ,
the indices is clear.

.. , X., .

is denoted by {Xi}?"! or {Xi} if the range of

Let {Xi} be an infinite sequence of real numbers and suppose that there
is a real number S satisfying: (I) for every E > 0 there is an N such that for
all n > N, Xn < S + E, and (2) for every E > 0 and M > 0 there is an n > M
such that Xn > S - E. Then S is called the limit superior of{xn } and we write
S = lim sup X n . If {x.} is not bounded above we write lim sup Xn = + 00.
n'" 00

The limit inferior of Xn is lim inf Xn

inf Xn = S, we write lim Xn = S.

=-

lim sup ( - x n ). If lim sup

= lim

Xn

Functions
The function sgn (pronounced sig-num) of a real variable is defined by
X>o

x=O
x<O

sgn (x) = (
-1

The Kronecker delta function

oij is defined by ..
i=j
i=j

The Dirac delta function {) is used occasionally in heuristic discussions.

It is defined by the relation

f /(t)1;(t) dt =/(0)
b

for every continuous function/provided that 0

Ca, b).

If 9 is a real-valued fUnction of a real variable we write S

= lim sup g(x)

X-"Xo

if: (1) for every s> 0 there is 0 > 0 such that for all X satisfying Ix - xol
< 0, g(x) < S + E, and (2) for every E > 0 and 0 > 0 there is anxsuchthat
Ix - xol
and g(x) > S - E. (See the corresponding definitions for
sequences.)

<"

If 9 is a real-valued function of a real variable, the notation g(x)

means that
.
Ig(X)
K=hmsup
X'" 0

is finite. The notation g(x) = o(x) means that K, above, is zero.

=O(x)

NOTATION

xvii

Matrices and Vedors

A vector X with n eomponentsis written x = (Xl' X;p ... , xn), but when used
in matrix calculations it is represented as a column vector, i.e.,

The corresponding row vector is

X' = Xl X2 .. Xn
,

An n x m matrilx A with entry ail in its i-th row andj-th column is written
A = [aij]' If X = (Xl' X2, ... , xn), the product Ax is the vector y with components Yt = Li"'l al r i , i = 1, 2, ... , m.
Letj(xl' X2, ... ,. xn) be a function of the n real variables Xi' Then we
write I" for the row vector
oj

oj

oj

ax;' OX2'"'' Oxn ',

If F = (fl , f2' ... ,1m) is a vector function of
F" for the m x n Jacobian matrix [ofi/oxj]'

(Xl, ,

xn), we write

1
INTRODtrCTION

1.1 Motivation
During the past twenty years mathematics and engineering have been
increasingly directed towards
IS of decision making in physical or
organizational systems. This trend has been inspired primarily by the
significant economic benefits which often result from a proper decision
concerning the distribution of expensive resources, and by the repeated
demonstration that such problems can be realistically formulated and
mathematically aMlyzed to obtain good decisions.
The arrival of high-speed digital computers has also played a major
role in the development of the science of decision making. Computers
have inspired the development of larger systems and the coupling of
previously separate systems, thereby resulting in decision and control
problems of correspondingly increased complexity. At the same time,
have revolutionized applied mathematics and solved
however,
many of the complex problems they generated.
It is perhaps natural that the concept of best or optimal decisions should
emerge as the fundamental approach for formulating decision problems.
In this approach Ii single real quantity, summarizing the performance or
value of a decision, is isolated and optimized (Le., either maximized or
minimized depending on the situation) by proper selection among available
alternatives. The resulting optimal decision is taken as the solution to the
decision problem. This approach to decision problems has the virtues of
simplicity, precisell1ess, elegance, and, in many cases, mathematical tractability. It also has obvious limitations due to the necessity of selecting a
single objective by which to measure results. But optimization has proved
its utility as a mode of analysis and is firmly entrenched in the field of
decision making.
_Nllcll...of the classical theory of optimization, motivated primarily by
problems
of physics, is associated with great I mathematicians: Gauss,
.
..
Lagrange, Euler, the Bernoulis, etc. During the recent development of
optimization in decision problems, the classical techniques have been reo
examined, extended, sometimes rediscovered, and applied to problems

INTRODUCTION

having quite different origins than those responsible for their earlier
development. New insights have been obtained and new techniques have
been discovered. The computer has rendered many techniques obsolete
while making other previously impractical methods feasible and efficient.
These recent developments in optimization have been made by mathematicians, system engineers, economists, operations researchers, statisticians, numerical analysts, and others in a host of different fields.
The study of optimization as an independent topic must, of course, be
regarded as a branch of applied mathematics. As such it must look to
various areas of pure mathematics for its unification, clarification, and
general foundation. One such area of particular relevance is functional
analysis.
Functional analysis is the study of vector spaces resulting from a
merging of geometry, linear algebra, and analysis. It serves as a basis for
aspects of several important branches of applied mathematics including
Fourier series, integral and differential equations, numerical analysis, and
any field where linearity plays a key role. Its appeal as a unifying discipline
stems primarily from its geometric character. Most"Of the principal results
in functional analysis are expressed as abstractions of intuitive geometric
properties of ordinary three-dimensional space.
Some readers may look with great expectation toward functional
analysis, hoping to discover new powerful techniques that will enable them
to solve important problems beyond the reach of simpler mathematical
analysis. Such hopes are rarely realized in practice. The primary utility
of functional analysis for the purposes of this book is its role as a unifying
discipline, gathering a number of apparently diverse, specialized mathematical tricks into one or a few general geometric principles.

1.2 Applications
The main purpose of this section is to illustrate the variety of problems
that can be formulated as optimization problems in vector space by introducing some specific examples that are treated in later chapters. As a
vehicle for this purpose, we classify optimization problems according to
the role of the decision maker. We list the classification, briefly describe
its meaning, and illustrate it with one problem that can be formulated in
vector space and treated by the methods described later in the book. The
classification is not intended to be necessarily complete nor, for that matter,
particularly significant. It is merely representative of the classifications
often employed when discussing optimization.
Although the formal definition of a vector space is not given until
Chapter 2, we point out, in the examples that follow, how each problem

1.2

APPLICATIONS

can be regarded as formulated in some apprc . . . riate vector space. However,

the details of the formulation must, in many cases, be deferred until later
chapters.

1. Allocation. In aIle,cation problems there is typically a collection of

resources to be distributed in some optimal fashion. Almost any optimization problem can be placed in this broad category, but usually the term
is reserved for problems in which the resources are distributed over space
or among various activlities.
A typical problem of this type is that faced by a manufacturer with an
inventory of raw materials. He has certain processing equipment capable
of producing n different kinds of goods from the raw materials. His
problem is to allocate the raw materials among the possible products so
as to maximize his profitt.
In an idealized version of the problem, we aSsume that the production
and profit model is linear. Assume that the selling price per unit of product
j is PJ' j = 1, 2, ... , n. If x J denotes the amount of product j that is to be
produced, b i the amount of raw material i on hand, and aij the amount of
material i in one unit of product j, the manufacturer seeks to maximize
his profit
.PIXI

+ P2 X2 + ... + PnXn

subject to the production constraints on the amount of raw materials

allx1

+ a 12 x 2 + ... + a1nXn

bl

a21 x I

+ a 2 2 x 2 + ... + a 2n X n

b2

amlx1

+ a m2 X 2 + ... + amnXn

bm '

...

and

This type of problem, characterized by a linear objective function subject

to linear inequality constraints, is a linear programming problem and is
used to illustrate aspects of the general theory of optimization in later
chapters.
We note that the problem can be regarded as formulated in ordinary
n-dimensional vector space. The vector x with components Xi is the unknown. The constraints define a region h_ the vector space in which the
in that region
selected vector must lie. The optimal vector is the
maximizing the
.
The manufacturing problem can be generalized to allow for nonlinear

INTRODUCTION

objectives and more general constraints. Linearity is destroyed, which may

make the solution more difficult to obtain, but the problem can still be
regarded as one in ordinary Euclidean n-dimensional space.
2. Planning. Planning is the problem of determining an optimal procedure for attaining a set of objectives. In common usage, planning refers
especially to those problems involving outlays of capital over a period of
time such as (1) planning a future investment in electric power generation
equipment for a given geographic region or (2) determining the best hiring
policy in order to complete a complex project at minimum expense.
As an example, consider a problem of production planning. A firm
producing a certain product wishes to plan its production schedule over a
period of time in an optimal fashion. It is assumed that a fixed demand
function over the time interval is known and that this demand must be met.
Excess inventory must be stored at a storage cost proportional to the
amount stored. There is a production cost associated with a given rate of
production. Thus, denoting x(t) as the stock held at time t, ret) as the rate
of production at time t, and d(t) as the demand at time t, the production
system can be described by the equations!
x(t) = r(t) - d(t),

x(O)

given

and one seeks the function r satisfying the inequality constraints

r(t) ;;:::
x(O)

[r(r) - d(r)] clr = X(/) ;;:::

0,]

for 0::::; t ::; T

and minimizing the cost

.T

J=

10 {c[r(t)] + h . x(t)} dt

where c[r] is the production cost rate for the production level rand h . x
is the inventory cost rate for inventory level x.
This problem can be regarded as defined on a vector spffCe consisting of
continuous functions on the interval [0, T] of the real line. The optimal

production schedule r is then an element of the space. Again the constraints define a region in the space in which the solution r must lie while
minimizing the cost.
3. Control (or Guidance). Problems of control are asso.::iated with dynamic systems evolving in time. Control is quite similar to planning;
1

x(t)

== dx(t)fdt.

1.2

APPLICATIONS

indeed, as we shall see, it is often the source of a problem rather than its
mathematical structure which determines its category.
Control or guidance usually refers to directed influence on a dynamic
system to achieve desired performance. The system itself may be physical
in nature, such as a rocket heading for Mars or a chemical plant processing
acid, or it may be operational such as a warehouse receiving and filling
orders.
Often we seek fee:dback or so-called closed-loop control in which decisions of current control action are made continuously in time based on
recent observations of system behavior. Thus, one may imagine himself as
a controller sitting at the control panel watching meters and turning knobs
or in a warehouse ordering new stock based on inventory and predicted
demand. This is in contrast to the approach described for planning in
which the whole series of control actions is predetermined. Generally,
however, the terms planning or control may refer to either possibility.
As an example of a control problem, we consider the launch of a rocket
to a fixed altitude h in given time T while expending a minimum of fuel.
For simplicity, we assume unit mass, a constant gravitational force, and
the absence of aerodynamic forces. The motion of a rocket being propelled
vertically is governed by the equations
ji(t) = u(t) - 9

where y is the vertical height, u is the accelerating force, and 9 is the

gravitational force. The optimal control function u is the one which
forces yeT) = h while minimizing the fuel expenditure
lu(t)1 dt.
This problem too might be formulated in a vector space consisting of
functions u defined on the interval [0, .J. J. The solution to this problem,
however, is that u(t) consists of an impulse at t = 0 lnd, therefore, correct
problem formulation and selection of an appropriate vector space are
themselves interestilllg aspects of this example. Problems of this type, including this specific example, are dilicussed in Chapter 5.

JJ

4. Approximation!. Approximation problems are motivated by the desire

to approximate a gcmeral mathematical entity (such as a function) by one
of simpler, specified! form. A large class of such approximation problems
is important in numerical analysis. For example, suppose we wish, because
of storage limitatiol1ls or for purposes of simplifying an analysis, to approximate a function, say x(t), over an interval [a, b] of the real line by a polynomial p(t) of ord(:r n. The best approximating polynomial p minimizes
the error e = x - p in the sense of some criterion. The choice of criterion
determines the approximation. Often used criteria are:

INTRODUCTION
IJ

1: JeZ(t) dt
a

2.

max le(t)1

a,;;r,;;b
b

3. fle(t)1 dt.
a

The problem is quite naturally viewed as formulated in a vector space

of functions over the interval [a, b]. The problem is then viewed as finding
a vector from a given class (polynomials) which is closest to a given vector.

5. Estimation. Estimation problems are really a special class of approximation problems. We seek to. estimate some quantity from imperfect
observations of it or from observations which are statistically correlated
but not deterministically related to it. Loosely speaking, the problem
amounts to approximating the unobservable quantity by a combination of
the observable ones. For example, the position of a random maneuvering
airplane at some future time might reasonably be estimated by a linear
combination of past measurements of its position.
Another example of estimatio'n arises in connection with triangulation
problems such as in location of forest fires, ships at sea, or remote stars.
Suppose there are three lookout stations, each of which measures the angle
of the line-of-sight from the station to the observed object. The situation is
illustrated in Figure 1.1. Given these three angles, what is the best estimate
of the object's location?

Figure 1.1

A triangulation problem

To formulate the problem completely, a criterion must be precisely

prescribed and hypotheses specified regarding the nature of probable

1.2

APPLICATIONS

measurement errors and probable location of the object. Approaches can

in vector space; such probbe taken that-result.in a problem
lems are discussed in Chapter 4.

6. Games. Many problems involving a competitive element can be

regarded as games. In the usual formulation, involving two players or
protagonists, there is an objective function whose value depends jointly on
the action employed by both players. One player attempts to maximize
this objective while the other attempts to minimize it.
Often two problems from the categories discussed above can be competitively intermixed to produce a game. Combinations of categories that
lead to interesting games. include: allocation-allocation, allocation-control,
control-control, and estimation-control.
As an example, consider a
game. Most problems of this
type are of the pursuer-evader type such as a fighter plane chasing a
bomber. Each player has a system he
but one is trying to maximize
the objective (time to intercept for instance) while the other is trying to
minimize the objective.
As a simpler example, we consider aproblem of advertising or campaigning which is essentially an allocation-allocation game. 2 Two opposing
candidates, A and B, am running for office and must plan how to allocate
their advertising resources (A and B dollars, respectively) among n distinct
geographical areas. Let Xj and y{ denote, respectively, the resources committed to area i by candidates A and B. We assume that there are currently
a total of u undecided votes of which there are U j undecided votes in area i.
The number, of votes going to candidates A and B from area'i are assumed
to be
X{Uj

---,
XI+

Y{

respectively. The total difference between the number of votes received by

A and by B is' then
XI-YI

L." - - -

1= 1 ;'(j'"

Yj

UI'

Candidate A seeks to maximize this quantity while B seeks to minimize it.

This problem is obviously finite dimensional and can be solved by
ordinary calculus in a fe:w lines. It is illustrative, however, of an interesting
class of game problems .
2

This problem is due to L. Friedman [57J.

INTRODUCTION

1.1 The Main PrInciples

The theory of optimization presented in this book is derived from a few
simple, intuitive, geometric relations. The extension of these relations to
infinite-dimensional spaces is the motivation for the mathematics of
functional analysis which, in a sense, often enables us to extend our threedimensional geometric insights to complex infinite-dimensional problems.
This is the conceptual utility of functional analysis. On the other hand,
these simple geometric relations have great practical utility as well because
a vast assortment of problems can be analyzed from this point of view.
In this section, we briefly describe a few of the important geometric
principles of optimization that are developed in detail in later chapters.

1. The Projection Theorem. This theorem is one of the simplest and

nicest results of optimization theory. In ordinary three-dimensional
Euclidean space, it states that the shortest line from a point to a plane is
furnished by the perpendicular from the point to the plane, as illustrated
in Figure 1.2.

Figure 1.2

The projection theorem

This simple and seemingly innocuous result has direct extensions in

spaces of higher dimension and in infinite-dimensional Hilbert space. In
the generalized form, this optimization principle forms the basis of all
least-squares approximation, control, and estimation procedures.

2. The Hahn-Banach Theorem. Of the many results and concepts in

functional analysis, the one theorem dominating the theme of this book
and embodying the essence of the simple geometric ideas upon which
the theory is built is the Hahn-Banach theorem. The theorem takes several
forms. One version extends the projection theorem to problems having
nonquadratic objectives. In this manner the simple geometric interpretation
is preserved for these more complex problems. Another version of the

1.3

THE MAIN PRINCIPLES

Hahn-Banach theorem states (in simplest form) that given a sphere and a
point not in the sphere there is a hyperplane separating the point and the
sphere; This version of the theorem, together with the associated notions
of hyperplanes, is the basis for most of the theory beyond Chapter 5.

3. Duality. There are several duality principles in optimization theory

that relate a probllem expressed in terms of vectors in a space to a problem
expressed in terms of hyperplanes in the space. This concept of duality is
a recurring theme in this book.
Many of these duality principles are based on the geometric relation
illustrated in Figure 1.3. The shortest distance from a point to a convex
set is equal to the: maximum of the distances from the point to a hyperplane separating the point from the convex set. Thus, the original minimization over vectors can be converted to maximization over hyperplanes.

Figure 1.3

Duality

4. Differentials. Perhaps the most familiar optimization technique is

the method of dHferential calculus-setting the derivative of the objective
function equal to zero. The technique is discussed for a single or, perhaps,
finite number of variables in the most elementary courses on differential
calculus. Its extension to infinite-dimensional spaces is straightforward and,
in that form, it can be applied to a variety of interesting optimization problems. Much of the classical theory of the calculus of variations can be
viewed as a consequence of this principle.
The geometric interpretation of the technigue for one-dimensional
problems is obviQus. At a maximum or minimum the tangent to the graph
of a function is horizontal. In higher dimensions the geometric interpretation is similar: at a maximum or mmimum the tangent hyperplane to the

10

INTRODUCTION

graph is horizontal. Thus, again we are led to observe the fundamental

role of hyperplanes in optimization.
1.4

Organization of the Book

Before our discussion of optimization can begin in earnest, certain fundamental concepts and results of linear vector space theory must be introduced. Chapter 2 is devoted to that task. The chapter consists of material
that is standard, elementary functional analysis background and is essential for further pursuit of our objectives. Anyone having some familiarity
with linear algebra and analysis should have little difficulty with this
chapter.
Chapters 3 and 4 are devoted to the projection theorem in Hilbert space
and its applications. Chapter 3 develops the general theory, illustrating it
with some applications from Fourier approximation and optimal control
theory. Chapter 4 deals solely with,the applications of the projection
theorem to estimation problems including the recursive estimation and
prediction of time series as developed by Kalman.
Chapter 5 is devoted to the Hahn-Banach theorem. It is in this chapter
that we meet with full force the essential ingredients of the general theory
of optimization: hyperplanes, duality, and convexity.
Chapter 6 discusses linear transformations on a vector space and is the
last chapter devoted to the elements of linear functional analysis. The
concept of duality is pursued in this chapter through the introduction of
adjoint transformations and their relation to minimum norm problems.
The pseudoinverse of an operator in Hilbert space is discussed.
Chapters 7, 8, and 9 consider general optimization problems in linear
spaces. Two basic approaches, the local theory leading to differential
conditions and the global theory relying on convexity, are isolated and
discussed in a parallel fashion. The techniques in these chapters are a direct
outgrowth of the principles of earlier chapters, and geometric visualization
is stressed wherever possible. In the course of the development, we treat
problems from the calculus of variations, the Fenchel conjugate function
theory, Lagrange multipliers, the Kuhn-Tucker theorem, and Pontryagin's
maximum principle for optimal control problems.
Finally, Chapter 10 contains an introduction to iterative techniques for
the solution of optimization problems. Some techniques in this chapter
are quite different than those in previous chapters, but many are based on
extensions of the same logic and geometrical considerations found to be so
fruitful throughout the book. The methods discussed include successive
approximation, Newton's method, steepest descent, conjugate gradients,
the primal-dual method, and penalty functions.

LINEAR SPACES

2.1 Introduction
The first few sections of this chapter introduce the concept of a vector
space and explon: the elementary properties resulting from the basic
definition. The notions of subspace, linear independence, convexity, and
dimension are developed and illustrated by examples. The material is
largely review for most readers since it duplicates the first part of standard
courses in linear algebra.
The second part of the chapter discusses the basic properties oLnormed
linear spaces. A normed linear space is a vector space having a measure of
distance or length defined on it. With the introduction of a norm, it
becomes possible to define analytical or topological properties such as
convergence and open and closed setS. Therefore, that portion of the
chapter introduces and explores these basic concepts which distinguish
functional analysis. from linear algebra.
VECTOR SPACES

2.2 Definition and Examples

Associated with every vector space is a set of scalars used to define scalar
multiplication on the space. In the most abstract Jetting these scalars are
required only to he elements of an algebraic field. However, in this book
the scalars are always taken to be either the set of real numbers or of
complex numbers. We sometimes distinguish between these possibilities
by referring to a vector space as either a real or a complex vector space. In
this book, however, the primary emphasis is on real vector spaces and,
although occasional reference is made to complex spaces, many results
are derived only for real spaces. In case of ambiguity, the reader should
assume the space to be real.
Definition. A vector space X is a set of elements called vectors together
with two operations. The first operation is addition which associates with

11

12

LINEAR SPACES

any two vectors X, y E X a vector X +Y E X, the sum of X and y. The

second operation is scalar multiplication which associates with any vector
x E X and any scalar IX a vector IXX; the scalar multiple of x by IX. The set X
and the operations of addition and scalar mUltiplication are assumed to
satisfy the foIlowing axioms:
1.
2.

3.
4.

5.
6.

7.

x + y = y + x.
(x + y) + z = x + (y + z).
There is a null vector 0 in X such
that x + 0 = x for all x in X.
IX(X + y) = IXX + lXy. }
(IX + f3)x = IXX + [3x.
(1X[3)X = 1X([3X).
Ox=O,
1x=x.

(commutative law)
(associative law)

(distributive laws)
(associative law)

For convenience the vector -1 x is denoted -x and called the negative

of the vector x. We have oX + (-x) = (1 - l)x = Ox = O.
There are several elementary but important properties of vector spaces
that foIlow directly from the axioms listed in the definition. For example
the following properties are easily deduced. The details are left to the
reader.

Proposition 1. III any vector space:

1.

2.
3.
4.
5.
6.

= x + z implies y = z.
= lXy and IX t= 0 imply x
y.
IXX = [3x and x t= 0 imply IX = [3.
(IX - [3)x = IXX - [3x. }
IX(X - y) = IXX - lXy.
IXO = O.

+y

IXX

(cancellation, laws)
(distributive laws)

Some additional properties are given as exercises at the end of the

chapter.

Example 1. Perhaps the simplest example of a vector space is the set of

real numbers. It is a real vector space with addition defined in the usual
way and multiplication by (real) scalars defined as ordinary multiplication.
The null vector is the real number zero. The properties of ordinary addition
and multiplication of real numbers satisfy the axioms in the definition of a
vector space. This vector space is called the one-dimensional real coordinate space or simply the real line. It is denoted by Rl or simply R.
Example 2. An obvious extension of Example 1 is to n-dimensional real
coordinate space. Vectors in the space consist of sequences (n-tupJe'J) of
n real numbers so that a typical vector has the form x =
... ,

'k

2.2

DEFINITION AND EXAMPLES

13

The real number is referred to as the k-th component of the vector. Two
vectors are equal if their corresponding components are equal. The null
vector is defined as e = (0, 0, ... ,0). If x = (el, e2, ... , en) and y =
(th, Yl2, ... , YIn), the vector x + y is defined as the n-tuple whose k-th
component is ek + Ylk' The vector ax, where a is a (real) scalar, is the
n-tuple whose k-th component is aek. The axioms in the definition are
verified by checking for equality among components. For example, if
x = (e 1, e2 , " ., en), the relation ek + 0 = ek implies x + = x.
This space, n-dimensional real coordinate space, is denoted by Rn. The
corresponding complex space consisting of n-tuples of complex numbers
is denoted by en.
At this point
are, strictly speaking, somewhat prematurely introducing the term dimensionality. Later in this chapter the notion of dimension is defined, and it is proved that these spaces are in fact n dimensional.

Example 3. Several interesting vector spaces can be constructed with

vectors consisting of infinite sequences of real numbers so that a typical
vector has the form x = (e 1 , e2, ... , ek> ... ) or, equivalently, x = {ek}
Again addition and multiplication.re defined componentwise as in
Example 2. The collection of all infinite sequences of real numbers forms
a vector space. A sequence {ek} is said to be bounded if there is a constant
M such that lekl -< M for all k. The collection of all bounded infinite
sequences forms a vector space since the sum of two bounded sequences
or the scalar multiple of a bounded sequence is again bounded. This space
is referred to as the space of bounded real sequences.
Example 4. The collection of all sequences of real numbers having only a
finite number of terms not equal to zero is a vector space. (Different
members of the spa<ee may have different numbers of nonzero components.)
This space is called the space of finitely nonzero sequences.

Example 5. The collection of infinite sequences of real numbers which

converge to zero is a vector space since the sum of two sequences converging to zero or the scalar multiple of a sequence converging to zero also
converges to zero.
Example 6. Consider the interval [a, b] on the real line. The collection of
all real-valued continuous functions on this interval forms a vector space.
Write x = y if x(t) := y(t) for all t E [a, b]. The null vector e is the function
identically zero on [a, b]. If x and yare vectors in the space and (J( is a
= (J(x(t). These are
(real) scalar, write (x + y)(t) = x(t) + y(t) and
obviously continuous functions. This space is referred to as the vector
space of real-valued continuous functions on [a, b].

14

LINEAR SPACES

Example 7. The collection of all polynomial functions defined on the

interval [a, b] with complex coefficients forms a complex vector space.
The null vector, addition, and scalar multiplication are defined as in
Example 6. The sum of two polynomials and the scalar multiple of a
polynomial are themselves polynomials.
We now consider how a set of vector spaces can be combined to produce
a larger one.

Definition. Let X and Y be vector spaces over the same field of scalars.
Then the Cartesian product of X and Y, denoted X x Y, consists of the
collection of ordered pairs (x, y) with x E X, Y E Y. Addition and scalar
multiplication are defined on X x Y by (Xl' Yl) + (X2' Y2) = (Xl + X2 ,
Yl + Y2) and (X(x, y) = (Xx, (Xy).
That the above definition is tonsistent with the axioms of a vector space
is obvious. The definition is easily generalized to the product of n vector
spaces Xl' X 2 , , X n. We write Xn for the product of a vector space
with itself n times.
2.3

Subspaces, Linear Combinations, and Linear Varieties

Definition. A nonempty subset M of a vector space X is called a subspace

of X if every vector of the form (Xx + f3y is in M whenever x and yare
both in M.
Since a subspace is assumed to be nonempty, it must contain at least
one element x. By definition, it must then also contain Ox = so every
subspace contains the null vector. The simplest subspace is the set consisting of alone. In three-dimensional space a plane that passes through
the origin is a subspace as is a line through the origin.
The entire space X is itself a subspace of X. A subspace not equal to the
entire space is said to be a proper subspace.
Any subspace contains sums and scalar multiples of its elements;
satisfaction of the seven axioms in the entire space implies that they are
satisfied within the subspace. Therefore a subspace is itself a vector space
and this observation justifies the terminology.
If X is the space of n-tuples, the set of n.tuples having the first com
ponent equal to zero is a subspace of X. The space of convergent infinite
sequences is a subspace of the vector space of bounded sequences. The
space of continuous functions on [0, 1] which are zero at the point one-

e,

half is a subspace of the vector space of continuous functions.

In three-dimensional space, the intersection of two distinct planes

2.3

SUBSPACES, LINEAR COMBINATIONS, AND LINEAR VARIETIES

15

containing the origin is a line containing the origin. This is a special case
of the following result
Proposition 1. Let M and N be subspaces of a nee tor space X. Then the
intersection, M n N, of M and N is a subspace of x.

Proof M n N contains 0 since 0 is contained in each of the subspaces

M and N. Therefore, M n N is nonemptj . If x and yare in M n N, they
are in both M and N. For any scalars IX, f3 the vector ('IX + f3y is contained
in both M and N since M and N are subspaces. Therefore, IXX + f3y is

contained in the intersection M n N.

The union of two subspaces in a vector space is not necessarily a subspace. In the plane, for example, the union of two (noncolinear) lines
through the origin does not contain arbitrary sums of its elements. However, two subspaces may be combined to form a larger subspace by introducing the notion of the sum of two sets.
Definition. The sum of two subsets Sand T in a vector space, denoted
S + T, consists of all vectors of the form s + t where s E Sand t E T.

The sum of two sets is illustrated in Figure 2.1.

Figure 2.1

The sum of two sets

Proposition 2. Let M and N be subspaces of a vector space X. Then their

sum, M

+ N, is a subspace of X.

Proof Clearly M + N contains

e.

Suppose x and yare vectors in

+ N. There are vectors m 1 , m z in M and vectors nl' nz in N such.that

16

LINEAR SPACES

x = tnl + nr,y
m 2 + nz. For any scalars a, p; ax + py
(aml + Pm2)
(anl + Pn2)' Therefore, ax + py can be expressed as the sum of a vector

in the subspace M and a vector in the subspace N.

1n two-dimensional Euclidean space the sum of two noncolinear lines

through the origin is the entire space. The set of even continuous functions
and the set of odd continuous functions are subspaces of the space X of
continuous functions on the real line. Their sum is the whole space X
since any continuous function can be expressed as a sum of an even and
an odd continuous function.
Definition. A linear combination of the vectors Xl' X 2 ,
space is a sum of the form alx l + a2x2 + ... + anxn .

, Xn

in a vector

Actually, addition Ihas been defined previously only for a sum of two
vectors. To form a sum consisting of n elements, the sum must be performed
two at a time. It follows from the axioms, however, that analogous to the
corresponding operations with real numbers, the result is independent of
the order of summation. There is thus no ambiguity in the simplified
notation.
It is apparent that a linear combination of vectors from a subspace is
also in the subspace. Conversely, linear combinations can be used to
construct a subspace from an arbitrary subset in a vector space,
Definition. Suppose S is a subset of a vector space X. The set [S], called
the subspace generated by S, consists of all vectors in X which are linear
combinations of vectors in S.

The verification that [S] is a subspace in X follows from the obvious

fact that a linear combination of linear combinations is also a linear
combination.
There is an interesting characterization of the subspace [S]. The set S
is, in general, wholly contained in a number of subspaces. Of these, the
subspace [S] is the smallest in the sense that if M is a subspace containing
S, then M contains [SJ. This statement is proved by noting that if the
subspace M contains S, it must contain all linear combinations from S.
In three-dimensional space the subspace generated by a two-dimensional
circle centered at the origin is a plane. The subspace generated by a plane
not passing through the origin is the Whole space. A subspace is a generalization of OUI intuitive notion of a plane or line through the origin. The
translation of a subspace, therefore, is a generalization of an arbitrary
plane or line.
Definition. The translation of a subspace is said to be a linear variety.

2.4

CONVEXITY AND CONES

17

A linear varietyl V can be written as V = Xo + M where M is a sub:space. In this representation the subspace M is unique, but any vector in
V can serve as xo. This is illustrated in Figure 2.2.
Given a subset S, we can construct the smallest linear variety containing S.
\M
\

\
\

\
8

Figure 2.2 A linear variety

Definition. Let S be a nonempty subset of a vector space X. The linear

variety generated by S, denoted v(S) is defined as the intersection of all

linear varieties in X that contain S.

We leave it to the reader to justify the above definition by showing that
v(S) is indeed a llinear variety.
2.4

Convexity and Cones

We come now to the topic that is responsible for a surprising number of

the results in this book and which generalizes many of the useful properties
of subspaces andllinear varieties.
Definition. A set K in a linear vector space is said to be convex if, given
Xl' X2

e K, all points of the form

/Xx!

+ (1

- ct)X2 with 0 S;

IX S;

1 are in K.

This definition merely says that given two points in a convex set, the line
segment between them is also in the set. Examples are shown in Figure 2.3
Note in particular that subspaces and linear varieties are convex. The
empty set is con!lidered convex.
1 Other names for a linear variety include: flat, affine subspace, and linear manifold.
The term linear manifold, although commonly used, is reserved by many authors as an
alternative term for subspace.

18

LINEAR SPACES

09
Convex

Nonconvex

Figure 2.3 Convex and nonconvex sets

The following relations for convex sets are elementary but important.
Their proofs are left to the reader.
Proposition 1. Let K and G be convex sets in a vector space. Then

1. rxK = {x: x = rxk, k

2. K + G is convex.

K} is convex for any scalar rx.

We also have the following elementary property.

Proposition 2. Let Ctf be an arbitrary collection of convex sets. Then
K E 'I K is convex.

nK

Proof Let C =
e'G K. If C is empty, the lemma is trivially proved.
Assume that Xl' X2 E C and select rx, 0::;; Or: ::;; 1. Then Xl' x 2 E K for all
K E Ctf, and since K is convex, rxXl + (1 - rx)X2 E K for all K E Ctf. Thus
rxx l + (1 - rx)X2 E C and C is convex. I

Definition. Let S be an arbitrary set in a linear vector space. The convex

cover or convex hull, denoted co(S) is the smallest convex set containing
S. In other words, co(S) is the intersection of all convex sets containing S.
Note that the justification of this definition rests with Proposition 2
since it guarantees the existence of a smallest convex set containing S.
Some examples of a set and its convex hull are illustrated in Figure 2.4.

Definition. A set C in a linear vector space is said to be a cone with vertex

at the origin if x E C implies that rxx E C for all rx O.

Figure 2.4 Convex hulls

2.S

LINEAR INDEPENDENCE AND DIMENSION

19

Several cases are shown in Figure 2.5. A cone with vertex p is defined as a
translation p + C of a cone C with vertex at the origin. If the vertex of a
cone is not e1{plicitly mentioned, it is assumed to be the origin. A convex
cone is, of course, defined as a set which is both convex and a cone. Of
the cones in Figure 2.5, only (b) is a convex cone. Cones again generalize
the concepts of subspace and linear variety since both of these are convex
cones.

(c)

(b)

(a)

Figure 2.5 Cones

Convex cones usually arise in connection with the definition of positive

vectors in a vector space. For instance, in n-dimensional real coordinate
space, the set
p

= {x : x =

gl'

... ,

all i},

defining the positive orthant, is a convex cone. Likewise, the set of all
nonnegative continuous functions is a convex cone in the vector space of
continuous functions.
2.5 Linear Independence and Dimension
In this section we first introduce the concept of linear independence, which
is important for any general study of vector space, and then review the
essential distinguishing features of finite-dimensional space: basis and
dimension.
Definition. A vector x is said to be linearly dependent upon a set S of
. vectors if x can be expressed as a linear combination of vectors from S.
Equivalently, x: is linearly dependent upon S if x is in [S], the subspace
generated by S. Conversely, the vector x is saJ.d to be linearly independent
of the set S if it is not linearly dependent on S; a set of vectors is said to
be a linearly independent set if each vector in the set is linearly independent
of the remaind,er of the set.

20

LINEAR SPACES

Thus, two vectors are linearly independent if they do not lie on a common line through the origin, three vectors are linearly independent if they
do not lie in a plane through the origin, etc. It follows from our definition
that the vector () is dependent on any given vector x since () = Ox. Also,
by convention, the sd consisting of () alone is understood to be a dependent
set. On the other hand, a set consisting of a single nonzero vector is an
independent set. With these conventions the following major test for linear
independence is applicable even to trivial sets consisting of a single vector.

Theorem 1. A necessary and sufficient condition for the set of vectors

Xl' X2 , . , Xn to be linearly independent is that the expression
1 (J.kXk = e
implie.\!J.k = Ofor all k = 1,2,3, ... , n.

Iz=

Proof To prove the necessity of the condition, assume that

Ik= 1 CXkXk = 0, and that for some index 1', CX r 1: O. Since the coefficient of
Xr is nonzero, the original relation may be rearranged to produce
Xr = Lk*r (-CXdCXr)Xk which shows xr to be linearly dependent on the
remaining vectors.
To prove the sufficiency of the condition, note that linear dependence
among the vectors implies that one vector, say X r , can be expressed as a
linear combination of the others, xr = Lk;tr Cf.kXk' Rearrangement gives
Lk*r (J.kXk - Xr = e which is the desired relation. I
An important consequence of this theorem is that a vector expressed as
a linear combination of linearly independent vectors can be so expressed
in only one way.

Corollary 1. If xl> x 2 , , XII are linearly independent vectors, and

Ik= 1 CXk Xk = Lk=' 1 13k Xb then (J.k = f3kfor all k = 1, 2, ... , n.
Proof. IfLk=1 (J.kXk = Lk=1 f3k Xb then D=1 (CXk - f3k)X k =
ak - 13k = 0 according to Theorem I. I

if

and

We turn now from the general notion of linear independence to the

special topic of finite dimensionality. Consider a set S of linearly independent vectors in a vector space. These vectors may be used to generate
[S], a certain subspace of the vector space. If the subspace [S] is actually
the entire vector space, every vector in the space is expressible as a linear
combination of vectors from the original set S. Furthermore, according
to the corollary to Theorem 1, the expression is unique.
Definition. A finite set S of linearly independent vectors is said to be a
basis for the space X if S generates X. A vector space having a finite basis
is said to be finite dimensional. All other vector spaces are said to be

infinite dimensional.

2.S

LINEAR INDEh..NDENCE AND DIMENSION

21

Usually, we charact(!rize a finite-dimensional space by the number of

elements in a basis. Thus, a space with a basis consisting of n elements
is referred to as
space. This practice would be undesirable
on grounds of ambiguity if the number of elements in a basis for a space
were not unique. The following theorem is, for this reason, a fundamental
result in the study of finite-dimensional spaces.

Theorem 2. Any two bases for a finite-dimensional vector space contain the
same number of elements.
Proof Suppose that {Xl' X2' ... , X.} and {Yl' Y2' ... , Ym} are bases
for a vector space X. Suppose also that m n. We shall substitute Y vectors
for X vectors one by one in the first basis until all the X vectors are replaced.
Since the x;'s form a basis, the vector Yl may be expressed as a linear
combination of them, say, Yl = II=l (XiXi' Since Yl::F e, at least one of
the scalars (Xi must be nonzero. Rearranging the x;'s if necessary, it may
be assumed that (Xl ::F O. Then Xl may be expressed as a linear combination
of Y1' x 2 , x 3 , , x. by the formula

The set Yl' x 2 , ... , x. generates X since any linear combination of the
original x/s becomes an equivalent linear combination of this new set
when Xl is replaced according to the above formula.
Suppose now that k - 1 of the vectors Xi have been replaced by the
first k - 1 y;'s. The vector Yk can be expressed as a linear combination of
Yl' Y2' ... , Yk-l, x k , ... , X n , say,
k-l
Yk

L (Xi Yi + i=k
L Pi Xi'
i= 1

Since the vectors Yl,)'2 , ... ,)'k are linearly independent, not all the P;'s
can be zero. Rearrangiing Xk, Xk+ l' ... , X. if necessary, it may be assumed
that Pk::F O. Then X k can be solved for as a linear combination of
Yl')'2, .. , Yk, X k + 1> , X., and this new set of vectors generates X.
By induction on k then, we can replace all n of the x;'s by y;'s, forming a
generating set at each step. This implies that the independent vectors
Yl')'2, ... , Yn generate X and hence forrr: a basis for X. Therefore, by the
assumption of linear independence of {Yl' Y2, ... , Ym}, we must have
n=m.

Finite-dimensional spaces are somewhat simpler to analyze t:lan

infinite-dimensional spaces. Fewer definitions are required, fewer pathological cases arise, and our native intuition is contradicted in fewer instances

22

LINEAR SPACES

in finite-dimensional spaces. FurthermQre, there are theorems and concepts in finite-dimensional space which have no direct counterpart in
infinite-dimensional spaces. Occasionally, the availability of these special
characteristics of finite dimensionality is essential to obtaining a solution
to a particular problem. It is more usual, however, that results first derived
for finite-dimensional spaces do have direct analogs in more general
spaces. In these cases, verification of the corresponding result in infinitedimensional space often enhances our understanding by indicating precisely which properties of the space are responsible for the result. We
constantly endeavor to stress the similarities between infinite- and finitedimensional spaces rather than the few minor differences.
NORMED LINEAR SPACES

2.6 Definition and Examples

The vector spaces of part\cular interest in both abstract analysis and applications have a good deal more structure than that implied solely by the
seven principal axioms. The vector space axioms only describe algebraic
properties of the elements of the space: addition, scalar multiplication,
and combinations of these. What are missing are the topological concepts
such as openness, closure, convergence, and completeness. These concepts
can be provided by the introduction of a measure of distance in a space.

DefinitiDn. A normed linear vector space is a vector space X on which

there is defined a real-valued function which maps each element x in X
into a real number Ilxll called the norm of x. The norm satisfies the following axioms:

1.
2.
3.

Ilxll;::: 0 for all x EX, Ilxll = 0 if and only if x = e.

Ilx + yll :s;; Ilxll + Ilyll for each x, y E X.
(triangle inequality)
Ilctxll = Ictl'llxll for all scalars ct and each x EX.

The norm is clearly an abstraction of our usual concept of length. The

following useful inequality is a Mrect consequence of the triangle inequality.
Lemma 1. In a normed linear space
vectors x, y.

Ilxll - Ilyll :::;; Ilx - yll

Proof.

Ilxll - Ilyl\

Ilx - y + yll - I\yll :s;; \\x - y\\ + \\y\\ - \\y\\

for any two

=\\x - y\\. I

By introduction of a suitable norm, many of our earlier examples of

vector spaces can be converted to nonned spaces.

2.6

DEFINITION AND EXAMPLES

23

Example 1. The

space C[a, b] consists of continuous functions on the real interval [a, b] together with the norm Ilxll = max Ix(t)l.
astSb

This space was considered as a vector space in Section 2.2. We now verify
that the proposed norm satisfies the three required axioms. Obviously,
Ilxll 0 and is zero only for the function which is identically zero. The
triangle inequality follows from the relation
max Ix(t)

+ y(t)1

max [lx(t)1

+ ly(t)l]

max Ix(t)1

+ max ly(t)l.

Finally, the third axiom follows from the relation

max

IlXx(t)I = max 1IXIIx(t)1 = loci max Ix(t)l

Example 2. The

linear space D[a, b] consists of all functions on

the interval [a, b]
are continuous and have continuous derivatives
on [a, b]. The norm on the space D[a, b] is defined as

Ilxll

max Ix(t)1
astsb

max Ix(t)l.
aStSb

We leave it to the reader to verify that D[a, b] is a normed linear space.

Example 3. The space: of finitely nonzero sequences together with the

norm equal to the sum of the absolute values of the nonzero components
0, 0, ... }
is a normed linear space. Thus the element x = {el> e2 , ,
has its norm defined as Ilxll = L'i=1 leil. We may easily verify the three
required properties by inspection.

en,

Example 6. The space: of continuous fl lctions on the interval [a, b]

becomes a normed space with the norm of a function x defined as
Ilxll = S! Ix(t)1 dt. This is a different normed space than C[a, b].
Example 5. Euclidean n-space, denoted En, consists of n-tuples with the
norm of an element x = {el> e2 ,., en} defined as IIxll = (L7=1 leiI 2)1/2.
This definition obviously satisfies the first and third axioms for norms. The
triangle inequality fot this norm is a well-known result from finitedimensional vector spaces and is a special case of the Minkowski inequality
discussed in Section 2.10. The space En can be chosen as.a real or complex
space by considering rleal or complex n-tuples. We employ the same notation En for both because it is generally apparent from context which is
meant.
I

Example 6. We consider now the space BV[a, b] consisting of functions

of bounded variation on the interval [a, b]. By a partition of the interval
[a, b], we mean a

a = to < 11 <

12

set of points ti E [a, b], i =0, 1,2, ... , n, such that

< ... :< 1n = b. A function x defined on [a, b] is said to

24

LINEAR SPACES

be of bounded variation
of [a, bJ

if there is a constant K so that for any partition

L,lx(t;) - x(tj_I)1

:$

K.

;=1

The total variation of x, denoted T.Y.(x), is then defined as

T.Y.(x)

= sup L Ix(t
j=

j) -

x(ti_I)1

where the supremum is taken with respect to all partitions of [a, bJ. A
convenient and suggestive notation for the total variation is

JIdx(t)l
b

T.Y.(x) =

The total variation of a constant function is zero and the total variation of
a monotonic function is the absolute value of the difference between the
function values at the end points a and b.
The space BV[a, bJ,is defined as the space of all functions of bounded
variation on [a, bJ together with the norm defined as
Ilxll

2.7

= Ix(a)1 +T.Y.(x).

Open and Closed Sets

We come now to the concepts that are' . Jundamental

to the study of
<,
topological properties.

Definition. Let P be a subset of a normed space X. The point PEP is

said to be an interior point of P if there is an 6 > 0 such that all vectors x
satisfying Ilx - pil < 6 are also members of P. The collection of all interior points of P is called the interior of P and is denoted P.
We introduce the notation S(X,6) for the (open) sphere centered at x
with radius 6; that is, sex, 6) = {y : Ilx - yll < S}. Thus, according to the
above definition, a point x is an interior point of P if there is a sphere
sex, 6) centered at x and contained in P. A set may have an empty interior
as, for example, a set consisting of a single point or a line in E2.

Definition. A set P is said to be open if P =P.

The empty set is open since its interior is also empty. The entire space
is an open set. The unit sphere consisting of all vectors x with Ilxll < 1 is
an open set. We leave it to the reader to verify that P is open.

2.7

Definition. A point x

OPEN AND CLOSED SETS

25

X is said to be a closure point of a set P if, given

> 0, there is a point PEP satisfying Ilx - pll <

B. The collection of all

closure points of P is called the closure of P and is denoted P.
B

In other words, a point x is a closure point of P if every sphere centered

at x contains a point of P. It is clear that P c:: P.
Definition. A set P is said to be closed if P =

P.

The empty set and the whole space X are closed as well as open. The
unit sphere consisting of all points x with Ilxli =:;; 1 is a closed set. A single
point is a closed set. It is clear that P = P.

Proposition 1. The complement of an open set is closed and the complement

of a closed set is open.
Proof Let P be an open set and 15 = {x: x P} its complement. A
point x in P is not a closure point of 15 since there is a sphere about x
disjoint from P. Thus 15 contains all its closure pomts and is therefore
closed.
Let S be a closed set. If XES, then x is not a closure point of Sand,
hence, there is a sphere about x which is disjoint from S. Therefore x is
an interior point of S. We conclude that S is open. I

The proofs of the following two complementary results are left to the
reader.

Proposition 2. The intersection of a finite number of open sets is open; the

union of an arbitrary collection of open sets is open.
Proposition 3. The union of a finite number of closed sets is closed; the intersection of an arbitrary collection of closed sets is closed.

We now have two topological operations, taking closures and taking

interiors, that can
applied to sets in normed space. It is natural to
investigate the effect of these operations on convexity, the fundamental
algebraic concept of vector space.

Proposition 4. Let C be a convex set in a normed space. Then E and Care

convex.
Proof If C is empty, it is convex. Suppose xo, Yo are points in C.
Fix ex, 0 < ex < 1. We must show that Zo = exxo + (1 - <X)Yo e C. Given
e> 0, let x, y be selected from C such that IIx - II < 8, Ily -Yoll < B.
Then lIax + (1 - a)y - axo - (1 - a)Yoll =:;; e and, hence, Zo is within a
distance B of z == (J.X + (1 - a)y which is in C. Since e is arbitrary, it
follows that Zo is a closure point of C.

26

LINEAR SPACES

If C is empty, it is convex. Suppose x o, Yo E (; and fix a, < a < l. We

must show that Zo = axo + (I - a)yo E C. Since x o , Yo E C, there is an
e > 0 such that the open spheres S(x o , e), S(Yo, e) are contained in C. It
follows that all points of the form Zo + w with IIwll < e are in C since
Zo + w = a(xo + w) + (I - a)(yo + w). Thus, Zo is an interior point of C. I

Likewise, it can be shown that taking closure preserves subspaces,

linear varieties, and cones.
Finally, we remark that all of the topological concepts discussed above
can be defined relative to a given linear variety. Suppose that P is a set
contained in a linear variety V. We say that p E, P is an interior point of P
relative to V if there is an e > 0 such that all vectors x E V satisfying
Ilx - pil < e are also members of P. The set P is said to be open relative
to V if every point in P is an interior point of l' relative to V.
In case V is taken as the closed linear variety generated by P, i.e., the
intersection of all closed linear varieties containing P, then x is
referred to as a relatice interior point of P if it is an interior point of P
relative to thtt variety V. Similar meaning is given to relatively closed, etc.

2.8

Convergence

In order to prove the existence of a vector satisfying a desired property,

it is common to establish an appropriate sequence of vectors converging
to a limit. In many cases the limit can be shown to satisfy the required
property. It is for this reason that the concept of convergence plays an
important role in analysis.

Definition. In a normed linear space an infinite sequence of vectors {x"}

is said to cont'erge to a vector x if the sequence {llx - x"ll} of real numbers
converges to zero. In this case, we write x" x.
If x" x, it follows that Ilx"1I
Ilxll because, according to Lemma 1,
Section 2.6, we have both Ilx"11 - IIxll ::; IIx" - xII and IIxll - IIx"" ::;
IIx" - xII which implies that \Ix" II - lixlll ::; IIx. - xii -+ O.
In the space E" of n-tuples, a sequence converges if and only if each
component converges; however, in other spaces convergence is not always
easy to characterize. Tn the space of finitely nonzero sequences, define the
vectors e{ = {O, 0, ... , 1,0, ... }, the i-th vector having each of its components zero except the i-th which is I. The sequence of vectors {ei}
(which is now a sequence of sequences) converges to zero componentwise,
but the sequence does not converge to the nulll vector since Ileill = I for
.
all i.
An important observation is stated in the following proposition.

2.9

TRANSFORMATIONS AND CONTINUITY

27

Proposition 1. If a sequence converges, its limit is unique.

Proof Suppose xn - t x and xn - t y.

yll
Thus, x =y. I
IIx -

= IIx -

Xn

+ Xn -

YII:5: I!x - xnll

+ Ilxn -

yll-t O.

Another way to state the definition of convergence is in terms of spheres.

A sequence {xn} converges to x if and only if given e > 0; the sphere
S(x, e) contains Xn for all n greater than some number N.
The definition of convergence can be used to characterize closed sets
and provides a useful alternative to the original definition of closed sets.
Proposition 2. A set F is closed if and only if every convergent sequence
with elements in F has its limit in F.

Proof The limit of a sequence from F is obviously a closure point of

F and, therefore, must be contained in F if F is closed. Suppose now that F
is not closed. Then there is a closure point x of F that is not in F. In each
of the spheres S(x, lin) we may select a point xn e F since x is a closure
point. The sequence {xn} generated in this way converges to x F. I

2.9 Transformations and Continuity

The objects that make the study of linear spaces interesting and useful
are transformatiorts.

Definition. Let X and Y be linear vector spaces and let D be a subset of

X. A rule which associates with every element XED an element y E Y
is said to be a transformation from X to Y with domain D. If y corresponds
to x under T, we write y = T(x).
Transformations on vector spaces become increasingly more important
as we progress through this book; they are treated in some detail beginning
with Chapter 6. It is the purpose of this section to introduce some common
terminology that is convenient for describing the simple transformations
encountered in the early chapters.
Ifa specific domain is not explicitly mentioned when discussing a
transformation on a vector space X, it is understood that the domain is
X itself. If for every y e Y there is at most one xeD for which T(x) = y,
the transformation T is said to be one-to-one. If for every y e Y there is at
least one xeD for which T(x) = y, T is said to be onto or, more precisely,
to map D onto Y. This terminology, as the notidn of a transformation
itself, is of course an extension of the familiar rtotion of ordinary functions.
A transformation is simply afunction c...;;fined on one vector space X while

28

LINEAR SPACES

taking values in another vector space Y. A special case of this situation is

that in which the space Y is taken to be the real line.
Definition. A transformation from a vector space X into the space of real
(or complex) scalars is said to be a functional on X.
In order to distinguish functionals from more general transformations,
they are usually denoted by lower case letters such asfand g. Hence,f(x)
denotes the scalar that f associates with the vector x E X.
On a normed space, f(x) = IIxll is an example of a functional. On the
space C[O, IJ, examples of functionals aref,(x) = X(-l),J2(X) = fbx(t)dt,
f3 (x) = max X4 (t), etc. Real-valued functionals are of direct interest
OS!SI

to optimization theory, since optimization consists of selecting a vector

in minimize (or maximize) a prescribed functional.
Definition. A transformation T mapping a vector space X into a vector
space Y is said to be linear if for every Xl' X2 E X and all scalars CL I , (.(2 we
have T(O'.lxI + 0'.2 X2 ) = 0'.1 T(x l ) + 0'.2 T(x 2 )
The most familiar example of a linear transformation is supplied by a
rectangulal'm x n matrix mapping elements of Rn into Rm. An example
of a linear transformation mapping X = C [a, b] into X is the integral
operator T(x) =
k(t, T)x('r) dT where k(t, T) is a function continuous on
the square a s; t s; a S; t S; b.
Up to this point we have considered transformations mapping one
abstract space into another. If these spaces happen to be normed, it is
possible to define the notion of continuity.
Definition. A transformation T mapping a normed space X into a normed
space Y is continuous at Xo E X if for every e > 0 there is a f> > 0 such that
Ilx - xoll < f> implies that IIT(x) - T(xo)11 < e.
Note that continuity depends on the norm in both the spaces X and Y.
If T is continuous at each point Xo E X, we say that T is continuous everywhere or, more simply, that T is continuous.
The following characterization of continuity is useful in many proofs
involving continuous transformations.
Proposition 1. A transformation T mapping a normed space X into a normed
space Y is continuous at the point Xo E X if and only if xn --+ Xo implies
T(xn) --+ T(xo).

Proof. The" if" portion of the statement is obvious; thus we need

only proofthe "only if" portion. Let {x n} be a sequence such that Xn - Xo,
T(x n) T(xo). Then, for some Il > 0 and every N there is an n > N such

2.1O

THE

Ip AND Lp SPACES

29

that IIT(xn) - T(xo)1I 8. Since xn -+ xo, this implies that for every 8 > 0
there is a point Xn with Ilxn - xoll < 8 and IIT(xn) - T(x)1I > s. This proves
the" only if" portion by contraposition. I
*2.10 The'.. and L.. Spaces
In this section we discuss some classical normed spaces that are useful
throughout the book.

Definition. Let p be a real number I P < oo.The space Ip consists of all

sequences of scalars
for which
<Xl

L leil

i= 1

The norm of an element x

<

00.

Ip is defined as

The space l<Xl consists of bounded sequences. The norm of an element

x=
in 1<Xl is defined as

It is obvious that the norm on Ip satisfies lIexxll = lexlllxll and that

In this section, we establish two inequalities concerning the Ip norms, the second of which gives the triangle inequality for
these Ip norms. Therefore, the Ip riorm indeed satisfies the three axioms
required of a general norm. Incidentally, it follows from these properties
of the norm that Ip is in fact a linea" vector space because, if x = gil,
Y = {711} are vectors in Ip, then for any scalars ex, p, we have lIexx + Pyll
lexlllxll + IPIllyll < 00 so that' (XX + py is a
in Ip. Since Ip is a
vector space and the norm satisfies the three required axioms, we may
justifiably refer to Ip as a normed linear vector space.
The following two theorems, although of fundamental importance for a
study of the Ip spaces, are somewhat tangential to our main purpose. The
reader will lose little by simply scanning the proofs.

IIxll > 0 for each x :f: O.

Theorem 1. (The Holder Inequality) If p and q are positive numbers

1
00, I q 00, such that IIp + llq = I ant! ifx =
2 , ... } e /p
. y = {71t> 712, .. } e lq , then

30

LINEAR SPACES

Equality holds

if and only if

(M)
Ilxllp

1/ q

(J1J..)

l/p

Ilyllq

Jor each i.
Proof The cases p = 1, 00; q = 00, I are straightforward and are left
to the reader. Therefore, it is assumed that I < p < 00, I < q < 00. We
first prove the auxiliary inequality: For a >- 0, b >- 0, and 0 < A < I, we
have
AO

aWl-A)

+ (l

- A)b

with equality if and only if a = b.

For this purpose, consider the function
J(t) = I). - At

+ A-

defined for t >- o. Then f'(t) = A(t)-l - I). Since 0 < A < I, we have
f'(t) > for < t < I and f'(t) < for t > 1. It follows that for t >- 0,
J(t)
= 0 with equality only for t = 1. Hence

tA

At + I - A

with equality only for t = 1. If b =1= 0, the substitution t = alb gives the
desired inequality, while for b = 0 the inequality is trivial.
Applying this inequality to the numbers

leil

)p.

a = ( IIxll p ,

b-

(J1J..)Q

with A=

Ilyllq

1 - A=

we obtain for each i

(JQ)P

(1hl.)q

lei lTd <

+
Ilxllpllyllq - P II x l p
q IIYllq .
Summing this inequality over i, we obtain the Holder inequality
tJ)

Jlle, lId

1 1

-'--'--=---<-+-=1

Ilxllpllyllq - P

The conditions for equality follow directly from the required condition
o = b in the auxiliary inequality. I
The special case p = 2, q = 2is of major importance. In this case,
Holder's inequality becomes the well-known Cauchy-Schwarz inequality
for sequences:

2.IO

THE

Ip

AND

Lp

SPACES

31

Using the Holder inequality, we can establish the triangle inequality

for Ip norms.

Theorem 2. (The Minkowski Inequality) If x and yare in IP' I :::; p :::; 00,
then so is x + y, and Ilx + yllp :::; Ilxllp+ IIYllp' For I < P < 00, equality holds
if and only ifk 1x = k 2 y for some positive constants kl and k2 .

Proof The cases p == 1 and p = 00 are straightforward. Therefore, it

is assumed that 1 < p < 00. We first consider finite sums. Clearly we may
write
n

L lei + '7d
i=1

:::;

+ L lei + '7d P- I I'7J

L lei +
1=1

i=1

Applying Holder's inequality to each summation on the right, we obtain

+ '7d P (Itllei: + '7d(P-I)q) Ilq [ (JlledP) lip + (Itl l'7d P ) liP]

p

+ Ctl '7d Pf/1

= Ctllel + '7t1 YIQ

Dividing both sides of this inequality by
account of I - llq = lip, we find

(tllel + '7II Pf'P

Letting n --+
value, so

00

+ (tl''7llpf'P.

on the right side of this inequality can only increase its

Ctl 1el + '7;1p) lip

for each n. Therefore, letting n --+

Ilxll p+ IIYllp'

'

(Li'= I Ie + '7iI P)I /Q and taking

00

Ilxllp + IIYllp

on the left side produces IIx

+ yllp

The conditions for equality follow from the conditions for equality in
the Holder inequality and are left to the reader. I
The Lp and Rp spaces are defined analogously to the Ip spaces. For
p I, the space Lp[a, b] consists of those real-valued measurable functions x on the interval [a, b] for which Ix(tW is Lebesgue integrable. The
norm on this space is defined as
lP
II x l p=
. ,
W

((IX(t dtr

Unfortunately, on this space Ilxllp = 0 does not necessarily imply x = e

since x may be nonzero on a set of measure zero (such as a set consisting
of a countable number of points). If, however, we do not distinguish

32

LINEAR SPACES

between functions that are equal almost everywhere l 2 then Lp[al bJ is a .

normed linear space.
The space Rp[a, bJ is defined in an analogous manner but with attention
restricted to functions x for which Ix(t)1 is Riemann integrable. Since all
functions encountered in applications in this book are Riemann integrable,
we have little reason, .at this point in the development, to prefer one of
these spaces over the other. Readers unfamiliar with Lebesgue measure
and integration theory lose no essential insight by considering the Rp
spaces only. In the next section, however, when considering the notion of
completeness, we find reason to prefer the Lp spaces over .the Rp spaces.
The space Loo[a, bJ is the function space analog of 100 It is defined as
the space of all Lebesgue measurable functions on [a, bJ which are bounded,
except possibly on a set of measure zero. Again, in this, space, two functions
differing only on a set of measure zero are regarded as equivalent.
, Roughly speaking, the norm of an element x in Loo[a, bJ is defined as
sup Ix(t)l. This quantity is ambiguous, however, since an element x does
a,Stsb

not correspond uniquely to any given function but to a whole class of

functions difftring on a set of measure zero. The value sup Ix(t)1 is
astsb

different for the different functions equivalent to x. The norm of a function

in Loo[a, bJ is therefore defined as

Ilxll oo = essential supremum of IxCt)1

== infimum [sup ly(t)IJ

y(t)=x(t) a.e.

For brevity, we write

Ilxll oo

= ess sup Ix(t)l.

Example 1. Consider the function

1 - t2

x(t) = {2

E [

-1, 1J,

t;:f:O

t=O

shown in Figure 2;6. The supremum of this function is equal to 2 but

the essential supremum is 1.
There are HOlder and Minkowski inequalities for the Lp spaces analogous
to the corresponding results for the Ip spaces.
Theorem 3. (The Holder Inequality) If x
llq= l,p,q> i,then

Lp[a, bJ, y
.

Lq{a, bJ, IIp

LIX(t)y(t)1 dt::;;

IIxllpllYllq.

2 Two fUnctions are said to be equal almost everywhere (a.e.) on [a, h] if they differ
on a set of Lebesgue measure zero.

2.11

Equality holds

BANACH SPACES

33

if and only if

IX(t)I)P = (ly(t)I)q
( Ilxll p
IiYllq
almost everywhere on [a, b].

Theorem 4. (The Minkowski Inequality) If x and yare in Lp[a, b], then so

is x + y, and Ilx + yllp :::;; Ilxll p + lIyllp'
The proofs of these inequalities are similar to thJse for the Ip-spaces
and are omitted here.
x(t)

-1

Figure 2.6

o
The junction jor Example 1

2.11 Banach Spaces

Definition. A sequence {x n} in a normed space is said to be a Cauchy
sequence if Ilx. - xmll-" 0 as n, m .... OCJ; i.e., given B > 0, there is an integer
N such that Ilx. - xmll < B for all n, m > N.

In a normed space,. every convergent sequence is a Cauchy sequence

since, if Xn .... x, then
Ilx. - xmll = Ilx. - x + x - xmll :::;; Ilxn - xii + Ilx - xmll .... O.
In general, however, a. Cauchy sequence may not be convergent.
Normed spaces in which every Cauchy sequence is convergent are of
particular interest in analysis; in such spaces it ?s possible to identify
convergent sequences without explicitly identifying their limits. A space
in which every Cauchy sequence has a limit (and is therefore convergent)
is said to be complete.

34

LINEAR SPACES

Definition. A normed linear vector space X is complete if every Cauchy

sequence from X has a limit in X. A complete normed linear vector space
is called a Banach space.
We frequently take great care to formulate problems arising in applications as equivalent problems in Banach space rather than as problems in
other, possibly incomplete, spaces. The principal advantage of Banach
space in optimization problems is that when seeking an optimal vector
maximizing a given objective, we often construct a sequence of vectors,
each member of which is superior to the preceding members; the desired
optimal vector is then the limit of the sequence. In order that the scheme
be effective, there must be available a test for convergence which can be
applied when the limit is unknown. The Cauchy criterion for convergence
meets this requirement provided the underlying space is complete.
We now consider examples of incomplete normed spaces.
Example 1. Let X be the space of continuous functions on [0, 1] with
norm defined by IIxll = Sb Ix(t)1 dt. One may readily verify that X is a
normed linear space. Note that the space is not the space C [0, lJ since the
norm is different. We show that Xis incomplete. Define a sequence of
elements in X by tM.e equation

'\ 0
"
x n (t) = '.\ nt - -2 + 1

[I

for

O<t<--- - 2 n

for

---<t<2 n- - 2

for

1
t>-:-2

This sequence of functions is illustrated in Figure 2.7. Each member

of the sequence is a continuous function and thus a member of the
space X. The sequence is Cauchy since, as is easily verified, Ilxn - xmll =
tll/n - l/ml-4 0. It is obvious, however, that there is no continuous
function to which the sequence convergd.
Example 2. Let X be the vector space of finitely nonzero sequences
x = gi' 2 , , en, 0,0, ... }. Define the norm of an element x as
IIxll = max Ie;\. Define a sequence of elements in X by

2.1l

BANACH SPACES

35

Each x. is in X and has its norm equal to unity. The sequence {x.} is a
Cauchy sequence sinc(:, as is easily verified,
11x.. - xmll

= max {lIn, 11m} -+0.

It is obvious,
that there is no element of X (with a finite
number of nonzero components) to which this sequence converges.
x

Figure 2.7 Sequence for Example 1

The space E 1, the real line, is the fundamental example of a complete

space. It is assumed here that the completeness of E1, a major topic in
elementary analysis, is well known. The completeness of E1 is used to
establish the completeness of various other normed spaces.
We establish the completeness of several important spaces that are used
throughout this book. For this purpose the following lemma is useful.

Lemma 1. A Cauchy sequence is bounded.

Proof. Let {xn} be a Cauchy sequence and let N be an integer such
that Ilx. - xN11 < 1 flOr n > N. For 11. > N, we have
II x nll

= Ilxn-

XN

+ xN11

IlxNII

+ IIx n-

xNII < IlxNII

+ 1. I

Example 3. C[O, 1] is a Banach spac., We have previously considered

this space as a normed space. To prove that C ro, 1J is complete, it is

only necessary to shlOw that every Cauchy sequence in C [0, IJ has a limit.
Suppose {xn} is a Cauchy sequence in C [0, 1]. For each fixed t E [0, 1],
IXn(t) - Xm(t) I Ilxn - xmll -+ 0, so {xit)} is a Cauchy sequence of real
Since the :set of real numbers is complete, there is a real number
x(t) to which the sequence converges; x.(t) -+ x(t)'ITherefore, the functions
Xn converge pointwiise to the function x.
We prove next that this pointwise convergence is actually uniform in
t E [0, IJ, i.e., given e > 0, there is an N such that
- x(t)\ < e for all

36

LINEAR SPACES

t E [0, IJ and n :2: N. Given e > 0, choose N such that

Ilxn - xmll < e/2 for

n, m > N. Then for n > N

Ixn(t) - x(t)1 S Ixn(t) - Xm(t) I + Ixm(t) - x(t)1
S

Ilxn - xmll + IXm(t) -

x(t)l.

By choosing m sufficiently large (which may depend on t), each term on

the right can be made smaller than e/2 so IXn(t) - x(t)1 < e for n > N.
We must still prove that the function x is continuous and that the
sequence {xn} converges to x in the norm of C [0, 1]. To prove the continuity of x, fix e > 0. For every <5, t, and n,
Ix(t

+ <5) -

x(t)1 S Ix(t

+ <5) -

xn(t

+ <5)1

+ Ixn(t + <5) -

xn(t)1

+ IXn(t) -

x(t)l.

Since {xn} converges uniformly to x, n may be chosen to make both the

first and the last terms less than 8/3 for all <5. Since xn is continuous, <5 may
be chosen to make the second term less than e/3. Therefore, x is continuous. The convergence of Xn tox in the C [0,1] norm is a direct consequence
of the uniform convergence.
It is instructive to reconcile the completeness of C [0, 1] with Example 1
in which a sequence of continuous functions was shown to be Cauchy but
nonconvergent Jith respect to the integral norm. The difference is that,
with respect to the C [0, 1] norm, the sequence defined in Example 1 is
not Cauchy. The reader may find it useful to compare these two cases in
detail.
Example 4. Ip, 1 :s: p :s: 00 is a Banach space. Assume first that 1 S P < 00.
Let {xn} be a Cauchy sequence in Ip. Then, if xn = {e'i,
we have

lip

ez," .},

Hence, for each k, the sequence

is a Cauchy sequence of real numbers
and, therefore, converges to a limit ek' We show now that the element
x = 1 , 2, ... } is a vector in Ip' According to Lemma 1, there is a constant
M which bounds the sequence {x n }. Hence, for alln and k

{e e

L:
i= 1

:$;

IIxnll P :$; MP.

.

Since the left member of the inequality is a finite sum, the iaequality

remains valid as

11

00;

therefore,
k

L
i=
I

S MP.

2.1l

BANACH SPACES

37

Since this inequality holds uniformly in k, it follows that

and x E lp; the norm e.f x is bounded by Ilxl\ :::;; M.

It remains to be shown that the sequence {x n } converges to x in the lp
horm. Given e > 0, there is an N such that
k

Liei' - eil P :::;; Ilxn - xmll P :::;; e

1= 1

for n, m > N and for ,each k. Letting m

we conclude that

-+ 00

in the finite sum on the left,

Liei' - eil P :::;; e

1= 1

for n > N. Now letting k -+ 00, we deduce that Ilxn - xil P :::;; e for n > N
and, therefore, Xn -+ x. This completes the proof for 1 :s; P < 00.
Now let {x n } be a Cauchy sequence in lex)' Then Ie;: - e;:'1 :::;; Ilxn - xmll -+ 0.
Hence, for each k there is a real number ek such that e;: -+ ek' Furthermore,
this convergence is uniform in k. Let x = {el> e2 , }. Since {xn} is Cauchy,
there is a constant M such that for all , , Ilxnll :::;; M. Therefore, for each k
and each n, we have le;:1 :::;; Ilxnll :::;; M from which it follows that x E Ie<)
and IlxlI:::;; M.
The convergence of Xn to x follows directly from the uniform convergence
of e;: -+ ek

n.

Example 5. Lp[O,
1 !:.p!:. 00 is a Banach space. We do not prove the
completeness of the Lp spaces because the proof requires a fairly thorough
familiarity with Leb()sgue integration theory. Consider instead the space
Rp consisting of all functions x on [0, IJ for which Ixl P is Riemann
integrable with norm defined as

Ilxll

= {( Ix(t)IP dt} lIP.

The normed space Rp is incomplete. It may be completed by adjoining

to it certain additional functions derived from Cauchy sequences in Rp.
In this way, Rp is imbedded in a larger normed space which is complete.
space containing Rp is Lp. A general method for
The smallest
completing a normed space is discussed in Problem 15.
Example 6. Given two normed spaces X, Y, we consider the product space
X x Y consisting of ordered pairs (x, y) as defined in Section 2.2. The
space X x Y can be normed in several ways such as II(x, y)11 = Ilxll + Ilyll

38

LINEAR SPACES

or II(x,y)11 = max {llxll, Ilyll} but, unless specifically noted otherwise, we

define the product norm as II(x, y) II = Ilxll + Ilyli. It is simple to show that
if X and Yare Banach spaces, the product space X x Y with the product
norm is also a Banach space.
2.12

Complete Subsets

The definition of completeness has an obvious extension to subsets of a

normed space; a subset is complete if every Cauchy sequence from the
subset converges to a limit in the subset. The following theorem states that
completeness and closure are equivalent in a Banach space. This is
not so in general normed space since, for example, a normed space is
always closed but not necessarily complete.
Theorem 1. In a Banach space a subset is complete

if and only if it is closed.

Proof A complete subset is obviously closed since every convergent

(and hence Cauchy) sequence has a limit in the subset. A Cauchy sequence
from a closed subset has a limit somewhere in the Banach space. By
closure the limit must be in the subset. I

The following theorem is of great importance in many applications.

Theorem 2. In a lI10rmed linear space, any finite-dimensional subspace is
complete.
Proof The proof is by induction on the dimension of the subspace.
A one-dimensional subspace is complete since, in such a subspace, all
elements have the form x = lXe where IX is an arbitrary scalar and e is a
fixed vector. Convergence of a sequence IXn e is equivalent to convergence
of the sequence of scalars {lXn} and, hence, completeness follows from the
completeness of E.
Assume that the theorem is true for subspaces of dimension N - 1.
Let X be a normed space and M an N-dimensional subspace of X. We
show that M is complete.
Let {e 1 , e2' ... , eN} be a basis for M. For each k, define

Ok = inf Ilek a/s

'2:>jejll.

j*k

The number Ok is the distance from the vector ek to the subspace Mk

generated by the remaining N - 1 basis vectors. The number Ok is greater

than zero because otherwise a sequence of vectors in the N - 1 dimensional

subspace Mk could be constructed converging to ek' Such a sequence
cannot exist since Mk is complete by the induction hypothesis.

2.l3

EXTREME VALUES OF FUNCTIONALS AND COMPACTNESS

39

Define J > 0 as the minimum of the Jk , k = 1,2, ... , N. Suppose that

{x.} is a Cauchy sequence in M. Each x. has a unique representation as

For arbitrary n, m

Ilx. - xmll

II itp7 -

Aj)el\1 ;;::

for each k, I k N. Since Ilx. - xmll-+ 0, each IA7 O. Thus,

1 is a Cauchy sequence of scalars and hence convergent to a scalar
Ak' Let x = Lk=1 Akek' Obviously, x EM. We show that x. -+ x. For all
n, we have

Ilx. - xii

II k=1
f(Ak -

Ak)ekll s:; N max IAk - Akl' Ilekll,

1skSN

but since IAk - Akl-+ 0 for all k,

XEM. I

Ilx. - xii

-+

O. Thus, {x.} converges to

*2.13 Extreme Values of Functionals and Compactness

Optimization theory is largely concerned with the maximization or
minimization of real functionals over a given subset; indeed, a major
portion of this book is concerned with principles for finding the points at
which a given functional attains its maximum. A more fundamental
question, however, is whether a functional has a maximum on a given set.
In many cases the answer to this is easily established by inspection, but in
others it is by no means obvious.
In finite-dimensional spaces the well-known Weierstrass theorem, which
states that "a continuous function defined on a closed and bounded (compact) set has a maximum and a minimum, is of great utility. Usually, this
theorem alone is sufficient to establish the existence of a solution to a
given optimization problem.
In this section we generalize the Weierstrass theorem to compact sets
in a normed space, thereby obtaining a simple and yet general result
applicable to infinite-dimensional problems. Unfortunately, however, the
restriction to compact sets is so severe in infinite-dimensional
normed
I
spaces that the Weierstrass theorem can in fact only be employed in the
minority of optimization problems. The theorem, however, deserves
special attention in optimization theory if only because of the finitedimensional version. Th(: interested reader should also consult Section S.lO.

40

LINEAR SPACES

Actually, to prove the Weierstrass theorem it is not necessary to assume

continuity of the functional but only upper semicontinuity. This added
generality is often of great utility.
Definition. A (real-valued) functional/defined on a normed space X is
said to be upper sC!micontinuous at Xo if, given B > 0, there is aD> Osuch.
that f(x) - /(x o) < B for Ilx - XO I < D. A functional/ is said to be lower
semicontinuous at Xo if - / is upper semicontinuous at Xo.

As the reader may verify, an equivalent definition is that/is upper semicontinuous at Xo if 3 lim suPx .... xJ(x) s,/(x o)' Clearly, if / is both upper
and lower semicontinuous, it is continuous.
Definition. A set K in a normed space X is said to be compact if, given an
arbitrary sequence {x;} in K, there is a subsequence {Xi.,} converging to an
element x E K.

In finite dimensions, compactness is equivalent to being closed and

bounded, but, as is shown below, this is not true in general normed space.
Note, however, that a compact set K must be complete since any Cauchy
sequence from K must have a limit in K.

Theorem 1. (Weierstrass) An upper semicon tinuous functional on a compact

subset K 0/ a normed linear space X achieves a maximum on K.
I

Proof. Let M = sup/ex) (we allow the possibility M::;: 00). There is
xeK

a sequence {Xi} from K such that /(x;) --+ M. Since K is compact, there
is a convergent subsequence Xj" --+ x E K. Clearly, /(Xi,,) --+ M and, since /
is upper semicontinuous,/(x) lim/(x i ,,) = M, Thus, since/ex) must be
finite, we conclude that M < 00 and that/ex) ,b. M. I
We offer now an example of a continuous functional on the unit sphere
of C [0, IJ which does not attain a maximum (thus proving that the unit
sphere is not compact).

Example 1. Let the functional/be defined on

fo

1/2

f(x) =

x(t) dt -

cEO, IJ

by

1/2

x(t) dt.

It is easily verified that / is continuous since, in fact, 1/(x)1 s IIxli. The

supremum of / over the unit sphere in C [0, I J is I, but no continuous
function of norm less than unity aCllieves this supremum. (If the problem
3 The lim sup of a functional on a normed space is the obvious extension of the
corresponding definition for functions of a real variable.

2.14

QUOTIENT SPACES

41

were formulated in Loo[O, IJ, the supremum would be achieved by a

function discontinuous at t = l)

Example 2. Suppose that in the above example we restrict our attention

to those continuous functions in CEO, IJ within the unit sphere which are
polynomials of degree n or less. The set of permissible elements in C [0, IJ
is now a closed, bounded subset of the finite-dimensional space of n-th
degree polynomials and is therefore compact. Thus Weierstrass's theorem
guarantees the existence of a maximizing vector from this set. .
*2.14

Quotient

Suppose we select a subspace M from a vector space X and generate

linear varieties V in X by translations of M. The linear varieties obtained
can be regarded as the elements of a new vector space called the quotient
space of X modulo M and denoted XI M. If, for example M is a plane
through the origin in three-dimensional space, XI M consists of the family
of planes parallel to M. We formalize this definition below.

Definition. Let M be: a subspace of a vector space X. Two elements

X are said to be equivalent modulo M if Xl - X 2 E M. In this case,
we write Xl == X 2 .
Xl' X 2 E

This equivalence relation partitions the space X into disjoint subsets, or

classes, of equivalent elements: namely, the linear varieties that are distinct
translates of the subspaceM. These classes are often called the eosels of M.
Given an arbitrary
X E X, it belongs to a unique coset of M which
we denote4 by [x].

Definition. Let M be a subspace of a vector space X. The quotient spaee

XI M consists of all cosets of M with addition and scalar multiplication

defined by [xa

= [Xl + x2J, ct[xJ = [ctx],

Several things, whkh we leave to the reader, need to be verified in order

to justify the above definition. It must be shown that the definitions for
addition and scalar multiplication are independent of the choice of
representative elements, and the axioms for a vector space must be verified.
These matters are easily proved, howevP,r, and it is not difficult to see that
addition of cosets [Xl] and [x 2] merely amounts to addition of the corresponding linear varieties regarded as sets in X. Like.vise, multiplication of
a coset by a scalar oc (except for oc = 0) amounts to multiplication of the
corresponding linear variety by ct.
4 The notation [xl is also used for the subspace generated by x but the usage is always
clear from context.

42

LINEAR SPACES

Suppose. now that Xis a notmed space

that M is a, closed subspace
of X. We define the norm of a coset [x] E X/M by

II [x] I

= ",eM
infllx + mil,

i.e., I [x] II is the infimum of the norms of all elements in the coset [x]. The
assumption that M is closed insures that II [x] I > 0 if [x] :f: e. Satisfaction
of the other two axioms for a norm is easily verified. In the case of X being
two dimensional, M one dimensional, and X/M consisting of parallel
lines, the quotient norm of one of the lines is the minimum distance of the
line from the origin.
Proposition 1. Let X be a Banach space, M a closed subspace of X, and X/ M
the quotient space l-t'ith the quotient norm defined as above. Then X/M is also
a Banach space.
The proof is left to the reader.
*2.15 Denseness and Separability
We conclude this chapter by introducing one additional topological concept, that of denseness.

Definition. A set D is said to be dense in a normed space X if for each

element x E X and each e > 0 there exists dE D with /Ix - d/l < e.
I

If D is dense in X, there are points of D arbitrarily close to each x E X.

Thus, given x, a sequence can be constructed from D which converges to
x. It follows that equivalent to the above definition is the statement: D
is dense in X if 15, the closure of D, is X.
The definition converse to the above is that of a nowhere dense set.

Definition. A set E is said to be nowhere dense in a normed space X if E

contains no open set.
The classic example of a dense set is the set of rationals in the real line.
Another example, provided by the well-known Weierstrass approximation
theorem, is that the space of polynomials is dense in the space C [a, b1

Definition. A normed space is separable if it contains a countable dense set.

Most, but not all, of the spaces considered in this book are separable

Example 1. The space E" is separable. The collection of vectors

eel' ez , ... , en) having rational components is countable

x=
in En.

and dense

2.16

PROBLEMS

43

Example 2. The Ip spaces, 1 p < 0Cl, are separable. To prove separability,

let D be the set consi!;ting of all finitely nonzero sequences with rational
components. D is easily seen to be countable. Let x = {el'
exist in
Ip, 1 P < 0Cl' and fix e > O. Since
1 leil P < 0Cl' there is an N such that
Ei'?:N+ 1 led P < e/2. For each k, 1 k N, let rk be a rational such that
- rkl P < e/2N; let d = {r1> r2' ... , rN, 0, 0, ... }. Clearly, de D and
IIx - dil P < e. Thus, D is dense in Ip.
The space 100 is not separable.

e2, ... }

Example 3. The spac1e C [a, bJ is separable. Indeed, the countable set

consisting of all polynomials with rational coefficients is dense in C [a, bJ.
Given x E C [a, bJ and e > 0, it follows from the well-known Weierstrass
approximation theorem that a polynomial p can be found such that
IxCt) - p(t)1 < e/2 for all t E [a, bJ. Clearly, there is another polynomial r
with rational
such that Ip(t) - r(t)1 < e/2 for all t E [a, bJ
can be constructed by changing each coefficient by less than s/2N where
N - 1 is the order of the polynomial p). Thus
Ilx - rll

= max Ix(t) te[a,b]

r(t)1

max Ix(t) - p(t)1

+ max Ip(t) -

r(t)1

Example 4. The Lp spaces 1 p < 0Cl are separable but Loo is not separable.
The particular space L2 is considered in great detail in Chapter 3.

2.16 Problems
1. Prove Proposition I, Section 2.2.
2. show that in a vector space (-(X)x = (X( -x) = -(Xx), (X(x - y) =
(Xx - (XY, (X - (3)x := (Xx - [3x.
3. Let M and N be subspaces in a vector space. Show that [Mu NJ =
M+N.
4. A convex combination of the vectors Xl' x 2 , .. " Xn is a linear combination of the form !XlX l + (X2 X2 + ... + (Xnxn where (Xi 0, for each
i; and (Xl + (X2 + ... + (Xn = 1. Given a set S in a vector space, let K be
the set of vectors consisting of all convex combinations from S. Show
that K = co(S).
5. Let C and D be co:nvex cones in a vector space. Show that C nD and
C + D are convex cones.
6. Prove that the union of an arbitrary collection of open sets is open and
that the intersection of a finite collection of open sets is open.

44

LINEAR SPACES

7. Prove that the intersection of an arbitrary collection of closed sets is

closed and that the union of a finite collection of closed sets is closed.
8. Show that the closure of a set S in a normed space is the smallest closed
set containing S.
9. Let X be a normed linear space and let Xl' Xl' .. , Xn be linearly
independent vectors from X. For fixed y E X, show that there are
coefficients ai' a2, ... , an minimizing Ily - a l x 1 - a2 x 2 ... - anxnll.
10. A normed space is said to be strictly normed if Ilx + yll = Ilxll + Ilyll
implies that y = or X = ay for some a.
(a) Show that Lp[O, 1] is strictly normed for 1 < P < 00.
(b) Show that if X is strictly normed the solution to Problem. 9 is
unique.
11. Prove the Holder inequality for p = 1,00.
12. Suppose X = g1'
... ,
... } E lp<> for some Po' 1 5, Po < 00.
(a) Show that X E lp for all P Po .
.
(b) Show that Ilxll", = lim Ilxllp'

p .... '"

13. Prove that the normed space D[a, b] is complete.

14. Two vector spaces, X and Y, are said to be isomorphic if there is a
one-to-one mapping T of X onto Y such that T(alx l + a2 x 2) =
alT(x l ) + a2 T(x;J. Show that any real n-dimensional space is isomorphic to En.
15. Two normed linear spaces, X and Y, are said to be isometrically
isomorphic if theY' are isomorphic and if the corresponding one-to-one
mapping T satisfies IIT(x)11 = Ilxll. The object of this problem is to
show. that any normed space X is isometrically isomorphic to a dense
subset of a Banach space %.
(a) Let X be the set of all Cauchy sequences {xn} from Xwith addition
and scalar multiplication defined coordinatewise. Show that X is a
linear space.
(b) If y {x n } E X, define Ilyll = sup Ilxnll. Show that with this

definition, X becomes a normed space.

(c) Let M be the subspace of X consisting of all Cauchy sequences
convergent to zero, and let j( = X /M. Show that j( has the required
properties.
16. Let S, T be open sets in the normed spaces X and Y, respectively.
Show that S x T = {(s, t) : s E S, t E T} is an open subset of X x Y
under the usual product norm.
17. Let M be an tn-dimensional subspace of an n-dimensional vector
space X. Show that XIM is (n - m) dimensional.
18. Let Xben dimensional and Ybe m dimensional. What is the dimension
of Xx Y?

2.l6

19. A real-valued functional

seminorm if:
1. Ixl:2: 0
all x E X
2. IlXxl = 11X1'lxl

3. Ix + y I Ixl
Let M = {x: Ixl
inf

meM

Ix + ml

lxi,

PROBLEMS

45

defined on a vector space X, is called a

+ Iyl

= O}

and show that the space X/M with

is llormed.

l/[xJI/

20. Let X be the space of all functions x on [0, IJ which vanish at all
but a countable lllumber of points and for which

IIxll

00

1: Ix(tn) I < 00,

n= 1

where the tn are the points at which x does not vanish.

(a) Show that X is a Banach space.
(b) Show that X is not separable.
21. Show that the Banach space I"" is not separable.

REFERENCES
There are a numbeJr of excellent texts on linear spaces that can be used to
supplement this chaptl:r.
2.2. For the basic concepts of vector space, an established standard is Halmos
[68].
2.4. Convexity is an important aspect of a number of mathematical areas. For
an introductory discussion, consult Eggleston [47] or, more advanced,
Valentine [148].
2.6-9. For general discussions of functional analysis, refer to any of the following. They a.re listed in approximate order of increasing level: Kolmogorov
and Fomin [88], Simmons [139], Luisternik and Sobolev [10lJ, Goft'man
and Pedrick [59], Kantorovich and Akilov [79], Taylor [145], Yosida [157],
Riesz and Sz-Nagy [123], Edwards [46], Dunford and Schwartz [45], and
Hille and Phillips [73].
2.1O. For a readable discussion of lp and Lp spaces and a general reference on
analysis and integration theory, see Royden [131].
2.13. For the Weierstrass approximation theorem, see Apostol [10].

3
HILBERT SPACE

3.1 Introduction
Every student of high school geometry learns that the shortest distance
from a point to a line is given by the perpendicular from the point to the
line. This highly intuitive result is easily generalized to the problem of
finding the shortest distance from a point to a plane; furthermore one might
reasonably conjecture that in n-dimensional Euclidean space the shortest
vector from a point to a subspace is orthogonal to the subspace. Thi's is, in
fact, a special Case of one of the most powerful and important optimization
principles-the projection theorem.
The key concept in this observation is that of orthogonality; a concept
which is not generally available in normed space but which is available in
Hilbert space. A Hilbert space is simply a special form of normed space
having an inner product defined which is analogous to the dot product of
two vectors in analytic geometry. Two vectors are then defined as
orthogonal if their inner product is zero.
Hilbert spaces, equipped with their inner products, possess a wealth of
structural properties generalizing many of our geometrical insights for two
and three dimensions. Correspondingly, these structural properties imply a
. wealth of analytical results applicable to problems formulated in Hilbert
space. The concepts of orthonormal bases, Fourier series, and leastsquares minimization all have natural settings in Hilbert space.
PRE-HILBERT SPACES

3.2 Inner Products

Definition. A pre-Hilbert space is a linear vector space X together with an
inner product defined on X x X. Corresponding to each pair of vectors
x, y in X the inner product (x Iy) of x and y is a scalar. The inner product
satisfies the following axioms;
1.

(xly) = (Ylx).
+ ylz) = (xlz)

2. (x

+ (Ylz).
46

3.2

3. (Ax Iy) = A(x Iy).

4. (x I x) ;::: 0 and (x I x)

= 0 if and only if x =

INNER PRODUCTS

47

e.

The bar on the right: side of axiom I d, .lotes complex conjugation. The
axiom itself guarantees that (x I x) is real for each x. Together axioms 2 and
3 imply that the inner product is linear in the first entry. We distinguish
between real and complex pre-Hilbert spaces according to whether the underlying vector space is real or complex. In the case of real pre-Hilbert spaces,
it is required that the iltlner product be real valued; it then follows that the
inner product is linear in both entries. In this book the pre-Hilbert spaces
are almost exclusively considered to be real.
The quantity J(XTJ0 is denoted IIxll, and our first objective is to verify
that it is indeed a norm in the sense of Chapter 2. Axioms I and 3 together
give II ax II = lo:lllxll and axiom 4 gives IIxll > 0, x # e. It is shown in
Proposition I that I II satisfies the triangle inequality and, hence, defines a
norm on the
space.
Before proving the triangle inequality, it is first necessary to prove an
important 'lemma whkh is fundamental throughout this chapter.

Lemma 1. (The Cauchy-Schwarz Inequality) For all x, y in an inner product

space, l(xly)l::;;; Ilxllllyll. Equality holds if and only ifx = Ay or y = e.

e;

Proof If y = the inequality holds trivially. Therefore, assume y #

For all scalars A, we have
0::;;; (x - AY I x -- AY) = (x I x) - A(y I x) - 1(x I y)

e.

+ IAI2(y Iy).

In particular, for A = (x I y)/(y I y), we have

o : ; ; (x I x) _I(x I y)12 ,
(y Iy)

or

l(xly)1 ::;;;J(xlx)(Yly) = IIxlillyll

Proposition 1. On a pre-Hilbert space X the junction IIxll = J(x I x) is a

norm.
Proof The only requirement for a norm which has not already been
established is the triaillgle inequality. For any x, y E X, we have

IIx + yl12 = (x + ylx + y) = (xix) + (xly) + (Ylx) + (Yly)

::;;; IIxll2 +21(xly)1 +llyIl2.
By the Cauchy-Schwarz inequality, this becomes

Ilx + Yll2

IIxll2 + 211xllilyll I- lIyll2

(lIxll + lIyll)2.

The square root of the above inequality is the deshd result.

48

HILBERT SPACE

Several of the normed spaces that we have previously considered can be

converted to pre-Hilbert spaces by introducing an appropriate inner
product.
Example 1. The space En consisting of n-tuples of real numbers is a
pre-Hilbert space with the inner product of the vector x =
and the vector y = ('11' '12"", '1n) defined as (xly) = 2J=1 ei'1i' In this
case it is clear that (x Iy) = (y I x) and that (x Iy) is linear in both entries.
The norm defined as (x Ix) is

Ilxll

Ct

1e"1 2 f/2

which is the Euclidean norm for En.

Example 2. The (real) space 12 becomes a pre-Hilbert space with the inner
product of the vectors x = gl' e2' ... } and y = {'11' '12, ... } defined as
(xIY) = Lf';,1 i '1i' The Holder inequality for 12 (which becomes the
Ilxii . lIy\\ and,
Cauchy-Schwarz inequality) guarantees that I(x Iy)
thus, the inner product has finite value. The norm defined by (x I x)
is the usual 12 norm.

Example 3. The (real) space L 2 [a, b] is a pre-Hilbert space with the inner
product defined as (x Iy) = S: x(t)y(t) dt. Again the Holder inequality
guarantees that (x Iy) is finite.
Example 4. The space of polynomi1tlfunctions on [a, b] with inner product
(x I y) = x(t)y(t) dt is a pre-Hilbert space. Obviously, this space is a subspace of the pre-Hilbert space L 2 [a, b].

There are various properties of inner products which are a direct consequence of the definition. Some of these are useful in later developments.
Lemma 2. In a pre-Hilbert space the statement (x Iy) = 0 for all y implies
that x = O.
Proof Putting y

=:

x implies (x I x)

= 0.

Lemma 3. (The Parallelogram Law) In a pre-Hilbert space

Ilx + yll2 + Ilx -

yll2 =

211xl12 + 211y112.

Proof. The proof is made by direct expansion of the norms in terms of

the inner product. I

This last result is a generalization of a result for parallelograms in twodimensional geometry. The sum of the squares of the lengths of the
diagonals of a parallelogram is equal to twice the sum of the squares of
two adjacent sides. See Figure 3.1.

3.3

THE PROJECTION THEOREM

49

Since a pre-Hilbert space is a special kind of normed linear space, the

concepts of convergence, closure, completeness, etc., apply in these spaces.

Definition. A complet1e pre-Hilbert space is called a Hilbert space.

A Hilbert space, then, is a Banach space equipped with an inner product
which induces the norm. The spaces En, 12 , and L 2 [a, b] are all Hilbert
spaces. Inner products. enjoy the following useful continuity property.

Figure 3.1 The parallelogram law

Lemma 4. (Continuity of the Inner Product) Suppose that Xn -+ x andy n -+ Y

in a pre-Hilbert space. Then (xn IYn) -+ (x Iy).

Proof Since the sequence {xn} is convergent, it is bounded; say

Ilxnll M. Now
I(x nIYn) - (x Iy)l == l(x IYn) - (xn Iy)
II

+ (xn IY) -

(x ly)1

I(x n IYn - y)1

+ l(xn -

x ly)l

Applying the Cauchy-Schwarz inequality, we obtain

I(x nIYn) - (x Iy)1

Since

/lxn/lllYn - yll

+ /lxn -

x/lllyI/

/lxnll is bounded,
l(xn IYn) - (x Iy)l

MIlYn - y/I f- /lxn -

x/llIylI

-+

O.

3.3 The Projection Theorem

We get a lot of analytical mileage from the following definition.

Definition. In a pre-Hilbert space two vectors x, yare said to be orthogonal if (x IY) = O. We symbolize this by x .Ly. A vector x is said to be
orthogonal to a set S (written x.L S) if x.L s for each s E S.
The concept of orthogonality has many of the c!onsequences in preHilbert spaces that it has in plane geometry. For example, the Pythagorean
theorem is true in pre-Hilbert spaces.
Lemma 1. Ifx .Ly, then IIx

+ YII2

IIxll2 + lIylI 2.

50

HILBERT SPACE

3.

Proof
Ilx

+ Yl12

= (x

+ y I x + y) =

IIxl12

+ (x Iy) + (y Ix) + Ilyll:! = IIxl12 + IIYI12 .

We turn now to our first optimization problem and the projection

. 'theorem which characterizes its solution. We prove two slightly different
versions of the theorem: one valid in an arbitrary pre-Hilbert space and
another. with a stronger conclusion, valid in Hilbert space.
The optimization problem considered is this: Given a vector x in a preHilbert space X and a subspace M in X, find the vector mE M closest to x
in the sense that it minimizes Ilx - mil. Of course, if x itself lies in M, the
solution is trivial. In general, however, three important questions must be
answered for a complete solution to the problem. First, is there a vector
m E M which minimizes IIx - mil, or is there no m that is at least as good as
all others? Second, is the solution unique? And third, what is the solution
or how is it characterized? We answer these questions now.
Theorem 1. Let X be a pre-Hilbert space, M a subspace of X, and x an
arbitrary vector in X. If there is a vector mo EM such that Ilx - moll ::;;
IIx - mil for all mE M, then mo is unique. A necessary and sufficient

condition that rno E M be a unique minimizing vector in M is that the error

vector x - mo be orthogonal to M.
Proof We show first that if mo is a minimizing vector, then x- m() is
orthogonal to M. Suppose to the contrary that there is an m E M which is
not orthogonal to x - mo. Wfthout loss of generality, we may assume that
Ilmil == I and that (x - mo 1m) = f> =;6 O. Define the vector m1 in M as
m1 = mo + f>m. Then
Ilx - ml11 2 = Ilx - mo - f>m112 = Ilx - mol1 2 - (x - mo I f>m)
-(f>m I x - mo) + 1f>12
2
= IIx - moll 2 -1f>1 2 < Ilx - m ol .
Thus, if x - mo is not orthogonal to M, mo is not a minimizing vector.
We show now that if x - mo is orthogonal to M, then mo is a unique
minimizing vector. For any mE M, the Pythagorean theorem gives

IIx - mll 2 = IIx - mo + mo - ml12 = IIx - m oll 2 + limo - m1l2.

Thus, IIx - mil> IIx - moll for m =;6 mo .
The three-dimensional version of this theorem is illustrated in Figure 3.2.
We still have not established the existence of the minimizing vector. We
have shown that if it exists, it is unique and that x - rno is orthogonal to the
subspace M. By slightly strengthening the hypotheses, we can also guarantee the existence of the minimizing vector.

3.3

THE PROmCTION THEOREM

51

Figure 3.2 The projection theorem

Theorem 2. (The Classical Projection Theorem) Let H be a Hilbert space

and M a closed subspace of H. Corresponding to any vector x e H, there is a
unique vector mo e M such that IIx - moll S IIx - mil lor all me M. Furthermore, a necessary and sufficient condition that mo E M be the unique minimizing vector is that x - mo be orthogond to M.
Proof The uniqueness and orthogonality have been established in
Theorem 1. It is only required to establish the existence of the minimizing
vector.
If x E M, then m(l = x and everything is settled. Let us assume x M and
define 0 = inf \Ix - mil. We wish to produce an mo e Mwith II x-moll = O.
meM

For this purpose, let {m i } be a sequence of vectors in M such that

IIx - mill -. O. Now, by the parallelogram law,

lI(m) - x) + (x -- mi)1I 2 +1I(m) - x) - (x - ml)11 2

= 211m) - xl12 + 211x - m1112.
Rearranging, we obtain

For all i, j the vector (mj + mj)/2 is in M since M is a linear subspace.

Therefore, by definition of /), Ilx - (mi + mj)/211 /) and we obtain

11m) -- mill 2 S 211m) - xl1 2 + 211x - mill 2 - 40 2

Since IImi - xll2

->

02 as i -+

00,

we conclude

as i, j -.

00.

Therefore, {m l } is Ii Cauchy sequence, and since M is a closed subspace of a

52

HILBERT SPACE

. complete space, the sequence {m j } has a limit 1110 in M.By continuity of the
norm, it follows that IIx - moll =
I

o.

It should be noted that neither the statement nof the proof of the
existence of the minimizing vector makes explicit reference to the inner
product; only the norm is used. The proof, however, makes heavy use of
the parallelogram law, the proof of which makes heavy use of the inner
product. There is an extended version of the theorem, valid in a large class
of Banach spaces, which is developed in Chapter 5, but the theorem cannot
be extended to arbitrary Banach spaces.

3.4

Orthogonal Complements

In this section we apply the projection theorem to establish some additional

structural properties of Hilbert space. Our primary object is to show that
given any closed subspace of a Hilbert space, any vector can be written as
the sum of two vectors: one in the given subspace and one orthogonal to it.

Definition. Given a subset S of a pre-Hilbert space, the set of all vectors

orthogonal to S is called the orthogonal complement of S and is denoted
S1..

The orthogonal complement of the set consisting only of the null vector
is the whole space. For any set S, S 1- is a closed subspace. It is a subspace
because a linear combination of vectors orthogonal to a set is also orthogonal to the set. It is closed since if {x,J is a convergent sequence from S1-,
say Xn x, continuity of the mner product implies that 0 = (xn Is) (x I s)
for all s E S, so XES 1-. The following proposition summarizes the basic
relations between a set and its orthogonal complement.

Proposition 1. Let Sand T be subsets of a Hilbert space. Then

I.
2.
3.
4.

S 1- is a closed subspace.
SC SH,
If SeT, then T1- c S 1-.
SH1- = S1-.

5.

SH = [SJ, i.e., SH is the smallest closed subspace containing S.

Proof

1.
2.
3.
4.

This was proved in the text above.

If XES, then x J.y for all y E S1-; therefore, x E SH.
If Y E T1-, then Y 1. x for all XES since SeT. Therefore, y E S1-.
From 2, S.L c S.LH. Also S c SH which from 3 implies SH1- cs.L.

Therefore, S 1.
S.

::=

S H1..

This statement is most easily proved by using the result of Theorem 1

below. We leave it as an exercise. I

3.5

THE GRAM-SCHMIDT PROCEDURE

53

Definition. We say that a vector space X is the direct sum of two subspaces
M and N if every vector x E X has a unique represeniationof the form
x = m + n where m E M and n E N. We describe this situation by the notation X = M EB N. (The 4)nly difference between direct sum and the earlier
definition of sum is the added requirement of uniqueness.)
We come now to the theorem that motivates the expression" orthogonal
complement" for the set of vectors orthogonal to a set. If the set is a
closed subspace in a Hilbert space, its on:hogonal complement contains
enough additional vectors to generate the space.
Theorem 1. If M is a closed linear subspace of a Hilbert space H, then
H = MEt> M 1. and M == M H.

Proof The proof follows from the projection theorem. Let x E H.

By the projection theorem, there is a unique vector mo E M such that
IIx - moll ::s; IIx - mil for all mE M, and no = x - mo E M1.. Thus,
x = 10 + no with 10 e M and no E M J..
To show that this
is unique, suppose x = m1 + n1 with
m1 E M, n1 E M 1.. Then 9 = m1 - mo + n1 - no but m1 - mo and n1 - no
are orthogonal. Hence, by the Pythagorean theorem, 11911 2 = IIm1 - m oi\2+
IIn1 - noll 2 This
that mo == m1 and no = n1'
To show that M = Al H, it is only necessary to show M H c: M since by
Proposition I we have M c: M H. Let x E M H. By the first part of this
theorem, x = m + n where mE M, n E M1.. Since both x E MH and
mE MH, we have x - mE MH; that is, n E MH. But also n E M1., hence
n .Ln which implies n = 9. Thus, x = mE M and MH c: M. I
In view of the above results, given a vector x and a closed subspace Min
a Hilbert space, the vector mo E M such that x - mo E M 1. is called the
orthogonal projection of x onto M.
3.5 The Gram-Schmidt Procedure

Definition. A set S of vectors in a pre-Hilbert space is said to be an

orthogonal set if x .L)' for each x, YES, x : y. The set is said to be orthonormal if, in addition, each vector in the set has norm equal to unity.
As we shall see, orthogonal sets are extremely convenient in various
problems arising in Hilbert space. Partially, this is due to the property
stated below.
I
Proposition 1. An orthogonal set of nonzero vectors is a linearly independent
set.

54

HILBERT SPACE

Proof Suppose {Xi> X2., ... , xn} is a finite subset of the given orthogonal set and that there are n scalars lXi' i = 1, 2, ... , n, such that
:Li'= 1 ctiXi = O. Taking the inner product of both sides of this equation with
Xk produces

or

Thus, ctk = 0 for each k and, according to Theorem 1, Section 2.5, the
vectors are independent. I
In Hilbert space, orthonormal sets are greatly favored over other
linearly independent sets, and it is reassuring to know that they exist and
can be easily constructed. The next theorem states that in fact an orthonormal set can be constructed from an arbitrary linearly independent set. The
constructive method employed for the proof is referred to as the GramSchmidt orthogonalization procedure and is as important in its own right
as the theorem itself.

Theorem 1. (Gram-Schmidt) Let {Xi} be a countable or finite sequence of

linearly independent vectors in a pre-Hilbert space X. Then, there is an
orthonormal sequence {e i} such that for each n the space generated by the
first n e;'s is the same as the space generated by the first n x;'s; i.e.,jor each n
we have [e l , e2' ... , en] [Xl' X2' ... , XnJ.
Proof For the first vector, take

which obviously generates the same space as Xl' Form e2 in two steps.
First, put

Z2

= X2 -

(x21 el)el

and then e2 = z2/11z 211. By direct calculation, it is verified that Z2 .1. e1 and
e2 .1. el . The vector Z2 cannot be zero since X2 and el are linearly independent; furthermore, e2 and e1 span the same space as Xl and X2 since X2
may be expressed as a linear combination of el and ez.
The process is best understood in terms of the two-dimensional diagram
of Figure 3.3. The vector Z2 is formed by subtracting the projection of X2 on

3.6

THE NORMAL EQUATIONS AND GRAM MATRICES

el from x 2 The remaining ej's are defined by induction. The vector

55
Zn

is formed according to the equation

n-l
Zn

Xn -

L1(xn Iej)ej

j=

and

Again it is easily verified by direct computation that Zn .L e / for all i < n,

and Zn is not zero since it is a linear combination of independent vectors. It
is clear by induction that the e;'s generate the same space as the x/'s. If the
original collection of x I 's is finite, the process terminates; otherwise the
process produces an infinite sequence of orthonormal vectors. I

(.\'21 el )el

Figure 3.3

The Gram-Schmidt procedure

The study of the Gram-Schmidt procedure and its relation to the projection theorem, approximation, and equation solving is continued in the
following few secti'Dns.
APPROXIMATION

3.6 The Normal lr.quations and Gram Matrices

In this section we investigate tHe following approximation problem.

Suppose Yto Y2' ... , Yn are elements of a Hilbert space H. These vectors
generate a (closed) finite-dimensional subspace M of H. Given an arbitrary vector x E H, we seek the vector in M which is closest to x. If is
expressed in terms. of the vectors Yi as = IXIYI + 1X2Y2 + ., . + IXnYn, the
problem is equivalent to that of finding the n scalars lXi' i == I, 2, ... , n,
that minimize Ilx,- (XIYI - (X2Y2 - .. , - IXnYnll

56

HILBERT SPACE

According to the projection theorem, the unique minimizing vector is

the orthogonal projection of x on M, or equivalently the difference vector
xis orthogonal to each of the vectors Yi' Therefore,
(x

-IXIYI

=0

-1X2Y2 -'"

for i = I, 2, ... , n. Or, equivalently,

+ (Y21 YI) 1X 2 + ... + (Yn IYI)lX n = (x IYI)

(YII Y2)IX I + (Y21 Y2)1X 2 + ... + (Yn IY2)er: n = (x IY2)

(YII

YI)IX I

These equations in the n coefficients IX; are known as the normal equations
for the minimization problem.
Corresponding to the vectors YI' Y2, ... , Yn, the n x n matrix
(YII Yt)
(Y21 YI)
G=G(Yt.Y2, .. ,Yn):=
:
[
(YnIYt)

(YII Y2)

is called the Gram matrix of YI' Y2, ... , Yn' It is the transpose of the coefficient matrix of the normal equations. (In a real pre-Hilbert space the
matrix is symmetric. In a complex space its transpose is its complex conjugate.) The determinant g = g(Yb Y2 , ... , Y.) of the Gram matrix is known
as the Gram determinant.
I
The approximation problem is solved once the normal equations are
solved. In order that this set of equations be uniquely solvable, it is necessary and sufficient that the Gram determinant be nonzero.

Proposition 1. g(Yl> Y2 , ... , Yn) = 0 if and only if the vectors Yl> Y2 , ... , Yn
are linearly independent.

Proof. The equivalent contrapositive statement is that

g(Yl, Y2 , ... , Yn)

=0

if and only if the vectors YI' Y2, ... , Y. are linearly dependent. Suppose that
the Yi'S are linearly dependent; that is, suppose there are constants IX/) not
aU zero. such that 2:7-. f't,)J, ::= fJ. It then clear that the rows of the Gram
determinant have a corresponding dependency and, hence, the determinant
is zero.

3.6

1rHE NORMAL EQUATIONS AND GRAM MATRICES

57

Now assume that the Gram determinant is zero or, equivalently, that
there is a linear dependency among the rows. In that case there are-constants r:J.j, not all zero, such that
1 r:J.j(Yi IYJ) = 0 for all j. From this it
follows that
for all ,j
and hence that

or

Thus,

L7=

r:J.iYI = () and the vectors Yl, Y2, ... , Y. are linearly dependent.

Although the normal equations do not possess a unique solution if the

YI'S are linearly dependent, there is always at least one solution. Thus the
degeneracy that arises as a result of g = 0 always results in a multiplicity of
solutions rather than an inconsistent set of equations. The reader is asked
to verify this in Probkm 4.
We turn now to the evaluation of the minimum distance between x and
the subspace M.

Theorem 1. Let Yl' hI ... , Y. be linearly independent. Let be the minimum

distance from a vector x to the subspace M generated by the y;'s; i.e.,

Then

== g(ylo Y2, . , Y., x).

g(yto Y2, , Yn)
Proof By definition,
:::: IIx ::;= (x I x) - (x - I
projection theorem, x - is orthogonal to M so, in particular, (x O. Therefore,

By the

I =

or

This equation, together with the normal equations, gives n + 1 linear

58

HILBERT SPACE

equations for the n + 1 unknowns

rule to determine 6 2 , we obtain

(ytl Y1)
(YIIY2)

(Y21 YI)

...

IXt, 1X2' . , IXn'

(Yn IYt)

(x IYI)

(Yt I Yn) , .
(Yn I Yn) (x I Yn)
(YI I x)
I-=.....::....:---C-_
_ _ _ _----"'(Y-::n-'-Ix-,)_-,(_x.:.....1X---'-)_I
.:5 2 =
(YIIYI) (Y21 Yt)
(Yn I Yt)
0
(hi Y2)
(Yt I Yn)
(Yt I x)

6 2 Applying Cramer's

.. , (Yn I Yn)
.. , (Ynlx)

= g(y 1, Y2 , ... , Yn' x).

g(YI' Y2' ... , Yn)

0
1

This explicit formula, although of some theoretical interest, is of little

practical importance because of the impracticality of evaluating large
determinants. The evaluation of an n + I-dimensional determinant is in
general about as difficult as inverting an n-dimensional matrix; hence, the
determinant relation is not particularly attractive for direct computations.
An alternative approach is developed in Section 3.9.
3.7 Fourier Series
Consider the problem of finding the best approximation to x in the subspace M generated by the orthonormal vectors et , e2' ... , en' This special
case of the general approximation problem is trivial since the Gram
matrix of the e /s is simply the identity matrix. Thus the best approximation is
n

i = LlXiei
i= t

where (Xi = (x lei)'

In this section we generalize this special approximation problem slightly
by considering approximation in the closed subspace generated by an
infinite orthonormal sequence. This leads us naturally to a general discussion of Fourier series.
To proceed we must define what is meant by an infinite series.

Definition. An i'lfinite series of the form

Xi

is said to converge to the

element x in a normed space if the sequence of partial sums

converges to x. In that case we write x =
I Xi'

Sn

Li= I

XI

3.7

FOURIER SERIES

S9

We now establish a necessary and sufficient condition for an infinite

series of orthogonal vectors to converge in Hilbert space.

Theorem 1. Let {el} be an orthonormal sequence in a Hilbert space H. A

series of the form
I i e i converges to an element x
led 2 < 00 and, in that case, we have = (xl ei)'
froof. Suppose that
sums s" =
Then,

ele,.

H if and only if

Ie 112 < 00 and define the sequence of partial

Iisn - sml1 = II
2

lei ei 112 =

-4 0

as n, m -4 00. Therefore, {sn} is a Cauchy

so by the completeness
of H there is an element x E H such that Sn -). x.
Conversely, if s" converges, it is a Cauchy sequence 10 Lf=n+ 1
o.
Thus,
-0 and
< 00.
Obviously, (sn lei) -4- I as n -+ 00 which, by the continuity of the inner
product (Lemma 4, Section 3.2), implies that (x I e i) = I' I

In the above theorem we started with an arbitrary square-summable

1 i e i in
sequence of scalars i} and constructed the element x =
the !;pace. It was found that I = (x I e I)' We now consider the possibility of
start'ng with a given x and forming an infinite summation with (x I e l ) as
coefficients of the orthonormal vectors ei' These coefficients are called the
Fourier coefficients of x with respect to the ei and are fundamental in the
theory of generalized Fourier series in Hilbert space.

Lr;. e

Lemma 1. (Bessel's Inequality) Let x be an element in a Hilbert space H

and suppose {etl is an orthonormal sequence in H. Then,

Proof. Letting

IXI

= (x I ei), we have for all n

n

eill 2 = IIxl12 - L IlXil2.

Os Ilx - i'"1>Xl
1
1= 1

Thus,
n

L \cxi \2 S; I\xl\2

i. == 1

for all n.

Hence,
00

L IIXil 2S
i= 1

Ilx112. I

60

HILBERT SPACE

I<x

Since Bessel's inequality guarantees that

1
Ie i)iZ < 00, Theorem 1
guarantees that the series
1 (x I ej)e i converges to some element. We
now characterize that element.

Ir;.

Definition. If S is a subset of a Hilbert space, the closed subspace generated

by Sis [S].
Theorem 2. Let x be an element in a Hilbert space H and suppose {e;} is a
orthonormal sequence in H. Then the series

converges to an element
The difference vector x -

in the closed subspace M generated by the e;'s.

is orthogonal to M.

Proof The fact that the series converges is a consequence of

Theorem 1 and Bessel's inequality. Obviously, E M. The sequence of
partial sums Sn =
(xl ej)e j converges
For each) and for n >},
we have

Ii=,

(x -

Sn

I e) = (x -

L1 (x I e;)ei I e) = (x I e) -

(x I e)

j=

= O.

Therefore, by the continuity of the inner product lim (x n.... IX)

Sn

lej) ==

Iej ) = 0 for each }. Thus, x - is orthogonal to the subspace

generated by the e,'s. Again using the continuity of the inner product, we
deduce that x - is perpendicular to the closed subspace generated by the
e;'s. I

(x -

From Theorem 2 it is apparent that if the closed subspace generated by

the orthonormal set of e/s is the whole space, any vector in H can be
expanded as a series of the e;'s witIt coefficients equal to the Fourier
coefficients (x I ei)' In the next section, we turn to the problem of constructing an orthonormal sequence that generates the whole space.
*3.8

Complete Orthonormal Sequences

Suppose {e;} is a sequence of orthonormal vectors in a Hilbert space H.

Generally this sequence generates a linear subspace in H and, as we have
seen, any vector in the closure M of this subspace can be expressed uniquely
as the limit of an infinite series of the form
1 !Xi e j To express every
vector in H this way, it is necessary that the closed subspace M generated
by the e/s be the whole space.
Dofinililm. An orthonormalscqucncc {eJ in a Hilbert space H is said to be
complete if the closed subspace generated by the e;'s is H.

3.8

COMPLETE ORTHONORMAL SEQUENCES

61

The following result giv1es a simple criterion for an arbitrary orthonormal

sequence to be complete. We leave the proc.;' to the reader.

Lemma 1. An orthonormal sequence {el} in a Hilbert spa;;;e H is complete if

and only if the only vector orthogonal to each of the e / s is the null vector.
Example 1. Consider the Hilbert space L 2 [ -1, IJ consisting of squareintegrable functions on the interval [-1,1]. The independent functions
1, t, t 2 , , tn, ... generate the subspace of polynomials. The GramSchmidt procedure can be applied to the sequence 1, t, t 2 , '" to produce an
orthonormal sequence
Obviously the e/s, being linear combinations of polynomials, are themselves polynomials. It turns out that
2n + 1 .
- 2 - p"(t),

en<t) =

n = 0,1,2, ... ,

where the Pit) are the well-known Legendre polynomials

(_1)" d"
2"
p"(t) = -2"
n., d"
t {(I - t ) }.
We wish to show that this sequence is a complete orthonormal sequence in
L 2 [ -1, 1]. According to Lemma 1, it is sufficient to prove that there is

no nonzero vector orthogonal to the subspace of polynomials.

Assume that there exists an f E L 2 [ -1, 1] orthogonal to each t"; i.e.,
.
assume that

L/''f(t) dt
1

for n = 0,1,2, ....

The continuous function

t

F(t) = L/(-r) d-r

has F( -1) == F(1)

f-1

= 0, and it follows from integration by parts that

t"F(t) dt

t"+1
-f -1n+l
--f(t) dt =

t"+1

= n+l

for n = 0, 1, 2, ... , Thus the continuous function F is also orthogonal

to polynomials.
Since F is continuous, the Weierstrass approximation theorem applies:
Given B > 0, there is a polynomial

Q(t) =

La, t'
1=0

62

HILBERT SPACE

such that
IF(t) - Q(t)1 <
for all t

E [

-1, 1]. Therefore, since F is orthogonal to polynomials,

the last inequality being a special case of the Cauchy-Schwarz inequality.

It follows that

L
1

IIF(t)1 2 dt :::; 28 2

Since e is arbitrary and Fis continuous, we deduce that F(t) = and, hence,
thatf(t) = (almost everywhere). This proves that the subspace of polynomials is dense in 2[ -1, 1J.

Example 2. Consider the complex space 2[0, 211:]. The functions ek(t) =
(l/jfn)e ikt for k = 0, 1, 2, ... are easily seen by direct calculation to be
an orthonormal sequence. Furthermore, it may be shown, in a manner
similar to that for the Legendre polynomials, that the system is complete.
Therefore, we obtain the classical complex Fourier expansion for an
arbitrary function x E 2[0, 211:J
elkt

<Xl

x(t)=
where the Fourier coefficients

Ck

Ck = (x I ek)

I fkJk=-oo
211:

are evaluated according to the formula

1

= ;;;;: J
.....; 211:

21t

x(t)e

-ikt

dt .

3.9 Approximation and Fourier Series

Suppose again that we are given independent vectors Yl' Yz, ... , Yn
generating a subspace M of a Hilbert space H and wish to find the vector :
in M which minimizes IIx Rather than seeking to obtain : directly
as a linear combination of the Yi'S by solving the normal equations, we
can employ the Gram-Schmidt orthogonalization procedure together with
Fourier series.

3.9

APPROXIMATION AND FOURIER SERIES

63

First we apply the Gram-Schmidt procedure to {Ylo Y2, ... , Yn}, obtaining an orthonormal set {e 1 , e2' ... ', en} generating M. The vector is then
given by the Fourier series

" (x I ej)el
= 1=L
1
since x - is orthogonal to M. Thus, our original optimization problem is
. easily resolved once
independent vectors YI are orthonormalized. The
advantage of this method is that once the et's are found, the best approximation to any vector is easily computed.
Since solution to the approximation problem is equivalent to solution of
the normal equations, it is clear that the
procedure can be
interpreted as a procedure for inverting the Gram matrix. Conversely,
many effective algorithms for solving the normal equations have an interpretation in terms of the minimum norm problem in Hilbert space.
We have seen that the Gram-Schmidt procedure can be used to solve
a minimum norm approximation problem. It is interesting to note that
the Gram-Schmidt procedure can itself be viewed as an approximation
problem. Given a sequence {Yl' Y2, Y3' ... , Yn} of independent vectors, the
Gram-Schmidt procedure sets

k-l
Yk ek ='

IIYk -

The vector Yk problem:

L (Yk I ej)e

j= 1
k 1

L (Yk I ej)ejll

j=1

(Yk Iej)ej is the optimal error for the minimum

minimize

IIy"

where the minimum is over all in the subspace [Ylo Y2' ... , Yk-I] =
rei, e2' ... , ek-l]. The vector ek is just a normalized version of this error.
Thus, the Gram-Schmidt procedure consists of solving a series of minimum
norm approximation problems by use of the projecti(;m theorem.
Alternatively, the minimum norm approximation of x on the subspace
[YI, Y2, ... , YnJ can be found by applying the Gram-Schmidt procedure to
is found at
the sequence {Yl' Y2' Y:I' ... , Yn' x}. The optimal error x the last step.

64

HILBERT SPACE

OTHER MINIMUM NORM PROBLEMS

3.10 The Dual Approximation Problem

In Section 3.6 we considered in some detail the problem of approximating
arbitrary vector in a Hilbert space by a vector in a given finite-dimensional subspace. The projection theorem led to the normal equations which
could be solved for the best approximation. A major assumption in such
problems was the finite dimensionality of the subspace from which the
approximation was chosen. Finite dimensionality not only guarantees
closure (and hence existence of a solution) but leads to a feasible computation procedure for obtaining the solution.
In many important and interesting practical problems the subspace in
which the solution mllst lie is not finite dimensional. In such problems it is
generally not possible to reduce the problem to a finite set of linear equations. However, there is an important class of such problems that can be
reduced by the projection theorem to a finite set of linear equations similar
to the normal equations. In this section we study these problems and their
relation to the earlier approximation problem. We begin by pointing out a
trivial modification of the projection theorem applicable to linear varieties.

all

Theorem 1. (Restatement of Projection Theorem) Let M be a closed subspace of a Hilbert space H. Let .\ be a fixed element in H and let V be the
linear mriety x + M. Then there is a unique rector Xo in V of minimum norm.
Furthermore, Xo is orthogonal to M.
Proof The theorem is proved by translating V by - x so that it
becomes a closed subspace and the,n applying the projection theorem.
See Figure 3.4. I
A point of caution is necessary here. The minimum norm solution Xo is
not orthogonal to the linear variety V but to the subspace M from which V
is obtained.

Figure 3.4

Minimum 1I0rm to a linear variety

3.10

THE DUAL APPROXIMATION PROBLEM

65

Two special kinds of linear varieties are of particular interest in optimization theory because they lead to finite-dimensional problems. The first
is the n-dimensional linear variety consisting of points of the form
x + Li': I aj XI where {Xl' X2, ... , Xn} is a linearly independent set in Hand
X is a fixed vector in H. Problems which seek minimum norm vectors in an
n-dimensional variety can be reduced to the solution of an n-dimensional
set of normal equations as developed in Section 3.6.
The second special type of linear variety consists of all vectors X in a
Hilbert space H satisfying conditions of the form
(x IYI) = CI

= C2

(XIY2)

where Ylo Y2' ... , Yn are a set of linearly independeQ,t vectors in Hand
Cl , C2' ... , Cn are fixed ';:onstants. Ifwe denote by M the subspace generated
by YI' Y2, ... , Yn' it is dear that if each C I == 0, then
linear variety is the
subspace M1.. For nonzero CI'S the resulting linear variety is a translation
of M 1.. A linear variety of this form is said to be of codimension n since
the orthogonal complement of the subspace producing it has dimension n.
We now consider the minimum norm problem of se'eking the closest
vector to the origin lying in a linear variety of finite codimension.

Theorem 2. Let H be Cl Hi/bert space and {YI' Y2' ... , Yn} a set of linearly
independent vectors in H. Among all vectors X E H satisfying
(XIYl) =

C1

(XIY2) =

C2

(X I Yn)

= Cn'

let Xo have the minimum norm. Then

"

Xo = LPIYI
1= I

where the coefficients PI satisfy the equations

(Yl I Yl)Pl
(Yl I Y2)Pl

+ (Y21 Yl)P2 + . " + (Yn I Yl)Pn, =

+ (Y21 Y2)P2 + ... + (Yn I Y2)Pn =

CI

C2

66

HILBERT SPACE

Proof. Let M be the (n-dimensional) subspace generated by the vectors

Yi' The linear variety defined by the constraints of the minimization
problem is a translation of the subspace M 1.. Since M 1. is closed, existence
and uniqueness of an optimal solution follow from the modified projection
theorem. Furthermore, the optimal solution Xo is orthogonal to M 1.. Thus,
Xo E MH. However, since M is closed, ME == M. We therefore conclude
that Xo E M or, equivalently, that
n

L f3iYi'
i= 1

Xo =

The n coefficients f3 i are chosen so that Xo satisfies the n constraints. This

leads to the n equations displayed in the
statement. I
Example 1. The shaft angular velocity w of a d-c motor driven from a
variable current source u is governed by the following first-order differential
equation:
w(t)

+ wet) = u(t),

where u(t) is the field current at time t. The angular position of the motor
shaft is the time integral of w. Assume that
motor is initially at rest,
8(0) = w(O) = 0, and that it is desired to find the field current function u
of minimum energy which rotates the shaft to the new rest position = I,
w = 0 within one second. The energy is assumed to be proportional to
g u2(t) dt. This is a simple control problem in which the cost criterion
depends only on the control function u. The problem can be solved by
treating it as a minimum norm problem in the Hilbert space L 2 [O, 1].
We can integrate the first-order differential equation governing the motor
to obtain the explicit formula

(1)

for the final angular velocity corresponding to any control. From the equation wet) +O(t) = u(t) follows the equation

Jo

(2)

8(1) =

u(t) dt - w(1)

or, combined with equation (1),

(3)

0(1)

Jo {I - e(t-l)}u(t) dt.
1

3.10

THE DUAL APPROXIMATION PROBLEM

67

If the function u is regarded as an element of the Hilbert space L 2 [0, 1],

the above relations may be expressed as
(0(1)

= (Y1 I u)

8(1) = (Y21

u)

where Yl = e(t-1l, Y2 = I With these definitions the original problem is equivalent to that of finding
u E L 2 [0, 1] of minimum norm subject to the constraints
e(t- 1 l.

(4)

(Y11 u)

1 = (Y21 u).

According to Theorem 2, the optimal solution is in the subspace

generated by Yl and Y2' Equivalently,

u(t)

=OC1

+ OC2 et

where OC 1 and OC2 are chosen to satisfy the two constraints (4). Evaluating the
constants leads to the solution
1

u(t) == -3-

-e

[1

+e-

2et ].

We have observed that there are two basic forms of minimum norm
problems in Hilbert space that reduce to the solution of a finite number of
simultaneous linear equations. In their general forms, both problems are
.concerned with finding the shortest distance from a point to a linear variety.
In one case the linear variety has finite dimension; in the other it has
finite codimension. To understand the relation between these two problems, which is one of the simplest examples of a duality relation, consider
Figure 3.5. Let M be a closed subspace of a Hilbert space H and let x be an
arbitrary vector in H. We may then formulate two problems: one of
projecting x onto M and the other of projecting x onto M 1. The situation
is completely symmetrical since M l.l. = M. If we solve one of these problems, the other is automatically solved in the process since if, for instance,

Figure 3.5 Dual projectIOn problems

68

HILBERT SPACE

mo is the projection of x onto M, then x - mo is the projection of x onto

M 1.. Therefore, if either M or M 1. is finite dimensional, both problems can
be reduced to that of solving a finite set of linear equations.
*3.11

A Control Problem

In this section we indicate how the projection theorem can be applied to

more complicated problems to establish existence and uniqueness results.
We consider an optimal control problem that is more difficult than the one
discussed in the previous section but which is of great practical importance.
Suppose we seek to minimize the quadratic objective

Jo {x (t) + u (t)} dt
T

J =

where x and u are connected by the differentiaR equation

x(O) given.

x(t)=u(t);

(1)

Problems of this form arise when it is desired to reduce x to zero quickly

by suitable application of control u. The quadratic objective represents a
common compromise between a desire to have x small while simultaneously conserving control energy.
For notational ease we have selected the simplest possible differential
equation to represent the controlled systems, but the techniques developed
below apply equally well to more complex differential equations.
Let us begin by replacing equation (1) by the equivalent constraint

(2)

x(t) = xeD)

+ fu(r) dr.
o

We now may formulate our problem in the Hilbert space

H

= L 2 [O, TJ x L 2 [O, TJ

consisting of ordered pairs (x, u) of square-integrable functions on [0, T]

with inner product defined by
T

fo [X 1(t)X2(t) + ul(t)uz(t)] dt

Xl' u t ) I (X2' u 2 =

and the corresponding norm

II(x, u)11 2

fo [x 2 (t)

+ u 2(t)] dt.

The set of elements (x, u) E H satisfying the constraint (2) is a linear variety
V in H. Thus, abstractly the control problem is one of finding the element
(x, u) E V having minimum norm.

, 3.12

MINIMUM DISTANCE TO A CONVEX SET

69

To establish the existence and uniqueness of a solution, it is sufficient to

prove that the linear variety V is closed. For this purpose, let {(Xn' Un)} be a
sequence of elements from V converging to an element (x, u). To prove that
Vis closed, we must show that (x, u) e V. Letting y(t) = x(O) + u(t) dr,
we must show that .x = y. We havey(t) - xit) = [u('t) - ui't)] d't. Thus,
by the Cauchy-Schwarz inequality (applied to the functions 1 and u - un),
Iy(t) - Xn(t)12 ::::; t

{Iu('t) - un(t)12 d't::::; Tllu - unl1 2

o

xnll ::::; Tllu - unll. It follows that

lIy - xII ::::; lIy - xnll + IIxn- xII ::::; Tllu - unll + IIxn - xII.

and, hence, integrating from 0 to T, Ily -

Both terms on the right tend to zero as n -+ 00, from which we conclude
that x = y, the desired result.
The general quadratic loss control problem is treated further in Section 9.5.
3.12 Minimum Distance to a Convex Set
Much of the discussion in the previous sections can be generalized from
linear varieties to convex sets., The following theorem, which treats the
minimum norm problem illustrated in Figure 3.6, is a direct extension of
the proof of the projection theorem.

Figure 3.6 Minimum distance to a convex set

Theorem 1. Let x be a vector in a Hilbert space H and let K be a closed

convex subset 0/ H. Then there is a unique vector ko e,K such that

IIx-ko'lI::::; IIx-kll

for all k E K. Furthermore, a necessary and sufficient condition that ko he

the unique minimizing vector is that (x - ko 1k - k o) ::::; O/or all k e K.

70

HILBERT SPACE

Proof. To prove existence, let {kJ be a sequence from K such that

Ilx - kill -+ 0 = inf IIx - kll.
kEK

By the parallelogram law,

Ilk; - kj

l1

2 IIk i - xl12 + 2 IIkj - xllz - 4/1 x - kl

By convexity of K, (k;

+ k j )/2 is in K; hence,

and therefore
Ilk; - kjl12 :$ 2 IIk i - Xll2

+ 2 IIk j

- xllz - 40 2-+ O.

The sequence {k i } is Cauchy and hence convergent to an element ko E K.

By continuity, IIx - koll = 0.
To prove uniqueness, suppose kl E K with IIx - k t II ::::: 0. The sequence

n even
n odd
has llx - knll-+ 0 so, by the above argument, {k n} is Cauchy and convergent. This can only happen if kl = k o .
We show now that if ko is the unique minimizing vector, then
(x - ko Ik - k o) :$ 0

for all k E K. Suppose to the contrary that there is a vector kl E K such

that (x - ko I kl - k o) = B> O. Conside\- the vectors k,. = (1 -rt)k o + rtk 1 ;
o : ;:; rt :$ 1. Since K is convex, each k,. E K. Also
IIx - k,.11 2 = 11(1 - a) (x - k o) + a(x - k 1)11 2
z
= (1 -a)ZIIx - koll + 2rt(l- a)(x - kolx - k )+o:21Ix-k I1 2.
1

The quantity IIx - k,.11 2 is a differentiable function of rt with derivative at

ex = 0 equal to
d
. da

IIx -

k,,11 2/ a=O = -2 IIx - koll 2+ 2(x - ko Ix - k 1 )

= -2(x-k o lk t

-k o)= -2B<0.

Thus for some small positive ex, IIx - k,.11 < IIx - koll which contradicts the
minimizing property of k o . Hence, no such kl can exist.

3.12

MINIMUM DISTANCE TO A CONVEX SET

71

Conversely, suppose that ko e K is such that (x - ko I k - k o) :s; 0 for all

K. Then for any k E K, k =F ko, we have

Ilx - kl1 2

= Ilx - ko

and therefore ko is

+ ko - kll 2 = IIx - k oll 2

+ 2(x - ko I ko - k) + Illko\- kl1 2 >
a unique minimizing vector. I

IIx - k oll 2

Example 1. As an applic:ation of the above result, we consider an approximation problem with restrictions on the coefficients. Let {Yt, Y2, ... ,Yn} be
linearly independent vectors in a Hilbert space H. Given x e H, we-seek
to minimize IIx - IXtYt -- 1X2Y2 - - IXnYnll where we require IX; 0 for
each i. Such a restriction is common in many physical problems.
This general problem can be formulated abstractly as that of finding the
minimum distance from the pointx to the convex cone

K = {y: Y =

IXtYt

+ ... + IXnYn, IX;

0 each i}.

This cone is obviously dosed and there is therefore a unique minimizing

vector. The minimizing vector = IXtYt + ... + 1X.,Yn must satisfy

Ik

(x --

Setting k =

+ Y; leads to
for i

(x-jcly;)sO

and setting k

all k eK.

:s; 0,

= -

IX/Y;

= 1,2, ... , n

leads to

(x -

if IX; > O.

Iy;)

Thus it follows that

for i = 1, 2, ... , n
with equality if IX; > O.
Letting G be the Gra.m matrix of the y;'s and letting b; = (x IYI), we
obtain the equation
(1)

GIX -

b =z

for some vector z with components Zl O. (Here IX and b are vectors with
components IX; and b;, respectively.) Furthermore, IX;Z/ = 0 for each i or,
more compactly,
(2)

IX'Z

= O.

Conditions (1) and (2) are the analog of the normal equations. They
represent necessary and sufficient conditions for the solution ofthe approximation problem but are not easily solved since they are nonlinear.

72

HILBERT SPACE

The dual of the approximation problem discussed above is that of

minimizing /lxll when xis
to inequality constraints of the. form
(xl Yi) ;::;: C i This topic is treated in the problems and in Chapter 5.

3.13

Problems

1. Let x andy be vectors in a pre-Hilbert space. Show that I(x Iy)1 = Ilxllllyll
if and only if i1.X + fiy = for some scalars i1., {3.
2. Consider the set X of all real functions x defined on (- 00, 00) for
which

lim T->Cfj 2T

J Ix(tW dt <
T

00.

-T

It is easily verified that these form a linear vector space. Let M be the
subspace on which the indicated limit is zero.
(a) Show that the space H::= XjM becomes a pre-Hilbert space when
the inner product is defined as

([x] I [y]) =

1 T
2T L/(t)y(t) dt

(b) Show that H is not separable.

3. Let H consist of all m x n real matrices with addition and scalar multiplication defined as the usual corresponding operations with matrices,
and with the inner product of two matrices A, B defined as
(A I B)

= Trace

[A' QB]

where A I denotes the transpose of tbe matrix A and Q is a symmetric,

positive-definite m x m matrix. Prove that H is a Hilbert space.
4. Show that if g(Xl' x 2 , ... , x n ) = 0, the normal equations possess a
solution but it is not unique.
2
5. Find the linear function x(t) = a + bt minimizing
1 [t - x(t)]2 dt.
6. Given a function x E L 2 [0, 1], we seek a polynomial p of degree n or
less which minimizes
Ix(t) - p(t)1 2 dt while satisfying the requirement p(t) dt = 0.
(a) Show that this problem has a unique solution.
(b) Show that this problem can be solved by first finding the polynomial q of degree n or less which minimizes Sb Ix(t) - q(t)12 dt and
then finding p of degree n or less which minimizes
Iq(t) -p(tW dt
while satisfying SA p(t) dt = 0.

S6

J6

7. Let M and N be orthogonal closed subspaces of a Hilbert space Hand

let x be an arbitrary vector in H. Show that the subspace M

Ee N

is

3.13

PROBLEMS

73

closed a.nd that the orthogonal projection of x onto M EEl N is equal to

m + n, wherem is the orthogonal projection of x onto M and n is the
. orthogonal projection of x onto N.
8. Let {Xl' X2, ... ,xm} and {Yl' Y2, ... ,Yn} each be sets of linearly
independent vectors in a Hilbert space generating the subspaces M and
N, respectively. Given a vector x E H, it is desired to find the best
approximation (in the sense of the r.orm) to x in the subspace M n N.
(a) Give a set of equations analogous to the normal equations which
when solved
(b) Give a geometrical interpretation 01 the method of solution.
(c) Give a computational procedure for producing
Hint: Consider the dual problem and use Problem 7.
9. Prove part 5 of Proposition 1, Section 3.4.
10. A Hilbert space H of functions on a set S is said to bea reproducing
kernel Hilbert space iif there is a function K defined on S x S having the
properties: (1) K( . , t) E H for each t E S, and (2) x(t) = (x IK( . , t))
for each x E H, t E S. Such a function K is, called a reproducing kernel.
Prove that a reproducing kernel, if it exists, is unique.
11. Suppose H with reproducing kernel K is a closed subspace of a Hilbert
is the projection
space X. Show that for x E X the function (x I K( . ,
of x onto H.
12. Suppose randomly varying data in the form of a function x(t) is
observed from t = 0 to the present time t = T. One proposed metnod
for predicting the future data (t > T) is by fitting an (n - l)-th degree
polynomial to the past data and using the extrapolated values of the
polynomial as the e:stimate.
Specifically, suppose that the polynomial,

peT, t) == a1(T)

+ a2(T)t + a3(T)t 2 + ... + an(T)t n - 1

minimizes
T

fo [x(t) -

peT, t)]2 dt.

Show that the coefficients ai(T) need not be completely recomputed for
each T but rather can be continuously updated according to a formula
of the form

. where e(T) = x(T)

fixed constants.

-- p(T, T) is the instantaneous error and the bi's are

74

IDLBERT SPACE

13, Show that the Gram determinant g(x" X2, ... , x n) is never negative
(thus generalizing the Cauchy-Schwarz inequality which can be
expressed as g(Xl, X2) ;;::; 0).
14. Let {Yl' Y2, ... , Yn} be linearly independent vectors in a pre-Hilbert
space X and let x be an arbitrary element of X. Show that the best
approximation to 'x in the subspace generated by the y/s has the
explicit representation
(Yn I YI) (x I YI)
(Yn I Y2) (x I Y2)

(Yl I Yl)
(YI I Y2)

(Y21 Yt)
(Y21 Y2)

(YI I Yn)
YI

(Y21 Yn)
(Yn I Yn)
Y2
Yn
-gCYI' Y2,"" Yn)

(x I Yn)
()

where the determinant in the numerator is to be expanded algebraically

to yield a linear combination of the Y /s. Show that the minimum error
- x is given by the identical formula except for x replacing 0 in the
determinant.
15. (Muntz's Theorem) It was shown in Example 1, Section 3.8, that the
functions l,t,t 2 , generate a dense subspace of L 2 [-I,I] (or
L2[0, 1] by a simple translation and scalar factor). In this problem we
prove that the functions tn" (n2, . ; I n l < n2 < ... , generate a
dense subspace of L2 [0, I] if and only if the integer's ni satisfy

00

2: - = 00.
i'" ni
I

(a) Let Mk = [tn" tnz, ... , t nk ]' The result holds if and only if for each
m ;;::; 1 the minimal distance dk of from Mk goes to zero as k goes to
infinity. This is equivalent to

t.:

. g(ln" tn2, ... , tnk ,

I1m
n,
n2
nk'
k-o 00
g( t , t , ... , t )

= O.

Show that
k

d k2

ITI (ni -

m)2

i=

(2m

+ 1)TI(m + ni + 1)2
i= 1

(b) Show that a series of the form 2:r; I log (1 + ai) diverges ifand only
1 a I diverges.
if
(c) Show that lim log d/ = - 00 jf and only if 'L;';ll/ni = 00.

L{;

k'" 00

3.13

PROBLEMS

75

16. Prove Parseval's equality: An orthonormal sequence {ei}i"=l is complete in a Hilbert space H if and only if for each x, y in H
(x Iy)

OC>

=1=L1(x I et)(etly)

17. Let {Yl'Y2' ... ,Yn} be independent and suppose {eto e2, ... , en} are

obtained from the y;'s by the Gram-Schmidt procedure .. Let

n

L (xl el)el =i=L lXiYi'

1= 1

Show that the coefficients lXi can be easily obtained from the Fourier
coefficients (x lei).
18. Let w(t) be a positive (weight) function defined on an interval [a, b] of
the real line. Assume that the integrals
exist for n = 1, 2, ....
Define the inner product of two polynomials p and q as

(p I q)

= f p(t)q(t)w(t) dt.
b

Beginning with the sequence {I, t, t 2 , ., .}, we can employ the GramSchmidt procedure to produce a sequence of orthonormal polynomials
with respect to this weight function. Show that the zeros of the real
orthonormal polynomials are real, simple, and located on the interior
of [a, b].
19. A sequence of orthonormal polynomials {en} (with respect to a
weighting function on a given interval) can be generated by applying
the Gram-Schmidt procedure to the vectors {I, t, (2, }. The procedure is straightforward but becomes progressively more complex with
each new term. A superior technique, especially suited to machine computation, is to exploit a
relation of the form
n

= 2, 3, ....

Show that such a recursive relatIon I;;xists among the orthonormal polynomials and det.ermine the coefficients an' bn,
.
20. Suppose we are to set up a special manufacturing company which will
operate for only ten months. During the ten m"nths the company is to
produce one million copies of a single product. We assume that the
manufacturing facilities have been leased for the ten-month period, but
that labor has not yet been hired. Presumably, employees will be hired

76

HILBERT SPACE

and fired during the ten-month period. Our problem is to determine how
many employees should be hired or fired in each of the ten months.
It is assumed that each employee produces one hundred items per
month. The cost of labor is proportional to the number of productive
workers, but there is an additional cost due to hiring and firing. If
u(k) workers are hired in the k-th month (negative u(k) corresponds to
firings), the processing cost can be argued to be u2 (k) because, as u
increases, people must be paid to stand in line and more nonproductive
employees must be paid.
At the end of the ten-month period all workers must be fired. Find
u(k) for k = 1,2, ... ,10.
21. Using the projection theorem, solve the finite-dimensional problem:
minimize x' Qx
subject to Ax = b
where x is an n-vector, Q a positive-definite symmetric matrix, A an
< n), and b an m-vector.
22. Let x be a vector in a Hilbert space H, and suppose {Xl' X2, . , xn}
and {Yl' Yz , ... , Ym} are sets of linearly independent vectors in H. We
seek the vector minimizing Ilx - xii while satisfying:
m x n matrix (m

x E M = [xl> X2, , XII]

i = 1,2, ... , m.

(xly;)=c;,

(a) Find equations for the solution which are similar to the normal
equations.
(b) Give a geometric interpretation of the equations.
23. Consider the problem of finding the vector X of minimum norm
satisfying

(X IYi)

ci ,

..

= 1,2, ... ,11

where the Yi'S are linearly independent.

(a) Show that this problem has a unique solution.
(b) Show that a necessary and sufficient condition that
n

X=

Lai)'i

i= f

be the solution is that the vector a with components a l satisfy

G'a

and that aj = 0 if (x IYi) > c i G is the Gram matrix of {Yll Y2' .. " Yn}.

3.13

PROBLEMS

77

24. The following the:orem is valid in a Hilbert space H. If K is a closed

convex set in H and x E H, x K; there is a unique vector ko e K such
that Ilx - koll :$ Ux - kll all k E K. Show that this theorem does not
apply in arbitrary Banach space.

REFERENCES
3.1-3.8. Most of the texts on functional analysis listed as references for
Chapter 2 contain discussions of pre-Hilbert space and Hilbert space. In
addition, see Berberian [21], or Akhiezer and Glazman [3] for introductory

treatments.
3.9. Fot" a discussion of Gram matrices and polynomial approximation, see

Davis

136] or Natanscm [109].

3.l0. For some additional applications of the dual problem, see Davis [36].

3.ll. For a different proof of the existence and uniqueness to this type of
control problem, see Bellman, Glicksberg, and Gross [17].
3.13. For Problem 12, see Luenberger [99]. Reproducing kernel Hilbert space
introduced in Problems 10 and 11 is a useful concept in approximation and
estimation problems. See Aronszajn [11] and Parzen [116]. For additional
material on polynomial approximation and a treatment of Problems 14, 15,
18, and 19, see Davis [36].

4
LEAST-SQUARES ESTIMATION

4.1

Introduction

Perhaps the richest and most exciting area of application of the projection
theorem is the area of statistical estimation. It appears in virtually all
branches of science, engineering, and social science for the analysis of
experimental data, for control of systems subject to random disturbances,
or for decision making based on incomplete information.
All estimation problems discussed in this chapter are ultimatlely formulated as equivalent minimum norm problems in Hilbert space and are
resolved by an appropriate application of the projection theorem. This
approach has several practical advantages but limits our estimation criteria
to various forms of least squares. At the outset, however, it
be
pointed out that there are a number of different least-squares estimation
procedures which as a group offer broad flexibility in problem foqnulation.
The differences lie primarily in the choice of optimality criterion a\nd in the
statistical assumptions required. In this chapter three basic forms of least
squares estimation are distinguished and examined.
Least squares is, of course, only one of several established approaches to
estimation theory, the main alternatives being maximum likelihood ann
Bayesian techniques. These other techniques usually require a
statistical description of the problem variables in terms of joint probability
distribution functions, whereas least squares requires only means, variances,
and covariances. Although a thorough study of estimation theory would
certainly include other approaches as well as least squares, we limit our
discussion to those techniques that are derived as applications of the projection theorem. In complicated, multivariable problems the equatiol)s
resulting from the other approaches are often nonlinear, difficult to solve,
and impractical to implement. It is only when all variables have GaussLln
statistics that these techniques produce linear equations, in which case the
estimate is identical with that obtained by least squares. In many practical
situations then, the analyst is forced by the complexity of the problem to
either assume Gaussian statistics or to employ a least-squares approac:h.
Since the resulting estimates are identical, which is used is primarily a
matter of taste.

78

4.2

HILBERT SPACE OF RANDOM VARIABLES

79

The first few sections of this chapter are devoted to constructing various
Hilbert spaces of random variables and to finding the appropriate normal
equations for three bask estimation problems. No new optimization techniques are involved, but these problems provide interesting applications
of the projection theorem.

4.2 Hilbert Space of R.andom Variables

A complete and rigorolls definition of a random variable and associated
probabilistic concepts is in the realm of measure and integration theory.
The topic of least-squar1es estimation, however, makes only casual, indirect
reference to the machinery available for a thorough study of probability.
Consequently, the notions of probability measure and related concepts
can be largely suppressed from the studies cfthis chapter. We assume only
a rudimentary familiarity with random variables, probability distributions,
and expected value. Of these, only the concept of
value is used
explicitly in least-squares estimation. With this in mind, we now briefly
review those probabilistic concepts pertinent to our development.
If x is a real-valued random variable, we define the probability distribu-

tionPofxby

= Prob (x :s;
In other words,
is the probability that the random variable x assumes
a value less than or equal to the number
The expected value of any
function g of x is then
as
E[g(x)]

which may in general not be finite. Of primary interest are the quantities
the expected value of x,

E(x),
Z

E(x ),

the second moment of x,

E[(x -- E(X2],

the variance ofx.

Given a finite collec1tion of real random variables {Xl> X2,' .. , x n }, we

define their joint probability distribution P as
= Prob (Xl :s;

1,

X2 :s;

Xn :s;

i.e., the probability of the simultaneous occurrence of Xi :s;

The expected value of any function g of the xt's is defirled as

for all i.

80

LEAST-SQUARES ESTIMATION

All of the second-order statistical averages of these variables are

described by then expected values
i;:: 1,2, ... , n,

and the n x n covariance matrix cov (Xl' X2 , , xn) whose ij-th element is
defined as
E {[Xi - E(Xi)] [Xj - E(Xj )]}
which in: case of zero means reduces to E(xixj)'
Two random variables Xi' Xj are said to be uncorrelated if E(XiXj) =
E(xj) E(x) or, equivalently, if the ij-th element of the covariance matrix
is zero.
With this elementary background material, we now construct a Hilbert
space of random variables. Let {Yl, Y2, ... , Ym} be a finite collection of
random variables with E(y/) < C1) for each i. We define a Hilbert space H
consisting of all random variables that are linear combinations of the y;'s.
The inner product of two elements x, Y in H is defined as
(x I y)

= E(xy).

Since x, yare linear combinations of the y/s, their inner product can be
calculated from the second-order statistics of the Yi'S. In particular, if
X = LCtiYi,Y = LPiYi' then
E(xy) = E{

Ct j

Yi)

Pj

yj)}'

The space H is a finite-dimensional Hilbert space with dimension equal to

at most m. (If the matrix G = [E(YiY)] is not positive definite, there
will be nontrivial linear combinations L CtiYi having zero norm in the space
H. These combinations must be considered equivalent to the zero element,
thereby reducing the dimension of H.)
If in the Hilbert space H each random variable has expected value equal
to zero, then two vectors x, z are orthogonal if they are uncorrelated:
(x I z) = E(x) E(z) = O.
The concept of a random
can be generalized in an important
direction. An n-dimensional vector-valued random variable X is an ordered
collection of 11 scalar-valued random variables. Notationally, x is written
as a column vector

4.2

HILBERT SPACE OF RANDOM VARIABLES

81

(the components being random variables). For brevity, a vector-valued

random variable is referred to simply as a random vector. In applications,
of a random vector may, for
the random variables which are
example, be associated with the parameters of a single experiment. Or,
they may correspond to results of repeated trials uf the same experiment
or be elements in a random time series such as barometric pressures on
five consecutive days.
A Hilbert space of random vectors can be generated from a given set of
random vectors in a manner analogous to that for random variables.
Suppose {h, Y2, ... , Ym} is a collection of n-dimensional random vectors.
Each element YI has n components Yij' j = I, 2, ... ', n, each of which is a
random variable with finite variance. We define the Hilbert space .lit' of
n-dimensional random vectors as consisting of all vectors whose components are linear combinations of the components of the y/s. Thus an
arbitrary element y in this space can be expressed as
Y

= Kl Yl + K2 Y2 + ... + Km Ym

where the K/s are real 11 x n matrices. The resulting space is, of course,
generally larger than that which would be obtained by simply considering
linear combinations of the y/s. This specific compound construction of.llt',
although rarely
to explicitly in our development, is implicitly
responsible for the simplicity of many standard results of estimation theory.
If x and z are elements of:?t, we define their inner product as

which is the expe<:ted value of the n-dimensional inner product. A convenient notation is (x I z) = E(x'z).
The norm of an element x in the space of n-dimensional random vectors
can be written

II x II

= {Trace E(xX'W I 2,

where

xn)]

E(XI Xl) E(XI X2) ... E(x t

E(X2 Xt) E(X2 X2) ... E(x 2 x n)

E(xx') =

[
E(X"X2)

.. I

E(xnxn)

is the expected value of the random latrix dyad xx'. Similarly, we have
(x Iz) ,; Trace E(xz').

82

LEASTSQUARES ESTIMATION

In the c'ase Of zero means the matrix E(xx') is simply, the covariance
matrix of the random variables Xi which form the components of x, These
components can, of course, be regarded as elements of a " smaller" Hilbert
space of random variables; the covariance matrix above then is seen to be
the Gram matrix corresponding to {Xl' X2, . , x n }.
Corresponding to the general case with nonzero means, we define the
corresponding covariance matrix by
cov (x) = E [(x - E(x (x - E(x'].

4.3 The Least.Squares Estimate

The purposes of this and the next two sections are to formulate and to solve
three basic linear estimation problems. The three problems are quite
similar, differing only slightly in the choice of optimality criterion and in the
underlying statistical assumptions,
Suppose a quantity of data consisting of m real numbers has been
collected and arranged as the m components of a vector y. Often the nature
of the source of the data leads one to assume that the vector y, rather than
consisting of m independent components, is a given linear function of a
few unknown parameters. If these parameters are arranged as the com
ponents of an n-dimensional vector 13 (where 11 < m), such a hypothesis
amounts to assuming that the vector y is of the form y = Wf3. The m x n
matrix W is, by assumption, determined by the particular experiment or
physical situation at hand and is assumed to be known. The data vector y
is known. The problem is to determine the vector 13. Since n < m, however,
it is generally not possible to determine a vector 13 exactly satisfying y = Wf3.
A useful alternative consists of determining the value of 13 which best
approximates a solution in the sense of minimizing the norm II y .;... Wf3ll.
If this norm is taken as the standard Euclidean mspace norm, this
approach leads to a simple least-squares estimate.
As formulated above, this problem is not a statistical one. It simply
lying in the subspace spanned by,
amounts to approximating y by a
the n column vectors of the matrix W. However, the technique is often
used in a statistical setting and it provides a useful comparison for other
statistical methods. For example, the procedure might be applied to finding
a straight-line approximation to the data in Figure 4.1 (which may represent
experimental data gathered on distance traveled versus time for a body
under constant acceleration). We hypothesize a model of the form s = tf3
and choose 13 by least squares. The y vector in this case is made up of the
measured s values; the W matrix consists of a single column made up of the

4.3

THE LEAST-SQUARES ESTIMATE

83

corresponding t values. The error is due to the unaccounted nonlinearity

and the measurem(:ht errors. If the hypothesized model were originally
of the forms = tPl + t 2P2' a better fit could be made.
s

Figure 4.1

Straight-line approximation to data

Theorem 1. (Least-Squares Estimate) Suppose y is an m vector and Wan

m x n matrix with linearly independent columns. Then there is a unique n
vector P which minimizes Ily - WPII over all P (the norm taken as the
Euclidean m-space norm). Furthermore,

Proof. As pointed out above, this problem amounts to approximating

y by a linear combination of columns of W. The existence and uniqueness
follow immediately from the projection theorem and the independence of
the columns of W. Furthermore, the Gram matrix corresponding to the
column vectors of Wis easily seen to be W'W. The vector W'y has as its
components inner products of the columns of W with the vector y. Hence
the normal equations become

W'WP= W'y.
Since the columns of Ware assumed to be linearly independent, the Gram
matrix W' W is nonsingular and the result follows. I
There is an extensive theory dealing with the case where W' W is singular.
See the references at the end of this chapter and the discussion of pseudoinverses in Section 6.11.
,
Although Theorem 1 is actually only a simple finite-dimensional version
of the general approximation problem treated in Chapter 3, the solution is
stated here explicitly in matrix notation so that the result can be easily
compared with other estimation techniques derived in this chapter.

84

LEAST-SQUARES ESTIMATION

4.4 Minimum-Variance Unbiased Estimate (Gauss-Markov Estimate)

Assume now that we have arranged an experiment that leads to an
m-dimensional data vector y of the form
y

= Wp + B.

In this model W is a known matrix, p is an It-dimensional unknown vector

of parameters, and e is an m-dimensional random vector with zero mean
and with covariance (ee') = Q which we assume to be positive definite.
The vector y can be interpreted as representing the outcome of m inexact
measurements, the random vector e representing the measurement errors.
For example, repeated measurements of a single quantity p might be
represented as Yi = P+ ei , in which case W would contain a single column
with each component equal to unity.
In this section we consider a method for estimating the unknown vector
of parameters Pfrom the vector y. In particular we seek a linear estimate P
oftheform
P=Ky,
where K is a constant n x m matrix.
Since y is the sum of random vector B and the constant vector Wp, it is
itself a random vector. Therefore, both the estimate Pand the error P- P
are random vectors with statistics determined by those of e and the choice
of K. A natural criterion for the optimality of the estimation scheme is
minimization of the norm of the error, expressed in this case as

E[II P- PII 2 ],
where 1/ I is the ordinary Euclidean n-space norm. If, however, this error
is written explicitly in terms of the problem variables, we obtain
(1)

E[IIP - P11 2 ]

= E[IIKy - P11 2 ] = E[IIK Wp + Ke - P1I 2]

= 11K Wp -

PI1 2 + Trace (KQK'),

The matrix K minimizing this expression is obviously a function of the

(unknown) parameter vector p. Therefore, a practical estimation scheme
for the parameters P cannot be derived as a solution to the proposed
problem.
We observe, however, that if K W = I (the identity matrix), the norm
of the error is independent of p. This observation suggests the alternative
problem: find the estimate P = Ky minimizing E[IIP - P1I 2 ] while
satisfying
(2)

KW=I.

The solution to this problem is independent of p.

4.4

MINIMUM-VARIANCE UNBIASED ESTIMATE (GAUSS-MARKOV)

85

The additional requirement K W = I may at first seem to be highly

arbitrary and pel:haps inappropriate. The requirement has a simple interpretation, however, that tends to defend its introduction. The expected
value of the estimate is E(P) == E(Ky) = E(KWP + Ke) = KWp. If KW= I,
the expression reduces to E(P) = p. Thus the expected value of the estimate
of P is itself p. In general, estimators with this property are said to be
unbiased. Thus the requirement K W = I leads to an unbiased estimator.
Conversely, if we require that an estimator of the form P== Ky be unbiased
in the sense that E(P) == Pfor all p, it follows that K W = I. We seek the
unbiased linear estimate of P which minimizes E(IIP _ PII 2 ).
Before plunging into the detailed analysis of this problem, it behooves
us to make some: elementary observations concerning its structure. The
problem can be written out in terms of the components of P as that of
finding Pto minimize

subject to

i = 1,2, .... n,
and
Pi
A

== k'iY

i = 1,2, ... , n,

where k/ is the i-th row of the matrix K. Therefore, the problem is really
n separate problems, one for each Pi' Minimizing each E(PI - Pi)2
minimizes the sum. Each subproblem can be considered as a constrained
minimum norm problem in a Hilbert space of random variables.
An alternative viewpoint is to consider the equivalent deterministic
problem of selecting the optimal matrix K. Returning to equations (1) and
(2), the problem is: one ofselecting the n x m matrix K to
minimize
sul::dect to

Trace {KQK'}
KW=I.

This problem can be regarded as a minimum norm problem in the space

of matrices (see Problem 3, Chapter 3), or it may be decomposed in a
manner analogous to that described for the components of p. The j-th
subproblem is
minimize
subject to
where

Wj

k/Qki

k/wj == Oi}'

is the j-th column of Wand

Oij

j == 1,2, ... , n,
is the Kronecker delta function.

86

LEAST-SQUARES ESTIMATION

Introducing the inner product (x IY)Q ::::: x' Qy, the problem becomes
minimize
subject to

= 1,2, ... , n.

This is now in the form of the standard minimum norm problem treated in
Section 3.10. It is a straightforward exercise to show that

kj

= Q-1W(W'Q-1W)-lei

where ei is the n-vector with i-th component unity and the rest zero. Note
that W' Q W is the Gram matrix of the columns of W with respect to (I)Q'
Finally, combining all m of the subproblems, we obtain

K'

= Q-I W(W'Q- 1 W)-1.

We summarize the above analysis by the following classical theorem.

Theorem 1. (Gauss-Markov) Suppose y =

Wp + /l where

(/l) = 0
(/l/l')

=Q

with Q positive definite. The linear minimum-variance unbiased estimate

olP is
with corresponding error covariance

Proof The derivation ofthe estimate is given above. It only remains to

calculate the corresponding error covariance
E[(P - fJ)(p - PY]

= E[(Ky - P)(Ky =E[K/l/l'K'J

PY]

= KQK' = (W'Q- 1W)-1W'Q-IQQ- 1W(W'Q- 1W)-1

=(W'Q- 1 W)-11
As an aside, it might be mentioned
that the justification for the terminI
ology "minimum variance unbiased" rather than "minimum covariance
trace unbiased" is that for each i,
is the minimum-variance unbiased
estimate of PI' In other words, the Gauss-Markov estimate provides a
minimum-variance unbiased estimate of each component rather than
merely a vector optimal in the sense of minimizing the sum of the individual
vanances.

Pi

4.5

MINIMUM-VARIANCE ESTIMATE

87

A striking property of the result of the Gauss-Markov theorem is that if

= I, the linear, minimum-variance unbiased estimate is identical
with the least-squares estimate of Section 4.3. It is clear that the leastsquares technique and the Gauss-Markov technique are intimately related.
However, a fundamental difference between the two approaches is that
least squares is a singJle elementary minimum norm problem while the
Gauss-Markov problem is actually n separete minimum norm problems.

E(88')

4.5

Minimum-Varianc:e Estimate

In the preceding two sections, the vector {3 was assumed to be a vector of

unknown parameters. Presumably these parameters could have assumed
any value from - 00 to + 00; we, as experimenters, had absolutely no
prior information conc:erning their values. In many situations, however,
we do have prior information and it is then perhaps more meaningful to
regard the unknown ve:ctor {3 as a random variable with known mean and
covariance. In such situations, this a priori statistical information can be
exploited to produce an estimate oflower error variance than the minimumvariance unbiased estimate.
In view of this observation, in this section we again consider estimation
of {3 from measurements of the form

= W{3 + e

but in this case both p and e are random vectors. The criterion for optimality is simply the minimization of E[IIS - {3112].
We begin by establishing an important theorem which applies in a somesetting than that described above but which is really
what more
only an application of the normal equations.
Theorem 1. (Minimum-Variance Estimate) Let y and {3 be random vectors
(not necessarily 0/ the same dimension). Assume [E(yy')] -1 exists. The
linear estimate S0/{3, based ony, minimizing E[IIP - {3112] is

P = E({3y')[E(yy')]-1 y ,
with corresponding error covariance matrix
E[(P - {3) (p

.- {3)'] =

E({3{3') - E(PP')
= E({3{3'} - E({3y') [E(yy')] -1 E(y{3'}.

Proof. This problem, like that in the last sectioh, decomposes into a
separate problem for leach component {3 i' Since there are no constraints,
subproblem is simply that of finding the best approximation of {3 i
the
within the subspace ge:nerated by the y;'s.

88

,LEAST-SQUARES ESTIMATION

Wrhing the optimal estimate as fl = Ky, where K is an n x m matrix,

the i-tIi subproblem is equivalent to the problem of selecting the i-th row
of K which, in turn, gives the optimal linear combination of the y/s for
Pi' Hence, each row of K satisfies the normal equations corresponding to
projecting 131 into the y/s. The normal equations for all n subproblems
can be written simultaneously in the matrix form

[E(yy')]K' = E(yf3')
from which

that
K = E(f3y')[E(yy')J -1,

which is the desired result. Proof of the formula for the error covariance
is obtained by direCt substitution. I
Note that, as in the previous section, the term minimum variance applied
to these estimates can be taken to imply that each component of 13 is
estimated by a minimum-variance estimator. Also note that if both 13 and y
have zeto means, the estimate is unbiased in the sense that E(P) = E(f3) O.
If the means are not zero, we usually broaden the class of estimators to
include. the form P= Ky + b where b is an appropriate constant vector.
This matter is considered in Problem 6.
Returning to our original purpose, we now apply the above theorem to
a revised form of our standard problem.

Corollary 1. Suppose
y = Wf3

+8

where yis a known m-dimensional data vector"f3 is an n-dimensional unknown

random vector, 8 is an unknown m-dimensional random error vector, W is
a known m x n constant matrix, and
E(88') = Q

E(f3f3') = R,

E(ef3')

= O.

We assume that Rand Q are positive-semidefinite matrices (of appropriate

size) and that WRW' + Q is nonsingu!ar. Then the linear estimate Pof 13
minimizing E[IIP - f311 2 J is

..

P = RW'(WRW'

+ Q)-I y

with err@r covariance

E[(f3 - p)(f3 -

Pn =

+ QrIWR.
E(yy') = WRW' + Q

R - RW'(WRW'

Proof It is easily computed that

and that
E(f3y') = RW' from which the result follows by the minimum-variance
theorem. I

4.S

MINIMUM-VARIANCE ESTIMATE

89

A significant difference between the estimation problem treated above

and those of previous sections is that the number of observations, m,
need not be at least as large as the number of unknowns, n. The estimate
in Corollary 1 exists if the m x m matrix (WRW' + Q)-I exists; this is the
case for any m if, for instance, Q is positive definite. Physically the reason
for this property is that since Pis now a random variable, it is possible to
form a meaningful estimate (namely zero) even with no measurements.
Every new measurement simply provides additional information which may
modify our original estimate.
Another unique feature of this estimate is that for m < n there need be
no measurement error. Thus, we may have Q = 0, because as long as
WRW' is positive definite the estimate still exists.
Example 1. Consider Pto be a random two-dimensional vector with mean
zero and covariance I (the identity matrix). Such a vector can be thought
of as having an expected length of unity but with a completely random
angle as measured from any given axis. Suppose that a single perfect
measurement y of the first component PI of Phas been obtained and from
this we seek the best estimate of p. The situation is illustrated in Figure 4.2.

............

,..'"
-::::::.-()

....

----

"-,,"-

"-

"-

"-,,-

....

Figure 4.2 Estimation from one component

Intuitively, we choose our estimate somewhere 01. the vertical line at

PI = y. Having no reason to favor one angle over another, we take the
shortest vector since it is closest to the mean. Hence we select the horizontal
vector that meets the vertical line. This solution can be verified by application of Corollary 1.
By some matrix ma.nipulations, the result of Corollary 1 can be translated
into a form that more closely resembles our earlier results.

90

LEAST-SQUARES ESTIMATION

Corollary 2. The estimate gil'etl by Corollary 1 call be expressed ill the

alternative form

lI'ith corresponding error covariance

[(/3 - {J)(# - {J)'J

= (W'Q-I W + R-1)-I ..

Proof The matrix identity

RW'(WRW'

+ Q)-I

= (W'Q-I W

+ R-1)-1 W'Q- 1

is easily established by postmultiplying by (WR W' + Q) and premultiplying by (W'Q-I W + R- 1 ). This establishes the equivalence of the
formula for the estimate p.
From Corollary I we have

[(/3 - (3)(p - (J)']

=R-

RW'(WRW'

+ Q)- 1 WR

which becomes, by application of the above matrix identity,

[(p - {J)(/3 - (J)']

=R -

(W'Q- 1W + R- 1)-1 W'Q-I WR

= (W'Q-1W

+ R-1)-I{(W'Q-1W + R-1)xR

- W'Q-1WR}

=(W'Q-1W+W1)-I1

In this form the minimum-variance estimate can be easily compared with

the least-squares and Gauss-Markov estimates. In particular, note that
if R- 1 = 0 (corresponding to infinite variance of the a priori information
concerning (J), the result of Corollary 2 is identical with the Gauss-Markov
estimate. Thus the Gauss-Markov estimate is simply a limiting case of
the minimum-variance estimate and, hence, further study of minimumvariance estimates, for the most part, also applies to Gauss-Markov
estimates.

4.6

Additional Properties of Minimum-Variance Estimates

In this section we investigate minimum-variance estimation in more detail

by considering three problems: the first is to deduce the optimal estimate
of a linear function of {J, the second is to determine how the optimal
estimate of fi changes if the optimality criterion is a more general quadratic
form, and the third is to determine how an estimate of {J is changed if
additional measurement data become available. We shall see that these
problems are related and that all three have strikingly simple solutions.

4.6

ADDITIONAL PROPERTIES OF MINIMUM-VARIANCE ESTIMATES

91

Theorem 1. The minimum-variance linear estimate of a linear function of

{3, based on the random vector y, is equal to the same linear function of the
minimum-variance linear estimate of {3; i.e., given an arbitrary p x n matrix

T, the best estimate ofT{3 is rp.

Proof The result can be obtained from the observation, made in the
last section, that the linear estimate P minimizing E[II{3 - P11 2] actually
minimizes E({3i - Pi for each component {3i of {3. We leave it to the reader
to complete a proof along these lines.
An alternate proof can be obtained directly from the projection theorem
by deriving the optimal estimate of T{3 and comparing the result with TP.
If ry is the optimal estimate of T{3, we must have
E[y(T{3 - ry)'J

=0

and, hence, in matrix form the normal equations for the columns of rare

[E(yy')]r' = E[y{3'T'J
so that

= T E({3y') [E(yy')] -1

which, by comparison with the minimum-variance estimate of {3, yields

the desired result. I
Another property of linear minimum-variance estimates, which is closely
related to the property considered above and which again can be regarded
asa simple consequence of the componentwise optimality of the estimate,
is that the estimate is optimal for any positive-semidefinite quadratic
optimality criterion.
Theorem 2. If P = Ky is the linear minimum-varianr:e estimate of {3, then

P is also the linear estimate minimizing E[({3 - P)'P({3 - p)]for any positivesemidefinite n x n matrix P.
Proof Let p I /2 be the unique positive-semidefinite square root of P.
According to Theorem 1, p I/ 2p is the minimum-variance estimate of
pl/2{3 and, hence, Pminimizes
Finally, we consider the problem of updating an optimal estimate of {3
if additional data become available. This result is of fundamental practical
importance in a number of modern sequential estidtation procedures such
as the recursive estimation of random processes discussed in Section 4.7.
Like the two properties discussed above, the answer to the updating problem
is extremely simple.

92

LEAST-SQUARES ESTIMATION

The procedure is based on the simple orthogonality properties of projection in Hilbert space. The idea is illustrated in Figure 4.3. If Y 1 and Y 2
are subspaces of a Hilbert space, the projection of a vector {3 onto the
subspace Y 1 + Y 2 is equal to the projection onto Y l plus the projection
onto Yz wtjere Y2 is orthogonal to Y I and is chosen so that Y 1 El3 Y2 =
Y I + Y 2 FUrthermore, if Y 2 is generated by a finite set of vectors,
the diffe"'!llces between these vectors and their projections onto Y l
generate Yz .

Figure 4.3

For clarity and simplicity, the following theorem is stated in terms of a

single unknown random variable {3 rather than an n-dimensional random
vector. Since minimum-variance estimation of an n-dimensional random
vector is merely n separate problems, the theorem statement easily carries
over to the more general case.

Theorem 3. Let {3 be a member of a Hilbert space H of random variables

and let PI denote its orthogonal projection on a closed subspace Y 1 of H.
(Thus, PI is the best estimate of {3 in Y I ) Let yz be an m-vector of random
variables generating a subspace Y 2 of H, and let yz denote the m-dimensional
vector of the projections of the components of yz onto Y I . (Thus, Y2 is the
vector of best estimates ofYz in Y I ) Let Y2 = Yz - Y2
Then the projection of {3 onto the subspace Y l + Y 2 , denoted P, is

P= PI + E({3Y2')[E(Y2Y2')]-ly2
In other words, Pis

PI plus the best estimate of{3 in the subspace Y2 generated

by Y2 .
Proof. It is clear that Y 1 + Y z = Y 1 El3 2 and that Yz is orthogonal
to Y 1 The result then follows immediately since the projection onto the
sum of subspaces is equal to the sum'of the individual projections if the
subspaces are orthogonal. I

An intuitive interpretation of this result is that if we have an estimate PI

based on measurements generating Y1 , then when receiving another set of

4.7

RECURSIVE ESTIMATION

93

measurements we should substract out from these measurements that part

that could be anticipated from the result of the first measurements. In
other words, the updating must be based on that part of the new data which
is orthogonal to the old data.

Example 1. Suppose that an optimal estimate iJ of a random n-vector [3

has been formed on .the basis of past measurements and that
E[([3 - iJ)([3 - iJ)'J

= R.

Given additional measurements of the form

y

= Wp + 6

where e is a random vector which has zero mean and which is uncorrelated
we seek the updated optimal
with both p and the past
estimate of [3.
The best estimate of y based on th'" past measurements is p = wiJ and
thus y = y - wiJ. Hence, by Theorem 3,

P= iJ + E([3y')[E(YY'>r 1y
which workf> out to be

/1 = iJ + RW'[WRW' + Qrl(y - wiJ).

The error covariance is
E[([3 - p)([3 -

4.7

/3)'] = R -

RW'[WRW'

+ Q]-IWR.

Recursive Estimation

In many applications we are led, quite naturally, to consider a sequence of

random variables that occur consecutively in time. For example, daily
temperature measurements, the Dow-Jones stock averages, or a sequence of
measurements of the position of a maneuvering aircraft can be regarded as
sequences of random variables. We define a discrete random process as
any sequence of random variables.
There are a number of important estimation problems associated with
random processes including: prediction, which is estimation of future
values of the process from past observations; filtering, which is estimation
of the present value of a random process from inexact measurements of
the process up to the present time; or, more generally, estimation of one
random process from observations of a different but related process. If we
require linear minimum-variance estimates, these1estimation problems are
only special cases of the theory developed earlier in this chapter.
It is customary Ito depart from our previous notation slightly by indexing
the sequence of random variables that compose a discrete process by the

94

LEAST-SQUARES ESTIMATION

notation x(k) rather than by subscripts. Unless explicitly stated otherwise,

all random variables in this section are assumed to be real and to have
zero means.
The starting point for the recursive approach to estimation is a representation of random processes that explicitly characterizes the manner in which
the process is generated in time. We begin by describing this representation.
Definition. A random process {x(k)} is said to be orthogonal or white if

E[x(j)x(k)]

= rt j 0jk'

The process is orthonormal if, in addition, exj

= 1.

We assume that underlying every observed random process is an orthogonal process in the sense that the variables of the observed process are
linear combinations of past values of the orthogonal process. In other
words, the given observed process results from the operation on an orthogonal process by a linear-processing device acting in real time.
Example 1. (Moving Average Process) Let {u(k)}f" _ ex:> be an orthornormal random process, and suppose that {x(k)}k'= -ex:> is generated by the
formula
ex:>

x(k):= Laju(k-j)
j=1

where the constants aj satisfy LJ= I la j/2 < 00. The process can be regarded
as a moving average of past values of the underlying orthonormal process.
Example 2. (Autoregressive Scheme of Order 1) Let {u(k)}k'o:.:.ex:> be an
orthonormal process and suppose that {x(k)} is generated by the formula
x(k) = ax(k - 1)

+ u(k - 1),

This process, defined by a

the moving average process

lal < 1.

difference equation, is equivalent to

<Xl

x(k)

=L

aj-Iu(k - j).

j=1

Example 3. (Finite-Difference Scheme or Autoregressive Scheme of

Order n) As a generalization of the previous example, we imagine that
{x(k)} is generated from the orthonot*mal process {u(k)} by the finitedifference formula
x(k)

+ a1x(k -

1)

+ ... + anx(k -

n) = b1 u(k - 1)

+ ... + bnu(k -

n).

In order that the formula represents a stable system (so that E[x 2 (k)]
remains finite when the formula is assumed to have been operating over the

4.7

RECURSIVE ESTIMATION

95

infinite past), we require that the characteristic pOlynomial equation

s" + a1 s"-1

+ ... + an = 0

has its roots within the unit circle in the complex plane. Alternatively,
we may assume that n initial random variables, say x(O), x( -1), .... ,
x( -n + I), have been specified and regard the difference formula as
generating [x(k)] for positive k only. In that case, stability is immaterial,
although in general E[x 2 (k)] may grow without bound as k -L!X). By
hypothesizing time..varying coefficients in a finite-difference scheme such
as this, we can generate a large class of random processes.
There is a great deal of physical motivation behind these models for
random processes. It is believed that basic randomness at the microscopic scale including electron emissions, molecular gas velocities, and
elementary particle fission are basically uncorrelated processes. When
their effects are observed at the macroscopic scale with, for example, a voltmeter, we obtain some average of the past microscopic effects.
The recursive approach to estimation of random processes is based on
a model for the process similar to that in Example 3. However, for convenience and generality, we choose to represent the random process as
being generated by a first-order vector difference equation rather than an
n-th order scalar difference equation. This model accommodates a larger
number of practic:al situations than the scalar model and simplifies the
notation of the analysis.
Definition. An n-dimensional dynamic model of a random process consists
of the following three parts:

1. A vector difference equation

+ 1) = <ll(k)x(k) + u(k),

k = 0,1,2, ... ,
where x(k) is an n-dimensional state vector, each component of
which is a random variable, <ll(k) is a known n x n matrix, and u(k)
is an n-dime:nsional random vector input of mean zero satisfying
x(k

E[u(k)u'(l)] =

2. An initial random vector x(O) together with an initial estimate

having covariance
-x(O))'] = P(O).
3. Measurements of the process, assumed to be ofthe form
vi(k) = M(k)x(k)

+ w(k),

= 0, 1,2, ... ,

where M(k) is an m x n matrix and w(k) is at} m-dimensional random

measurement error having me:l.n zero and satisfying
E[w(k)w'(j)] = R(k)DkJ

where R(k) is; positive definite.

96

LEAST-SQUARES ESTIMATION

In addition, it is assumed that the random vectors x(O), u(j) and w(k) are

all uncorrelated for j

;?

0, k

O.

This model covers most of the examples discussed previously as well as

several other important situations. In many applications, the matrices
<l>(k), Q(k), M(k), and R(k) are independent of k.
The estimation problem is that of obtaining the linear minimum-variance
estimate of the state vector x from the measurements v. We introduce the
notation! x(k I j) for the optimal estimate of x(k), given the observations v
up to instant j. Thus x(k I j) is the appropriate projection of x(k) onto the
space V(j) generated by the random 111-vectors v(O), v(l), ... , v(j).
We are concerned exclusively with the case k
j-the case corresponding to either prediction of some future value of the state or estimation of
the present state. Estimation of past values oCthe state vector is referred
to as the smoothing problem; although in principle it is not substantially
different than prediction, the detailed calculations are a good deal more
messy. Solution to the estimation problem is quite straightforward, requiring primarily the successive application of the updating procedure of
Section 4.6.

Theorem 1. (Solution to Recursive Estimation Problem-Kalman) The

optimal estimate x(k + 11 k) of the random state vector may be generated

recursively according to the equation

(1)

x(k

+ 11 k) = <l>(k)P(k)M'(k)[M(k)P(k)M'(k) + R(k)r I
X [v(k) - M(k)x(k I k - 1)] + <l>(k)x(k I k -

1)

where the n x n matrix P(k) is the covariance of x(k' k - 1) which itself is

generated recursively by
(2)

P(k

+ 1) =

<l>(k)P(k){l- M'(k)[M(k)P(k)M'(k)

+ R(k)r 1 M(k)P(k)}<l>'(k) + Q(k).

The initial conditions for these recurrence equations are the initial estimate
x(O I - 1) = X(O) and its associated covariance P(O).
Proof Suppose that v(O), v(l), ... , vtk - 1) have been measured and
that the estimate x(k I k - 1) together with the covariance matrix P(k) =
E[(x(k I k - 1) - x(k(x(k' k - 1) - x(k)'J have been computed. In other
words, we have the projection of x(k) opto the subspace V(k - 1). At timek,
we obtain the measurements
v(k)

= M(k)x(k) + w(k)

which gives us additional information about the

I

random vector x(k). This

In this section this notation should not be confused with that of an inner product.

4.8

PROBLEMS

97

is exactly the situation considered in Example I, Section 4.6. The updated

estimate of x(k) is
x(k I k) == x(k I k - 1)

(3)

+ P(k)M'(k)[M(k)P(k)M'(k) + R(k)r 1
x [v(k) - M(k)x(k I k - 1)]

with associated error covariance

(4)

P(k I k) = P(k) -. P(k)M'(k)[M(k)P(k)M'(k)

+ R(k)r IM(k)P(k).

Based on this optimal estimate of x(k), we may now compute the optimal
estimate x(k + 11 k) of x(k + 1) = <1>(k)x(k) + u(k), given the observation
v(k). We do this by noting that by Theorem 1, Section 4.6, the optimal
estimate of <1>(k)x(k) is <1>(k)x(k I k), and since u(k) is orthogonal (uncorrelated) to v(k) and x(k), the optimal estimate of x(k + 1) is
x(k

(5)

+ II k) =

<1>(k)x(k I k).

The error covariance of this estimate is

(6)

P(k

+ 1) =

<1>(k)P(k I k)<1>'(k)

+ Q(k).

Substitution of equation (3) into (5) and of (4) into (6) leads directly to (1)
and (2). I
Equations (1) and (2) may at first sight seem to be quite complicated,
but it should be easily recognized that they merely represent the standard
minimum-variance formulae together with a slight modification due to the
updating process. Furthermore, although these equations do not allow for
simple hand computations or analytic expressions, their recursive structure
is ideally suited to machine calculation. The matrices P(k) can be precomputed from equation (2); then, when filtering, only a few calculations
must be made as each new measurement is received.
Theorem 1 treats only estimates of the special form x(k + 1 I k) rather
than xU I k) for arbitrary j. Solutions to the more general problem, however,
are based on the estimate x(k + 11 k). See Problems 15 and 16.

4.8 Problems
1. A single scalar f3 is measured by m different people. Assuming uncorrelated measurement errors of equal variance, show that the
minimum-variancle unbiased estimate of f3 is equal to the average of
the m measurement values.
I
2. Three observers, [ocated on a single straight line, measure the angle
of their line-of-sight to a certain object (see Figure 4.4). It is desired
to estimate the true position of the object from these meaSllrements.

98

LEAST-SQUARES ESTIMATION

Figure 4.4 A triangulation problem

Assuming that the angles measured, () i, i = I, 2, 3, are sufficiently
accurate so that "small angle approximations" apply to the deviations, show that the estimated position defined by 0/, i = 1, 2, 3,
which minimizes
I (}i - 0;)2 is given by

Li;

+ k(R 2 O2 = (}2 + k(R 3 = 0 3 + k(R , A

0, =

(}1

R 3 )S,

RI)S/
R 2 )S/'

where Si = secant ()/ and k is chosen so that 01, O2 , &3' define a single
point of intersection.
3. A certain type of mass-spectrometer produces an output graph similar
to that shown in Figure 4.5. Ideally, this curve is made up of a linear
combination of several identical but equally displaced pulses. In other
words,

s(t) =

n-I

L fJia(t -

i)

i=O

where the fJ;'s are unknown constants and a(t) is a known function.
Assuming that
a(t)a(t - i) dt =i. show that the least-squares
estimate of the fJ;'s, given an arbitrary measurement curve s(t), is
.
1
<Xl
Pi = -_12 b;<t)s(t) dt
P -<Xl
where

bi(t) =

a(t - i) - pa(t - i - 1)
(

(1 -I- p2)a(t - i) - pa(t - i - 1) - pa(t - i

a(t - i) - pa(t - i + 1)

i= 0

+ 1)

O<i<n-l
i=n-l

4.8

PROBLEMS

99

S1.t)

>' t

Figure 4.5 Mass-spectrometer data

4. Let P be an n-dimelilsional
vector of zero mean and positivedefinite covariance matrix Q. Suppose measurements of the form
y = Wp are made where the rank of W is m. If is the linear minimum
variance estimate of P based on y, show that the covariance of the
error P- has rank n - m.
5. Assume that the measurement vector y is obtained from Pby

y == Wp

+e

where

E(f313') = R,

E(ee') = Q,

E(f3e')

= S.

Show that the minimum-variance estimate of Pbased on y is

P= (RW' + S)(WRW' + WS + S'W' + Q)-ly.

6. Let p, y be random vectors with E(P) = Po, E(y) = Yo, and finitecovariance matrices. Show that the minimum-variance estimate of P
ofform
P=Ky+ h,
where b is a constant vector, is

7.

P= Po + E[(P - Po)(Y - Yo)']{E[(y - yo)(y - Yo)']} -ley - Yo).

Let P= Ky be the minimum-variance linear estimate of a random

vector Pbased on the random vector y. Show that

E[(P - p)(P - P)'] = EEPP'] - EEPP'J.

8. In this problem we develop the rudiments of the theory of linear

regression. Suppose that associated with an experiment there are two
random variables y and x. If the outcomes of several measurements of
y and x are plotted on a two-dimensional graph, the result may look
somewhat like that shown in Figure 4.6. These results could be
effectively summarized by saying that y is approximately a linear function of x. So y would be described by the equation y = a + bx which

100

LEAST-SQUARES ESTIMATION

is represented by the dashed line in the figure. A natural way to choose

the appropriate dashed line is to choose that line which minimizes
the total sum of the squared errors Lf= 1 ei2 where ei = Yi - (a + bXj)
is the vertical distance between an observation point on the graph and
the dashed line.

. .,.... -...,..-

. .....,;- ..,..-

...-'""

Figure 4.6 Regression

(a) Show that the best linear approximation is given by

y

ji

+ b(x- x)

where
N

1 N

ji= -N LYj,
i= 1

x=LX,
N j=1

LYjX j - Njix

= ;:...I=-.r:'--_ __
L,X,2 -Nx 2
i= I

(b) Show that b may be alternatively expressed as

N

r(Yi - ji)(x; - x)

b=

;:...1=.....:!c.......,.;N-----

L(x; - X)2
i= 1

9. With the same terminology as in Problem 8, suppose now that the

random variable y can be represented as

y =, IX

+ px + e

where e is a random variable with zero mean, variance (T2, and independent of x. We are hypothesizing the existence of a linear relation
between x and y together with an additive error vallable. The values
of a and b found in Problem 8 are estimates of the parameters IX and p.

Suppose the outcomes of N observations produce

(The x/s are thus fixed in the discussion that follows.)

Xl. X2, ... ,

xN

4.8

(a) Show that

PROBLEMS

10 1

L (61.- e)(xi -

x)

b = P + I._=....:I,N..-----

L (XI -

so that E(b) = p.
(b) Show that E(a)
(c) Show that

X)2

1= 1

= ct..
(12

Var (b) =

(Xi i= 1

X)2

(d) Whatis Var (b)?

10. Suppose a large experiment producing data of the form
Y = Wp

+6

is planned with y an m-dimensional data vector, p an n-dimensional

vector of unknown parameters, and 6 a random error vector with
E(6) = 0, E(66') = I. Before the experiment the matrix (W'W)-l,
which is of high order, is calculated. Unfortunately, during the experiment the last component of y is not obtained and the Gauss-Markov
estimate of p must be computed on the basis of the remaining components. Show that the inverse need not be recalculated but that P
can be determined using (W' W) -1 and the last row of W.
11. Show that, analogous to Theorem 1, Section 4.6, the Gauss-Markov
estimate of TP is 1'fJ where fJ is the Gauss-Markov estimate of p.
12. Show that, analogous to Theorem 2, Section 4.6, the Gauss-Markov
estimate P of P is the linear minimum-variance unbiased estimate
minimizing E [(P -- /3) PCP - /3)'] for any positive-semidefinite matrix P.
13. It is the purpose of this problem to show that Gauss-Markov estimates
can be updated in a manner analogous to that of minimum-variance
estimates. As a by.product, a formal connection between the two kinds
of estimation techniques is obtained.
Let Y1 and Y2 be two measurement vectors of an n-dimensional
vector of unknown parameters P; say,

= W1 P+ 61
Y2 = W 2 P+ 62
= e, E(6 1 61') = QI'
Yl

where E(8 1 ) = (), E(6 2 )

E(6 2 62 ') = Q2' and
E(6 162') = O. Assuming that the dimension m1 of Yl is at least nand
that WI' QI -1 WI is nonsingular, show that the Gauss-Markov
.estimated of Pbased on Y1 and Y2 can be obtained by first obtaining the

102

ESTIMATION

estimate, as well as the corresponding error covariance

of fJ based on Yb and then updating it with the
estimator based on Y2 .
14. A satellite is put into orbit about the earth. Its initial state vector (whose
components are position and velocity coordinates) is not known
precisely and it is desired to estimate it from later measurements of the
satellite's position. The equations of motion are:
x(k
where x(k) is the
It is assumed that

+ I) =

<l>x(k)

state vector and <I> is an n X n matrix.

E[x(O)] =

e,

E[x(O)x'(O)]

= P.

The measurements are of the form

v(k) == Mx(k)

+ e(k)

where Mis an m x n matrix (m < n) and E(e(k = f), E[e(k)e'(j)] =

QfJ jk' Develop the recursive equations for the minimum- variance
linear estimate of x(O) given the observations.
15. Given the general dynamical model of a random process as in Section
\ j) = <I>(k - l)<I>(k - 2) ... <I>(j +
+ 1\.i) for
4.7, show that
k > j.
16. Given the general dynamical model of a random process as in Section
\.i), .i > O.
4.7, show how to calculate recursively
REFERENCES
4.1. For more general approaches to estimation, not restricted to least squares,
consult the text by Deutsch [40).
4.3. The method of least-squares estimation goes back to Legendre and Gauss,
the latter having made extensive use of the method for planetary orbit
calculations.
4.6. For a detailed study of minimum-variance unbiased estimates and a
general foundation of statistical estimation, see Freeman [56] Or Graybill [65].
For additional applications and extensions of the theory, see Chipman [29]
and Mann [103].
,
4.7. The theory of estimation of random processes was initiated by Wiener [154]
and independently by Kolmogoroff who used a Hilbert space approach.
This section closely follows Kalman [77] who developed the recursive
approach.
4.8. The mass-spectrometer problem (Problem 3) is treated in Luenberger [98].
For a solution of the smoothing problem (Problem 16), see Rauch, Tung, and
Striebel [122].

DUAL SP

5.1 Introduction
The modern theory of optimization in normed linear space is largely
centered about the interrelations between a space and its corresponding
dual-the space consisting of all continuous linear functionals on the
original space. In this chapter we consider the general construction of
dual spaces, give some e:xamples, and develop the most important theorem
in this book-the Hahn-Banach theorem.
In the remainder of the book we witness the interplay between a normed
space and its dual in a number of distinct situations. Dual space plays a
role analogous to the inner product in Hilbert space; by suitable interpretation we can develop results extending the projection theorem solution
of minimum norm problems to arbitrary normed linear spaces. Dual space
provides the setting for an optimization problem "dual" to a given
problem in the origina.l (primal) space in the sense that if one of these
problems isa minimization problem, the other is a maximization problem.
The two problems are equivalent in the sense that the optimal values of
objective functions are equal and solution of either problem leads to
solution of the other. Dual space is also essential for the development of
the concept of a gradient, which is basic for the variational analysis of
optimization problems. And finally, dual spaces provide the setting for
Lagrange multipliers, fundamental for I:' study of constrained optimization
problems.
.
Our approach in this chapter is largely geometric. To make precise
mathematical statements, however, it is necessary to translate these geometric concepts into concrete algebraic relations. In this chapter we follow
two paths to a final set of algebraic results by consitiering two different
geometrical viewpoints, corresponding to two versions of the HahnBanach theorem. The first viewpoint parallels the development of the
projection theorem, While the second is based on the idea of separating
convex sets with hyperplanes.

104

DUAL SPACES

LINEAR FUNCTIONALS

5.2

Basic Concepts

First we recall that a functional fOil a vector space X is linear if for any
two vectors x, Y E X and any two scalars a, fJ there holds f(ax + fJy) ==

af(x)

+ Pf(y)

Example 1. On the space En a linear functional can be expressed in the

form f(x) =
1 11k
for x =
... ,
where the 11k are fixed
scalars. Such a functional is easily seen to be linear. Furthermore, it can be
shown that all linear functionals on E" are of this form.
Example 2. On the space C[O, I] the functional f(x) = xG-) is a linear
functional.
Example 3. On the space Lz[O, I] the functionalf(x) = Sby(t)x(t) dt, for a
E L 2 [O, IJ, is a linear functional.

fixed y

We are primarily concerned with continuous linear functionals. The

reader may verify that the functionals in the above three examples are all
continuous.

If a linear functional on a normed space X is continuous at a

single point, it is continuous throughout X.

Proposition 1.

Proof Assume that f is linear and continuous at Xo E X. Let {xn} be

a sequence from X converging to x E X. Then, by the linearity off,
If(x,,) - f(x) I = If(x n - x

+ xo) - f(xo)l

However, since XII - x + Xo -+ Xo and sincefis continuous at xo, we have

f(x n - x + xo) -+ f(x o). Thus, If(x,,) - f(x) I -+ O. I

The above result is most often applied to the point and continuity
thus verified by verifying that f(x n) -+ for any sequence tending to
Intimately related to the notion of continuity is the notion of boundedness.

e.

Definition. A linear functional f on a normed space is bounded if there

is a constant M such that lI(x) I S M IIxll for all x EX. The smallest
such constant M is called the norm of.( and is denoted IIf11. Thus, Ilfll =
inf {M: If(x) I ::::;; Mllxll, all x EX}.
It is shown below that this definition satisfies the usual requirements of
a norm.

Proposition 2. A linear functional on a normed space is bounded if and only

if it is continuous.

5.2

BASIC CONCEPTS

105

Proof Suppose first that the linear functional/is bounded. Let M be

such that II(x) I ::;; Mllxll for all x E X. Then if Xn -4 e, we have I/(x n) I ::;;
Mllxnll _0. Thus, I is, continuous at e. From Proposition 1 it follows that
lis continuous everywhere.
Now assume that f is continuous at
Then there is aD> 0 such that
II(x) I < I for Ilxll ::;; D. Since for any nonzero x E X, Dxlllxli has norm
equal to D, we have

e.

II(x) I =
and M

<

= liD serves as a bound for f I

We offer now an example of an unbounded linear functional.

Example 4. On the space of finitely nonzero sequences with norm equal to

the maximum of the absolute values of the components, we define, for

x = {el'

e2' ... , en, 0, 0, " .},

f{x)

= k=t
L kek'

The functional lis clearly linear but unbounded.

The linear functionals on a vector space may themselves be regarded as
elements of a vector space by introducing definitions of addition and scalar
multiplication. Given two linear functionals/t '!2 on a space X, we define
their sum/l + 12 as the functional on X given by (/1 + 12)(X) = II (x) + lix)
for all x E X. Similarly, given a linear functional!, we define rJ./by
=
rJ.[f(x)]. The null ehement in the space ')f linear functionals is the functional that is identically zero on X. The space of linear functionals defined
in this way is called the algebraiC dual of X. Its definition is independent of
any topological structure on X such as might be induced by a norm on X.
Of greater importance for our purposes is the subspace of the algebraic
dual consisting of all bounded (i.e., continuous) linear functionals on a
normed space X. The space becomes a normed space by assigning the norm
according to the last definition.
The norm of a functional I can be expressed in several alternative ways.
We have

Ilfll

= inf {M: If(x)1 ::;;

M

= sup Il(x)1
x <1<8

IIxll

= sup If(x) I
IIxll :s 1

= sup If(x)l.
Ilxll =1

Mllxll, for

all x EX}

106

DUAL SPACES

The reader should verify these equivalences since they are used throughout
this chapter. The norm defined in this way satisfies the usual requirements
of a norm: 11/\\ > 0; 11/\\ = 0 if and only if 1= e; 11 alii = lalll/ll; 1111 +
1211 = sup I/l(x)+/2(x)1
sup I/I(x)1 + sup I/l(x) I = Il/tll+1I/211
II xli :s; I
IIxll:s; I
lIxli s 1
In view of the preceding discussion, the following definition is justified:
Definition. Let X be a normed linear vector space. The space of

an bounded

linear functionals on X is called the Ilorllled dllal of X and is denoted X *.

The norm of an element I EX'" is

11/11 =

sup

IIxll:s; 1

I/(x)l.

Given a normed space X, we usually refer to its normed dual X* simply

as the dual of X. As a general rule we denote functionals, linear or not, on a
normed space X by J, g, h, etc. However, when in the context of a particular
development, certain bounded linear functionals are regarded as elements
of the space X *; they are usually denoted by x1, x; , etc. The value of the
linear functional x* E X* at the point x E X is denoted by x*(x} or by the
more symmetric notation (x, x*) which is introduced in Section 5.6.
Theorem 1. X * is a Banach space.

Proof Since it has already been established that X* is a normed linear

space, it remains only to show that X* is complete. For this purpose, let
{x:} be a Cauchy sequence in X*. This means that Ilx: - x!1I
0 as
fl, m -+ 00. Now for any x e X, {x:(x)} is a Cauchy sequence of scalars since
Ix:(x) - x!(x)1
Ilx: - x!1I . \lxll. Hence, for each x, there is ascalar x*(x)
such that x:(x) -+ x*(x). The functional x* defined on all of X in this way
is certainly linear since X*(IXX + f3y) = lim x:(ax + f3y) = lim (IXX:(X) +
f3x:(y)] = a lim x:(x) + f3lim x:(y) = IXX*(X) + f3x*(y).
Now since {x:} is Cauchy, given a > 0, there is an M such that Ix:(x)x!(x) I < al\xll all n, m> M and all x; but since x:(x) -+ x*(x), we have
Ix*(x) - x!(x)1 < allxl1, m > M. Thus,
Ix*(x)1 = Ix*(x) - x!(x)

+ x!(x) I

Ix*(x) - x!(x) I + Ix!(x)1

(a + IIx!IDllxll
and x* is a bounded linear functional. Also from Ix*(x) - x!(x) I <
m > M, there follows IIx* - x!1I < a
that x! x* in X*. I

al\xll,

5.3 Duals of Some Common Banach Spaces

In this section we develop representations of the duals of En, lp, L p, co'
and Hilbert space. The dual 'of C[a, b] is discussed in Section 5.5. These

5.3

DUALS OF SOME COMMON BANACH SPACES

107

concrete representations of the duals of various Banach spaces enable us

to apply the abstrac:t theory of functional analysis to specific practical
problems.

The Dual of Eft. In the space En, each vector is an n-tuple of real scalars
x=
2' ... ,
with norm Ilxll = (Li _,
12. Any functional fOf the
formf(x) = L,= I til with each tljareal numberisc1earlylinear. Also,from
the Cauchy-Schwarz inequality for finite sequences.
If(x)1 =

\lt tl1el \ Ctlt112) 1/2(Jl e?) 1/2 = Cttl/2) '/2 1Ixll ,

we see that f is bounded with Ilfll

tlI2)'/2. However, since for
x = (til' tl2' ... , tin) equality is achieved in the Cauchy-Schwarz inequality,
we must in fact have IIfII =
Now letfbe any bounded linear functional on En. Define the basis vectors
ei in X by el = (0, 0, ... ,0, 1,0, ... ,0) with the i-th component 1 and all
others 0. Let til = x*(e i). A vector x = (ell e2, ... , en) may be expressed in
terms of the basis as x =
eie/. Sincefis linear, we have
Thus the dual space X* of X = En is itself En in the sense that the space
X* consists of all functionals of the formf(x) = Ltliei and the norm on
X* is Ilfll =
11'11 2 )1 / 2.

The Dual of lp, 1 p < 00. The Ip spaces were discussed in Section 2.10.
For every p, 1 p < 00, we define the conjugate index q = p/(p - 1), so
that l/p + l/q = 1; if p = 1, we take q = 00. We now show that the dual
space of Ip is Iq

Theorem 1. Every bounded linear functional on lp, 1

p<

00,

is represent.

able uniquely in the form

00

(1)

f(x)

= 1=L1'1i el

where y == {'1f} is an element of lq. Furthermore, every element of lq defines a

member of (1,,)* in this way, and we have
(2)

f l'1il
Ilfll = Ilyllq = {(

) I/q

1=1

sup l11kl
k

if 1 <

<

00

if p = 1.

Proof Suppose f is a bounded linear functional on 1". Define the

i = 1,2, ... , as the sequence that is identically zero except
element el E

I".

108

DUAL SPACES

for a 1 in the i-th component. Define lJi =!(el)' For any x ={el} E lp, we
have, by the continuity of/,f(x) = Ir'=l
Suppose first that I < P < 00. For a given positive integer N define the
vector XN E Ip having components
i

5, N

i >N.

Then

and

But I/(XN) I 5,
follows that

11/llllxNII; therefore, from the above two expressions, it

(tlllJd q )

ljq

5, 11111

for all N.

Hence the sequence y = {lJi} is an element of lq, and Ilyllq 5, II/II.

Suppose now that y = {IJ I} is an element of lq. If x = {, I} E 1p' then
f(x) = 2:r;, 1 'IIJ I is a bounded linear functional on Ip since, by the Holder
inequality, I/(x) 1:::;; 2:(;II'ilJd 5, Il xl1 p llyllq, and thus 11/115, Ilyllq' Since
/(el) = IJI in this case, it follows from the previous analysis that lIyllq :::;; II/II
Therefore, 111\1 = lIyllq'
For p = 1, q = 00, define XN by
i :i= N

i=N.
Then

IlxN11 5, I and

IIJNI =/(xN) 5, 11/lllIxNII 5, II/II.

Thus the sequence y = {IJJ is bounded by Ilfll. Hence, Ilyll", :::;; Ilfll.
Conversely, if y = {l1J E la;, the relation (1) obviously defines an element
of (/1)* with II/II :::;; IIYlloo Sinte again /(el) = lJi' it follows from above
that Ilyll", :::;; 11/11 and, hence, 11/11 = Ilyll",. I

The dual of I", is not II'

The Dual of Lp(O, 1), 1!;p <. 00. The Lp spaces were discussed in Section 2.10 and are the function space analogs of the lp spaces. Arguments
similar to those given in Theorem 1 show that for 1 ::; P < 00, the dual

5.3

DUALS OF SOME COMMON BANACH SPACES

109

space of Lp is Lq; (lIp + llq = 1) in the sense that there is a one-to-one

correspondence between bounded linear functionals f and elements y e Lq
such that
I

f(x)
and Ilfll

= fax(t)y(t) dt

= lIyllq'

The Dual of
sequences x =
is IIxll = max
i

The space Co is defined as the space of all infinite

of real numbers converging to zero. The norm on Co
Thus, Co is a subspace of 100 .

Co.

We leave it to the reader to verify that the dual of Co is II in the usual

sense with bounded linear functionals represented as

and IIfll = lIyll.

The Dual of Hilhert Space. On a Hilbert space the functionalf(x) = (x Iy)

for a fixed y is a linear functional in the variable x. The Cauchy-Schwarz
inequality I(x Iy)1 Ilxlillyll shows that the functional f is bounded with
IIfll Ilyll; the relation f(y) = (y I y) shows that in fact Ilfll = lIyll. Ob
viously, distinct vectors y produce distinct functionals f Thus, in Hilbert
space, bounded linear functionals are generated by elements of the space
itself. Here we show that all bounded linear functionals on Hilbert space
are of this form. The examples for the Hilbert spaces En, 12 , and L1considered above illustrate this general result.
Theorem 2. (Riesz-Frechet) Iff is a bounded linear functional on a Hilbert

space H, there exists a unique vector y e H such that for all x e H, f(x) =
(x I y). Furthermore, we have IIfll = lIyll and every y determines a unique
bounded linear functional in this way.

Proof Given a bounded linear functional f, let N be the set of all

vectors n e H fClr which f(n) = O. The set N is obviously a subspace of
H. It is closed since if ni -+ x is a sequence in H with ni eN, we have 0 =
f(ni) -+ f(x) by the continuity off.
If N = H, then/= 0 and the theorem is proved by taking y = e.
If N:F H, we may write, according to T.heorem 1, Section 3.4,
H = N $ N.J., and sinceN # H, there is a nonzero vector Z E N.J.. Since z is
nonzero and z N, necessarily f(z) .,; O. Since N.J. is a subspace, we may
assume that z has been appropriately scaled so that /(z) = 1. It will be
shown that the vector z is a scalar multiple of tht: desired vector y.

110

DUAL SPACES

Given any x e n, we have x - f(x)z eN since f[x - f(x)z] = f(x) f(x)f(z) = O. Sincez ..L N, we have (x - f(x)z Iz) = 0 or (x Iz) = f(x)llzI1 2 or
f(x) = (x Iz/lIzI12). Thus, defining y = z/llzIl2, we havef(x) = (x Iy).
The vector y is clearly unique since if y' is any vector for whichf(x)=
(x I y') for all x we have (x I y) = f(x) = (x Iy'), or (x Iy - y') = 0 for all x
which according to Lemma 2, Section 3.2 implies y' ::::: y.
It was shown in the discussion preceding the theorem that IIfll = lIyll I
EXTENSION FORM OF THE HAHN-BANACH THEOREM

5.4 Extension of Linear Functionals

The Hahn-Banach theorem, the most important theorem for the study of
optimization in linear spaces, can, like so many important mathematical
results, be stated in several equivalent ways each having its own particular
conceptual advantage. The two classical versions of the theorem, referred
to as the" extension form" and the" geometric form," playa fundamental
role in the theory of this book. The extension form proved in this section
serves as an appropriate generalization of the projection theorem from
Hilbert space to normed space and thus provides a means for generalizing
many of our earlier results on minimum norm problems.
Definition. Letfbe a linear functional defined on a subspace M of a vector
space X. A linear functional F is said to be an extension off if F is defined
on a subspace N which properly contains M, and if, on M, F is identical
with! In this case, we say that F is an extension of ffrom M to N.

In simple terms, the Hahn-Banach theorem states that a bounded linear

functional f defined only on a subspace M of a normed space can be
extended to a bounded linear functional F defined on the entire space and
with norm equal to the norm off on M; i.e.,
IIFII =

IlfllM =

If(m)1

lmf'

Actually, we are able to prove a somewhat more general version of the

theorem, replacing norms with sublinear functionals. This generalization
is used later to prove the geometric form of the
theorem.
Definition. A real-valued function p defined on a real vector space X is said
to be a sub linear functional on X if

1. P(Xl 1" X2) S P(Xl) -t P(X2),

2. p(Cl.x) = Cl.p(x),

for aU XI' X2 X.
for all CI. ;;:: 0 and x e X.

Obviously, any norm is a sublinear functional.

S.4

EXTENSION OF LINEAR FUNCTIONALS

III

Theorem 1. (Hahn-Banach Theorem, Extension Form) Let X be a real

linear normed space' and p a continuous sublinear functional on X. Let f be
a linear functional defined on a subspace M of X satisfying f(m) p(m)
for all m e M. Then there is an extension F off from M to X such that
F(x) p(x) on X.
Proof. The theorem is true in an arbitrary normed linear space, but
our proof assumes that X is separable. The general result, however, is

obtained by exactly the same method together with a simple application

of Zorn's lemma. The reader familiar with Zorn's lemma should have little
difficulty generalizing the proof. The basic idea is to extend f one dimension
at a time and then apply induction.
Suppose y is a vector in X not in M. Consider all elements of the subspace
[M + y]. Such an element x has a unique representation of the form
x = m + ocy, where m e M and oc is a real scalar. An extension 9 off from
M to [M + y] has the form
g(x) = f(m)

+ ocg(y)

and, hence, the extension is specified by prescribing the constant g(y). We

must show that this constant can be chosen so thatg(x) p(x) on [M + y].
For any two elements m 1, m2 in M, we have

or

and hence
inf [p(m

sup [f(m) - p(m - y)]

meM

+ y) -

f(m)].

""oM

Therefore, there is a constant c such that

sup [f(m) - p(m - y)]
meM

inf [p(m
meM

+ y) -

f(m)].

For the vector x = m + ocy e [M + y], we define g(x) = f(m)

must show that gem + ocy) p(m. + ocy).
If oc > 0, then

+ occ.

We

112

DUAL SPACES

If ex =

-p < 0,

then

- pc + f(m) = p[ - c + f(7i) ]

p[p(; - Y) - f(7i) + f(7i)]

Y) = p(m -

py).

Thus g(m + exy) p(m + exy) for all ex and 9 is an extension off from M
to [M+yJ.
Now let {XI' Xl, , X n , } be a countable dense set in X. From this
set of vectors select, one at a time, a subset of vectors {YI, Y2 , ... , Yn, ... }
which is independent and independent of the subspace M. The set
{YI' Y2' ... , Yn, ... } together with the subspace M generates a subspace S
dense in X.
The functionalfcan be extended to a functional 9 on the subspace S by
extendingffrom M to [M + YI], then to [[M + YI] + Y2]; and so on.
Finally, the resulting 9 (which is continuous since p is) can be extended
by continuity from the dense subspace S to the space X; suppose X E X,
then there exists a sequence {sn} of vectors in S converging to x. Define
F(x) = lim g(sn)' F is obviously linear and F(x) +- g(sn) p(sn) p(x) so
n.... co

F(x)

p(x) on X.

The version of the Hahn-Banach extension theorem given above is by

no means the most general available. It should be noted in particular that
since f and its extension F are dominated by the continuous sublinear .
functional p, both rand F are continuous linear functionals. A more general
version of the theorem requires X only to be a linear vector space and,
hence, neither continuity of the functionals nor separability of the space
plays a role. Neither of these restrictions is of practical importance to us;
in all applications considered in this book, the linear functions are bounded
and the Hahn-Banach theorem is applied only to separable normed spaces,
although the dual spaces may be nonseparable.
Corollary 1. Let f be a bounded linear functional defined on a subspace M

of a real normed vector space X. Then there is a bounded linear functional

F defined on X which is an extension off and which has norm equal to the
norm of f on M.
Proof Take p(x)

= IIfll11 IIxll

in the Hahn-Banach Theorem.

The following corollary establishes the existence of nontrivial bounded

linear functionals on an arbitrary normed space.
Corollary 2. Let X be an element of a normed space X. Then there is a
nonzero bounded linear functional F on X such that F(x) = IIF 1lllxli.

5.5

THE DUAL OF C[a,

b]

113

Proof Assume X:F (J. On the one-dimensional subspace generated by

x, define f(lXx) = lXllxll. Then f is a bounded linear functional with norm
unity which, by Corollary 1, can be extended to a bounded linear functional
F on X with norm unity. This functional satisfies the requirements.
If x = (J, any bounded linear functional (existence of one is now
established) will do. I
The converse of Corollary 2 is not generally true even in Banach space,
as the following example illustrates.
Example 1. Let X = II' X* = 100 , For x =

f(x)

gl'

.. } E

X, we define

= i=f 1 (1 - I

It is clear that fE X*, IIfll = 1. However, the reader may verify by

elementary analysis thatf(x) < IIxll for all nonzero x Ell '

The Hahn- Banach theorem, particularly Corollary 1, is perhaps most

profitably viewed as an existence theorem for a minimization problem.
Given anf on a subspace M of a normed space, it is not difficult to extendf
to the whole space. An arbitrar:\' extension, however, will in general be
unbounded Olr have norm greater than the norm off on M. We therefore
pose the problem of selecting the extension ot minimum norm. The HahnBanach theorem both guarantees the existence of a minimum norm extension and tells us the norm of the best extension.

5.5 The Dual of C[a, b]

The Hahn-Banach theorem is a useful tool for many problems in classical
analysis as well as for optimization problems. As an example of its use, we
characterize the dual space of C[a, b]. This result is of considerable interest
for applications since many problems are naturally formulated on C[a, b).

Theorem 1. (Riesz Representation Theorem) Let f be a bounded linear

functional on X = qa, b]. Then there is a function v of bounded variation
on [a, b] such that for all x E X

f x(t) dv(t)
b

f(x) =

and such that the norm off is the total variation of v on [a, b). Conversely,
every function of bounded variation on [a, b] defines a bounded linear
functional on X in this way.
Proof Let B be the space of bounded functions on [a, b] with the
norm of an element x e B defined as !lxliB = sup Ix(t)l. The space C[a, b)
a:St:sb

l14

DUAL SPACES

can be considered as a subspace of B. Thus, if/is a bounded linear functional

on X = C[a, b], there is, by the Hahn-Banach theorem, a linear functional
F on B which is an extension of/and has the same norm.
For any S E [a, b], define the function Us by Ua = 0, and by
if a::.;; I::';; S
ifs<t::.;;b
for a < S ;S; b. Obviously, each Us E B.
We define v(s) = F(u.) and show that v is of bounded variation on [a, b].
For this purpose, let a = 10 < t I, 12 , . , < In = b be a finite partition of
[a, b]. Denoting 81 = sgn [V(tl} - v(tl-I)]' we may write
n

L \v(ti } -

1=1

V(t l _ I )\ ==

L 81[V(tI) -

v(t,_t)]

1= 1
n

=1=1
L 81[F(u r)

- F(u r,_,)]

Thus,

since

IIFII

IIJII

and

and hence v is of bounded variation with T.V.(v) ::.;; Ilfll.

Next we derive a representation for / on X. If x E X, let
n

z(t) =

L X(tI-I)[Ur,(t) i=

ur,_,(t)]

where {II} is again a finite partition of [a, b]. Then

liz - xllB = max

1

max IX(ti-I)- x(t)1

r'-<:Sl''''

which (by the uniform continuity of x) goes to zero as the partition is made
arbitrarily fine. Thus, since F is continuous, F(z) -+ F(x} = /(x). But
n

F(z) ==

L X(ti-I)[V(t/) -

1= 1

V(ti-I)]

5.6

THE SECOND DUAL SPACE

115

and, by the definition of the Stieltjes integral,

F(z) -+

Jx(t) dv(t).

f(x) =

Jx(t) dv(t).

Therefore,
b

It is a standard property of the Stieltjes integral that

\ (X(t)dV(t)\

Ilxll . T.V.(v)

and, hence, II/II :s;; T.V.(v). On the other hand, we have Ilfll T.V.(v) and,
consequently, IIfll = T.V.(v).
Conversely, if v is a function of bounded variation on [a, b], the functional

f(x) =

J x(t) dv(t)
b

is obviously linear. Furthermore,fis bounded since If(x)1

IlxIIT.V.(v).

It should be noted that Theorem 1 does not claim uniqueness of the

function of bounded variation v representing a given linear functional f
since, for example, the functional x(1/2) can be represented by a v which
is zero on [0, 1/2), unity on (1/2, 1], and has any value between zero and
unity at the point t = 1/2. To remove the ambiguity, we introduce the
following subspace of BV[a, b].
Definition. The normalized space of functions of bounded variation denoted
NBV[a, b] consists of all functions of bounded variation on [a, b] which
vanish at the point a and which are continuous from the right on (a, b).
The norm of an element v in this space is IIvll = T.V.(v).

With the above definition the association between the dual of C [a, b]
and NBV[a, b] is unique. However, this normalization is not necessary in
most applications, since when dealing with a specific functional, usually
any representation is: adequate.
5.6 The Second Dual Space
Let x* e X*. We often employ the notation (x, x*) for the value of the
functional x* at a point x e X. Now, given x e XI the equation f(x*) =
(x, x*) defines a functional on the space X*. The functionalfdefined on
X* in this way is linear

f(rxx!

+ px!} = (x,rxx! +

= rx(x, xf) + P(x, xi> = rxf(x!) + P/(x!).

116

DUAL SPACES

Furthermore, since I/(x*)1 = I(x, x*)1 s IIxll . IIx*lI, it follows that

11/11 :::;; IIxli. By Corollary 2 of the Hahn-Banach theorem, there is a nonzero
x* e X* such that (x, x*) = Ilxllllx*lI, so in fact 11/11 = !lxll. We see then
that, depending on whether x or x* is considered fixed in (x, x*), both
X and X* define bounded linear functionals on each other and this
motivates the symmetric notation (x, x*).
The space of all bounded linear functionals on X* is denoted X** and is
called the second dual of X. The mapping cp:X -+ X** defined by x**= cp(x)
where (x*, x**) = (x, x*) is called the natural mapping of X into X**.
In other words, cp maps members of X into the functionals they generate
on X* through the symmetric notation. This mapping is linear and, as
shown in the preceding paragraph, is norm preserving (i.e., !!cp(x)\I = \Ix!!).
Generally, however, the natural mapping of X into X** is not onto.
There may be elements of X** that cannot be represented by elements in
X. On the other hand, there are important cases in which the natural
mapping is onto.

Definition. A normed linear space X is said to be reflexive if the natural

mapping cp: X -4 X** is onto. In this case we write X = X**.
Example 1. The Ip and Lp spaces, 1 < p < 00, are reflexive since lp * = lq
where lip + l/q == 1 and, thus, lp ** = lq * = lp.
Example 2. II and LI are not reflexive.
Example 3. Any Hilbert space is reflexive.
Reflexive spaces enjoy a number of useful properties not found in
arbitrary normed spaces. For instance, the converse of Corollary 2 of the
Hahn-Banach theorem holds in a reflexive space; namely, given x* E X*
there is an x e X with (x, x*) = Ilxl! IIx*lI.
5.7 Alignment and Orthogonal Complements
In general, for any x e X and any x*e X* we have (x, x*):::;; Ilx*llllxll.
In Hilbert space we have equality in this relation if and only if the
functional x* is represented by a nonnegative multiple of x, i.e., ifand only if
(x, x*) = (x I a:x) for some a: O. Motivated by the Hilbert space situa.
tion, we introduce the following definition ...

Definition. A vector x* e X* is said to be aligned with a vector x

(x, x*) = IIx*lIlIxll.

X if

Alignment is a relation between vectors in two distinct vector spaces: a

normed space and its normed dual.

S.7

ALIGNMENT AND ORTHOGONAL COMi-LEMENTS

117

Example 1. Let X = Lp[a, b), I < p < 00, and X = Lla, b), l/p + l/q= 1.
The condition for two functions x E L p, y E Lq to be aligned follow directly
from the conditions for Ilquality in the Holder inequality, namely,

S:X(t)y(t) dt = U:IX(tW dt} lIP. U:\y(tW dt} llq

if and only if x(t) == K[sgn y(t)]ly(t)lqlp for some constant K.
Example 2. Let x eX = C[a, b) and let r be the set of points t E [a, b) at
which Ix(t)1 = Ilxll. In glmeral, r may be infinite or finite but it is always
nonempty. A bounded linear functional x*(x) = $: x(t) dv(t) is aligned with
x if and only if v varies only on r and v is nondecreasing at t if x(t) > 0
and nonincreasing if x(t) < O. (We leave the details to the reader.) Thus, if
r is finite, an aligned functional must consist of a finite number of step
discontinuities. See Figure 5.1

r-----,
,...
I
L ___ "''/

/ - - - v(t)

Figure 5.1 Aligned junctions

The notion of orthogol1ality can be introduced in normed spaces through

the dual space.
Definition. The vectors x e X and x*
(x, x*) = O.

X* are said to be orthogonal if

Since the dual of a Hilbert space X is itself X, in the sense described in

theorem, the definition of orthogonality
Section 5.3 by the
given above can be regarded as a generalization of the corresponding
Hilbert space definition.
Definition. Let S be a subset of a normed linear space X. The orthogonal
complement l of S, denoted SJ., consists of all elements x* E X* orthogonal

to every vector in S.
Given a subset U of the dual space X*, its orthogonJI complement uJ.
is in X**. A more useful concept in this case, however, is described below.
I

The term annihilator is often used in place of orthogonal complement.

118

DUAL SPACES

Definition. Given a subset U of the dual space X*, we define the orthogonal
complement of U in X as the set .L U c: X consisting of all elements in X
orthogonal to every vector in U.
The set .L Umay be thought of as the intersection of U.L with X, where X
is considered to be imbedded in X** by the natural mapping. Many of the
relations among orthogonal complements for Hilbert space generalize to
normed spaces. In particular we have the following fundamental duality
result.
Theorem 1. Let M be a closed subspace of a normed space X. Then
.L[M.L] = M.

Proof. It is clear that M c: .L[M.L]. To prove the converse, let x rf= M.

On the subspace [x + MJ generated by x and M, define the linear functionalf(exx + m) = ex for m EM. Then
,
f(x

11111

Ilx

+ m)
1
+ mil = inf Ilx + mil
m

and since M is closed, IIfll < 00. Thus by the Hahn-Banach theorem, we
can extendfto an x* E X*. Sincefvanishes on. M, we have x* E M.L. But
also (x, x*) = 1 and thus x rf= .L[M.L]. I
5.8 Minimum Norm Problems
In this section we consider the question of determining a vector in a subspace M of a normed space which best approximates a given vector x in
the sense of minimum norm. This section thus extends the results of Chapter 3 for minimum norm problems in Hilbert space.
We recall that if M is a closed subspace in Hilbert space, there is always a
unique solution to the minimum norm problem and the solution satisfies
an orthogonality condition. Furthermore, the projection theorem leads to
a linear equation for determining the unknown optimizing vector. Even
limited experience with nonquadratic optimization problems warns us that
the situation is likely to be more complex in arbitrary normed spaces. The
optimal vector, if it exists, may not be unique and the equations for the
Nevertheless, despite these
optimal vector will generally be
difficulties, we find that the theorems of Chapter 3 have remarkable analogs
here. As before, the key concept is that of orthogonality and our principal
result is an analog of the projection theorem.
As an example of the difficulties encountered in arbitrary norrned space,
we consider a simple two-dimensional minimum norm problem that does
. not have a unique solution.

S.S

119

MINIMUM NORM PROBLEMS

Example 1. Let X be the space of pairs of real numbers x =

with
Ilxll = max d. Let M be the subspace of X consisting of all those vectors
1=1,2

....

having their second component zero, and consider the fixed point x = (2, 1).
The minimum distance from x to M is obviously 1, but any vector in M of
the form m = (a, 0) where 1 S; as; 3 satisfies Ilx - mil = 1. The situation
is sketched in Figure 5.2.
x

..

"'i4.L...

vectors

Figure 5.2 Solution to Example 1

The following two theorems essentially constitute a complete resolution
of minimum norm problems of this kind. These theorems, when specialized
to Hilbert space, contain all the conclusions of the projection theorem
except the uniqueness of the solution. When uniqueness holds, however, it
is fairly easy to prove separately.
Furthermore, the following two theorems contain even more information
than the projection theorem. They introduce a duality principle stating the
equivalence of two extremization problems: one formulated in a normed
space and the other in its dual. Often theransition from one problem to its
dual results in significant simplification or enhances physical and mathematical insight. Some infinite-dimensional problems can be converted to
equivalent finite-dimensional problems by consideration of the dual problem.

Theorem 1. Let x be an element in a real normed linear space X and let d

denote its distance from the subspace M. Then,
d = inf

(1)

meM

Ilx - mil

max (x, x*)

IIx1I s

xeM'!'

where the maximum on the right is achieved for some xci E Ml..
If the infimum on the left is achieved for some mo E M, then xci is aligned
with x - mo.
Proof For 8 > 0, let m. E M satisfy IIx - m.1I S; d + 8. Then for any
MJ.., IIx*1I S; 1, we have
(x, x*) == (x - m., x*) S; Ilx*lllIx - m.11 S; d + 8.
Since 8 waS arbitrary, we conclude that (x, x*) ::;; d. Therefore, the proof
of the first part of the theorem is complete if we exhibit any xci for which
(x, xci> = d.

x*

120

DUAL SPACES

Let N be the subspace [x + M]. Elements of N are uniquely representable in the form n = ax + m, with m E M, a real. Define the linear functional f on N by the equation fen) = ad. We have

Ilfll =

If(n)1

laid

Tnf = sup Ilax + mil

laid

r r
d

-,up 1.lll" +: infll' +:

1.

Now form the Hahn-Banach extension, xci, of ffrom N to X. Then II xci 11 = 1

and xci =f on N. By. construction, we have xci E M.l and (x, xci> = d;
hence xci satisfies the requirements of the first part of the theorem.
Now assume that there is an mo EM with \Ix - moll = d and let xci be
any element such that xci E MJ., IIxci\l = 1, and <x, xci> = d-the xci constructed above being one possibility. Then

<x - mo, xci> = (x, xci> = d = IIx6'lIlIx - moll

and x6' is aligned with x - mo I
The reader should attempt to visualize the problem and the relation (1)
geometrically. The theorem becomes quite clear if we imagine that the
error x - mo is orthogonal to M.
.
Theorem 1 states the equivalence of two optimization problems: one in X
called the primal problem and the other in X* called the dual problem.
The problems are related through both the optimal values of their respective
objective functionals and an alignment condition on their solution vectors.
Since in many spaces alignment can be explicitly characterized, the solution
of either problem often leads directly to the solution of the other. Duality
relations such as this are therefore often of extreme practical as well as
theoretical significance in optimization problems.
lfwe take only a portion of the above theorem, we obtain a generalization
of the projection theorem.

Corollary 1. Let x be an element of a real normed linear vector space X

and let M be a subspace of X. A vector mo E ,M satisfies Ilx - moll ::; IIx - mil
for all m E M if and only if there is a nonzero vector x* E MJ. aligned with
x-mo
Proof The" only if" part follows directly from Theorem 1. To prove
the" if" part, assume that x - mo is aligned with x* E M.l. Without loss
of generality, take llx*ll = 1. For all mE M we have

<x, x*> = <x - m, x*) ::; IIx - mil

5.8

121

MINIMUM NORM PROBLEMS

whereas

(x, X*) = (x - mo, x*) = IIx - moll.

Thus, Ilx -

IIx - mil. I

moll

As a companion to Theorem 1, we have:

Theorem 2. Let M be a subspace in a real normed space X. Let x*
. be a distance d from Ml.. Then

(2)

= m*eML
min Ilx* - m*11

X*

= sup (x, x*)

XliiM

IIxll SI

where the minimum on the left is achievedfor mri E Ml..lfthe supremum on

the right is achieved for some Xo EM, ;'hen x* -m is aligned with Xo.
Proof We denote the right-hand side of relatior (2) by Ilx*IIM because
it is the norm of the functional x* restricted to the subspace M. For any
m* E Ml., we have
Ilx* - m*11 = sup [(x, x*) - (x, m*)J
IIxlI:s:!

sup [(x, x*) - (x,

XGM

m*)J

IIxll S!

sup (x, x*) = IIx*IIM'

XliiM

IIxlI:s; !

Thus,

Ilx* - m*11

Ilx*IIM and the first part of the theorem is proved ifan

Ml. can be found giving equality.

Consider x* restricted to the subspace M. The norm of x* so restricted
is IIx*IIM' Let y* be the Hahn-Banach extension to the whole space of the
restriction of x*. Thus, IIY*II = Ilx*IIM and x* - y* = 0 on M. Put mri =
x*- y*. Then
E Ml. and Ilx* = Ilx*IIM'
If the supremum on the right is achieved by some Xo EM, then obviously
llxoll = 1 and I\x* = (xo, x*) = (xo, x* Thus x* - mri is
aligned with Xo I
E

Theorem 2 guarantees the existence of a solution to the minimum norm

problem if the probl,em is appropriately formulated in the dual of a normed
space. This result simply reflects the fact that the Hahn-Banach theorem
establishes the existence of certain linear functionals rather than vectors
and establishes the general rule, which we adhere to in applications, that
minimum norm problems must be formulated in a dual space if one is to
guarantee the existence of a solution.
The duality of the above theorems can be displa.yed more explicitly in
terms of some fairly natural notation. Assuming M to be a closed subspace,
we write

IlxilM = inf Ilx - mIl

meM

122

DUAL SPACES

since this is the norm of the coset generated by x in the space

element x* e X*, we write, as in the proof of Theorem 2,

Ilx*IIM ==

XI M.

For an

sup <x, x*)

IIxll,; 1
xeM

since this is the norm of x* considered as a functional on M. In this notation,

equations (1) and (2) become
(3)
(4)

IlxilM = IlxllMl
IIx*IIMl = Ilx*IIM

where on the right side of(3) the vector x is regarded as a functional on X*.

5.9 Applications
In this section we present examples of problem solution by use of the theory
developed in the last section. Three basic guidelines are applied to our
analyses of the problems considered: (1) In characterizing optimum solutions, use the alignment properties of the space and its dual. In the Lp and
lp spaces, for instance, this amounts to the conditions for equality in the
Holder inequality. (2) Try to guarantee the existence of a solution by
formulating minimum norm problems in a dual space. (3) Look at the
dual problem to see if it is easier than the original problem. The dual may
have lower dimension or be more transparent.

Example 1. (Chebyshev Approximation) Let I be a continuous function

on an interval [a, b] of the real line. Suppose we seek the polynomial p of
degree n (or less) that best approximates I in the sense of minimizing
max I/(t) - p(t)I, i.e., minimizing the maximum deviation of the two
dSISb

functions. In the Banach space X = C[a, b], this problem is equivalent to

finding the P in the n + I-dimensional subspace N of n-th degree polynomials that minimizes III - pil. We may readily establish the existence of
an optimal p, say Po, since the subspace N is finite dimensional.
Suppose III - Poll = d> 0, and let r be the
of points t in [a, b] for
which I/(t) - Po(t) I = d. We show that r contains at least n + 2 points.
According to Theorem 1 of Section 5.8, the optimal solution Po must be
such that f - Po is aligned with an element in NJ. c X* = NBV[a, b].
Assume that r contained m < n + 2 points ti : a:::;; tl < t2 < ... < t m :::;; b.
If v E N B V[a, b] is aligned with I - Po, v varies only on these points (see

Example 1, Section 5.7), and hence must consist of jump discontinuities

at the t/s. Let tk be a point of jump discontinuity of v. Then the polynomial
q(t) = ilik (t - t i ) is in N but has S:q dv =1= 0 and consequently, v NJ..

S.9

APPLICATIONS

123

Therefore,f - Po is not aligned with any nonzero element of NJ. and hence
r must contain at least n + 2 points. We have therefore proved the classic
result of Tonelli: If J is continuous on [a, b] and Po h the polynomial oj
degree n (or less) minimizing max IJ(t) - p(t)l, then IAt) - Po(t) I
t e[a, b]

achieves its maximum at at least n + 2 points in [a, b].

Many problems amenable to the theory of the last section are most
naturally formulated a.s finding the vector of minimum norm in a linear
variety rather than as finding the best approximation on a subspace. A
standard problem of this kind arising in several contexts is to find an element
of minimum norm satisfying a finite number of linear constraints. To
guarantee existence of a solution, we consider the unknown
in a dual
space X* and express the constraints in the form

x\

(Yl' x*) = C1

(Y2' x*)

= c2

If x* is any vector satisfying the constraints, we have

d ==

min
(YI.X*)=CI

IIx*1I

min
m*eMJ.

IIx* - m*1I

where M denotes the space generated by the Yi'S. From Theorem 2,

. Section 5.8, this becomes

d = min

IIx* - m*1I

Any vector in M is of the form x

since M is finite dimensional,

d=

min

IIx*1I =

(YI. x*;. =CI

sup (x,

xeM
IIxll Sl

= l:1 =

x*).

alYlor, symbolically, Ya; thus,

max (Ya,x*) = max c'a,

11 Yall S

II Yall:S t

the last equality following from the fact that x* satisfies the constraints.
The quantity c'a denotes the usual inner product of the two n-vectors a
and c with components ai and ci . Furthermore, we have the alignment
properties of Theorem 2, Section 5.8, and thus the following corollary.

Corollary 1. Let Y 1 E X, i = I, 2, ... , n, and suppose the system oj linear

equalities (YI' x*) = ('I' i = 1,2, ... , n is consistent; i.e., the set
D = {x* e X *: (y I, x*) = c i' i = I, 2, ... , n}

124

DUAL SPACES

is nonempty. Then
min

x'eD

Ilx*11

max c'a.

IIYalls!

Furthermore, the optimal x* is aligned with the optimal Ya.

Example 2. (A Control Problem) Consider the problem of selecting the

field current u(t) on [0, 1J to drive a motor governed by
e(t)

+ O(t) =

u(t)

from the initial conditions e(O) = 0(0) = to e(1) == 1, O(l) == in such a

way as to minimize max lu(t)l. This example is similar to Example 1,
o st::>: 1
Section 3.10, but now our objective function reflects a concern with
possible damage due to excessive current rather than with total energy.
The problem can be thought of as being formulated in CEO, 1J, but
since this is not the dual of any normed space, we are not guaranteed that
a solution exists in CEO, 1]. Thus, instead we take X = Ll [0, 1], X* =
Loo[O, 1] and seek u E X* of minimum norm. The constraints are, as in
Example 1, Section 3.10,
1

So e(t- J)u(t) dt = 0

Jo {1 - e(l-l)}u(t) dt == 1.
1

From Corollary 1,
min

lIuli

max

a2'

!lalYI +a2Y211 S 1

The norm on the right is the norm in X = L 1[0, 1J. Thus, the two constants aI' a2 must satisfy
!

So I(a

1 -

a2)e(t-l)

+ a21 dt s; 1.

Maximization of a 2 subject to this constraint is.a straightforward task,

but we do not carry out the necessary computations. Instead we show
that the general n:ature of the optimal control is easily deduced from the
alignment requirement. Obviously, the function atYt(t) + a2yz(t), being
the sum of a constant and an exponential term, can change sign at most
once, and since the optimal u is aligned with this function, u must be

"bang-bang" (Le., it must have values M for some M) and changes

sign at most once. We leave it to the reader to verify this by characterizing
alignment between L 1[0, 1] and Loo[O, 1].

5.9

APPLICATIONS

125

Example 3. (Rocket Problem) Consider the problem of selecting the

thrust program u(t) for a vertically ascending rocket-propelled vehicle,
subject only to the f,orces of gravity and rocket thrust in order to reach a
given altitude with minimum fuel expenditure. Assuming fixed unit mass,
unit gravity, and zero initial conditions, the altitude x(t) is governed by a
differential equation of the form

j,(t) == u(t) - 1

x(o)

= x(o)

= 0.

This equation can be! integrated twice (once by parts) to give

T2

x(T)

= fo (T -

t)u(t) dt -

'2 .

Our problem is to attain a given altitude, say x(T) = 1, while minimizing

the fuel expense.
T

fo lu(t)1 dt.
The final time T is in general unspecified, but we approach the problem by
finding the minimum fuel expenditure for each fixed T and then minimizing
over T.
For a fixed T the optimization problem reduces to that of finding u
minimizing
T

fo lu(t)1 dt
while satisfying the single linear constraint
T

Io (T - t)u(t) dt

T2

= 1 + -.

At first sight we might regard this problem as one in Ll [0, T]. Since,
however, L1 [0, T] is not the dual of any normed space, we imbed our
problem in the space NBV[O, T] and associate control elements u with
the derivatives of ellements v in NBV[O, T]. Thus the problem becomes
that of finding the I) e NBV[O, I] minimizing

fo Idv(t)1 = T.V.(v)
T

I\vl\

subject to
T

fo

(T - t) dv(t)

T2

=1+- .
2

126

DUAL SPACES

According to Corollary 1,
min

Ilvll =

max
II(T-r)alls;'

[a(l + T2)],
2

which is only a one-dimensional problem. The norm on the right-hand

side is taken in CEO, T], the space to which NBV[O, T] is dual. In CEO, T]
we have
II(T - t)all = max I(T - t)al
os;rs;T

== T lal

since the maximum occurs at t = 0. Thus the optimal choice is a = liT

and
min

IIvll

T2)

= (1 + T T'

The optimal v must be aligned with (T - t)a and, hence, can vary only
at t = 0. Therefore, we conclude that v is a step function and u is an
impulse (or delta function) at t = 0. The best final time can be obtained by
differentiating the optimal fuel expenditure with respect to T. This leads
to the final result

J2
min IIvll = J2
T=

anp

v=

{ft

t=

0< t

J2.

Note that our early observation that the problem should be formulated
in NBV[O, T] rather than L, [0, T] turned out to be crucial since the
optimal u is an impulse.
*5.10 Weak Convergence

...

An interesting and important concept that arises naturally upon the introduction of the dual space is that of weak convergence. It is important for
certain problems in analysis and plays an indirect role in many optimization problems.

Definition_ A sequence {x nl in a normed linear vector space X is said to

converge weakly to x E X if for every X* E X* we have (xn' x*) -+ (x, x*).
In this case we write Xn -+ x weakly.

5.IO

WEAK CONVERGENCE

127

Our earlier notion of convergence, convergence in norm, is sometimes

referred to as strong convergence. We have the following result.
Proposition 1. If Xn -+ x strongly, then

Xn -+

x weakly.

Proof I(x n , x*) - (x, x*)I:::;; Ilx*llllxn - xll-+O.

There are, however,. sequences that converge weakly but not strongly.

Example 1. In X = 12 consider the elements Xn = {O, 0, ... ,0, 1,0, ... }

with the 1 in.the n-th place. For any y = {lh, '12' .. } E 12 = X*, we have
(xn Iy) = '1n -+ as n.-;> 00. Thus Xn -+ e weakly. However, X n ++ e strongly
since II xn 1\ = 1.

Starting with a normed space X, we form X* and define weak convergence on X in terms of X*. The same technique can be applied to X*
with weak
being defined in terms of X**. However, there is a
more important notion of convergence in X* defined in terms of X rather
than X**.

Definition. A sequence {x:} in X* is said to converge weak-star (or weak*)

to the element x* if for every x E X, (x, x:) -+ (x, x*). In this case we
write x! -+ x* weak*.
Thus in X* we have three separate notions of convergence: strong,
weak, and weak*; fUirthermore, strong implies weak, and weak implies
weak* convergence. In general, weak* convergence does not imply weak
convergence in X *.

Example 2. Let X = co, the space of infinite sequences convergent to

zero, with norm equal to the maximum absolute value of the terms. Then
X* = 'I' X** = ' 00 , (See Section 5.3.) In X* = 'I' let x: = {O, 0, 0, ... ,
0, 1,0,0,0, " .}, the term 1 being at the n-th place. Then x: -+ e weak*
but x: ++ e weakly since (x:, x**) ++ for x** = {I, 1, 1, ... }.

It was shown in Section 2.13 that a continuous functional on a compact

set achieves a maximum and a minimum. However, the usual definition of
compactness, i.e., compactness with respect to strong convergence, is so
severe that this property can be used only in very special circumstances
such as in finite-dimensional spaces. Weak compactness and weak* compactness are less severe requirements on a set; indeed, such compactness
criteria (especially weak*) provide alternative explanAtions for the existence of solutions to optimization problems.

Definition. A set K c X* is said to be weak* compact if every infinite

sequence from K contains a weak* convergent subsequence.

128

DUAL SPACES

Theorem 1. (Alaoglu) Let X be a real normed linear space. The closed unit

sphere in X* is weak* compact.

Proof Although the theorem is true in general, we only prove it for
the case where X is separable (although X* need not be separable), which
is adequate for all of the examples treated in this book.
1. Let {Xk} be a
Let {x:} be an infinite sequence in X* such that Ilx:11
sequence from X dense in X. The sequence {(Xl' x:>} of real numbers is
bounded and thus contains a convergent subsequence which we denote
{(XI' X:l>}' Likewise the sequence {(x 2 , x: 1>} contains a convergent subsequence {(X2' 2>}. Continuing in this fashion to extract subsequences
{(Xb X:k>}' we then form the diagonal sequence {x:n} in X *.
The sequence {x:n} converges on the dense subset {Xk} of X; i.e., the
sequence of real numbers {(Xk, x:,,>} converges to a real number for
each Xk' The proof of the theorem is complete if we show that {x:n} converges weak* to an element x* E X *.
Fix x E X and e > O. Then for any n, m, k,

x:

I(x, x:n> - (x, x!m>1

I(x, x:n> - (Xb x:n>1

+ I(Xb x:n> -

(Xb X!m> I

+ I(Xk, x!m> - (x, x!m>1

211xk - xii + I(xk' x:n> - (Xk' x!m>l
Now choose k so that Ilxk - xii < 8/3 and then N so that for n, m > N,
I(Xb x:n> - (Xb x!m>1 < 8/3. Then for n, m > N, we have I(x, x:.> (x, x!m> I < 8. Thus, {(x, x:n >} is a Cauchy sequence and consequently
converges to a real number (x, x*>.
The functional (x, x*> so defined is clearly linear and has Ilx*11 :s 1. I

Definition. A functional (possibly nonlinear) defined on a normed space

X is said to be weakly continuous at xoif given 8 > 0 there is a 0 > 0 and
a finite collection {xf,xi', ... , x:} from X* such that I/(x)-f(xo)1 <8
for all x such thatl(x-xo., x7>1< fJ for i =1,2, . .. ,n. Weak * continuity of
a functional defined on X* is defined analogously with the roles of X
and X* interchanged.
...
The reader can easily verify that the above definition assures that if/is
weakly continuous, Xn - x weakly implies thatf(xn } - f(x). We also leave
it to the reader to prove the following result.
Theorem 2. Let I be a weak* continuous real-valued lunctional on a weak*

compact subset Sol X*. Then/is bounded on S and achieves its maximum

on S.

A special application of Theorems 1 and 2 is to the problem of maximizing (x, x*> for a fixed x E X while x* ranges over the unit sphere in X*.

S.II

HYPERPLANES AND LINEAR FUNCTIONALS

129

Since the unit sphere in X* is weak* compact and <x, x*) is a weak*
continuous functional on X*, the maximum is achieved. This result is
equivalent to Corollary 2 of the Hahn-Banach theorem, Section 5.4.
GEOMETRIC FORM OF THE HAHN-BANACH THEOREM

S.11

Hyperplanes and Linear Functionals

In the remaining sections of this chapter we generalize the results for

from linear varieties to convex sets. The foundaminimum norm
tion of this development is again the Hahn-Banach theorem but in the
geometric rather than extension form.
There is a major conceptual difference between the approach taken in
the remainder of this chapter and that taken in the preceding sections.
Linear functionals, rather than being visualized as elements of a dual space,
are visualized as hyperplanes generated in the primal space. This difference
in viewpoint combines the relevant aspects of both the primal and the
dual into a single geometric image and thereby frees our intuition of the
burden of visualizing two distinct spaces.

Definition. A hyperplane H in a linear vector space X is a maximal proper

linear variety, that is, a linear variety H such that H X, and if V is any
linear variety containing H, then either V = X or V = H.
This definition of hyperplane is made without explicit reference to linear
functionals and thus stresses the geometric interpretation of a hyperplane.
Hyperplanes are intimately related to linear functionals, however, as the
following three propositions demonstrate.
Proposition 1. Let H be a hyperplane in a linear vector space X. Then there

is a linear Junctional! on X and a constant c such that H = {x :J(x) = c}.

Conversely, iJ J is a nonzero linear Junctional on X, the set {x :J(x) == c}
is a hyperplane in X.

Proof Let H be a hyperplane in X. Then H is the translation of a

subspace M in X, say H == Xo + M. If Xo M, then [M + xo] = X, and
for x = IXXo + m, with m EM we defineJCx) == IX. Then H = {x : J{x) = I}.
If Xo EM, we take XI M, X =[M +
H = M, and define for
x = IXXt + m,J(x) = IX. Then H = {x ;f{x) = OJ.
Conversely, letJbe a nonzero linear functional on X and let M = {x:
J(x) == O}. It is clear that M is a subspace. Let Xo e X with J(xo) = l.
Then for any x E X" J [x - J(x)xo] = 0 and, hence, x - f(x)xo EM.
Thus, X = [xo + M], and M is a maximal proper subspace. For any
real c, let Xl be any element for which f(xj) = c. Then {x: f(x) = c} =
{x :f(x - Xl) = O} = M + Xl which is a hyperplane. I

xa,

130

DUAL SPACES

Hyperplanes that contain the origin represent a somewhat special case,

but barring these it is possible to establish a unique correspondence
between hyperplanes and linear functionals.

Proposition 2. Let H be a hyperplane in a linear vector space X. If H does

not contain the origin, there is a unique linear functional f on X such that
H = {x :f(x) = I}.
Proof By appropriate scaling, Proposition I guarantees the existence
of at least one such functionalf Let 9 be any other such functional, so that
in particular H = {x :f(x) = I} = {x: g(x) = I}. It is then clear that
He {x :f(x) - g(x) = O}. Since the smallest subspace of X containing H
is X, it follows thatf = g. I
The above considerations are quite general in that they apply to hyperplanes in an arbitrary linear vector space. A hyperplane H in a normed
space X must be either closed or dense in X because, since H is a maximal
linear variety, either H = H or H = X. For our purposes we are primarily
interested in closed hyperplanes in a normed space X. These hyperplanes
correspond to the bounded linear functionals on X.

Proposition 3. Let f be a nonzero linear functional on a normed space X.

Then the hyperplane H = {x :f(x) = c} is closed for every c if and only if
f is continuous.
Proof, Suppose first thatfis continuous. Let {x.} be a sequence from
E X. Then c = f(x.)
f(x) and thus x E Hand H is
closed. Conversely, assume that M = {x :f(x) = O} is closed. Let X =
[xo + M] and suppose x. -+ x in X. Then x. = 0(. Xo + m., x = O(x o + m,
and letting d denote the distance of Xo from M (which is positive since M
is closed), we have 10(. - O(ld ::;; Ilx. - xII -+ 0 and hence 0(.-+0(. Alsof(x.) =
O(nf(xo) + f(m.) = O(J(xo) -+O(f(x o) =f(x). Thusfis continuous on X. I

H. convergent to x

Iffis a nonzero linear functional on a linear vector space X, we associate

with the hyperplane H = {x :f(x) = c} the four sets
{x :f(x) ::;; c},

{x :f(x) < c},

{x :f(x);::: c},

{x :f(x) > c}

called half-spaces determined by H. The first two of these are referred

to as negative half-spaces determined by f and the second two as positive half-spaces. If f is continuous, then the half-spaces {x :f(x) < c},
{x :f(x) > c} are open and {x :f(x)::;; c}, {x :f(x);::: c} are closed.

The results of this section establish a correspondence between hyper

planes and linear functionals, particularly between the closed hyperplanes
and members of the dual X*. The fact that the correspondence is unique

S.12

HYPERPLANES AND CONVEX SETS

131

for hyperplanes not containing the origin suggests that virtually all concepts in which X* plays a fundamental role can be visualized in terms of
closed hyperplanes or their corresponding half-spaces.
5.12 Hyperplanes and Convex Sets
In this section we prove the geometric form of the Hahn-Banach theorem
which in simplest form says that given a convex set K containing an
interior point, and given a point Xo not in 1<, there is a closed hyperplane
containing Xo but disjoint from 1<..
If K were the unit sphere, this result would follow immediately from our
earlier version of the Hahn-Banach theorem since it establishes the
existence of an
aligned with Xo' For every x in the interior of the unit
sphere, we then have (x, xti>
<
= (xo,
or
(x,
< (xo,
which implies. that the hyperplane {x: (x,
=
(xo,
is disjoint from the interior of the unit sphere. If we begin with
an arbitrary convex set K, on the other hand, we might try to redefine the
norm on X so that K, when translated so as to contain e, would be the
unit sphere with
to this norm. The Hahn-Banach theorem could
then be applied on thiH new normed space. This approach is in fact successful for some special convex sets. To handle the general case, however,
we must use the
Hahn-Banach theorem stated in terms of sublinear functionals instead of norms.

Definition. Let K be a cop.vex set in a normed linear vector space X and

suppose e is an interior point of K. Then the Minkowskifunctional p of K
is defined on X by

p(X)=inftr :;EK,r>o}.
We note that for K equal to the unit s1,here in X the Minkowski functional is IIxli. In the general case, p(x) defines a kind of distance from the
origin to x measured with respect to Kj it is the factor by which K must
be expanded so as to include x. See Figure 5.3.
Lemma 1. Let K be a convex set containing

the M inkowski functional p of K satisfies:

1.

2.
3.
4.
5.

00 > p(x)
0for all x E X,
p(rxx) = rxp(x) for rx > 0,

P(XI

+ X2)

p is continuous"
K = {x: p(x):-:; I}.

+ P(X2),
K=

{x : p(x) < I}.

e as an

interior pOint. Then

132

DUAL SPACES

.,..
/'

,/

,/
,/

I
(

I
\

"

.......

Figure 5.3

_- -The Minkowski Junctional oj a convex set

Proof

1.
2.

Since K contains a sphere about e, given x there is an r> 0 such

that xlr E K. Thus p(x) is finite for all x. Obviously, p(x) O.
For IX > 0,
p(IXX) = inf {r:

e K, r > O}

= inf{lXr': :, E K, r' > O}

r

IX

inf {r':

K, r' > O} = IXp(X).

Given Xl' X2 and 6 > 0, choose r l , r2 such thatp(xj) < rj < p(Xj) + 6,
i = I, 2. By No.2, p(x!r/) < 1 and so x;/ri E K. Let r = r 1 + r 2 .
By convexity of K, (rt/r)(xt/r 1) + (r2!r)(x2h) = (Xl + x2)!r E K.
Thus, P(X/+X2)!r::;;1. Or by No.2, P(Xt-!-x2)sr<p(xl)+
P(X2) + 26. Since e was arbitrary, pis subadditive.
4. Let e be the radius of a closed sphere centered at e and contained
in K. Then for any X E X, 6x!lIxll E K and thus p(sx!llx!l) s 1.
Hence, by No.2, p(x) ::;;(lfe)llxli. This shows that p is continuous at
e. However, from No.3, we have p(x) = p(x - y + y) ::;; p(x - y) +
p(y) and p(y) = p(y - X + x) s p(y - x) + p(x) or -p(y
p(x) - p(y) ::;; p(x - y) from which continuity on X follows from
continuity at fJ.
5. This follows readily from No.4. I

3.

We can now prove the geometric form of the Hahn-Banach theorem.

5.12

HYPERPLANES AND CONVEX SETS

133

Theorem 1. (Mazur's Theorem, Geometric Hahn-Banach Theorem) Let K

be a convex set
a nonempty interior in a real normed linear vector
space X. Suppose V is a linear variety in X containing no interior points
of K. Then there is a closed hyperplane in X containing V but containing no
interior points of K; i.e., there is an element x* E X* and a constant c
such that (v, x*) = c for all v E V and (k, x*) < c for all k E K.

Proof By an appropriate translation we may assume that (} is an

interior point of K. Let M be the subspace of X generated by V. Then V
is a hyperplane in M and does not contain (}; thus there is a linear functionalf on M such that V = {x :f{x) = 1}.
Let p be the Minkowski functional of K. Since V contains no interior
point of K, we have f{x) == 1 p{x) for x E V. By homogeneity, f{rxx) =
rx p(rxx) for x E V and rx > O. While for rx < 0, f{rxx) 0 p{rxx). Thus
f{x) p{x) for all x EM. By the Hahn-Banach theorem, there is an
extension F off from M to X with F{x)
Let H= {x: F{x) = I}.
Since F{x) p{x) on X and since by Lemma 1 p is continuous, F is
continuous, F(x) < 1 for x E K, therefore, H is the desired closed hyperplane1
There are several corollaries and modifications of this important
theorem, some of whiich are discussed in the remaind,:r of this section.

Definition. A closed hyperplane H in a normed space. X is said to be a

support (or a supporting hjlperplane) for the convex set Kif K is contained
in one of the closed half-spaces determined by li and H contains a point
of K.
Theorem 2. (Support Theorem) If x is not an interior point of a convex
set K which contains interior points, there is a closed hyperplane H containing x such that K lies on one side of H.
As a consequence of the above theorem, it follows that, for a convex
set K with interior points, a supporting hyperplane can be constructed
containing any boundary point of K.

Theorem 3. (Eidelheit Separation Theorem) Let Kl and K2 be convex sets

in X such that Kl has interior points and K2 contains no interior point of K I.
Then there is a closed hyperplane H separating KI and K2; i.e., there is an
x* E X* such that sup <x, x*) S; inf <x, x*). In other words, Kl and K2
xeKt

"eK2

lie in opposite half-spaces determined by H.

Proof Let K == KI - K2 ; then K contains an interior point and is

not one of them. By Theorem 2 there is an x* e X*, x* ::p. such that
(x, x*) 0 for x K. Thus for XI E K l , X2 E K2 , (Xl' x*) (X2 ,x*).

e,

134

DUAL SPACES

Consequently, there is a real number c such that sup <k 1 , x*)

KI

inf <k2 , x*). The desired hyperplane is H = {x ; <x, x*) = c}.

kz

Theorem 4. If K is a closed convex set and x K, there is a closed halfspace that

K but does not contain x.
Proof. Let d = inf

keK

Ilx - kll. Then d> 0 since K is closed. Let S be

the open sphere about x of radius d/2. Then apply Theorem 3 to Sand K.

The previous theorem can be stated in an alternative form.

Theorem 5. If K is a closed convex set in a normed space, then K is equal
to the intersection of all the closed half-spaces that contain it.
Theorem 5 is often regarded as the geometric foundation bf duality
theory for convex sets. By associating closed hyperplanes (or half-spaces)
with elements of X*, the theorem expresses a convex set in X as a collection
of elements in X *.
The appeal of the above collection of theorems is that they have simple,
geometrically intuitive interpretations. These apparently simple geometric
facts lead to some fairly profound and useful principles of optimization.
*5.13 Duality in Minimum Norm Problems
In this section we generalize the duality principle for minimum norm
problems to include the problem of finding the minimum distance from a
point to a convex set. Our development of this generalization is based on
the geometric notions of separating hyperplanes formulated in the last
section.
The basic principle of duality is illustrated in Figure 5.4: the minimum
distance from a point to a convex set K is equal to the maximum of the

Figure 5.4

Duality

S.13

DUALITY IN MINIMUM NORM PROBLEMS

135

distances from the point to hyperplanes separating the point and the
convex set K. We translate this simple, intuitive, geometric relation
into algebraic form and show its relation to our earlier duality result.
Since the results of thi:s section are included in the more general theory of
Chapter 7, and since the machinery introduced is not explicitly required
for later portions of the book, the rea:ler may wish to skip this section.

Definition. Let K be a. convex set in a real normed vector space X. The

functional h(x*) = sup (x, x*) defineJ on X* is called the support
xeK

functional of K.
In general, h(x*) may be infinite.
The support functional is illustrated in Figure 5.5. It can be interpreted
geometrically as follows: Given an element x* E X*, we consider the
family of half-spaces {x: (x, x*) ::;; c} as the constant c varies. As c
increases, these half-spaces get larger and h(x*) is defined as the infimum

(x, x*) = h(x*)

Figure 5.5 The support functional

of those constants c such that K is contained within the half-space. (The
reader should verify tllle equivalence of this last statement with the defini
tion above.) In view of this interpretation, it is clear that the support
functional of a convex set K completely specifies the set-to within closure
-since, according to Theorem 5 of the last section,

K=

n{x: (x, x*) :s: h(x*)}.

x'

136

DUAL SPACES

The support functional has other interpretations as well, some of which

are discussed in the problems. Note, in particular, that if K is the unit
sphere in X, then h(x*) is simply the norm in X*.
There is a final interpretation of the support functional that directly
serves our original objective of expressing in analytical form the duality
principle for the minimum norm problem iIIustrated in Figure 5.4. Let
K be a convex set which is a finite distance from e, let x* E X* have
Ilx* II = 1, and suppose the hyperplane H = {x : (x, x*) :::: h(x*)} is a support hyperplane for K separating e from K; then the distance from e to H
is -h(x*). This interpretation is both verified and applied to the minimum
norm problem in the proof of the next theorem. Figure 5.6 iIIustrates the
result and the method of proof.

(x, x*) = II(x*)

Figure 5.6

< 0,

IIx* \I

=I

Proof of Theorem 1

Theorem 1. (Minimum Norm Duality) Let Xl be a point in a real normed

vector space X and let d> 0 denote its distance from the convex set K
having support functional h; then

d= infllx-xlll= max [<xl,x*)-h(x*)]

IIx*ll:S;l

xeK

where the maximum on the right is achieved by some

E X*.
If the infimum 011 the left is achieved by some Xo E K, then - xri is aligned
with Xo - Xl'

Proof For simplicity we take x t

= 0 since the general case can then

be deduced by translation. Thus we must show that

d == inf Ilxll = max - h(x*).

xeK

Ilx*II:s;1

5.14

}ROBLEMS

137

We first show that for any x* e X*, IIx*11 S; 1, we have d -h(x*).

For this we may obviously limit our attention to those x*'s which render
h(x*) negative. If, however, h(x*) is negative, K is contained in the halfspace {x: (x, x*) S; h(x*)}. And since (e, x*) = 0, this half-space does
not contain e. Therefore, if h(x*) is negative, the hyperplane H = {x :
(x, x*> = h(x*)} separates K and e.
Let See) be the sphere centered at e of radius e. For any x* e X*, having
h(x*) < 0 and Ilx* I == 1, let e* be the supremum of the e's for which the
hyperplane {x : (x, x*) = h(x*)} separates K and See). Obviously, we have
Os; s* s; d. Also h(x*):= inf (x, x*) = -s*. Thus for every x* e X*,

Ilx*11

IIxll <

1, we have -h(x*) S; d.
On the other hand, since K contains no interior points of Sed), there
is a hyperplane separating Sed) and K. Therefore, there is a X6 e X*,
= 1, such that -h(X6) = d.
To prove the statement concerning alignment, suppose that Xo E K,
IIxoll = d. Then (xo, X6> S;
= -d since Xo E K. However,
-(xo ,
S; IIx61111xoil = d. Thus -(xo,
= IIx61111xoII and -X6 is
aligned with Xo' I
Example 1. Suppose that K = M is a subspace of the normed space X
and that x is fixed in X. Theorem 1 then states that
inf Ilx - mil = max [(x, x*) - h(x*)].
S;

IIx*Usl

mllM

It is clear, however, that, corresponding to M, h(x*) is finite only for

x* E Ml., in which case it is zero. Therefore, we obtain

inf

Ilx - mil

meM

max (x, x*)

IIx*lIsl
xeMl.

which is equivalent to our earlier duality result: Theorem 1, Section 5.8.

The theorem of this section is actually only an intermediate result in
the development of duality that continues through the next few chapters
and as such it is not of major practical importance. Nevertheless, even
though this result is superseded by more general and more useful duality
relations, its simplicity and geometric character make it an excellent first
example of the interplay between maximiz8tion, minimization, convexity,
and supporting hyperplanes.
5.14

Problems

1. Define the linear functional! on L 2 [0, 1] by

f(x) =

f aCt) f b(s)x(s) ds dt
1

138

DUAL SPACES

where a, bE L 2 [0, 1]. Show that

is a bounded linear functional on

= (xly).
2. Define the Banach space c as the space of all sequences x =
{el' e2 , } which converge to a limit (i.e., lim ek exists), with
L 2 [0, 1] and find an element y E L2 such thatf(x)

IIxll

3.
4.
5.
6.

lekl.

sup
I :;;k<

k-+oo

Define Co as the space of all sequences which con-

00

verge to zero (same norm as in c). Characterize the dual spaces of Co

and c (with proofs). Warning: the dual spaces of Co and c are not
identical.
Let X* be the dual of the normed space X. Show that if X* is separable, then X is separable.
Show that the normed space C[a, b] is not reflexive.
Verify the alignment criteria given in Examples 1 and 2, Section 5.7.
Suppose we wish to bring a rocket car of unit mass, and subject only
to the force of the rocket thrust, to rest at x = in minimum time by
proper choice of the rocket thrust program. The available thrust u is
limited to lu(t) 1 I for each t. Assume that initially x(O) = 0, x(O) = 1.
See Figure 5.7.

Figure 5.7

A rocket car

(a) Produce an argument that converts this problem to a minimum

norm problem on a fixed interval [0, T].
(b) Solve the problem.
7. The instantaneous thrust u(t) produced by, a rocket is a two-dimen. sional vector with components ul(t), u2 (t). The instantaneous thrust
magnitude is lu(t)12 == JUI2(t) + U2 2(t). Consideration of the maximum magnitude over an interval of time leads to a definition of a
norm as

IIuli

= max lu(t)12'
O:;;tSI

Let C 2 [0, IJ = X be defined as the space of all ordered pairs u(t) =

(ul(t), U2(t of continuous functions with norm defined as above.
2

Show that bounded linear functionals on X can be expressed in the

form

J(u) =

ul(t) dvl(t)

+ U2(t) dV2(t)

5.14

where

Vl

and

V2

PROBLEMS

139

are fu:nctions of bounded variation and

LJldvt\2 + Idvzl
1

IlfiII =

8. Let X = CP"[O, 1J be defined as the space of all continuous functions

from [0, IJ to n-dime:nsional space; i.e., each x E X is of the form
x

:5

(Xt(t), X2(t), ... , xn(t

where each Xj(t) is a continuous function on [0, 1], The norm on X is

IIxli

= sup
Ostsl

Ix(t)lp

where

Find X*.
9. Let Xl and X2 denote horizontal and vertical position components in
a vertical plane. A rocket of unit mass initially at rest at the point
Xl = X2 =
is to be propelled to the point Xl = X2 = 1 in unit time
by a single jet with components of thrust u l , U2' Making assumptions
similar to those of Example 3, Section 5.9, find the thrust program
Ul(t), U2(t) that accomplishes this with minimum expenditure of fuel,

L
1

Ju/(t)

+ ul(t) dt.

10. Let X be a normed space and M a subspace of it. Show that (within
an isometric isomorphism) M* = X*/MJ. and MJ. = (XIM)*. (See
Problem 15, Chapter 2.)
11. Let {xt} be a bounded sequence in X* and suppose the sequence of
scalars {(x, xt>} converges for each x in a dense subset of X. Show
that {xt} converges weak* to an element x* E X*.
12. In numerical computations it is often necessary to approximate certain
linear functionals by simpler ones. For instance, for X = C[O, IJ, we
might approximate
1

L(x) =

by a formula of the form

Sox(t) dt

140

DUAL SPACES

The coefficients might be found by requiring that Ln(P) be exact for

all polynomials P of degree n. Following the above scheme, find L2(X)
(Simpson's Rule).
13. Let X = C[O, I]. Suppose there is defined the two triangular arrays
of real numbers

a l1

tl1
t21

t22

t31

t32

t33

a21

a22

a31

a32

a33

a41

t41

and we construct the quadrature rules

n

LnCx) = Lank X(tnk)

k=1

which have the property that

I

Ln(P) = tp(t) dt
for any polynomial p of degree n -1 or less. Suppose also that
for all n.
Show that for any x

X
1

Ln(x)

fo x(t) dt.

14. Let K be a convex set in a real normed linear space X. Denote by

v(K) the intersection of all closed linear varieties containing K. A
point x in K is said to be a relative interior point of K if x is an interior
point of K, regarded as a subset of v(K). See Section 2.7. Let y be a
point in v(K) which is not in the relative interior of K and suppose K
has relative interior points. Show that there is a closed hyperplane
in X containing y and having K on one side of it.
15. Let X be a real linear vector space and let 11'/2' ... ,In be linear
functionals on X. Show that, for fixed IX/S, the system of equations
fl(x) = IXI

has a solution x

the relation

1 s; is; n

X if and only if for any n numbers AI' A2 , ... , An

S.14

PROBLEMS

141

implies

Hint: In the space en consider the subspace formed by all points of

the form (ft(x),fix), ... ,j,,(x as x varies over X and apply the
Hahn-Banach
16. Letgl> g2' ... , gIl be linearly independent linear functionals on a vector
space X. Let I
another linear functional on X such that for every
x e X satisfying {heX) = 0, i = 1,2, ... , n, we have/(x) = O. Show that
there are constants At, A2, ... , An such that

17. Let X be a real l:inear vector space and let It, 12 , ... ,j" be linear funtionals on X. Show that, for fixed cx/s, the system of inequalities
fi(x)

CX i

has a solution x e X if and only if fOf any n nonnegative numbers

AI' A2 , ... , An the relation

implies

18. Many duality properties are not completely symmetric except in

reflexive spaces where X = X** (under the natural mapping). The
theory of weak duality with a deemphasis of the role of the norm is
in many respects more satisfying although of somewhat less practical
importance. This problem introduces the foundation of that theory.
space and let X* be its dual. Show that the weakly
Let X be a
continuous linear functionals on X are precisely those of the form
I(x) = (x, x*) where x* E X* and that the weak* continuous linear
functionals on X* are precisely those of the form g(x*) = (x, x*)
where x e X. Hilu: Use the result of Problem 16.
19. Show that the support functional h of a convcyx set is sublinear.
20. Show that the set in X* where the support functional is finite is a
convex cone.
21. Let h be the support functional of the convex set K. Find the support
functional of the set K + x l'

142

DUAL SPACES

22. Prove the duality theorem of Section 5.13 for the case Xl =F O.
23. Let X be a real normed linear space, and let K be a convex set in X,
having 0 as an interior point. Let h be the support functional of K
and define KO = {x* E X*: h(x*) s I}. Now for X E X, let p(x) =
sup (x, x*). Show that p is equal to the Minkowski functional of K.
xeK

24. Let X, K, KO, p, h be defined as in Problem 23, with the exception that
K is now an arbitrary set in X. Show that {x: p(x) I} is equal to
the closed convex hull of K u to}.

REFERENCES
5.2-6. The presentation of these sections closely follows that of standr.rd works
such as Yosida [157], Goffman and Pedrick [59], and Taylor [145]. These

references can be consulted for additional examples of normed duals and for
a proof using Zorn's lemma of the more general Hahn-Banach theorem.
5.8-9. The duality theorems for minimum norm problems are given in the form
presented here in Nirenberg [114]. This material has a parallel development
known as the L-problem in the theory of moments, which was studied extensively by Akhiezer and Krein [4]. The solution of the L-problem has been
applied to optimal control problems by several authors. See Krasovskii [89],
Kulikowski [92], Kirillova [86], Neustadt [110], [111], and Butkovskii [26];
for an extensive bibliography, see Sarachik [137]. For additional applications
of the Hahn-Banach theorem to approximation theory, see Deutsch and
Maserick [39].
5.l1-12. The theory of convexity and hyperplanes has been studied extensively.
For some extensions of the development given in this chapter, see Klee [87],
Lorch [97], and Day [37]. For a number of counter-examples showing nonseparability of disjoint convex sets when the interior point condition is not
met, see Tukey [147].
5.l3. The duality theorem is proved in Nirenberg [11h
5.l4. Problems 7-9 are based on material from Neustadt [111].

6
LINEAR OPERATORS
AND ADJOINTS

6.1 Introduction
A study of linear operators and adjoints is essential for a sophisticated
approach to many problems oflinear vector spaces. The associated concepts
and notations of operator theory often streamline an otherwise cumbersome analysis by eliminating the need for carrying along complicated
explicit formulas and by enhancing one's insight of the problem and its
solution. This chapter contains no additional optimization principles but
instead develops results of linear operator theory that make the application
of optimization principles more straightforward in complicated situations.
Of particular importance is the concept of the adjoint of a linear operator
which, being defined in dual space, characterizes many aspects of duality
theory.
Because it is difficult to obtain a simple geometric representation of an
arbitrary linear operator, the material in this chapter tends to be somewhat
more algebraic in character than that of other chapters. Effort is made,
however, to extend some of the geometric ideas used for the study of linear
functionals to general linear operators and also to interpret adjoints in
terms of relations among hyperplanes.

6.2 Fundamentals
A transformation T is, as discussed briefly in Chapter 2, a mapping from
one vector space to another. If T maps the space X into Y, we write
T: X -+ Y, and if T maps the vector x F.: X into the vector Y E Y, we write
y = T(x) and refer to y as the image of x under T. As before, we allow that
a transformation may be defined only on a subset D C IX, called the domain
of T, although in most cases D = X. The .. ,)Uection of an vectors Y E Y for
which there is an xED with y = T(x) is called the range of T.
If T : X -+ Yand S is a given set in X, we denote by T(S) the image of S
in Y defined as the subset of Y consisting of points of the form y = T(s)
143

144

LINEAR OPERATORS AND ADJOINTS

with s E S. Similarly, given any set P c: Y, we denote by T -1(p) the inverse

image of P which is the set consisting of all points x E X satisfying
T(x) EP.
Our attention in this chapter is focused primarily on linear transformations which are alternatively referred to as linear operators or simply
operators and are usually denoted by A, E, etc. For convenience we often
omit the parentheses for a linear operator and write Ax for A(x). The range
of a linear operator A : X - Y is denoted
and is obviously a subspace
of Y. The set {x : Ax = iI} corresponding to the linear operator A is called
the nullspace of A and denoted .;V(A). It is a subspace of X.
Of particular importance is the case in which X and Yare normed
spaces and A is a continuous operator from X into Y. The following result
is easily established.

Proposition 1. A linear operator on a normed space X is continuous at every

point in X ifit is continuous at a single point.
Analogous to the procedure for constructing the normed dual consisting
of continuous linear functionals on a space X, it is possible to construct
a normed space of continuous linear operators on X. We begin by defining
the norm of a linear operator.
Definition. A linear operator A from a normed space X to a normed space

Y is said to be bounded if there is a constant M such that IIAxlls M Ilxll

for all x E X. The smallest such M which satisfies the above condition is
denoted II A I and called the norm of A.
Alternative, but equivalent, definitions ofthe norm are

IIAII
IIAII

sup IIAxll

Ilxll S

= sup IIAxll.
x8

IIx!!

We leave it to the reader to prove the following proposition.

Proposition 2. A linear operator is bounded (f and only if it is continuous.

If addition and scalar multiplication are defined by
(AI

+ A 2 )x =
(aA)x

==

Alx

+ A2 x

a(Ax)

the linear operators from X to Y form a linear vector space. If X and Yare
normed spaces, the subspace of continuous linear operators can be
identified and this becomes a normed space when the norm of an operator

6.2

FUNDAMENTALS

145

is defined according to the last definition. (The reader ciln easily verify that
the requirements for a norm are satisfied.)

Definition. The norffiied space of all bounded linear operators from the
normed space X into the normed space Y is denoted B(X, Y).
We note the following result which generalizes Theorem 1, Section 5.2.
The proof requires only slight modification of the proof in Section 5.2
and is omitted here.

Theorem 1. Let X and Y be normed spaces with Y complete. Then the space
B(X, Y) is complete.

In general the space B(X, Y), although of interest by its own right, does
not play nearly as dominant a role in our theory as that of the normed
dual of X. Nevertheless, certain of its elementary properties and the definition itself are often convenient. For instance, we write A e B(X, Y) for,
"let A be a continuous linear operator from the normed space X to the
normed space Y."
Finally, before turning to some examples, we observe that the spaces
of linear operators have a structure not present in an arbitrary vector space
Y,
in that it is possible Ito define products of operators. Thus, if S : X
T: Y -+ Z, we define the operator TS : X -+ Z by the equation (TS)(x) =
T(Sx) for all x e X. For bounded operators we have the following useful
result.

Proposition 3. Let X, Y, Z be normed spaces and suppose S e B(X, Y),

Te B(Y, Z). Then IITSII :s; IITIIIISIi.

Proof I TSxl1 :s; IITllllSxll :s; IITllllSllllxl1

for all x e X.

Example 1. Let X =: C[O, 1] and define the operator A: X

X by

Ax = SA K(s, t)x(t) dt where the function K is contmuous on the unit

square 0 :s; s :s; 1,0 :s; t :s; 1. The operator A is clearly linear. We compute

IIAII. We have
IIAxll= max I{K(S,t)X(t)dt\
o sss 1

:s; max {f1IK(S, t)1 dt} max Ix(t)1

OS.S1

OS!Sl

f IK(s, t)1 dt . IIxll.

1

= max

o s.s 1

Therefore,

f IK(s, t)1 dt.

1

IIAII :s; max

OSsSl

146

UNBAR OPERATORS AND ADJOINTS

We can show that the quantity on the right-hand side is actually

norm of A. Let

So

the

be the point at which the continuous function

1

fo IK(s, 1)1 dt
achieves its maximum. Given e > 0 let p be a polynomial which approximates K(so , . ) in the sense that
max IK(so, t) - p(t)1 < e
1

and let x be a function in C [0, 1) with Ilxll 1 which approximates the

discontinuous function sgn p(t) in the sense that

IS: p(t)x(t) dt -

s: Ip(t)1 dt I< e.

This last approximation is easily constructed since p has only a finite

number of sign changes.
For this x we have

IS: K(so , t)x( t) dt I

IS: p(t)x(t) dt I-I s: [K(so, t) - p(t)Jx(t) dt I

J:

IS: p(t)x(t) dt /- e
(IK(So, t)1 dt -/

s:

Ip(t)1 dt - 2e

[IK(so, t)1 - lp(t)l] dt /- 2e

J IK(so, t)1 dt -

3e.

Thus, since

IIxll ::; 1,
1

IIA"

Jo IK(so, t)1 dt -

3e.

But since e was arbitrary, and since the reverse inequality was established
above, we have

IIAII

f. IK(s, t)1 dt.

1

= max
o

Example 2. Let X = En and let A : X

X. Then A is a matrix acting on the

components of x. We have llAx1l2 = (x I A' Ax) where A' is the transpose of
the matrix A. Denoting A' A by Q, determination of IIAll is equivalent to
maximizing (x I Qx) subject t6 IIxl12 1. This is a finite-dimensional
optimization problem. Since Q is symmetric and positive semidefinite, it

6.3

LINEARITY OF INVERSES

147

has nonnegative eigenvalues and the solution of the optimization problem

is given by x equal to the eigenvector of Q corresponding to the largest
eigenvalue.
We conclude that IIAII =-jXmax
Example 3. The operator Ax = did! x(t), defined on the subspace M of
CEO, 1] consisting of all continuously differentiable functions, has range
CEO, 1]. A is not bounded, however, since elements of arbitrarily small
norm can produce elements of large norm when differentiated. On the
other hand, if A is regarded as having domain D [0, IJ and range C [0, IJ,
it is bounded with II A II = 1.

OPERATORS
6.3 Linearity of Invers,!s
Let A : X -+ Y be a linear operator between two linear spaces X and Y.
Corresponding to A we consider the equation Ax = y. For a given Y E Y
this equation may:
1.
2.
3.

have a unique solution x E X,

have no solution,
have more than one solution.

Many optimization problems can be regarded as arising from cases 2 or

3; these are discussed in Section 6.9. Condition 1 holds for every Y E Y
if and only if the mapping A from X to Y is one-to-one and has range equal
to Y, in which case the operator A has an inverse A - I such that if Ax = y,
then A -I(y) = x.
Proposition 1.
A - I is linear.

If

a linear operator A : X -+ Y has an inverse, the inverse

Proof. Suppose A -1(YI) = Xl' A -1(Y2) = x 2 , then

A(Xl) = Yl'

A(x2) = Y2'

and the linearity of A implies that A(OCIXI + OC2 X2) =octYt

A-I(OC1Yt + OC2Y2) = ocIA-I(YI) + oc2A-1(Y2)' I

+ OC2Y2'

Thus

The solution of linear equations and the determination of inverse

are, of course, important areas of pure and applied mathematics.
For optimization theory,. however, we are not so much interested in solving
equations as formulating the equations appropriate for characterizing an
optimal vector. Once the equations are formulated, we may rely on standard
techniques for their solution. There are important exceptions to this point

148

LINEAR OPERATORS AND ADJOINTS

of view, however, since optimization theory often provides effective procedures for solving equations. Furthermore, a problem can never really be
regarded as resolved until an efficient computational method of solution is
derived. Nevertheless, our primary interest in linear operators is their
role in optimization problems. We do not develop an extensive theory
of linear equations but are content with establishing the existence of a
solution.

6.4 The Banach Inverse Theorem

Given a continuous linear operator A from a normed space X onto a
normed space Y and assuming that A has an inverse A-I, it follows that
A -1 is linear but not necessarily continuous. If, however, X and Yare
Banach spaces, A -1 must be continuous ifit exists. This result, known as the
Banach inverse theorem, is one of the analytical cornerstones of functional
analysis. Many important, deep, and sometimes surprising results follow
from it. We make application of the result in Section 6.6 and again in
Chapter 8 in connection with Lagrange multipliers. Other applications to
problems of mathematical analysis are discussed in the problems at the end
of this chapter.
This section is devoted to establishing this one result. Although the proof
is no more difficult at each step than that of most theorems in this book,
it involves a number of steps. Therefore, since it plays only a supporting
role in the optimization theory, the reader may wish to simply scan the
proof and proceed to the next section.
I
We begin by establishing the following lemma which itself is an important and celebrated tool of analysis.

Lemma 1. (Baire) A Banach space X is not the union of countably many

nowhere dense sets in X.
Proof Suppose that {E.} is a sequence of nowhere dense sets and let

Fn denote the closure of En' Then Fn contains no sphere in X. It follows that

each of the sets; n is open and dense in X.

Let S(X1' r 1) be a sphere in F\ with center at Xl and radius r 1. Let

S(X" , r,,) be a sphere in P" (') S(x!, rd2). (Such a sphere exists since P2
is open and dense.) Proceeding inductively, let S(x., rn) be a sphere in
S(x._1> r._ 1/2) (') p.
The sequence {xn} so defined is clearly a Cauchy sequence and, thus,
by the completeness of X, there is a limit x; Xn - t x. This vector x lies in
each of the S(xn' r.), because, indeed, x. +k ES(X., rn/2) for k 1.
Hence X lies in eachPn Therefore, x

nn Fn.

6.4

THE BANACH INVERSE THEOREM

149

It follows that the union of the original collection of sets {En} is not X
since

Theorem 1. (Banach InlJerse Theorem) Let A be a continuous linear operator

from a Banach space X onto a Banach space Y and suppose that the inverse
operator A -1 exists. ThEm A -1 is continuous.

Proof In view of the linearity of A and therefore of A -1, it is only

necessary to show that A -1 is bounded. For this it is only necessary to show
that the image A(S) in Y of any sphere S centered at the origin in X contains
a sphere P centered at the origin in Y, because then the inverse image of P
is contained in S. The proof amounts to establishing the existence of a
sphere in A(S).
Given a sphere S, for any x E X there is an integer n such that x/n E S
and hence A(x/n) E A(8) or, equivalently, A(x) E nA(S). Since A maps X
onto Y, it follo:ws that
00

y =

UnA(S).

"=1

According to Baire's l,emma, Y cannot be the union of countably many

nowhere dense sets and, hence, there is an n such that the closure of nA(S)
contains a sphere. It follows that A(S) contains a sphere whose center y
may be taken to be in A(S). Let this sphere N(y, r) have radius r, and let
y = A(x). Now as y' varies over N(y, r), the points y' - y cover the sphere
N(}, r) and the points of a dense subset of these are of the form A(x' - x)
where A(x' ) = y', x' E S. Since x', XES, it follows that x' - x E 2S. Hence,
the closure of A(2S) contains N(}, r) (and by linearity A(S) contains

N(}, r/2.
We have shown tha.t the closure of the image of a sphere centered at
the origin contains such a sphere in Y, but it remains to be shown that
the image itself, rather than its closure, contains a sphere. For any i:> 0,
let S(8) and P(8) be the spheres in X, Y, respectively, of radii 8 centered
at the origins. Let 8 0 > 0 be arbitrary and let '10 > 0 be chosen so that
P('1o) is a sphere contained in the closure of the image of S(80). Let y
be an arbitrary point in P('1o). We show that there is an x E S(280 ) such
that Ax = y so that the image of the sphere of radius 2110 contains the
I
sphere P('1o).
Let {Si} be a sequence of positive numbers such that Ii";,! 6; " 60' Then
there is a sequence {'11}, with '11 > 0 and '11-+ 0, such that

150

LINEAR OPERATORS AND ADJOINTS

Since A(S(llo)) is dense in peso), there is an Xo e S(60) such that y - Axo e

P('1t). It follows that there is an XI e 5(61) with y - Axo - AXI eP('12)'
Proceeding inductively, a sequence {x.} is defined with X. e S(6.) and
y-A(l:r<=ox/)eP('1.+')' Let z.=xo+x,+"+x. Then evidently
{z.} is a Cauchy sequence since form > n, liz,. - z.11 = Ilx.-, + X. -2 + ... +
xmll <6.+1+6.+ 2 ++6m Thus there is an xeX such that
Zn -+ x. Furthermore, Ilxll < 60 + 6, + ... + 6. + ... < 260; so X e 5(260)'
Since A is continuous, Az. -+ Ax, but since Ily - Az.11 < '1.+ 1 -+ 0,
Az. -+ y. Therefore, Ax = y. I
ADJOINTS

6.5

Definition and Examples

The constraints imposed in many optimization problems by differential

equations, matrix equations, etc., can be described by linear operators.
The resolution of these problems almost invariably calls for consideration
of an associated operator: the adjoint. The reason for this is that adjoints
provide a convenient mechanism for describing the orthogonality and
duality relations which permeate nearly every optimization analysis.

Definition. Let X and Y be normed spaces and let A E B(X, Y). The adjoint
operator A*: y* -+ X* is defined by the equation
(x, A*y*

>= (Ax, y*).

This important definition requires a bit of explanation and justification.

Given a fixed y* eY*, the quantity (Ax, y*) is a stalar for each x eX
and is therefore a functional on X. Furthermore, by the linearity of y* and
A, it follows that this functional is linear. Finally, since
I(Ax, y*)1 :s;

Ily*IIIIAxll :s; Ily*IIIIAllllxll,

it follows that this functional is bounded and is thus an element x* of X*.

We then define A*y* = x*. The adjoint is obviously unique and the reader
can verify that it is linear. It is important to remember, as illustrated in
Figure 6.1, thatA*: y* -+ X*.
A*

Figure 6.1

r*

An operator and its adjoint

6.5

DEFINITION AND EXAMPLES

151

In terms of operator, rather than bracket, notation the definition of the

adjoint satisfies the equation
y*(Ax)

for each x

E X.

=(A*y*)(x)

Thus we may write

y*A = A*y*

where the left side denotes the functional on X which is the composition
of the operators A and y* and the right side is the functional obtained by
operating on y* by A *.

Theorem 1. The adjoint ,operator A* of the linear operator A

linear and bounded with II A * II = II A II

B(X, Y) is

Proof The proof of linearity is elementary and left to the reader.

From the inequalities

I(x, A*y*)1

== I(Ax,

Y*)I

Ily*IIIIAxll

1IY*IIIIAllllxll

it follows that

IIA*y*11

IIAIlIIY*11

which implies that

IIA*II

IIAII

Now let Xo be any nonzero element of X. According to Corollary 2 of

the Hahn-Banach theorem, there exists an element
E Y*,
== 1,
such that (Axo ,
= I Axo II. Therefore,

IIAxo11

I(xo,

IIA*llllxoll

from which we conclude that

IIA*II
It now follows that

IIA*II

IIAII I

In addition to the above result, adjoints enjoy the following algebraic

relations which follow easily from the basic definition.

Proposition 1. Adjoints satisfy the following properties:

2.

If I is the identity operator on a normed space X,

If A 1, A2 E B(X, Y), then (A1 + A 2 )* = A! +

3.
4.
5.

If Al EB(X, Y), E B(Y, Z), then (A2 A 1)* = A1A!.

If A E B(X, Y) and A has a bounded inverse, then (A -1)* =

1.

then 1* == I.

If A E B(X. Y)andrx isa real scalar, then (rxA)* == e<A*.

(A *) -1.

152

LINEAR OPERATORS AND ADJOINTS

Proof Properties 1-4 are trivial. To prove property 5, let A E B(X, Y)

have a bounded inverse A -1. To show that A* has an inverse, we must show
that it is one-to-one and onto. Let yt =/:
Y*, then

, ,'. \,' c' ':1,"".

= (Ax, yt

<x, A*yt) - <x,

=/:

for some x E X. Thus, A*yf =/: A*yi and A* is one-to-one. Now for any
x* E X* and any x E X, Ax = y, we ha,ve
(x, x*)

= <A- 1y, x*) =

<y, (A- 1)*x*)

= <Ax, (A- 1 )*.x*)

which shows that x* is in

= <x, A*(A- 1 )*x*)

and also that (A*) -1 = (A -1)*.

An important special case is that of a linear operator A : H G where

Hand G are Hilbert spaces. If Hand G are real, then they are their own
duals in the sense of Section 5.3, and the operator A* can be regarded as
mapping G into H. In this case the adjoint relation becomes (Ax Iy) =
(x I A*y). If the spaces are complex, the adjoint, as defined earlier, does not
satisfy this relation and it is convenient and customary to redefine the
Hilbert space adjoint directly by the relation (Ax Iy) = (x I A *y). In our
study, however, we restrict our attention to real spaces sO that difficulties
of this nature can be ignored.
Note that in Hilbert space we have the additional property: A** = A.
Finally, we note the following two definitions.

Definition. A bounded linear operator A mapping a real Hilbert space into

itself is said to be self-adjoint if A * = A.
Definition. A self-adjoint linear operator A on a fIilbert space H is said to
be positive semidefinite if (x I Ax) 2 for all x E H.

Example 1. Let X = Y = En. Then A: X -t X is represented by an n x n

matrix. Thus the i-th component of Ax is
n

(Ax);= 'LaijXj.
J=1

WecomputeA*. Fory EYwe have

n

(Axl y)

= 1=1
L LYialjXj =
j=1

where A* is the matrix with elements

of A.

Example 2. Let X

j= 1

n
Xj

au =

l>ijYi

1"'1

ajl'

= (xIA*Y)

Thus A* is the transpose

= Y = L 2 [0, 1] and define

Ax =

K(t, s)x(s) ds,

t E [0, 1J

6.S

DEFINITION AND EXAMPLES

153

where
1

fo fo IK(t, sW ds dt <

00.

Then

(Ax I y) = (y(t):J; K(t, S)X(s)dsidt

=fo XeS) f0 K(t, s)y(t) dtds.

1

Or, by interchanging the roles of sand t,

(Ax I y)

= fo x(t) fo K(s, t)y(s) ds dt = (x I A*y)

where
1

A*y = fo K(s, t)y(s) ds.

Therefore, the adjoint of A is obtained by interchanging sand tin K.
Example 3. Again let X

= Y = L 2 [0, 1] and define

t

Ax = foK(t, s)x(s) ds,

t E [0, 1J,

with
1

sW dt ds < 00.

fo fo IK(t,
Then
1

(Ax I y)

= fo yet) fo K(t, s)x(s) ds dt

1

= fo
,

fa y(t)K(t, s)x(s) ds dt.

t

The double integration represents integration over the triangular region

shown in Figure 6.20, integrating vertically and then horizontally. Alternatively, the integration may be performed in the reverse order as in Figure
6.2b, leading to

(Ax I y)

= fo fa y(t)K(t, s)x(s) dt ds
1

= (X(S)({K(t, s)y(t) dt) ds.

154

LINEAR PPERATORS AND ADJOINTS

(b)

(0)

Figure 6.2 Region of integration

Or, interchanging the roles of t and s,

(Ax\y)= S:X(t)({K(S,t)y(S)dS) dt=(x\A*y)

where

A *y =

JK(s, t)y(s) ds.

1

This example comes up frequently in the study of dynamic systems.

Example 4. Let X = C [0, 1], Y

= en and define A : X -+ Y by the equation

Ax = (x(t l ), x(t 2), ... , x(tn

: ;

where
t 1 < t2 < t2 < '" < tn :::;; 1 are fixed. It is easily verified that A
is continuous and linear. Let y* = (YI' Y2' .. , Yn) be a linear functional
on En. Then

(Ax, y*)

L YIX(t = f x(t) dv(t) = (x, A,*Y*)

I

1= 1

l)

where vet) is constant except at the points tt where it has a jump of magnitude Yi' as illustrated in Figure 6.3. Thus A*: en -+ NBV[O, 1] is defined
by A*y* = v.
v

' - _ L - -_ _ _

tl

t2

. .r

2I

tn

Figure 6.3 The function v

>- t

6.6

RELATIONS BETWEEN RANGE AND NULLSPACE

1SS

6.6 Relations Between Range and Nullspace

Adjoints are extremely useful in our recurring task of translating between
the geometric properties and the algebraic description of a given problem.
The following theorem and others similar to it are of particular interest.
Theorem 1. Let X and Y be normed spaces and let A
:0"

B(X, Y). Then

';v(A*).

Proof. Let y* E .;V(A*) and y E 9P(A). Then y = Ax for some x EX.

The calculation <y, y*) = <Ax, y*) = <_'i:', A*y*) = 0 shows that .;V(A*) c
[9P(A)].l.
Now assume y* e [9P(A)].l. Then for every x E X, <Ax, y*) = O. This
implies <x, A*y*) = 0 and hence that [9P(A)].l c ';v(A*). I

Example 1. Let us c:onsider the finite-dimensional version of Theorem 1.

Let A be a matrix; A: En -+ Em. A consists of n column vectors aj,

i = 1,2, ... , n, and

is the subspace of Em spanned by these vectors.
[9l(A)].l consists of those vectors in Em that are orthogonal to each al.
On the other hand, the matrix A* (which is just the transpose of A) has
the a/s as its rows; hence the vectors in Em orthogonal to a/s comprise the
nullspace of A*. The:refore, both [9P(A)].l and .;V(A*) consist of all vectors
orthogonal to each Gf l
Our next theorem is a dual to Theorem 1. It should be noted, however,
that the additional hypothesis that 9P(A) be closed, is required. Moreover,
the dual theorem is much deeper than Theorem 1, since the proof requires
both the Banach inverse theorem and the Hahn-Banach theorem.
Lemma 1. Let X and Y be Banach spaces and let A e B(X, Y). Assume that
9P(A) is closed. Then there is a constant K such that for each y E 9P(A) there

is an x satisfying Ax = y and IIxlt

Kllyll.

Proof Let N = .;V(A) and consider the space XI N consisting of equivalence classes [x] modulo N. Define A: XIN -+ 9P(A) by A[x] = Ax. It is
easily verified that ,4 is one-to-one, onto, linear; and bounded. Since 9P(A)
closed implies that 9l(A) is a Banach space, it follows from the Banach
inverse theorem th:at A has a continuous inverse. Hence, given y E 9l(A),
there is [x] e XIN with \I[x]\I
\lJ-l\1\1y\l. Take x e [x] with
\lxll 2 II [x] II and then K = 211A- 1 11 satisfies the
stated in the
lemma. I
N oW we give the dual to Theorem 1.

156

LINEAR OPERATORS AND ADJOINTS

Theorem 2. Let X and Y be Banach spaces and let A E B(X, Y). Let fA(A) be
closed. Then
= [%(A)]J..

Proof Let x* E
%(A), we have

Then x* = A*y* for some y*

(x, x*) = (x, A*y*) = (Ax, y*)

Y*. For any

= O.

Thus x* E [%(A)]J. and it follows that

*) c:: [%(A)]J..
Now assume that x* E [%(A)]J.. For y E &leA) and each x satisfying
Ax = y, the functional (x, x*) has the same value. Hence, define fey) =
(x, x*) on
Let Kbe defined as in the lemma. Then for eachy E
there is an x with Ilxll Kllyll, Ax = y. Therefore, If(y) I Kllx*llllyll
Extendfby the Hahnand thusfis a bounded linear functional on
Banach theorem to a functional y* E Y*. Then from
(x, A*y*) = (Ax, y*) = (x, x*),

it follows that A*y* = x* and thus

.:J

[%(A)]J..

In many applications the range of the underlying operator is finite

dimensional, and hence satisfies the closure requirement. In other problems,
however, this requirement is not satisfied aid this generally leads to severe
analytical difficulties. We give an example of an operator whose range is
not closed.

Example 2. Let X = Y = II with A: X -+ Y defined by

Then
However,

contains all finitely nonzero sequences and thus

and thus

= { 1, 2I2 ' I32 " ' "

n 2 ""

= Y.

is not closed.

In Hilbert space there are several additional useful relations between

range and nullspace similar to those which hold in general normed space.
These additional properties are a consequence of the fact that in Hilbert
space an operator and its adjoint are defined on the same space.

6.7

DUALITY RELATIONS FOR CONVEX CONES

157

Theorem 3. Let A be a bounded linear operator acting between two real

Hilbert spaces. Then
1.

[&l(A)JJ. = %(A*).

2. &leA) = [%(A*)]J..
3. [&l(A*)]J. = %(A).
4. &l(A*) = [%(A)JJ..
Proof. Part 1 is just Theorem 1. To prove part 2, take the orthogonal
complement of both sides of 1 obtaining [&l(A)]J.J. == [%(A*)JJ.. Since
&leA) is a subspace, the result follows. Parts 3 and 4 are obtained from
1 and 2 by use of the relation A** = A. I

e2.. ,},defineAx = {O, el, e2,.. .}.

2,
A is a shift operator (sometimes referred to as the creation operator because
a new component is created). The adjoint of A is easily computed to be the
operator taking y == {111> 11", .. } into A*y = {11", 113' ... }, which is a
shift in the other direction (referred to as the destruction operator). It is
clear that [&l(A)JJ. consists of all those vectors in I" that are zero except
possibly in their first component; this subspace is identical with %(A*).

Example3. Let X = Y = 1 For x = {el>

6.7

Duality Relations for Convex Cones

The fundamental algebraic relations between nullspace and range for an

operator and its adjoint derived in Section 6.6 have generalizations which
often playa role in the analysis of problems described by linear inequalities
analogous to the role of the earlier results to problems described by linear
equalities.

Definition. Given a set S in a normed space X, the set SfD = {x* e X*:
(x, x*) ::::: 0 for all xeS} is called the positive conjugate cone of S. Likewise the set Sa == {x* e X* : (x, x*) s; 0 for all xeS} is called the
negative conjugate cone of S.
It is a simple matter to verify that SfD and Sa are in fact convex cones.
They are nonempty since they always contain the zer9 functional. If S is a
subspace of X, then obviously SfD = Sa = SJ.; hence, the conjugate cones
can be regarded as generalizations of the orthogonal complement of a set.
The definition is illustrated in Figure 6.4 for the Hilbert space situation
where SfD and Sa can be regarded as subsets of X. The basic properties

158

LINEAR OPERATORS AND ADJOINTS

Figure 6.4 A set and its conjugate cones

of the operation of taking conjugate cones are given in the following

proposition.

Proposition 1. Let Sand T be sets in a nor'{led space X. Then

1.

S$ is a closed conVeX cone in X*.

2.

IfS c: T, then T$ c: S$.

In the general case the conjugate cone can be interpreted as a collection

of halfspaces. If x* E S $, then clearly inf <x, x*) 0 and hence the
xeS

hyperplane {x: <x, x*) = o} has S in its positive half-space. Conversely,

if x* determines a hyperplane having S in its positive half-space, it is a
member of S $. Therefore, S $ consists of all x* which contain S in their
positive half-spaces.
The following theorem generalizes Theorem I of Section 6.6.

Theorem 1. Let X and Y be normed linear spaces and let A

R(X, Y). Let S

be a subset of X. Then

(where the inverse denotes the inverse image of SEll).

s E S. Then
arbitrary in S, y*

Proof Assume y* E [A(S)]$ and

<s, A*y*)

O. Thus, since
ment is reversible. I

s is

<As, y*)
E

0 and hence
A*-l(S$). The argu-

6.8

GEOMETRIC INTERPRETATION OF ADJOINTS

Note that by putting S = X, SGl

[B(A)]i = ,Ai(A*).
*6.8

= {O},

159

the above result reduces to

Geometric Inlterpretation of Adjoints

It is somewhat diffic:ult to obtain a clear simple visualization of the relation

between an operat()r and its adjoint since if A : X Y, A* : Y* .... X*,
four spaces and two operators are involved. However, in view of the unique
correspondence between hyperplanes not containing the origin in a space
and nonzero elements of its dual, the adjoint A* can be regarded as mapping hyperplanes in Y into hyperplanes in X. This observation can be used
to consolidate the adjoint relations into two spaces rather than four. We
limit our discussion here to invertible operators between Banach spaces.
The arguments can be extended to the more general case, but the picture
becomes somewhat more complex.
Let us fix our attention on a given hyperplane HeX having 0 H.
The operator A maps this hyperplane point by point into a subsetL of Y.
It follows from the linearity of A that L is a linear variety, and since A is
assumed to be invertible, it follows that L is in fact a hyperplane in Y not
containing 0 e Y. Therefore, A maps the hyperplane H point by point into
a hyperplane L.
The hyperplanes Hand L define unique elements x! e X* and y! e y*
through the relations H {x: (x, x!> = I}, L = {y: (y, y'l)
I}. The
adjoint operator A* can then be applied to y! to produce an xt or,
equivalently, A * maps L into a hyperplane in X. In fact, A * maps L back
to H. For if A*yt := xt, it follows dim;tly from the definition of adjoints
that
{x: (x, xt> = I} = {x: (x, A*yt) = I} = {x: (Ax, y!> '= l} = H.
Therefore, A * maps the hyperplane L, as a unit,
to the hyperplane H.
This interpretation is illustrated in Figure 6.5 where the dotted line arrows
symbolize elements of a dual space.
Another geometI'ic interpretation is discussed in Problem 13.

---

'\
.

...

..............
------ --""......
........ "
1

'

Figure 6.5 Geometric interpretation of adjoints

160

LINEAR OPERA TORS AND ADJOINTS

OPTIMIZATION IN HILBERT SPACE

Suppose A is a bounded linear operator from a Hilbert space G into a

Hilbert space H; A : G - H. Then, as pointed out previously, the linear
equation Ax = y may, for a given Y E H,
1.
2.
3.

possess a unique solution x E G,

possess no solution,
possess more than one solution.

Case 1 is in many respects the simplest. We found in Section 6.4 that in

this case A has a unique bounded inverse A - 1. The other two cases are of
interest in optimization since they allow some choice of an optimal x to
be made. Indeed, most of the problems that were solved by the projection
theorem can be viewed this way.
6.9

The Normal Equations

When no solution exists (case 2), we resolve the problem by finding an

approximate solution.

and let A E B(G, lJ). Then

Theorem 1. Let G and H be Hilbert
E H the vector x E G minimizes lIy - Axil if and only if

for a fixed y
A*Ax = A*y.

Proof The problem is obviously equivalent to that of minimizing

yll where y E
Thus, by Theorem 1, Section 3.3 (the projection
theorem without the existence part), y is a minimizing vector if and only if
y - yE
Hence, by Theorem 3 of Section 6.6, y - YE ';v(A*). Or
lIy -

() = A*(y -

y)

= A*y -

A*Ax.

Theorem I is just a restatement of the first form of the projection theorem

applied to the subspace
There is no statement of existence in the
may not be closed. Furthermore, there is no
theorem since in general
statement of uniqueness of the minimizing vector x since, although
y = Ax is unique, the preimage of y may not be unique. If a unique
solution always exists, Le., if A *A is invertible, the solution takes the form
x = (A*A)-IA*y.

Example 1. We consider again the basic approximation problem in Hilbert

space. Let {xl' X2' ... , x n} be an independent set of vectors in a real
Hilbert space H. We seek the best approximation to y H of the form

y=t?=la,Xj.

6.JO

THE DUAL PROBLEM

161

Define the operator A : E" -+ H by the equation

Aa

L a/xI,

i= I

where a = (ai' "', a.). The approximation problem is equivalent to

minimizing lIy - Aall. Thus, according to Theorem 1, the optimal solution
must satisfy A *Aa =: A *y.
It remains to compute the operator A*. Clearly, A*:H-tE. For any
x E H, a E E", we have
"

(x I Aa) = (x I L a/ XI) =
i= 1

n
L
al(x I Xi) = (z I a)En
i=l

'.

where z = x I XI)' ... , (x I x.)). Thus, A*x == x I XI)' (x I x 2 ), " ' , (x I x.)).
The operator A*A maps E into E and is therefore represented by an
n x n matrix. It is then easily deduced that the equation A *Aa = A*y
equivalent to

(XIIXI) (x2Ixl)
(XI I x2)
[

.:
(xlix.)

... (X.IXI)]al]
a2
:

(X. I x") a.

(YIXI)]
(y I X2)
=

'

(y I x.)

the normal equations.

The familiar arguments for this problem show that the normal equations
possess a unique solution and that the Gram matrix A *A is invertible.
Thus, a = (A*A)-IA*y.
The above example illustrates that operator notation can streamline an
optimization analysis by supplying a compact notational solution. The
algebra required to compute adjoints and reduce the equations to expressions involving the original problem variables is, however, no shorter.

6.10 The Dual Problem

If the equation Ax = y has more than one solution, we may choose the
solution having minimum norm.
Theorem 1. Let G and H be Hilbert spaces and let A E B(G, H) with
range closed in H. Then the vector x of minimum. norm satisfying Ax = y
is gh'en by x = A*z where z is any solution of AA*z = y.
Proof If XI hi a solution of Ax = y, the general solution is x = XI + u
where U E %(A). Since %(A) is closed, it follows that there exists a unique

162

LINEAR OPERATORS AND ADJOINTS

vector x of minimum norm satisfying Ax = y and that this vector is orthogonal to %(A). Thus, since
is assumed closed,

[%(A)]l ==

Hence x = A*z for some z E II, and since Ax

AA*z=y. I

= y,

we conclude that

Note that if, as is frequently the case, the operator AA* is invertible,
the optimal solution takes the form
X

== A*(AA*)-ly.

Example 1. Suppose a linear dynamic system is governed by a set of

differential equations of the form
x(t) = Fx(t)

+ bu(t)

where x is an n x I vector of time functions, F is an n x n matrix, b is an

n X 1 vector, and u is a scalar control function.
Assume that x(o) = () and that it is desired to transfer the system to
x(T) = x I by application of suitable controL Of the class of controls which
accomplish the desired transfer, we seek \he one of minimum energy
Jl u2 (t) dt. The problem includes the motor problem discussed in Chapter 3.
The explicit solution to the equation of motion is
T

x(T)

= J eF(T-t)bu(t) dt.
o

Thus, defining the operator A : L2 [0, TJ

-+

En by

Au

= fo eF(T-t)bu(t) dt,

the problem is eqUivalent to that of determining the u of minimum norm

satisfying Au == Xl'
Since
is finite dimensional, it is closed. Thus the results of Theorem
1 apply and we write the optimal solution as

u = A*z
where

AA*z =

Xl'

It remains to calculate the operators A* and AA*. For any u E L

YE En

(y I AU)En = y'

fo eF(T-tlbu(t) dt = fo y' eF(T-t)bu(t) dt

== (A*Y I U)L2

2'

6.11

PSEUDOINVERSE OPERATORS

163

where
A*y = b'eF'(T-t)y.

Also, AA* is the n x n matrix,

AA* =

fo eF(T-t)bb'eF'(T-t) dt.

If the matrix AA* is inv,ertible, the optimal control can be found as

u = A*(AA*)-IX I .
6.11

Pseudoinverse Operators

We now develop a
general and more complete approach to the problem of finding approximate or minimum norm solutions to Ax = y. The
approach leads to the c:oncept of the pseudo inverse of an operator A.
Suppose again that G and H are Hilbert spaces and that A e B(G, H)
with E/9(A) closed. (In applications the Closure of E/9(A) is usually supplied
by the finite dimensionality of either G or H).

Definition. Among all vectors Xl e G satisfying

IIAxI - yll

= min IIAx - yll,

"

let Xo be the unique vector of minimum norm. The pseudoinverse A t of A

is the operator mapping y into its corresponding Xo as y varies over H.
Xo

To justify the above definition, it must be verified that there is a unique

corresponding to each y e H. We observe first that min IIAx - yll is

" in the closed

achieved since this amounts to approximating y by a vector
subspace E/9(A). The approximation y = AXI is unique, although Xl may
not be.
The set of vectors Xl satisfying Ax 1 = y is a linear variety, a translation
of the subspace .;V(A). Thus, since this variety is closed, it contains a
unique Xo of minimum norm. Thus At iS,well defined. We show below that
it is linear and bounded.
The above definitio!tl of the pseudo inverse is somewhat indirect and
algebraic. We can develop a geometric interpretation of At so that certain
of its properties become more transparent.
According to Theorem 1, Section 3.4, the space G can be expressed as
I

G = %(A) E6' %(A)J..

Likewise, since E/9(A) is assumed closed,

H = 9l!(A) $ :;P(Al.

164

LINEAR OPERATORS AND ADJOINTS

Figure 6.6 The pseudo inverse

The operator A restricted to ,AI'(A).L can be regarded as an operator from

the Hilbert space ,AI'(A).L onto the Hilbert space &l(A). Between these
spaces A is one-to-one and onto and hence has a linear inverse which,
according to the Banach inverse theorem, is bounded. This inverse
operator defines A t on &leA). Its domain is extended to all of H by defining
Aty = () for y E &l(A).L, Figure 6.6 shows this sChematically and Figure 6.7
gives a geometric illustration of the relation of the various vectors in the
problem.
It is easy to verify that this definition of A t is in agreement with that of
the last definition. Any y E H can be expressed uniquely as y == y + Yl
where ye &leA), Yl E &l(A).L, Thus y is the best approximation to y in
&leA). Then Aty == At(y + YI) = Aty. Define Xo = Aty. Then by definition
Axo = y. Furthermore, Xo E ,AI'(A).L and is therefore the minimum norm
solution of Ax 1 = y.
The pseudoinverse possesses a number of algebraic properties which are
generalizations of corresponding properties for inverses. These properties
are for the most part unimportant from our viewpoint of optimization;
therefore they are not proved here but simply stated below.

Figure 6.7 Relation between y and Xo

6.12

PROBLEMS

165

Proposition 1. Let A be a bounded linear operator with closed range and

let A t denote its pseua'oinverse. Then
1.
2.
3.

4.
5.
6.

7.
8.
9.

A t is linear.
A t is bounded.
(At)t = A.
(A*)t = (At)*.
AtAAt = At.
AAtA = A.
(AtA)*=AtA.
At = (A*A)tA*.
At = A*(AA*)t.

In certain limiting cases it is possible to give a simple explicit formula

for At. For instance, if A *A is invertible, then At = (A *A) - I A *. If AA * is
invertible, then At ==A*(AA*)-I. In general, however, a simple formula
does not exist.
Example 1. The pseudoinverse arises in connection with approximation
problems. Let {Xl> X2'" xn} be a set of vectors in a Hilbert space H.
In this example, however, these vectors are not assumed to be independent.
As usual, we seek the best approximation of the form y = Li"1 al XI to
the vector y. Or by dl:fining A: En -+ Hby
n

Aa = Laixj,
1= 1

we seek the approximation to Aa = y. If the vector a achieving the best

approximation is not unique, we then ask for the a of smallest norm which
givesy. Thus
ao

= Aty.

The computation of A t can be reduced by Proposition 1 to

At = (A*A)tA*
so that the problem reduces to computing the pseudoinverse of the n x n
Gram matrixA* A =: G(Xj' X2' ... , xn).
6.12

Problems

1. Let X = L 2 [0, 1] and define A on Xby

Ax

K(t. S)X(S) ds

166

LINEAR OPERATORS AND ADJOINTS

where
1

fo fo IK(t,

s)1 2 dt ds <

00.

Show that A : X ....... X and that A is bounded.

2. Let X = LlO, I], I < p < 00, Y =:: LlO, 1], lip

+ I/q = 1.

Define

Aby
1

Ax

= 50 K(t, s)x(s) ds.

Show that A E R(X, Y) if SA HIK(t, s)lq dt ds < 00.

3. Let A be a bounded linear operator from Co to 1(fJ' Show that corresponding to A there is an infinite matrix of scalars IXij' i, j =
1, 2, ... , such that y = Ax is expressed by the equations

where y = {Ili}' x =

g i}, and the norm of A is given by

00

IIAII

= sup
i

I: IIXul.
j= 1

4. Prove the two-norm theorem: If X is a Banach space when normed

by 11111 and by I liz and if there is a constant c such that Ilxlll S c Ilxllz
for all x E X, then there is a constant C such that Ilxll z s CIlxll l for
all x EX.
5. The graph of a transformation T: X ....... Y with domain D c X is the
set of points (x, Tx) E X X Y with xED. Show that a bounded linear
transformation with closed domain has a closed graph.
6. Let X = C [a, b] = Y and let D be the subspace of X consisting of all
continuously differentiable functions. Define the transformation T on
D by Tx = dx/dt. Show that the graph of Tis closed.
7. Prove the closed graph theorem: If X and Yare Banach spaces and T
is a linear operator from X to Y with closed domain and closed graph,
then Tis bounded.
8. Show that a linear transformation mapping one Banach space into
another is bounded if and only if its nulls pace is closed.
9. Let H be the Hilbert space of n-tuples with inner product (x IY)Q =
x' Qy where Q is a symmetric positive-definite matrix. Let an operator
A on H be defined by an n x n matrix [aijJ in the usual sense. Find
the matrix representation of A *.

6.12

10. Let X = Lp [0, 1], 1 < p <

A e B(X, Y) be d.efined. by Ax

00,

iF

167

+ llq =

1. Let

Y = Lq [0,1], lip

J& K(t, s)x(s) ds where

fo f IK(t, s)l dt ds <

1

PROBLEMS

00.

(See Problem 2.) Find A*.

11. Let X and Y be normed spaces and let A e B(X, Y). Show that
.L[Bl(A*)] = %(A). (See Section 5.7.)
12. Let X, Y be Banach spaces and let A e B(X, Y) have closed range.
Show that
inf

IIxll

Ax=b

13.
14.

15.

16.

max (b, y*).

IIAyII:sI

Use this result to reinterpret the solution of the rocket problem of

Example 3, Section 5.9.
Let X and Y be normed spaces and let G be the graph in X x Y of an
operator A E B(X, Y). Show that G.L is the graph of -A* in X* x Y*.
Prove the Minkowski-Farkas lemma: Let A be an m x n matrix and
b an n-dimensicmal vector. Then Ax:;:; (J implies b'x :;:; 0 if and only
ifb = A'A for s>ome m-dimensional vector A:2: (J. Give a geometric
interpretation of this result. (Ineq.,alities among vectors are to be
interpreted comjponentwise.)
Let M be a closed subspace of a Hilbert space H. The operator P
defined by Px == m, where x = m + n is the unique representation of
x E H with m e M, n e M.L, is called the projection operator onto M.
Show that a projection operator is linear and bounded with IIPII = 1
if M is at least OIt1e dimensional.
Show that a bounded linear operator on a Hilbert space H is a projection operator if and only if:

1. pZ = P (idempotent)
p* = P (self-adjoint).

2.

17. Two projection operators PI and P z on a Hilbert space are said to be

orthogonal if PIPZ = O. Show that two projection operators are
orthogonal if and only if their ranges are orthogonal.
18. Show that the sum of two projection operators is a projection operator
if and only if
are orthogonal.
19. Let G and H be Hilbert spaces and suppose A E B(G, H) with Bl(A)
closed. Show that

At

lim (A* A
1:-+0+

+ 81)-1 A* =

lim A*(AA*
-+0+

where the limits represent convergence in B(H, G).

+ 81)-1

168

LINEAR OPERATORS AND ADJOINTS

20. Find the pseudoinverse of the operator A on E 3 defined by the matrix

[i i bj.
J

1 0

21. Let G, H, K be Hilbert spaces and let B E B(G, K) with range equal to
K and C E B(K, H) with nulls pace equal to {O} (i.e., B is onto and C
is one-to-one). Then for A = C B we have At = Bt ct.

REFERENCES
6.4. The Banach inverse theorem is intimately related to the closed graph
theorem and the open mapping theorem. For further developments and applications, see Goffman and Pedrick [59].
6.7. The concept of conjugate cone is related to various alternative concepts
including dual cones, polar
and polar sets. See Edwards [46], Hurwicz
[75J, Fenchel [52), or Kelley, Namioka, et al. [85].
6.1O. See Balakrishnan [14].
6.11. The concept of a pseudo inverse (or generalized inverse) was originally
introduced for matrices by Moore [105], [l06] and Penrose [117], [118J.
See also Greville [67]. For an interesting introduction to the subject, see
Albert [5]. Pseudoinverses on Hilbert space have been discussed by Tseng
[1461 and Ben Israel and Charnes [20]. Our treatment closely parallels that of
Desoer and Whalen [38].

7
OPTIMIZATION
OF FUNCTIONALS

7.1

Introduction

In previous chapters, while developing the elements of functional analysis,

we often considered minimum norm optimization problems. Although the
availability of a large variety of different norms provides enough flexibility
for minimum norm problems to be of importance, there are many important optimization problems that cannot be formulated in these terms. In
this chapter we consider optimization of more general objective functionals. However, much of the theory and geometric insight gained while
studying minimum norm problems are of direct benefit in considering
these more general problems.
Our study is guided by two basic geometric representations of nonlinear
functionals. Each of these representations has its own particular advantages,
and often it is enlightening to view a given situation in both ways. The
first, and perhaps most obvious, geometric representation of a nonlinear
functional is in terms of its graph. Suppose f is a functional defined on a
subset D of the vector space X. The space
is then imbedded in the
product space R x X where R is the real line. Elements of this space consist
of ordered pairs (r, x). The graph off is the surface in R x X consisting
of the points (f(x), x) with XED. Usually the R axis (Le., points of the
form (r, 0)) is regard,ed as the vertical axis, and the value of the functional
at x is then regarded as the vertical distance of the graph above the point
x E X. In this representation a typical functional is visualized as in Figure
7.1. This representation is certainly familiar for one- or two-dimensional
space. The point her,e is that it is convenient conceptually even in infinitedimensional space.
The second representation is an extension of thl; technique of represe'nting a linear functional by a hyperplane. A functional is described by
its contours in the space X. A typical representation is shown in Figure 7.2.
Iff is sufficiently smooth, it is natural to construct hyperplanes tangent to
the contours and to define the gradient of f Again, this technique is

170

OPTIMIZATION OF FUNCTIONALS

Figure 7.1

The graph of a functional

\

familiar in finite-dimensional space and is easily ,extended to infinitedimensional space. The principal advantage of the second representation
over the first is that for a given dimensionality of X, one less dimension
is required for the representation.
The first half of this chapter deals with generalizations of the concepts of
differentials, gradients, etc., to normed spaces and is essentially based on
the second representation stated above. The detailed development is largely
algebraic and manipulative in nature, although certain geometric interpretations are apparent. From this apparatus we obtain the local or variational theory of optimization paralleling the familiar theory in finite
dimensions. The most elementary portion of the classical calculus of
variations is used as a principal example of these methods.
The second half of the chapter deals with convex and concave funetionals from which we obtain a global theory of optimization. This

f(x) = c

Figure 7.2

The contours of a functional

7.2

GATEAUX AND FRECHET DIFFERENTIALS

171

development, based essentially on the first representation for functionals,

is largely geometric and builds on the theory of convex sets considered in
Chapter 5. The interesting theory of conjugate functionals produces
another duality theorem for a class of optimization problems.
LOCAL THEORY

7.2

Gateaux and Fll'echet Differentials

In the following discussion let X be a vector space, Ya normed space, and

T a (possibly nonlinear) transformation defined on a domain D c X and
having range R c Y.
Definition. Let xED c X and let h be arbitrary in X. If the limit_

(1)

c5T(x; h) = lim! [T(x

"'0 rx

+ rxh) -

T(x)]

exists, it is called the Gateaux differential of T at x with increment h. If

the limit (1) exists for each hEX, the transformation T is said to be
Gateaux differentiable at x.
We note that it makes sense to consider the limit (1) only if x + rxh E Dfor
all 0( sufficiently small. The limit (1) is, of course, taken in the usual sense .
of norm convergenc:e in Y. We observe that for fixed XED and h reg/l.rded
as variable, the Gateaux differential defines a transformation from X to
Y. In the particular case where T is linear, we have oT(x; h) = T(h).
By far the most frequent application of this definition is in the case
where Y is the real line and hence the transformation reduces to a (realvalued) functional on X. Thus if f is a functional on X, the Gateaux
differential of J, if it exists, is
8f(x; h) = dd f(x
01.

and, for each fixed x

variable hEX.

+ O(h)\

=0

'

X, of(x; h) is a functional with respect to the

Example 1. Let X =En and letf(x) = f(x l , X2, . , x h ) be a functional on

En having continuous partial derivatives with respect to each variable xi.
Then the Gateaux differential off is
I
c5f(x; h) =

of
L -;hi.
n

1=1 vXi

172

OPTIMIZATION OF FUNCTIONALS

Example 2. Let X = C[O, 1] and Jet I(x) = fA g(x(t), t) dt, where it is

assumed that the function Ux exists and is continuous with respect to x
and t. Then

(jj(x; h)

= :rx (g(X(t) + rxh(t), t) dtla=o

Interchange of the order of differentiation and integration is permissible

under our assumptions on U and hence

fo uix, t)h(t) dt.

1

()f(x; h) =

Example 3. If X = En, Y = Em, and T is a continuously differentiable

mapping of X into Y, then bT(x; h) exists. It is the vector in Yequal to
the vector It E X multiplied by the matrix made up of partial derivatives
of Tat x.
'
The Gateaux differential generalizes the concept of directional derivative
familiar in finite-dimensional space. The existence of the Gateaux differential is a rather weak requirement, however, since its definition requires
no norm on X; hence, properties of the Gateaux differential are not easily
related to continuity. When X is normed, a more satisfactory definition is
given by the Frechet differential.
Definition. Let T be a transformation defined on an open domain D in a
normed space X and having range in a normed space Y. If for fixed XED
and each hEX there exists ()T(x; h) E Y which is linear and continuous
with respect to h such that

lim IIT(x

+ h) -

T(x) - bT(x;

Ilh I

IIhll->O

h)11 _ 0
- ,

then T is said to be Frechet differentiable at x and ()T(x; h) is said to be the

Frechet differential of T at x with increment h.
We use the same symbol for the Frechet and Gateaux differentials since
generally it is apparent from the context which is meant.
Proposition 1. If the transformation T has a Frechet differential, it is unique.

Proof Suppose both bT(x; /1) and b'T(x; h) satisfy the requirements
of the last definition. Then

IIbT(x; It) - o'T(x; 11)11

IIT(x

+ 11) -

T(x) - oT(x; h) I

+ IIT(x + h) - T(x) - o'T(x; h)11

or II{)T(x; h) - ()'T(x; h)11 = o(llhll). Since ()T(x; h) - ()'T(x; h) is bounded
and linear in h, it must be zero. I

7.2

GATEAUX AND FREcHET DIFFERENTIALS

173

Proposition 2. If the Frechet differential of T exists at x, then the Gateaux

differential exists at x and they are equal.
the Frechet differential by t5T(x; h). By definition we
tor any h,
1

-IIT(x + exh) - T(x) - t5T(x; exh)ll ...... 0,

ex

as ex ...... 0.

Thm, by the linearity of t5T(x; exh) with respect to ex,

lim T(x

+ exh) -

T(x)

= t5T(x; h). I

ex

... 0

A:!inal property is given by the following proposition.

Prop'Jsition 3. If the transformation T defined on an open set D in X has a

Frecltet differential at x, then T is continuous at x .

.Proof. Given s > 0, there is a sphere about x such that for x

+ h) -

+ h in

< sllhll.
Thus IIT(x + h) - T(x) II <sllhll + IIt5T(x; h)1I < Mllhll I
IIT(x

T(x) - t5T(x; h) II

W! are primarily concerned with Frechet differentials rather than

Gateaux differentials and often assume their existence or, equivalently,
aSSU:iYle the satisfaction of condition, which are sufficient to imply their
existence.

Example 4. We show that the differential

t5f(x; h) =

/I

of

L -;:- hi

.=1 vX.

in Example 1 (of the functionalf on En) is a Frechet differential. Obviously,

t5f(x;h) is linear and continuous in h. It is therefore only necessary to
verify the basic limit property required of Frechet differentials.
Given s > 0, the continuity of the partial derivatives implies that there
is a neighborhood S(x; 8) such that

of(x) _ of(y) <

ox.

oXi

foce all ye S(x; t5) and i = 1,2, ... , n.

Define the unit vectors in the usual way,
a:od for h = Li'=l hie. define go = e and

e.

for k

e; = (0,0, ... , 1,0, ... ,0),

= 1,2, ... , n.

174

OPTIMIZATION OF FUNCTIONALS

We note that Ilok"

f(X

+ h) -

"hll

f(x) -

for all k. Then

:1 It

1=1 uXI

hil =

k=1

+ gk) -

{f(X

f(x

+ gk-I) -

:f hk}1

uX k

Now we examine the k-th term in the above summation. 1;'he vector x + gk
differs from x + gk-l only in the k-th component. In fact, x + gk = X +
gk-t + hk ek . Thus, by the mean value theorem for functions of a single
variable,
\

for some (x, 0 (X hk

Also, x + gk-t + (Xek E Sex; c5) if Ilhll < c5. Thus

jJ(X

+ gk) -

+ gk-t) -

f(x

hk <

Ilhll.

Finally, it follows that

f(x

+ h) -

f(x) -

of

OXk hi < ellhll

for all h, Ilhll < c5.

Example 5. We show that the differential

c5f(x; h)

= fotgxCX, t)h(t) dt

in Example 2 is a Frechet differential. We have

If(x

+ h) -

f(x) - c5f(x; h)1

= I( {g(x + h, t) -

g(x, t) - gx(x, t)h(t)} dt ,.

For a fixed t we have, by the one-dimensional mean value theorem,

g(x(t)

+ h(t), t) -

g(x(t), t) = gx(x(t), t)h(t)

where Ix(t) - x(t) I Ih(t)l. Given e > 0, the uniform continuity of g" in x
and t implies that there is a /j > 0 such that for IIhl! < c5, Igx<x + h, t) g,,(x, t)1 < e. Therefore, we have

1/(.... + J.) -1(;0) - o/(x, h)(

forllhll < c5. The result follows.

I(fl1.(X,

r) -

uix, l);h(C) dtl

ellhll

7.3

FRECHET DERIVATIVES

175

Example Let X = CI[O, I], the space of continuous n-vector functions

on [0, 1]. tet Y = Cm[O, 1] and define T: X -+ Y by
T(x)

or] dor

where F = (fl,f2, ... ,fm) has continuous partial derivatives with respect
to its arguments. The Gateaux differential of T is easily seen to be

By combining the analyses of Examples 4 and 5, we can conclude that this

is actually a Frechet dilfferential. Note also that since the partial derivatives
of F are continuous, c5T(x; h) is continuous in the variable x.

7.3 Frechet Derivatives

Suppose that the tramformation T defined on. an open domain D c X is
Frechet differentiable throughout D. At a fixed point XED the Frechet
differential c5T(x; h) is then, by definition, of the form c5T(x; h) = Axh,
where Ax is a
linear operator from X to Y. Thus, as x varies
over D, the correspondence x -+ Ax defines a transformation from D into
the normed linear space B (X, Y); this transformation is called the Frechet
derivative T' of T. Thus we have, by definition, c5T(x; h) = T'(x)h.
If the
x -+ T'(x) is continuous at the point Xo (i.e., if given
B> there is c5 > such that Ilx - xoll < c5 implies I\T'(x) - T'(xo)1I < B),
we say that the Frechet derivative of T is continuous at xo. This should
not be confused with the statement that T'(xo) is a continuous mapping
from X to Y, a property that is basic to the definition of the Frechet
derivative. If the derivative of T is continuous on some open sphere
S, we say that T is continuously Frechet differentiable on S. All of the
examples of Frechet differentiable transformations in the last section are
in fact continuously Frechet differentiable.
In the special case where the original transformation is simply a functional f on the space X, we have c5f(x; h) =f'(x)h where f'(x) E X* for
each x. The elementf'(x) is called the gradient offat x and is sometimes
denoted Vf(x) rather thanf'(x). We sometimes write (h,f'(x for c5f(x; h)
since f'(x) E X*, but usually we prefer f'(x)h which is consistent with the
notation for differentials of arbitrary transformations.
Much of the theory of ordinary derivatives can be generalized to Frechet
derivatives. For instance, the implicit function theorem and Taylor series
have very satisfactory extensions. The interested reader should consult the

176

OPTIMIZATION OF FUNCTIONALS

references cited at the end of the chapter. In this section we discuss the
elementary properties of Frechet derivatives used in later sections.
It follows immediately from the definition that if T( and T2 are Fnechet
differentiable at xED, then rJ.( T( + CX2 T2 is Frechet differentiable at x and
(a(T( + CX2 T 2 )'(x) = cx(T('(x) + CX 2 T 2 '{x). We next show that the chain rule
applies to Frechet derivatives.

Proposition 1. Let S be a transformation mapping an open set D c: X into

an open set E c:: Y and let P be a transformation mapping E into a normed
space Z. Put T = PS and suppose S i} Frechet differentiable at XED and P
is Frechet differentiable at y = Sex) E E. Then Tis Frechet differentiable
at x and T'(x) = P'(y)S'(x).
Proof For hEX, X + h E D, we have
T(x + h) - T(x) = P[S(x + h)] - P[S(x)] = P(y + g) - P(y)
where g = Sex + h) - Sex). Thus IIT(x
Since, however,

Ilg - S'(x)hll

+ h) - T(x) - P'(y)gll

o(llgll).

o(llhll),

we obtain

II T(x + II) - T(x) ,- P'(y)S'(x)hll

= o(IIhll) + o(lIgll).

Since, according to Proposition 3 of Section 7.2, S is continuous at x, we

conclude that IIgll = O(lIhll) and hence

T'(x)h

= P'(y)S'(x)h. I

We now give a very useful inequality that replaces the mean value
theorem for ordinary functions.

Proposition 2. Let T be Frechet differentiable on an open domain D. Let

xED and suppose that x
IIT(x

+ cxl! E D for all cx, 0 :S cx :S

+ h) - T(x)"

:S

IIhll

sup 1IT'(x
0<<<< 1

I. Then

+ cxh)lI.

Proof Let y* be a nonzero element of y* aligned with the element

T(x + h) - T(x). The function cp(cx) = y*[T(x + ah)] is defined on the
interval [0, \] and, by the chain rule, has derivative
cp'(rx) = y*[T'(x

+ rxh)h].

By the mean value theorem for functions of a real variable, we have

0< CXo < 1,

7.4

EXTREMA

177

and hence

ly*[T(x

+ h) - T(x)] I ::;; IIy"1I

and since y* is aligned with T(x

sup IIT'(x

0<,,< 1

+ cth)II "hil,

+ h) - T(x),

IIT(x -I- h) - T(x)11 :::; Ilhll sup IIT'(x

0<,,<1

+ cth)ll I

If T: X -+ Y is F'rechet differentiable on an open domain D c: X, the

derivative T' maps D into B(X, Y) and may itself be Frechet differentiable
on a subset Dl c: D. In this case the Frechet derivative of T' is called the
second Frechet derivative of T and is denoted by T".

Example 1. Let f be a functional on X = En having continuous partial

derivatives up to second order. Then !"(xo) is an operator from En to En
having matrix form

f"(xo) =

[02f (X)] ,
OXl ax x=xo
j

where Xi is the i-th component of x.

The following inequality can be proved in a manner paralleling that of
Proposition 2.
Proposition 3. Let T be twice Frechet differentiable on an open domain D.
Let xED and suppose that x + cth E D for all ct, 0 ::;; ct :::; 1. Then

IIT(x

+ h) - T(x) - T'(x)hll :::; illhll2

sup IIT"(x

0<,,<1

+ cth)ll.

7.4 Extrema
It is relatively simple to apply the concepts of Gateaux and Frechet
differentials to the problem of minimizing or maximizing a functional on a
linear space. The technique leads quite naturally to the rudiments of the
calculus of variatio11s where, in fact, the abstract concept of differentials
originated. In this section, we extend the familiar technique of minimizing
a function of a single variable by ordinary calculus to a similar technique
based on more general differentials. In this way we obtain analogs of the
classical necessary conditions for local extrema and, in a later section, the
Lagrange technique for constrained extrema.
I

Definition. Let f be a real-valued functional defined on a subset n of a

normed space X. A point Xo E is said to be a relative minimum off on
if there is an open sphere N containing Xo such th:..tf(xo) ::;;f(x) for all
x E r. N. The point Xo is said to be a strict relative minimum off on

178

OPTIMIZATION OF FUNCTIONALS

if f(xo) <f(x) for all x xo, x e

similarly.

n Ii N.

Relative maxima are defined

We use the term extremum to refer to either a maximum or a minimum

over any set. A relative extremum (over a subset of a normed space) is
also referred to as a local extremum. The set n on which an extremum
problem is defined is sometimes calIed the admissible set.

Theorem 1. Let the real-valued functional f have a Gateaux differential on

a vector space X. A necessary condition for 'y to have an extremum at Xo e X
is that lJf(xo; h) = 0 for all heX.
Proof For every heX, the functionf(xo + ('J.h) of the real variable ('J.
must achieve an extremum at ('J. = O. Thus, by the ordinary calculus,
d f(xo
d('J.

+ ('J.h) I

,,=0

= O.

A point at which of(x; h) = 0 for all h is called a stationary point;

hence, the above theorem merely states that extrema occur at stationary
points. It should be noted that a similar result holds for a local extremum
of a functional defined on an open subset of a normed space, the proof
being identical for both cases.
The simplicity of Theorem 1 can be misleading for it is a result of great
utility. Much of the calculus of variations can be regarded as a simple
consequence of this one result. Indeed, many interesting problems are
solved by careful identification of an appropriate vector space X and
some algebraic manipulations to obtain the differential. There are a
number of useful generalizations of Theorem 1. We offer one of the
simplest.

Theorem 2. Let f be a real-valued functional defined on a vector space X.

Suppose that Xo minimizes f on the convex set n c: X and that f is Gateaux
differentiable at Xo. Then
for all x e n.
Proof Since n is convex, Xo
by ordinary calculus,

+ ('J.(x -

:('J. f(xo +()!(x for a minimum at xo'

xo) e

n for 0 S
0

('J. S 1 and hence,

7.S

179

EULER-LAGRANGE EQUATIONS

*7.5 Euler-Lagrange Equations

'-The classical problem in the calculus of variations is that of finding a
function x on the interval [t l' t 2J minimizing an integral functional of
the form

J =

'2

I f[x(t), x(t), tJ dt.

1\

To specify the problem completely, we must agree on the class of

functions within which we seek the extremum-the so-called admissible
set. We assume that the function.f is continuous in x, X, and t and has
continuous part'ial derivatives with respect to x and X. We seek a solution
in the space D[t\l t2l In the simplest version of the problem, we assume
that the end points x(t 1) and x(t 2) are fixed. This further restricts the
admissible set.
Starting with a given admissible vector x, we consider vectors of the
form x + h that are admissible. The class of such vectors h is called the
class of admissible variations. In the case of fixed end points, it is clear
that the class of admissible variations is the subspace of D[t 1 , t 2J, consisting of functions which vanish at tl and t 2 . The necessary condition
for the extremum problem is that for all such h, JJ(x; h) = o.
The differential of J is

M(x; h)

d
= -d
I f(x + IXh, x + IXh, t) dt I
12

IX I.

,,=0

or, equivalently,

(1)

M(x; h)

= tfx(x, x, t)hU) dt + tfix, x, t)h(t) dt,

I.

"

and it is easily verified that this differential is actually Prechet. Equating

this differential to zero and assuming that the function fx has a continuous derivative with respect to t when the optimal solution is substituted for x, we may integrate by parts to obtain

M(x; h)

= f12 [fix, x, t) - dd
1\

fix,

x, t)Jh(t) dt + fix, x, t)h(t) \'2 =

O.

1\

The boundary terms vanish for admissible h and thus the necessary
condition is
I

([fiX, x, t) - :t fix, x, t)]h(t) dt = 0

for all h E D[t 1, t 2J vanishing at

tl

and t 2

180

OPTIMIZATION OF FUNCTIONALS

Since the term multiplying h(t) in the integrand is continuous, it readily

follows (see Lemma I below) that it must vanish identically on [tt, t 2 ].
Thus we conclude that the extremal x must satisfy the Euler-Lagrange
equation
(2)

fix,

x, t) -

:t fix,

x, t) = o.

The above derivation of the Euler-Lagrange equations suffers from the

weakness of the assumption that

is continuous at the optimal solution. Actually we have no basis on which

to assume that the solution is in fact smooth enough for this assumption
to hold. The alternate derivation, given after the following three lemmas,
avoids this drawback.

Lemma 1. If aCt) is continuous on [t l , t2 ] and S:ia(t)h(t) dt = 0 for every

he D[tt, t 2 ] with h(tl) = h(t 2 ) = 0, then aCt) == 0 on [t1> t 2 ].
Proof Assume that aCt) is nOQzero, say positive, for some t
Then aCt) is positive on some interval [tt', t 2 '] c [tt, t 2 ]. Let

[tl> t 2 ].

tE[tt',t 2 ']
otherwise.
The function h satisfies the hypotheses of the lemma and

12

a(t)h(t) dt> O.

11

Lemma 2. If aCt) is continuous in [tl> t2 ] and

dt = 0 for every
he D[/ t , 12 ] with hCtt) = h(t 2 ) = 0, then aCt) == c in [tl> t 2 ] where c is a
constant.
Proof Let c be the unique constant satisfying
and let
h(t) =
Then
'2

[aCt) -

C]2

II

dt =

f" [ae-r) -

c] d1:.

I [aCt) - c)]h(t) dt
12

II

=f'Ci(t)h(t) dt 11

and hence aCt) == c.

[aCt) - c] dt = 0

c[h(t 2 )

h(tt)] = 0,

7.5

EULER-LAGRANGE EQUATIONS

181

Lemma 3. If aCt) and P( t) are continuous in [t 1> t 2] and

\

[!X(t)h(t)

(3)

tl

+ (J(t)h(t)] dt = 0

for every hE D[tlo t 2 ] with h(tt) = h(t 2 ) == 0, then (J is differentiable and

fi(t) =:: O(t) in [t 10 t21Proof Define

ACt) =

aCt) dt.

ft

Then by integration by parts we have

tl

!X(t)h(t) dt

tl

=-

Itl A(t)h(t) dt.

tl

Therefore (3) becomes

t(

-A(t)

+ fi(t)Jh(t) dt =

tl

which by Lemma 2 implies

(J(t)

= A(t) + c

for some constant c. Hence, by the definition of A,

= O(t). I

Now, in view of Lemma 3, it is clear that the Euler-Lagrange equation

(2) follows directly from equation (1) without an a priori assumption of
the differentiability of It .

Example 1. (Minimum Arc Length) Given t 1 , t,. and x(tt), x(t 2 ), let us
employ the Euler-Lagrange equations to determine the curve in D[tt, t2 ]
connecting these points with minimum arc length. We thus seek to minimize

From (2) we obtain immediately

dtox J 1 + (X)2 = 0
or, equivalently,

x = const.
Thus the extremizing arc is the straight line connecting the two points.

182

OPTIMIZA nON OF FUNCTIONALS

Example 2. (Estate Planning) What is the lifetime plan of investment and

expenditure that maximizes total enjoyment for a man having a fixed
quantity of savings S? We assume that the man has no income other
than that obtained through his investment. His rate of enjoyment (or
utility) at a given time is a certain function U of r, his rate of expenditure.
Thus, we assume that it is desired to maximize

where the e- P1 term reflects the notion

less today.
If x(t) is the total capital at time t, then

future enjoyment is counted

x(t) = ax(t) - ret)

where

(X

is the interest rate of investment. Thus the problem is to maximize

T

fo e-P1V[ax(t) -

x(t)] dt

subject to x(O) = Sand x(T) = 0 (or some other fixed value). Of course,
there is the additional constraint x(t) ;;::: 0, but, in the cases we consider,
this turns out to be satisfied automatically.
From the EulerLagrange equation (2), we obtain

ae-P1U'[ax(t) - x(t)]

+ dt e-P1U'[ax(t) -

x(t)] = 0

where U' is the derivative of the utility function U. This becomes

(4)

:t U'[cxx(t) - x(t)]

= (P -

cx)U'[cxx(t) - x(t)].

Hence, integrating (4), we find that

(5)

U'[r(t)] = U'[r(O)]e(P-a)l.

Hence, the form of the time dependence of r can be obtained explicitly

once U is specified.
A simple utility function, which reflects both the intuitive notions of
diminishing marginal enjoyment (i.e., V' decreasing) and infinite marginal
enjoyment at zero expenditure (i.e., U'[O] =(0), is U[r] == 2r1/2. Substituting this in (5), we obtain

cxx(t) - x(t)

= ret) =

r(0)e 2 (a- P)I.

7.6

PROBLEMS WITH VARIABLE END POINTS

183

Integrating this last equation, we have

x(t)

= etZtx(O) +

a - 2P

[eal _ e2 (a- P)I]

(6)
= {X(O) _

r(O) --'ea.

af

2P -

e2(a-P)I.

2P - a

Assuming a> fJ > a/2, we may find r(O) from (6) by requiring x(T) = O.
Thus

(2P - a)x(O)
-e -(2p-a)T'

r(O) = 1

The total capital grows initially and then decreases to zero.

*7.6 Problems with Variable End Points
The class of admissible functions for a given problem is not always a
linear variety of functions nor even a convex set. Such problems may,
however, generally be approached in essentially the same manner as
before. Basically the technique is to define a continuously differentiable
one-parameter family of admissible curves that includes the optimal curve
as a member. A necessary condition is obtained by equating the derivative
of the objective functional, with respect to the parameter, equal to zero.
In the case
the admissible class is a linear variety, we consider the
family x + 8h and differentiate with respect to 8. In other cases, a more
general family X(8) of admissible curves is considered.
An interestil1lg class of problems that can be handled in this way is
calculus of variations problems having variable end points. Specifically,
suppose we seek an extremum of the functional

J f(x, ie, t)
12

J =

dt

1\

where the interval [t1' t2], as well as the function x(t), must be chosen.
In a typical problem the two end points are constrained to lie on fixed
curves in the x - t plane. We assume here that the left end point is fixed
and allow the right end point to vary along a curve S described by the
function x =g(t), as illustrated in Figure 7.3.
Suppose X(8, t) is a one-parameter family of Ifunctions emanating from
the point 1 and terminating on the curve S. The termination point is
X(8, t2(8 and satisfies X(8, t 2(8 = g(t2(8). We assume that the family is
defined for 8 e [-a, a] for some a> 0 and that the desired extremal is the

184

OPTIMIZATION OF FUNCTIONALS

7
glt)

s
Figure 7.3

A variable end-point problem

curve corresponding to e = O. The variation oj of the functional J is defined

as the first derivative of J with respect to e and must vanish at e =.0.
Defining, in the obvious way,

ox(t)

= dd

x(e, t)/
8=0

we have

J(e) =
OJ

J'2) f[x(t, e), x(t, e), t] dt

'1

= f[x(t z), x(t z), tZJ01 2 + {2{f,,(X, x, t)ox + fix, x, t)ox} dt = O.

'1

Arguments similar to those of the last section lead directly to the

necessary conditions
d
(1)
f,,(x, x, t) = dt fix, x, t)
and

(2)

f[x(t z ), x(t z ), tz]ot z + f,,[x(t z ), x(t z ), tz]Dx(t z) = O.

Condition (1) is again the Euler-Lagrange equation, but in addition we

have the transversality condition (2) which must be employed to determine the termination point. The transversality condition (2) must hold for
all admissible ox(tz) and 0/ 2 , These two quantities are not independent,
however, since
(3)
Upon differentiating (3)

with respect to

DX(t 2 )

B.

we obtain the relation

+ X(tz)Dt 2 == g(t 2 )Dt Z

7.7

PROBLEMS WITH CONSTRAINTS

185

and hence the complete set of necessary conditions is (1) and the transversality condition

(4)

{f(x,

x, t) + [g -

x]f,.{x,

x, t)}lt=tz = o.

Example 1. As an I!xample of the transversality condition, we consider the

simple problem of finding the differentiable curve of minimum
length
from the origin (0,0) to the curve g. Thus we seek to extremize the
functional

t Jt +
h

J=

(x)2 dt,

where t2 is the point of intersection of the line.

As in Example I of the last section, the Euler-Lagrange equation leads
to x = const and, hence, the extremal must be a straight line. The transversality condition in this case is

or
.
1
X(t2) = - -.- .
g(t 2 )

Thus the extremal arc must be orthogonal to the tangent of gat t 2 (See
Figure 7.4.)
:<

" ...

""

"

Figure 7.4 Minimum distance to a curve

7.7

Problems with Constraints

In most optimization problems the optimal vector is required to satisfy

some type of constraint. In the simplest calculus of variations problem,
for instance, the end points of the curve are constrained to fixed points.

186

OPTIMIZATION OF FUNCTIONALS

Such problems, and those of greater complexity, often can be resolved

(in the sense of establishing necessary conditions) by considering a oneparameter family of vectors satisfying the constraints which contains the
optimal vector as a member. A complication arises, however, when the
constraint set is defined implicitly in terms of a set of functional equations
rather than explicitly as a constraint surface in the space. In such cases
the one-parameter family must be constructed implicitly.,In this section
we carry out the necessary construction for a finite number of functional
constraints and arrive at our first Lagrange multiplier theorem.
A more general discussion of constrained optimization problems, including the material of this section as a special case, is given in Chapter 9.
The later discussion parallels, but is much deeper than, the one given here.
The problem investigated in this section is that of optimizing a functional
f subject to n nonlinear constraints given in the implicit form:
gl(X) = 0

gix)

=0

(1)

These n equations define a region n in the space X within which the

optimal vector Xo is constrained to lie. Throughout this section it is
assumed that all functionalsf, OJ are continuous and Frechet differentiable
on the normed space X.
Before embarking on the general resolution of the problem, let us
briefly consider the geometry of the problem. The case where X is two
dimensional is depicted in Figure 7.5. If Xo is optimal, the functional f
has an extremum at Xo with respect to small displacements along n.
Under sufficient smoothness conditions, it seems that f has an extremum
at Xo with respect to small displacements along T, the tangent to n at Xo .

T
g(x) = 0

Figure 7.5 Constrained extremum

7.7

PROBLEMS WITH CONSTRAINTS

187

The utility of this observation is that the exact form of the surface n near
Xo is replaced by the simple description of its tangent in the expression for
the necessary conditions. In order for the procedure to work, however, it
must be possible to express the tangent in terms of the derivatives of the
constraints. For this reason we introduce the definition of a regular point.

Definition. A point Xo satisfying the constraints (1) is said to be a regular

point of these constraints if the n linear functionals gl'(XO), g2'(XO), ... ,
g"'(xo) are linearly independent.
The following theorem gives the necessary conditions satisfied by the
solution of the constrained extremum problem.
Theorem 1. If Xo is an extremum of the functional f subject to the constraints
g/(x) = 0, i = 1,2" ... , n; and if Xo is a regular point of these constraints,
then
(jf(xo; h) = 0

= 0, i = 1,2, ... , n.
Choose heX such that (jgj(xo; h) = 0 for i =

for all h satisfyin!), (jg/(xo; h)

Proof
1,2, ... , n. Let
YI, Y2, ... , Y" e X be 11 linearly independent vectors chosen so that the
n x n matrix

M=

(jgl(XO; YI) (jgl(XO; Y2) ... Ogl(XO; Y")]

c5 g2 (Xr YI) c5g 2(xo; Y2) ... 15g2(xo; Y")
[
.(jg"(xo; Y I)

(jg"(xo; Y")

is nonsingular. The existence of such a set of vectors Yi follows directly

from the regularity of the point Xo (see Problem 11).
We now introduce the n + 1 real variables e, CPI' CP2, ... , cp" and consider
the n equations

+ eh + CP1Yl + CP2Y2 + ... + cp"y") = 0

g2(X O+ eh + CPIYl + CP2 Y2 + ... + cp" Y") = 0
gl(XO

(2)

g"(xo

+ eh + CP1Yt + CP2Y2 + ... + cp"y") = O.

The lacobianofthis set with respectto the variables CPj, at e = 0 and CPi == 0,
is just the determi.nant of M and is therefore nonzel;O by assumption. Hence,
the implicit function theorem (see, for example, Apostol [1 OJ and also Section

9.2) applies and

the existence of n functions CPi(e) satisfying (2)
and defined in some neighborhood of e = O.

188

OPTIMIZATION OF FUNCTIONALS

l:?=

Denote by y(e) the vector

1 ep/(e)Yi and by ep(e) the n-dimensional
vector having components epiCS). For each i we have

(3)

0 = g/(xo
= g/(xo)

+ eh +

epjYj)

+ sog/(xo;

h)

+ ogj(xo ; y(e + o(e) + o[lly(e)llJ.

Or, writing all n equations simultaneously and taking into account the
fact that the first two terms on the right side of (3) are zero, we have,
after taking the norm of the result,

0= II Mep(e) II

(4)

+ o(e) + o[lly(e)llJ.

Since M is nonsingular, there are constants Cl > 0, C2 > 0 such that

S; II Mep(e) II :;; c21Iep(e) II; and since
y/s are linearly independent,
there are constants d , > 0, d2 > 0 such that dtlly(e) I S; Ilep(e)II S; d21Iy(e)ll.
Hence, c1d 1 1Iy(e)1I S; IIMep(e)1I S; c2d21Iy(e)11 and, therefore (4) implies
Ily(e)11 = o(e). Geometrically, this result states that if one moves along the
tangent plane of the constraint surface n from Xo to Xo + eh, it is possible
to get back to n by moving to Xo + eh + y(e), where y(e) is small compared
with sh.
The points Xo + eh + y(e) define a one-parameter family of admissible
vectors. Considering the functional f at these points, we must have

clllep(e)1I

d f[xo
dS

+ eh + yes)]

Thus, since Ily(e)11 = o(e), of(xo; h)

8=0

= o.

= o. I

From Theorem 1 it is easy to derive a finite-dimensional version of the

Lagrange multiplier rule by using the following lemma.

Lemma 1. Let fo ,fl' ... ,/" be linear functionals on a vector space X and
suppose thatj'o(x) = Ofor every x E X satisfyingfl(x) =fJfor i = 1,2, ... , n.
Then there are constants ..1.1' . 1. 2 , .. , An such that
fo

+ Adl + ..1.212 + . " + Anfn = O.

Proof. See Problem 16, Chapter 5 (which is solved by use of the

Hahn-Banach theorem). I

Theorem 2. If Xo is an extremum of the functional f subject to the constraints

g/(x) = 0,

i = 1,2, ... , n,

7.7

PROBLEMS WITH CONSTRAINTS

189

and Xo is a regular point of these constraints, then there are n scalal

AI> i = 1, 2, ... , n, that render the functional

f(x)
stationary at

II

L A/ub)

/=1

Xo

Proof. By Theorem 1 the differential (jf(xo; h) is zero whenever ea

of the differentials (jg/(xo; h) is zero. The result then follows immediately
from Lemma 1. I
Example 1. Constrained problems of the form above are called is ope
metric problems in the calculus of variations since they were originally
studied in connection with finding curves of given perimeter which rna: .
mized some objective such as enclosed area. Here we seek the curve in t
t -x plane having end points (-1,0), (1,0); length I; and enclosing
maximum area between itself and the t-axis. Thus we wish to maximize 1

subject to
1

f JX2 + 1 dt

-1

1.

Therefore, we seek a stationary point of

= f-1 (x + AJ x2 + 1) dt = f-1 f(x, x, t) dt.

1

J(x)

Applying the Euler equations, we require

=0
f " - df"
dt
or

d
x .= 0
2
dtJx +1

1- A-

or

+ 1 =,. t + c.

J
1

X2

See Problem 12 for a different formulation of this problem.

190

OPTIMIZATION OF FUNCTIONALS

It is easily verified that a solution to this first-order differential equation

is given by the arc 0:' a circle
(x -

Xl)1.+

(t - tlYZ = r2.

The parameters Xl' t1> alld r are chosen to satisfy the boundary conditions
and the condition on total length.

GLOBAL THEORY

7.8 Convex and Concave Functionals

Definition. A real-valued functional I defined on a convex subset C of a

linear vector space is said to be convex if

__- L ____- L ______

Figure 7.6

______

A convex function

C and all ex, 0 < ex < 1. If strict inequality holds whenever

,lis said to be strictly convex. A functional 9 defined on a convex
set is said to be (strictly) concave if - 9 is (strictly) convex.

for all

Xl' X2 E

Xl ::j: X2

Examples of convex functions in one dimension are I(x)

for X > 0; and
discontinuous function

x=o

x>O

= x 2 ; I(x) =

eX

7.8

CONVEX AND CONCAVE FUNCTIONALS

191

defined on [0, 00). The functional

1

f(x) =

So {x

(t)

+ Ix(t)l}

dt

defined on L 2 [0, 1] is convex and continuous (the reader may wish to

verify this). Any norm is a convex functional.
A convex functional defined on an infinite-dimensional normed space
may be discontinuous everywhere since, for example, any linear functional
is convex, and we constructed discontinuous linear functionals earlier.
Convex functionals play a special role in the theory of optimization
because most of the theory of local extrema for general nonlinear functionals can be strengthened to become global theory when applied to
convex functionals. Conversely, results derived for minimization of convex
functionals often have analogs as local properties for mOre general problems. The study of convex functionals leads then not only to an aspect of
optimization important in its own right but also to increased insight for
a large portion of optimization theory.
The following proposition illustrates the global nature of results for
minimization problems ill1volving convex functionals.

Proposition 1. Let f be a convex functional defined on a convex subset C

of a normed space. Let !J. = inf f(x). Then
xeC

1.
2.

The subset n of C where f(x) = !J. is convex.

If Xo is a local minimum off, then f(xo) = !J. and, hence Xo is a global
minimum.
Proof

1.

2.

Suppose Xl' x2 E n. Then for X = IXX I + (l - IX)X 2 , < IX < 1, we

have f(x) S;; IXf(x:!) + (1 - lX)f(X2) =!J.. But for any x E C necessarily f(x) ;;::: 11 Thus, f(x) = 11
Suppose N is a neighborhood about Xo in which Xo minimizes f
For any XI E C, there is an X E N such that X = IXXo + (l - IX)XI for
some IX, < IX < I. We have f(xo) S;; f(x) S;; IXf(xo) + (1 - r:t.)f(xI)'
Or f(xo) s;;f(xI)' I

The study of convex functiO'nals is quickly and effectively reduced to

the study of convex sets by considering the region above the graph of the
functional.
Definition. In correspondence to a convex functional 1 defined on a convex
set C in a vector space X, we define the convex set [I, C] in R x X as

[f, C]

= {(r, x) E R X

X: x E C,J(x) s; r}.

192

OPTIMIZATION OF FUNCTIONALS

Usually we think of the space R x X as being oriented so that the R

axis, i.e., all vectors of the form (r, 0), is the vertical axis. Then the set
[J, C] can be thought of as the region above the graph of J, as illustrated
in Figure 7.7. This set [J, C] is sometimes called the epigraph off over C.

-+---'---------------'--x
Figure 7.7

The convex region above the graph

Although we have no occasion to do so, we could imagine forming the

set [f, C] as in the above definition even iff were not convex. In any case,

however, we have the following proposition.

Proposition 2. The functional f defined on the convex domain C is convex

if and only if [J, C] is a convex set.

The major portion of our analysis of convex functionals is based on
consideration of the corresponding convex set of the last definition. To
apply our arsenal of supporting hyperplane theorems to this set, however,
it must be determined under what conditions the set [J, C] contains
interior points. The next section is devoted primarily to an analysis of this
question. Since the results are so favorable-namely continuity of the
functional at a single point guarantees an interior point-the reader need
only glance through the proposition statements at first reading and proceed
to the next section.
*7.9

Properties of the Set [f, C]

Proposition 1. If f is a convex functional on the convex domain C in a

normed space and C has nonempty relative interior C, then the COnvex set
[f, C] has a relative interior point (ro, xo) if and only iff is continuous at
the pOint Xo e C.

7.9

PROPERTIES OF THE SET

[f, e]

193

Proof First assume that f is continuous at a point Xo E C. Denote by

N(xo, J) the open spherical neighborhood of Xo with radius tJ. We note
that v([f, e]), the linear variety generated by [f, e], is equal to R x v( e).
Given e, 0 < e < I,
is a 15 > 0 such that for x E N(xo, (5) (1 vee) we
havex E Cand If(x) - /(xo) I < e. Let ro = f(xo) + 2. Then the point (ro, xo)
E [f, e] is a relative interior point of [f, e] since (r, x) E [f, e] for Ir rol < I and x E N(xo, (5) (1 vee).
Now suppose that (ro, xo) is a relative interior point of [f, e]. Then
there is eo> 0, 15 0 > 0 such that for x E N(xo, tJ o) (1 vee) and Ir - rol < eo
we have r ;;::'f(x). Thusfis bounded above by f(xo) + eo on the
hood N(xo, (5 0 ) (1 vee).
We show now that the above implifs thatfis continuous at Xo. Without
loss of generality we may assume x 0 = 0 and fe", 0) = O. For any e, 0 < e < I,
and for any x E N(x o , e15 o) (1 vee), we have

where eo is the bound on f in N (x 0' (jo)

0=f(8)=f[-_1
1+e

(1

v( e). Furthermore,

x+ (1 _1+e
_
1 )(_!x)] :::;_1 f(x)
e
1+e

or

f(x);;::'

-eeo.
'I.

Therefore, for xEN(x o ,e15 o)(1v(e), we have If(x)l:::;eeo, Thusfis

continuous at xo, I
Convex functionals enjoy many of the properties of linear functionals.
As an example, the following proposition is a generalization of Proposition I, Section 5.2,

Proposition 2. A convex functional f defined on a convex domain e and

continuous at a single point in the relative interior C of e is continuous
throughout C.
Proof Without loss of generality we may assume thatf is continuous
at (J E C and thatf(O) = O. Furthermore, by restricting attention to vee),
we may assume that e has interior points rather than relative interior
points.

194

OPTIMIZATION OF FUNCTIONALS

Let y be an arbitrary point in C. Since Cis (relatively) open, there is a

C. Given e > 0, let 0 > 0 be such that Ilxll < 0 implies
liz - yll < (1 - P-l)O, we have

p > 1 such that py e

If(x) I < e. Then for
I

z
for some

xe

= Y + (l -

P-l)X == P-l(py)

+ (1

- P-l)X

Cwith Ilxll < O. Thus z e C and

f(z) S P-lf(Py)

+ (1

- P-l)f(x) < p-1j(py)

+ (1

- p-l)e.

Thus f is bounded above in the sphere liz - yll < (1 - P-l)<5. It follows
that for sufficiently large r the point (r, y) is an interior point of [J, C];
hence, by Proposition 1,J is continuous at y. I
The proof of the following important corollary is left to the reader.
Corollary 1. A convex functional defined on a finite-dimensional convex set C

is continuous throughout

6.

Having established the simple relation between continuity and interior

points, we conclude this section by noting a property off which holds if
[J, C] happens to be closed. As illustrated in Figure 7.8, closure of
[I, C] is related to the continuity properties of f on the boundary of C.

.r

Figure 7.8 A nonclosed epigraph

Proposition 3. If [J, C] is closed, then f is lower semicontinuous on C.

Proof The set {(a, x) e R x x: x e X} is obviously closed for each

a e R. Hence, if [J, C] is closed, so is
[f, C] n {(a, x) : x e X} == {(a, x); x e C,f(x)

a}

7.10

CONJUGATE CONVEX FUNCTIONALS

195

for each a E R. It follows that the set

Ta = {x : x E C,J(x) :5: a}
is closed.
Now suppose {Xi} is a sequence from C converging to x E C. Let
b = lim inf f(x;). If b = - 00, then x E Ta = Ta for each a E R which is
X,"" X

impossible. Thus b> - 00 and x E Tb +. = Tb+. for all B > O. In other

words,J(x):5: lim inff(x,) which proves thatfis lower semicontinuous. I
Figure 7.9 shows thl! graph of a convex functional f defined on a disk C
in E2 that has closed [j, C] but is discontinuous (although lower semicontinuous) at a point x.
f

x.
I

)'
X2

Figure 7.9

7.10

Conjugate Convex Functionals

A purely abstract approach to the theory of convex functionals, including

a study of the convex set [j, C] as in the previous section, leads quite
naturally to an investigation of the dual representation of this set in terms
of closed hyperplanes. The concept of conjugate functionals plays a natural
and fundamental role in such a study. As an important consequence of this
investigation, we obtain a very general duality principle for optimization
problems which extends the earlier duality results for minimum norm
problems.
Definition. Let f be a convex functional defined on a convex set C in a
normed space X. The conjugate set C * is defined as

C*

= {x* E X*:

sup [(x, x*) - f(x)] < oo}

xeC

196

OPTIMIZATlvN OF FUNCTIONALS

and the junctional j* conjugate

.0 j

is defined on C * as

f*(x*) = sup [(x, x*> - f(x)],

xeC

Proposition 1. The conjugate set C * and the conjugate junctional f* are

convex and [f*, C*] is a closed convex subset of R x X*.
Proof For any xr,
sup {(x, exx!
xeC

X* and any ex, 0 < ex < 1, we have

+ (1 - ex)xi> - f(x)}

= sup {ex[ (x, xi> - f(x)]

xeC

+ (1 -

ex) [ (x, xi> - f(x)]}

;s; ex sup [(x, xD - f(x)]

xeC

. + (1 -

ex) sup [(x, xi> - f(x)]

xeC

from which it follows immediately that C* andf* are convex.

Next we prove that [j*, C*] is closed. Let {CSt, xi)} be a convergent
sequence from [j*, C*] with (st> xi) -+ (s, x*). We show now that
(s, x*) E [j*, C*]. For every i and every x e C, we have

f*(xi)

Sl

Taking the limit as i

cx:>, we obtain

s
for all x

(x, xi> - f(x).

(x, x*> - f(x)

C. Therefore,

sup [(x, x*> - f(x)]

xeC

from which it follows that x* e C* and s

We see that the conjugate functional defines a set [j*, C*] which is of
the same type as [f, C]; therefore we write [f, C]* = [f*, C*]. Note that
if f = 0, the conjurate functional f* becomes the support functional of C.

Example 1. Let X = C = En 'd define, for x;;:::: (Xl' Xz, ... , x n), f(x) =
lip Li=l Ixi/P, I < P < cx:>. Then for x* = (el> e2 , , en),

7.10

CONJUGATE CONVEX FUNCTIONALS

197

The supremum on the right is achieved by some x since the problem is

finite dimensional. We find, by differentiation, the solution

(;i = Ixi lp f*(x*)

where lip

L Ixd
n

i=l

sgn

P(

Xi

1) = -1L led
n

1- -

qi=l

+ llq = 1.

Let us investigate the relation of the conjugate functional to separating

hyperplanes. On the space R x X, closed hyperplanes are represented by
an equation of the form

sr

+ (x, x*) =

where s, k, and x* determine the hyperplane. Recalling that we agreed to

refer to the R axis as vertical, we say that a hyperplane is non vertical if it
intersects the R axis at one and only one point. This is equivalent to the
requirement that the deJlining linear functional (s, x*) have s #- o. If attention is restricted to nonvertical hyperplanes, we may, without loss of
generality, consider only those linear functionals of the form (-1, x*).
Any nonvertical closed hyperplane can then be obtained by appropriate
choice of x* and k.
To develop a geometric interpretation of the conjugate functional, note
that as k varies, the solutions (r, x) of the equation

(x, x*) - r = k
describe parallel closed hyperplanes in R x X. The number f*(x*) is the
supremum of the values of k for which the hyperplane intersects [f, C].
Thus the hyperplane (x, x*) - r = f*(x*) is a support hyperplane of
[I, C].
In the terminology of Section 5.13, f*(x*) is the support functional
h[( -1, x*)] of the functional (-1, x*) for the convex set [f, C]. The
special feature here is that we only consider functionals of the form
(-1, x*) on R x X and thereby eliminate the need of carrying an extra
variable.
For the application to optimization problems, the most important geometric interpretation of the conjugate functional is that it measures vertical
distance to the support hyperplane. The hyperplane

(x, x*) - r = f*(x*)

intersects the vertical axis (i.e., x = 8) at ( - J*(x*), 8). Thus, - J*(x*) is
the vertical height of the: hyperplane above the origin. (See Figure 7.10.)

198

OPTIMIZATION OF FUNCTIONALS

1' ....

r =-f*(x*)

"-

.... ....

, ...

....

.... .$.x, x*) - r = f*(x*)

"

Figure 7.](; A conjugate convex functional

Another interpretation more clearly illuminates the duality between
[f, C) and [f*, C *] in terms of the dual representation of a convex set
as a collection of points or as the intersection of half-spaces. Given the
point (s, x*) R x X*, let us associate the half-space consisting of all
(r, x) R x X satisfying

(x, x*) - r

s.

Then the. set [f*, C *] consists of those (nonvertical) half-spaces that contain the set [f, C]. Hence [f*, C*] is the dual representation of [f, C].
Beginning with an arbitrary convex functional ({J defined on a convex
subset r of a dual space X*, we may, of course, define the conjugate of ({J
in X** or, alternatively, following the standard pattern for duality relations (e.g., see Section 5.7), define the set *r in X as

*r =

{x: sup [(x, x*) - ({J(x*)] < oo}

xOe r

and the convex functional

*({J(x) = sup [(x, x*) - ({J(x*)]

xOer

on *r. We then write *[({J, r] = [*({J, *r]. With these definitions we have
the following characterization of the duality between a convex functional
and its conjugate.
Proposition 2. Let f be a convex functional on the convex set C in a normed
space X. If [f, C] is closed, then [J, C] = *[[f, C]*].

Proof We show first that [J, C] c: *[f*, C*] = *[[J, C]*]. Let
(r, x) E [J, C]; then for all x* E C *, f*(x*) (x, x*) - f(x). Hence, we
have r
(x, x*) - f*(x*) for all x* C*. Thus
r;;:: sup [(x, x*) - f*(x*)]

.nee

and (r, x)

*[f*, C *].

7 . 11

CONJUGATE CONCAVE FUNCTIONALS

199

We prove the converse by contraposition. Let (r o , xo) [f, C]. Since

[f, CJ is closed, there is a hyperplane separating (r o , x o) and [f, C].
Thus there exist x* E X*, S, and c such that

sr

+ <x, x*>:::;; c < sro +<xo, x*>

for all (r, x) E [f, C]. It can be shown that, without loss of generality, this
hyperplane can be assumed to be nonvertical and hence s=!:O (see Problem 16). Furthermore, since r can be made arbitrarily large, we must have
s < O. Thus we take s = - 1. Now it follows that <x, x*> - f(x) :::;; c
for all x e C, which implies that (c, x*) e [f*, C*J. On the other
hand, c < <xo, x*) - r0 implies <xo, x*) - c > ro, which implies that
(ro,xo)*[f*,C*].1
7.11

Conjugate Concave Functionals

A development similar to that of the last section applies to concave funetionals. It must be stressed, however, that we do not tre;;t concave functionals by merely multiplying by - I and then applying the theory for
convex functionals. There: is an additional sign change in part of the
definition. See Problem 15.
Given a concave functional 9 defined on a convex subset D of a vector
space, we define the set

[g, D] = {(r, x) : xED, r :::;; g(x)}.

The set [g, DJ is convex and all of the results on continuity, interior points,
etc., of Section 7.9 have direct extensions here.

Definition. Let 9 be a concave functional on the convex set D. The conjugate set D* is defined as
D*

= {x* e X*:

inf[<x, x*> - g(x)] > - OO},

xeD

and the functional g* conjugate to 9 is defined as

g*(x*)

= inf[<x, x*> -

g(X)].

xeD

We can readily verify that D* is convex and that g* is concave. We

write [g, DJ* = [g*, D*].
Since our notation does not completely distinguish between the development for convex and concave functiona\s, it is important to make clear
which is being employed in any given context. This is particularly true when
the original function is linear, since either definition of the conjugate
functional might be employed and, in general, they are not equal.

200

OPTIMIZATION OF FUNCTIONALS

The geometric interpretation for concave conjugate functionals is

similar to that for convex conjugate functionals. The hyperplane <x, x*) r = g*(x*) supports the set [g, D]. Furthermore, - g*(x*) is the intercept
of that hyperplane with the vertical axis. The situation is summarized in
Figure 7.11.
/"

/" =-g*(.\"*)

"'"

"'" "'"

"'"

-+------.. . . . .

Figure 7.11

.\'

A conjugate concave functional

7.12 Dual Optimization J'lroblems

We come now to the application of conjugate functionals to optimization.
Suppose we seek to minimize a convex functional over a convex domain.
Or more generally, iff is convex over C and 9 is concave over D, suppose
we seek
inf [f(x) - g(x)].
C"D

In standard minimization problems, 9 is usually zero. But as we shall see,

the present generalization is conceptually helpful. The general problem is
illustrated in Figure 7.12. The problem can be interpreted as that of finding
the minimum vertical separation of the sets [J, C] and [g, D]. It is
reasonably clear, from geometric intuition, that this distance is equal to
---the maximum vertical separation of two parallel hyperplanes separating
[j, C] and [g, D]. This relation, between a given minimization problem
and an equivalent maximization problem, is a generalization of the duality
principle for minimum norm problems.
Conjugate functionals are precisely what is needed for expressing this
duality principle algebraically. Since - f*(x*) is the vertical distance to a

7.12

DUAL OPTIMIZATION PROBLEMS

201

,.

> x

Figure 7.12

support hyperplane below [j. C], and - g*(x*) is the vertical distance to
the parallel support hyperplane above [g, D], g*(x*) - f*(x*) is the vertical separation of the two hyperplanes. The duality principle is stated in
detail in the following theorem.

Theorem 1. (Fenchel Duality Theorem) Assume that f and 9 are, respectively, convex and concave functjonals on the convex sets C and D in a
normed space X. Assume that C n D contains points in the relative interior
of C and D and that either [j; C] 01' [g, D] has nonempty interior. Suppose
further that p. = inf {f(x) - g(x)} is finite. Then
xeCnD

p.

inf {f(x) - g(x)} =

xeCr.D

max

{g*(x*) - /*(x*)}

x"'eCnD'"

where the maximum on the right is achieved by some X6 E C* n D*.

If the infimum on the left is achieved by some Xo E C n D, then

max [<x, xri) - f(x)] = <xo, xri) - f(xo)

xeC

and

min [(x, X6) - g(x)]

xeD

= <xo , X6)

- g(Xo)

Proof By definition, for all x* E C * n D*, x E C n D,

f*(x*)

(x, x*) - f(x)

g*(x*)

(x, x*) - g(x).

Thus,
f(x) - g(x)

g*(x*) - f*(x*)

202

OPTIMIZATION OF FUNCTIONALS

and hence
sup [g*(x*) - f*(x*)J.

inf [f(x) - g(x)]

CnD

COnDO

Therefore, the equality in the theorem is proved if an X6 e C * n D* can

be found for which inf [f(x) - g(x)] =
CnD

The convex set [f - Il, C] is a vertical displacement of [f, C); by

definition of Il the sets [f - Jl, C) and [g, D] are arbitrarily close but have
disjoint relative interiors. Therefore, since one of these sets has nonempty
interior, there is a closed hyperplane in R x X separating them. This
hyperplane cannot be vertical because otherwise its vertical projection
onto X would separate C and D. Since the hyperplane is not vertical, it
can be represented as {(r, x) e R x X: (x;
- r = c) for some
e X*
and c R. Now since [g, D] lies below this hyperplane but is arbitrarily
close to it; we have

c == inf [(x,

- g(x)] = g*(X&).

xeD

Likewise,
c == sup
xeC

[<X,

f(x) + Il] =

+ Il.

Thus Jl =
If the infimum Il on the left is achieved by some Xo e C ("\ D, the sets
[f - Jl, C] and [g, D] have the point (g(xo), xo) in common and this
point lies in the separating hyperplane. I
In typical applications of this theorem, we consider minimizing a convex
functional f on a convex domain D; the set D representing constraints.
Accordingly, we take C = X and 9 = 0 in the theorem. Calculation of f*
is itself an optimization problem, but with C = X (t is unconstrained.
Calculation of g* is an optimization problem with a linear objective
functional when 9 = O.

Example 1. (An Allocation Problem) Suppose that there is a fixed quantity Xo of some commodity (such as money) which is to be allocated among
n distinct activities in such a way as to maximize the total return from
these activities. We assume that the return associated with the i-th activity
when allocated XI units is ;."JXI) where g/, due to diminishing marginal
returns, is assumed to be an increasing concave function. In these terms
the problem is one of finding x ::= (X., xz ... , x n ) so as to
= tlg/(X/)

(1)

subject to LXI == Xo,

1=1

XI
'

i = 1,2, ... , n.

7.12

DUAL OPTIMIZATION PROBLEMS

203

To solve this problem by conjugate functionals, we set D equal to the

positive orthant,Jequal to the zero functional, and C = {x: L;'= 1 Xi = xo}.
The functional conjugate to f is (for Y = (YI' Y2 , ... , Yn
j*(y) = sup y'x.
I:Xi=,xO

This is finite only jf Y = A(I, I, ... , I) in which case it is equal to AXo

Thus

C*={Y:Y=A(I, 1, ... , I)}

f*(A(l, 1, ... , 1, 1

= AXo .

Also, for each i we define the conjugate functions (of a single variable)
(2)

gi(y/)

= x,.,o
inf [Xi y/ -

g/(x/)]

and it is clear that

n

g*(y) ::;::

I 1 gi(y/).

i=

The problem conjugate to (1) is therefore

(3)

min [AXo "

gi(A)].

i= 1

In this form we n01le that to solve the allocation problem requires

evaluation of the conjugate functions 91 and then solution of (3) which is
minimization with respt:ct to the single variable A. Once the optimal value
of A is determined the x;'s which solve (1) can be found to be those which
minimize (2) with each Yi = A.
This analysis can be modified in order to apply to a multistage allocation
problem where there is the possibility of investment growth of uncommitted resources. See Problem 19.

Example 2. (Horse-Racing Problem) What is the best way to place bets

totaling Xo dollars in a race involving n horses? Assume we know Pi' the
probability that the i-th horse wins, and s/, the amount that the rest of
the public is betting on the i-th horse. The track keeps a proportion 1 - C
of the total amount belt (0 < I - C < 1) and distributes the rest among
the public in proportion to the amounts bet on the winning horse.
Symbolically, if we
Xj on the i-til horse, i = 1, 2, ... , n, we receive
c(xo+

1=1

Sj

+ Xi

204

OPTIMIZATION OF FUNCTIONALS

if the jth horse wins. Thus the expected net return, R, is

(4)

R
I

p.x.} = C(Xo + Ln){

51
Ln _'_I
1=1.

i'=1 51

+ XI

Xo.

Our problem is to find Xi' i:::::: 1,2, ... , n, which maximize R subject to
n

(5)

L XI::::: Xo,

Xi;::: 0,

1=1

i = 1,2, ... , n

or, equivalently, to maximize

(6)

subject to (5), where

gi(Xi)::::: Pi XI .
51 + XI

This problem is exactly the type treated in Example I since each gj is

conCave for positive Xi (as can be verified by observing that the second
derivative is negative). Solution to the problem is obtained by calculating
the functions conjugate to the g/s.
A typical gi is shown in Figure 7.13. Its slope at Xi::::: 0 is Pi/SI and it
approaches the value PI as XI
The value of the conjugate functional
at Ais obtained by finding the lowest line of slope ..1. which lies above gj.
It is clear from the diagram that forA ;::: Pi/SI, we have gt(),) = 0, and that
forA < 0, gt(..1.) is not defined. For 0 < ..1. < pdSI' we have
g,/,(..1.) = min [..1.Xi - Olxj)J.

(7)

,,/>0

Pi

Figure 7.13

7.12

DUAL OPTIMIZATION PROBLEMS

205

Performing the minimization by differentiation, we have the equation

A_
- (SI

Si Pi

+ X/)2

or
(8)

Substitution back into (7) yields

for 0< A <
(9)

Pi
Si

p.
for A ;;::: ....!
sI

where XI is determined from equation (8).

We can now deduce the form of the answer. Suppose that A has been
found from (3). To simplify the notation, rearrange the indices so that
Pl/SI > P2/S2 > .. , > Pn/sn (assuming strict inequality). For the given A,
we define m as the largest index for which pt/s, :?! ).. Then, from equations
(8) and (9), our solution may be written
for i = 1,2, ... , m

(10)

for i

= m + 1, ... , n.

The parameter A in this section can be found from (3) or from the constraint
(5). In other words, A is chosen so that
(11)

Now SeA) is easily verified to be continuous and S(O) = 00, S( (0) = O. Thus
there is a A satisfying (11).
Note that for small XCI (xo < < max Sj), the total amount should be
on a single horse, the horse corresponding to the maximum p;/Sj,
or equivalently, the maximum Pi ri where rj := C Lj Sj/sl is the track odds.
Example 3. Consider the minimum energy control problem discussed in
Sections 3.10 and 6.10. We seek the element U E L 2 [O, I] minimizing

feu) = t

tu
1

(t) dt,

206

OPTIMIZATION OF FUNCTIONALS

while satisfying the linear constraints

Ku=c
where K: L 2 [0,' 1] -+ En.
In Theorem 1, let C = L 2 [O, 1], 9 = 0, and D = {u: Ku = c}. We
assume that Dis nonempty. Then we can readily verify that

C* =L 2 [O, 1]
and

f*<u*) =

fo [U*(t)]2 dt.

Since 9 = 0, the conjugate functional of the set D is equal to the support

functional. Thus

D*

= {u* ; u* = K*a, a e En}

and g*(K*a) = (a I c).

The dual problem is therefore the finite-dimensional problem in the
vector a:
ma {(a I c) - !(K*a I K*a)}
which is solved by finding the n vector a satisfying

KK*a

=c

where KK* is an n x n matrix.

Finally, the solution to the original problem can be found in terms of
the solution of the dual by application of the second part of Theorem 1.
Thus
.
Uo

= K*a.

*7.13 Min-Max Theorem of Game Theory

In this section we briefly introduce the classical theory of games and prove
the fundamental min-max theorem. Our purpose is to show that the
min-max theorem can be regarded as an example of the Fenchel duality
theorem.
Let X be a normed space and X* its normed dual. Let A be a fixed
subset of X and B a fixed subset of X*. In the form of game that we
consider, one player (player A) selects a vector from his strategy set A
while his opponent (player B) selects a vector x* from his strategy set B.
When both players have selected their respective vectors, the quantity

7.13

MIN-MAX THEOREM OF GAME THEORY

207

(x, x*) is computed and player A pays that amount (in some appropriate
units) to player B. Thus A seeks to make his selection so as to minimize
(x, x"'> while B seeks to maximize (x, x"'>.
Assuming for the moment that the quantities
fJ,0

= min max (x, x"'>

xeA xOeB

fJ,0

= max min (x, x"'>

xOeB xeA

exist, we first take the viewpoint of A in this game. By selecting x E A, he

loses no more than max (x, x"'>; hence, by proper choice of x, say xo,
x"eB

he can be assured oflosing no more than fJ,0. On the other hand, player B
by selecting x'" E B, wins at least min (x, x"'>; therefore, by proper choice
xeA

of x"', say X6, he can be assured of winning at least fJ,o. It follows that
fJ,0
(xo, X6> :s;; /1-0, and the fundamental question that arises is whether
fJ,o = fJ,0 so that there is determined a unique pay-off value for optimal
play by both players.
The most interesting type of game that can be put into the form outlined above is the classical finite game. In a finite game each player has a
finite set of strategies and the pay-off is determined by a matrix Q, the
pay-off being qij if A uses strategy i and B uses strategy j. For example,
in a simple coin-matching game, the players independentlY select either
" heads" or " tails." If their choices match, A pays B 1 unit while, if they
differ, B pays A 1 unit. The pay-off matrix in this case is
Q= [

1 -1]

-1

Finite games of this kind usually do not have a unique pay-off value.
We might, however, consider a long sequence of such games and" randomized" strategies where each player determines his play in anyone
game according to a fixed probability distribution among his choices.
Assuming that A has n basic strategies, he selects an n vector of probabilities x = (Xl' X2' , x n) such that Xi;::: 0, 2:i,'1 Xi = 1. Likewise, if B has m
strategies, he selects Y = (YI' Y2 , ... , Ym) such that Yi ;::: 0, Ii"= 1 Yi = 1. The
expected (or average) pay-off is th.en (x IQy).
DefiningA = (x
O,LI'= 1 XI = I} c En,andB = {x'" :x'" = QY'Yi;::: 0,
1 YI = I} c En, the randomized game takes the standard form given at
the beginning of the section. Note that A and B are both bounded closed
convex sets. Other game-type optimization problems with bilinear objectives other than the classical randomized finite game also take this form.
We now give a simple proof of the min-max theorem based on duality.
For simplicity, our proof is for reflexive spaces, although more general

Liz

208

OPTIMIZATION OF FUNCTIONALS

versions of the result hold. For the generalizations, consult the references
at the end of the chapter.
Theorem 1. (MiniMax) Let X be a reflexive normed space and let A and B
be compact convex subsets of X and X*, respectively. Then
min max (x, x*)
"eA "oeB

= max min (x, x*).

"oeB "eA

Proof. Define the functional f on X by

f(x) = max (x, x*).
"Oe B

The maximum exists for each x e X since B is compact. The functional is

easily shown to be convex Ilnd continuous on X. We seek an expression for
min f(x)
"sA

which exists by the compactness of A and the continuity of f We now

apply the Fenchel duality theorem with the associations: f -+ J, C -+ X,
g -. 0, D -. A. We have immediately:
(1)

D*=X*

(2)

g*(x*) = mill (x, x*).

"eA

We claim that furthermore

C* =B

(3)
(4)

f*(X*)

= o.

To prove (3) and (4), let x! B, and by using the separating hyperplane
theorem, let Xl e X and IX be such that (XI' xi) - (XI' x*) > IX > 0 for
all x* E B. Ther. (x, xi> - max (x, x*) can be made arbitrarily large by
"Oe B

taking x = kXl with k > 0. rhus

sup [(x, xi) - f(x)] ==
and xi C*.
Conversely, if x!

00

"
E

B, then (x, xi) - max (x, x*) achieves a maximum

II'&B

value of 0 at x = 9. This establishes (3) and (4).

The final result follows easily from the equality

minf(x) = max
""A

g*(x*) =

,,eBnXo

max min (x,

"oeB "eA

x*).

7.14

PROBLEMS

209

An interesting special case of the min-max theorem is obtained by

taking B to be the unit sphere in X*. In that case we obtain
min

Ilxll

"eA

max - h(x*)

11"*11 s

where h is the support functional of the convex set A. This result is the
duality theorem for minimum norm problems of Section 5.8.

7.14 Problems
1. On the vector space of continuous functions on [0, 1], define
f(x) = max x(t).
OSrSI

Determine for which x the Gateaux differential l>j(x; h) exists and is

linear in h.
2. Repeat Problem 1 for

f(x)::;

Ix(t)1 dt.

3. Show that the functional

if XI = 0
is Gateaux differentiable but not continuous at XI = Xl = o.
4. On the space X = C [0, 1], define the functional f(x) = [X(!)]l. Find
the Frechet differential and Frechet derivative off
5. Let qJ be a function of a single variable having a continuous derivative
and satisfying IqJWI < Klel. Find the Gateaux differential of the
functionalj(x) =
I qJ(e i) where X = gil Ell. Is this also a Frechet
differential?
6. Suppose the real-valued functionalj defined on an open subset D of a
normed space has a relative minimum at Xo E D. Show that ifjis twice
Gateaux differentiable at x o , then (h,f"(x o )h) 0 for all hEX.
7. Let f be a real-valued functional defined on an open region D in a
normed space X. Suppose that at Xo E D the first Frechet differential
vanishes identically on X; within a sphere, S(xo, e), f"(x) exists
and is continuous in x; and the lower bound of (h,J"(xo)h) for
IIhll = 1 is positive. Show that f obtains a relative minimum at Xo.
8. Let A be a nonempty subset of a normed space X and let Xo EA.
Denote by C(A, xo} the closed cone generated by A - Xo, i.e., the

Ir;

210

OPTIMIZATION OF FUNCTIONALS

intersection of all closed cones containing A - Xo' The local closed

cone of A at Xo is the set
LC(A, xo)

==

n C(A

Ii N,

xo)

N e.f'

where. is the class of all neighborhoods of Xo. Suppose that over A

the Frechet differentiar,;e functional f achieves a minimum at Xo.
Show that

9. In some problems in the calculus of variations, it is necessary to

consider a broader class of functions than usual. Suppose that we
seek to extremize

= fa P[x, x, t] dt
b

among all functions x for which x(a) = A, x(b) = B and which have
continuous derivatives on [a, b] except possibly at a single point in
[a, b]. If the extremum has a continuous derivative except at C E [a, b],
show that Fx = dF;i;/dt on the intervals [a, c) and (c, b] and that the
functions F;i; ,and F - XP;i; are continuous at c. These are called the
Weierstrass-Erdman corner conditions. Apply these considerations to
the functional

x(-l) = 0,

x(l) = 1.

10. Consider the effect of an additional source of constant income in the

estate-planning problem.
11. LetfI' fl' ... ,In be linearly independent linear functionals on a vector
space X. Show that there are n elements Xl> Xl, , Xn in X such that
the n x n matrix [/i(x})] is nonsingular.
12. Formulate and solve the isoperimetric problem of Example I of
Section 7.7 by using polar coordinates.
13. Solve the candidates allocation problem described in Chapter 1.
14. Let X = L 2 [O, 1] and define f(x) = SA (tx 2(t) + 1x(t)1} dt on X. Find
the conjugate functional off.
15. Exhibit a convex set C having the property that the convex conjugate
functional of 0 over C is not equal to the negative of the concave
conjugate functional of 0 over C.
16. Let M be a nonempty closed convex set in R x X, and assume that
there is at least one nonvertical hyperplane containing M in one of its
half-spaces. Show that for any (r o. xo) M there is a nonvertical hyper-

7.14

PROBLEMS

211

plane separating (ro, xo) and M. Hint: Consider convex combinations

of hyperplanes.
17. Let f be a convex functional on a convex set C in a normed space
and let [1*, C*] = [f, C]*. For x E C, x* E C*, deduce Young's
inequality
(x, x*) !;,f(x) + f*(x*).
Apply this result to norms in Lp spaces.
18. Derive the minimum norm duality theorem (Theorem 1, Section 5.8)
directly from the Fe:nchel duality theorem.
19. Suppose a fixed quantity Xo of resource is to be allocated over a given
time at n equally spaced time instants. Thus XI !;, Xo is allocated first.
The remaining resource Xo - XI grows by a factor a so that at the
second instant X2 s; a(xo - XI) may be allocated. In general, the
uncommitted resource grows by the factor a between each step. Show
that a sequence of allocations {xiH= I' Xi ;;:: 0 is feasible if and only if
an-ixi

+ an-2x2 + '" + aXn_ 1 + xn!;, an-Ixo'

Hence, show how to generalize the result of Example 1, Section 7.12,

to multistage problems.
20. The OWner of a small food stand at the beach is about to order his
weekend supply of food. Mainly, he sells' ice cream and hot dogs
and wishes to optimize his allocation of money to these two items.
He knows from past experience that the demands for these items
depend on the weather in the following way:
Ice cream
Hot dogs

Hot Day
1000
400

Cool Day
200
200

He believes that a hot or a cool day is equally probable. Anything he

doesn't sell he may return for full credit. His profit on ice cream is
10 cents and on hot dogs it is 30 cents. He, of course, wishes to
maximize his expected profit while remaining within his budget of
$100. Ice cream costs him 10 cents and hot dogs 20 cents.
(a) Formulate his problem and reduce it to the form
maximizefl(xl) + fix2)
subject to XI + X2 !;, Xo, XI ;;:: 0, X2 ;;:: O.
(b) Solve the problem using conjugate functionals.
21. Consider a control system governed by the n-dimensional set of
differential equations
x(t)

= Ax(t) + hu(t)

212

OPTIMIZATION OF FUNCTIONALS

which has solution

T

x(T)

= <I>(T.)x(O) + fo <I>(T -

t)bu(t) dt,

where <I>(t) is a fundamental matrix of solutions. Using conjugate

function theory and duality, find the control u minimizing
J = tllx(T)ll l

+ t So u2 (t) dt.

Hint: Assume first that x(T) is known and reduce the problem to a
finite-dimensional one. Next optimize x(T). Alternatively, formulate
the problem in E" X L2[O, T].

REFERENCES
7.1-4. For general background on differentials, see Graves [62], [63], Hildebrandt and Graves [72 and Hille and Phillips [73, Chapters 3 and 26]. A

fairly comprehensive account of several extensions of ordinary calculus is

contained in Luisternik and Sobolev [101J. The inequality extension of the
mean value theorem is discussed by Kantorovich and Akilov [79]; also see
Antosiewicz and Reinboldt [9].
7.5-6. The classic treatise on the calculus of variations is Bliss [22]. Also see
Fox [55], For a somewhat more modern approach to the subject, see Gelfand
and Fomin [58] or Akhiezer [2].
7.7. The implicit fUnction theorem used here has generalizations to abstract
spaces .. See Hildebrandt and Graves [72) and also Section 9.2 of this book.
There are a number of abstract versions of the Lagrange multiplier theorem,
e.g., Goldstine [60), (61) and Blum [23).
7.9-12. Conjugate functionals on finite-dimensional space were introduced by
Fenchel [52]. For an excellent presentation of this topic, consult Karlin [82].
Some extensions and related discussions are to be found in Rockafellar [126],
RfLdstrom [121), and Whinston (151]. Extensions to more general linear
topological spaces were made by Brondsted [251, Dieter [41], Rockafellar
[127], and Moreau [107]. See Kretschmer [90] for an example of a case
where the values of the primal and dual problems are unequal. An equivalent
approach, the maximum transform, was developed by Bellman and Karush
[19]. They treat the allocation problem. The horse-racing problem is due to
Rufus Isaacs and was analyzed in detail by Karlin using the Neyman-Pearson
lemma.
7.13. See Edwards [46], Karlin [81], [82], [83], Dresher [44], and McKinsey
[104]. For more general min-max theorems, see Fan [51] and Sion [141].
7.l4. For further developments along the lines of Problem 8, see Varaiya (149].
For a solution to Problem 16, see Brondsted [25].

GLOBAL THEORY OF
CONSTRAINED OPTIMIZATION

8.1 Introduction
The general optimization problem treated in this book is to locate from
within a given subset of a vector space that particular vector which minimizes a given functional. In some problems the subset of admissible vectors competing for the optimum is defined explicitly, as in the case of a
given subspace in minimum norm problems; in other cases the subset of
admissible vectors is defined implicitly by a set of constraint relations. In
previous chapters we: considered examples of both types) but generally we
reduced a problem with implicit constraints to one with explicit constraints
by finding the set of solutions to the constraint relations. In this chapter and
the next we make a more complete study of problems with implicit Constraints defined by nonlinear equality or inequality relations.
Deservedly dominating our attention) of course, are Lagrange multipliers
which somehow almost always unscramble a difficult constrained problem.
Although we encountered Lagrange multipliers at several points in previous
chapters) they were treated rather naively as a convenient set of numbers or
simply as the result of certain duality calculations. In a more sophisticated
approach) we do not speak of individual Lagrange multipliers but instead of
an entire Lagrange mUltiplier vector in an appropriate dual space. For
example, the problem
minimize f(x)
(1)
{subject to H(x) =

e,

where H is a mapping from a normed space X into a normed space Z, has

Lagrangian
L(x, z*) = f(x) + (H(x), z*)
and the Lagrange multiplier is some specific z*
Lagrangian in the functional notation

L(x) z*)

Z*. (We also write the

=I(x) + z* H(x)

since this is similar to the convention for the finite-dimensional theory.)

214

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

There are several important geometric interpretations of constrained

problems and their corresponding Lagrange multipliers. For example,
problem (1) can \,e viewed in the space X by studying the contours off,
or in R x Xby studying the graph off Alternatively, however, the problem
can be viewed in the constraint space Z or in R x Z, and these representations are in many respects more illuminating than those in the primal space
X because the Lagrange multiplier is an element of Z * and appears directly
as a hyperplane in Z. Lagrange multipliers cannot be thoroughly understood without understanding each of these geometric interpretations.
Because of the interpretation of a Lagrange multiplier as a hyperplane,
it is natural to suspect that the theory is simplest and most elegant for
problems involving convex functionals. Indeed this is so. In this chapter we
therefore consider the global or convex theory based on the geometric
interpretation in the constraint space where the Lagrange mUltiplier
appears as a separating hyperplane. In the next chapter we consider the
local theory and the geometric interpretation in the primal space X.
8.2

Positive Cones and Convex Mappings

By introducing a cone defining the positive vectors in a given space, it is

possible to consider inequality problems in abstract vector spaces.

Definition. Let P be a convex cone in a vector space X. For x, y E X, we

write x y (with respect to P) if x - yeP. The cone P defining this
relation is called the positive cone in X. The cone N = - P is called the
negative cone in X and we write y x for y - x e N.
For example, in E" the convex cone
(1)

P = {xeE": x

= (e l , e2'"'' e,,); el ofor all i}

dt:@es the ordinary positive Jrthant. In a space of functions defined on an

interval of the real line, say [tu 12], it is natural to define the positive cone
as consisting of all functions in the space that are nonnegative everywhere
on the interval [t1> 12],
We can easily verify that x ;::= y, y;::= z implies x;::= z and, since 0 e P,
x ;::= x for all x e X.
In a normed space it is sometimes important to define the positive cone
by a closed convex cone. For example, in En the cone defined by (1) is
closed, but if one or more of the inequalities in the description of the set
were changed to strict inequality (and the point 0 adjoined), the resulting
cone would not be closed.
In the case of a normed space, we write x > () if x is an interior point of
the positive cone P. For many applications it is essential that P possess an

8.2

POSITIVE CONES AND CONVEX MAPPINGS

215

interior point so that the separating hyperplane theorem can be applied .

. Nevertheless, this is not possible in many common normed spaces. For
instance, if X = Ll [11' t 2 ] and P is taken as the subset of nonnegative
functions on the interval [tl, t 2], we can easily show that P contains no
interior point. On the other hand, in C[t l , t 2 ] the cone of nonnegative
functions does possess interior points; for this reason the space C [fl' t 2 ] is
of particular interest for problems involving inequalities.
Given a normed space X together with a positive convex cone Pc X, it is
natural to define a corre:sponding positive convex cone p$ in the dual space
X* by
(2)

p$ = {x*

X*: <x, x*)

for all x

P}.

Example 1. If P is taken as given by (I), then p$ consists of all linear

functionals represented by elements of En with nonnegative components.
Example 2. If in the space C [t l' t 2] p is taken as the set of all nonnegative
continuous functions on [tl, t 2 ], then p$ consists of all linear functionals
on C[tl' t 2 ] represented by functions v of bounded variation and nondecreasing on [tl> t2l
We can easily show thatp$ is closed even if p is not. If P is closed, P and
p$ are related through the following result.

Proposition 1. Let the positive cone P in the normed space X be closed. If

x E X satisfies
<x, x*)
then x

for all x*

e,

e.

Proof Assume x P. Then by the separating hyperplane theorem

there is an x* E X* such that <x, x*) < <p, x*) for all pEP. Since P
is a cone, <p, x*) can never be negative because then <x, x*) > <ap, x*)
for some a> O. Thus x* eP$. Also, since inf <p, x*) = 0, we have
(x, x*) < O.

peP

Since we have generalized the notion of inequality between vectors, it is

possible to introduce a general definition of convexity for mappings.
Definition. Let X be a vector space and let Z be a vector space having a
cone P specified as the positive cone. A mapping G : X -+ Z is said to be
convex if the domain n of G is a convex set and if G(exXl + (1 - ex)x 2 ) :::;;
aG(x l ) + (1 - 0:)G(X2) for all Xl> X2 E n and all ex, 0 < 0: < 1.
We note that convexity is not an intrinsic property of a mapping but is
dependent on the specified positive cone in the range space.

216

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

The following elementary property of convex mappings is obviously of

particular interest in constrained optimization problems.

Proposition 2. Let q be a convex mapping as in the last definition. Then for
every Z EZ the set {x: x EO, Cfx)::;; z}is convex.
8.3 Lagrange Multipliers
The basic problem considered in the next few sections is:
.. . "
minimize f(x)
{subject to x E 0, G(x) :s: 0,
(1)

where is a convex subset of a vector space X, f is a real-valued convex

functional on 0, and G is a Convex mapping from into a normed space Z
having positive cone P. Problem (1) above is referred to as the general
convex programming problem.
We analyze problem (1) and develop the Lagrange multiplier theorem
essentially by imbedding it in the family of problems
minimize f(x)
subject to x E 0, G(x) ::;; z
where z is an arbitrary vector in Z. The solution to these problems depends
on z and it is an examination of this dependency that guides our analysis.
In view of the above remark, we define the set r c: Z as
r = {z: There is an x EO with G(x) ::;; z}.
The set r is convex since ZI' Z2 E r implies the existence of Xl' X2 E n
with G(XI):S;; Zl' G(X2):S;; Z2; hence, for any (x, 0 < (X < 1, G(xx i +
(l - (X)x 2)::;; (XZl + (1 - (X)Z2 which implies (XZI + (l - (X)Z2 E r.
On the set r, we define the primal functional w (which may not be
finite) as
w(z) = inf {I(x): x eO, G(x) z}.
The original problem (1) can be regarded as determining the single value
ro(). The entire theory of this chapter is based on a study of roo
Proposition 1. The functional w is convex.

Proof
W(XZI

+ (I

(X)Z2)

= inf {f(x) : x E 0, G(x):s;; (XZI + (1 - (X)Z2}

::;; inf {f(x) : x = (xXI + (I - (x)x;z , Xl EO, X2 E 0,
G(Xl) :s;; Z1' G(X2) ::;; Z2}
:s: (X inf {I(xl ) : Xl E 0, G(Xl) :s;; Zl}
+(1 - (X) inf {f(x2): X2 e.Q, G(X2) Z2}
:s;; (XW(ZI) + (1 - (X)W(Z2)' I

8.3

LAGRANGE MULTIPLIERS

Proposition 2. The functional co is decreasing; i:e., if

Zl

217

Z2' then

CO(Zl) :$ CO(Z2)'

Proof. This is immediate.

A typical co for Z

dimensional is shown in Figure 8.1.

I'

w(Z)

o
8.1

The primal functional

Conceptually the Lagrange multiplier theorem follows from the simple

observation that since co is convex, there is a hyperplane tangent to co at
Z =
and lying below w throughout its region of definition. If one were
to tilt his head so that the tangent hyperplane became the new horizontal,
it would appear that co was minimized at or, said another way, by adding
an appropriate linear functional <z, Z6) to co(z), the resulting combination
co(z) + <z, Z6) is minimized at Z= The functional Z6 is the Lagrange
multiplier for the problem; the tangent hyperplane illustrated in Figure 8.1
corresponds to the element (1, zri) E R x Z*.
The discussion above is made precise by the following theorem.

e.

Theorem 1. Let X be a linear vector space, Z a normed space, n a convex

subset of X, and P the positive cone in Z. Assume that P contains an interior
point.
Letfbe a real-valued convexfunctional on nand G a convex mapping from
n into Z. Assume the existence of a point XI E n jor which G(XI) < () (i.e.,
G(x l ) is an interior point oj N = -P).
Let
(2)

and assume
(3)

1-10 =
1-10

infj(x)

subject to x

n, G(x) ::;; e

is finite. Then there is an element

Po

;;::

= xeO
inf {f(x) + <G(x), Z6)}.

e in Z * such that

218

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

Furthermore, the infimum is achieved in (2) by an Xo

achieved by Xo in (3) and
.

(4)

n, G(xo) :::;; e, it is

(G(xo),
Proof. In the space W = R x Z, define the sets
A = {(r, z) : r '?:.!(x), z '?:. G(x) for some x E n}

B = {(r, z) : r:::;; JJ.o, z:::;; e}.

Since! and G are convex, both A and B are convex sets. (It should be
noted that the set A is the convex region above the graph of the primal
functional co.) The definition of JJ.o implies that A contains no interior points
of B. Also, since N contains an interior point, the set B contains an interior
point. Thus, according to the separating hyperplane theorem, there is a
nonzero element
= (ro , E w* such that

rOr1

+ (Zlo

'?:. rOr2

+ (Z2'

for (rlo Zl) E A, (rz, zz) E B.

From the nature of B it follows immediately that
'?:. or, equivalently,
that ro '?:. 0,
'?:. e. We now show that ro > O. The point (JJ.o, e) is in B;
hence

ror

+ (z,

'?:. roJJ.o

for all (r, z) EA. If ro were zero, it would follow in particular that
(G(Xl)'
0 and that zci :p. However, since G(Xl) is an interior point
of N and
'?:. it follows that (G(X1),
< 0 (the reader should verify
this), which is a contradiction. Therefore, ro > 0 and, without loss of
generality, we may assume ro = 1.
Since the point (JJ.o, e) is arbitrarily close to A and B, we have (with ro = 1)

e.

e,

JJ.o

= (r.z)eA
inf [r + (z,

:::;; inf[f(x) + (G(x),

xeD

:::;; inf f(x) == JJ.O .

xeD
G(x):S8

Hence the first part of the theorem is proved.

If there exists an Xo such that G(xo) :::;; JJ.o

e, =!(xo), then

JJ.o :::;;!(xo) + (G(xo),

and hence (G(xo),

= O.

4) :::;;!(xo) =

JJ.o

The proof of Theorem 1 depends partially on the convexity of set A.

This set, being the region above the graph of the primal functional co, is

8.4

SUFFICIENCY

219

ConVex if and only if co is convex. This in turn is implied by, but weaker
than, the assumption of convexity off and G.
Aside from convexity there are two important assumptions in Theorem 1
that deserve comment: the assumption that the positive cone P contains an
interior point and the assumption that G(X'l) < for some Xl E n. These
assumptions are .introduced to guarantee the existence of a nonverticat"
separating hyperplane. The requirement that the positive cone possess an
interior point is fairly severe and apparently cannot be completely
omitted. Indeed, in many applications this requirement is the determining
factor in the choice of space in which a pro'blem is formulated and, of
C[a, b] is a natural candidate for problems involving functions on
the interval [a, b]. The condition G(x l ) < e is called a regularity condition
and is typical of the assumptions that must be made in Lagrange multiplier
theorems. It guarantees that the separating hyperplane is nonvertical.
We have considered only convex constraints of the form G(x) e. An
equality constraint H(x)
with H(x) = Ax - b, where A is linear, is
equivalent to the two convex inequalities H(x) e and - H(x) e and can
thus be cast into the Dorm G(x) e. One expects then that, as a trivial
corollary to Theorem 1, linear equality constraints can be included together
with their resulting Lagrange multipliers. There is a slight difficulty, however, since there never exists an Xl which simultaneously renders H(xt) <
and - H(xl ) < e. A composite theorem for inequality constraints and a
finite number of equality constraints is given in Problem 7.
Theorem 1 is a geometric version of the Lagrange multipliet theorem for
convex problems. An equivalent algebraic formulation of the results is
given by the following saddle-point statement.

=e

Corollary 1. Let everything be as in Theorem 1 and assume that Xo achieves

the constrained minimum. Then there is a Z&
such that the Lagrangian
L(x, z*) = f(x) + <G(x), z*)
has a saddle point at Xo ,. z&; i.e.,

L(xo, z*)

L(xo ,

L(x,

for all X En, z*

e.

Proof Let
diately L(xo, Z&)

be ddined as in Theorem 1. From (3) we have immeL(x, z;n. By equation (4) we have

L(xo, z*) - L(x o ,

= <G(xo), z*) -

<G(x o),

= <G(xo), z*)

::; O.

8.4 Sufficiency
The conditions of convexity and existence of interior points cannot be
omitted if we are to guarantee the existence of a separating hyperplane in the

220

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

space R x Z. If, however, the appropriate hyperplane does exist, despite

the absence of these conditions, the Lagrange technique for location of the
optimum still applies. The situation is illustrated in Figure 8.2, where again
1

co{z) = inf {I(x) : x en, G(x)::s;; z}

w(z)

w(z)

(a)

(b)

Figure 8.2 Nonconvexity

is plotted, but the convexity of co is not assumed. If, as in Figure 8.2a,
an appropriate hyperplane exists, it is fairly clear that f(x) + (G(x),
attains a minimum at xo' In Figure 8.2b, no supporting hyperplane exists
at z = e and the Lagrange statement cannot be made.
These observations lead to the following sufficiency theorems.

functional defined on a subset n of a

Theorem 1. Let f be a
linear space X. Let G be a mapping from n into the normed space Z having
nonempty positive ;;one P.
Suppose there exists an elen-ent e Z*,
0, and an element XoE n
such that
f(xo) + G(xo) ,
::s;; f(x) + G(x),
for all x E n. Then Xo solves: .
minimize f(x)
subject to G(x) ::s;; G(xo),
x e n.
Proof Suppose there is an Xl en with !(Xl) <f(xo), G(Xl) ::s;; G(xo).
Then, since
e, it follows that

<

<

<G(x

1 ),

and hence that

<G(xo) ,

[(Xl) + \v(X l ), zo> < f(xo) + <.:G(xo), zo>

which contradicts the hypothesis of the theorem. I

8.5

SENSITIVITY

221

n, P, f, G be as above and assume that P is

and an Xo E n such that the
closed. Suppose there exIsts a zri E Z *, zri
Lagrangian L(x, z*) = f(x) + (G(x), z*) possess a saddle point at xo, zri;
i.e.,
L(xo, z*) L(xo ,
L(x,
Theorem 2. Let X, Z,

for all x

e n, z*

e,

8. Then Xo solves:
minimize f(x)
subject to G(x)

8,

xen.

Proof The saddle-point condition with respect to z* gives

for all z*

(G(xo), z*) :::;; (G(xo),

zri)

8. Hence, in particular, for all zi

(G(xo),

zt +

:::;; (G(xo),

or
(G(xo), z!) :::;; O.

We conclude by Proposition I, Section 8.2, that G(xo) 8. The saddlepoint condition therefore implies that (G(xo), zri) = O.
Assume now that Xl e n and that G(XI) 8. Then, according to the
saddle-point condition with respect to x,
f(xo) =f(xo)

+ (G(xo),

Thus Xo minimizesf(x) subject to

+ (G(XI)'
X En, G(x) 8. I

The saddle-point condition offers a convenient compact description of

the essential elements of the Lagrangian results for convex programming.
Iff and G are convex, the positive cone P c Z is closed and has nonempty
interior, and the regularity condition is satisfied, then the saddle-point
condition is necessary a,nd sufficient for optimality of Xo .
8.5 Sensitivity

The Lagrange theorems of the preceding sections do not exploit all of the
geometric properties evident from the representation of the problem in
R x Z. Two other properties, sensitivity and duality, important for both
theory and application, are obtainable by visualization of the functional

w.
For any Zo the hyperplane determined by the Lagrange multiplier for the
problem
minimize f(x)
subject to G(x) Zo ,
xen

222

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

is a support hyperplane at ro(zo), and this hyperplane serves as a lower

bound for ro as
in Figure 8.3. This observation produces the
following sensitivity theorem.
!

w(::)

Figure 8.3 Sensitivity

Theorem 1. Let f and G be convex and suppose Xo.,

Xl

are solutions to the

problems
minimize f(x)
subject to x e nand G(x)::;; zo.

G(x)::;;

Zl.

respectively. Suppose zri. zi are Lagrange multipliers corresponding to these

problems. Then

Proof The Lagrange multiplier zri makes

!(xo)

+ (G(xo) -

::;;!(x) + (G(x) - ZO,

Zo,

for all x en. In particular, setting x

zo,
= 0 yields

!(xo) -

!(Xl) ::;;

<G(Xl) -

A similar argument applied to Xl'

= Xl and taking account of (G(xo) Zo, zri> ::;; (z 1 - Zo,

zT produces the other inequality. I

A statement equivalent to that of Theorem 1 is that

ro(z) - ro(zo)

(zo - z,

Hence, if the fun.;;;tional ro is Frechet differentiable at Zo, it follows (see

Prql>l....,. 9) thAt

*
ro '(Zo) -- -zo.

8.6

DUALITY

223

Therefore, the Lagrange: multiplier Z6 is the negative of the first-order

sensitivity of the optimal objective with respect to the constraint term Zo

8.6 Duality
Consider again the basic convex programming problem:
minimize f(x)
subject to G(x)

0,

XEn

where/, G and n are convex. The general duality principle for this problem
is based on the simple geometric properties of the problem viewed in R x Z.
As in Section 8.3, we define the primal functional on the set r
(1)

w(Z)

= inf {f(x)

: G(x)

z, X En}

and let flo = w(O). The duality principle is based on the observation that (as
illustrated in Figure 8.4) flo is equal to the maximum intercept with the

--tI

w(z)

Figure 8.4 Duality

vertical axis of all closed hyperplanes that lie below w. The maximum
intercept is, of course, attained by the hyperplane determined by the
Lagrange mUltiplier of the problem.
To express the above duality principle analytically, we introduce the
dual functional <p corresponding to (1) to be defined on the positive cone in
Z* as

(2)

<p(Z*)

=:

inf [f(x)
xeO

+ <G(x), z*) J.

In general, cp is not finite throughout the positive cone in Z * but the region
where it is finite is convex.

224

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

Proposition 1. The dual functional is concave and can be expressed as

q>(z*) = inf [w(z)

(3)

:er

+ (z, z*)].

Proof The concavity of q> is easy to show and is left to the reader. For
any z* () and any z e r, we have

q>(z*) = inf (f(x) + (G(x), z*)J

;S;

inf {f(x)

+ (z, z*) : G(x);S; z, X E O}

xell

= w(z)

+ (z, z*).

On the other hand, for any

f(xl)

+ (G(x 1), z*>

Xl

EO we have, with

inf {f(x)
= W(Zl)

= G(Xl),

+ (Zl> z*) : G(x);S; Zl' x eO}

+ (Zl> z*)

and hence
q>(Z*)

Zl

inf [w(z)

zer

+ (z,

z*)].

Therefore, equality must hold in (3).

The element (1, z*) e R x Z * determines a family of hyperplanes in

R x Z, each hyperplane consisting of the points (r, z) which satisfy
r + (z, z*) = k where k is a constant. Equation (3) says that for k = q>(z*)
this hyperplane supports the set [00,
the region above the graph of w.
Furthermore, at z = () we have r = q>(z*); hence, q>(z*) is equal to the
intercept of this hyperplane with the vertical axis. It is dear then that the
dual functional is precisely what is needed to express the duality principle.
Referring to Section 7.10 we see that q> is essentially the conjugate
functional of w. Unfortunately, as is often the case with theories that are
developed independently and later found to be intimately linked, there is a
discrepancy in sign convention between the dual functional and the conjugate functional but the essential concepts are identical.
In view of the above discussion the following result establishing the
equivalence of two extremization problems-the minimization of a convex
functional and the maximization of a concave functional-should be selfevident.

n,

Theorem 1. (Lagrange Duality) Let f be a real-valued convex functional

defined on a convex subset n of a vector space X, and let G be a COnvex

mapping of X into a normed space Z. Suppose there exists an Xl such that
G(Xl) < e and that ito = inf {f(x) : G(x) ;S; e, x e n} is finite. Then

(4)

inf [(x)
(1(X) SCI

= max <p(z*)

xd}

and the maximum on the right is achieved by some

8.

8.6

If the infimum on the left is achieved by some Xo

DUALITY

225

fl, then

<G(xo) , Z6) = 0
and Xo minimizes f(x)

+ <G(x),

x E n.

Proof For any z* 2'! 8 we have

inf [J(x)

xen

+ <G(x), z*)]

inf [J(x)

S;;

xen

+ <G(x), z*)]

G(x) sO

s;; inf f(x) == J.lo

xen

G(x) so

Therefore, the right-hand side of the equation in the theorem statement is

less than or equal to Ito. However, Theorem 1, Section 8.3, establishes the
existence of an element Z6 which gives equality. The remainder of the
theorem statement is givlen in Theorem 1, Section 8.3. I
Since w is decreasing, w(8) ::s; w(z) for all z::S; 8; hence, an alternative
symmetric formulation of (4) which emphasizes the duality (but which is
one step removed from the original problem), is
min w(z) = max cp(z*).

(5)

zso

z*<!O

As a final note we observe that even in nonconvex programming problems the dual functional can be formed according to (2). From the geometric interpretation of cp in terms of hyperplanes supporting [w, r],
it is clear that cp formed according to (2) will be equal to that which would
be obtained by considering hyperplanes supporting the convex hull of
[w, r]. Also from this interpretation it follows (provided cp(z*) is finite
for some z* 2'! 8) that for any programming problem
max cp(z*) ::s; min w(z)

(6)

z*<!O

zso

and hence the dual functional always serves as a lower bound to the value
of the primal problem.

Example 1. (Quadratic Programming) As a simple application of the

duality theorem, we calculate the dual of the quadratic program:
minimize lx'Qx - b'x
subject to Ax S;; c
where x is an n vector to be determined, b is an n vector, Q is an n x n
positive-definite symmetric matrix, A is an m x n matrix, and c is an m
vector.
Assuming feasibility (i.e., assuming there is an x satisfying Ax < c), the
problem is equivalent to

(7)

max min Hx'Qx - b'x

).2!O

+ A.'[Ax - c]}.

226

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

The minimization over x is unconstrained and is attained by

= Q -l(b -

A').).

Substituting this in Problem (7), the problem becomes

max {-tA'PA. - A'd - tb'Q-1b},
where

Thus the dual is also a quadratic programming problem. Note that the dual
problem may be much easier to solve than the primal problem since the
constraint set is much simpler and the dimension is smaller if m < n.

Example 2. (Quadratic Programming) Consider now the quadratic programming problem:

minimize tx' Qx - b' X
subject to Ax ::s;; c,
where Q is assumed to be only positive semidefinite rather than positive
definite. The dual is again
(8)

max min {tx'Qx - b'x + A,[Ax - c]},

x

but now the minimum over x may not be finite for every A. c;
minimization over; is finite if and only if there is an x

e. In fact, the

Qxb+A'A.=e
and all x's satisfying this equation yield the minimum. With this equation
substituted in (8), we obtain the dl.!al in the form
maximize - A' c - tx' Qx
subject to Qx - b + A'A. = e,
where the maximum is taken with respect to both A and x.

8.7 Applications
In this section we use the theory developed in this chapter to solve several
allocation and control problems involving inequality constraints.
Although the theory of convex programming developed earlier does not
require even continuity of the convex functionals involved, in most problems the functionals are not only continuous but Frechet differentiable. In
such problems it is convenient to express necessary or sufficient conditions

8.7

APPLICATIONS

227

in differential form. For this reason we make frequent use of the following lemma which generalizes the observation that if a convex funct,ion
f on [0, ex:: achieves a minimum at an interior point Xo , then !'(xo) = 0,
while if it achieves a minimum at Xo = 0, then !'(xo) ;:::: 0; in either case
xo!,(x o) = 0.

Lemma 1. Let f be a Frechet differentiable convex functional on a real

normed space X. Let P be a convex cone in X. A necessary and sufficient
condition that Xo E P minimize f over P is that
(1)

(jf(x o; x) ;:::: 0

(2)

of(xo; xo) = 0.

all x

Proof Necessity: If Xo minimizes/, then for any x

P we must have

Hence

(3)
Setting x = xo/2 yields

(4)
while setting x = 2xo yields

(5)

Together, equations (3), (4), and (5) imply (1) and (2).
Sufficiency: For xo, x E P and < IX < I we have

or
1
f(x) - f(xo) ;:::: - [f(xo
IX

+ IX(X -

- f(xo)].

xo

As IX -+ +, the right sidle of this equation tends toward of(xo; x - xo);

hence we have

f(x) - f(xo) ;:::: of(xo; x - Xo).

(6)

If (1) and (2) hold, then of(xo; x - Xo) ;::::

from (6)
f(x) - f(xo) ;::::
for all x

P.

for all x

P and, hence,

228

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

Note that the Frechet differentials can be replaced everywhere by

Gateaux differentials in the lemma, provided the Gateaux differentials
are linear.
In many of the applications considered in this section, the unknown
vector x is constrained to be positive so the constraint set n is a cone. In
these cases, Lemma 1 is used to express the condition that/ex) + (G(x), z*)
is minimized.
Example 1. (Optimal Control) Consider a system governed by the set of
differential equations

(7)

x(t) = A(t)x(t)

+ b(t)u(t),

where x(t) is an n x 1 state vector, A(t) is an n x n matrix, bet) is an n x 1

distribution matrix, and u(t) is a scalar control.
Given the initial state x(to), we seek the control Uo minimizing
(8)

1-

J=

I u (t) dt,
tl

to

while satisfying the terminal inequalities

X(t1)

c,

where c is a fixed n x 1 vector and t1 ;?! to is fixed. This problem might

represent the selection of a thrust program for a rocket which must exceed
in a given time or the selection of a
certain altitude and velocity
production program for a plant with constraints on the total amount
produced in a given time.
We can write the solution to equation (7) in the form
(9)

x(t 1 ) == <1>(t 1 , to)x(t o) +

tI

I <1>(t1' t)b(t)u(t) dt,

to

where <1> is the fundamental solution matrix of the corresponding homogeneous equation. We assume <1>(t1' t) and b(t) to be continuous. The
original problem can now be expressed as: minimize

subject to
(10)

where d = c - <1>(t1> to)x(to) and K is the integral operator defined as

ll

Ku

= f <1>(tl' t)b(t)u(t) dt.

to

8.7

APPLICATIONS

229

This is a convex programming problem defined on, say, L 2 [t O , t1 ] with

the constraint space being finite dimensional. Using the duality theorem,
this infinite-dimensional problem can be reduced to a finite-dimensional
one. 1
Denoting the minimum of (8) under the constraints (10) by 110' the
duality theorem gives
110 == max min {J(u)
u

).,,0

+ X(d -

Ku)},

where A is an n x 1 vector. More explicitly,

(11)

110 = max min t[-!-u 2 (t) - A'(J>(t 1 , t)b(t)u(t)] dt

).,,0

+ Xd

to

and hence
110 = max A'QA

(12)

).,,0

+ It'd,

where

Q=

-1-

f (J>(t
t,

j ,

t)b(t)b'(t)(J>'(tj> t) dt.

to

Problem (12) is a simple finite-dimensional maximization problem. Once

the solution 11.0 is determined, the optimal control uo(t) is then given by the
function that minimizes the corresponding term in (11). Thus
uo(t) = Ao'(J>(tl, t)b(t).

The Lagrange multiplier vector 11.0 has the usual interpretation as a sensitivity. In this case it is the gradient of the optimal cost with respect to the
target c.
Example 2. (Oil Drilling) A company has located an underground oil
deposit of known quantity IX. It wishes to determine the long-term oil
extraction program that will maximize its total discounted profit.
The problem may
formulated as that of finding the integrable function x on [0, (0), repre:senting the extraction rate, that maximizes

Ia<XlF[X(t)]V(t) dt

subject to
{"'x(t) dt::::;;

IX,

x(t)

o.

1 A problem almost identical to this is solved in Section 7.12 by use of the Fenchel
duality theorem.

230

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

F[x] represents the profit rate associated with the extraction rate x, By
assuming diminishing marginal returns, the function F can be argued to be
strictly concave atld increasing on [0, co) with F[O] = O. We assume also
that Fhas a continuous derivative. The" function vet) represents the discount
factor and can be assumed to be continuous, positive, strictly decreasing
toward zero, and integrable on [0, co).
To apply the differentiability conditions, we first consider the problem
with the additional restriction that x be continuous and x(t) = 0 for
t T where T is some fixed positive time. The constraint space Z then
corresponds to the inequality g x dt :s; (X and is thus only one dimensional.
The Lagrangian for this problem is

L(x, A) = (F[x(t)]V(t) dt - A[(x(t) dt -

(13)

(X

where the minus sign is used because we are maximizing a concave

functional. In this case we want xo(t) 0 and Ao such that
min max L(x, A) = L(xo, Ao).
),:1;0

X:1;O

In view of the concavity of F and Lemma 1, maximization of L(x, Ao) over

x (J is equivalent to
T

fa {Fx[xo(t)]v(t) -

(14)

Ao}X(t) dt :s; 0

for all x(t) ::::: 0, and

T

fa {Fx[xo(t)]v(t) -

(15)

AO}Xo(t) dt == 0

where Fx is the derivative of F;

It is clear that the solution will have Ao > 0; therefore we seek xo(t), AO'
J.l.(t) satisfying, for t e [0, T],

(16)

xo(t)

0,

Ao > 0,

J.l.(t)

0,

J.l.(t)xo(t)

= 0,

faxo(t) dt

= <x,

(17)
Since Fx is condnuous and strictly decreasing, it has an inverse Fx -1 which
is continuous and strictly decreasing. From (16) and (17) we see immediately that on any interval Nhere xo(t) > 0 we have xo(t) = Fx -l{Ao/t>(t)}
which is decreasing. Since we seek a continuous x, it follows that the
maXimizing xo, if it exists, m'lst have the form

to:::;; t:::;; T.

8.7

APPLICATIONS

231

In this solution we may have to == T or, if to < T, continuity demands that

Fx- 1 {Ao/V(to)}
Defining

= 0 and heillce that Ao =

J(t ) =
o

ft0oFx -

Fx[O]v(to).

{F x[O]v(to)} dt
v(t)
,

one can show (we leave it to the reader) that J is continuous and that lim
to"" 0

J(t 0) = 0 and lim J(t 0) = co. Hence, let us choose to such that J( t 0) = 0(, Then
to-t- 00

for T> to, the solution

F; 1 {Fx[O]V(to)
xo(t) =

v(t)

o
o ::; t ::; to
to ::; t ::; T

satisfies (16) and (17). Since the Lagrangian is concave, these conditions
imply that Xo is optimal for the problem on [0, T].
Given 8 > 0, however, any function y(t) for which
( ' y(t) dt =

IX

< co

can be approximated by a continuous function x(t) vanishing outside a

finite interval in such a way that
( ' {F[y(t)] - F[x(t)]} v(t)dt < e
o

and hence xo(t) is optimal with respect to this larger class.

The Lagrange multiplier Ao in this case is the derivative of the total discounted profit with respec:t to IX.

Example 3. (The Farmer's Allqcation Problem) A farmer produces a

single crop such as wheat. After harvesting his crop, he may store it or sell
and reinvest it by buying additional land and equipment to increase his
production rate. The farmer wishes to maximize the total amount stored
up to time T.
Letting
x 1 (t) = rate of production
X2(t) = rate of reinvestment
X3(t) = rate of storage

234

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

changing the constraint

to
Xz(t)

s Xl(t) + b(t -

't')

where b is the delta function. According to the sensitivity result (which is

exact for this problem), the corresponding change in the objective, total
storage at t = T, is
ill

= (b(t -

't') dv(t) = dv(t)

dt t"'!

Thus dvldtlt=! represents the value, in units of final storage, of an opportunity to reinvest one unit of storage at time 't'.
Example 4. (Production Planning) This example illustrates that the

interior point assumptions in the Lagrange theorem can be important in

even simple problems. Consider the following problem:
1

minimize

1- fo r2(t) dt
set),

subject to i(t) = ret),

given z(O) > O.

This problem may be regarded as a production-planning problem with
no inventory costs but with known demand d(t) = s(t) (see Chapter 1).
Thus at any t we must have
z(o)

or, equivalently, z(t)

+ {r('t') d't'

{d('t') d't'

set). We solve this problem for the specific case:

z(O)

=!

set)

= (1

2t

O:::;;tS1-

-1<t:::;;1.

Let us consider r as a member of a space X. The choice of X is somewhat

free but we wish to choose it so that the corresponding constraint space Z
can be taken to be CEO, 1]. In other words, we want a class of r's such that
z(t) = z(O)

+{

reT) d't'

will be continuous. The choice X = L 1[O, IJ is awkward because

J& r2 dt

does not exist for all r E L 1 [O, 1]. The choice X = CEO, 1] is feasible but,

8.7

235

APPLICATIONS

as we see from Example 3, there may be no solution in this space. Two other
possibilities are: (1) the space of piecewise continuous functions and (2) the
space of bounded measurable functions. We may define the supremum
norm on these last two spaces but this is unnecessary since Lemma 1 is
valid for linear Gateaux differentials as well as Frechet differentials and
this example goes through either way. Let us settle on piecewise continuous
functions.
The Lagrangian for this problem is

L(r, v) =
where v E BV[O,
Vo

e.

IJ

f r2(t) dt + f [set) 1

z(t)] dv(t),

and is nondecreasing. We seek a saddle point ro

e,

A more explicit formula for the Lagrangian is

L(r, v)

=t

fo r2(t) dt

fo r2(t) dt

fa fa r(-c) dT dv(t)
1

+ fo [set) -

z(O)J dv(t) -

+ fa [set) -

z(O)J dv(t)

+ fo r(t)v(t) dt Using Lemma 1 to minimize L(r, vo) over r

v(l)

dt.

e, we require that

+ vo(t) - vo(l) ; 0
ro(t)[ro(t) + vo(t) - vo(l)J = 0
ro(t)

for all t
for

I'

t.

Since set) S z(t) for all t, in order to maximize L(ro, v) we require that
vo(t) varies only for those t with z(t) = set).
The set offunctions satisfying these requirements is shown in Figure 8.5.
Note that v is a step function and varies only at the single point t = 1. If a
Lagrangian of the form

L(r, A)

=t

fo r2(t) dt

+ fo [set) -

Z(t)]A(t) dt

were considered, no saddle point could be found unless functions A. having

delta functions were allowed.
The economic
of the Lagrange multiplier can be obtained
rather directly. Suppose that the demand function were changed by lid(t).
Letting liJ and lis be the corresponding changes in the total production cost
and in the function s,
have to first order
1

I:iJ == fo l:is(t) dv(t)

= - fo Mv(t) dt + I:is(l)v(l)'-l:is(O)v(O).

236

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

J* ) o

r(t)

;;. t

r
I

"2

-11-----1

vo(t)

Figure 8.5

The solution to a production problem

Since As(O) = 0, v( 1) = 0, the boundary

vanish and hence

AJ = -

fo Ad(t)v(t) dt.

Thus - vet) is the price per unit that must be charged for small additional
orders at time t to recover the increased production cost. Note that in the
particular example considered above this price is zerO for t > t. This is
because the production cost has zero slope at r == 0.

8.8 Problems
1. Prove that the subset P of Lp[a, b], I

$,

p <

00,

consisting of functions

nonnegative almost everywhere on [a, b], contains no interior point.

2. Prove that the subset P of C[a, b] or L\X)[a, b], consisting of nonnegative functions on [a, b], contains interior points.
3. Let P be the cone in C[a, bJ consisting of nonnegative (continuous)
functions on [a, b]. Show that p'iIJ consists of those functions v of
bounded variation on [a, bJ that are non decreasing.
4. Show that the positive cone p'iIJ in X* is closed.
5. Show that the sum of two convex mappings is also Convex.
6. A cone is said to be pointed if it contains no one-dimensional subspace.
If X is a vector space with positive cone P, show that x ;;:: () and x$,(}
imply x = () if and only if P is pointed.
7. Let Po = inf {I(x) : x n, G(x) $, 0, H(x) = O}, where all entities are
defined as in Theorem 1, Section 8.3, with the addition that H(x) =
Ax + Yo (where A is linear) is a map of X into the finite-dimensional
normed space Y. Assume that Po is finite, that () e Y is an interior
pointof{y E Y:H(x)= y forsomexen}, and that there exists Xl ensuch
that G(xl) < 0, H(x t ) == e. Show that there is
e, such that
Po

= inf {I(x) + <G(x), z&) + <H(x),y&) : x e il}.

8.8

PROBLEMS

237

8. Let I be a functional defined on a normed space X. An element

e X * is said to be a subgradient of I at Xo if

f(x) - f(xo) (x - Xo,

for all x eX. Show that if/has a gradient at xo, any sub gradient is in
fact equal to the gradient.
9. Show that in the sense of Theorem 1, Section 8.6, the dual to the linear
programming problem:
minimize b' x
subject to Ax ;:::: c,

x;::::

e,

where x is an n vector, b is an n vector, c is an m vector, and A is an

m x n matrix is the: problem:
maximize c' A
subject to A' A. ::5: b,
10. Derive the minimum norm duality theorem from the Lagrange duality
theorem.
11. Derive necessary conditions for the problem: minimize I(x) , subject to
x E n, G(x) ::5: z, z e: A, where both x and z are variable. Assume that
all functions and sets are convex.
12. Assuming that the interior point condition is satisfied, show that the
dual of a convex-programming problem is also a convex-programming
problem. Show that the dual of the dual, with variables restricted to X
rather than X**, is, in some sense, the primal problem.
13. Let G be a convex mapping from n c: X into a normed space Z and
assume that a positive cone having nonempty interior is defined in Z.
Show that the two regularity conditions are equivalent:
(a) There is an Xl E n such that G(xl ) < e.
(b) For every z* ;:::: 0, z* ::F there is an x En such that (G(x), z*) <0.
14. Ifwe let
z(t) = production rate
d(t) = demand rate
yet) = inventory stock,
a simple production system is governed by the equation

e,

yet) = yeO)

+ fo[z(r) -

d(r)] dr.

Find the production plan z(t) for the interval 0::;; t::;; T satisfying
z(t) ;:::: 0, y(t) ;:::: 0, 0::;; t::5: T, and minimizing the stlm of the production
costs and inventory costs

f [tz (t) + h . yet)] dt.

T

J=

238

GLOBAL THEORY OF CONSTRAINED OPTIMIZATION

Assume'that
d(t) == 1,

2y(O)h> 1,

Answer:

z(t) =

2h

o
(h . (t -

+ y(O) > T,

y(O)

< T.

t1)

where
tl

== T -

[T - y(O)].

IS. Repeat Problem 14 for

d(t)

== 1,

1
2h + y(O) S T,

2y(O)h> I,

y(O) < T.

It is helpful to define
tl

== y(O) - - , t z == y(O) + -.
2h
2h
REFERENCES

8.1-4. Lagrange multipliers for problems having inequality constraints were

fitst treated explicitly by John [76] and Kuhn and Tucker [91]. The approach
presented in this chapter, which requires,no differentiability assumptions, was
developed by Slater [142},md extended to infinite-dimensional spaces by
Hurwicz [75]. In fact, our presentation closely follows Hurwicz. An excellent
treatment of convex programming in finite-dimensional spaces is contained in
Karlin [82]. For some interCbting extensions, see Arrow, Hurwicz, and
Uzawa [12]. A similar theory can be developed for quasi-convex programming.
See Luenberger [100].
8.5-6. The sensitivity property was observed by Everett [48]. The general
duality theorem evolved from Dom [42], Wolfe [155], Hanson [69], Huard
[74],.. and Mangasarian [102]. Also see Ritter [125] .and Rissanen [124].
8.7. Much of the solution to the oil-drilling problem is based on an analysis
by Karlin {83, pp. 210-214] who obtains a solution derived from the NeymanPearson Lemma. The farmer's allocation problem was devised by Phillip
Wolfe and David Gale as an example of continuous linear programming.
Leonard Berkovitz noted an easy solution using control theory (see Section
9.6). Problems such as the production problem were solved using convex
programming theory by Lack [93].
S.S. Problems 14 and 15, together with numerous other cases, are solved in
Arrow, Karlin, and Scarf [13, especially pp. 37-40].

LOCAL THEORY OF
CONSTRAINED OPTIMIZATION

9.1 Introduction

Historically, the local theory of Lagrange multipliers, stated in differential

form, predates the global theory presented in Chapter 8 by almost a
century. Its wider range of applicability and its general convenience for
most problems continue to make the local theory the better known and
most used of the two.
The general underlying principles of the two theories are substantially
the same. In fact the Lagrange multiplier result for inequality constraints
goes through almost identically for both the local and global theories. But
there are some important differences, particularly for equality constraints,
stemming from the fact that the local theory is based on approximations in
the primal space X and hence auxiliary analysis is required to relate these
approximations to the constraint space Z or to Z *, the space in which the
Lagrange multiplier is defined. For this reason, adjoint operators playa
significant part in the development of the local theory since they enable
us to transfer results in X * back to Z *.
For problems with equality constraints only, the nicest result available
is the general Lagrange multiplier theorem established by means of an
ingenious proof devised by Liusternik. This theorem, which is by far the
deepest Lagrange multiplier theorem in this book, is proved in Section 9.3.
The difficult analysis underlying the theorem, however, is contained in a
generalized inverse function theorem proved in Section 9.2.
In Section 9.4 an analogous theorem is given for optimization problems
subject only to inequality constraints. The proof of this result is similar to
the proof of the global Lagrange multiplier theorem of Chapter 8.
Following these general theorems on Lagrange multipliers, there are two
sections devoted to optimal control theory. To some extent, this topic can
be treated as a simple application of the general Lagrange multiplier
theory. The structure of control problems, however, is worthy of special
239

240

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

attention because additional results can be. derived for this important class
of problems and additional insight into Lagrange multipliers is obtained
from their a,nalysis.
LAGRANGE MULTIPLIER THEOREMS

9.2 Inverse hnction Theorem

In this section we prove a rather deep generalization of the classical inverse
function theorem that enables us to derive the generalized Lagrange
mUltiplier theorem in the next section. The inverse function theorem is of
considerable importance in its own right in analysis; some variations of it
are discussed in the problems at the end of this chapter. The proof of the
theorem is quite complex and may be omitted at first reading, but it is
important to understand the statement of the theorem and, in particular,
to be familiar with the notion of a regular point.
Defillition. Let T be a continuously Frechet differentiable transformation
from an open set D in a Banach space X into a Banach space Y. If Xo e: D
is such that T'(x o) maps X onto Y, the point Xo is said to be a regular point
of the transformation T.

Example 1. If T is a mapping from En into Em, a point Xo e En is a regular

point if the Jacobian matrix of T has rank m.
Theorem 1. (Generalized Inverse Function Theorem) Let Xo be a regular
point o/a transformation T mapping the Banach space X into the Banach
.space Y. Then there is a neighborhood N(yo) of the point Yo = T(xo) (i.e., a
sphere centered at Yo) and a constant K such that the equation T(x) = y has a
solution for every y E N(yo) and the solution satisfies Ilx - Xo I :s; K Ily- Yo U

Proof Let Lo = nullspace of T'(x o)' Since Lo is closed, the quotient

space X/Lo is a Banach space. Define the operator- A on this space by
A[x] = T'(xo)x, where [x] denotes the class of elements equivalent to x
modulo Lo. The operator A is well defined since equivalent elements x
yield identical elements y E Y. Furthermore, this operator is linear, continuous, one-to-one, and onto; hence, by the Banach inverse theorem, A
has a continuous linear inverse.
Given y E Y sufficien, 'J close to Yo , we construct a sequence of elements
{L,,} from XILo and a corresponding sequence {gn} with gn E Ln such that
Xo + gn converges to a solution of T(x) = y. For fixed y E Y, let go = () E Lo
ana aeflne tne sequences {Lni and {Un} recursively by
(1)

L" - L,,-1 = A- 1(y - T(xo

+ gn-1))'

9.2

INVERSE FUNCTION THEOREM

241

and from the coset L. select g. such that

Ilg. - g.-lll :::; 2 IlL. - L.- 111

since IlL. - L.-111 = inf Ilg - g.-lll).

(which is possible

Rewriting (1)

geL"

slightly, we have

Ln == A- 1(y - T(xo

+ 9n-1) + T'(XO)Un-l)

and similarly

== A- 1(y - T(xo + gn-2) + T'(X O)g.-2)

L.-1

Therefore,
L. - L.-1

== -A -l(T(xo + g.-l) - T(xo + g.- 2) - T'(Xo)(g.-l - g.-2)).

Define g, =:: fg n - 1 + (1 - t)gn-2. By the generalized mean value inequality

(Proposition 2, Section 7.3) applied to the transformation
r(x)

== - A -l(T(x) - T'(xo)x),

we obtain

Since T' is continuous at x o , given B > 0, there is an r > 0 such that

IIT'(x) - T'(x o) I < ,s for Ilx - xoll < r. Assuming Ilg.-lll < r, Ilgn-211 < r,
we have Ilg,ll < r; therefore (2) implies that

IlL. - Ln- 111 :::; BIIA- 1111Ign_1 - g.-211.

Furthermore,

Ilg. - gn-111 :::; 2IlL. - L.-111 :::; 2BIIA- 1111Ign_1 - g.-211

and hence for sufficiently small

(3)
Since IIg111 :::; 211L111
ciently small we

211A -llllly - Yo II,

it follows that for

lIy - Yo II

suffi-

IIg1 11 < tr.

Thus the conditions for (3) to be valid hold for n = 2 and, in fact, hold for
all n since, by induction, the relation
IIgnll = IIgl

+ (g2

- gl)

$(1 + 1- + ... +
shows that

Ilgnll :::; r for

aU n.

+ ... + (gn -

gn-I)II

211glll5. r

242

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

Thus

for all n and hence the sequence {gn} converges to an element g. Correspondingly, the sequence {Ln} converges to a coset L. For these limits we
have
L = L I- A-1(y - T(xo

+ g

or, equivalently,
. T(xo

+ g) = y;

and

IIgll :S: 211g111:s; 411A- 1 1111y - YolI,

thus K can be taken to be 4I1A- 1 1I. I
9.3 Equality Constraints
Our aim now is to develop necessary conditions for an extremum of/subject
to H(x)
where/is a real-valued functional on a Banach space X and H
is a mapping from X into a Banach space Z.

='e

Lemma 1. Let/achieve a local extremum subject to H(x) = e at the point

Xo and assume that / and H are continuously Frechet differentiable in an
open set containing Xo and that Xo is a regular point 0/ H. Then/'(xo)h = 0
for all h satisfying H'(xo)h = e.
Proof To be specific, assume that the local .extremum is a local
minimum. Consider the transformation T: X -+ R x Z defined by
T(x) = (f(x), H(x. If there were an h such that H'(xo)h = e,f'(xo)h : 0,
then T'(xo) = (!'(xo), H'(xo: X -+ R x Z would be onto R x Z since
H'(xo) is onto Z. By the inverse function theorem, it would follow that for
any B > 0 there exists a vector x and 15 > 0 with IIx - xoll < B such that
T(x) = (f(xo) - 15, e), contradicting the assumption that Xo is a local
'
minimum. I

The above result can be visualized geometrically iIi the space X in terms of
the concept of the tangent space of the consttaint surface. By the tangent
space at xo, we mean the setof all vectors h for which H'(xo)h = e (Le., the
nUllspace of H'(xo' It is a subspace of X which, when translated to the
point xo, can be regarded as the tangent of the surface N = {x : H(x) = 8}
near Xo . An equivalent statement to that of Lemma 1 is that/ is stationary
at Xo with respect to variation in the tangent plane. This is illustrated in
Figure 9.1 where contours of/as well as the constraint surface for a single

9.3

EQUALITY CONSTRAINTS

243

functional constraint h(x) == 0 are drawn. The Lagrange multiplier theorem

now follows easily from the duality relations between the range and nullspace of an operator and its adjoint.
\
\

\
\

ex\

\
\
\
\

Tangent space
Figure 9.1

lI(x)

=e

Constrained optimization

Theorem 1. (Lagrange Multiplier) If the continuously Frechet differentiable

functional f has a local extremum under the constraint H(x) = () at the regular
point x o , then there exists an element Z6 E Z * such that the Lagrangian
functionatl
L(x) =f(x) + Z6H(x)
is stationary at x o , i.e.,!,(xo)

+ Z6H'(Xo) = e.

Proof From Lemma I it is clear that!'(xo) is orthogonal to the nullspace of H'(x o)' Since, however, the range of H'(xo) is closed, it follows
(by Theorem 2, Section 6.6) that

!'(xo) E PA [H'(xo)*J.
Hence there is a

Z6 E Z * such that
f'(x o) == - H'(x o)*z6

or, in an alternative notation,

The second term of this last expression is to be interpreted in the usual
sense of composition of linear transformations. I
We may immediately rephrase this result to include the case when the
constraint is not regular.
1 When considering the local theory the necessary conditions are written !'(xo) +
zo H'(xo) = 6. Therefore we usually ,write the Lagrangian in the form indicated in the
theorem statement rather than L(x) =I(x) (H(x), Zo

*>.

244

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

Corollary 1. Assuming all the hypotheses of TheQrem I with the exception

-7TiQ:t the range 0/ H'(x o) is closed but perhaps not onto, there exists a nonzero
element (ro,
e R x Z* such that the/unctional
I
rof(x) +
is stationary at Xo .
Proof. If Xo is regular, we may take ro == 1 and use Theorem 1. If Xo
is not a regular point, let M = !J4(H'(xo There is a point z E Z such that
inf

meM

liz - mil > O.

and hence by Theorem 1, Section 5.8, there is Ii

by Theorem 1, Section 6.6, MJ.=.K(H'(x o)*), this
satisfies the requirements of the corollary. I

e.

e.MJ.,
:F Since
together with ro == 0

Reviewing the arguments that led to the Lagrange multiplier theorem, it

should be noted that the difficult task, requiring the regularity assumption,
was to show that at an extremum/is stationary with respect to variations in
the tangent pl"ne; in other words, justifying that a nonlinear constraint can
be replaced by a lineariud version of it. From this result it is clear that
!'(xo), the gradient of f, must be orthogonal to the tangent space. The
Lagrange multiplier theorem then follows from the familiar adjoint relations.
A slightly different but useful interpretation yields a more direct identi
fication of the Lagrange multiplier. In Figure 9.1 the
is described
by a single functional equation h(x) = 0; it is clear on geometric grounds
that, at the optimum, the gradient of h must be parallel to the gradient off.
Thus/'(xo) + h'(xo)z = 0, where z is a scalar: the Lagrange multiplier. A
similar figure can be drawn in three dimensions for the case where H
consists of two functionals h1, hz . An example is shown in Figure 9.2. For
optimality the gradient of/must lie in the plane generated by h1' and h/;
hence !'(xo) + Z1hl'(XO) + Zz h/(xo) = 6.

Figure 9.2 Two constraints

9.3

EQUALITY CONSTRAINTS

245

Example 1. (Constrained Problems in the Calculus of Variations) We

consider the problem of finding x E Dn[fo, fa, the space of n-vector
functions on [to,
possessing continuous derivatives, having fixed end
points x(to), X(t1), which minimizes

ta

t,

(1)

J =

x, t) dt

f(x,

to

while satisfying the constraint

cp(x, t) =

(2)

O.

Both f and cp are realvalued functions and are assumed to have continuous partial derivatives of second order. Since the end points are fixed,
we restrict OUr attention to variations lying in the subspace Xc:: Dn[fo,
consisting of those functions that vanish at to and 11 As demonstrated in
Section 7.5, the Frechet differential of J is

ta

(3)

M(x; h) =

[fx(x,

to

x, t)h(t) + fx(x, x, t)h(t)Jdt.

The function cp can be considered as a mapping H from X into Y where Y is

the subspace of D[to, 11] consisting of those functions vanishing at 10 and
fl. The Frechet differential of H is

(4)

bH(x; h) = CPx h(/).

We assume that along the minimizing curve the n partial derivatives of cp

(with respect to the n components of x) do not all simultaneously vanish at
any point in [to,
In this case the Frechet differential (4), evaluated at the
minimizing x, defines a bounded linear mapping from X onto Y. To verify
this we note that, since cp has continuous second partial derivatives, (4)
defines an element of Y for each hEX. Also, given Y E Y, the selection

la.

(where CPx is represented as an n x 1 row vector and CPx' is its transpose)

satisfies CPxh = y. Thus, according to Theorem 1, there exists a Lagrange
multiplier for the problem. The multiplier is in this case an element of Y
which in gener,al can be represented in the form

f y(t) dz(t)
t,

<y, y*) =

to

246

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

where Z e NBV[to, 11]' However,

it can be shown that the multiplier for

this problem actually takes the special form

. f'l y(t)A(t) dt

(y, y*) =

to

for some continuous function A. Hence the necessary conditions become

(5)

fix,

x, t) + A(t)cpx(X, x, t) = dtd ft(x, x, t).

Example 2. (Geodesics) Let x, y, z denote the coordinates of an arbitrary

point in three-dimensional space. The distance between two given points
along a smooth arc x(t), y(t), z(t) (parametrized by t, tl :::; t :::; t 2) is

Given a smooth surface defined by cp(x, y, z) = 0 and two points on this

surface, the arc of minimum length lying inthe surface cp(x, y, z) = 0 and
connecting the two points is called the geodesic between the points. If
ICPxl + Icp,,1 + ICPzl =1= 0 along the geodesic, the method of Example 1 may be
applied.
Writing the Lagrangian in the form

the equations corresponding to equation (5) are

d t J.2

x
'2

'2

x +y +z

dt J

+ A(t)cpx = 0

x2 + y2 + Z2 + A{t),I,.'1'" --

Z
.2

dtJ x

'2

'2

+y +z

+ A(t)cpz =0

which, together with the constraint; can be solved for the arc.
In the special case of geodesics on a sphere, we have

cp(x, y, z) = x 2 + y2

CPx = 2x,

+ Z2 -

Cp" = 2y,

r2 = 0

cpz = 2z.

9.4

INEQUALITY CONSTRAlNTS (KUHN-TUCKER THEOREM)

247

Let the constants a, b, c be chosen so that the plane through the origin
described by
ax

+ by + cz =

contains the two given points on the sphere.

Corresponding to the extremal arc x(t), y(t), z(t), let p(t)
by(t) + cz(t). Then it can be seen that p satisfies

-dt

= ax(t) +

J x2 +py2 + ;i2 + 2A(t)p(t) = 0

and P(t1) =p(t z) = O. It follows from the uniqueness of solutions to

differential equations, that p(t) == 0 and hence that the geodesic lies in the
plane. Therefore, the geodesic is a segment of a great circle on the sphere.
The explicit dependence on t is, of course, somewhat arbitrary since any
parametrization of the arc is allowed.
9.4 Inequality Constraints (Kuhn-Tucker Theorem)
In this section we derive the local necessary conditions for the problem
minimize f(x)
{subject to G(x)

(I)

e,

wherefis defined on a vector space X and G is a mapping from X into the

normed space Z having positive cone P.
To see how the Lagrange multiplier technique can be extended to
problems of this type, consider a problem in two dimensions having three
scalar equations gj(x) 0 as constraints. Figure 9.3a shows "the constraint
it is assumed that the minimum occurs at a point Xo
region. In 9.3b,
in the interior of the region, it is clear thatf'(x o) = O. In 9.3c, where it is
assumed that the minimum occurs on the boundary gl (x)' = 0, it is clear
thatf'(x o) must be orthogonal to the boundary and point inside. Therefore,
in this case, f'(x o} + Alg1'(XO) = (J for some A1 O. Similarly, in 9.3d,
where it is assumed that the minimizing point Xo satisfies both gl(xO) = 0
and gz(x o} == 0, we must have f'(x o) + A1g1 '(x o) + A2 g2'(X O) = (J with A1
0, A2 O. All of these cases can be summarized by the general statement

where A*
and Ajgj(Xo) == 0, i = 1,2,3. The equality Ajgj(Xo) = 0 merely
says that if gi(X O) < 0, then the corresponding Lagrange multiplier is absent
from the necessary condition.

248

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

By considering various positive cones (such as P = {O}) or by considering

constraints of the form H(x)
-H(x) 0, problem (1), as stated.
includes minimizat\on problems having equality constraints. It is therefore
clear that a general attack on-roblem (1) would be at least as difficult to
carry through as the corresponding attack on problems having only
equality constraints. To avoid these difficulties, our approach excludes the
possibility of equality constraints. As a consequence of this restriction, the
results for problem (1). although analogous to those for problems having
equality constraints, are far easier to obtain. This approach closely parallels
the development of the global Lagrange multiplier theorem of Section 8.3.

e.

(b)

(c)

(d)

Figure 9.3 Inequality constraints

Defilfition. Let X be a vector space and let Z be a normed space with a

positive cone P having nonempty interior. Let G he a mapping G : X - t Z
which has a Gateaux differential that is linear in its increment. A point
Xo E X is said to be a regular point of the inequality G(x) :s: e if G(xo) :s: 0
and there is an hEX such that G(xo) + c5G(xo; h) < e.
Thill d.Minitio1"l

Lf 11

J'e81llllJ' "oint is a natural a114108 to the i11terior point

condition employed for im'luality constraints in the global theory

(Theorem 1. Section 8.3). Note that the definition excludes the possibility

9.4

INEQUALITY CONSTRAINTS (KUHN-TUCKER THEOREM)

249

of incorporating equality constraints by reducing the cone to a point or by

including a constraint and its negative.
The regularity condition essentially eliminates the possibility of the
constraint boundary forming a cusp at a point. An example where the
regularity condition fails is shown in Figure 9.4 where gl (x) = - X2 ,

Figure 9.4

The regularity condition violated

g2(X) = -Xl' g3(X) = X2 + (Xl - 1)3. The point Xl = 1, X2 = 0 is not

regular since the gradients of g2 and g3 point in opposite directions there.

Theorem 1. (Generalized Kuhn-Tucker Theorem) Let X be a vector space

and Z a normed space having positive cone P. Assume that P contains an
interior point.
Let f be a Gateaux dU.rerentiable real-valued functional on X and G a
Gateaux durerentiable mapping from X into Z. Assume that the Gateaux
dWerentials are linear in their increments. l Suppose Xo minimizes f subject to
G(x)::s;; and that Xo is a regular point of the inequality G(x) ::s;; Then there
is a
E Z *,
;;:; such that the Lagrangian

e.

f(x)
is stationary at Xo; furthermore,

+ <G(x),

<G(xo),

o.

Proof In the space W == R x Z, define the sets

A = {(r, z) : r ;;:; of(xo; h), z
B

{(r, z) : r

0, z

G(xo)

+ oG(xo; h) [or some hEX}

O}.

1 As discusseed in Problem 9 at the end of this chapter, these hypotheses can be somewhat weakened.

250

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

The sets A and B are obviously convex; in fact, both are convex

cones
although A does not necessarily have its vertex at the origin. The set B
contains interior points since P does.
.
The set A does not contain any interior points of B because if (r, z) e A,
with, < 0, z < 9, then there exists heX such that

()f(xo; h) < 0,
The point G(xo) + ()G(x o ; h) is the center of some open sphere contained
in the negative cone N in Z. Suppose this sphere has radius p. Then for
0< oc < 1 the point oc[G(xo) + ()G(xo; h)] is the center of an open sphere of
radius oc' p contained in N; hence so is the point (1 - oc)G(xo)+oc[G(xo) +
()G(xo; h)] = G(xo) + oc . ()G(x o ; h). Since for fixed h
IIG(.Ji)

+ och) -

G(xo) - oc' ()G(xo; h)1I = o(oc),

it follows that for sufficiently s''''Lall oc, G(xo + och) < 8. A similar argument
shows thatf(xo + och) <f(xo) for sufficiently small oc. This contradicts the
optimality of Xo; therefore A contains no interior points of B.
to Theorem 3, Section 5.12, there is a closed hyperplane
() such that
separating A and B. Hence there are '0,
ro . r + (z,

()

for all (r, z) e A

+ (z,

()

for all (r, z) e B.

ro' r

Since (0. 8) is in both A and B, we have () = O. From the nature of B it

0,
8. Furthermore, the hyperplane cannot be
follows at once that
vertical because of the existence of h such that G(xo) + ()G(xo; h) < 8.
Therefore, we take = 1.
From the separation property, we have

'0
'0

()f(xo; h)

+ (G(xo) + ()G(xo; h),

for all heX. Setting h == 9 gives (G(xo),

0 but G(xo) 8,
;;;:= 9
0 and hence (G(xo),
= O. It then follows from
implies (G(xo),
the linearity of the differentials with respect to their increments that
()f(xo; h) + ({)G(xo; h),
== o. I

Example 1. Suppose that X is a normed space (rather than simply a

vector space) and that f and G are Frechet differentiable. Then if the solution is at a regular point, we may write the conclusion of Theorem 1 as
!'(xo)

== 8

<G(xo),

== O.

9.4

INEQUALITY CONSTRAINTS (KUHN-TUCKER THEOREM)

251

Example 2. Consider the n-dimensional mathematical programming

problem
minimize f(x)
subject to G(x) ::s; 0
X"2!O
where x E En, G(x) E Em, ;and f and G have continuous partial derivatives
with respect to the components of x. The constraint for this problem can be
written in partitioned form as

which has E"+m as the constraint space. Assuming satisfaction of the

regularity condition, we have the existence of two Lagrange multiplier
vectors ,.1,0 E Em, Jlo E En with ,.1,0 "2! 0, Jlo "2! 0 such that at the solution Xo
(1)

fixo)

(2)

+ ,.1,0 Gixo) -

Jlo ::;:: 0

,.1,0 G(xo) - Jlo Xo =

o.

Since the first term of (2) is nonpositive and the second term is nonnegative,
they must both be zero. Thus, defining the reduced Lagrangian

L(x, A) == f(x)

+ A'G(x),

the necessary conditions can be written as

Lixo, ,.1,0)

Lx(x o , Ao)x o = 0

L;.(xo, Ao)

L;.Cxo, Ao)Ao = 0

X"2!O

The top row of these equations says that the derivative of the Lagrangian
with respect to Xi must vanish if Xi> 0 and must be nonnegative if Xi = O.
The bottom row says that Aj is zero if the j-th constraint is not active, i.e.,
if the j-th component of G is not zero.

Example 3. Consider the: constrained calculus of variations problem

t

(3)

minimize J =

f f(x, x, t) dt
'

to

(4)

subject to 4>(x, t) ::s;

o.

Here to, t1 are fixed and X is a function of t. The initial value x(to) is fixed
and satisfies (x(to), to) < o. The real-valued functions f and 4> have continuous partial derivatives with respect to their arguments. We seek a
continuous solution xU) having piecewise continuous derivative. We
assume that, along the solution, 4>x ::j:. O.

252

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

We incorporate the fact that x(to) is fixed by restricting attention, in the

variational analysis, to variations in the space X consisting of continuous
functions
at to and having piecewise continuous derivatives on
[to, t1]. We consider the range of the constraint (4) to be C [to, 11]. The
regularity condition is then equivalent to the existence of an heX such that
cp(xo(t), t)

+ CP;r;(XO(t), t) . h(t) < 0

for all t e [to, t 1]. This condition is satisfied since it is assumed that
cp(xo(to), to) < 0 and CP;r;(XO(t), t) :F O.
We now obtain, directly from Theorem 1, the conditions
(5)

f''I0 U;r;h(t) +!;i;h(t)] dt + J''01 CP;r;h(t) d-t(t) = 0

for all heX with h(to) = 0, and

(6)

f'1 cp(x, t) d-t = 0

to

where -t e NBV[to, ttl and is nondecreasing.

Integrating (5) by parts, we have

The function

is bounded on [to, t1] and has at most a countable number of discontinuities. However, with the ex:;;eption of the right end point, M must be
continuous from the right. Therefore,by a slightly strengthened version of
Lemma2, Section 7.5, and by considering (7) for those particular he Xthat
vanish at tl as well as at to, we have
{S}

for t e [to, t 1). If A does not have a jump at tl> then (8) substituted into (7)
yields
!;i;(X,

x, t)1

t=t1

=0

because h(t1) is arbitrary. On the other hand, if -t has a jump at tl> then
(0] ImpliCI5 tl1i:lt

cp(x, t)

''''II

= O.

9.4

CONSTRAINTS (KUHN-TUCKER THEOREM)

Therefore, we obtain in

case the boundary condition

q>(x, t) . fx(x,

(9)

253

x, t) \1=11 = O.

Together (6), (8), and (9) are a complete set of .necessary conditions for the
problem. For a simple application of this result, see Problem 6.
Example 4. As a specific instance ofthe above result, consider the problem
1

minimize J =

fa [x(t) + !x(t)2] dt

subjeict to x(t) set).

Here s is a given continuous function and the initial condition x(O) > s(O) is
given.
Such a formulation might result from considering the problem of maintaining a work force X sufficient to handle a work level s when there is a
linear salary cost and a quadratic cost for hiring and firing.
From (8) we obtain
(10)
t - A(t) - x(t) = c
where we have taken ),(0) = O. Thus initially, while x(t) > set), we have
in view of equation (6)
(11)
x(t) = x(O) + t.
Let us hypothesize that the constraint x(t) set) is never achieved by
equality. Then equation (11) must hold throughout [0, IJ and the terminal
condition (9) implies in this case that x(l) = O. In other words, x(t) is a
parabola with second derivative equal to unity and having horizontal slope
at the right end point.
If to is a point where the solution meets the cO)lstraint, it is clear from
equation (10) that the derivative must not have a positive jump at to; hence,
unless s(to) has a corner at to , x must be tangent to s at to . A typical solution is shown in Figure 9.5.

)0

Fi{;/ure 9.5 Solution to Example 4.

254

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

OPTIMAL CONTROL THEORY

9.5 Basic Necessary Conditions
On an interval [to, t 1] of the real line, we consider a set of differential
equations of the form
(1)

x(t) = !(x(t), u(t,

where x(t) is an n-dimensional" state" vector, u(t) is an m-dimensional

"control" vector, and! is a mapping of En X Em into En. Equation (1)
describes a dynamic system which, when supplied with an initial state
x(to) and a control input function u, produces a vector-valued function

x.
We assume that the vector-valued function! has continuous partial
derivatives with respect to x and u. The class of admissible control functions
is taken to be Cm[to, td, the continuous m-dimensional functions on
[to ,td, although there are other important alternatives. The space of
admissible control functions is denoted U.
Given any u E U and an initial condition x(to), we assume that equation
(Ir-defines a unique continuous solution x(t), t > to. The function x
resulting from application of a given control u is said to be the trajectory of
the system produced by u. The class of all admissible trajectories which we
take to be the continuous n-dimensional functions on [to , t1] is denoted X.
In the classical optimal control problem, we are given, in addition to the
dynamic equation (1) and the initial condition, an objective functional of
the form
(2)

tl

J=

J,to l(x,u)dt

and a finite number of terminal constraints

gi(X(t 1 = c/

i = 1,2, ... , r

which we write in vector form as

(3)

The functions I and G are assumed to possess continuous partial derivatives

with respect to their arguments. The optimal control problem is then that
of finding the pair of funct: Jns (x, u) minimizing J. while satisfying (1)
and (3).
TUI,;II,; arc: a number of
of
problem, many of which
can be reduced to this form by appropriate transformations. For example,

9.5

BASIC NECESSARY CONDITIONS

255

problems in which the obj<!ctive contains a function of the terminal state

or problems in which the trajectoryis constrained to satisfy a finite number
of relations of the form
tk(x, u) dt
to

=d

can be transformed into

form considered above by adjoining additional
state variables (see Problem 13). Problems in which the control variables are
restricted by inequalities such as lu(t)1 ::;; I are discussed in Section 9.6.
When attempting to abstract the control problem so that the general
variational theory can be applied, we discover several alternative viewpoints. Perhaps the most natural approach is to consider the problem as
one formulated in X x U and to treat the differential equation (1) and the
terminal constraint (3) as constraints connecting u and x; we then apply
the general Lagrange multiplier theorem to these constraints. Another
approach, however, is to note that (I) uniquely determines x once u is
specified and hence we really only need to select u. The problem can thus be
regarded as formulated in U; the Lagrange multiplier theorem need only
be applied to the terminal constraints. Still another approach is to view the
problem in X by considering the implicitly defined set of all trajectories that
can be obtained by application of admissible controls. Finally, in Section
10.10 it is seen that it is sometimes profitable to view the problem in E', the
finite-dimensional space corresponding to the constraint (3). Each of these
approaches has theoretical advantages for the purpose of deriving necessary
conditions and practical advantages for the purpose of developing computational procedures for obtaining solutions. In this section we approach the
problem in the space X x U and in the next section in the space U.
The differential equation (1) with initial condition x(t o) is equivalent to
the integral equation

f f(x(-c), u(-c)) d-c = e

t

(4)

x(t) - x(t o) -

to

which we write abstractly as

(5)

A(x, u)

= e.

The transformation A is a mapping from X x U into X. If we take

X = C"[to , ftJ, U = Cm[to, fd, then the Frechet differential of A exists, is
continuous under our assumptions, and is given by the formula
(6)

c5A(x,u;h,v)=h(t)- ffxh(-c)d-C- ffuv(-c)d-c

to

for (h, v)

U.

to

256

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

The terminal constraint (3) is a mapping from X into e with Frechet

differen tial
(7)

Together the transformations A and G define the constraints of the

pl:oblem, and we must investigate the question of regularity of these
We must ask if, at the optimal trajectory, the Frechet differentials (6) and (7) taken as a pair map onto X x E' as (h, v) varies over

Xx U.
From the differential (7) it is immediately clear that we must assume that
the r x n matrix G,,(X(tI has rankr. In addition we invoke a controllability
assumption on (6). Spedfically we assume that for any n-dimensional
vector e it is possible to select a continuous function v such that the
equation

Jrl fxh(r:)dr: -

h(t)-

II !"v(r:)dr:=O

10

10

has solution h with h(t1) = e. An intuitive way of describing this assumption

is to say that the original system (1), linearized aboutthe optimal trajectory,
can be driven from the origin to any point in En.
With the above two assumptions we can show that the constraints are
regular. For this it is sufficient to show that for any e E E" and any function
ye X there is an (h, v) e X x U such that

(8)

h(t) -

I f"h(r:) dr: - I fuv(r:) dr: =

I

10

10

h(tl)

(9)

y(t)

= e.

First, for v = 0 in equation (8), there is, by the fundamental existence

theorem for linear Volterra integral equations (see Example 3, Section
10.2), a solution 1i of (8). We may then write (8) as
(10)

where w(t) = h(t) -h(t). By the controllability assumption there is a v

such that the solution to (10) has W(tl) = e -h(tl)' Then h(t) = w(t) + h(t)
is the desired solution to equations (8) and (9).
Having examined the question of regularity, we now give the basic necessary conditions satisfied by the solution to the optimal control problem.
Theorem 1. Let xo, Uo minimize

Jl(x, u) dt
tl

(2)

J =

10

9.5

BASIC NECESSARY CONDITIONS

257

subject to

(1)

*(t) =/(X, U),

(3)

and assume that the regularity conditions are satisfied. Then there is an
n-dimensional vector-valued/unction A(t) and an r-dimensional vector J1 such
that for all t E [to , t 1]
(I I)

- ,l.(t) = [/x'(xo(t), Uo(t]A(t)

+ fx'(xo(t), uo(t

';'(t l ) = Gx'(XO(tlJ1

(12)

uo(t

(13)

+ lu(xo(t), uo(t = e.

Proof The Lagra.nge multiplier theorem (Theorem I, Section 9.3)

yields immediately the existence of A E NBvn[to, t l ], J1 E E r such that
(14)

(lx(X o , uo)h(t) dt

+ (dA'(t)[ h(t) -

(f,(x o , uo)h(-c) d-c]

+ J1'G x(xo(tlh(tl) = 0

J l.(xo, uo)v(t) dt - J dA'(t) J/u(xo, uo)v(-c) d-c =

"

(15)

I,

for all (h, v) E X X U. Without loss of generality, we may take A(tl)

Integrating (14) by parts, we have

= e.

J fixo, uo)h(t) dt + J dA'(t)h(t) + J A'(t)/x(xo, lIo)h(t) dt

I,

'II

I,

10

10

10

+ J1'G x h(tt) = o.

(16)

It is clear that A can have no jumps in [to, t l ) since otherwise a suitable h

could be constructed to make the second term of (16) large compared with
the other terms. There must, however, be a jump at tl of magnitude
-Gix o(tIJ1. Since (16) holds for all continuous h, it holds in particular
for all continuously dijferentiable h vanishing at to . Therefore, integrating
the second term by parts, we have for such functions

f {lx(xo, uo)ft(t) - A'(t)h(t) + A'(t)/x(xo , uo)h(t)} dt

I,

= O.

10

Hence, by Lemma 3, Section 7.5, it follows that A is differentiable on

[to, t l ) and that (11) holds.
Integrating equation (15) by parts, (13) follows from Lemma I, Section
7.5. Now by changing the boundary condition on A(tl) from A(lI) = to
A(11) = Gx'1l to account for the jump, A will be continuous throughout

[t 9 ,

ta I

258

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

Note that the conditions (II), (12), and (13) together with the original
constraints (1) and (3) and the initial condition form a complete system of
equations: 2n first-order differential equations, 2n boundary conditions,
r terminal constraints, and m instantaneous equations from which xo(t),
,t(t), jl., and uo(t) can be found.
Example 1. We now find the m-dimensional control function u that
minimizes the quadratic objective functional

=t f

tl

(17)

to

[x'(t)Qx(t) + u'(t)Ru(t)] dt

subject to the linear dynamic constraint

x(t) = Fx(t)

(18)

+ Bu(t),

x(to) fixed,

where Q is an n x n symmetric positive-semidefinite matrix, R is an m x m

symmetric positive-definite matrix, F is an n x n matrix, and B is an n x m
matrix. This problem is of considerable importance in optimal control
theory because it is sufficiently general to describe many practical problems
adequately and is one of the few problems that can be solved explicitly.
Applying the necessary conditions of Theorem 1, we have
(19)

-let)

= F',t(t) + Qx(t),

,t(tt) = ()

,t'(t)B + u'(t)R = O.

(20)

Since R is positive definite, we have

u(t)

(21)

= -R- 1B',t(t).

Substituting equation (21) into (18), we obtain

x(t) = Fx(t} - BR- 1 B' A.(t),

(22)

x(to) fixed.

Together (19) and (22) form a linear system of differential equations in the
variables x and A.. The system is complicated, however, by the fact that half
of the boundary conditions are given at each end. To solve this system, we
observe that the solution satisfies the relation
(23)

,t(t)

= P(t)x(t),

where pet) is the n x n matrix solution of the Riccati differential equation

(24)

Pet)

=-

P(t)F - F' pet)

+ P(t)BR -1 B'P(t) -

Q,

The verification follows by direct substitution and is left to the reader. It

nan h@ ihnwn that (24) haa a unique. symmetric. positive semidefinite
solution on [to, t1].

9.5

BASIC NECESSARY CONDITIONS

259

From equations (21) and (23) we then have the solution

u(t)

(25)

= -R-IB'P(t)x(t)

which gives the control input in feedback form as a linear function of the
state.
This solution is of great practical utility because if the solution pet) of
equation (24) is found (as, for example, by simple backward numerical
integration), the optimal control can be calculated in real time from
physical measurements of x(t).

Example 2. A rocket is to be launched from a point at time t

= 0 with
fixed initial velocity and direction. The rocket is propelled by thrust
developed by the rocket motor and is acted upon by a uniform gravitational
field and negligible atmospheric resistance. Given the motor thrust, we
seek the thrust direction program that maximize'S the range of the rocket on
a horizontal plane.
in Figure 9.6. Note that the final time T is
The problem is
determined by the impact on the horizontal plane and is an unknown
variable.
y

Figure 9.6

Letting

VI

= oX, Vz =

>" x

The rocket example

y, the equations of motions are

VI = fU) cos

e,

VI (0)

Vz = r(t) sin

e- g,

V2(O) given,

given,

where f(t) is the instantaneous ratio of rocket thrust to mass and g is the
acceleration due to gravity. The range is
J

VI(t)

dt = x(T),

where T> 0 is the time at which y(T) = 0 or, equivalently, the time at
which vz(t) dt = O.
We may regard the problem as formulated in the space CZ[O, To] of twodimensional continuous time functions where To is some fixed time greater

260

LOCAL THEORY OF CONSTRo\INED OPTIMIZATION

than the impact time of the optimal trajectory. Assuming a continuous

nominal trajectory, VI' v2 with impact time T,. we can compute the Frechet
differential of J by reference to Figure 9.7 which shows the continuous
nominal and a perturbed trajectory V1 = V1 + h1' V2 = V2 + h 2 The perturbed trajectory crosses the X axis at a different time T. Denoting by
x(T), y(T) the x and y coordinates of the perturbed trajectory at time 1, we
have, to first order,

and
,
J(v

+ h) -

J(v)

= x(T) -

=t

x(T)+(T- T)V1(T)

h 1(t) dt

+ (T -T)Vl<T).

Combining these we have, to first order,

M(v; h) =: J(v

+ h) -

J(v)

= fa h 1(t) dt -

'T

fo h2(t) dt.

Figure 9.7 Calculation 0/ Frechet differential

Therefore; the original problem is equivalent to finding a stationary point

of the functional

([v.(t) -

V2

(t)] dt

which is an integral on the fixed interval [0, T]. This problem can be solved
in the standard fashion.
Following Theorem 1, we introduce the Lagrange variables

J. 1 = -1
J. _ V1(T)
2 -

v2(T)

A,1(T) = 0

A,2(T) =

o.

9.6

THE PONTRYAGIN MAXIMUM PRINCIPLE

Then AI(t) = (T - t), Ait) =

[VI

Cf)/v 2(T)](t -

261

T). The optimal O(t) satisfies

-A1(t)f(t) sin O(t) + A2(t)f(t) cos O(t) =

o.

Thus we obtain the equation

ii,(T)

tan O(t) = - ii (T) .

2

We conclude that 0 is constant in time. The constant is determined

implicitly by the particular nature of nt).

*9.6 The Pontryagin Ma);imnm Principle

The Pontryagin maximum principle gives a set of necessary conditions for
control problems in which the control u(t) is constrained to a given set. In
this section we develop one form of the maximum principle from an
abstract viewpoint by exploiting the basic definition of an adjoint operator.
Motivated by the framework of the optimal control problem discussed
in Section 9.5, we let X and U be normed linear spaces and g[x, u] a cost
functional on X x U, and we consider a constraint of the form
(1)

A[x, uJ = 0

where A is a mapping from X x U into X. The transformation A describes

the system equations and may represent a set of differential equations,
integral equations, partial differential equations, difference equations, etc.
We assume that (1) defines a unique implicit function x(u). Furthermore,
we assume that A and 9 are Frechet differentiable with respect to x and that
the derivatives Ax[x, uJ andgx[x, uJ are continuous on X x U. Finally, we
assume that the implicit function x(u) satisfies a Lipschitz condition of the
form
(2)

Ilx(u) - x(v)11 ::;; Kllu -

vii.

The control problem is to find (x, u) minimizing J = g[x, uJ while

satisfying A[x, u] = 0 and u e n where n is a prescribed subset of U. Since
x is uniquely determined from u, the objective functional J can be considered
to be dependent only on u, it being understood that J(u) = g[x(u), u].
We now introduce the Lagrangian functional of our constrained optimization problem. For x eX, u e U, A* e X*, we define
(3)

L[x, u, A*] = A* A[x, u]

+ g[x, u].

The following proposition can be regarded as the basis of a number of

necessary conditions for control problems connected with various types
of systems.

262

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

Proposition 1. For q:ny u e n let l* be a solution oj the equation

A* A x[x(u) , u]

(4)

+ gx[x(u), u] = O.

Then for v E 0,
(5)

J(u) - J(v)

= L[x(u), u, A*] -

L[x(u), v, A*]

+ o(lIu -

viI).

Proof By definition
J(u) - J(v)

= g[x(u), u] =g[x(u), u] =g[x(u), u] =

g[x(v), v]

+ g[x(u), v] - g[x(v), v]
g[x(u), v] + gx[x(u), u] . [x(u) - x(v)]
+ (gx[x(v), v] -: gx[x(u), u])[x(u) - x(v)] + o<llx(u) g[x(u), u] - g[x(u), v] + gx[x(u), u][x(u) - x(v)]
+ o(lIu - viI),
g[x(u), v]

x(v)11)

where the last two steps follow from the continuity of gx and the Lipschitz
condition (2),
Likewise,
IIA[x(u), u] - A[x(u), v] - Ax[x(u), u][x(v) - x,(u)]!1 = o(llv - ull).

Therefore,
J(u) - J(v) = L[x(u), u, A*] - L[x(u), v, A*]

+ o(lIu- viI). I

The significance of the above result is that it gives, to first order, a way of
determining the change in J- due to a change in u without reevaluating the
implicit function x. The essence of the argument is brought out in Problem 18. We next apply this result to a system described by a system of
ordinary differential equations of the form

x =/(x, u),
and an associated objective functional
J=

tl(X, u) dt.
to

It is assumed that the functions/and I are continuously differentiable with

respect to x and that/satisfies a uniform Lipschitz condition with respect
to x and u of the form

11/(-'11, II) -/(y, v)1I

where

II II

M[lIx

-.vII

denotes the finite-dimensional norm.

-I-

lIu - vII],

9.6

THE PONTRY AGlN MAXIMUM PRINCIPLE

263

Unlike in the previous section, we now take the admissible control

functions u to be the piecewise continuous functions on the interval
[to, t 1 ], and require that for each t, u(t) E n where n is a prescribed subset
of Em.

Theorem 1. Let x o , U o be optimal for the problem of minimizing

f lex, u) dt
I.

J =

10

subject to x(t)
equation

(6)

=f(x,

u), x(t o) fixed, u(t) E n. Let A be the solution of the

-.le(t) =fx'A(t)

+ lx',

where the partial derivatives are evaluated along the optimal trajectory,
and define the Hamiltonian function

(7)

H(x, u,

t)

= A'(t)f(x, u) + lex, u).

Then for all t E [to, t1 ],

(8)
for all u E

H(xo(t), uo(t), AU :::.; H(xo(t), u, A(t

n.

Proof. For notational simplicity we assume m == 1; i.e., the controls

are scalar functions. We take X = C"[to, (d, and for U we take the space
of piecewise continuous functions with the L1 norm. Defining

f f(x(-r:), u(-r: d-r:

I

A[x, u]= x(t) - x(to) -

10

g[x,

uJ =

l.

f lex, u) dl,
10

we see that A and g are continuously Frechet differentiable with respect to x

(although not with respect to u with the norm we are using).
If x,x + t5x correspond to u, u + t5u, respectively, in {(x, u) : A[x, uJ =8},
we have

+ lt5u(-r:)I} d-r:

IIt5x(t)IIEn =:;;

from which it follows that

Ilox(t)IIE":::;; MeM(t1-t O )

f'lou(-r:)1 d-r:.
10

Therefore, IIt5xll =:;; K Iioull and the transformation A[x,

Lipschitz condition
of Proposition 1.

uJ satisfies the

264

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

It is clear that (6) is equivalent to the adjoint equation A* Axlx, u] +

gx[x, u]

= e. The functional

f H(x, u, A) dt
11

10

is then identical with the Lagrangian (3) except for a term

dt,
which is not important since it does not depend explicitly on u. Proposition
I gives us
tl

(9)

J(uo) - J(u) =

f [H(xo, uo , A)':'" H(xo, u, A)] dt + o(lIu - uoll)

'0

We now show that equation (9) implies (8). Suppose to the contrary that
there is 1 e [to, t 1] and u e n such that

In view of the piecewise continuity of u and the continuity of x, A,f, and I,

it follows that there is an interval [t', til] containing 1 and an 8 > 0 such that
H(x o(t), I it), A(t - H(xo(t),

u, A(t)

>e

for all t e [t', til].

Now let u(t) be the piecewise continuous function equal to uo(t) outside
[t', tit] and equal to ii on [t', t"]. From (9) we have
J(uo) - J(u) > 8(t" - t')

+ o(llu -

uolI).

But lIu - uoll =O([t" - t']); hence, by selecting [t', til] sufficiently small,
J(uo) - J(u) can be made positive, which contradicts the optimality of Uo I
Before considering an example, several remarks concerning this result
and its relation to other sets of necessary conditions are appropriate.
Briefly, the result says that if a control function minimizes the objective
functional, its values at each instant must also minimize the Hamiltonian.
(Pontryagin's adjoint equation is defined slightly differently than ours with
the result that his Hamiltonian must be maximized, thus accounting for
the name maximum principle rather than minimum principle.) It should
___ immediately be obvious that if the Hamiltonian is differentiable with
respect to u as well as x and if the region n is open, the conditions of
Theorem 1 are identical with those of Theorem 1, Section 9.5. The
mum principle can also be extended to problems having terminal constraints, but the proof is by no means elementary. However, problems of
this type arising from applications are often convex and can be treated by
the global theory developed in Chapter 8.

9.6

THE PONTRY AGIN MAXIMUM PRINCIPLE

265

Example 1. We now solve the farmer's allocation problem (Example 3,

Section 8.7) by the maximum principle. To formulate the problem as one
of optimal control, we let u(t) denote the fraction of production rate that is
reinvested at time t. The: problem then is described by'
x(t)

(10)

= u(t)x(t),

x(O)

>0

(11)

J =

(12)

So (1 -

u(t))x(t) dt

O::5:u(t)::5:1.

Here, as in Chapter 8, the farmer wishes to select u so as to maximize the

total storage J.
The adjoint equation for this problem is

(13)

U(t)A(t)

+ 1-

A(T) = 0

u(t),

and the Hamiltonian is

(14)

H(x, u, A)

= A(t)U(t)x(t) + [1 -

u(t)]x(t).

An optimal solution x o , uo, A must satisfy (10), (12), and (13) and (14)
must be maximized with respect to admissible u's. Since x(t) ;;::: 0 for all
t E [0, T], it follows from (14) that
A(t) > 1
A(t)

< 1.

Then sinceA(T) = 0, we have uo(T) = 0, and equation (13) can be integrated

backward from t = T. The solution is shown in Figure 9.8. We conclude
that the farmer stores nothing until T - 1, at which point he stores all
production.

o
Figure 9.8

Solution to adjoint equation

266

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

9.7 Problems
1. Prove the following inverse function theorem. Let D .be an open subset
of a Banach space X and let T be a transformation from D into X.
Assume that Tis continuously Frechet differentiable on D and that ata
point Xo eD, [T'(XO)]-l exists. Then:
(i) There is a neighborhood P of Xo such that T is one-to-one on P.
(ii) There is a continuous transformation F defined on a neighborhood
N of T(xo) with range R c: P such that F(T(x = x for all x e R.
Hint: To prove uniqueness of solution to T(x) == yin P, apply the
mean value inequality to the transformation q>(x) = [T'(xo)] -IT(x) - x.
2. Prove the following implicit function theorem. Let X and Y be Banach
spaces and let T be a continuously Frechet differentiable transformation from an open set D in X x Y with values in X. Let (xo, Yo) be a
point in D for which T(xo ,Yo) = (J and for which [T/(xo, Yo)] -1 exists.
Then there is a neighborhood N of Yo and a continuous transformation
F mapping N into X such that F(yo) = Xo and T(F(y), y) = () for all
yeN.

3. Show that if all the hypotheses of Theorem 1, Section 9.4, are satisfied,
except perhaps the regularity condition, then there is a nonzero,
e R x Z* such that ro/(x) + <G(x),
is
positive element (ro,
stationary at xo, and <G(x),
= o.
4. Let gl' gz , ... , gn be real-valued Frechet differentiable functionals on a
normed space X. Let Xo be a point in X satisfying
(1)

== 1,2, ... ,no

Let I be
set of indices i for which gi(XO) = 0 (the so-called binding
constraints). Show
Xo is a regular point of the constraints (I) if and
only if there is nO set of A./s, i e I satisfying
for all i E I,
5. Show that if in the generalized Kuhn-Tucker theorem X is normed, /
and G are Frechet differentiable, and the vector x is required to lie in a
given convex set 0 c: X (as well as to satisfy G(x) ::;; (J), then there is a
Zd 8 such that <G(xo),
= 0 and!'(x o) +
e [0 - xo] al.
6. A bomber pilot at a certain initial position above the ground seeks the
path of shortest distance to put him over a certain target. Considering
only two dimensions (vertical and horizontal), what is the nature of
the solution to his problem when there are mountain ranges between
him and his target?

9.7

PROBLEMS

267

7. Let X be a normed linear space and let Z be a normed linear space

having positive cone P. Let G be a Frechet differentiable mapping of
X" into Z. A point Xo is said to 'satisfy the Kuhn-Tucker constraint
qualification relative to the inequality G(x) :::; e if G(xo) :::; e and if for
every hEX satisfying G(xo) + G'(xo)h :::; there is a differentiable arc
x(t) defined for t E [0, I] such that

(i)

G(x{t)) :::;
dx(t)
dt

(ii)

:=

for all t E [0, 1]

(iii)

x(O) = xo.

1=0

Give an example of a finite-dimensional mapping G and a point Xo

that satisfies the Kuhn-Tucker constraint qualification but is not
regular. Give an example of a finite-dimensional mapping G and
an xo, G(xo) :::; e that does not satisfy the Kuhn-Tucker constraint
qualification.
8. Let X, Z, P, and G be as in Problem 7 and let / be a real-valued
functional on X. Suppose Xo minimizes / subject to the constraint
G(x) :::; and that Xo satisfies the Kuhn-Tucker constraint qualification.
Show that /'(xo)h 2: for every h satisfying G(xo) + G'(xo)h :::; e.
Using this result, prove a Lagrange multiplier theorem for finitedimensional spaces.
9. Let T be a transformation mapping a vector space X into a normed
space Z with positive cone P. We say that T has a convex Gateaux

differential
<5+T(x; h) = lim
"->0+01:

[T(x

+ rxh) -

T(x)]

if the limit on the right exists for all hEX and if <5+T(x; h) is convex in
the variable h. Let/be: a functional on X and G a transformation from
X into Z. Assume that both / and G possess convex Gateaux differentials. Let Xo be a solution'to the problem:
minimize /(x)
subject to G(x) ::;;

e.

Assume that there exists an h such that G(xo)

that there is a
2: e such that

+ <5+G(xo , h) < e. Show

for all hEX. Give an example of a functional having a convex Gateaux

differential but not a linear Gateaux differential.

268

LOCAL THEORY OF CONSTRAINED OPTIMIZATION

10. After a heavy military campaign a certain army requires many new
shoes. The ,?'lartermaster can order three sizes of shoes. Although he
does not know precisely how many of each size are required, he feels
that the demands for thl three sizes are independent and the demand
for each size is uniformly distributed between zero and three thousand
pairs. He wishes to allocate his shoe budget of four thousand dollars
among the three sizes so as to maximize the expected number of men
properly shod. Small shoes cost one dollar per pair, medium shoes
cost two dollars per pair, and large shoes cOst four dollars per pair.
__How many pairs of each size should he order?
11. Because of an increasing average demand for its product, a
is
considering a program of expansion. Denoting the firm's capacity at
time t by e(t) and the rate of demand by d(t), the firm seeks the nondecreasing function e(t) starting from e(O) that maximizes
T

fo {min [e(t), d(t)] -

[C(t)]2} dt

where the first term in the integrand represents revenue due to sales and
the second represents expansion costs. Show that this problem can be
stated as a convex programming problem. Apply the considerations of
Problem 9 to this problem.
12. Derive the necessary conditions for the problem of extremizing
t.

f I(x, x, t) dt
to

subject to
cf>(x,

x, t) $

0,

making assumptions similar to those in Example 3, Section 9.4.

13. Consider these two optimal control problems:
t.

minimize

J lex, u, t) dt
to

(
A subject to x(t) = I(x, u, t),

x(to) fixed,

f K(x, u, t) dt
'

= b

to

minimize ",(X(tl
B{subject to x(/) = I(x, u, t),

x(to) fixed,

Show that by the introduction of additional components in the state

vector. a problem of type A can be converted to one of type B. Show

that if Gand I/J have con:nuous partial derivatives, a problem of type B

can be converted to one of type A.

9.7

PROBLEMS

269

14. A discrete-time system is governed by the set of difference equations

x(k

+ 1) =f(x(k),

u(k,

where x(k) is an n vector,. u(k) is m-vector control, andfhas continuous

partial derivatives. Find a set of necessary conditions for the problem
of controlling the system from a given x(O) so as to minimize
N

L l(x(k), u(k,

k=O

where the function I has continuous partial derivatives.

15. Using the results of Problem 14, find an optimal feedback control law
when
f(x(k), u(k = Ax(k) + Bu(k)
l(x(k), u(k = x'(k)Qx(k)

+ u'(k)Ru(k),

where Q is positive semidefinite and R is positive definite.

16. Show that in the one-dimensional optimal control problem:

fo lex, u) dt
1

minimize

subject to x(t) = -x(t)

+ u 2 (t)

x(O) = 1
x(l)

= e- 1 ,

the constraints are not r<:::gular.

17. Show that for the general optimal control problem discussed in Section 9.5, a Lagrangian statement with an additional scalar multiplier
can be made even if the system is not regular.
18. Let X and V be normed spaces and let A [x, uJ = Bx + Cu where Band
C are bounded linear operators with range in X. Assume that the
equation A[x, uJ = () defines a unique implicit solution x(u). Show that
for any pair (x, u) satisfying A[x, uJ = () and any b * E X*, c* E U*,
we have <x, b*) +<u, c*) = (u, c* - C*A*) where B*A* = b*. Compare with Proposition 1, Section 9.6.

REFERENCES
9.2. This form of the inverse function theorem is apparently new, but the proof
is based on a technique in Luisternik and Sobolev [101, pp. 202.208]. For
standard results on inverse function theorems, see Apostol [10] and Hildebrandt and Graves [72].

270

LOCAL THEORY OF CONSTRAINED OPTIMIZA:rION

9.3. The general Lagrange multiplier theorem is due to Luisternik (in

Luistemik and Sobolev [101]). Another similar result is due to Goldstine
.
9.4. Extremizahon problems subject to inequality constraints in finite dimension were first treated systematically by John {761 and Kuhn and Tucker [91].
There are a number of extensions of the Kuhn-Tucker result to more general
spaces, but these results are a good deal weaker than the corresponding results
for equalities (see Problems 7 and 8 and Russell [135]). The approach we
take is similar to that employed by Balakrishnan [15] for a finite number of
inequalities. For an interesting development of necessary conditions in finitedimensional spaces, see Canon, Cullum, and Polack [28].
9.5. For a nice discussion of necessary conditions for optimal control, see
Blum [23].
9.6. See Pontryagin, Boltyanskii, Gamkrelidze, and Mishchenko [119] or Lee
and Markus [95]. For an approach to the maximum principle Closely connected with
programming, see Neustadt [112], [113]. For an
analysis similar to the one in this section, see Rozonoer [132], [133]. [134].
[60], [61].

10
ITERATIVE METHODS
OF OPTIMIZATION

10.1

Introduction

Although a number of interesting optimization problems can be completely

resolved analytically, or reduced to simple finite-dimensional problems, the
great majority of probkms arising from large industrial, aerospace, or
governmental systems must ultimately be treated by computer methods.
The reason for this is not that the necessary conditions are too difficult to
derive but rather that solution of the resulting nonlinear equations is
usually beyond analytic tractability.
There are two basic approaches for resolving complex optimization
problems by numerical
(I) formulate the necessary conditions
describing' the optimal solution and solve these equations numerically
(usually by some iterative scheme) or (2) bypass the formulation of the
necessary conditions and implement a direct iterative search for the
optimum. Both methods have their merits, but at present the second
appears to be the most ejlfective since progress during the iterations can be
measured by monitoring the corresponding values of the objective functional. In this chapter WI! introduce some of the basic concepts associated
with both procedures. Sections 10.2 and 10.3 discuss methods for solving
nonlinear equations; the remaining sections discuss methods for minimizing
objective functionals.
The relevance of the material in the previous chapters to implementation
of the first approach is obvious. In the second approach, however, since the
necessary conditions ar'e abandoned, it is perhaps not clear that any
benefit is derived from classical optimization theory. Nevertheless,
adjoints, Lagrange multipliers, and duality nearly always enter any detailed
analysis of an iterative technique. For instance, the Lagrange multipliers
of a problem are often by-products of an iterative search procedure. The
most important tie between the two aspects of optimization, however, is that
much of the analytical machinery and geometric insight developed for the
271

272

ITERATIVE METHODS OF OPTIMIZATION

10

theory of optimization underlies much of the reasoning that leads to new,

effective, computational procedures.
\\mTHODS FOR SOLVING EQUATIONS

10.2 Successive Approximation

In its general form the classical method of successive approximation applies
to equations of the form x = T(x). A solution x to such an equation is said
to be a fixed point of the transformation T since T leaves x invariant. To
find a fixed point by successive approximation, we begin with an initial trial
vector Xl and compute X:z = T(XI)' Continuing in this manner iteratively,
we compute successive vectors Xn+ l = T(xn). Under appropriate conditions
the sequence {xn} converges to a solution of the original equation.

Definition. Let Sbe a subset ofa normed space X and let Tbe a transformation mapping S into S. Then T is said to be a contraction mapping if there is
an oc, 0 S oc < 1 such that \IT(XI) - T(x:z)1I Soc \lx l - x:zII for all Xl' X:z E S.
Note for example that a transformation having IIT'(x)1I Soc < 1 on a
convex set S is a contraction mapping since, by the mean value inequality,
IIT(Xl) - T(x:z)1I S sup IIT'(x)lllIxl - X:z1I Soc IIxl - x:zlI.
Theorem 1. (Contraction Mapping Theorem) 1fT is a contraction mapping on

a closed
S of a Banach space, there is a unique vector Xo e S satisfying
Xo = T(xo) Furthermore, Xo can be obtained by the method of successive
approximation starting fn.n an arbitrary initial vector in S.
Proof Select an arbitrary element Xl e S. Define the sequence {xn}
by the formula Xn+l = T(xn}. Then IIXn+1 - xnll == IIT(xn) - T(xn-I)11
oclixn - xn-ili. Therefore,

It follows that
IIxn+p - xnll S Ilxn+p - xn+p-lll
S (ocn+p-:Z + ocn+p00

+ Ilxn+p-1 - Xn+p-2\1 + ... + Ilxn+ I

3 + '" + ocn-I)llx:z - xt\1

- x.11

n- 1

- XIII ;:::; _ oc IIx:z - xdl,

and hence we conclude that {xn} is a Cauchy sequence. Since S is a closed

subset of a complete space, there is an element Xo E S such that Xn Xo

IO.2

SUCCESSIVE APPROXIMATION

273

We now show that Xo == T(xo). We have

Ilxo -

T(xo) I

= Ilxo -

+ Xn -

x.

T(xo) I

IIxo - x.1I

+ Ilx. -

T(xo) I

:::;; IIxo - x.11 + cxllxn _ , - xoli.

By appropriate choice of 11 the right-hand side of the above inequality can
be made arbitrarily small. Thus Ilxo - T(xo)1I = 0; Xo = T(x o).
It remains only to show that Xo is unique. Assume that Xo and Yo are
fixed points. Then

Thus Xo = Yo'

Ilxo - Yoll
I

= IIT(xo) - T(yo) II

:::;; allxo - Yoll

The process of successive approximation is often illustrated by a diagram

such as that of Figure 10.1. The figure represents the process of solving the
one-dimensional equation x = f(x). On the diagram this is equivalent to
finding the point of intersection of f(x) with the forty-five degree line
through the origin. Starting with XI' we derive X2 = f(x ,) by moving along
the curve as shown. Thefshown in the figure has slope less than unity and
is thus a contraction. Figure 10.2 shows a case where successive approximation diverges.
[(x)

II

//

/
/
/

/1

1
1

.f- - ......"..
I
/

/
/

Figure 10.1

Successive approximation

Example 1. (Linear Algebraic Equations) Consider the set of equations

Ax = b where A is an 11 >< 11 matrix. A sequence of approximate solutions
=
, ...
k = 1,2, ... , can be generated by solving the equations

Xk

+
+

+ ... +
+ ... +

= b,
= b2

274

ITERATIVE METHODS

('7

OPTIMIZATION

10

f(x)

Figure 10.2 Divergent case of successive approximation

for Xk+ I given X k . In other words, the p-th equation is solved for the new
component
I by first setting all other components equal to their values
at the last iteration. We analyze this method and show that it converges if A
has a property called strict diagonal dominance.

Definition. A matrix A is said to have strict diagonal dominance if

latil >

L: laijl

j*1

for each i.
In what follows we assume that A has strict diagonal dominance and that
each of the n equations represented by Ax;: b has been appropriately
scaled so that ajj = 1 for each i. The equation may be rewritten as
x

= (I -

A)x + b.

Defining (I - A)x + b = T(x), the problem is equ.ivalent to that of finding a

fixed point of T. Furthermore, the method of successive approximation
proposed above for this problem is equivalent to ordinary successive
approximation applied to T. Thus it is sufficient to show that T is a contraction mapping with respect to some norm on n-dimensional space.
Let X be the space of n-tuples with norm defined by

Ilxll

= Imax
IXII
SISn

This norm on X induces a horm on n x n matrices B;

UBII

= max
L: Ibljl
I
j= 1

For the mapping T defined above we have

\IT(x) - T(y) I

= II(A -

I)(x - y)11 ::;;; IIA -

IllUx - YII

IO.2

SUCCESSIVE APPROXIMATION

275

However, since ali = 1, the norm of A - I is

IIA - III

L lau l == ee.

= max

J*I

By the assumption of strict diagonal dominance, ee < I and thus T is a

contraction mapping.

Example 2. Consider the itntegral equation

J
b

x(t) ,= f(t)

+ A K(t, s)x(s) ds.

a

Let
Kl(S, t) dt ds = p2 < co, and assume that f EX:::= Ll[a, bJ. Then
the integral on the righthand side of the integral equation defines a
bounded linear operator on X having norm less than or equal to p. It follows
that the mapping

T(x) = f( t) + A.

JK(t, s)x(s) ds
b

is a contraction mapping on X provided that 1.11 < liP. Thus, for this range
of the parameter A the equation has a unique solution which can be determined by successive approximation.
The basic idea of successive approximation and contraction mappings
can be modified in
ways to produce convergence theorems for a
number of different situations. We consider one such modification below.
problems at the end of this chapter.
Others can be found in
Theorem 2. Let T be a continuous mapping from a closed subset S of a
Banach space into S, and suppose that Tn is a contraction mapping for some
positive integer n. Then T htls a unique fixed point in S which can be found by

successive approximation.
Proof. Let

Xl

be arbitrary in S. Define the sequence {Xi} by

Xi+l = T(xJ
Now since Tn is a contraction, it follows by Theorem 1 that the subsequence
{Xnk} converges to an eleme:nt Xo E S which is a fixed point of Tn. We show
that Xo is a unique fixed point of T.
By the continuity of T, the element T(xo) can be obtained by applying
Tn successively to T(xl)' Therefore, we have Xo = lim T"k(X t ) and
k ... oo

276

ITERATIVE METHODS OF OPTIMIZATION

T(xO) = T[ lim Tnk(x l )]

k.. oo
tinuity of T,

10

= lim T"k[T(x l )]. Hence, again using the conk"oo

I\xo- T(x o)1\ == lim I\T nk(Xl) - Tnk[T(x])]1\

k-+oo
== lim I\T"{T ft(k-l)(Xl) - T"(k-l)[T(Xl)]}1\
k .. oo
S;

odim I\T"(k-l)(Xl) - T"(/<-$>[T(Xl)]1\

k"oo

,... ccllx o- T(xo)l\,

where cc < 1. Thus Xo = T(xo).
If xo, Yo are fixed points, then IIxo - Yo II = IIT"(xo) - T"(yo) I :::;
cc IIxo - Yo II and hence Xo = Yo. Thus the Xo found by successive approximation is a unique fixed point of T. I

Example 3. (Ordinary Differential Equations) Consider the ordinary

differential equation
x(t)

= f[x(t), t].

The function x may be taken to be scalar valued or vector valued, but for
simplicity we assume here thatit is scalar valued. Suppose that x(to) is
specified. We seek a solution x(t) for to S; t S; t 1
We show that under the assumption that the function f satisfies a
Lipschitz condition on [to, t l ] of the form
If[x" t] - f[Xl' t]1 S Mlxl - xli,
a unique solution to the initial value problem exists and can be found by
successive approximation.
The differential equation is equivalent to the integral equation

x(t) = Xo
On the space X = C [to, t

to

f[x(t), t] dt.

a let the mapping T be defined as

T(x) = fI[x(t), t] dt.

to

Then

II T(Xl) - T(x2)11

II {[[Xl' t] - [[Xl' t]} dt II

to
t

S;

IIJ Mllxl - x211 dt I S; M(t 1 - to)llxl - x211.

to

10.3

Thus T is a contraction mapping if M < l/(t 1

however, shows that

NEWTON'S METHOD
-

277

to). A simple calculation,

Since n! increases faster than any geometric progression, it follows that for
sufficiently large n, Tn :is a contraction mapping. Therefore, Theorem 2
applies and there is a unique solution to the differential equation which can
be obtained by successive approximation.
A slight modification of this technique can be used to solve the EulerLagrange differential t:quation arising from certain optimal control
problems. See Problem 5.
For a successive approximation procedure applied to a contraction mapping T having fixed point xo, we have the inequalities

(1)

ex

< 1,

and

Ilx" - xoll ::; ex Ilx

(2)

lI -

1 -

xoll

A sequence {xn} is said to converge linearly to Xo if

lIm sup

Ilx" - xoll
II X - - xolI
n

= ex

for some ex, 0 < ex < 1. Thus, in particular, (2) implies that a successive
approximation procedure converges linearly. In many applications, however, linear convergence is not sufficiently rapid so faster techniques must
be considered.
.

10.3 Newton's Method

Newton's method is an iterative technique for solving an equation of the
form P(x} ::;: As originally conceived, it applies to equations of a single
real variable but it has a direct extension applicable to nonlinear transformations on normed spaces.
The basic technique for a function of a real variable is illustrated in
Figure 10.3. At a given pOint the graph of the function P is approximated
by its tangent, and an approximate solution to the equation P(x) = 0 is
taken to be the point where the tangent crosses the x axis. The process is

e.

278

ITERATIVE METHODS OF OPTIMIZATION

10

f(x)

Figure 10.3 Newton's method

then repeated from this new point. This procedure defines a sequence of
points according to the recurrence relation
P(Xn)
Xn+1= Xn - P'(x )'
n
Example 1. Newton's method can be used to develop an effective iterative
scheme for computing square roots. Letting P(x) = X2 - a, we obtain by
Newton's method
Xn+l

= Xn _

xn

2xn

a=

t[x + !:].
xn
n

This algorithm converges quite rapidly, as illustrated below, for the computation of .jiO, beginning with the initial approximation Xl = 3.0.
Iteration

Xn

X2
n

3.0000000000
3.1666666667
3.1622807018
3.1622776602

9.0000000000
10.0277777778
10.0000192367
10.0000000000

2
3
4

When applied to equations of form P(x) = 8, where P is a nonlinear

operator between Banach spaces, Newton's method becomes
X,,+l = Xn - [P'(xn)]-lp(x n).

An interpretation of the method is, of course, that the original equation is

linearized about the point

Xn

and then solved for

Xn+l'

Alternatively, the

method can be viewed as the method of successive approximation applied

IO.3

NEWTON'S METHOD

279

to the operator T(x) == x - [P '(x)] -lp(X). A fixed point of T is a solution

of P(x) = e.
When analyzing the convergence of Newton's method, we assume that
P is twice Frechet differentiable throughout the region of interest. Corresponding to a point X n, we denote by Pn the bounded linear operator P'(xn)
and by P;; 1 its inverse [P '(x,,)] -1 ifitexists. Since Newton's method amounts
to successive approximation with T(x) = x - [P '(x)] -1 P(X) , an initial approach at an analysis of convergence would be to determine if IIT'(x)1I < 1.
Since

T'(xn)

P"(xn)p;; 1 [P(x n)],

if lip;; 111 :;;; /3, liP "(X.) II ::; K, lip;; 1[P(xn)]ll :;;; 11, h = /3YJK, we have
IIT'(xn) II :;;; h. Therefore, by the contraction mapping principle, we expect
to obtain convergence if h -< 1 for every point in the region of interest. In
the following theorem it is shown that if h < t at the initial point, then
h < t holds for all points in the iteration and Newton's method converges.
Just as with the principle of contraction mapping, study of the convergence of Newton's method answers some questions concerning existence
and uniqueness of a solution to the original equation.

Theorem 1. Let X and Y be Banach spaces and let P be a mapping from X to

Y. Assume further that:

1. P is twice Frechet differentiable and that IIP"(x)1I :;;; K.

There is a point XI E X such that PI = P '(x 1) has a bounded inverse
PI 1 with Ilpi 1 11 :;;; /31' IIpl l [P(X 1)] II :;;; 1]1'
3. The constant hi = /31thK satisfies hI < t.

2.

Then the sequence Xn+1 = Xn - p;; 1 [P(xn)] exists for all n > 1 and converges
to a solution of P(x) = e.
Proof We show that if the point Xl satisfies 1, 2, and 3, the point
Xz = Xl - Pl 1p(x 1) satisfies the same conditions with new constants /3z,
YJz, h z .
Clearly, x z is well defined 'and II x z value inequality,

XI

I :;;; YJ l ' We have, by the mean

IIpl [pl - Pz:J1I ;:; /31 sup 1IP"(x)llIl x z - XIII,

x

where x = Xl + IX(X z - Xl), 0;:; IX:;;; 1. Thus Ilpl 1 [pl - PzJII ;:; /31KI]I = hi'
Since hI < 1, it follows that the linear operator

= 1- Pl 1 [Pl

- Pl] = pjlpl

280

ITERATIVE METHODS OF OPTIMIZATION

has a bounded inverse satisfying liB-III

have PlH = pz and (PIH)-1
bound is

Ilpilli

10

1/(1- h1)(see Problem 4). We

= H- 1pl 1 so pi 1 exists.

IIH- l llllpl 1 11

An estimate of its

= f3z

To obtain a bound for IIp;lp(xz)ll, we consider the operator Tl(x)

x
Clearly, Tl(X l ) = Xz and Tl'(Xl) = e. Thus
Pl1p(xz) = Tl(xt)

Tl(xz)

Tl'(Xl)(x Z

Xl)'

By applying Proposition 3, Section 7.3, we obtain

IIpl1p(xz)1I

z
t sup I T"(x) II Ilxz - xlll
= t sup ilPl1p"(x)lllIxz - xlli z
::s;; tf31K'11

= th 1'11'

Therefore,
l
IIPilP(xz)11 = IIH- pl 1p (Xz)1I

Finally, setting hz = f3z'1zK, we have

h z
h z ::s; t (1 -\1)Z <

='1Z<t'11'

Hence the conditions 1, 2, and 3 are satisfied by the point X z and the
constants f3z, '1z, and hz . It follows by induction that Newton's process
defines {xn }.
SinCe'1n+l < t'1n ,it follows that'1n < (t)"-l '11' Also since IIXn+1 -xnll <'1n
follows that IIxn+k - xnll < 2'1n and hence that the sequence {xn} con
verges to a point Xo E X.
To prove that Xo satisfies P(xo) = we note that the sequence {IIPnll} is
bounded since

e,

and the sequence {IIxn - Xli\} b bounded since it is convergent. Now for
each n, Pn(Xn+ 1 - Xn) + P(xn) := e; and since IIxn+ 1 - xnll ..:... 0 and IIPnll is
bounded, it follows that IIP(xn)II -+ O. By the continuity of P, P(xo) = O. I
It is assumed in the above theorem that P is defined throughout the
Banach space X and that IIP"(x)1I :;;;; K everywhere. It is clear that these

requirements are more severe than necessary since they are only used in the
neighborhood of the points of the successive approximations. It can be

10.3

NEWTON'S METHOD

281

shown that jf P{x) is defined and liP "(x) II s; K in a neighborhood of Xl with

radius
1
r> h; (l - )1 - 2h 1)'11'
the successive approximations of Newton's method converge and remain
within this neighborhood.
The above theorem can be paraphrased roughly by simply saying that
Newton's method converges provided that the initial approximation x 1 is
sufficiently close to the solution Xo' Because if [P 'ex)] -I is bounded near
xo, the quantity 11 = [p'(x)r1p(x) goes to zero as
therefore
h = /311K is small for x close to Xo.
The most attractive feature of Newton's method, the reward for the
high price paid for the possibly difficult job of solving a linearized version at
each step, is its rate of convergence. Suppose that Newton's method converges to a solution Xo E X where [P '(XO)] -I exists. Furthermore, assume
that, within an open region R containing Xo and the sequence {xn}, the
quantities I [P'(x)] -111, liP "(x) II, and liP II/(x) I are bounded. Then again
defining T(x) == x - [P 'ex)] -IP(X), we have
x n+ 1

Xo ::::; xn - [P'(xn)]"clp(x n) - Xo

=xn Since T'(xo) ::::;

e,

::::; T(x n) - T(xo)

Ilxn+ 1

where x ::::; Xn

+ IX(Xn -

(1)
where c::::;

[P'(x n)] -I P(x n) - {xo - [p'(xo)r l P(xo)}

xoll :::;

2
sup IIT"(x)llllxn - xo11 ,
x

Xo), 0 :::; IX :::; 1. Hence,

Ilxn + 1

xoll :::; c Ilxn

- XOIl2,

sup IIT"(x)II,.a bound depending upon liP I/I(X) II. Relation (1)
xeR

is referred to as quadratic convergence.

The overall conclusion of the analysis of Newton's method is that, under
mild restrictions, the method converges quadratically provided that the
initial approximation is sufficiently near the solution. In practice the detailed
criteria for convergence stated in this section are difficult to check, and it is
often simpler to carry out the process than to verify beforehand that it will
converge. Moreover, the method may converge even though the sufficiency
conditions are violated. One device useful in these situations is to begin
iterating with a slower but surer technique and then change over to
Newton's method to gain the advantage of quadratic convergence near the
end of the process.

282

ITERATIVE METHODS OF OPTIMIZATION

10

Example Z. (Two-Point Boundary Value Problem) Newton's method can

be used very effectively to compute solutions of nonlinear two-point
boundary value Pfoblems such as those arising in connection with optimal
control problems.
Consider first 'Che linear two-point boundary value problem

x(t, = A(t)x(t) + vet),

(2)

where x(t) is an n-dimensional vector function of t subject to the boundary

conditions

(3)

Cx(t l ) =

(4)

Dx(t2)

CI

= d2 ,

where dim (el) + dim (d2 ) = n. Since the system is linear, we may write the
superposition relation
(5)

where 4>(t2 , t) is the matrix function of t satisfying

1>(t2 , t) = -4>(t2' t)A(t)

4>(t2' t 2) = [.
Note that 4>(t2' t) can be found by integrating backward from t 2 .
Defining b = 4>(t2' t)v(t) dt, the boundary conditions (3) and (4) can be
expressed entirely in terms of t I

Cx(t l ) =
D4>(t2 , tl)X(tl)

CI

= d2 -

Db.

Therefore, assuming the existence of the appropriate inverse, we have

(6)

1
XUl)

t1)r

d2 _

Having determined x(t l), the original equation (2) can be solved by a single
forward integration.
Now consider a similar, nonlinear, two-point boundary problem:

x(t) = F(x, t)

CX(tl):: cl
Dx(t2 ) = d2

10.4

GENERAL PHILOSOPHY

283

Although this problem cannot be solved by a single integration or by superposition, it can often be solved iteratively by Newton's method. We start
with an initial approximation Xt(t) and define
Xn+l(t):::: F(x n, t)

+ Fix,p t)(X,,+I(t) -

xn(t))

= Cl
DXn +I(t2) = d2
CXn+I(tI)

At each step of the iteration the linearized version of the problem is solved
by the method outlined above. Then provided that the initial approximation is sufficiently close to the solution, we can expect this method to converge quadratically to the solution.
DESCENT METHODS

10.4

General Philosophy

Successive approximation, Newton's method, and other methods for

solving nonlinear equations, when applied to an optimization problem,
iterate on the equations derived as necessary conditions for an optimal
solution. A major disadvantage of this approach is that these iterative
techniques may converge only if the initial approximation is sufficiently
close to the solution. With these methods only local convergence is
guaranteed.
A more direct approach for optimization problems is to iterate in such a
way as to decrease the cost functional continuously from one step to the
next. In this way global convergence, convergence from an arbitrary
starting point, often can be insured.
As a general framework for the method, assume that we seek to minimize
a functional f and that an initial point Xl is given. The iterations are
constructed according to an equation of the form

where IXn is a scalar andpn is a (direction) vector. The procedure for selecting
the vector Pn varies from technique to technique but, ideally, once it is
chosen the scalar IXn is selected to minimize f(x" + IXPn), regarded as a
function of the scalar IX. Generally, things are arranged (by multiplyingpn
by - 1 if necessary) so that f(xn + aPn) < f(xn) for small positive IX. The
scalar IXn is then often take:n as the smallest positive root of the equation

d
da f(xn + IXPn)

= O.

284
In

ITERATIVE METHODS OF OPTIMIZATION

10

practice, of course, it is rarely possible to evaluate the minimizing a

exactly. Instead, some iterative

search or approximation is required. The

essential point, however, is that after an OC n is selected, I(xn + ocnPn) is
evaluated to veritY that the objective has in fact decreased from I(xn). If
I has not decreased, a new value of OCn is chosen.
The descent process can be visualized in the space X where the functional
I is represented by its contours. Starting from a point Xl' one moves
along the direction vector PI until reaching, as illustrated in Figure 10.4,
the first point where the line Xl + OCPl is tangent to a contour of f Alternatively, the method can be visualized, as illustrated in Figure 10.5, in the
space R x X, the space containing the graph off

Figure 10.4 The descent process in X

If/is bounded below, it is clear that the descent process defines a bounded
Qecreasing sequence of functional values and hence that the objective values
tend toward a limit/o . The difficulties remaining are those ofinsuring that
10 is, in fact, the minimum off, that the sequence of approximations {xn}
converges to a minimizing vector, and finally, the most difficult, that convergence is rapid enough to make the whole scheme practical.
Example 1. Newton's method can be modified for optimization problems
to become a rapidly converging descent method. Suppose again that we seek
to minimize the functional Ion a Banach space X. This might be accomplished by the ordinary Newton's method for solving the nonlinear equation F(x) = where F(x) =I'(x), but this method suffers from the lack of a
global convergence theorem. The method is modified to become the
Newtonian descent method by selecting the direction vectors according to
the ordinary Newton's method but moving along them to a point mini.
mizing/in that direction. Thus the general iteration formula is

and

OCn

is chosen to minimize/(xn + 1).

lO.S

I
I
I

Iii

X2

Figure 10.5

STEEPEST DESCENT

285

Xl

The descent process in R

10.5 Steepest Descent

The most widely used descent procedure for minimizing a functionalf, the
method of steepest descent, is applicable to functionals defined on a Hilbert
space X. In this method the direction vector Pn at a given point Xn is chosen
to be the negative of the gradient off at x n If X is not a Hilbert space, the
method can be modified by selecting P n to be a vector aligned with, or
almost aligned with, the negative gradient. In this section, however, we
restrict our attention to functionals on a Hilbert space.
An application of the method is to the minimization of a quadratic
functional
f(x) = (x I Qx) - 2(b I x),
where Q is a self-adjoint positive-definite operator on the Hilbert space X.
This problem is of particular theoretical interest because it is the only
problem for which a detailed convergence analysis of steepest descent and
other iterative methods is available. The problem therefore provides a
methods. Of course the quadratic problem
comparison point for the
is of practical interest as well, as illustrated in Chapters 3 and 4.
In analyzing the quadratic problem it is assumed that the constants
. f (x I Qx)
(x I x)

m=m
M

= sup (x I Qx)
(xix)

286

ITERATIVE METHODS OF OPTIMIZATION

10

are positive, finite numbers. Under these 90nditions/is minimized

by the

unique vector Xo satisfying the equation

Qxo

(1)

=b

and. indeed. minimization of / is completely equivalent to solving the

linear equation (1). It is convenient therefore to regard any approximation
x to the point minimizing / as an approximate solution to equation (1).
The vector

r=b- Qx
is called the residual of the approximation; inspection of/reveals that 2r is
the negative gradient of/at the point x.
The method of steepest descent applied to / therefore takes the form

where rn = b - QXn and an is chosen to minimize/(xn+ 1). The value of exn

can be found explicitly since
/(Xn+l) = (xn + ar,. IQ(xn + ar,.
= a2(rn I Qrn) - 2ex(rn I rn)

- 2(x,. + r,.lb)

+ (Xn I QXn) -

2(xn I b),

which is minimized by

(2)

an =

(rn I rn)
.
(rnlQrn)

Steepest descent for /(x) = (x I Qx) - 2(x I b) therefore progresses according to .

(3)
where rn = b - Qxn

Theorem 1. For any Xl e X the sequence {XII} defined by (3) converges (in
norm) to the unique solution Xo 0/ Qx = b. Furthermore, defining
F(x)

= (x -

Xo I Q(x - xo

the rate 0/ convergence satisfies

(Yn IY,.)
where YII

= Xo -

Xn

1
m
F(x,.)

m)n-l

1(
m 1 - MF(Xt)

1O.5

STEEPEST DESCENT

287

Proof. Note that

F(x) = (x -

Xo I Q(x

- xo = (x I Qx) - 2(x 1b)

+ (xo 1 Qxo)

=lex) + (xo I Qxo)

so that both/and Fachieve a minimum at Xo and the gradients of/and F
are equal.
We have

F(x n) - F(x" + 1) 2etir nIQYn) - etn2(r nI Qrn)

F(x n)
=
(Yn I QYn)
by direct calculation. Now

'n =

QYn' so in terms of 'n

2(r n lr")2 (rn l r n)2

F(x n) - F(xn+ I) (r,,1 Qrn)
(rn I Qrn)
=
F(x n)
(Q Irnlrn)
=

Using ('n I Q'n)

(rn I rII)
(r nI rn)
(,,,I Qrn) (Q-I rn I rn)

M('n I rn) . and (Q

-1 'n

I 'n)

('n I 'n), which follows

from the definition of m (see Problem 10), we obtain

F(x n) - F(x n+ 1)
F(x n)

m
M

And finally,

This process of steepest descent is illustrated for a two-dimensional

problem in Figure 10.6. Note! that, according to Theorem 1 and from the
figure, the rate of convergence depends on the eccentricity of the elliptical
contours off For m = M the contours are circular and convergence occurs
in one step.

Example 1. Consider the problem of solving the set of linear equations

Ax = b where A is an N x N positive-definite matrix. We assume that the

288

ITERATIVE METHODS OF OPTIMIZATION

10

Figure 10.6 Steepest descent

equations. have been scaled so that all = 1, i = 1,2, ... , N. According to

the method of steepest descent (with Q = A), the approximate solutions are
generated by

x" + 1 = x" + IX" r" .

Suppose that for simplicity (X"' instead of being calculated according to
equation (2), is taken as !X" == 1. Then the method becomes

x" + I

= x" + r" = x" + b -

Ax",
or X"+l = (I - A)x" + b. This last equation is equivalent to the method of
successive approximation given in Example 1, Section 10.2.

The method of steepest descent is frequently applied to nonquadratic

problems with great success; indeed, the method rarely fails to converge to
at least a local minimum. The following theorem establishes conditions for
which success is guaranteed.
Theorem 2. Let f be a functional bounded below and twice Frechet differentiable on a Hilbert space H. Given Xl e H, let S be the closed convex hull of
{x :f(x} <f(XI}}' Assume that r(x} is self-adjoint and satisfies 0 < mI
r(x} S; MI throughout S (i.e., r(x) is uniformly bounded and uniformly
positive definite}. If {x"} is the sequence generated by steepest descent
applied to f starting at Xl' then f'(x"} --. e. Furthermore, there exists an
Xo e S such that x" Xo andf(xo} = inf {f(x} : x e H}.
Proof. Given xeS, let us apply Taylor's expansion with remainder to
the function g(t} = f(tx + (l - t)XI} obtaining
g(1} - g(O) - g'(O) =

t g"(l}

for some t, 0 < t < 1. This leads immediately to

f(x) - f(x t )

- f'(XI)(X ..,. Xl)

t m \Ix -

XI liZ

from which it follows that S is bounded. Therefore the steepest-descent

process defines a bounded sequence {x"}.

10.5

289

STEEPEST DESCENT

the corresponding sequence {f(x.)} is nonAs for any descent

increasing and bounded bdow and therefore converges to some value [0 .
Now assume that the sequ1ence of gradients {['(x.)} does not converge to
zero. Then there is an e > 0 such that for any N there is an n > N with
II ['(x.) II e. Thus, choose n so large that 1I['(xn)1I e and 1[(xlI) - [01 <
e2/4M. For IX > 0 let XIX = x. - IX['(X.). We have by Taylor's expansion
with remainder

x = tx. + (l

- t)xa , 0

Ilr(x)II II[,(x n)II 2

M)IIf'(X.)1I 2,

(-IX +
where

--IX IIf'(x.)11 2+

f(x a) - f(x.)

1. Therefore, for

r:t.

11M we have

f(x a )

e
- 2M'

f(x.)

which implies that f(x.+ 1) < fo. Since this is impossible, it follows that
1I[,(x.) II --+ O.
For any x, YES we have, by the one-dimensional mean value theorem,

(f'(x) - ['(y) I x - y) = (x - y I f"(X)(X - y

where x = tx

+ (1

- t)y, 0 :s; t

IIx.+k -

m Ilx

_ yll2,

1. Thus

x.11 2 .5: 2.
(f'(x.+ k) - r(x.) I X.+k - xn)
m

or

IIX.+k -

x.1I

2.m IIf'(x.+k) -

f'(x.)II.

Since {f'(x.)} is a Cauchy sequence, so is {x.}; thus there exists Xo

S with

x. --+ Xo.
Obviously,f'(xo) =
such that

f(xo

e. Given h such thatxo +

E S,

there is t, 0 < t < 1,

+ h) =f(x o) + Hh If"(xo + th)h)

?;f(xo)

+ '2 IIhi/2

so Xo minimizes f in S and hence in H.

290

ITERATIVE METHODS OF OPTIMIZATION

10

Example 2. Steepest descent can be applied to the optimal control problem

minimize J

=j

tl

l(x(t), u(t dt

to

subject to x(t) = !(x(t), u(t,

(4)

x(to} given, where x(t) is n dimensional and u(t) is, r dimensional. Under

appropriate smoothness conditions it follows (see Section 9.5) that the

gradient of J (with respect to u) is given by the function

l,.{x(t), u(t

(5)

+ A'(t)!..(x(t), u(t,

where the x(t) resulting from (4) when using the given control u is substituted in (5). The function ).,(t) is the solution of
-,1,(t)

(6)

=fx'(x, u).,(t) + lx'(x, u),

Thus, in summary, given a contro' function u the corresponding gradient of

x, then integrating (6)
backward to find.t, and finally substituting the results in {5). This technique
for calculating the gradient followed by the standard steepest.descent
procedure is one of the most practical and efficient methods for solving
unconstrained control problems.

J can be found by integrating (4) forward to. find

CONJUGATE DIRECTION METHODS

10.6 Fourier Series

The problem of minimizing a quadratic functional 011 a Hilbert space can,
by an appropriate transformation, be fonnulated as a Hilbert ,space
minimum norm problem. It is then natural to look to the machinery
of orthogonalization, the Gram-Schmidt procf.dure,and Fourier series to
obtain a solution. This philosophy underlies conjugate direction methods.
Consider the quadratic objective functional j, defined on a Hilbert
space H,
f(x) = (x I Qx) - 2(x I b),

where Q is a self-adjoint linear operator satisfying

(x I Qx) ::;,;; M(x I x)

(I)
for all
vector

(xl Qx)
X
Xo

m(x I x)

e H and some M, m > O. Under these conditions the unique

minimizing f is the unique solution of the equation Qx = b.

1O.6

FOURIER SERIES

291

We can view this problem as: a minimum norm problem by introducing the
new inner product

[xIY]

= (xl Qy),

since the problem is then equivalent to minimizing

Ilx -

== (x - Xo I Q(X - XO

Suppose we have, or can generate, a sequence of vectors {PI' P2' ... }

that are orthogonal with respect to the inner product [ I ]. Such a sequence
is said to be Q-orthogonal or to be a sequence of conjugate directions.
The vector xo can be expa.nded in a Fourier series with respect to this
sequence. If the n-th partial sum of such an expansion is denoted X n ,
then we have, by the fundamental approximation property of Fourier
series, that Ilx. - Xo II Q is minimized over the subspace [PI' pz , ... , P.].
Therefore, as n increases, the value of IIx. - Xo IIQ and the value of f
decrease. If the sequence {Pi} is complete, the process converges to Xo.
Of course, to compute the Fourier series of Xo with respect to the {Pi}, we
must be able to compute inner products of the form [Pi I xo]. These are
computable even though Xo is unknown since [Pi I x o] = (Pi I Qxo) = (Pi I b).

Theorem 1. (Method of Conjugate Directions) Let {pd be a sequence in H

such that (p i I Qp j) = 0, i:f= j, and such that the closed linear subspace
generated by the sequence is H. Then for any XI E H the sequence generated
by the recursion.
(2)

X.+ I = X.

+ (X.P.

(p.lr.)
=
(P. I Qp.)

(3)

(X

(4)

r. = b - Qx.

satisfies (r. IPk)

Qx=b.

= 0, k

Proof Define Y.

= I, 2, ... , n - I, and X. -+ xo-the unique solution of

= X.

-- Xl. The recursion is then equivalent to Yl = ()

and
(5)

Y.+I=Y.+

(P. I b - Qx I - Qy.)
(P.IQPn)
P.,

or in terms of the inner product [

I ],

(6)

[pRI Yo - Y.]
[ I ]
P.
Pn Pn

Y.+l=Y.+

292

ITERATIVE METHODS OF OPTIMIZATION

10

Since Yn e [PI, P'}., ' . , ,Pn- tJ and since the p/s are Q-orthogonal, it follows
that [Pn I Yn] = 0 and hence equation (6) becomes
[Pnl Yo]
Yn+ 1 == Yn + [ I ] Pn'
Pn Pn
Thus
[Pkl.Yo]
Yn+ 1 = k=
"- [p I p ] Pk'
1
k
k
which is the n-th partial sum of a Fourier expansion of Yo' Since with our
assumptions on Q, convergence with respect to II I is equivalent to convergence with respect to II lIa, it follows that Yn -+ Yo and hence that
xn-+ xo'
The orthogonality relation (rn IpJ = follows from the fact that the error
Yn-YO=xn-xO is Q-orthogonal to the subspace [Pl,P'}., ... ,Pn-lJ.1

Example 1. Consider once again the basic approximation problem in a

Hilbert space X. We seek the vector
in the subspace [Yl' Y'}., ., . , Yn] which best approximates a given vector x.
This leads to the normal equations

Ga=b,
where G is the Gram matrix of {Yl' Y'}., . , ., Yn} and b is the vector with
components b , = (xIYI)'
The n-dimensional linear equation is equivalent to the unconstrained
minimization of

a'Ga - 2a'b
with respect to a e En. This problem can be solved by using the method
of conjugate directions. A set of linearly independent vectors {Pl'P'}." , "Pn}
satisfying p,'GPJ = 0, i ::pj, can be constructed by applying the GramSchmidt procedure to any independent set of n vectors in En, For instance,
we may take the vectors e, = (0, ... , 1,0, ... ,0) (with the 1 in the i-th component) and orthogonalize these with respect to G, The resulting iterative
calculation for a has as its k-th approximation ak the vector such that
n

LalYI

1=1

is the best approximation to x in the subspace [Yl' Y'}., ... , Yk]' In other
words, this conjugate direction method is equivalent to solving the original

10.7

293

ORTHOGONALIZATION OF MOMENTS

a.ppl'oximation problem by a Gram-Schmidt orthogonalization of the

vectors {Yl' Y2, ... , Yn}
*10.7

Ortbogonalization of Moments

The Q-orthogonal direction vectors for a conjugate direction method can

be obtained by applying the Gram-Schmidt procedure to any sequence of
vectors that generate a dense subspace of H. Thus, if {el} is such a sequence
in H, we define

(n> 1),
where again [x Iy] == (x I Qy). This procedure, in its most general form, is in
practice rarely worth the efi'olrt involved. However, the scheme retains some
of its attractiveness if, as practical considerations dictate, the sequence {en}
is not completely arbitrary but is itself generated by a simple recurrence
scheme. Suppose, in particullar, that starting with an initial vector el and a
bounded linear self-adjoint operator B, the sequence {et} is generated by
the relation en+I = Ben. Such a sequence is said to be a sequence of
moments of B. There appeal' to be no simple conditions guaranteeing that
the moments generate a dense subspace of H, so we ignore this question
here. The point of main interest is that a sequence of moments can be
orthogonalized by a proc1edure that is far simpler than the general
Gram-Schmidt procedure.
Tbeorem 1. Let {el} be a sequence of moments of a self-adjoint operator B.

Then the sequence

P2

= BPI -

Pn+1 =

Pn -

[PIIBPI]
- - - - PI
[PI I PI]
[Pn 1BPn]

[Pn-I 1BPn]

-[-1-]
Pn - [
I
] Pn-I
Pn Pn
Pn - I Pn - I

(n

2)

defines a Q-orthogonal sequence in H such that for each n, [PI' P2, ... , Pn] =
[el' e2 , ... , en]'
Proof Simple direct verification shows the theorem is true for PI' P2'
We prove it for n> 2 by induction. Assume that the result is true for
{p/}7= I ' We prove that it is true for {PI}7:; f .

294

ITERATIVE METHODS OF OPTIMIZATION

10

It is clear by inspection that PlIt! is nonzero and is in the subspace

[el' e2' ... , en+ l ]. Therefore, it is only necessary to establish that Pn+1 is
orthogonal to each PI' i :s; n. For any i :s; n we have

[PnIBPn] [ I ]
[PnIBPn-l] [ I
]
[ PI I Pn+ I] = [ PI I B]
Pn - [ I ] Pi Pn - [p
I P ] PI Pn-I .
Pn Pn
n-I . n-I
For i :s; n - 2 the second two terms in the above expression are zero by the
induction hypothesis and the first term can be written as [BPIIPn] which is
zero since BpI lies in the subspace [PI' P2 , ... , PI +I]. For i =n -1 the first
and the third term cancel while the second term vanishes. For i = n the
first and the second term cancel while the third term vanishes. I
10.8 The Conjugate Gradient Method
A particularly attractive method of selecting direction vectors when
minimizing the functional
/(X)

= (x I Qx) -

2(b I x)

is to choosepi = r l = b - QXI (the direction of the negative gradient of/at

XI) and then, after moving in this direction to X2' consider the new negative
gradient direction r2 = b - QX2 and choose P2 to be in the space spanned
by rl' r2 but Q-orthogonal to PI' We continue by selecting the other p/s
in a similar way. In other words, the sequence of p/s is a Q-orthogonalized
version of the sequence of negative gradients {rl' r 2 , . } generated as the
descent process progresses. The method leads to the simple recursive form
( 1)

(2)

(rnIPn)

Xn + I

= Xn + (Pn I QPn) Pn

Pn + I

= r n+ I -

(rn+11 QPn)
(Pn I QPn) Pn .

This two-term formula for the next member of the Q-orthogonal

sequence {Pj} can be considered a consequence of the theorem in the last
section on orthogonalized moments. In the present case it is easily seen
that rn +1 is in the subspace [rl' Qru ... , QnrtJ. Furthermore, because
the direction vectors {Pn} are generated from the negative gradients, the
resulting closed subspace generated by them is .always large enough so
that the xn's converge to th0 optimal solution.
Theorem 1. Let XI e H be given. Define PI
(3)

(4)

I"n

=b -

= b - QXn

QXl and

lO.S

(5)

P.+l

(6)

0(.

THE CONJUGATE GRADIENT METHOD

= r.+ 1 -

295

P.P.

.:....:.i..:..p.:.::..)_

(P. I Qp.)

[3. = (r .+11 Qp.) .

(7)

(P. IQp.)

Then the sequence {x.} converges to Xo = Q- 1b.

Proof We first show that this is a method of conjugate directions.

Assume that this assertion is true for {Pk}k= l '
t ; we shall show
that it is true for one mone step.
Since 0(.+ 1 is chosen in accordance with a method of conjugate directions
we must only show that P.+l is Q-orthogonal to the previous direction
vectors. We have from (5)
(8)

For k = n the two terms on the right of (8) cancel. For k < n the second
term on the right is zero and the first term can be written as (QPk I r.+ 1 ).
But QPk E [PI' P2' ... , PH 1] C [PI' P2 , ... ,P.] and for any conjugate
direction method (r.+ 1 I Pi) = 0, i n. Hence the method is a conjugate
direction method.
Next we prove that the sequence {x.} converges to xo. Define the functional E by
E(x) ,= (b - Qx I Q-l(b - Qx)).

We have, by direct evaluation,

E(x.) - E(x.+ 1 )

= O(.(r. Ip.).

But by (5) and (r. Ip.-l) =: 0 we have (r. Ip.) = (r. I r.) and hence

(9)

E(x. ) -

x.+ 1)

(r.lr.)
1) E(x.).

= 0(. (I
r. Q

r.

Now from (5) and the Q-orthogonality of Pn and P.-l we have

(10)

(r. I Qr.)

= (P. I Qp?) + P;-I(P.-l I Qp.-l)

;;:: (P. I Qp.).

Also, by definition of m, Section 10.6,

(11)

(r. I r.)
(r. I Q-1r.) ;;:: m.

296

ITERATIVE METHODS OF OPTIMIZATION

10

Hence, combining (9), (10), and (11), we obtain

E(Xn+l)

s: (1 - ':)E(Xn).

Thus E(xJ -+ 0 which implies rn -+ O.

The slight increase in the amount of computation required for the conjugate gradient method over that of steepest descent can lead to a significant
improvement in the rate of convergence. It can be shown that for m(x I x)
(x I Qx) M(x
the con ,.,rgence rate is
2

IIxli +1 - xoll S

in E(x1)

1+

Jc '

where c = m/M, whereas for steepest descent the best estimate (see
Problem 11) is

IlxUl - xoll

2s: m
1E(xJ (1C)'h
1 +c

In an n-dimensional quadratic problem the error tends to zero geometrically with steepest descent, in one step with Newton's method, and within n
steps with any conjugate direction method.
The method of conjugate directions has several extensions applicable to
the minimization of a nonquadratic functional f One such method, the
method of parallel tangents (PARTAN), is based on the easily established
geometric rel!ltion which exists among the direction vectors for the
quadratic version of the conjugate gradient method (see Figure 10.7).
B

XII

1'"

.\'11+1

Figure 10.7 PARTAN

Point A, the intersection of the line between (xn , X n + 2) with the line
(xn+l> B) determined by the negative gradient offat Xn+lt is actually the
minimum off along the line (x n + 1 , B). To carry outthe PARTAN procedure
wimwizan6 "" ","bit'""'"Y funvtiQnulf, the point xnH b found from Xn
and Xn+l by first minimizingf along the negative gradient direction from

F<:u:

lO.9

PROJECTION METHODS

297

to find the point A and then minimizing f along the (dotted) line
determined by Xn and A to find x n + 2 . For a quadratic functional this
method coincides with the conjugate gradient method. For nonquadratic
functionals the process determines a decreasing sequence of functional
values f(xn); practical experience indicates that it converges rapidly,
although no sharp theoretical results are available.
X n+1

METHODS FOR SOLVING CONSTRAINED PROBLEMS

Devising computati9nal procedures for constrained optimization problems, as with solving any difficult problem, generally requires a lot of
ingenuity and thorough familiarity with the basic principles and existing
techniques of the area. No general, all-purpose optimization algorithm
has been devised, but a number of procedures are effective for certain
classes of problems. Essentially all of these methods have strong connections with the general principles discussed in the earlier chapters.

10.9 Projection Methods

One of the most common techniques for handling constraints is to use a
descent method in which the direction of descent is chosen to decrease the
cost functional and to re:main within the constraint region.
The simplest version of the method is designed for problems of the form
minimize f(x)
subject to Ax = b,
wherefis a functional on the Hilbert space X, A is a bounded linear operator
from H into a Hilbert space Y, and b is fixed in Y. We assume that A has
closed range. The procedure begins by starting from a point XI satisfying
the constraint. An ideal direction vector PI is found by some standard
technique such as steepest descent or Newton's method. This vector is
projected onto the nullspace of A, JV(A), giving the new direction vector
91' The next point X2 is then taken as

where (X) is chosen in the usual way to minimize f(x 2 ). Since 9) E %(A),
the point X 2 also satisfies the constraint and the process can be continued.
To project the negative gradient at the n-th step onto %(A), the component of f'(xn) in al(A) must be added to the negative gradient. Thus the
required projected negative gradient has the form
9n

= - f'(xn) + A*An

298

ITERATIVE METHODS OF OPTIMIZATION

10

and An is chosen so that AU'(xn) - A*An} = (). Therefore,

9n = -[1- A*(AA*)-1AJ!'(xn)

At the solution xo, !'(xo) will be orthogonal to .A'(A) and hence

j'(xo) - A*,A,o

= O.

This last equation is the necessary <:ondition in terms of the Lagrange

multiplier. Thus, as the method progr,esses, the computation of the projection of the gradient gives, in the limit, the Lagrange multiplier for the
problem.
Similar considerations apply to nonlinear constraints of the form
H(x) = e and to inequality constraints of the form G(x) :$ e, but a number
of additional calculations are required at each step. We do not discuss these
techniques in detail here but merely indicate the general idea of one possible
method. Additional information is c:ontained in the references and the
problems.
Suppose we seek to
minimize j( x)
'ubject to H(x) = 0,
and that we have a point Xl satisfying the constraint. To obtain an improved
vector X2 , we project the negative gradient of j onto the tangent space
{x: H'(X1)x = O} obtaining the direction vector 91' Then XZ(l) is taken as
Xl + IX l91 where IXl is chosen to minimilzej(xl(l. This new vector Xl(l) may
not satisfy the constraint so it must be modified. One way is to employ a
successive approximation technique to generate a sequence {Xl (k) } originating
atxZ(l) ,which converges to a vector Xl satisfyingH(xl) = e. Of course, once
Xz is finally obtained, it must be verified that j(Xl) s.j(Xt) so that Xl is
indeed an improvement. Ifj(Xl) > j(JC1), /Xl must be reduced and a new X;z
found. The method is illustrated in Figure 10.8.
\

-j'
\

Xl

H(x)

Figure 10.8 Gradient projection

={}

10.JO

THE PRIMAL-DUAL METHOD

299

There are a number of techniques of successive approximation that can

be employed to move from the tangent plane to the constraint surface. The
most common methods are modifications of Newton's method (see Problem 9).

10.10 The Primal-Dual Method

Duality theorems supply the basis for a number of computational procedures. These procedures, like the duality theorems themselves, often can
be justified only for convex problems, but in many cases the basic idea can
be modified so as to be effective for other problems.
We consider a dual method for the convex problem
(1)

minimize f(x)
{subject to G(x)

s e,

XEQ.

Assuming that the constraint is regular, this problem is equivalent to

(2)

max inf {I(x)

+ (G(x),

z*)}.

z";:O;O xeO

Or, defining the dual functional

(3)

qJ(Z*) = inf {f(x)

xeO

+ (G(x), z*)},

the problem is equivalent to the dual problem

(4)

maximize qJ(z*)
{subject to z* ;;:::: e.

e,

The dual problem (4) has only the constraint z* ;;:::: hence, assuming that
the gradient of qJ is available, the dual problem can be solved in a rather
routine fashion. (Note that if the primal problem (1) has only equality
constraints of the form Ax = b, the dual problem (4) will have no
constraints.) Once the dual problem is solved yielding an optimal Z6,
the primal problem can be solved by minimizing the corresponding
Lagrangian.

Example 1. (Hildreth's Quadratic Programming Procedure) Consider the

constrained quadratic minimization problem:
(5)

minimize -lx'Qx - b'x

{subject to Ax s c.

Here x is an n vector, Q an n x n positive-definite matrix, A an m x n

matrix, and band c are given vectors of dimension nand m, respectively.

300

ITERATIVE METHODS OF OPTIMIZATION

10

The dual problem (see Example 1, Section 8.6) is

maximize -tA'PA - A'd - tb'Q- 1b
subject to A

e,

where P == AQ- 1A', d == AQ- 1b + c. Or, equivalently,

{minimize tA'PA + A'd
subject to A :2: O.

(6)

After solving Problem (6), obtaining Ao , the solution to (5) is

Xo

= Q-l(b - A'Ao).

To solve (6) we employ a descent procedure with direction vectors equal

to the usual basis vectors ej == (0, 0, ... , 1,0, ... ,0). Specifically we let the
infinite sequence of direction vectors be {elo ez, ... , en' eu ez , ... , en, ... }.
Thus we vary the vector A one component at a time.
At a given step in the process, having obtained a vector A 8, we fix our
attention on a single component Aj The objective functional may be
regarded as a quadratic function of this one component. We adjust Ai to
minimize the function, or if that would require Ai < 0, we set Aj == 0. In any
case, however, the objective functional is decreased. Then we consider the
next component Aj+l'
If we consider one complete cycle through the components to be
one iteration taking the vector Ak to Ak+l, the method can be expressed
explicitly as

where

k+ 1 ==

W,

Convergence of the

1 (
1- 1
k+ 1
- - d j + LPjJAj

Pii

J=1

k)

L pjJAJ .

J=I+l

is easily proved.

Although the dual functional can be evaluated analytically in only a few

special cases, solving the dual rather than the primal is often an efficient
procedure. Each evaluation of the dual functional, however, requires
solving an unconstrained minimization problem.
An evaluation of the dual functional by minimization yields the gradient
of the dual as well as its value. Suppose that Xl is a minimizing vector in
(3) corresponding to zT. Then for arbitrary z*

cp(z!) == !(Xl) + (G(Xl)' zr)

cp(z*) -s.!(Xl) + (G(Xl)' z*).

1O.1O

THE PRIMAL-DUAL METHOD

301

Therefore,
(7)
and hence G(x l ) defines a hyperplane that bounds q> from above. If q> is
differentiable, G(XI) is the (unique) gradient of q> at
(See Problem 8,
Chapter 8.)
In view of the above observation, it is not difficult to solve the dual
problem by using some gradient-based technique modified slightly to
account for the constraint z* e. At each step of the process an unconstrained minimization is performed with respect to x in order to evaluate q>
and its gradient. The minimization with respect to x must, of course,
usually be performed by some iterative technique.

zt.

Example 2. (Optimal Control) Suppose a dynamic system is governed by

an nth order set of ditre:rential equations

x(t)

= I(x, u, t).

Given an initial state x(to), we consider the problem of selecting the

m-dimensional control u(t), to t t I' such that u(t) E U c Rm so as to
minimize the convex functional

subject to the terminal constraints

where G is a convex mapping of En into E Y

Let us partition the state vector into two parts, x = (y, z), where y is that
part of the state vector that enters explicitly into the cost functional t/J and
the constraints G. We write t/J(y), G(y) for t/J(x) , G(x), respectively. For
many problems the dimension p of y is equal to m + I. For example, in
optimizing the flight of a rocket to a given target, the components in x(t I)
representing velocity do not explicitly enter the terminal position constraint.
To apply the primal. dual algorithm to this problem, we define the set

r == {y E EP : (y,

z) = XUI), where X(tl) is the terminal point of some

trajectory generated by a feasible control input u}.

The control problem is then equivalent to the finite-dimensional problem

minimize t/J(y)
to G(y) ::; 0,

302

ITERATIVE METHODS OF OPTIMIZATION

10

This is a convex programming problem if r is convex. For many nonlinear

systems, r can be argued to be convex if the dimension p of y is sufficiently
low. By duality tl).e
, problem is then equivalent to the problem
max {min [",(y)
"Gr

+ ),'G(Y)]} .

For fixed), the inner minimization over y is equivalent to an optimal control

problem having terminal cost

"'(x(tt

+ ),'G(X(tl)

but having no terminal constraints. This latter type of control problem can
be solved by a stll!.ldard gradient method (see Example 2, Section 10.5).

10.11 Penalty Functions

It has long been common practice among optimizers to attack a problem
such as
(1)

minimize f(x)
{subject to hi(x)

= 0,

i = 1,2, ... , p,

by solving instead the unconstrained approximating problem

(2)

minimizej(x) + K

L hi 2(X)
i

for some large positive constant K. For sufficiently large K it can be

reasoned that the solutions to problems (1) and (2) will be nearly equal.
The term KLi h;2(X) is referred to as a penalty function since in effect it
assigns a specific cost to violations of the constraints.
In the practical implementation of the penalty function method, we are
driven on the one hand to select K as large as possible to enhance the degree
of approximation, and on the other to keep K somewhat small so that when
calculating gradients the penalty terms do not completely swamp out the
original objective functional. A common technique is to progressively
sowe problem (2), the unconstrained approximation, for a sequence of K's
which tend toward infinity. The resulting sequence of approximate solutions
can then be expected to converge to the solution of the original constrained
problem. In this section we investigate this type of scheme as applied to
inequality as well as equality constraints.
At first the penalty function method may appear to be a simple algebraic
device-a somewhat crude scheme for overcoming the difficulties imposed
by constraints. There is a geometric interpretation ofthe method, however,

10.11

PENALTY FUNCTIONS

303

which illuminates its intimate relation with other optimization techniques

and lends a degree of elegance to the scheme. Problem (1) is clearly
equivalent to
minimize

(3)

subject to

f(x)

Lt h/(x) ::;; 0,

and by this transformation we reduce the n constraints to a single constraint. This constraint, it should be noted, is not regular; i.e., there is no x
such that
h/(x) < O. The primal function for problem (3),

Li

w(z) == inf {I(x) :

L h 2(X) ::;; z},

i

looks something like that,:;hown in Figure 10.9. It is nonincreasing with

z = 0 as a boundary point of its region of definition. The hyperplane
(which in this case is merely a line since z is a real variable) supporting the
shaded region at the point (w(O) , 0) may be vertical.
Specifying K> 0 and minimizing I(x) + KLi h/2(X) determines, as
shown in Figure 10.9, a supporting hyperplane and a value ({JK for the dual
functional corresponding to problem (3). Provided w is continuous, it is
clear that as K is increased, ({JK will increase monotonically toward w(O).
No convexity requirements need be imposed; since 0 is a boundary point of
the region of definition for w, a (perhaps vertical) support hyperplane
always exists.
There are numerous variations of this scheme. For instance, since the n
original equality constraints are also equivalent to the single constraint

'",

Figure 10.9 The primal function

)0

::

304

ITERATIVE METHODS OF OPTIMIZATION

10

,the penalty term K L /hlx)/ can be used. If the h;'s have nonvanishill! first
derivatives at the solution, the primal function will, in this case, as in Figure
1O.10,havefinite slope at z = 0, and hence some finite value of K will yield
a support hyperplane. This latter feature is attractive from a computational
point of view but is usually offset by the difficulties imposed by nonexistence of a gradient of

w
\

______
Figure 10.10 The primal function using Llhd

For inequality constraints a similar technique applies. Given a functional

9 on a vector space X, we define
g+(x) = max {O, g(x)}.

Let G be a mapping from X into RP, i.e., G(x) ::;: (g1(X), g2(X), ., ., gp(x.
We define

= (g1 +(x), g2 +(x), ... , gp +(x.

inequalities G(x) s: 0 are equivalent

G+(x)
It is then clear the p

inequality

G+(x)'G+(x);::

[g/(X)]2

to the single

s: 0.

Again this inequality does not satisfy the regularity condition. Since this
form includes equality constraints, we consider only inequalities in the
remainder of the section. Hence we analyze the method in detail for the
problem
minimize f(x)
{subject to G(x)

(4)

e.

The geometric interpretation of this problem is identical with that for

equalities.

'

Throughout the following it is assumed that {KII } is an increasing

10.11

PENALTY FUNCTIONS

305

sequence of positive constants tending toward infinity and that problem (4)
has a solution.

Lemma 1. Define
Jlo

= min {I(x)

: G(x) :5 O}.

For each n let Xn be a point minimizing

Then
1. fn+l(X n+1 ) '?:.!,,(x n).
2. flo '?:. !,,(xn)
3. lim Kn G+(xn),G+(x.)

= O.

n-+OO

Proof
1. fn+I(X Q + 1)

2.
3.

+l} + Kn+lG+(Xn+l)'G+(xn+l)
'?:.f(X.+l) + K. G+(X.+l)'G+(Xn+l)
'?:. !,,(xn).

= f(X

Let Xo solve (4).

Jlo = f(xo) = fn(xo) '?:. f,,(x.).
For convenience define g(x) = G+(x)'G+(x). Since K. g(x) '?:. 0, it is
only necessary to prove that lim sup K" g(xn) = O. Assume to the
contrary that lim sup Kn g(x.) = 38 > O. Since by parts 1 and 2 the
sequence {!,,(x n)} is nondecreasing and bounded, it therefore has a
limit y. Select N such that

y <fN(XN )
Since Kn - t
such that

00

+ 8.

and since lim sup Kn g(xn)

:=

38, there is an integer M

and

28 < KM g(XM) < 48.

We then obtain a contradiction from the following string of inequalities:
Y <!N(XN)

+ 8 :5!(XM) + KNg(XM) + 8

=!(XM)

<!(XM)

+ 28 <!(XM) + KMg(XM) =!M(XM):5 y. I

+8

<!(XM)

+ !KMU(X M) + 8

306

ITERATIVE METHODS OF OPTIMIZATION

10

Notice that part 3 of Lemma 1 is stronger than the statement

G+(x,,)'G+(x,,}

O.

Notice also that' no continuity, convergence, or convexity assumptions

are required by Lemma 1.
Theorem 1. Letf(x} and O+(x)'G+(x) be lower semicontinuousfunctionals.
Ifxo is any limit point of the sequence {xn} defined in Lemma 1, then Xo solves
problem (4).

Proof Since g(xn} == G+ (x,,),G +(x n) -+ 0 by Lemma 1, it follows by

tnelower semicontinuity of 9 that g(xo} = 0 and hence
We have
Jl.o
since G(x o} e. Alsof(xn}
Jl.o and hence by the lower
semicontinuity off, f(xo} Jl.o I

e,

It is remarkable and yet perhaps inevitable that the penalty function

method, a method so simply conceived, is strongly connected with the
theory of Lagrange multipliers. The connection is greatest for convex
problems where, in fact, the penalty function method emerges as a particularly nice implementation of the primal-dual philosophy.
Lemma 2. Let f and G be convex and continuous on the normed space X.

If

Xo minimizes

(5)
then it also minimizes
f(x}

(6)

where .,1.0

+ .,1.o'G(x),

= 2G+(xo}.

Proof The reader should find it simple but instructive to supply a

proof assuming differentiab.'ity off and G.
Without assuming differentiability, we construct a proof by contraposition. Suppose Xo does not minimize (6); specifically suppose there is Xl and
8 > 0 such that
f(Xl}
Let x = a:X I
f(x}

+ (1

+ .,1.o'G(x l } <f(xo) + Ao'G(Xo} -

- a:}xo. We write the identity (for 0

8.

a:

1)

+ G+(x)'G+(x} =f(xo) + G+(xo)'G+(xo}

+ f(x} + AO'G(X) - f(xo} - AO'G(XO}
+ IIG+(x} - G+(xo)1I2 - A.o'[G(x} - G+(xJ].

Because of the definition of Xl the second line of the expression is less than
-a:8. Since a continuous convex functional satisfies a Lipschitz condition at

1O.11

PENALTY FUNCTIONS

307

every point on the interior of its domain (see problem 19), the first term in
the third line is O(1X 2 ). The last term is identically zero for ex sufficiently
small. Thus
f(x) + G+(x"YG+(x) $,f(xo) + G+(xo)'G+(xo) - ecx + O(1X 2),
and hence the left side is less than the right for sufficiently small ex showing
that Xo does not
(5). I
Suppose we apply the penalty function method to a problem wherefand
G are convex and continous. At the n-th step let Xn minimize

and define

An = 2Kn G+(xn)
Then applying Lemma 2 with G

Kn1/2G, it follows that Xn also minimizes

or, in terms of the dual functional q> corresponding to problem 4,

In other words, x n ' the nesult of the n-th penalty function minimization,
determines a dual vector ,1.n and the corresponding value of the dual functional. This leads to the interpretation that the penalty function method
seeks to solve the dual.

Theorem 2. Let f and G be convex and continuous. Suppose Xn minimizes

f(x) + Kn G+(x)'G+(x) and define An = 2Kn G+(xn) If Ao is any limit point
of the sequence {An}, then
solves the dual problem
maximize q>(A)
subject to A e.
Proof. The dual functional q> is concave and, being a conjugate
functional (except for sign), it is upper semicontinuous (see Sections 7.9
and 7.10). Hence q>(,1,o);;::: lim sup q>(An). Now given any A e and any
6> 0, select N large enough so that for all n
N we have A.'G(xn) < 6. This
choice is possible since G+(xn} e. Then, since An'G(xn) 0, we have
q>(A) $,f(xn) + A'<J(xn) $,f(xn) + An'G(xn)
for n N. Therefore,
q>(Ao) I

+ e = q>(An) + e

e was arbitrary, we have q>(A.)

$,

lim sup q>(A.n) $,

308

ITERATIVE METHODS OF OPTIMIZATION

10

10.12 Problems
1. Let S be a closed subset of a Banach space. A mapping T from S
onto a region r contai'ling S is an expansion. mapping if there is a
constant K> 1 such that lIT(x) - T(Y)II :?: Kllx - yll for x =1= y. Show
that an expansion mapping has a unique fixed point.
2. Let S be a compact subset of a Banach space X and let T be a mapping
of S into S satisfying IIT(x) - T(y)1I < IIx - yll for x =1= y. Show that T
has a unique fixed point in S which can be found by the method of
successive approximation.
3. Let X = L 2 [a, b]. Suppose the real-valued function [is such that for
each xe X

f [(t, s, x{s ds
b

"

is an element of X. Suppose also that I[{t, s, e) - [(t, s, 01 :s;

K{t, s) Ie - e'l, where
K(t, S)2 ds dt < 1. Show that the integral
equation

J!J!

x(t) = y(t) +

J!(t, s, x(sds
b

has a unique solution x e X for every y e X.

4. Let X be a Banach space and let A be a bounded linear operator from
X into X. Using the contraction mapping theorem, show that if
IIAII a < 1, then (I - A)-l exists (where I is the identity operator)
and 11(1 - A)-lll < I/O - a).
5. Using a technique similar to that employed for solving Volterra
integral equations, devise a successive approximation scheme with
guaranteed convergence for solving the two-point boundary value
problem associated with minimizing

ll

f [x'(t)x(t) + u (t)] dt
2

to

subject to X(t) = [[x(t)] + bu(t), x(to) fixed.

6. Show that Newton's method applied to a function [ of the single
real variable x converges monotonically after the second step if
['(x) > 0 andf"{x) > o everywhere.
7. In a modified Newton's method for solving P{x) = lJ, we iterate
according te>
(1)

10.12

PROBLEMS

309

Assume that P is Ftechet differentiable in a convex region D and that

for some xl E D, [P'(X I )]-l exists. Assume that X2 calculated according to equation (1) is in D, that
p

= II[P'(xl)J-lll sup IIP'(xl) -

PI(x)11 < 1,

xeD

and that the sphere

p

S = {x: Ilx - x211 < - l l x l

1-p

x211}

is contained in D. Show that the modified method converges to a

solution Xo E S.
8. Use the modified Newton's method to calculate
starting from
Xl = 3 and from Xl :::; 1.
9. Let X and Y be Hilbert spaces and let P be a transformation from X
into Y. Suppose that P has a Frechet derivative P 'ex) at each point
X E X and that P'(x) has closed range. Show that, under conditions
similar to those of the standard Newton's method theorem, the
sequence Xn +1 = Xn - [P'(xn)]tp(x n) converges to a point Xo satisfying P(xo) :::; e. (At denotes the pseudoinverse of A.)
Suppose the boundt:d, self-adjoint operator Q on a Hilbert space X
satisfies

J10

10:

. (x I Qx)
lnf (xix) = m > O.

Show that Q has an inverse and that for all

1
m

(x I Q-1x):$; - (xl x).

11. Use Kantorovich's inequality

to obtain an improved estimate for the convergence rate of steepest

descent over that given in Theorem 1, Section 10.5. Prove the inequality
for positive-definite (symmetric) matrices. .
12. Letfbe a functional defined on a normed space X and bounded below.
Given Xl E X, let S be the closed convex hull of the set {x : I(x) </(x I)}'
Assume that S is bounded and that I has a uniformly continuous
Frt!chet derivative on S. Show that the method of steepest descent
applied to ffrom XI generates a sequence {xn } such that III'(xn)1I -+

e.

310

ITERATIVE METHODS OF OPTIMIZATION

10

13. Under the hypothesis of Theorem 2, Section 10.5, show that the simplified steepest-descent process defined by Xn+1 = Xn -(l/M)f'(xn)
converges tt> the point minimizing f
14. Suppose a sequence {xn} in anormed space converges to a point Xo.
The convergence is said to be weakly linear if there exists a positive
integer N such that
.
Ilxn+N - xoll
hm sup II Xn - Xo II < 1.

n.... oo

(a) Show that in Theorem 1, Section 10.5, the convergence is weakly

linear.
(b) Show I,hat in Theorem 2, Section 10.5, the convergence is weakly
linear.
15. Let B be a bounded hnear operator mapping a Hilbert space H into
itself and let el be an arbitrary element of H. Let {ek} be the sequence
of moments ek+l = Be,fe' Show that if en E Eel' e2' ... , en-l], then
em E Eel' e2' ... , en-l] for all m > n.
16. Show that in the method of conjugate gradients there is the relation
(rn IPn) = (rl IPn)'
17. Verify that the geometric relation of Figure 10.7 holds for the method
of conjugate gradients.
18. Show that even iff and G are not convex, if the primal-dual method
converges in the sense that z: -I'
where
e,
e, and
<p'(4)> == 0, then the corresponding Xo minimizingf(x) +(G(x),
is optimal.
19. Show that a continuous convex functional satisfies a Lipschitz condition at every point in the relative interior of its domain of definition.
Hint: See the end of the Proof of Proposition 1, Section 7.9.
20. An alternative penalty function method for minimizing f(x) over the
constraint region

n=

{x: g,(x) ;:: 0, i = 1,2, ... , m}

is to find a local minimum of

over the interior of the region n. For small r the solutions to the two
problems will be nearly equal. Develop a geometric interpretation of

this method.

IO.12

PROBLEMS

311

REFERENCES
1O.2. The contraction mapping theorem is the simplest of a variety of fixed-

point theorems. For an introduction to other results, see Courant and Robbins
[32], Graves [64], and Bonsall [24].
of Newton's method in Banach spaces was developed
10.3. Much of the
by Kantorovich [78]. See also Kantorovich and Akilov [79], [80], Bartle [16],
Antosiewicz and Rheinboldt [9], and Collatz [30]. Some interesting extensions
and modifications of the method have been developed by Altman [6], [7], [8].
An interesting account of various applications of Newton's method is found
in Bellman and Kalaba [18].
1O.4-5. For material on steepest descent, see Ostrowski [115], Rosenbloom
[130], Nashed [108], Curry [33], and Kelley [84),
1O.6-8. For the original development of conjugate direction methods, see
Hestenes and Stiefd [71] and Hestenes [70]. The underlying principle of
orthogonalizing moments, useful for a number of computational problems, is
discussed in detail by Vorobyev [150] and Faddeev and Faddeeva [50].
There are various extensions of the method of conjugate gradients to problems
with nonquadratic objective functionals. The most popular of these are the
Fletcher and Reeves [54] method and the PARTAN method in Shah, Buehler,
and Kempthorne [138]. For another generalization and the derivation of the
convergence rate see Daniel [34], [35]. For application to optimal control
theory, see Lasdon, Mitter, and Waren [94] and Sinnott and Luenberger
[140].
1O.9-11. The gradient-projection method is due to Rosen [128], [129]. A
closely related method is that of Zoutendijk [158]. For an application of the
primal-dual method, see Wilson [154]. The penalty function method goes
back to a suggestion of Courant [31]. See also Butler and Martin [27], Kelley
[84], Fabian [49], and Fiacco and McCormick [53]. For additional general

material, bibliography, and review of optimization techniques, see Saaty and

Bram [136], Wolfe [156], Zoutendijk [159], Leon [96], Wilde and Beightler
[153], Dorn [43], and Spang [143].
10.12. For extensions of Problem 9 to Banach space, see Altman [6] and the
generalized inverse fUnction theorem of Chapter 9. For Kantorovich's inequal
-ity (Problem 11),
Greub and Rheinboldt [66]. The method introduced in
Problem 20 has been studied extensively by Fiacco and McCormick [53J.

BmLIOGRAPBY

[1] Abadie, J., ed., Nonlinear Programming, North-Holland Pub. Co., Amsterdam, 1967.
[2} Akhiezer, N. I., The Calculus of Variations, Ginn (Blaisdell), Boston, 1962.
[3] Akhiezer, N. I. and I. M. Glazman, Theory of Linear Operators in Hilbert
Space, Yols. I, II, Frederick Ungar, New'York, 1961.
[4] Akhiezer, N. I. and M. Krein," Some Questions in the Theory of Moments,"
Article 4, Amer. Math. Soc. Publ., 1962.
[5] Albert, A., "An Introduction and Beginner's Guide to Matrix PseudoInverses," ARCON, July 1964.
[6] Altman, M., "A Generalization of Newton's Method," Bull. A cad. Polon.
Sci. CI. Ill, 3, 189, 1955.
.
[7] Altman, M., "On the Generalization of Newton's Method," Bull. A cad.
Polon. Sci. CI. III, 5, 789, 1957.
[8] Altman, M., "Connection between the Method of Steepest Descent and
Newton's Method," Bull. Acad. Polon. Sci. CI. III, 5, 1031-1036, 1957.
[9] Antosiewicz, H. A. and W. C. Rheinboldt, "Numerical Analysis and Functional Analysis," Chapter 14 in Survey oj Numerical Analysis, ed. by J. Todd,
McGraw-Hill, l'.'ew York, 1962.
[10] Apostol, T. M., Mathemr..'ical Analysis, A Modern Approach to Advanced
Calculus, Addison-Wesley, Reading, Mass., 1957.
[11] Aronszajn, N., "Theory of Reproducing KernelS," Trans. Amer. Math.
Soc., 68, 337-404, 1950.
[12] Arrow, K. J., L. Hurwicz, and H. Uzawa, Studies in Linear and Non-Linear
Programming, Chapter 4, "Programming in Linear Spaces," Stanford Univ.
Press, Stanford, Calif., 1964, pp. 38-102.
[13] Arrow, K. J., S. Karlin, and H. Scarf, Studies in the Mathematical Theory
oj Inventory and Production, Stanford Univ. Press, Stanford, Calif., 1958.
[14] Balakrishnan, A. Y., "An Operator.Theoretic Formulation of a Class of
Control Problems and a Steepest Descent Method of Solution, J. SIAM
Control, Ser. A, 1 (2), 1963.
[15] Balakrishnan, A. Y., "Optimal Control Problems in Banach Spaces,"
J. SIAM Control, Ser. A, 3 (1), 152-180, 1965.
[16J Bartle, R. G., "Newton's Method in Banach Spaces," Proc. Amer. Math.
Soc., 6, 827-831, 1955.
[I7] Bellman, R. E., I. Glicksberg, and O. A. Gross, Some Aspects oj the
Mathematical Theory of Control Processes, The Rand Corp., Santa Monica,
Calif., Jan. 16, 1958.

312

BIBLIOGRAPHY

313

(I8] Bellman, R. E. and R. E. Kalaba, QUQsilinearization and Nonlinear Boundary-Value Problems, American Elsevier, New York, 1965.
[19] Bellman, R. E. and W. Karush, "Mathematical Programming and the
Maximum Transform," J. Soc. Indust. Appl. Math., 10,550-567,1962.
[20] Ben-Israel, A. and A. Charnes, "Contributions to the Theory of Generalized Inverses," J. Soc. Indust. Appl. Math., 11 (3), Sept. 1963.
[21] Berberian, S. K., Introduction to Hilbert Space, Oxford Univ. Press, 1961.
[22) Bliss, G. A., Lectures on the Calculus of Variations, Univ. of Chicago Press,
1945.
[23] Blum, E. K., "The Calculus of Variations, Functional Analysis, and
Optimal Control Problems," Topics in Optimization, Ed. by G. Leitman,
Academic Press, New York, 417-461, 1967.
[24] Bonsall, F. F., Lectures on Some Fixed Point Theorems of Functional
Analysis, Tata Inst. of Fundamental Res., Bombay, 1962.
[25] Brondsted, A. "Conjugate Convex Functions in Topological Vector
Dan. Vid. Selsk. 34 (2), 1-27, 1964.
Spaces," Mat. Fys.
[26] Butkovskii, A. G., "The Method of Moments in the Theory of Optimal
Control of Systems with Distributed Parameters," Automation and Remote
Control, 24,1106-1113,1963.
[27] Butler, T. and A. V. Martin, "On a Method of Courant for Minimizing
Functionals," J. Math. and Physics, 41, 291-299, 1962.
[i8} Canon, M., C. Cullum, and E. Polak, "Constrained Minimization Problems in Finite-Dimensional Spaces," J. SIAM Control, 4 (3), 528-547, 1966.
[29] Chipman, J. S., "On Least Squares with Insufficient Observations,"
J. Amer. Stat. Assoc., S9, 1078-1111, Dec. 1964.
[30] Collatz, L., Functional Analysis and Numerical Mathematics, trans. by
H. Oser, Academic Press, New York, 1966.
[31) Courant, R., "Cakulus of Variations and Supplementary Notes and
Exercises" (mimeographed notes), Supplementary notes by M. Kruskal and
H. Rubin, revised and amended by J. Moser, New York Univ., 1962.
[32] Courant, R. and H. Robbins, What is Mathematics? Oxford Univ. Press,
London, 1941, pp. 251-255.
[33] Curry, H. B., "The Method of Steepest Descent for Nonlinear Minimization Problems," Quar. Appl. Math., 2, 258-261, 1944.
[34) Daniel, J. W., "The Conjugate Gradient Method for Linear and Nonlinear Operator Equations," J. SIAM Numer. Anal., 4 (1), 10-25, 1967.
[35] Daniel, J. W., "Convergence of the Conjugate Gradient Method with
Computationally Convtmient Modifications," Numerische Mathematik, 10,
125-131, 1967.
[36) Davis, P. J., Interpolation and Approximation, Blaisdell, New York, 1965.
[37] Day, M. M., Normed Linear Spaces, Springer, Berlin, 1958.
[38] Desoer, C. A. and B. H. Whalen, "A Note on Pseudoinverses," J. Soc.
Indust. Appl. Math., 11 (2), 442-447, June 1963.
(39) Deutsch, F. R. and JP. H. Maserick, "Applications of the Hahn-Banach
Theorem in Approximation Theory," SIAM Rev., 9 (3), 516-530, July 1967.

316

BIBLIOGRAPHY

[8I] Karlin, S. "Operator Treatment of Minimax Principle," Contributions to

the Theory of Games, ed. by H. W. Kuhn and A. W. Tucker, Princeton Univ.
I
Press, 1950, pp. 133-154.
[82) Karlin, S. Mathematical Methods and Theory in Games, Programming, and
Economics, Vol. I, Addison-Wesley, Reading Mass., 1959.
[83] Karlin, S. Mathematical Methods and Theory in Games, Programming, and
Economics, Vol. II, Addison-Wesley, Reading, Mass., 1959.
[84] Kelley, H. J., "Method of Gradients," Chapter 6 of Optimization Techniques, ed. by G. Leitmann, Academic Press, New York, 1962, pp. 206-252.
[85] Kelley, J. L. and I. Namioka, et ai., Linear Topological Spaces, Van
Nostrand, Princeton, N.J., 1963.
[86] Kirillova, F. M., "Applications of Functional Analysis to the Theory of
Optimal Processes," J. SIAM Control,S (1), 25-50, Feb. 1967.
[87] Klee, V. L., "Separation and Support Properties of Convex Sets," Seminar
on Convex Sets, The Institute for Advanced Study, Princeton, N.J., 1949-1950,
pp.77-87.
[88] Kolmogorov, A. N. and S.V. Fomin, Elements of the Theory of Functional
Analysis, Vol. I, Graylock Press, Rochester, N.Y., 1957.
[89] Krasovskii, N. N., "Approximate Computation of Optimal Control by
Direct Method," Appl. Math. and Mech., 24, 390-397, 1960.
[90) Kretschmer, K. S., "Programmes in Paired Spaces," Canadian J. Math.,
13 (1),221-238, 1961.
(91] Kuhn, H. W. and A. W. Tucker, "Nonlinear Programming," Proc. 2nd
Berkeley Symposium on Mathematical Statistics and Probability, Univ. of Calif.
Press, Berkeley, 1961, pp. 481-492.
[92] Kulikowski, R., "Optimizing Processes and Synthesis of Optimizing
Automatic Control Systems with Nonlinear Invariable Elements," Proc. First
IFAC Congress 1960, Butterworths, London, pp. 473-477.
[93] Lack, G. N. T., "Optimization Studies with Applications to Planning in the
Electric Power Industry and Optimal Control Theory," Rept CCS-5, Institute
in Engineering-Economic Systems, Stanford Univ., Aug. 1965.
[94] Lasdon, L. S., S. K. Mitter, and A. D. Waren, "The Conjugate Gradient
Method for Optimal Control Problems," IEEE Trans. on Automatic Control,
AC-12 (2), 132-138, Apr. 1967.
[95] Lee, E. B. and L. Markus, Foundations of Optimal Control Theory, John
Wiley, New York, 1967.
[96] Leon, A., "A Classified Bibliography on Optimization," in Recent Advances
in Optimization Techniques, ed. by A. Lavi and T. P. Vogl, John Wiley, New
York, 1965, pp. 599-645.
[97] Lorch, E. R., "Convexity and Normed Spaces," Publications L'Institute
MatMmatique, Academie Serbe des Sciences, 4, 109-112, 1952.
[98] Luenberger, D. G., "Reuolution of Mass Spectrometer Data," Report
----SU.SEL-64.129, Stanford Electronics Lab., Stanford, Calif., Nov. 1964.
[99) Luenberger, D. G., "A Simple Polynomial Estimator," IEEE Trans. on
Automatic Control, AC.12(2), 211-212, Apr. 1967.

BIBLIOGRAPHY

317

[100] Luenberger, D. G., "Quasi-Convex Programming," J. SIAM Appl. Math.

Sept., 1968
[101] Luisternik, L. and V. Sobolev, Elements of Functional Analysis, Frederick
Ungar, New York, 1961.
[102] Mangasarian, O. L., "Duality in Nonlinear Programming," Quar. Appl.
Math., 20, 300-302, 1962.
[103] Mann, H. B., Analysis and Design of Experiments; Analysis of Variance
and Analysis of Variance Designs, Dover, New York, 1949.
[104] McKinsey, J. C. c., Introduction to the Theory of Games, McGraw-Hill,
New York, 1952.
(lOS] Moore, E. H., Bull. Amer. Math. Soc., 26, 394-395, 1920.
[106] Moore, E. H., General Analysis, Part I, Memoirs, Amer. Philosophical
Soc. 1, 1935.
[107] Moreau, J. J., .. Fonctions convexes duales et points promimaux dans un
espace hilbertien," C.R. Acad. Sci., Paris, 255, 2897-2899, 1962.
[l08] Nashed, M. Z., ';'The Convergence of the Method of Steepest Descents
for Nonlinear Equations with Variational or Quasi-Variational Operators,"
J. Math. and Meek, 13 (5), 765-794, 1964.
[l09] Natanson, I. P., Constructive Function Theory, Volumes I, II, III, (trans.
by A. N. Obolensky and J. R. Schulenberger), Frederick Ungar, New York,
1964-1965.
[110] Neustadt, L. W., .. Minimum Effort Control," J. SIAM Control, 1, 16-31,
1962.
[111] Neustadt, L. W., "Optimization, A Moment Problem, and Nonlinear
Programming," J. SIAM Control, 2, 33-53, 1964.
[112] Neustadt, L. W., "An Abstract Variational Theory with Applications to
a Broad Class of Optimization Problems. I. General Theory," J. SIAM
Control, 4 (3), 505-527, 1966.
[113] Neustadt, L. W., "An Abstract Variational Theory with Applications to
a Broad Class of Optimization Problems. II. Applications," J. SIAM Control,
S (1), 90-137, 1967.
U14] Nirenberg, L., "Functional Analysis," lectures given in 1960-61, notes
by Lesley Sibner, New York Univ., 1961.
[lIS] Ostrowski, A. M., Solutions of Equations and Systems of Equations, Sec.
Ed., Academic Press, New York, 1966.
[116] Parzen, E., "An Approach to Time Series Analysis," Anal. Math. Stat., 32
(4), 951-989, 1961.
[117] Penrose, R., "A Generalized Inverse for Matrices," Cambridge Philosophical Soc., 51, 406-413, 1955.
(118] Penrose, R, "On Best Approximate Solutions of Linear Matrix Equations," Cambridge Philosophical Soc., 52, 17-19, 1956.
[119] Pontryagin, L. S., V. G. Boltyanskii, R. V. Gamkrelidze, and E. F.
Mishchenko, The Mathematical Theory of Optimal Processes (trans, by
K. N. Trirogoff, ed. by L. W. Neustadt), Interscience, John Wiley. New York,
1962.

318

BIBLIOGRAPHY

[120] Porter, W. A., Modern Foundations of Systems Engineering, Macmillan, New

York,1966.
{I21] Rastrom, H., "Polar Reciprocity," Seminar on Convex Sets, Institute
for Advanced Study, Princeton, NJ.,1949-19S0, pp. 27-29.
[122] Rauch, H. E., F. Tung, and C. T. Striebel, "On the Maximum Likelihood
Estimates for Linear Dynamic Systems," Lockheed Missiles and Space Co.
Tech. Rept. No. 6-90-63-62, Palto Alto, Calif., June 1963.
[123] Riesz, F. and B. Sz-Nagy, Functional Analysis, Frederick Ungar, New
York,1955.
[124] Rissanen, J., "On Duality Without Convexity," J. Math. Anal. and Appl.,
18 (2), 269-275, May 1967.
[125] Ritter, K., "Duality for Nonlinear Programming in a Banach Space,"
J. SIAM Appl. Math., 15 (2), 294-302, Mar. 1967;
[126] Rockafellar, R. T., "Duality Theorems for Convex Functions," Amer.
Math. Soc. Bull., 70, 189-192, 1964.
[127] Rockafellar, R T., "Extension of Fenchel's Duality Theorem for Convex
Functions," Duke Math. J., 33, 81-90, 1966.
[128] Rosen, J. B., "The Gradient Projection Method for Nonlinear Programming: Part I, Linear Constraints," J. Soc. Indust. Appl. Math., 8, 181217, 1960.
[129] Rosen, J. B., "The Gradient Projection Method for Nonlinear Programming, Part II, Nonlinear Constraints," J. Soc. Indust. Appl. Math., 9,
514-532, 1961.
[130] Rosenbloom, P. c., "The Method of Steepest Descent," Proc. of Symposia
in Appl. Math., VI, 127-176, 1956.
[131] Royden, H. L., Real Analysis, Macmillan, New York, 1963.
[132] Rozonoer, L. I., "Theory of Optimum Systems. I," Automation and
----Remote Control, 20 (10), 1288-1302, 1959.
[133] Rozonoer, L. I., "L. S. Pontryagin's Maximum Principle in Optimal
System Theory, II," Automation and Remote Control, 20 (11), 1405-1421,
1959.
[134] Rozonoer, L. I., "The Maximum Principle of L. S. Pontryagin in OptimalSystem Theory, III," Automation and Remote Control, 20 (12), 1517-1532,
1959.
[135] Russell, D. L., I' The Kuhn-Tucker Conditions in Banach Space with
an Application to Control Theory," J. Math. Anal. and Appl., 15, 200-212,
1966.
[136] Saaty, T. L. and J. Bram, Nonlinear Mathematics, McGraw-Hill, New
York,1964.
[137] Sarachik, P. E., "Functional Analysis in Automatic Control," IEEE
International Convention Record, Part 6, pp. 96-107, 1965.
[138] Shah, B. Y., R J. Buehler, and O. Kempthorne, "Some Algorithms for
Minimizing a Function of Several Variables," J. Soc. Indust. Appl. Math.,
12 (1), 74-91, Mar. 1964.
[139] Simmons, G. F., Introduction to Topology and Modern Analysis, McGrawHill, New York, 1963.

BIBLIOGRAPHY

319

[140] Sinnott, J. F. and D. G. Luenberger, "Solution of Optimal Control

Problems by the Method of Conjugate Gradients," Preprints of 1967 Joint
Automatic Control Conference, pp. 566-573.
[141] Sion, M., liOn General Minimax Theorems," Pacific J. 0/ Math., 8,
171-176,1958.
[142] Slater, M., "Lagrange Multipliers Revisited: A Contribution to NonLinear Programming," Cowles Commission Discussion Paper, Math. 403,
Nov. 1950.
[143] Spang, H. A III, "A Review of Minimization Techniques for Nonlinear
Functions, SIAM Rev. 4 (4),343, Nov. 1962.
[144] Taussky, 0., "A Recurring Theorem on Determinants," Amer. Math.
Monthly, 56, 672-676, 1949.
[145] Taylor, A. E., Introduction to Functional Analysis, John Wiley, New York,
1958.
[146] Tseng, Y. Y. "Virtual Solution and General Inversions," Uspehi Mat.
Nauk. (NS), 11, 213-215, 1956. (Reviewed in Math. Rev., 18, 749, 1957).
[147] Tukey, J. W., .,' Some Notes on the Separation of Convex Sets," Portugaliae Math., 3, 95-102, 1942.
[148] Valentine, F. A, Convex Sets, McGraw-Hill, New York, 1964.
[149] Varaiya, P, P., "Nonlinear Programming in Banach Space," J. SIAM
Appl. Math., 15 (2), 284-293, Mar. 1967.
[150] Vorobyev, Y. V., Method of Moments in Applied Mathematics, Gordon
and Breach Science Pub., New York, 1965.
[151] Whinston, A, .. Some Applications of the Conjugate Function Theory to
Duality," Chapter V of Nonlinear Programming, ed. by J. Abadie, Interscience, New York, 1967.
[152] Wiener, N., Extrapolation, Interpolation, and Smoothing of Stationary
Time Series, Technology Press, MIT, and John Wiley, New York, 1949.
[153] Wilde, D. J. and C. S. Beightler, Foundations of Optimization, PrenticeHall, Englewood CHIfs, N,J., 1967.
[154] Wilson, R. "Computation l'>f .Optima1. Controls," J. Math. Anal. and
Appl., 14, 77-82, 1966.
[155] Wolfe, P., "A Duality Theorem for Nonlinear Programming," Quar.
Appl. Math., 19, 239--244, 1961.
[156] Wolfe, P., "Methods of Nonlinear Programming," Chap. 6 of Nonlinear Programming, ed. by J. Abadie, Interscience, John Wiley, New York,
1967, pp. 97-131.
[157] Yosida, K., Functional Analysis, Springer-Verlag, Germany, 1965.
[158] Zoutendijk, G., Methods of Feasible Directions, Elsevier Pub. Co., 1960.
[159] Zoutendijk, G., "Nonlinear Programming: A Numerical Survey,"
J. SIAM Control, 4 (I), 194-210, 1966.

SYMBOL INDEX

la. bl, etc., xv

o. O. xvi

Y, 143
A*, ISO

At, 163
B(X. Y), 145
BVla. b], 23
c,I38
Co, 109

co, 18
cov(x),82
C*, 195
qa, b], 23

Ii if' 5(t), xvi

liT(x;h),171-172
D[a,b], 23

E,xv
'1,23
E(x),79

Ix,xvii
/*, 196
C], 191

v:

C+,304

inf, xv
lim inf, lim sup, xvi
p,29
11....,29
Lp[a,b],31
L .... [a. b], 32
.;iI'(A). 144
NBVla, b], 115

R.R",12-13
144

sup, xv
[S], 16
S(x. e), 24
0,12
T'(x),175

T.Y.(v),24
v(S),17
x2::0,x>O,214

{Xi) ,xvi
X" .... x, 26
[x], 41
X",14
X/M,41
X*,106

,xvii

-,xv
,24
-,25
J., 52,117,118
U,xv
n,xv
c,xv
1111,22
II lip, 29
(1),46
(,),106,115
X,I4
@,53
0,157
e,157

SUBJECT

Addition, 12
of sets, 15
Adjoint, 150, 159
Alaoglu's theorem, 128
Alignment, 116
Allocation, 3, 202-203, 211,230-234
Approximation, 5-6, 55-58, 62-68, 71,
122-123, 165, 292
Arc length problem, 181, 185
Autoregressive scheme, 94
Baire's lemma, 148
Banach inverse theorem, 149
Banach space, 34
Bessel's inequality, 59
Bounded linear functional, 104
Bounded operator, 144
Bounded sequence, 13, 35
Bounded variation, 23-24, 113-115
Cartesian product, 14
Cauchy-Schwarz inequality, 30,47, 74
Cauchy sequence, 33
Chain rule, 176
Chebyshev approximation, 122-123
Closed graph theorem, 166
Closed set, 25
Closure point, 25
Codimension, 65
Compactness, 39-40
weak *, 127
Complement, xv, 25
orthogonal, 52, 117-118, 157
Complete orthonormal sequ(mce, 60-62
Complete set, 38
Complete space, 34
Concave functional, 190

Cone, 18-19
conjugate, 157
positive, 214
Conjugate cone, 157
Conjugate directions, 291
Conjugate functional, 196, 199,224
Conjugate gradients, 294-297
Conjugate set, 195-199
Continuity, 28
of inner product, 49
Continuous functional, 28
linear, 104
maximum of, 39-40
ContinuouslY differentiable, 175
Contraction mapping, 272
Control, 4-5, 66, 68-69,124,162-163,
205-206,211,228-229,254-265,
290,301
Controllability, 256
Convergence, 26-27
weak, weak*, 126
Convex combination, 43
Convex functional, 190
Convex hull, 18, 142
Convex mapping, 215
Convex programming, 216
Convex set, 17, 25, 131-137
Cosets,41
Covariance matrix, 80, 82
Dense set, 42
Differential, 9, 227
Frechet, 172
Gateaux, 171
Dimension, 20-22
Direct sum, 53
Discrete random process, 93

323

324

INDEX

Domain, 27. 143

Gradient, 175

Dual functional, 223

Duality,9,67, 72,120-122,134-137,
200-209,713-226,299-302,307
Dual space, algebraic, 105
nonned; 106
Dynamic model of random process, 95

projection of, 297-299

sub,237
Gram determinant, 56
Gram-Schmidt procedure, 54-55, 62-63,
293
Graph,166

Eidelheit separation theorem, 133

Epigraph, 192
Equivalent elements, 41
Estate planning, 182-183, 210
Estimation, 6,73, 78-102
Gauss-Markov, 84-87, 90-93,101
least-squares, 82-83
minimum variance, 87-93
of linear function, 91, 101
updating procedure for, 91-97,101
Euclidean space, 23
Euler-Lagrange equation, 180, 184
Expected value, 79
Extension of linear functional, 110

Hahn-Banach theorem, 8, 111, 120-122,

133
Half-space, 130
Hildreth's procedure, 299-300
Hb1der inequality,
Horse racing problem, 203-205
Hyperplane, 129

Farmer'S problem, 231-234, 265

Fenchel duality theorem, 201
Fixed point, 272
Fourier coefficients, 59, 62
Fourier series, 2, 58-63, 290-293
Frechet derivative, 175
Frechet differential, 172
FUnctional,28
cOrijugate, 196, 199,224
. continuous, 193-194
convex, 190
dual,223
linear, 104
Minkowski,131
primal,216
support, 135, 141-142, 197
Game theory, 7,206-209
Gateaux differential, 171
Gaussilih statistics, 78
Gauss-Markov estimate, 84-87, 90-93,
101
Generated subspace, 16
closed,60
Geodesic, 246

Image, 143
Implicitfunction theorem, 187,266
Infunum, xv
Infinite series, 58
Inner product, 46
continuity of, 49
Interior point, 24
relative, 26
Intersection, xv, 15
Inverse function theorem, 240, 266
Inverse image, 144
Inverse operator, 147
Isomorphic, 44
Kalman's theorem, 96
Kuhn-Tucker theorem, 249, 267
Lagrange multiplier, 188-189, 213, 243
Lagrangian, 213
Least-squares estimate, 82-83
Legendre polynominals, 61
Linear combination, 16
Linear convergence, 277
Linear dependence, 19-20, 53
Linear functional, 104
extension of, 110
Linear programming, 3, 232
Linear transformation, 28
Linear variety, 16-17,64-65
Mass spectrometer problem, 98
Minkowski-Farkas lemma, 167
Minkowski functional, 131
Minkowski inequalitY,23, 31, 33

INDEX

Minimum norm problemH, 50-52,

55-58,64-72,118-126,134-137,
160-165
Minimum-variance estima.te, 87-93
Min-max theorem, 208
Moments, 293
Motor control problem, 66-67, 124,
162-163
Moving average process, 94
Muntz's theorem, 74
Newton's method, 277-284
Norm, 22,47
minimum (see Minimum norm
problems)
of coset, 42
of linear functional, 10:5
of linear operator, 144
of product space, 37-38
Normal equations, 55-58, 160-161,292
Nullspace, 144
relation to range, 155
Null vector, 12
Oil drilling problem, 299--231
One-to-one, 27
Onto, 27
Open set, 24
Orthogonal complement, 52, 117-118,
157
Orthogonal projection, 53
Orthogonal vectors, 49. 117
Orthonormal set, 53
Orthonormal sequence, 59
Parallelogram law, 48-49, 51-52, 70
Parallel tangents, 296
Parseval's inequality, 75
Partition, 23
_
Penalty function, 302-307
Polynomial approximation, 72-75,
122-123
Pontryagin maximum principle, 261-265
Positive cone, 214
Positive semidefinite, 152
Pre-HUbert space, 46
Primal-dual method, 299-302, 306
Primal functional, 216, 303
Probability distribution, 79

325

Production scheduling, 4. 234-238

Product space, 14,37-38,44
Projection, 53
Projection operator, 167
Projection theorem, 8,46,51-52,64,
69, 120
Pseudo inverse, 83, 163
Pythagorean theorem, 49
Quadratic convergence, 281
Quadratic programming, 225-226,
299-300
Quotient space, 41-42
Random variable, 79
Range, 143
closed, 156
relation to nullspace, 155-157
Reflexive space, 116
Regression, 99-101
Regular point, 187,219,240,244,248,
256
Relatively closed, 26
Relative minimum, 177
Reproducing kernel Hilbert space, 72
Riccati equation, 258
Riesz-Frechet theorem, 109
Riesz representation theorem, 113
Rocket problem,S, 125-126, 138-139,
259-261
Saddle-point condition, 219, 221
Scalar, 11
Scalar mUltiplication, 12
Self-adjoint, 152
Semicontinuity, 40, 194, 306-307
Seminorm, 45
Separable space, 42
hyperplane theorem, 133
Sequence, xvi
bounded, 13, 35
Cauchy, 33
complete, 60-61
convergent, 26
finitely nonzero, 13, 105
orthonormal, 59,61
Series, 58
Sphere, 24
Stationary point, 178

326

INDEX

Steepest descent, 285-290, 296

Subgradient, 237
functional, 110
Subspace, 14
Successive approximation, 272
Sum of sets, 15
direct, S3
Support functional, 135, 141-142, 197
Supporting hyperplane, 133
Supremum, xv
Tangent space, 242
Topological properties, 11
Total variation, 24, 115
Transformation, 27, 143
con';nuous, 28. 173
linear, 28
Triangle inequality, 1
Triangulation problem, 6, 97-98

Unbiased estimate, 85
Union, xv
Updating a linear estimate, 91-93,

101-102
Variance, 79
Vector space, 11-12
Vertical axis, 192
Weak continuity, 127
Weak convergence, 126
Weierstrass approximation theorem,

42-43,61
Weierstrass-Erdman corner conditions,

210
Weierstrass maximum theorem, 39-40,

128
Young's inequality, 211